content
stringlengths 1
15.9M
|
|---|
\section{Introduction}\label{sec:intro}
Archives and libraries throughout the world hold billions of
historical manuscripts. Many of these documents are already digitized
into images, but their access is limited because the contents are not
available in a symbolic format that would allow modern treatment of
textual matters such as editing, indexing, and retrieval. Handwritten
Text Recognition (HTR) is the cornerstone in this situation which
aims to provide automatic ways of transcribing these
documents~\cite{muehlberger2019transforming}.
In classical HTR laboratory experiments, the text lines are assumed
to be given. Therefore, the performance is evaluated at the line
level. Traditional evaluation measures for line-level HTR are the
Character Error Rate (CER) and the Word Error Rate (WER), borrowed
from the Automatic Speech Recognition field. These metrics indicate
the length-normalized number of elementary editing operations needed
to produce a reference (correctly transcribed) sequence from the HTR
hypothesis, at the character (CER) or word (WER) level. Under the
premise of a line-level formulation, it is generally acknowledged that
these metrics provide a good measure of performance.
Due to recent advances in the field, especially brought about by the
intensive use of deep neural networks, line-level HTR is considered
practically solved, or close to. Therefore, the field is experiencing
a paradigm shift towards end-to-end full-page scenarios. In a
page-level application, lines are not given. Instead, images are
usually processed to first extract single lines, under a process
generally known as Layout Analysis (LA).\!%
\footnote{Many present-day HTR systems use simplified forms of LA
which only focus on detecting the text-lines of each image. In the
sequel, the term LA will be used indistinctly to refer to proper LA
as well as to just line detection.}
Then, each line is transcribed independently with line-level
HTR. Furthermore, recent works do not explicitly include any LA step
and aim to obtain the transcription hypothesis by processing whole
pages or paragraphs~\cite{Bluche2016,Coquenet2021,Singh2021,Peng2022}.
Despite moving from the line-level to the page-level HTR scenario, the
traditional CER and WER metrics are still generally used for
assessment. However, this evaluation protocol is too naive: full-page
real applications do suffer from LA errors which systematically lead
to inconsistencies when evaluating the model using such metrics.
Figure \ref{fig:la-errors-for-htr} shows a real example of this kind
of issues related to LA (see other examples in
Figs.\,\ref{fig:complexExamples},~\ref{fig:splitLines},
and~\ref{fig:evalExample}). While all the words are perfectly
recognized, the WER is 70\%, which is absolutely misleading. Clearly,
if this figure is meant to reflect anything, it is a LA problem ---
nothing related with word recognition errors!
This kind of problems become even more insidious in approaches that
bypass the LA step. When researchers were hard-pressed to obtain
acceptable performance values, questioning the traditional evaluation
protocol did not seem relevant. However, with an increasing number of
effective page-level transcription workflows, we see the need to ask
ourselves about the nature of its evaluation and whether the
traditional line-level evaluation faithfully represents a proper
indicator of page-level transcription performance.
\begin{figure}
\vspace{-0.2em}
\def-.5ex{-0.3ex} \def-.2ex{+2.6ex} \newcommand{\f}[1]{\textsf{\small #1}}
~~~\fbox{\scalebox{.91}{
\begin{minipage}[t]{.25\textwidth}
\f{Two ways of coming}\\[-.5ex]
\f{at the (archetypes of)}\\[-.5ex]
\f{Geometrical abstract}\\[-.5ex]
\f{Quantities: 1. by}\\[-.5ex]
\f{decomposing Bodies:}
\end{minipage}
\begin{minipage}[t]{.26\textwidth}
1. application of\\[-.5ex]
metaphisics to\\[-.5ex]
mathematics.\\[-.5ex]
2. Method of facilitating\\[-.5ex]
the Study of mathematics
\end{minipage}}}~
\begin{minipage}[t]{0.4\textwidth}
\emph{Two-columns reference transcript}
\end{minipage}
\\[-1.6em]
\begin{minipage}[t]{\textwidth}
~\flushright\emph{Automatic transcript (WER=70.0\%)}~~~
\end{minipage}\\[-0.6em]
\raisebox{-17pt}{~~~\includegraphics[width=0.5\textwidth]{exampleRO}}
\fbox{\scalebox{.91}{
\begin{minipage}[b]{.49\textwidth}
\f{Two ways of coming 1. application of}\\[-.2ex]
\f{at the (archetypes of) metaphisics to}\\[-.2ex]
\f{Geometrical abstract mathematics.}\\[3ex]
\f{2. Method of facilitating}\\[-0.2ex]
\f{Quantities: 1. by}\\[1ex]
\f{decomposing Bodies: the Study of mathematics}\\[-0.6em]
\end{minipage}}}
\vspace{-0.8em}
\caption{\label{fig:la-errors-for-htr} Example of misleading WER
evaluation, caused by wrong reading order due to text-line
detection flaws.
%
While all the words in the automatic transcript are perfectly
correct, the WER is 70\% (13 matching words, 13 substitution
errors, 4 insertions and 4 deletions).
%
The image is part of a page of the Bentham Papers collection
(see Sec.\,\ref{sec:ro}).
}
\vspace{-1em}
\end{figure}
The difficulties underlying the evaluation of page-level HTR results
boil down to a Reading Order (RO)
problem~\cite{breuel2003,malerba2008,clausner2013,Pletschacher2015,%
naoum2019,quiros2022}.
A number of recent proposals try to heuristically weight and combine
both word recognition and LA geometric errors into a single scalar
value~\cite{LeifertLGL19,coquenet2022dan}. Unfortunately, this
hinders the capability to sort out the nature of the corresponding
errors and thereby making a comprehensive, useful assessment.
Here we instead advocate for a two-fold evaluation approach which
decouples the impact of word (and character) recognition errors from
the influence of wrong RO and, furthermore, it is largely agnostic to
geometry-related flaws.
One possibility to assess page-level word recognition accuracy
regardless of RO is to use the so-called Bag-of-Words Accuracy (bWAC)
or Error Rate, as proposed and used in early works by
Antonacopoulos, Clausner and Pletschacher~\cite{Antonacopoulos2013,%
Pletschacher2015,Clausner2017,ClausnerAP19}, and later also by other
authors~\cite{Strobel2020}.
Even though the bWAC has proved a useful assessment metric in a number
of cases, we will argue that a proper WER based on the bag-of-words
concept can not be trivially defined as $1-$bWAC, neither by just
counting how many words do not appear both in the reference and HTR
transcripts. So we (re-)define a bag-of-words WER (bWER) so that it
becomes faithfully comparable with the traditional WER and proves to
be a very convenient page-level RO-independent word error metric.
However, the bag-of-words WER approach does not allow measuring
character-level error, nor it provides the word alignment information
needed to compute RO assessment metrics. Instead, both word- and
character-level RO-independent recognition accuracy can be precisely
computed using the well-known Hungarian Algorithm
(HA)~\cite{kuhn1955hungarian,burkard2012assignment}.
Here we introduce a regularized version of the HA which provides
HA-based WER values (hWER) that are almost identical to those of bWER
and, moreover, are also practically equal to those of the classical
WER when the reference and HTR transcripts are in the same RO. In
addition, it further provides the information needed to compute RO
assessment metrics such as the Normalized Spearman's Footrule Distance
(NSFD) \cite{ravi2010,quiros2022}.
In this work, we study all these related approaches to separately
assess at the page-level both the HTR word (and character) recognition
accuracy and the quality of the RO . The problems considered and the
proposed solutions will be presented along with empirical results
obtained on a semi-artificial task, where the typically expected LA
errors and associated RO problems are simulated. The proposed
assessment methods will be then applied to a series of real page-level
end-to-end HTR experiments, considering both LA-based and holistic
page-level transcription approaches.
Our experiments will show that: i) in the traditional line-level
setting, bWER and WER are typically almost identical; ii) the WER
based on the regularized HA is almost the same as the bWER and both
accurately approach page-level WER in the traditional line-level
evaluation setup; iii) the difference between WER and bWER highly
correlates with the NSFD and is much more efficient than using the HA,
needed to compute the NSFD.
The remainder of this work is structured as follows. Classical WER and
CER measures are reviewed in Section \ref{sec:wcer}; the RO problem
and the NSFD measure are discussed in Section \ref{sec:ro}; the
proposed bWER and hWER metrics are described in Section \ref{sec:bow}
and \ref{sec:hungarian}, respectively; and a summary of the
different metrics considered is provided in Section
\ref{sec:summMetrics}. Then, simulated and real experiments are
reported and analyzed in Section \ref{sec:simuExp} and
\ref{sec:realEval}, respectively. We close the article by outlining
related works in Section \ref{sec:related} and concluding in Section
\ref{sec:concl}. Finally, \ref{ap:ex} presents detailed examples of
the computation of the different metrics proposed and
\ref{ap:software} provides details for public access to the datasets
and software tools used and developed in this work.
\section{Word \& Character Error Rates Based on the Edit Distance }
\label{sec:wcer}
Traditional HTR experiments are based on given line-level images,
typically produced or supervised manually as part of the generation of
ground truth (GT) of the text image collection considered. As
discussed below, WER and CER can then be easily computed for each
line and the line-level word and character error counts can be
accumulated for all the lines of a page image to straightforwardly
compute page-level WER and CER.
Of course, this approach neglects possible line detection and/or
extraction errors made by the LA stage in real automatic transcription
tasks.
This section reviews this traditional HTR evaluation approach, as an
introduction to the forthcoming sections, where we propose new
approaches for fair page-level end-to-end HTR assessment.
\subsection{Edit Distance, WER and CER for Word Sequences}
\label{sec:basicED}
Let $x$ and $y$
reference text and $y$ an HTR hypothesis.
The word \emph{edit distance}~\cite{wagner74} from $x$ to $y$,
$\edit(x,y)$, is the minimum number of word insertion, substitution
and deletion edit operations that transform $x$ into $y$.
Edit operations define a ``trace'' or alignment between word instance
positions of $x$ and $y$, which may be formulated in several
equivalent ways. Here we loosely follow the work of Marzal and
Vidal~\cite{marzal93} and define an alignment $\mathcal{A}(x,y)$ as a
sequence of ordered pairs of integers (word indices), $(j,k)$,
$1\!\leq\!j\!\leq\!|x|,~1\!\leq\!k\!\leq\!|y|$, such that for every
two distinct pairs
$(j,k),(j',k')\in\mathcal{A}(x,y),~j\!<\!j'\Leftrightarrow k\!<\!k'$.
In what follows, word alignments which fulfill this sequentiality
constraint will be denoted as $\mathcal{T}(\cdot,\cdot)$, leaving the
notation $\mathcal{A}(\cdot,\cdot)$ only for unconstrained alignments.
$\mathcal{T}(x,y)$ can be conveniently extended to explicitly
represent word insertions and deletions. To this end, a \emph{dummy}
position, denoted by $\epsilon$, is assumed in both $x$ and $y$ which
points to the \emph{``empty word''}, $\lambda$; that is,
$x_\epsilon=y_\epsilon\eqdef\lambda$.
The edit distance from $x$ to $y$ is thus formally defined as:
\vspace{-0.2em}
\begin{equation}\label{eq:editDist}
\edit(x,y) ~=~ \min_{\mathcal{T}(x,y)}
\sum_{(j,k)\in\mathcal{T}(x,y)}\!\!\! \updelta(x_j,y_k)~~~~~~~~
\vspace{-0.3em}
\end{equation}
where $\updelta(a,b)$ is defined to be $1$ if $a\neq b$ and $0$ otherwise.%
\footnote{If the cost of each edit operations is 1, as it is here
assumed, $\edit(\cdot,\cdot)$ is also called Levenshtein distance.}
To allow the above sequentiality constraint to still be meaningful, we
assume that for any $j,j'$ where $j$ or $j'$ are $\epsilon$, the
predicate $j\!<\!j'$ is \emph{true}.
By analyzing the pairs in the optimal trace $\mathcal{T}(x,y)$, the
sum in Eq.\,\eqref{eq:editDist} can be decomposed into separate counts for
insertions, substitutions and deletions; i.e., ~$\edit(x,y)=i+s+d\,$.
Example~1 in \ref{ex:1} illustrates the computation of of the
word edit distance and the corresponding trace for $x=\,\,$\textsf{\small
``To be or not to be, that is the question''} and $y=\,\,$\textsf{\small
``to be oh! or not to be: the question''},
with $i=1, s=2, d=2\,$ and $\,\edit(x,y)=5$.
Because of the sequentiality constraint, edit operations are applied
sequentially along the word sequences $x$ and $y$. This allows the
optimal $\mathcal{T}(x,y)$ according to~Eq.\,\eqref{eq:editDist}, the
corresponding counts $i,s,d$, and thereby $\edit(x,y)$, to be easily
computed by Dynamic Programming in $O(|x|\,|y|)$~\cite{wagner74}.
The WER of $y$ with respect to $x$ is defined as the edit distance,
normalized by the length of the reference text, $n=|x|$:
\vspace{-0.7em}
\begin{equation}\label{eq:lineWER}
\WER(x,y) ~=~ \frac{\edit(x,y)}{n}
~\equiv~ \frac{i+s+d}{c+s+d}~~~~~~~
\end{equation}
where $c$ is the number of correct words (those which do not need
editing).
In Example~1~(\ref{ex:1}), $\WER(x,y)$ $=(1+2+2)/(6+2+2)=5/10=50\%$.
Note that, defined in this way, it may happen that $\WER(x,y)\!>\!1$,
which prevents WER to be properly interpreted as an error probability.
For the same reasons a Word Accuracy can not be defined just as $1-\WER$.%
\footnote{The Normalized Edit Distance~\cite{marzal93,vidal95} would
overcome these drawbacks but, following time-honored tradition in ASR
and HTR alike, we stick with the conventional normalization
by the length of the reference sequence.}
The CER is defined similarly, by just assuming that $n$ is the total
number of characters in $x$ and $i,s,d,c$ are \emph{character}, rather
than \emph{word} edit operations and correct matching counts. In
Example~1, $\CER(x,y)=(4+2+8)/(30+2+8)=14/40=35\%$.
\subsection{Traditional, Line-based Page Level WER and CER}
\label{sec:traditionalWerCer}
Let $\mathcal{I}$ be a text image and $X$ the reference GT transcript
of $\mathcal{I}$. Let $Y$ be the transcription hypothesis provided by
an HTR system for $\mathcal{I}$. Both $X$ and $Y$ are made up of the
same number $M$ of individual text-lines $x^1,x^2,\dots x^M$ and
$y^1,y^2,\dots y^M$, respectively, where each text-line is a sequence
of words.
Each pair of text-lines $x^\ell$ and $y^\ell$ are transcripts of the
same image line, which is simply denoted as
$\ell,1\!\leq\!\ell\leq\!M$.
In the traditional setting, page-level WER is then computed as:
\begin{equation}\label{eq:pageLineWER}
\WER(X,Y) ~=~ \frac{\sum_{\ell=1}^M \edit(x^\ell,y^\ell)}{N}
~\equiv~ \frac{\sum_{\ell=1}^M(i_\ell+s_\ell+d_\ell)}
{\sum_{\ell=1}^M(c_\ell+s_\ell+d_\ell)}
\end{equation}
where $N=|X|$ is the total number of word instances of $X$.
Another way to compute $\WER(X,Y)$ is to concatenate all the $M$ lines
of $X$ and $Y$ in any arbitrary order (the same order for $X$ and $Y$)
and directly compute the edit distance between the concatenated
texts. Except for small possible differences in the text-line
boundaries, the editing operations obtained by this computation will
be essentially the same as those involved in the $M$ edit distances
$\edit(x^\ell,y^\ell), 1\!\leq\!\ell\!\leq\!M$ of
Eq.\,\eqref{eq:pageLineWER}. Therefore:
\begin{equation}\label{eq:fullPageWER}
\WER(X,Y) ~\approx~ \frac{\edit(X,Y)}{N}
~\equiv~ \frac{i'+s'+d'}{c'+s'+d'}
\end{equation}
where $i',s',d'$ and $c'$ are now counts of word edit operations and
matchings involved in the computation of $\edit(X,Y)$ for the
\emph{whole} texts $X$ and $Y$.
As in Sec.\,\ref{sec:basicED}, the CER is defined similarly by just
assuming that $N$ is the total number of characters in $X$ and
$i,s,d,c,i',s',d',c'$ are \emph{character}, rather than \emph{word}
edit operation and correct matching counts.
The computational cost of Eq.\,\eqref{eq:pageLineWER}
and~\eqref{eq:fullPageWER} is dominated by the quadratic cost of
computing the edit distance.
This results in $O(Mm^2)$ for Eq.\,\eqref{eq:pageLineWER} and $O(N^2)$
for Eq.\,\eqref{eq:fullPageWER}, where $M$ is the number of text
lines, $m$ is the length of the longest line transcript, and $N$ is the
length of a full-page transcript.
\vspace{-0.3em}
\subsection{Page-level End-To-End Assessment Using Traditional WER and CER }
\label{sec:traditEndToEnd}
\vspace{-0.2em}
In a realistic scenario, image-lines may be given for the GT reference
transcript, $X$. But these lines may not correspond one-to-one with
lines automatically detected in the text image $\mathcal{I}$.
Moreover, the number of text-lines in $X$ and $Y$ might be different.
To overcome this hurdle, it is often ignored that the lines of $Y$ may
\emph{not} be in the same \emph{reading order} (RO) as those of $X$
and the WER is thus naively computed for the whole texts in $X$ and
$Y$ as in Eq.\,\eqref{eq:fullPageWER}.
This is the approach often followed in experiments which aim to
provide end-to-end performance assessment such
as~\cite{Bluche2016}~\cite{Peng2022} and~\cite{sanchez2019} (Sec.8,
Test-B2).
\section{The Reading Order Problem}\label{sec:ro}
Unfortunately, assuming that reference and hypothesis transcripts are
in the same RO is generally unrealistic because lines automatically
detected in a text image can seldom be provided in the same RO as the
lines of the reference transcript. Moreover, LA often fails to detect
some existing text lines or wrongly detects nonexistent lines. This
is in fact a typical situation for most historical handwritten text
images which generally exhibit complex and non-uniform layouts and
suffer from many types of distortions and noise.
Fig.\,\ref{fig:complexExamples} shows examples of these difficulties.%
\footnote{From the Bentham Papers collection. See, e.g.:
\url{http://prhlt-kws.prhlt.upv.es/bentham}}
\begin{figure}[htbp]
\centering
\includegraphics[height=.36\linewidth]{172_152_001crop}\\[0.35em]
\includegraphics[height=.26\linewidth]{134_034_002crop}%
\hspace{.5em}%
\includegraphics[height=.26\linewidth]{544_049_001crop}
\vspace{-0.5em}
\caption{Examples of frequent reading order issues, from the Bentham
Papers collection.}
\label{fig:complexExamples}
\vspace{-0.5em}
\end{figure}
As discussed in Sec.\,\ref{sec:traditEndToEnd}, a naive approach is to
just apply Eq.\,\eqref{eq:fullPageWER} to the full page transcripts
$X$ and $Y$, ignoring the RO issues possibly caused by LA errors.
In this scenario, the sequential restriction assumed for the
computation of WER will prevent to match those text fragments of $Y$
which are not in the RO given by $X$.
Consequently, the resulting WER figure will reflect an uncontrolled
combination of actual word recognition failures and errors due to
inaccurate RO.
Furthermore, it is important to realize that ``the correct'' RO
of a text image is seldom unique. The RO provided by reference
transcripts and/or other layout GT annotations is generally only one
among several possible RO annotations which would be all correct.
Therefore, mixing RO and word recognition errors into a single
assessment measure (as in~\cite{coquenet2022dan,LeifertLGL19}) does
mot seem the best idea for understanding which are the inner issues of
an end-to-end full-page HTR system.
These facts lead us to propose a two-folded evaluation approach which
completely decouples the RO from word recognition errors, while also
providing a simple, comprehensive picture of the end-to-end system
performance.
Research on RO has a relatively long tradition for printed
documents~\cite{lee2002,breuel2003,ceci2007,malerba2008,%
clausner2013,naoum2019,lee2021}.
More recently, RO analysis has also been considered for handwritten
documents, where RO issues are specially relevant. In the work of
Quiros and Vidal~\cite{quiros2021,quiros2022}, effective methods to
learn line RO in handwritten text images from examples are proposed
and empirically assessed.
\subsection{Assessing Reading Order: Normalized Spearman's Footrule Distance}
\label{sec:roAssess}
Evaluating the reading order is a non-trivial task that requires
measuring not only how many elements are not placed in the correct
position within the expected order, but also how far those elements
are from their correct positions. This is thoroughly discussed
in~\cite{quiros2021,quiros2022}, where two metrics are finally
proposed and used in the experiments: the Kendall's Tau rank distance
(also called bubble-sort distance)~\cite{kendall1938} and the
Normalized Spearman's Footrule Distance (NSFD)~\cite{ravi2010}. Here
we adopt the latter because it provides reasonable estimates of the
human effort that would be needed to render a sequence of elements in
a correct order given by a reference sequence.
The NSFD can be defined as:
\begin{equation}\label{eq:nsfd}
\rho(X,Y) ~=~ \frac{1}{\lfloor\frac{1}{2}N^2\rfloor}
\!\sum_{(j,k)\in\mathcal{A}(X,Y)}\!\!\! |j-k|
\vspace{-0.5em}
\end{equation}
where $\mathcal{A}(X,Y)$ is an alignment between the reference text
$X$ and the HTR hypothesis $Y$, $N=\max(|X|,|Y|)$ is the number of
words of the longest text.
Note that the alignment $\mathcal{A}(X,Y)$ does \emph{not} need to
fulfill the sequentiality constraint used in Sec.\,\ref{sec:basicED}
to define the word edit distance. In what follows, we
assume that the alignment used in Eq.\,\eqref{eq:nsfd} will be
provided by the methods discussed in Sec.\,\ref{sec:hungarian}.
From a user point of view, insertions and deletions do not typically
affect the RO in a substantial way. Therefore, in Eq\,\eqref{eq:nsfd}
we just assume that $|j-\epsilon|=|\epsilon-k|\,\eqdef\,1~\forall
j,k$.
However, insertions and deletions may indirectly affect significantly
the result of Eq\,\eqref{eq:nsfd}, because of the contribution of
subsequent values of $|j-k|$. This is illustrated in Example~2
(\ref{ex:2}), along with the approach we propose to circumvent this
problem by just renumbering the positions of words of $Y$ and/or $X$
according to the inserted or deleted words specified in
$\mathcal{A}(X,Y)$.
The sum in Eq\,\eqref{eq:nsfd} can only be larger than
$\lfloor\frac{1}{2}N^2 \rfloor$ if the alignment has many insertions
and/or deletions as compared with $N$. Therefore, except for
miscellaneous cases (for instance when $N$ is very small),
$\rho(\cdot,\cdot)$ is usually bounded between $0$ and $1$.
The computational cost of Eq.\,\eqref{eq:nsfd} is $O(N)$, in addition
to the cost to obtain $\mathcal{A}(X,Y)$.
Example~2 in \ref{ex:2} illustrates the computation of the NSFD for
$X\!=\,\,$\textsf{\small ``To be or not to be, that is the question''}
and $Y\!=\,\,$\textsf{\small ``The big question: to be or not to be''},
with $\rho(X,Y)=27/50~ (54\%)$.
\vspace{-0.6em}
\section{Bag of Words WER}\label{sec:bow}
\vspace{-0.2em}
In Sec.\,\ref{sec:wcer}, $X$ and $Y$ were considered sequences where
the order of text-lines and words is relevant for computing word
errors. However, in page-level performance assessment, once we have a
specific metric to measure RO, it is desirable to largely ignore the
order of words in $X$ and $Y$ to measure word recognition performance.
A simple way to achieve this goal is to rely on the ``Bag of Words''
concept~\cite{Antonacopoulos2013,Strobel2020}.
To this end, $X$ and $Y$ are now considered multi-sets (or ``bags'')
of words and the number of instances of each word can be used to
compute a metric which is fairly closely related to the WER.
Let $V_X$ and $V_Y$ be the respective sets of different words
(vocabularies) of $X$ and $Y$, and $V=V_X\cup V_Y$. For each word
$v\in V$ let $f_X(v)$ and $f_Y(v)$ be the number of instances of $v$
in $X$ and $Y$, respectively. The ``bag of words distance'' between
$X$ and $Y$ is defined as:
\vspace{-0.2em}
\begin{equation}\label{eq:BoWdist}
\mathrm{B}(X,Y) = \sum_{v\in V}|f_X(v)-f_Y(v)|~~~~~
\vspace{-0.5em}
\end{equation}
\noindent
Then, if $N$ is the number of words in the reference $X$, a simple
``BoW WER'' can be rather naively defined as
$\beta\!\WER(X,Y)=B(X,Y)/N$.
As defined in Eq.\,\eqref{eq:BoWdist} $\mathrm{B}(X,Y)$ is the number
of word instances of $X$ which fail to appear in $Y$ plus the number
of word instances in $Y$ which are not in $X$. This can be properly
interpreted in terms of editing operations just as the total number of
word insertions and deletions that would be needed to transform $X$
into $Y$, \emph{without allowing for word substitutions}.
In the classical WER, a combined deletion and insertion pair of edit
operations can be achieved by a single substitution. So, if $X$ and
$Y$ are in the same RO, the bag of words distance will always be
larger than or equal to the corresponding word edit distance; that is,
$\mathrm{B}(X,Y)\geq\edit(X,Y)$.
If word substitution were allowed, many pairs of the $\mathrm{B}(X,Y)$
insertions and deletions could be advantageously exchanged by a single
substitution. In the best case, the number of these word
substitutions would be exactly $\mathrm{B}(X,Y)/2$.
However, if $|X|\neq|Y|$ there must be a number
$b=\big|\,|X|-|Y|\,\big|$ of insertions and/or deletions which are
``unavoidable''.
Therefore, to define a ``bag of words WER'' which can be fairly
compared with the traditional WER, we assume that each
insertion/deletion pair, except those ``unavoidable'', is equivalent
to a single substitution.
Formally speaking, the above definition of bag of words distance needs
to be revamped into
$\mathrm{B}'(X,Y)=b+\lfloor(\mathrm{B}(X,Y)-b)/2\rfloor$. Since
$\mathrm{B}(X,Y)-b$ is always even, the \emph{bag of words WER} is
thus defined as:
\vspace{-0.3em}
\begin{equation}\label{eq:werBoW}
\bWER(X,Y) ~=~ \frac{\mathrm{B}'(X,Y)}{N}
~=~ \frac{1}{2N}\,\biggl(\,b~+~\sum_{v\in V}|f_X(v)-f_Y(v)|\,\biggr)
\vspace{-0.5em}
\end{equation}
where, as previously discussed, $N=|X|$ and $b=\big|\,N-|Y|\,\big|$.
The word frequencies $f_X(v)$ and $f_Y(v)$ can be easily obtained by
hashing in $O(|X|+|Y|)$. So, if $N=|X|$ and $|Y|$ is $O(N)$, the bWER
(and the $\beta$WER alike) can be computed in $O(N)$.
Through the computation of Eq.\,\eqref{eq:werBoW}, the \emph{number}
of word insertions, deletions and (implicit) substitutions can be
easily derived, even though which specific words are involved in the
different operations remain unknown.
This becomes a significant drawback, because it prevents to derive any
kind of word-to-word or position-to-position alignment that could be
used to compute the NSFD or any other metric to assess RO mismatch.
For the same reason, a CER associated with bWER can neither be
properly computed. (Note that a ``bag of characters'' error rate
would be overtly deceptive, and therefore is not an option).
The examples in \ref{ex:3} illustrate the computation of $\beta\,$WER
and the reformulated version here proposed bWER (Eq.\,\eqref{eq:werBoW}),
along with their relation with the classical WER.
It could be thought that the bWER defined in~Eq.\,\eqref{eq:werBoW} is
identical to or a very tight bound of the classical WER. However, it
should be taken into account that the WER is based on the edit
distance, which is computed under the sequentiality constraint
(cf. Sec.\,\ref{sec:basicED}). Since this restriction is fully
relaxed for bWER, the values of $\bWER(X,Y)$ can actually be much
lower than those of $\WER(X,Y)$, specially if the RO of $X$ and $Y$
are very different.
Even without RO issues, in the worst case, the bWER can considerably
underestimate what could be considered ``true'' word errors. A simple
case is also given in \ref{ex:3} (Example 3a). Nevertheless,
according to the empirical evidence presented in
Secs.\,\ref{sec:simuExp}--\ref{sec:realEval}. these cases are
extremely rare in practice.
These empirical results also show that, in the traditional
experimental setting, where RO is not an issue (because the detected
text lines are ensured to be all correct), the page-level bWER
\eqref{eq:werBoW} is in practice a very good approximation to the
corresponding WER given by~\eqref{eq:pageLineWER}
and~\eqref{eq:fullPageWER}.
This is interesting because bWER is much simpler and cheaper to
compute than WER.
\section{CER, WER and NSFD Based on Bipartite Graphs and the\\[-4pt]
Hungarian Algorithm}
\label{sec:hungarian}
\vspace{-0.2em}
As discussed above, determining RO-independent word and character
recognition accuracy at the full-page level, requires words and/or
word positions from the reference transcript $X$ to be freely aligned
or paired with corresponding words of the transcription hypothesis
$Y$.
Edit distance computation provides word alignments (traces) as a
byproduct, but the trace sequentiality restriction leads to alignments
which lack the freedom needed for RO-independent word pairing.
A proper formulation of the required kind of word alignments is given
by the so-called \emph{``minimum-weight matching or assignment
problem''}~\cite{burkard2012assignment,bijsterbosch2010solving}.
\bigskip\bigskip
Let $G=(V,E)$ be a \emph{bipartite graph}, where the set of
nodes $V$ is composed of two disjoint subsets $A,B$, $A\cup B\!=\!V$,
$A\cap B\!=\!\emptyset$, and the set of edges $E$ is a subset of
$V\!\times V$ such that
$(u,v)\in\!E\Rightarrow u\in\!A\,\wedge\,v\in\!B$.
A \emph{matching} $\mathcal{M}\!\subset\!E$ is a set of pairwise
non-adjacent edges; that is, no two edges share a common node.
A node is matched if it is an endpoint of one of the edges in the
matching. Otherwise, the node is unmatched.
$\mathcal{M}$ is said to be \emph{maximum} if it contains the largest
possible number of edges
and it is a \emph{perfect matching} if all the vertices of the
graph are matched.
Every perfect matching is also maximum.
A bipartite graph $G=(V,E)$ is \emph{weighted} if a real-valued
weight $g(u,v)$ is assigned to each edge $(u,v)\in\!E$. Then, the
weight of a matching $\mathcal{M}$ is the sum of the weights of the edges
in $\mathcal{M}$.
Given a weighted bipartite graph, the assignment problem is to find a
perfect matching with minimum weight.
An efficient solution to this problem is provided by the
\emph{Hungarian Algorithm} (HA) which runs in $O(N^3)$ time, where $N$
is the number of nodes in $A$ and $B$~\cite{kuhn1955hungarian}.
While the HA does not formally need the subsets $A$ and $B$ to be of
the same size, most of its implementations do assume equal sizes. In
any case, it is straightforward to support different sizes by adding
\emph{``dummy''} nodes to either (or both)
set(s)~\cite{burkard2012assignment,bijsterbosch2010solving,%
crouse2016implementing}.
\medskip
In our HTR assessment task, $A$ and $B$ are, respectively, the word
instances of the reference transcript $X$ and the HTR hypothesis $Y$
of a page image; $E\!=\{\!(X_j,Y_k), 1\!\leq\!j\!\leq\!|X|,
1\!\leq\!k\!\leq\!|Y|\}$ is the set of all pairs of word instances in
$X$ and $Y$, and
the weight $g(X_j,Y_k)$ is the character edit distance between the
$j$-th word of $X$ and the $k$-th word of $Y$.
Word insertions and deletions are represented by assignments to dummy
nodes, which represent the empty word $\lambda$. These nodes need to
be added to \emph{both} sets, not only because in general
$|X|\neq|Y|$, but also because we need to simultaneously support
\emph{both} \emph{insertions} and \emph{deletions} for any given pair
of transcripts.
The cost of an edge connecting a dummy node with a word $v$ is thus
defined as $g(v,\lambda) = g(\lambda,v)=\frac{1}{2}|v|$, where $|v|$
is the number of characters of $v$ and, as in Sec.\,\ref{sec:bow}, the
factor $\frac{1}{2}$ is introduced to balance the cost of a word
substitution with that of an equivalent combined word insertion and
deletion.
The assignment problem is to pair each (maybe empty) word instance of
$X$ with a (maybe empty) word instance of $Y$ so that the sum of
character edit distances between the paired words is minimum.
Therefore, the HA yields what could be called ``HA Character Edit
Distance'':
\begin{equation}\label{eq:hungCER}
\mathrm{d_h}(X,Y) ~=~ \min_{\mathcal{A}}
\!\!\sum_{(j,k)\in\mathcal{A}(X,Y)}\!\! g(X_j,Y_k)\,;
\vspace{-0.3em}
\end{equation}
The optimal alignment $\hat{\mathcal{A}}(X,Y)$ associated with
Eq.\,\ref{eq:hungCER} is a set of pairs $(j,k)$,
$1\!\leq\!j\!\leq\!|X|\!=\!N$, $1\!\leq\!k\!\leq\!|Y|$, along with two
additional sets of pairs of the form $(j,\epsilon)$ and $(\epsilon,k)$
to account for word deletions and insertions, respectively.
Let $D$ be the number of these dummy pairs in $\hat{\mathcal{A}}(X,Y)$
and, as in Eq.\,\eqref{eq:werBoW}, let $b=\big|\,|X|-|Y|\,\big|$.
Since both insertions and deletions are allowed in
$\hat{\mathcal{A}}(X,Y)$, $D\geq b$. So, as in the case of
Eq.\,\eqref{eq:werBoW} for the bWER, the (now typically few) $D-b$ excess
pairs of insertions and deletions, can be interpreted as single
substitutions.
Then, the ``HA WER'' (hWER) can be defined as:
\vspace{-0.5em}
\begin{equation}\label{eq:hungWER}
\hWER(X,Y) ~=~ \frac{1}{N}\!
\!\!\!\sum_{\:\:\:(j,k)\in\hat{\mathcal{A}}(X,Y)}
\!\!\!\updelta(X_j,Y_k) ~~-~~ \frac{(D - b)}{2N}
\vspace{-0.2em}
\end{equation}
where $\updelta(\cdot,\cdot)$ is the $0/1$ function introduced in
Sec\,\ref{sec:basicED}.
Also using $\hat{\mathcal{A}}(X,Y)$, the NSFD $\rho(X,Y)$ can be
computed straightaway as in Eq.\,\eqref{eq:nsfd}.
When comparing hWER with bWER, note that the optimization of
Eq.\,\eqref{eq:hungCER} ensures a word alignment with minimum sum of
\emph{character edit distances} between the paired words. But this
alignment may not always lead to a minimum \emph{word edit distance}.
Thus, while it can be easily shown that
$\bWER(X,Y)\leq\hWER(X,Y)~\forall X,\!Y\!$,\, the strict equality may
not hold in some cases.
\medskip
Consider for example $X\!\!=\,$\textsf{\small''xx a ba yy''},
$Y\!\!=\,$\textsf{\small''xx ac a yy''}, with
$\,\bWER=\lfloor\,0+(2-0)/2\rfloor/4=25\%$
The minimum sum of character errors is $2$, but it can be achieved
through two different alignments:
$(1,1),(2,3),(3,2),(4,4)$, with
$\,\hWER(X,Y)\!=\!1/4\!=\!25\%\!=\!\bWER(X,Y)$ and
$(1,1),(2,2),(3,3),(4,4)$, with
$\,\hWER(X,Y)\!=\!2/4\!=\!50\%\!>\!\bWER(X,Y)$.
\medskip
The examples in \ref{ex:4} further illustrate the
computation of \,hWER for the more realistic texts used in
Example~3.
It is worth noting that the values of hWER in these examples are
identical to the corresponding bWER values of Example~3~(\ref{ex:3}).
When multiple instances of some word exist in $X$ and/or in $Y$, as in
the examples of \ref{ex:4}, the HA is free to pair any matching
instances, as long as the values of $\mathrm{d_h(X,Y)}$ are the same.
In other words, there may be multiple alignments which provide the
same optimal result for Eq\,\eqref{eq:hungCER} and the HA has no means
to decide which one would be more consistent with the positions of
these words in the RO of the compared texts.
This is discussed in detail in \ref{ex:5} for one of the examples of
\ref{ex:4}. Because of unlucky tie breaks, the NSFD between two
example sentences $X$ and $Z$ which are almost in the same RO is
$\rho(X,Z)\!=\!13.3\%$. However, if ties are broken more favorably
(and in a more natural way), the resulting NSFD is
$\rho(X,Z)=1/\lfloor14^2/2\rfloor = 1.0\%$, which much better reflects
the very minor RO discrepancy between $X$ and $Z$.
\medskip
To avoid this kind of ties, we propose to regularize the HA cost with
a term which measures the contribution of each pairing to increase the
NSFD. That is, we propose changing Eq.\,\eqref{eq:hungCER} into:
\begin{equation}\label{eq:hungRegulCER}
\mathrm{d_h}(X,Y) ~=~ \min_{\mathcal{A}}
\!\!\sum_{(j,k)\in\mathcal{A}(X,Y)}\!\! \biggl(g(X_j,Y_k)
~+~ \gamma\,\frac{|j-k|}{N} \,\biggr)
\vspace{-0.3em}
\end{equation}
where $\gamma$ is the regularization factor and, as in Eq.\eqref{eq:nsfd},
it is assumed that $|j-\epsilon|=|\epsilon-k|=1~\forall j,k$.
If $\gamma$ is close to $0$, the HA will just behave as usual,
yielding hWER values very close or identical to those of bWER, but
alignments $\hat{\mathcal{A}}(\cdot,\cdot)$ not ideal for assessing RO
discrepancies.
On the other extreme, for large $\gamma$ the HA will tend to provide
alignments which do not change word order; that is, alignments close
to the sequential trace $\mathcal{T}(\cdot,\cdot)$ of the traditional
edit distance (cf. Eq.\eqref{eq:editDist}), with NSFD values close to
$0$.
For small values of $\gamma$ it is expected that the hWER result
provided by Eq,\,\eqref{eq:hungWER} and Eq.\,\eqref{eq:hungRegulCER}
will be very close or identical to those obtained with $\gamma=0$; but
the alignment $\hat{\mathcal{A}}(\cdot,\cdot)$, when used in
Eq.\,\eqref{eq:nsfd}, will result is NSFD values which more fairly
reflect RO discrepancies.
In the example discussed in \ref{ex:5}, the hWER value yield by
Eq,\,\eqref{eq:hungWER} using the alignment provided by
Eq.\,\eqref{eq:hungRegulCER} with $\gamma=1.0$, is exactly the same as
before (21.4\%). But now the alignment is the one sought,
which leads to $\rho(X,Z)=1.0\%$, rather than $\rho(X,Z)=13.3\%$
obtained with the original $\hat{\mathcal{A}}(X,Z)$ of
Example~4~(\ref{ex:4}).
\medskip
To define a proper ``HA character error rate'' (hCER), note that the
HA score $\mathrm{d_h}(X,Y)$ is not directly suitable because of the
regularization and the special treatment of word insertions and
deletions.
However, a simple approximation can be easily computed as
$\hCER(X,Y)\,\eqdef\,\CER(X,\tilde{Y})$, where $\CER(\cdot,\cdot)$ is
the standard character error rate (see
Sec.\,\ref{sec:traditionalWerCer}) and $\tilde{Y}$ is obtained by
reordering the word hypothesis $Y$ according to the optimal alignment
of Eq.\,\eqref{eq:hungRegulCER}.
The values obtained in this way for the examples in \ref{ex:4} are:
$~\hCER(X,Y)=8.1$, $~\hCER(X,Z)=16.1$,
\vspace{1em}
\section{Summary of The Different Metrics Proposed}\label{sec:summMetrics}
\vspace{-0.2em}
This section summarizes the properties of the most important
evaluation metrics discussed above. In all the cases, it is assumed
that $X$ is a full-page reference transcript, with $N$ running words,
and $Y$ a corresponding HTR hypotheses with $O(N)$ running words.
\begin{description}\itemsep=-0em
\item $\WER(X,Y)$: The traditional Word Error Rate, defined in
Eqs.\,(\ref{eq:editDist}), and\,(\ref{eq:fullPageWER}), with a
computational cost in $O(N^2)$. If $Y$ is in the same RO as $X$,
the WER just measures the word recognition error rate. Otherwise,
this metric is expected to grow monotonically with the amount of RO
mismatch between $X$ and $Y$, with an offset that reflects the
actual word recognition failures. This offset can accurately be
estimated by the bWER or the hWER.
\item $\beta\!\WER(X,Y)$: An early, naive notion of ``bag of words
error rate'' defined as $\mathrm{B}(X,Y)/N$, where $\mathrm{B}(X,Y)$
measures text discrepancies in terms of only word insertions and
deletions (Eq.\,\eqref{eq:BoWdist}). It can be computed in $O(N)$
time. When $Y$ is in the same RO as $X$, the classical WER yields
(much) lower values than the $\beta$WER, but if the RO is very
different, WER is expected to be much larger. The use of this
metric is, therefore, not appealing.
\item $\bWER(X,Y)$: A redrafted version of $\beta$WER, given in
Eq.\,\eqref{eq:werBoW}, which exactly estimates how many word
insertions and deletions can be equivalently resolved with word
substitutions. It can be computed in $O(N)$ time. When $Y$ is in
the same RO as $X$, it is expected to yield values which are only
slightly lower than those of the classical WER but, in contrast to
WER, it is completely insensible to RO mismatch. A drawback of this
metric is that it does not provide any word-to-word alignment,
thereby preventing to compute a character error rate or to be used
as a basis to estimate a RO mismatch metric.
\item $\hWER(X,Y)$: The ``Hungarian Algorithm Word Error Rate'',
defined in Eq.\,\eqref{eq:hungWER} based on a RO-independent word
alignment obtained as a byproduct of computing
Eq.\,\eqref{eq:hungRegulCER}. Its computational cost is $O(N^3)$. In
terms of word error rate, hWER is almost identical to bWER, but it
may provide slightly higher values than bWER in some cases.
In contrast with bWER, hWER does provide word alignments which allow
computing a character error rate and can be used to estimate a RO
mismatch (with the NSFD, e.g.).
\item $\rho(X,Y)$: Normalized Spearman Footrule Distance (NSFD),
defined in Eq.\,\eqref{eq:nsfd} to explicitly estimate the amount of
RO mismatch between $X$ and $Y$. It requires a word-to-word
alignment which is assumed to be available as a byproduct of
computing the hWER. Its computational cost is $O(N)$, but taking
into account the cost of obtaining the required alignment, the
overall cost is $O(N^3)$. The values of NSFD are expected to grow
monotonically with the degree of RO mismatch. It is also expected
that these values be closely correlated with the values of the
classical WER, after discounting the offset due to actual word
recognition errors which, as previously mentioned, can be accurately
estimated by the bWER or the hWER.
\vspace{-0.3em}
\end{description}
\section{Simulation Experiments}\label{sec:simuExp}
\vspace{-0.2em}
A first series of experiments were carried out to check and
empirically analyze the properties of the proposed metrics under
controlled conditions.
To this end a simple HTR dataset was adopted and real full-page HTR
transcription results were artificially altered in order to simulate
typical conditions that are expected to affect the different evaluation
results.
\vspace{-0.2em}
\subsection{A Basic Dataset for Testing Different Assessment Approaches}
\label{sec:dataset}
The well known and widely used ICFHR14 dataset was adopted.
This is a small subset of selected manuscripts from the Bentham Papers
collection,\!%
\footnote{The full collection (searchable using PrIx~\cite{toselli2019a})
is here: \url{http://prhlt-kws.prhlt.upv.es/bentham}} %
mostly written by the English philosopher and reformer Jeremy Bentham.\!%
\footnote{\url{http://blogs.ucl.ac.uk/transcribe-bentham/jeremy-bentham}} %
The ICFHR14 dataset contains text-line images extracted from around
$433$ page images, some examples of which are shown in
Fig.~\ref{fig:BenthamSamples}.
It was first used in the ICFHR-2014 HTR competition~\cite{sanchez:2014}
and is now freely available for research purpose at \textsc{zenodo}
(see \ref{ap:software}).
\begin{figure}[htb]
\centering
\includegraphics[height=.35\linewidth]{071_151_003}%
\hspace{.5em}%
\includegraphics[height=.35\linewidth]{035_321_001}%
\hspace{.5em}%
\includegraphics[height=.35\linewidth]{116_064_001-2}
\vspace*{-0.7em}
\caption{Page images of the ICFHR14 dataset.}
\label{fig:BenthamSamples}
\vspace{-0.5em}
\end{figure}
This early dataset was carefully prepared by the ICFHR14 organizers so
as to avoid the need of LA and to simplify ``non-essential'' HTR
matters as much as possible. To this end, text lines were manually
detected and extracted and small pieces of text such as marginalia
were ignored.
Thus, all the benchmarking results reported so far for this dataset
have been based only on conventional WER, exactly as discussed in
Sec.\,\ref{sec:traditionalWerCer}. That is, the given training
text-line images and their corresponding GT transcripts were directly
used for model training and the WER was evaluated on the results
achieved for the independent set of test line images.
Here we will use the test-set line images to simulate different
shortcomings typically expected both from HTR and LA. Main statistics
of this test set are reported later in Table\,\ref{tab:datasets}.
\subsection{General Settings to Analyze the Proposed Metrics}
\label{sec:settSimul}
For each test-set page, the transcripts of the different text-lines
were concatenated into a single word sequence, following the RO
specified in the GT of that page. From this sequence, $\WER$, $\bWER$
and $\hWER$ can be computed according to
Eqs.\,(\ref{eq:fullPageWER},\,\ref{eq:werBoW}) and \eqref{eq:hungWER},
respectively.
NSFD, in turn, can be determined according to Eq.\,\eqref{eq:nsfd},
using the alignment derived from the computation of $\hWER$, after the
position renumbering described in Sec.\,\ref{sec:roAssess}. Finally,
$\CER$ and $\hCER$ can be calculated as explained in
Sec.\,\ref{sec:traditionalWerCer} and at the end of
Sec.\,\ref{sec:hungarian}.
To obtain global values of these metrics for the whole set of test
pages, we perform ``micro-averaging''. For $\WER$, $\bWER$ and
$\hWER$, respectively, the edit-distances $d(\cdot,\cdot)$
of~\eqref{eq:fullPageWER}, the revamped bag-of-word distances
$\mathrm{B}'(\cdot,\cdot)$ of~\eqref{eq:werBoW} and HA word-mismatches
$\updelta(\cdot,\cdot)$ of~\eqref{eq:hungWER}, are accumulated over
the test-set pages. Then, the accumulated values are normalized by
the total number of reference words (or characters for $\CER$ and
$\hCER$).
In the case of NSFD, note that because of the quadratic normalization
term, micro-averaging is not adequate. However, normal (or ``macro'')
average is not ideal either, because results might show tendencies
which differ from those of the other metrics, where the relative impact
of each page depends on its size (number of words). To attenuate
this potential mismatch, a weighted average of single-page NSFD values
is computed, with weights defined as the relative number of words in
each page.
Among the proposed metrics, only $\hWER$ has a tunable parameter;
namely, the regularization factor of Eq.\,\eqref{eq:hungRegulCER},
$\gamma$.
Throughout several tests, it has been consistently found that this
parameter does not require critical tuning.
For one of these typical tests, Table\,\ref{tab:reguTuning} reports
NSFD and $\hWER$ results for increasing values of $\gamma$.
These results were obtained in a controlled RO--alteration experiment,
described in Sec.\,\ref{sec:roRes}, where random swaps were applied to
4 text lines of each image, at distances ranging from 4 to 7 lines apart.
\vspace{-0.5em}
\begin{table}[htb]
\centering
\caption{ \label{tab:reguTuning}%
$\hWER$ and NSFD results (in percentage) for increasing values of
the regularization factor $\gamma$. The corresponding $\WER$ and
$\bWER$ results were $42.6\%$ and $12.4\%$, respectively.
}
\vspace{0.3em}
\begin{tabular}[b]{l|rrrrrrrrrr}
\toprule
$\gamma$
& 0~~ & 10$^{-4}\!\!$& 0.1 & 1~& 2~& 5~ & 10~ & 20~ & 50~ & 100 \\
\midrule
$\rho$
& 14.7 & 12.9 & 12.9 & 12.8 & 12.6 & 11.5 & 9.0 & 5.3 & 1.6 & 0.9 \\
$\hWER$
& 12.4 & 12.4 & 12.4 & 12.4 & 12.5 & 13.8 & 17.9 & 25.4 & 36.5 & 42.7 \\
\bottomrule
\end{tabular}
\vspace{-0.3em}
\end{table}
As discussed in Sec.\,\ref{sec:hungarian}, $\rho$ actually decreases
monotonically with $\gamma$, while $\hWER$ is almost constant and
identical to $\bWER$ for a wide range of $\gamma<5$.
According to these and other similar results, the regularization
factor was set to $\gamma=1.0$ for all the experiments presented in
this paper.
\subsection{Inducing Word-level Character Errors,
While the RO is Kept Essentially Unchanged}
\label{sec:icerRes}
In this experiment we applied increasingly higher character-level
insertions, deletions and substitution distortion to the test-set
reference transcripts, while keeping text lines in their original (correct) RO.
Two different settings were considered: 1) \emph{``line-level''},
where white-space editing operations are allowed to separate or join
words, and 2) \emph{``word-level''}, where white-space was excluded
from editing operations in order to keep the number of running words
unchanged.
The lowest distortion was chosen so as to induce a CER of $3.25\%$,
which is the CER of real HTR transcripts obtained in a regular
experiment (see Table\,\ref{tab:E2EevalRes}).
Increasing distortion was then progressively applied according to
$\tCER(n)\!=\!3.25n,\,n\!\in\!\{1,2,\ldots,6\}$, until reaching an
induced (or "theoretical'') tCER of $19.5\%$.
The distribution of the total tCER into the different character error
types was set proportional to the observed proportions of
substitutions, insertions and deletions.
Further, for line-level distortion, the proportion of white-space
characters was set according to the character error distribution
observed in the real HTR experiment.
Fig.\,\ref{fig:lineDistRes} plots the empirical $\WER$, $\bWER$ and
$\hWER$ results, along with the theoretical values of induced CER
(tCER, dotted-line) and WER (tWER, dashed-line, calculated according
to $\tWER(n)\!=\!4.65\cdot\tCER(n)$, where $4.65$ is the average word
length in the reference transcripts).
\begin{figure}[htb]
\centering
\includegraphics[width=.47\textwidth]{Hung_N_vs_WER-SPR_DstrWrd-Rfct}%
\hspace{2em}%
\includegraphics[width=.47\textwidth]{Hung_N_vs_WER-SPR_Dstr-Rfct}%
\vspace{-0.6em}
\caption{Evaluation results for increasingly distorted transcripts,
as a function of the CER artificially induced by the distortion
process.
%
Left: words are distorted individually avoiding induced white
space errors to break or join words.
%
Right: distortion is applied at full line level, allowing white
space to be deleted/inserted between/within words.
%
Curves with very similar or identical values are depicted with the
same color and/or point shape.
The prefix ``t'' in tCER, tWER and tNSFD indicates the corresponding
values are theoretically computed.}
\label{fig:lineDistRes}
\end{figure}
Results for the word-level distortion are shown in
Fig.\,\ref{fig:lineDistRes}-left. As the RO in this case is not
altered at all, the theoretical NSFD (tNSFD) is 0 (horizontal
dash-dotted line). As expected, all the empirical NSFD values are
also very close to 0.
Moreover, the empirical values of $\WER$, $\bWER$ and $\hWER$ all grow
almost identically for increasing tCER. This also holds for $\CER$ and
$\hCER$.
For line-level distortion the results are shown in
Fig.\,\ref{fig:lineDistRes}-right. In this case, for large tCER, the
empirical NSFD results become significantly larger than 0, and
$\CER$ is also somewhat larger than $\hCER$.
This is clearly due to the white-space editing operations which, for
large tCER, results in significant variations in the number of
words. The HA need to accommodate these variations by means of
insertions and/or deletions, which explicitly increases the NSFD,
albeit only moderately.
\subsection{Altering Text Line RO for HTR Transcripts with Fixed Word Errors}
\label{sec:roRes}
Here we evaluated the impact of altering the RO of
the real HTR transcripts produced in a regular HTR experiment (namely,
the one whose results are reported in the first row of
Table\,\ref{tab:E2EevalRes} of Sec.\,\ref{sec:e2eRes}).
For the sake of simplicity, alterations considered in this section are
limited to whole-line swapping. This aims to simulate typical
failures in text-line ordering, often caused by poor (implicit or
explicit) LA of images with multi-column text blocks, marginalia, etc.
For each test-set page with $M$ text lines, the order of line
transcription hypotheses is changed by swapping a given number of line
pairs, $S$, at a given distance or range, $r$. Line pairs are randomly
selected, but lines already swapped are not candidate for further
swapping.
For a given $S$, depending on the value of $r$, the actual number of
possible swapping on a page may be lower than $S$.
For example, the maximum number of swappable line pairs of a page with
$M=8$ lines, at a distance $r=7$, is only one: the first line with the
last one of that page.
\medskip
For a given range of swap distances $r\in[R',R]$, and a given number
of pages, $K$, the expected NSFD induced by this process,
$\tilde{\rho}$, can be approximated as:
\begin{equation}\label{eq:nsfd-wordLv4}
\tilde{\rho}(K,S,R',R) ~\approx~
\frac{S(R'+R)}{K}\,\sum_{k=1}^K\frac{1}{\lfloor M_k^2/2\rfloor}
\end{equation}
where $M_k$ is the number of lines of the $k$-th page.
In our experiments, $K=33,\,R'=4,\,R=7,$ and
$\,\sum_{k=1}^K1/{\lfloor M_k^2/2\rfloor}=0.136$, yielding:
$\tilde{\rho}(S) ~=~ 0.045\,S$. In the right plot of
Fig.\,\ref{fig:lineSwapsBreakRes} shows these expected NSFD values as
the dashed line labeled ``tNSFD swp'').
The left plot of Fig.\,\ref{fig:lineSwapsBreakRes} shows WER values
obtained for different (maximum) numbers of swapped lines, where each
value is the average over a range of swap distances $[4,7]$.
As expected, while $\WER$ increases quickly with the number of
swapped lines, the corresponding $\bWER$ and $\hWER$ remain almost
constant.
On the other hand, the right plot shows how the empirical NSFD values
also grow as the number of line swaps increases, more or less closely
following the expected linear tendency (tNSFD~swp).
Fig.\,\ref{fig:lineSwapsBreakRes} also includes WER and NSFD results
of the experiments discussed in the next subsection.
\vspace{0.5em}
\begin{figure}[htb]
\centering
\includegraphics[width=.44\textwidth]{Hung_N_vs_WER_SwpDst-Rfct}%
\hspace{2em}%
\includegraphics[width=.44\textwidth]{Hung_N_vs_SPR_SwpDst-Rfct}%
\vspace{-0.3em}
\caption{Evaluation results on actual HTR transcripts
where the line order is distorted by random line swaps and
breaks. Curves with almost identical values are depicted with the
same color and symbol. tNSFD corresponds to theoretically computed
values.}
\label{fig:lineSwapsBreakRes}
\end{figure}
\subsection{Impact of Text Line Splitting Errors}
\label{sec:lnBrkRes}
Finally we check the effect of randomly inserting line-breaks in the
HTR transcripts.
This aims to simulate (implicit or explicit) line detection errors
which lead to wrong intra-line text ordering.
To this end, the following procedure was carried out for the HTR
transcripts of each test page: 1) $S$ lines are randomly selected. 2)
For each selected line a splitting position is randomly chosen; it can
be at character or word level, with a chance of $1$ to $4$
respectively. 3) The split line fragments are relocated according
to one of these three equiprobable options:
i) the line suffix goes before the prefix, ii) the line suffix goes
after the line next to the selected one, or iii) the line prefix goes
after the line next to the selected one. These cases correspond to
relatively common flaws of (implicit or explicit) LA, which may happen
mainly with highly skewed text images, as illustrated in
Fig.\,\ref{fig:splitLines} (see also Fig.\,\ref{fig:complexExamples}).
\begin{figure}[htb]
\centering
\fbox{\includegraphics[width=.63\textwidth]{breakLine_case1}}\\[0.2em]
\fbox{\includegraphics[width=.75\textwidth]{breakLine_case2}}\\[0.2em]
\fbox{\includegraphics[width=.60\textwidth]{breakLine_case3}}
\vspace{-0.3em}
\caption{\label{fig:splitLines}%
Real examples from Bentham Papers images 010\_003\_002,
019\_004\_003, which illustrate text line splitting errors that
affect RO. In the simulation experiments these examples
correspond, top to bottom, to Cases 1, 2 and 3.}
\vspace{-0.3em}
\end{figure}
\bigskip\bigskip\bigskip\bigskip
As in Sec.\,\ref{sec:roRes}, we can estimate the impact of these RO
alterations on the NSFD metric. Ignoring the effect of word breaks,
the NSFD induced for $K$ page images can be approximated as:
\begin{equation}\label{eq:nsfd-wordLv4b}
\tilde{\rho}(S) ~\approx~
\frac{7}{6}\frac{S}{K}\sum_{k=1}^K\frac{N_k/ M_k}{\lfloor M_k^2/2\rfloor}
\end{equation}
where, as before, $M_k$ and $N_k$ are respectively the number of lines
and words of the $k$-th page image. For our $K=33$ page images, this
leads to $\tilde{\rho}(S)=0.0048\,S$. The dotted line labeled ``tNSFD
brk'' in the right plot of Fig.\,\ref{fig:lineSwapsBreakRes} shows
these expected NSFD values.
Note that, unlike the RO alteration simulation of
Sec.\,\ref{sec:roRes}, here not only the RO is changed (in this case
at a range distance $r=1$), but also some words are distorted because
a line split point may happen to fall within a word, thereby producing
two word fragments.
Such word splits happen with probability $S\,(K/4)/\sum_{k=1}^K\,N_k$
and for each split, two word errors are expected. In our case, $K=33$
and, for the transcription hypotheses, $\sum_{k=1}^KN_k=6955$.
Therefore, the expected increase of $\bWER$ (and $\hWER$) is
$0.0023\,S$ ($0.23\,S$ in \%), which explains the tiny increase of
$\bWER$-brk and $\hWER$-brk observed in
Fig.\,\ref{fig:lineSwapsBreakRes}.
\subsection{WER--NSFD Correlation and Computational Costs}
\label{sec:concluSimu}
In Sec.\,\ref{sec:icerRes} (Fig.\,\ref{fig:lineDistRes}) we have seen
that, when the amount of character (and word) errors increases without
changing the word order, the NSFD remains essentially constant and
close to 0. In contrast, all the word and character error metrics
grow almost linearly with the amount of induced character errors.
Moreover, the three word error metrics ($\WER$, $\bWER$ and $\hWER$)
yield almost identical values in all the cases.
On the other hand, we have seen in Sec.\,\ref{sec:roRes}
(Fig.\,\ref{fig:lineSwapsBreakRes}) that if the amount of word errors
is kept constant but the RO of the transcripts is increasingly
perturbed, both $\WER$ and NSFD (and also $\CER$) grow fairly linearly
with the amount of induced RO mismatch. In contrast, now $\bWER$ (and
$\hWER$) remain practically constant and equal to the value of $\WER$
when HTR and reference transcripts are in the same RO.
All these results (those of Fig.\,\ref{fig:lineSwapsBreakRes} in
particular) suggest a strong correlation between NSFD and $\WER$, with
an offset given by $\bWER$ (or $\hWER$). This is explicitly put
forward in Fig.\,\ref{fig:correlation}, where values of
$\Delta\!\WER\eqdef\WER-\bWER$ (and also $\WER-\hWER$) are plotted
against the corresponding values of NSFD.
We also include in this plot a few points corresponding to real
end-to-end evaluation results of some of the experiments that will be
presented in Sec.\,\ref{sec:realEval} (Table\,\ref{tab:E2EevalRes}).
It can be seen that these points also show a fair linear correlation
between $\Delta\!\WER$ and NSFD.
Regarding the relative costs of the different metrics, computing times
are plotted in Fig.\,\ref{fig:compTimesHung} as a function of the
number of words per page. All the times were measured on the same
computer, using the C++ implementations of $\WER$, $\bWER$, and
$\hWER$ described in \ref{ap:software}.
The points correspond to real end-to-end evaluation of individual pages
and the least-square fitted curves clearly show the different time
complexities of each method.
\begin{figure}[htb]
\centering
\begin{minipage}[t]{.48\textwidth}
\centering
\raisebox{0pt}{
\includegraphics[height=1.33\textwidth]{Hung_N_vs_SPR-WER_SwpDst-Rfct}}%
\vspace{-0.5em}
\caption{Correlation of $\WER\!-\!\bWER$ (and $\WER\!-$ $\!\hWER$)
with NSFD. Real results from Table~\ref{tab:E2EevalRes} are
included, along with a straight line fitted to these points.
Curves with almost identical values are depicted with the same
color and symbol.
\label{fig:correlation}}
\end{minipage}%
\hfill%
\begin{minipage}[t]{.48\textwidth}
\centering
\includegraphics[height=1.33\textwidth]{Hung_Times-Correlations}
\vspace{-0.5em}
\caption{Computing Times of $\hWER, O(N^3)$, $\WER, O(N^2)$ and
$\bWER, O(N)$, fitted respectively to polynomials of degrees 3,
2 and 1 (linear).
\label{fig:compTimesHung}}
\end{minipage}
\vspace{-0.5em}
\end{figure}
\vspace{-0.5em}
\section{Examples of Real End-to-End Evaluation}\label{sec:realEval}
\vspace{-0.2em}
The proposed evaluation metrics have been applied to assess end-to-end
HTR systems in real scenarios. The HTR datasets considered, the
empirical settings and the results obtained are presented in the
following subsections.
\vspace{-0.3em}
\subsection{Datasets and Methods}\label{sec:empiricalSettings}
\vspace{-0.2em}
Besides the historical dataset ICFHR14~\cite{sanchez:2014} already
used in the preceding sections,
four additional datasets were selected to test the proposed evaluation
metrics; namely: the traditional modern handwriting benchmark
IAMDB~\cite{marti2002iam}, and three historical handwriting datasets:
ICFHR16~\cite{sanchez2016} and ICDAR17~\cite{sanchez:2017}, compiled
for the ICFHR-2016 and ICDAR-2017 HTR competitions, and the
\emph{Finnish Court Records} dataset (FCR)~\cite{quiros2022} from the
``Renovated District Court Records'' held by the National Archives of
Finland. Information about how to download each of these datasets is
given in \ref{ap:software}.
IAMDB is a well known modern English handwritten text corpus, gathered
by the FKI-IAM Research Group on the base of the Lancaster-Oslo/Bergen
text Corpus (LOB)~\cite{johansson1978lancaster}. The last released
version (3.0) contains about $1\,500$ scanned text pages,
written by $657$ different writers.
The ICFHR16 dataset encompasses $450$ page images which are
a subset of the Ratsprotokolle collection, written in old German
and composed of handwritten minutes of council meetings held from 1470
to 1805. One remarkable characteristic of this dataset is that their
text lines are short, each one containing very few (long) words.
The ICDAR17 dataset comprises around $10$K page images,
most of which taken from the \emph{Alfred Escher Letter Collection}.
This collection is mostly written in German, but it also includes
pages in French and Italian. Here, the performance evaluation was
carried out on the pages corresponding to the partition called
``Test-B2'' in~\cite{sanchez:2017}, which was aimed to evaluate not
only text recognition accuracy but also (indirectly) LA performance.
Finally, the FCR dataset consists of $500$ manuscript images
which contain records of deeds, mortgages, traditional lifeannuity,
among others. They were written in Swedish by many hands during the
18th century. Here, the evaluation was done on $100$ images
($48$ are double-page images) which are a subset of the test partition
used in\,\cite{vidal2021hyp}.
For more details about ICFHR14, ICFHR16, ICDAR17 and FCR datasets,
refer to~\cite{sanchez2019,quiros2022}.
It is important to remark that no tokenization (e.g. to separate
punctuation marks from words) was applied to text references or HTR
transcripts, excepting ICDAR17 whose original references and HTR
results obtained in the associated competition were used.
Table\,\ref{tab:datasets} reports the main statistics
\emph{only for the test sets}, which are the focus of
the proposed evaluation metrics.
\vspace{-0.7em}
\begin{table}[htb]
\centering
\caption{
\label{tab:datasets}%
Test set main statistics of the evaluated datasets. Except for
ICDAR17, the running words and lexicon sizes correspond to
untokenized ``words'', which may include punctuation marks.}
\vspace{0.3em}
\small
\begin{tabular}{l|rrrrrr}
\toprule[1.5pt]
& ICFHR14 & IAMDB & ICFHR16 & ICDAR17 & FCR \\
\midrule
Number of pages & 33 & 336 & 50 & 57 & 100 \\
Number of lines & 860 & 2\,915 & 1\,138 & 1\,412 & 6\,183 \\
Running words & 6\,966 & 23\,406 & 3\,546 & 14\,460 & 33\,999 \\
Running chars & 38\,474 & 123\,090 & 22\,396 & 80\,568 & 214\,785 \\
Lexicon & 2\,278 & 6\,398 & 1\,834 & 4\,648 & 9\,890 \\
Alphabet size & 82 & 75 & 80 & 104 & 83 \\
\bottomrule[1.5pt]
\end{tabular}
\vspace{-0.3em}
\end{table}
Except ICDAR17 (for which the same transcripts as
in~\cite{sanchez:2017} were used), for each dataset we trained
specific optical character models using the provided training images
and the corresponding reference transcripts.
Character modeling was based on
\emph{Convolutional-Recurrent Neural Networks} (CRNN), trained using
the state-of-the-art freely available PyLaia
Toolkit\,\cite{puigcerver2018}.\!%
\footnote{\url{https://github.com/jpuigcerver/PyLaia}}
The same setup described in\,\cite{vidal2021hyp} was adopted here to
specify the CRNN topology and meta-parameters.
HTR transcripts of test images were obtained
through two different ways of line extraction:
1) use the line locations and RO given in the GT; and 2) use a
\emph{Region Proposal Network} (RPN)~\cite{he2017mask,lquiros:21}
trained to detect and extract lines with a RO given by their positions
on image, from top-to-bottom and left-to-right as in\,\cite{quiros2022}.
To this end, the same RPN topology and meta-parameter settings
as in~\cite{lquiros:21} was adopted.
For both ways of line extraction and each dataset, the corresponding
CRNN model trained with PyLaia was used to decode the extracted line
images. Finally, HTR full-page transcripts were produced by
concatenating the predicted text lines according to their RO
given by the GT or computed by the RPN.
In addition to the above ``classical'' HTR experiments, as an example
of what we consider the ultimate aim of the proposed metrics, we also
test the end-to-end LA\,+\,HTR approach named
\emph{Simple Predict \& Align Network} (SPAN)\,\cite{Coquenet2021}.%
\footnote{\url{https://github.com/FactoDeepLearning/SPAN}}
This model learns to transcribe paragraphs by aligning all the text
line representations via a horizontal feature map unfolding.
By training with the CTC loss strategy, this model learns how to align
input information with the feature map rows and produce a sequential
output, without requiring any specific LA preprocessing.
\subsection{Real End-to-End Evaluation Results}\label{sec:e2eRes}
Table\,\ref{tab:E2EevalRes} reports performance in terms of the
proposed evaluation metrics for different end-to-end HTR approaches,
tested on the datasets outlined before. The way of line extraction
and ordering, as well as the HTR system adopted, appear on the columns
labeled LA+RO and HTR, respectively.
Selected values of $\Delta\!\WER=\WER-\hWER$ and NSFD, highlighted
in boldface, are plotted in Fig.\,\ref{fig:correlation}, as already
mentioned in Sec.\,\ref{sec:concluSimu}.
\vspace{-0.5em}
\begin{table}[htbp]
\centering
\caption{
\label{tab:E2EevalRes}%
Real evaluation results for different datasets, LA and HTR
approaches. All values are percentages. $\Delta\!\WER$ denotes
$\WER\!-\!\bWER$. RO mismatch (NSFD) and $\Delta\!\WER$ values
corresponding to points shown in Fig.\,\ref{fig:correlation} are
marked in boldface.
$\WER,\bWER$ and $\hWER$ $95\%$ confidence intervals are narrower
than $\pm$1.6\% for ICFHR16 and $\pm$1\% for all the other
datasets. The LA+RO approach ``TRB'' for ICDAR17 stands for
Transkribus platform.}
\vspace{0.5em}
\small
\extrarowheight=-0.2pt
\begin{tabular}{l|ll||rr|rrr|rr}
\toprule[1.5pt]
\raisebox{-4pt}{Dataset}\raisebox{5pt}{\!\!\!\!\!\!Metric\!\!}
& LA+RO & HTR & NSFD &\!\!\!$\Delta\!\WER$ & $\WER$ & $\!\!\!\bWER$
& $\!\!\!\hWER$ & $\CER$ & $\!\!\!\hCER$ \\
\midrule[1.5pt]
\multirow{2}{*}{ICFHR14}
& GT & CRNN & 0.3 & 0.3 & 12.7 & 12.4 & 12.4 & 3.3 & 4.0 \\
& RPN & CRNN & {\bf 0.9}& {\bf 1.1}& 17.4 & 16.3 & 16.3 & 5.5 & 5.9 \\
\midrule
\multirow{3.5}{*}{IAMDB}
& GT & CRNN & 0.6 & 0.5 & 27.0 & 26.5 & 26.5 & 7.5 & 8.2 \\
& RPN & CRNN & {\bf 0.7} & {\bf 0.5} & 27.8 & 27.3 & 27.3 & 7.9 & 8.7 \\
\cmidrule{2-10}
&\multicolumn{2}{c||}{SPAN}& 0.5 & 0.6 & 26.7 & 25.9 & 26.0 & 7.5 & 8.3 \\
\midrule
\multirow{3.5}{*}{ICFHR16}
& GT & CRNN & 0.3 & 0.6 & 27.7 & 27.1 & 27.2 & 5.7 & 6.6 \\
& RPN & CRNN & {\bf 5.2}& {\bf 7.1} & 33.5 & 26.4 & 26.6 & 13.7 & 6.5 \\
\cmidrule{2-10}
&\multicolumn{2}{c||}{SPAN}& 1.3 & 1.6 & 31.5 & 29.9 & 30.0 & 10.7 & 10.9\\
\midrule
\multirow{2}{*}{ICDAR17}
& GT & CRNN & 1.4 & 2.2 & 18.6 & 16.4 & 16.5 & 6.3 & 6.6 \\
& TRB & CRNN & {\bf 1.6} & {\bf 2.5} & 20.1 & 17.6 & 17.7 & 7.0 & 7.1 \\
\midrule
\multirow{3}{*}{FCR
& GT & CRNN & 0.8 & 1.1 & 25.2 & 24.1 & 24.4 & 5.6 & 6.4 \\
& RPN & CRNN &{\bf 26.6}&{\bf 45.7}& 72.4 & 26.7 & 27.0 & 50.8 & 8.5 \\
& RPN1 & CRNN & {\bf 5.9} & {\bf 9.8} & 36.5 & 26.7 & 27.0 & 15.1 & 8.2 \\
\bottomrule[1.5pt]
\end{tabular}
\vspace{-0.2em}
\end{table}
In all the cases, $\hWER$ is slightly higher than or identical to
$\bWER$ and both are always smaller than $\WER$, as discussed in
Secs.\,\ref{sec:bow},\ref{sec:hungarian} (and summarized in
Sec.\,\ref{sec:summMetrics}) -- and as expected from the simulation
results of Sec.\,\ref{sec:simuExp}.
Also as expected, all the HTR approaches which use the (perfect) text
lines and RO given by the GT, achieve lower NSFD and $\Delta\!\WER$,
compared with other approaches involving automatic line detection.
Another important general remark is the fairly tight correlation
observed between NSFD and $\Delta\!\WER$. It is more clearly seen for
results more or less affected by RO issues, specifically those
highlighted in boldface which, as commented, are also plotted in
Fig.\,\ref{fig:correlation}. This further endorses the discussion in
Sec.\,\ref{sec:concluSimu} and adds empirical support to consider
$\Delta\!\WER$ as a suitable metric to put forward LA or, in general,
RO problems.
IAMDB has a very simple RO structure and no significant differences
exist among the different error rate metrics. To a lesser extent, the
same can be said for ICFHR14.
The case of ICFHR16 is worth commenting. The $\WER$ achieved by
RPN--CRNN ($33.5\%$) is significantly higher than the $\bWER$
($26.4\%$), leading to $\Delta\!\WER=7.1\%$. This makes it clear that
the RO provided by RPN LA is far from perfect, an issue directly
supported by the fairly high value of NSFD ($5.20\%$). As discussed
later in more detail, most RO errors are due to marginalia transcripts
for which the system fails to place in the correct RO.
Also interesting is the case of FCR, which contains a mixture of
single- and double-page images. For double-page images, the regular
RPN settings (denoted in the table just as ``RPN'') dramatically fail
to separate the lines of each page and render them in the correct RO.
So, even though the individual words are fairly well recognized (with
$\bWER=26.7\%$), the conventional $\WER$ is exorbitant ($72.4\%$).
This leads to a very large $\Delta\!\WER$ ($45.7\%$) which clearly
shows the massive RO mismatch, also reflected by the very large value
of NSFD ($26.6\%$).
Of course, this experiment was only aimed at providing a clear
illustration of the behavior of proposed metrics. So we also tested a
more reasonable LA approach (called ``RPN1'' in the table). In this
approach, when a double-page is identified, each detected text line is
classified as belonging to the left or to the right page and then the
usual RPN RO is applied page-wise.
This approach provides identical individual word recognition
performance ($\bWER=26.7\%$) and greatly solves the RO issues --
albeit not completely, as assessed by the still high values
$\Delta\!\WER=9.8\%$ and NSFD$=5.92\%$.
Regarding $\CER$ and $\hCER$ results, in general they reflect similar
tendencies as $\WER$ and $\hWER$ when RO issues are involved. Note
however that, as discussed in Sec.\,\ref{sec:hungarian}, $\hCER$ is
only an approximation and is not as directly and faithfully comparable
with $\CER$ as $\hWER$ is with $\WER$.
The SPAN (true full-page) approach, was tested on two datasets.
Results for IAMDB are comparatively good in terms of word and
character error metrics and also in terms of RO as assessed by NSFD
and $\Delta\!\WER$.
The SPAN results for ICDAR16 deserve a more detailed analysis. The
reading-order independent word recognition results
($\hWER\!\approx\!\bWER\!=\!29.9\%$) are sensibly worse than those of
RPN+CRNN discussed above ($\hWER\!\approx\!\bWER\!=\!26.4\%$), while
the conventional WER is somewhat better ($31.5\%$ vs. $33.5\%$). So
the $\Delta\!\WER$ for SPAN is significantly lower ($1.6\%$
vs. $7.1\%$) -- which is also consistent with NSFD ($1.3\%$
vs. $5.2\%$). This indicates that the transcripts provided by SPAN
have more word errors but are in significantly better RO than
RPN+CRNN.
To better understand these results, we can gather additional
evaluation clues from the distribution of $\bWER$ errors. In this
case, from $\bWER\!=\!29.9\%$, $25.7\%$ errors are substitutions,
$0.3\%$ insertions and $3.9\%$ deletions. So we observe that SPAN
makes many word deletions, around $10$ times more than RPN+CRNN (with
$0.4\%$ deletions, $0.6\%$ insertions and $25.4\%$ substitutions).
A closer look at the SPAN transcripts reveals that, indeed, SPAN
almost systematically delete (i.e., fails to detect and recognize) the
many marginalia lines existing in the ICDAR16 images. Clearly, while
the RO is hardly affected by this fact, there is a noticeable impact
on the reading-order independent recognition accuracy, evidenced by
the relatively larger values of $\bWER$ and $\hWER$.
As a specific example of this general fact,
Fig.\,\ref{fig:evalExample} shows an ICFHR16 page image, along with
its GT, RPN+CRNN and SPAN transcripts.
In the RPN+CRNN transcript, the lines corresponding to the marginal
note (in red) are correctly detected and all their words recognized
(with two errors). However, they are mixed with the lines of the last
paragraph (in blue). The total number of word errors is $22$
($31.4\%$), but because of the mixed marginal lines, the RO is rather
poor, as properly reflected by $\Delta\!\WER\!=\!10\%$.
For the SPAN transcript, $\hWER\!=\!\bWER\!=\!31.4$, exactly the same
as for RPN+CRNN. But, as suspected, it has completely failed to
transcribe the marginal note words. However, all the transcribed
words are in good RO, a fact faithfully reflected by
$\Delta\!\WER\!=\!1.4\%$.
\definecolor{lightgray}{rgb}{0.85, 0.85, 0.85}
\begin{figure}[htb]
\vspace{-0.3em}
\centering
\setlength{\fboxsep}{1pt}
\def-.5ex{-.5ex}
\def-.2ex{-.2ex}
\raisebox{-1pt}{
\includegraphics[width=.40\textwidth]{Seite0418}}%
\fbox{
\scalebox{.912}{
\begin{minipage}[b]{.28\textwidth}
210\\[-.5ex]
sein Behaūsūng khain\\[-.5ex]
Adelicher Ansiz, nit er¬\\[-.5ex]
heblichen, oder Zū benieg\\[-.5ex]
Angenomen. Dero¬\\[-.5ex]
wegen Zū biten, Nit\\[-.5ex]
mer wolerdeiten Fhn\\[-.5ex]
Zū Wolckhenstain vnd\\[-.5ex]
Haūbtmann Thanhaimer\\[-.5ex]
die Notwendig vnnd\\[-.5ex]
Ernnstlich verfiegūng\\[-.5ex]
Zetūen. vnd Zū geben da¬\\[-.5ex]
mit die schūldig Abstat¬\\[-.5ex]
ūng gelaisst: oder aber\\[-.5ex]
den Steūr laden sōuil\\[-.5ex]
Abgenomen werde.\\[-.5ex]
\blue{Dritens vnd schliesslich.}\\[-.5ex]
\blue{so ist mann in Erfar¬}\\[-.5ex]
\blue{ūng khomen, das der}\\[-.5ex]
\blue{Gmain Zū Keifs,.Mori}\\[-.5ex]
\red{\bf Traid ver¬}\\[-.5ex]
\red{\bf fierūng.}\\[-.5ex]
\red{\bf vermieg Patent.\\[-2ex]}
\end{minipage}}}%
\fbox{
\scalebox{.911}{
\begin{minipage}[b]{.28\textwidth}
210\\[-.5ex]
sein Behaūsūng khain\\[-.5ex]
\colorbox{lightgray}{Oodelicher} Ansiz, nit er¬\\[-.5ex]
heblichen, oder Zū \colorbox{lightgray}{benag}\\[-.5ex]
\colorbox{lightgray}{Anzenemen}, Dero¬\\[-.5ex]
wegen Zū biten, \colorbox{lightgray}{Vit}\\[-.5ex]
mer \colorbox{lightgray}{wolerd} \colorbox{lightgray}{eitem} \colorbox{lightgray}{Khln}\\[-.5ex]
Zū \colorbox{lightgray}{Grolckherstain.} vnd\\[-.5ex]
Haūbtmann \colorbox{lightgray}{Thonhaimer}\\[-.5ex]
die Notwendig vnnd\\[-.5ex]
Ernnstlich verfiegūng\\[-.5ex]
\colorbox{lightgray}{Zetūn.} vnd Zū geben da¬\\[-.5ex]
mit die schūldig \colorbox{lightgray}{Abstat.}\\[-.5ex]
\colorbox{lightgray}{nūng} gelaisst: oder aber\\[-.5ex]
den Steūr laden \colorbox{lightgray}{soūil}\\[-.5ex]
\colorbox{lightgray}{Abgemomen} werde.\\[-.5ex]
\blue{\colorbox{lightgray}{Gritens} vnd \colorbox{lightgray}{schliesslich}}\\[-.5ex]
\red{\bf Traid ver¬}\\[-.5ex]
\red{\bf fierūng.}\\[-.5ex]
\blue{so ist mann \colorbox{lightgray}{im} \colorbox{lightgray}{Erfar}
}\\[-.5ex]
\blue{ūng khomen, das der}\\[-.5ex]
\red{\bf \colorbox{lightgray}{vermig} \colorbox{lightgray}{Potent}}\\[-.5ex]
\blue{Gmain Zū \colorbox{lightgray}{Rriff.} \colorbox{lightgray}{Marj}\\[-2ex]}
\end{minipage}}}\\[.5ex]
\fbox{
\scalebox{.94}{
\begin{minipage}[b]{\textwidth}
\colorbox{lightgray}{10} sein Behaūsūng khain Adelicher Ansiz, \colorbox{lightgray}{niter¬} heblichen, oder Zū benieg \colorbox{lightgray}{Anzenemen,} Dero¬ wegen Zū biten, \colorbox{lightgray}{Nint} mer \colorbox{lightgray}{wolerd} \colorbox{lightgray}{eitem} \colorbox{lightgray}{khln} Zū \colorbox{lightgray}{irolckherstain.} vnd \colorbox{lightgray}{Haūbtmannnhai} \colorbox{lightgray}{Thomer} die Notwendig vnnd \colorbox{lightgray}{Ernnstliich} verfiegūng \colorbox{lightgray}{Zetūm.} vnd Zū geben da¬ mit die schūldig Abstat¬ ūng \colorbox{lightgray}{gelaisso:} oder aber den Steūr laden \colorbox{lightgray}{soūil} \colorbox{lightgray}{Abgemamen} werde. \blue{Dritens vnd schliesslich. so ist mann in Erfar¬ ūng khomen, das der \colorbox{lightgray}{Gmaim} Zū \colorbox{lightgray}{Reiff.} \colorbox{lightgray}{Marj} \colorbox{lightgray}{.}}
\end{minipage}}}\\[-0.3em]
\caption{ \label{fig:evalExample}%
A page from the ICHFR16 dataset (ID: Seite0418).
%
The red and blue texts and shadings correspond to the text blocks
affected by RO issues, while word recognition errors are marked
with shadowed boxes. The top-middle panel is the reference transcript.
%
The right-panel shows the RPN-CRNN's hypothesis, with
$\bWER=\hWER=31.4\%$ and $\Delta\!\WER=10\%$
which fairly reflects the RO errors caused by poor LA.
%
The bottom panel shows the SPAN's hypothesis, which clearly failed
to detect and recognize all the marginal note words (in red
color). As compared with RPN-CRNN, SPAN has produced the same
amount of word errors ($\bWER=\hWER=31.4\%$), but the transcript is
in better RO, a fact fairly reflected by $\Delta\!\WER=1.4\%$.
}
\end{figure}
\vspace{-0.5em}
\section{Related Works}\label{sec:related}
\vspace{-0.3em}
The problem of assessing the quality of full-page automatic
transcripts, taking into account LA and/or RO errors has been
addressed in many previous works. In this section, we briefly review
the literature both in HTR and other research fields where alternative
metrics have been proposed in this regard. The review is organized
into four topics, corresponding to what we consider the four main
contributions of our work.
\subsection{LA and/or RO Awareness in HTR Metrics}
Almost all the works cited in this section consider only
\emph{printed} (historical) documents and the task of full-page,
end-to-end evaluation is always more or less explicitly linked to
(geometric) issues caused by faulty LA -- see eg.~\cite{breuel2003}
for one of the earliest works in this category and~\cite{naoum2019}
for a recent one.
Going deeper in this direction, Antonacopoulos, Clausner and
Pletschacher are among the earlier authors who explicitly put forward
the importance of this problem and its relation with RO difficulties ---
and propose pioneering practical approaches for RO-aware
evaluation~\cite{clausner2013,Pletschacher2015}.
Other works which explicitly deal (only) with RO analysis methods
are~\cite{lee2002,breuel2003,ceci2007,malerba2008}.
Note, however, that with the exception
of~\cite{clausner2013,Pletschacher2015} (and others discussed in the
coming subsections), these works are \emph{not} directly concerned
with transcription \emph{evaluation}. Some produce end-to-end
transcription results, while others deal with LA and/or RO methods;
but all need evaluation metrics and some of them make proposals that
seem adequate to assess their results.
On the other hand, as far we can tell, none of the cited works makes
a convincing assessment of the adequateness of the proposed metrics
for general-purpose evaluation and benchmarking of full-page
transcripts of text images.
In comparison, our work explicitly analyses and proposes
general-purpose metrics which are agnostic of geometry and other
details of LA. We also report comprehensive results that support the
adequateness of these metrics for unbiased evaluation of the overall
quality of end-to-end transcription results of handwritten (or
printed) text images.
\subsection{Metrics Related With the Bag of Words}
The Bag of Words concept has been proposed in several works aiming to
achieve RO-independent evaluation of text image transcripts. It
appears in~\cite{Pletschacher2015} and~\cite{Strobel2020}, and it is
used in several ICDAR competitions organized by Antonacopoulos,
Clausner and Pletschacher~\cite{Antonacopoulos2013,antonacopoulos2015,%
Clausner2017,clausner2017a,ClausnerAP19}.
In most of these works, the proposed metric is the ``Bag of Words
\emph{Accuracy}'' which, using the notation of Sec.\,\ref{sec:bow},
can be expressed as:
\vspace{-0.5em}
\begin{equation}\label{eq:BoWacc}
\mathrm{bWAC}(X,Y) = \frac{1}{N}\sum_{v\in V}\min(f_X(v),f_Y(v))~~~
\vspace{-0.5em}
\end{equation}
Defined in this way, its value is $0$ if $Y$ does not have any word in
common with $X$ and $1$ ($100\%$) if all the words in $X$ appear in
$Y$, with the same frequency. So it is a reasonable measure of
RO-independent word recognition \emph{accuracy}.
However, a fundamental flaw of $\mathrm{bWAC}$ is that it ignores
(erroneous) word \emph{insertions}; that is, its value does not change
if arbitrary words (maybe relevant or not) are added to $Y$.
Therefore, one can just use the training set transcripts to generate a
(large) word sequence, and use this unique sequence as the ``system
transcript'' for every test image. This way $\mathrm{bWAC}$ can
approach $100\%$, without even having seen the test images!\,
Clearly, this makes this metric far from ideal for benchmarking.
In addition, as the $\WER$ may be greater than $1$ (see
Sec.\,\ref{sec:wcer}) and since the Bag of Words concept does not
directly account for substitution errors (c.f. Sec\,\ref{sec:bow}),
$\bWER(X,Y)$ can generally be very different from
$1\!-\!\mathrm{bWAC}(X,Y)$. Therefore, $\mathrm{bWAC}$ can not be
properly compared with the $\WER$ or with $1\!-\!\WER$ and,
consequently, it can not be used to define or compute anything like
the $\Delta\!\WER$.
Our definition of $\bWER$ in Eq.\,\eqref{eq:werBoW}
does estimate word substitutions, thereby making it almost identical
to $\hWER$ and allowing for a proper comparison with the $\WER$. This
leads to the introduction of $\Delta\!\WER$, which proves to be a very
convenient way to measure RO logical mismatch.
It is worth mentioning that our definition of $\bWER$ is not new. The
idea was first suggested in~\cite{ney2000} to obtain a rough measure
of the quality of Machine Translation (MT) results disregarding word
order. Under the name ``Position-independent Error Rate'' (PER),
that idea was later presented more formally
in~\cite{popovic2007,popovic2011}. By looking closely at the proposed
formulation, one can observe that the core computation is indeed
essentially the same as that of our Eq.\,\eqref{eq:werBoW}.
The formulation in~\cite{popovic2007,popovic2011} assume there may be
more then one GT reference with which the system hypothesis can be
compared. This is important in MT, where for an input sentence there
might be many correct MT outputs. In future works, we plan to borrow
this multi-reference idea for full-page evaluation of HTR transcripts,
where the correct transcription RO may be not be unique.
\subsection{Metrics Related With the Hungarian Algorithm}\label{sec:relatedHA}
The HA has been proposed and used for many document analysis and
recognition tasks, many of them related with full-page, end-to-end
training and/or text image
recognition~\cite{tensmeyer2019training,moysset2017full,%
moysset2018learning,long2022towards}.
It has been proposed as well for other miscellaneous tasks such as
invoice analysis~\cite{palm2017cloudscan}, pairing different versions
of historical manuscripts~\cite{kassis2017alignment}, and reassembling
shredded document stripes~\cite{liang2019eassembling}, to name a few.
All these works are completely unrelated with evaluation of HTR
transcripts, which is the topic of this paper.
Among the works which explicitly deal with evaluation, we should
mention an interesting early work in the field of Computer Vision,
which considers the evaluation of visual objects
detection~\cite{liu2002optimal}. More recently, several works on LA
make use of the HA to evaluate results of \emph{line detection} and/or
\emph{text region segmentation}~%
\cite{li2008linesegeval,yin2009linesegeval,garz2013linesegeval}.
While this task may seem similar to the one here considered, the
overall framework is quite different. In these proposals, the
elements to be paired are image regions, and the pairing criterion is
strongly based on region geometry information.
In contrast, our proposal is applied to transcripts, represented just
by character strings. And evaluation is completely blind to the
existence of text lines and explicitly ignores geometric features of
the text images and/or their GT annotations.
It is worth mentioning that our point of view in this matter is
similar to the one adopted in~\cite{soundararajan2006evaluation} for
assessment of video OCR results. However, the metric proposed
in~\cite{soundararajan2006evaluation} aims to assess not only the
quality of the transcripts (and their RO), but also the positions of
the detected and recognized words in the image. Therefore, this
evaluation approach mixes geometric and text criteria, which is
contrary to the principles adopted in our work.
Perhaps the most interesting proposal that is close to our work is the
so called ``Flexible Character Accuracy'' metric~\cite{ClausnerPA20}
(FCA). It is based on computing the character edit distance between
two chunks of text by iteratively comparing the lines with minimum
edit distance, following a greedy strategy.
The method is further based on several heuristics which need four
weighting factors
to control how much relevance is given to the offset and length
difference of the matched strings. Additionally, unmatched substrings
are considered insertion or deletion operations, so they are added
as a penalty to the whole result.
This metric was used to assess HTR transcription results in the ICDAR
2019 competition on Recognition of Documents with Complex
Layouts~\cite{ClausnerAP19}.
In our opinion, FCA does succeed in providing a reasonable word
accuracy score which is fairly RO-independent. Nevertheless, it has
two important drawbacks.
First, it is just based on a greedy, suboptimal solution to a line
matching or assignment problem, for which the here proposed
regularized HA would provide an optimal solution.
In comparison, the approaches here proposed ensure optimal word
pairings and, moreover, they do not need to assume any kind of
LA units such as text blocks or lines.
Second, FCA heavily depends on several tunable weights. Indeed, in
the experiments, the reported results correspond to
a best-scoring combination of parameters for each algorithm run.
Clearly, this makes the method too dependent on the datasets
considered, which would become problematic for general-purpose
benchmarking of full-page transcription results.
\medskip
In addition to the above discussions, perhaps our most important
contribution to the use of the HA for HTR evaluation is to introduce a
\emph{regularized} HA version.
Thanks to the proposed regularization term, the HA not only minimizes
the character edit distance between the paired words, but also avoids
as far as possible word order mismatch, as measured by the NSFD.
Such an enhancement has allowed us to define a HA WER (hWER) which
exhibits all the desired properties:
a) it yields essentially the same results as the bag-of-words WER
(bWER) and thereby provides a proper RO-independent evaluation of
individual word recognition performance;
b) it provides practically the same results as the conventional WER,
whenever reference and system transcripts are in the same RO;
and c) it produces the alignments needed to compute a RO-independent
character error rate and used by NSFD to explicitly measure RO
mismatch.
\subsection{Integrating Evaluation of WER and Reading Order Mismatch}
\label{sec:mixROandWER}
All the works dealing with full-page, end-to-end HTR need to assess not
only word recognition performance, but also the impact of errors due
to flaws in (explicit or implicit)
LA~\cite{Bluche2016,wigington2018start,Coquenet2021,%
Singh2021,Peng2022,coquenet2022dan}.
Of course, the main focus in these works is on the proposed HTR
methods; so they do not generally pay much attention to how to
properly measure the performance they achieve.
A popular idea is to measure LA errors using conventional LA metrics
and then make do with conventional WER or CER to measure word or
character recognition errors. Finally, both measures are somehow
combined to obtain a single scalar figure which hopefully represents an
``overall performance'' metric~\cite{coquenet2022dan}.
In a similar vein, but explicitly devoted to HTR evaluation, the work
presented in~\cite{LeifertLGL19},goes deeper in the metric combination
idea, with daunting mathematical formulation. However, this is a
utterly theoretical work which does not provide any empirical evidence
that would support the proposed formulation or methods in practice.
As we see it, the metric combination idea has several drawbacks: 1) as
discussed throughout this paper, if reference and system transcripts
are \emph{not} in the same RO, conventional WER or CER systematically
provide misleading word recognition performance values -- and any
combination of misleading values is obviously also misleading; 2)
metric combination requires adequately tuned weights which are
impossible to adjust for general-purpose benchmarking; and 3) the
required GT is expensive because of the effort entailed by manual
annotation of LA geometric details.
Another idea that has been adopted in some
works~\cite{sanchez:2017,wigington2018start,tensmeyer2019training} is
to assess the overall quality of system transcripts using the so
called ``BLEU'' measure~\cite{papineni2002bleu}. It is borrowed from
the field of Machine Translation and is based on matching n-gram
frequencies of the system transcript with those of the GT reference.
While this idea avoids the complications and exceedingly high cost of
taking into account LA geometric details, it does suffer form the same
problems of directly using the conventional WER; namely, it jumbles
errors coming from different flaws and it often fails to provide the
kind of insights needed for system improvement.
In contrast with the methods discussed above, the evaluation framework
we propose, develop and assess in this paper, favors a two-fold
evaluation approach which completely decouples intrinsic word
recognition errors from errors due to wrong RO caused by poor
(explicit or implicit) LA.
Before closing this section it is worth to cite the work presented
in~\cite{Strobel0CSHS22}, which aims at assessing HTR results without
resorting to GT reference annotations. While this is indeed an
interesting prospect, it is completely unrelated with the aims and
methods discussed in this paper.
\section{Concluding Remarks}\label{sec:concl}
In classical HTR experiments each relevant text-line image is given and
accuracy is adequately assessed using conventional WER and CER.
When moving to an end-to-end full-page transcription scenario,
page-level accuracy is often being assessed using two very different
metrics: geometric accuracy of layout analysis \emph{and} WER/CER.
We consider that this assessment approach is doubly misleading.
First, geometric accuracy seldom matches well with logical relation
between relevant image elements (text lines). Second, WER values are
systematically tainted with false word recognition errors caused by
well recognized words which are not placed in in the ``correct'' order
given by a page-level reference transcript.
We argue that methods which aim at end-to-end processing, or at full
integration of layout analysis with word recognition at page-image
level, need assessment criteria which do \emph{not} rely on any kind
of geometric accuracy.
Having this in mind, we have proposed page-level assessment approaches
which: a) are geometry agnostic, b) provide a measure of word
recognition accuracy which does not depend on word reading order, and
c) provides a measure of logical mismatch of transcription elements
(words or lines) which is largely independent on the accuracy with
which individual words are recognized..
As a basic, simple and computationally cheap method to assess word
recognition errors with independence of reading-order, we advocate for
a \emph{reformulated} version of the popular bag-of-words WER, which
we refer to as bWER.
In addition, we have introduced another reading-order independent WER,
called hWER, which is based on a \emph{new}, \emph{regularized}
version of the Hungarian Algorithm.
Both bWER and hWER provide almost identical results, but hWER is much
more computationally expensive. However, the proposed regularized
Hungarian Algorithm underlying the hWER also produces word alignments
which can be used to compute specific reading-order metrics such as
the Normalized Spearman Footrule Distance (NSFD). Moreover, if system
and references transcripts are in the same reading order, these
alignments very closely approach the traditional word-to-word
sequential ``traces'' underlying the word edit distance assumed in the
classical WER.
The proposed methods are analyzed both formally and with the help of
illustrative examples, as well as through a series of partially
simulated experiments.
Finally we have applied state-of-the-art line detection and HTR
methods to a good number of popular benchmark tasks and assessed the
achieved end-to-end accuracy using the proposed metrics.
An important conclusion from both simulated and real assessment
experiments is that the bWER is ideal in practice to assess the
performance of recognizing individual words, with full independence of
how these words are ordered in the reference transcripts or in the HTR
transcription hypotheses. Moreover, empirical evidence also shows
that bWER is almost identical to the classical WER in the classical,
simplified HTR experimental setting where the same reading order for
reference and system transcripts is (rather artificially) guaranteed.
Another important conclusion is that the difference between WER and
bWER ($\Delta\!\WER$) is a very good indicator of the amount of logical
or reading-order mismatch between reference and system transcripts.
Our experiments show that this difference graciously correlates,
almost linearly, with the NSFD, which explicitly measures the
reading-order mismatch.
Thanks to this correlation, the NSFD (which is rather complex and
requires alignments yield by the expensive Hungarian Algorithm)
becomes largely unnecessary. So, both the individual word recognition
accuracy \emph{and} the degree of logical or reading order mismatch
between (page-level) transcripts, can be assessed using just the
well-known WER and (the properly redefined version of) bWER.
\medskip
\emph{Therefore, our closing recommendation for benchmarking end-to-end
full-page transcription systems is to provide these two assessment
figures:} $\bWER$ and $\Delta\!\WER \,\eqdef\, \WER\!-\!\bWER$.\!\!
\medskip
Although both WER and bWER are simple and well known, in
\ref{ap:software} we provide publicly available software to reliably
compute these two metrics, along with the other auxiliary metrics we
have used in this work, the Regularized Hungarian Algorithm WER (hWER)
and the Normalized Spearman Footrule Distance (NSFD), based on the
hWER.%
Following the concepts and results here presented, in future works we
aim to develop adequate loss functions that allow training end-to-end
HTR systems which explicitly optimize the here proposed assessment
criteria.
\section*{Acknowledgments}
This paper is part of the I+D+i projects: PID2020-118447RA-I00
(MultiScore) and PID2020-116813RB-I00a (SimancasSearch), funded by
MCIN/AEI/10.13039/501100011033. The second author is supported by a
Mar\'{i}a Zambrano grant from the Spanish Ministerio de Universidades
and the European Union NextGenerationEU/PRTR. The third author is
supported by grant ACIF/2021/356 from the ``Programa I+D+i de la
Generalitat Valenciana''.
\bibliographystyle{plain}
{\small
|
\section{Introduction}\label{sec1}
The development of a new drug is a complex process with high cost and risk.
It usually takes billions of dollars and ten or more years for an innovative drug to be developed and finally put on the market~\cite{wouters2020estimated}.
Innovation in synthetic chemistry plays an essential role in accelerating the process of drug discovery, providing the opportunity to gain more rapid access to biologically active, complex molecular structures in a cost-effective manner that can change the practice of medicine~\cite{blakemore2018organic,campos2019importance}.
During the first half of the twentieth century, most synthetic routes were developed by selecting appropriate starting materials after a trial-and-error search for suitable reactions to transform the chosen starting materials to the desired product. In most cases, this approach was strongly dependent on the assumed starting point and did not allow an efficient search for the desired synthetic route~\cite{corey1991multistep}.
To design more reliable and efficient synthesis routes, retrosynthesis~\cite{corey1988robert,corey1991multistep}, which starts from the desired product and converts it to simpler precursor structures without defining the starting materials beforehand, has been an important technique for solving the planning of organic synthesis.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.98\textwidth]{images/overview_wider.pdf}
\caption{\textbf{The overview of single-step and multi-step retrosynthesis driven by artificial intelligence.}
The retrosynthesis reaction of 5-(Bromomethyl)-2,4-dichloropyridine is shown in the single-step retrosynthesis, and an incomplete AND/OR tree is shown in the multi-step retrosynthesis.
`M' nodes are OR nodes and represent molecules.
`R' nodes are AND nodes and represent the candidate reactions.
}
\label{fig:overview}
\end{figure*}
\begin{tcolorbox}[float*=t, floatplacement=t, width=\textwidth]
\subsection*{Box 1: Glossary of selected terms}
\textbf{Atom Mapping, Reaction Template, and Synthon.} The atom mapping numbers of the reactant and product represent the assignment of the atom before and after the reaction. By utilizing this information, researchers can extract the reaction template and synthon.
The reaction template is the substructures of molecules that actually participate in the reaction, and the synthon is the reactant substructure obtained by breaking the reaction center bond of the product.
\textbf{ECFP.} Extended-Connectivity Fingerprint (ECFP)~\cite{rogers2010extended} is a novel class of topological molecular fingerprints~\cite{todeschini2008handbook} and developed explicitly for activity modeling. It is a bit vector that contains indexed elements encoding physicochemical or structural properties, which allows it to be widely adopted for different analysis tasks, such as similarity searching, clustering, and virtual screening.
\textbf{SMILES.} Following the depth-first traversal over the molecular graph, a molecule can be represented by a string of ASCII characters called simplified molecular input line entry system~(SMILES). It can contain~(but not require) some information of stereochemistry, such as isomeric specification, configurations around tetrahedral centers, and many other types of chiral centers~\cite{david2020molecular}. A molecule can have multiple valid SMILES representations called ``randomized SMILES''.
\textbf{Molecular Graph.} Molecules can be intuitively represented by graph structures, where atoms are represented by nodes and bonds by edges. Mathematically, a molecular graph can be formally defined as $G=(A,F,E)$, where $|F| = N$, $|E| = M$, and $A \in \mathbb{R}^{N\times N}$ is the adjacency matrix.
Researchers usually build node and edge feature matrices $F \in \mathbb{R}^{N\times D}$ and $E \in \mathbb{R}^{M\times K}$ according to their needs, where $D$ and $K$ are the dimensions of node and edge features.
\textbf{Deep Neural Network.} A deep neural network is an artificial neural network built by multiple computational units called layers between the input and the output. By stacking simple but many transformations, it can be extremely powerful and widely adopted for different tasks nowadays.
\textbf{Graph Neural Network.} Graph neural networks~(GNN) are deep-learning methods that operate on the graph domain and can perform a series of graph analyses such as node classification, graph classification, and link prediction.
They usually consist of multiple GNN layers that follow the following computational units: propagation, sampling, and pooling modules.
These modules determine the mechanics of message passing on the graph.
\textbf{Transformer.}
Transformer~\cite{vaswani2017attention} is the most popular machine translation model so far.
It is an end-to-end model following a stepwise and autoregressive encoder-decoder fashion. The key idea of the Transformer is the attention mechanism, which allows each token to capture the global information, which makes it quite suitable for SMILES representation.
\textbf{Data Augmentation.} Data augmentation, usually achieved by artificially representing the data in multiple forms, is a useful technique to increase the data volume, which can enrich the diversity of the data and build a more general model.
\textbf{AND/OR tree.} A search tree for multi-step retrosynthesis.
The root node of the tree is the target molecule, and the leaf nodes are usually commercially available starting materials.
An OR node represents a molecule, and an AND node represents a set of precursor molecules.
\end{tcolorbox}
In the early days of retrosynthesis, expert synthetic chemists could design synthesis routes with their familiar reactions.
However, owing to the enormous scale of chemical space, it is quite challenging to design a synthetic route for an unfamiliar molecule with a purely manual search.
A recent study~\cite{szymkuc2016computer} estimates that more than 10, 000 transformations can be considered at each synthesis step, demonstrating how complicated the decision-making process is.
With the development of computer-aided technology since the 1970s, scientists started to design computer systems to integrate chemical reaction knowledge, including LHASA~\cite{pensak1977lhasa}, SYNLMA~\cite{johnson1989designing}, WODCA~\cite{gasteiger2000computer}, and Synthia~\cite{szymkuc2016computer}, to name a few.
These systems usually manually hard-coded reaction rules of specific areas to build a Network of Organic Chemistry~(NOC)~\cite{feng2018computational}, which largely hinders their wide application in real-world scenarios.
Recently, inspired by the success of data-driven artificial intelligence in various areas such as GO playing~\cite{chen2016evolution}, medical image classification~\cite{cai2020review}, and graph generation~\cite{guo2020systematic}, researchers began to solve retrosynthesis with deep learning approaches~\cite{baskin2016renaissance,chen2018rise} that automatically solve the single-step and multi-step planning both.
Heifets~\textit{et al.}~\cite{heifets2012construction} first introduced the proof-number search, a search algorithm of artificial intelligence for organic synthesis planning, showing the feasibility of applying an automatic planning technique to solve the multi-step retrosynthesis.
After that, Segler~\textit{et al.}~\cite{segler2017neural} first studied how to apply deep neural networks~\cite{yi2016study} to select an appropriate chemical rule for the product, which achieved a top-10 accuracy of 95\% in single-step retrosynthesis on a validation set of almost 1 million reactions.
Coupling this single-step retrosynthesis method with Monte Carlo tree search~(MCTS)~\cite{browne2012mcts}, they developed a fully automated multi-step retrosynthesis technique called 3N-MCTS~\cite{segler2018planning} that performed better and faster than any previous method.
These impressive advances all demonstrate the great potential of artificial intelligence in this field.
However, since the single-step solvers of these methods are either replaced with domain experts or specific to reaction rules, they cannot escape the limitations of high labor cost and poor generalization, which leads to the search for deep learning methods that do not rely on any expert experience or predefined reaction rule.
\begin{table*}[htbp]
\centering
\caption{A taxonomy of single-step retrosynthesis methods.}
\resizebox{\textwidth}{!}{
\begin{tabular}{rrlll}
\toprule
\multicolumn{1}{l}{\textbf{Category}} & \multicolumn{1}{l}{\textbf{Goal}} & \textbf{Molecule Representation} & \textbf{Architecture/ Algorithm} & \textbf{Reported Works} \\
\midrule
\multicolumn{1}{l}{Reactant Selection} & \multicolumn{1}{l}{Select appropriate reactants} & SMILES & Bayesian inference + Transformer & \cite{guo2020bayesian} \\
& & Molecular Graph & GNN & \cite{lee2021retcl}\\
\midrule
\multicolumn{1}{l}{Template Selection} & \multicolumn{1}{l}{Select appropriate reaction template} & Molecular Fingerprint & Highway network & \cite{segler2017neural,fortunato2020data} \\
& & Molecular Fingerprint & Molecular similarity & \cite{coley2017computer}\\
& & Molecular Fingerprint & Hopfiled Netwrok & \cite{seidl2022improving}\\
& & Molecular Graph & GNN & \cite{ishida2019prediction,dai2019retrosynthesis,chen2021deep}\\
\midrule
\multicolumn{1}{l}{Semi-Template Generation} & \multicolumn{1}{l}{Break the product to synthons, and } & SMILES & Transformer & \cite{wang2021retroprime}\\
& \multicolumn{1}{l}{then generate reactants based on } & Molecular Graph & GNN & \cite{shi2020graph,somnath2021learning}\\
& \multicolumn{1}{l}{synthons} & Molecular Graph + SMILES & GNN + Transformer & \cite{yan2020retroxpert}\\
\midrule
\multicolumn{1}{l}{Template-Free Generation} & \multicolumn{1}{l}{Generate reactants using an} & SMILES & LSTM & \cite{liu2017retrosynthetic}\\
& \multicolumn{1}{l}{end-to-end approach} & SMILES & Transformer & \cite{karpov2019transformer,zheng2019predicting,chen2019learning,yang2019molecular,lin2020automatic,tetko2020state,seo2021gta,kim2021valid,irwin2022chemformer,zhong2022root}\\
& & Molecular Graph & GNN & \cite{sacha2021molecule}\\
& & Molecular Graph + SMILES & Transformer & \cite{mao2021molecular} \\
& & Customized Strings & Transformer & \cite{mann2021retrosynthesis,ucak2021substructure,ucak2022retrosynthetic}\\
\bottomrule
\end{tabular}%
}
\label{tab:single_taxonomy}%
\end{table*}%
In recent years, there have been tremendous advances in AI-driven retrosynthesis techniques, both single-step and multi-step. In Fig.~\ref{fig:overview}, we provide an overview of these techniques, including their goals, methods, and evaluation metrics.
Although the work on single-step and multi-step is often published separately from each other, we believe that single-step and multi-step retrosynthesis are closely intertwined. If the single-step solvers could provide more accurate predictions, the success rate and search time of the multi-step methods would naturally decrease accordingly. In the meantime, since single-step ones ultimately serve the multi-step, researchers of multi-step retrosynthesis could propose new metrics for the single-step to guide its development. Therefore, these two technologies should be closely integrated and inseparable. In this review, we will focus on these existing AI-based retrosynthesis methods, aiming to provide a comprehensive overview of current advances and point out some possible future research directions.
The following sections are organized as follows: for single-step and multi-step retrosynthesis both, we first list their goal and provide a thorough taxonomy of existing methods. Subsequent sections will analyze these methods in terms of their mechanism and performance and then introduce popular evaluation metrics for them, in which we also provide a detailed comparison among representative methods on several public datasets. In the next part we introduce popular databases and established platforms for retrosynthesis. Finally, this review concludes with a discussion about promising research directions in this field. A glossary of selected terms is provided in Box 1.
\section{Single-step Retrosynthesis Methods}
Generally speaking, it is ideal for single-step retrosynthesis to directly obtain the complete structure of reactants by inputting a product molecule, but this is not necessarily the goal of single-step retrosynthesis.
In fact, existing single-step methods can be divided into two categories depending on whether to predict the complete reactant structure or not, namely the selection-based and generation-based.
1) Selection-based. Similar to the previous rule-based methods, by predefining reactant or reaction template candidates to be selected, these methods formulate retrosynthesis as a selection problem.
Since they utilize chemical knowledge to varying degrees, they are more likely to make valid and successful predictions on reactions that are similar to ones in the training set. However, it is impossible for them to make predictions outside the candidates, which largely limits their generalization ability.
2) Generation-based. Taking no chemical knowledge as priors, these methods usually generate the target reactants directly in the representation of string sequences or molecular graphs, which allows them to generalize to a wider variety of reactions. Based on the goal of single-step retrosynthesis, our proposed taxonomy of deep single-step retrosynthesis methods is shown in Table.~\ref{tab:single_taxonomy}. In this section, we aim to introduce the main idea of each category and specific methods.
\subsection{Selection-based Methods}
There are two types of selection-based retrosynthesis methods, reactant selection and template selection, which utilize chemical knowledge in varying degrees. Their ultimate goal is to select the appropriate reactant or reaction template instead of directly generating the reactant, which can be formulated as follows:
\begin{equation}
{{\{r_i\}}_{i=1}^{n}}^* = \mathop{\arg\max}\limits_{{\{r_i\}}_{i=1}^{n}}(f({\{r_i\}}_{i=1}^{n},p))
\end{equation}
and
\begin{equation}
t^* = \mathop{\arg\max}\limits_{t}(f(t,p))
\end{equation}
where $p$ is the desired product, ${\{r_i\}}_{i=1}^{n}$ is the set of reactant molecules, and $t$ is the reaction template.
\subsubsection{Reactant Selection}
These retrosynthesis methods directly select appropriate reactants from the molecule candidates by calculating the matching score between the product and the possible reactant.
There are some obvious advantages over other categories:
1) The candidates can be restricted to the commercially available molecules, which allows chemists to conduct practical experiments efficiently;
2) The validity of the molecules of the reactants is guaranteed;
3) It can be easily adapted to a specific set of molecular candidates.
Although there are currently few such methods, they have already achieved comparable results against others.
Since forward reaction prediction can be regarded as the inverse task of retrosynthesis prediction, Guo \textit{et al.}~\cite{guo2020bayesian} decided to use the forward prediction results to perform the retrosynthesis prediction.
They first proposed a workflow based on the inversion of the forward model into the backward one via Bayes's law of conditional probability. Specifically, the prediction result of retrosynthesis is obtained by the probability of getting the desired product through forward reaction prediction.
The reliability of the prediction relies on the high accuracy of the forward prediction~(top-1 accuracy of 90.4\% for Molecular Transformer~\cite{schwaller2019molecular}).
In the inference stage, they selected all reactant molecules in the dataset as the candidate set.
Similarly, Lee \textit{et al.}~\cite{lee2021retcl} also utilized forward prediction to aid retrosynthesis.
However, they believed that the method should not rely exclusively on forward predictions but rather integrate both forward and backward predictions.
They introduced a framework called RetCL that first searches the reactant set candidates by a specified score function, and then ranks these possible reactant sets by a combined score of both forward and backward predictions.
Albeit these two methods, especially RetCL~\cite{lee2021retcl}, have achieved excellent results, such methods are severely limited by the candidate set.
The authors took all the reactants in the dataset~(including the test set) as candidates, implying that the correct answer must be in the candidate set when evaluating performance. This ideal situation rarely exists in real-world scenarios.
In order to make a fair comparison with other methods, its candidate set should be restricted to include only the molecules in the training set when measuring performance, which would make it much less effective. We will further elaborate on it in the Section Evaluation Methodology.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.98\textwidth]{images/template-based_new.pdf}
\caption{\textbf{The general workflow of template-based methods.}
(a) Selecting the appropriate reaction template with extracted graph features.
(b) Selecting the appropriate local reaction template with extracted node or bond features.
(c) Modeling the retrosynthesis as a content-based retrieval problem.
}
\label{fig:template-based}
\end{figure*}
\subsubsection{Template Selection}
As described in Box 1, the reaction template is a subgraph pattern that describes the changes of atoms and bonds between a product molecule and its corresponding reactant(s).
Template-based methods formulate retrosynthesis as a classification problem of selecting a proper reaction template for the target product, which can be thought of as an interpolation of known reactions to novel substrates rather than an extrapolation to novel chemistry knowledge~\cite{coley2017computer}.
Although it is quite similar to the methods based on the selection of reactants, selecting the reaction template has the following advantages over selecting reactants:
1) Only one reaction template needs to be selected instead of selecting multiple reactants.
2) When only the reactions in the training set are considered as the candidate set, the coverage of reaction templates is much higher than that of reactants.
Because of the appearance of RDChiral~\cite{coley2019rdchiral} that can extract reaction templates automatically, there is no need to manually hard-encoding reaction templates by experts compared with the traditional reaction rules. In Fig.~\ref{fig:template-based}, we display the general workflow of existing template-based methods.
Neuralsym~\cite{segler2017neural} is not only the first template-based, but also the first attempt to apply neural networks to retrosynthesis. Although it was proposed as a rule-based method initially, it has been proven to work as an effective template-based method by Dai~\textit{et al.}~\cite{dai2019retrosynthesis}.
As illustrated in Fig.~\ref{fig:template-based}a, taking ECFP molecular fingerprint~\cite{rogers2010extended} as the input and highway network~\cite{srivastava2015highway} as the deep model, it could calculate an applicable probability for all reaction templates in the candidate set, which showed better performance than the previous expert systems and great potential of AI-driven retrosynthesis. This method has been widely adopted as the single-step solver of most recent multi-step retrosynthesis methods.
Since there are some templates with few examples in the dataset, Fortunato~\textit{et al.}~\cite{fortunato2020data} proposed a pretrain and data augmentation strategy for reaction templates to expand the chemistry scope covered by the training set.
The strategy was performed by matching reaction templates extracted from the training set with products one by one to obtain an augmented training set with a many-to-many relationship between products and reaction templates. Using the same model framework as Neuralsym, a model was first pretrained on the multi-label task of matching the input molecule to multiple optional reaction templates on the augmented dataset, and then finetuned on the classification task of selecting the appropriate reaction template on the original dataset.
This data augmentation strategy effectively improved the performance of rare templates, but still failed to deal with reaction templates that were missing in the training set.
To improve the performance of few-show and zero-shot reaction template prediction, Seidl~\textit{et al.}~\cite{seidl2022improving} developed a new method that did not consider templates as distinct categories, but could leverage structural information about template.
They first proposed the concept of recasting the classification problem of reaction templates into a content-based retrieval problem.
Most existing template-based methods calculated the template matching scores by the following equation:
\begin{equation}
\pmb{\hat{y}} = softmax(\pmb{W}\pmb{h}^m(\pmb{m}))
\label{eq:non-cross-tb}
\end{equation}
where $\pmb{h}^m(\pmb{m}))$ was a neural network called \textit{molecule encoder} that mapped a molecule representation to a vector of size $d_m$. $\pmb{W} \in \mathbb{R}^{K \times d_m}$ was the last layer of the molecule encoder and $K$ was the number of reaction templates. Seidl~\textit{et al.} believed that this approach ignored the similarity among reaction templates and prevented generalization over them.
Therefore, as shown in Fig.~\ref{fig:template-based}c, they proposed the \textit{template encoder} to map each template to a vector of size $d_t = d_m$, and converted Eq.~(\ref{eq:non-cross-tb}) to the following equation:
\begin{equation}
\pmb{\hat{y}} = softmax(\pmb{h}^t(\pmb{T})\pmb{h}^m(\pmb{m}))
\label{eq:cross-tb}
\end{equation}
where $\pmb{h}^t(\pmb{T}))$ was a neural network called \textit{template encoder} that mapped the template library to a matrix of size $K\times d_t$.
To solve such a retrieval problem, they adopted a fingerprint-based modern Hopfield network~(\textit{MHN})~\cite{ramsauer2020hopfield,widrich2020modern} to associate relevant templates to product molecules. Experiments conducted by them showed that their method had a considerable advantage over others when there were no or few examples in the training set, which proved that learning the similarity among templates was important for predicting rare templates. The authors also pointed out that this framework could be applied to other methods based on graph representation or SMILES.
There are two major drawbacks for Nerualsym and its derivations:
1) Neuralsym is a completely full-connected neural network that lacks the interpretability of the prediction reason.
2) Because of the lack of complete chemical structural information, molecular fingerprints are usually considered inadequate as input for deep learning.
Aiming to address the above two issues,
Ishida~\textit{et al.}~\cite{ishida2019prediction} proposed a new framework based on graph convolutional networks~(GCN)~\cite{kipf2016semi} with an architecture-free visualization method called integrated gradients~(IG)~\cite{sundararajan2017axiomatic} for highlighting reaction-related atoms.
By representing molecules as heterogeneous graphs with atom features and introducing a more advanced neural network, it not only achieved better performances, but also succeeded in finding some atoms related to the reaction with IG.
They defined IG $\pmb{I}$ as follows:
\begin{equation}
\pmb{I}(x_t)=\frac{\Delta x_t}{M}\sum_{k=1}^M\nabla_x S_l(\frac{k}{M}(\Delta x_t+x^0)
\label{eq:IG}
\end{equation}
where $x_t$ is an atom of a molecule, $x^0$ is a reference atom whose feature matrix is a zero matrix, $M$ is the number of a division of the reference atoms, and $S_l(\frac{k}{M}(\Delta x_t+x^0)$ is a score value of the softmax layer's $l$-th neuron for the predicted reaction template.
The concurrent work of Dai~\textit{et al.}~\cite{dai2019retrosynthesis} also adopted graph representation and introduced interpretability. They proposed Conditional Graph Logic Network~(GLN), a conditional graphical model defined with logic rules, where the logic variables were chemistry knowledge about reaction templates. It was designed to solve the joint probability of the reaction template and precursors. By calculating the inner product of each atom embedding in the target molecule with the matched template subgraph embedding, they visualized the prediction of probabilistic modeling of the reaction center to interpret their models.
Instead of relying on neural networks, Coley \textit{et al.}~\cite{coley2017computer} formulated the reaction template selection as a similarity comparison problem.
Based on the molecular similarity calculated by molecular fingerprints, they first search for the most similar product to the test molecule in the training set. By applying the reaction template of that product, the possible reactant can be easily obtained. Finally, the order of reaction templates was determined by a combined score of reactant and product similarities.
The work mentioned above is all based on the global features of the input molecules to make predictions.
However, for the reason that the changes during the chemical reaction occur mostly locally,
Chen \textit{et al.}~\cite{chen2021deep} pointed out that the use of global features may yield undesired focus on the details not directly related to the reaction.
As shown in Fig.~\ref{fig:template-based}b, they designed a new template-based workflow by locally deriving the reaction templates and evaluating the suitability of these local templates at all enumerated possible reaction centers for a molecule.
As the goal of the workflow was to predict a set of local templates at each possible reaction center, the desired model should focus more on local information.
Therefore, Chen~\textit{et al.} developed a model called \textit{LocalRetro}, which is mainly composed of message passing neural network~(MPNN)~\cite{gilmer2017neural} and global reactivity attention~(GRA).
MPNN was a spatial convolutional framework that allowed atoms and bonds to learn their local environment.
GRA was a variety of multi-head self-attention applied in Transformer~\cite{vaswani2017attention} to capture the information from the remote atoms and bonds.
LocalRetro, which made good use of the nature of chemical reactions, achieved the best performance among all existing template-based methods.
\subsection{Generation-based Methods}
Generation-based retrosynthesis methods rely on no chemical knowledge, where only a single product is taken as the input, and reactants are usually represented by heterogeneous molecular graphs or SMILES strings.
According to whether the generation is done in one step or not, they can be divided into template-free methods and semi-template methods.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.98\textwidth]{images/semi-template_new.pdf}
\caption{\textbf{The general workflow of semi-template generation methods.}
(a) Modeling the product-to-synthon stage as a link prediction problem.
(b) Modeling the product-to-synthon stage as a machine translation problem.
(c) Modeling the synthon-to-reactant stage as a graph generation problem. The model iteratively determines if the generation should terminate. If not, the model would select appropriate nodes and connect them with a appropriate bond.
(d) Modeling the synthon-to-reactant stage as a leaving group selection problem.
(e) Modeling the synthon-to-reactant stage as a machine translation problem.
Theoretically the free combination between product-to-synthon methods and synthon-to-reactant methods is possible.
}
\label{fig:semi_template}
\end{figure*}
\subsubsection{Semi-Template Generation}
Since the chemical reaction in the datasets is atom-mapped, the transformations of atoms and bonds during the reaction can be automatically identified by comparing the product to their corresponding reactants. The retrosynthesis can be resolved by predicting these transformations. Following this idea, semi-template methods divide the retrosynthesis into two steps: first, identify the reaction center to get intermediate molecules called synthons, and then complete synthons to reactants.
Here for simplicity, the product is abbreviated as P, the synthon as S, and the reactant as R.These two-step transformations are denoted as P2S and S2R.
When the P2S stage succeeds, due to the correctly recognized reaction center and the one-to-one mapping between synthons and reactants, the accuracy can be much higher than other methods, which makes the accuracy of the P2S stage much more important.
As the focus of the P2S stage is to identify the reaction center, it is usually formulated as a link prediction~\cite{daud2020applications,rossi2021knowledge} task, while the S2R stage varies.
The S2R stage is aimed at completing the synthons, which can be solved by a lot of different approaches, such as directly generation, attaching leaving groups, machine translation, and so on. The general workflow of semi-template methods is shown the Fig.~\ref{fig:semi_template}.
Theoretically, the P2S and S2R approaches can be freely combined.
Shi~\textit{et al.}~\cite{shi2020graph} proposed the first semi-template generation method that is completely based on graph representation.
First, they broke the target product into several synthons and then attached new atoms to the synthons to accomplish the retrosynthesis~(Fig.~\ref{fig:semi_template}a).
To obtain the atom and bond features, they applied the Relational Graph Convolutional Network (RGCN)~\cite{schlichtkrull2018modeling} to it.
As for the graph generation in the second step, Shi~\textit{et al.} leveraged the assumption of Markov Decision Process~(MDP)~\cite{puterman1990markov}, which satisfies the Markov property that $p(S^i\vert S^{i-1},z) = p(S^i\vert S^{i-1},...,S^0,z)$.
The MDP formulation means that each action is only conditioned on the graph that has been modified so far~(Fig.~\ref{fig:semi_template}c), which implies its generation process is autoregressive.
Somnath~\textit{et al.}~\cite{somnath2021learning} combined the ideas of semi-template generation and reactant selection, proposing the GraphRetro, whose first step is to identify the reaction center~(Fig.~\ref{fig:semi_template}a), while the second step is to pick a correct leaving group to attach to the synthon~(Fig.~\ref{fig:semi_template}d). They defined the leaving groups as the substructures that can help synthons expand into valid reactants by attaching them. Although this approach cleverly circumvents the problem of generating in the semi-template generation, it also inherits the disadvantages of both the semi-template and selection-based methods, which will be further elaborated in the Evaluation Methodology Section.
Yan~\textit{et al.}~\cite{yan2020retroxpert} adopted molecular graphs and SMILES in the two steps, respectively.
First they proposed a variant of GAT~\cite{velickovic2017graph}, called EGAT, to integrate the features of chemical bonds in molecule graphs to obtain better graph representation~(Fig.~\ref{fig:semi_template}a).
Since the previous practice of judging only one most probable reaction center was only applicable to reactions with a single or two reactants, they introduced an additional task to predict the number of reaction centers, which enabled them successfully handled the reactions with three or more reactants.
To be specific, they first predicted how many reaction centers would be needed.
Based on this prediction result, select the chemical bonds that should be broken according to the probability.
In the S2R stage, they followed the sequence generation formula by treating it as a machine translation task~(Fig.~\ref{fig:semi_template}e).
Wang~\textit{et al.}~\cite{wang2021retroprime} further adopted the machine translation for both steps of semi-template generation~(Fig.~\ref{fig:semi_template}b,e). They introduced several useful approaches to improve the performance without changing the model architecture of Transformer~\cite{vaswani2017attention}, which can be concluded as three points: (1) Since only the reaction centers need to be identified in the P2S stage, which means the translation from scratch is not required, they manually tagged the atoms of the reaction centers and identified the reaction centers by predicting these tags. (2) Based on the reaction center, they also added extra label information of whether it is a reaction center and whether it is connected to the reaction center to the input SMILES of the S2R stage, which can be utilized to align the input and output SMILES. (3) The previous method only selected the most probable synthon as the input of the S2R stage, which inevitably made the overall accuracy lower than the top-1 accuracy of the P2S stage. Aiming to address it, the authors first proposed the ``Mix and Match'' strategy, selecting multiple predicted synthons as the input of the S2R stage to increase the accuracy during the inference.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.98\textwidth]{images/template-free_new.pdf}
\caption{\textbf{The general workflow and difference of template-free generation methods.}
(a) Modeling the retrosynthesis as a fully machine translation problem. There are majorly three string representations for molecules: canonical SMILES, atomic environment fingerprint representation, root-aligned SMILES.
(b) Modeling the retrosynthesis as a fully graph edit problem. This is quite similar to the graph generation problem in the synthon-to-reactant stage. The model iteratively selects the graph action. If it is a editing action, update the intermediates and select next graph action, until the termination action is selected.
}
\label{fig:template_free}
\end{figure*}
\subsubsection{Template-Free Generation}
Template-free methods, \textit{i.e.}, fully end-to-end methods, formulate retrosynthesis as a sequence generation problem.
The general workflow of template-free methods is shown the Fig.~\ref{fig:template_free}.
Depending on the representation of molecules, the sequence can be a series of SMILES tokens or molecular edit actions on the molecule graph.
By adopting SMILES representation, a popular paradigm of existing retrosynthesis methods is to formulate retrosynthesis prediction as a sequence-to-sequence translation problem~\cite{yang2020survey}.
This translation process is usually token-by-token and autoregressive, \textit{i.e.}, based on the input product SMILES and the already decoded reactant SMILES, to obtain the next possible reactant token.
Researchers first create a token dictionary with the molecules in the training set.
Based on this dictionary and neural networks, they can get a probability distribution of the next possible token by inputting the product SMILES tokens and currently decoded reactant SMILES tokens.
In order to improve the performance of retrosynthesis, researchers usually make improvements in the following perspectives: applying more robust models, using more effective training strategies, or improving the input and output of the models.
By formulating retrosynthesis as a machine translation problem, Liu~\textit{et al.}~\cite{liu2017retrosynthetic} proposed the first template-free retrosynthesis model.
They designed an Encoder-Decoder architecture based on long short-term memory (LSTM)~\cite{van2020review} cells.
Inspired by the attention mechanism~\cite{parikh2016decomposable,kim2017structured}, they calculated the attention based on the tokens of the product and treated it as an input of the decoder, achieving better performance than the rule-based expert system.
But this model architecture was soon replaced by Transformer that is entirely driven by the attention mechanism.
Inspired by successful forward reaction prediction with Transformer~\cite{schwaller2019molecular}, Karpov~\textit{et al.}~\cite{karpov2019transformer} introduced the Transformer to the retrosynthesis and achieved better results than previous LSTM-based methods, which strongly demonstrated the strength of Transformer~(Fig.~\ref{fig:template_free}a).
As retrosynthesis is modeled as a translation problem, the output sequence may be grammatically invalid sentences, \textit{i.e.} prone to be invalid output molecules.
In fact, 12.2\% of top-1 predictions and 22.0\% of top-10 predictions by the vanilla Transformer are invalid~\cite{zheng2019predicting}.
To deal with it, Zheng \textit{et al.}~\cite{zheng2019predicting} proposed a self-corrected retrosynthesis predictor~(SCROP) that is a combination of two Transformers.
The first one served as the retrosynthesis predictor like the previous method, and the other as the SMILES syntax corrector.
When training the second model, they constructed a training library that consisted of a set of input-output pairs, where the inputs were invalid SMILES predicted by the former retrosynthesis predictor, and the outputs were the ground-truth reactants.
While this approach largely reduced the invalid rate of predictions, it did not provide a significant accuracy improvement.
As mentioned above, forward reaction prediction can be treated as an inverse task of retrosynthesis prediction. Unlike Guo~\textit{et al.}~\cite{guo2020bayesian} that relied entirely on forward predictions, Wang~\textit{et al.}~\cite{wang2020forward} proposed the concept of ``forward verification'', that is, using the high-accuracy forward reaction prediction to verify the retrosynthesis prediction.
They conducted experiments on multiple sets of combined models, all successfully improving the top-1 accuracy of retrosynthesis.
The concurrent work of Kim~\textit{et al.}~\cite{kim2021valid} employed a similar concept called ``Cycle Consistency'', but further incorporated it into the network architecture and training process. They developed a Transformer-based network to train the forward reaction prediction and retrosynthesis simultaneously. Since the vast majority of network parameters for both tasks were shared, the training process could be seen as adopting a unique data augmentation strategy.
In order to generate diverse reactant candidates, they additionally introduced a learnable multinomial latent variable $z \in {1,...,K}$.
During the inference stage, they scored both forward reaction and retrosynthesis by ordering the reactant candidates based on the following likelihood:
\begin{equation}
y^* = \arg\max(p(z\vert x)p(y\vert z,x)p(\tilde{x}=x\vert z,y))
\end{equation}
Experiments showed that the proposed method further reduced invalid rates and improved the diversity of predictions compared with previous ones.
Since a molecule can have multiple valid SMILES representations, many researchers performed data augmentation to improve the performance of seq2seq models~\cite{bjerrum2017smiles}.
Among the existing works, Tetko~\textit{et al.}~\cite{tetko2020state} conducted the most detailed study on different data augmentation strategies.
They first classified data augmentation strategies based on SMILES enumeration into four categories:
1) Augmentation of products only,
2) Augmentation of products and reactants/reagents,
3) Augmentation of products and reactants/reagents followed by shuffling of the order of reactant/reagents,
4) On the basis of 3), mixing forward and reverse reactions.
Their experiments demonstrated that the more complex the data augmentation strategy was applied, the better the generalization ability of the model.
In addition, they first proposed the data augmentation applied to the test set, \textit{i.e.}, adopting different SMILES strings of a molecule as the input and then ranking all the predictions uniformly.
In our research, we found that this approach can not only effectively improve the accuracy of model prediction but also significantly reduce the invalid rate, which will be further elaborated in Section Evaluation Methodology.
The pretrain-finetune strategy of Transformer for text generation has been proven to be extremely effective~\cite{devlin2018bert}, so some researchers tried to apply it to retrosynthesis as well.
The pretrain tasks can be majorly divided into two types:
1) The input and the output represent the same or similar content, which helps the model master the syntax;
2) The output is modified to make the pretrain task closer to the downstream task.
For the first type, the existing representative pretrain methods that can be applied on SMILES are BERT~\cite{devlin2018bert} and X-MOL~\cite{xue2021x}, though they are not proposed for retrosynthesis.
BERT proposed to mask a part of tokens in the input sentence and let the network predict the masked part, which could help the model understand the meaning of various types of tokens.
X-MOL selected another valid and equivalent SMILES representation of the input molecule as the output of the network, which was aimed to allow it to learn the complex grammar rules of SMILES.
Combining these two pretrain strategies, Irwin~\textit{et al.}~\cite{irwin2022chemformer} succeeded in largely improving top-1 accuracy. However, this approach also significantly reduced the top-10 accuracy, indicating that it simultaneously reduced the diversity of model predictions.
For the second type of pretraining, Chen~\textit{et al.}~\cite{chen2019learning} proposed two different strategies to generate outputs: random pretraining and template-based pretraining. The former only involved randomly breaking a bond of the input molecule to generate the output SMILES, while the latter involved breaking a bond and attaching a new group according to a matched reaction template, both of which yielded similar improvements.
The above studies all illustrate that Transformer performed well on the retrosynthesis task. However, some researchers questioned the direct application of SMILES to the vanilla Transformer, arguing that this approach did not fully utilize both the characteristics of the molecule.
By adding an extra graph encoder, Mao~\textit{et al.}~\cite{mao2021molecular} proposed the representation fusion of SMILES sequences and molecular graphs to enrich the features.
Seo~\textit{et al.}~\cite{seo2021gta} went one step further and proposed Graph-Truncated Attention~(GTA) that could modify the attention mechanism based on molecular graph topology. More specifically, by modifying the attention masks, they forced different attention heads in self-attention to pay attention to atom pairs at a different distance, and tried to make the cross-attention mask as close as possible to the atom-mapping mask.
\begin{table*}[htbp]
\centering
\caption{A taxonomy of multi-step retrosynthesis methods.}
\resizebox{\textwidth}{!}{
\begin{tabular}{rlll}
\toprule
\multicolumn{1}{l}{\textbf{Search Algorithm}} & \textbf{Method} & \textbf{Highlights} & \textbf{Single-Step Solver} \\
\midrule
\multicolumn{1}{l}{Proof-Number Search} & PNS~\cite{heifets2012construction} & The first application of artificial intelligence to multi-step retrosynthesis & Expert \\
& DFPN-E~\cite{kishimoto2019depth} & Balance the number of AND and OR nodes & Expert \\
\midrule
\multicolumn{1}{l}{Monte Carlo Tree Search} & 3N-MCTS~\cite{segler2018planning} & Evaluate the pathway quality via double-blind AB tests & Neuralsym \\
& AutoSyn~\cite{lin2020automatic} & Combine MCTS with the template-free model Transformer & Transformer \\
& ReTReK~\cite{ishida2022ai} & Introduce retrosynthesis knowledge to guide the search direction & GCN \\
\midrule
\multicolumn{1}{l}{A* Search} & Retro*~\cite{chen2020retro} & Consider both the informed cost and estimated future value of nodes & Neuralsym \\
& RetroGraph~\cite{xie2022retrograph} & Represent the search process as a directed graph rather than a search tree & Neuralsym \\
\midrule
\multicolumn{1}{l}{Others} & Self-Improved~\cite{kim2021self} & A end-to-end framework for improve single-step solver & Neuralsym \\
& Hyper-graph~\cite{schwaller2020predicting} & Apply forward model to filter suggestions & Transformer \\
& SimulatedExp~\cite{schreck2019learning} & Apply reinforcement learning to find a optimal search policy & Neuralsym \\
\bottomrule
\end{tabular}%
}
\label{tab:multi_taxonomy}%
\end{table*}%
Another group of researchers has worked on improving SMILES, arguing that there are more suitable inputs or outputs for Transformer than the original SMILES.
As a molecule has multiple valid SMILES representations, Sumner~\textit{et al.}~\cite{sumner2020levenshtein} selected the input-output pairs with the smallest edit distance as the training data, effectively aiding the cross-attention mechanism in the Transformer.
Mann~\textit{et al.}~\cite{mann2021retrosynthesis} and Ucak~\textit{et al.}~\cite{ucak2022retrosynthetic} defined the new hand-crafted representations of the molecule that contain more features and are more suitable for machine translation~(Fig.~\ref{fig:template_free}a).
However, a huge drawback of using hand-crafted representations is that there is no guarantee that the hand-crafted representations can be converted back to SMILES.
Ucak~\textit{et al.} converted their representations to SMILES by additionally training a translation model or by similarity matching, which not only did not guarantee the success of the translation, but also added additional time consumption.
In contrast to the manually designed representations, Zhong~\textit{et al.} decided to improve SMILES into a more suitable representation for reaction prediction.
There are two common problems in previous SMILES-based retrosynthesis methods: the inability to guarantee a one-to-one mapping between input and output, especially when applying data augmentation, and the large discrepancy between input and output SMILES. To address these two issues, Zhong~\textit{et al.}~\cite{zhong2022root} proposed root-aligned SMILES~(R-SMILES), which ensured a strict one-to-one mapping and small edit distance between input and output SMILES by aligning the root atoms of them.
As shown in Fig.~\ref{fig:template_free}a, the colored tokens in the input and output SMILES represent their common parts. Combining R-SMILES with the data augmentation strategy proposed by Tetko~\textit{et al.}~\cite{tetko2020state}, they achieved the best performance among template-free methods to date.
This approach is theoretically applicable to any SMILES-based retrosynthesis method.
Unlike the above SMILES-based template-free methods, Sacha \textit{et al.}~\cite{sacha2021molecule} proposed a novel framework using molecular graph editing. As shown in Fig.~\ref{fig:template_free}b, starting from the product molecule, they performed a sequence of editing operations, including \textit{EditAtom, EditBond, AddAtom, AddBenzene, Stop}. Compared with SMILES-based methods, the graph structure provides richer information for the model. Moreover, the graph editing operations avoids generating reactants from scratch, which allows the model to produce less invalid predictions and explore chemical space efficiently.
\section{Multi-step Retrosynthesis Methods}
While single-step methods are continuously being improved, most molecules in the real world cannot be synthesized within one step. The average industrial pharmaceutical synthesis route contains 8.1 steps~\cite{carey2006analysis}, and some complex academic targets or natural products may require over 100 steps~\cite{nicolaou1996classics}. Such a colossal space poses challenges for efficient searching and planning of multi-step retrosynthesis.
In multi-step retrosynthesis, the algorithm usually needs to build a search tree or a directed acyclic graph starting from the target molecule and ending to the commercially available building blocks. Based on the optional reactions provided by single-step retrosynthesis methods, the search is guided by the designed objective function and repeated until the raw materials are all commercially available. Considering the exponentially growing search space as the route grows, it poses a considerable challenge for researchers to find a solution within the limited route length.
As the number of possible synthesis routes is often astronomical, it is desirable to identify the route that minimizes some user-specified objective functions, such as the synthetic cost and the length of the reaction route.
Depending on the adopted search algorithm, our proposed taxonomy of multi-step retrosynthesis methods is shown in Table~\ref{tab:multi_taxonomy}.
This section aims to provide a detailed review of these methods.
\subsection{Proof-Number Search}
Unlike blind traversals like depth-first traversal and breadth-first traversal, the heuristic algorithm relies on the heuristic function and the current state to decide the next action.
With the estimated distance between the current node and the target calculated by the heuristic function, the algorithm can select the most urgent node to analyze and expand the search tree.
Heifets~\textit{et al.}~\cite{heifets2012construction} are the first to model multi-step retrosynthesis planning as a discrete state-space search problem, showing it is amenable to heuristic search techniques.
By adopting the proof-number search~(PNS)~\cite{allis1994searching}, an AND/OR tree search algorithm, they modeled it as a two-player zero-sum game where one must synthesize target molecules as possible, and the other tries to prevent it. In practice, the first player picks the reaction rule to synthesize the target molecule first, and the other player selects one of the precursor molecules generated by the reaction rule. Two players alternatively play moves until the end of the game. The first player wins if and only if he can construct all precursor molecules, implying a valid synthetic route.
As illustrated in Box 1, an AND/OR consists of AND nodes representing reactions and OR nodes representing molecules. In an AND/OR tree, a node is called \textit{proven} if it is guaranteed to win for the first player. In other words, if the molecule is a available building block, an OR node is proven. Similarly, a node is called \textit{disproven} if the first player cannot synthesize it.
Kishimoto~\textit{et al.}~\cite{kishimoto2019depth} pointed out a significant search space imbalance in PNS applied to multi-step retrosynthesis prediction. Since a molecule can be synthesized by various reactions, while most reactions contain only one precursor molecule, the branching factor of OR nodes is much larger than the one of AND nodes, resulting in a lopsided search space. In the lopsided search space, PNS have difficulties in identifying moves with higher chances of leading to proofs. To address this phenomenon, they propose a depth-first proof-number search with Heuristic Edge Initialization~(DFPN-E), which assigns a heuristic cost to an edge from an OR node to an And node. However, both of these PNS methods use human-designed heuristics and do not yet exploit the full potential of artificial intelligence.
\subsection{Monte Carlo Tree Search}
Based on the proposed Neuralsym~\cite{segler2017neural}, Segler~\textit{et al.}~\cite{segler2018planning} combined three different neural networks together with Monte Carlo Tree Search~(MCTS)~\cite{browne2012mcts} to perform multi-step retrosynthesis prediction.
MCTS is a heuristic search algorithm that repeats the selection, expansion, rollout, and update steps.
(1) Selection. Starting from the root of the search tree, the algorithm selects the most urgent node on the basis of the current position values. In the context of MCTS, each node represents a set of precursor molecules.
(2) Expansion. The authors adopted two neural networks to guide the search in promising reaction rules as well as to verify whether these reaction rules are feasible or not.
(3) Rollout. If the molecules in the current position are not building blocks, authors will use the third neural network to quickly find a solution to synthesize them within the maximal depth of the search tree.
(4) Update. Depending on the difficulty of the molecule synthesis in the previous step, the algorithm receives a reward and updates the position values in the tree.
Compared with PNS, MCTS focuses more on balance between exploration and exploitation~\cite{kishimoto2019depth}.
Afterwards, Lin~\textit{et al.}~\cite{lin2020automatic} also combined MCTS with the template-free method Transformer and successfully reproduced several synthetic routes in the literature.
Ishida~\textit{et al.}~\cite{ishida2022ai} proposed a promising research direction to integrate various portions of retrosynthesis knowledge into multi-step retrosynthesis planning. On the basis of MCTS architecture proposed by Segler~\textit{et al.}~\cite{segler2018planning}, they made two major improvements: 1) replaced the highway network with a more advanced GCN-based policy network to make better single-step predictions; 2) defined a series of retrosynthesis scores to evaluate promising search directions, including a convergent disconnection score~(CDScore), an available substances score~(ASScore), a ring disconnection score~(RDScore), and a selective transformation score~(STScore). The introduction of these scores led to a search preference for convergent reactions, reactions with available reactants, ring construction reactions, and reactions with few by-products, respectively. The domain knowledge does not necessarily improve the success rate, but it can guide the algorithm in the direction that the chemists want to search.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\textwidth]{images/RetroStarSearch.pdf}
\caption{\textbf{The workflow of A* Search with an AND/OR tree.}
The `M' nodes are OR nodes and represent molecules.
The `R' nodes are AND nodes and represent candidate reactions.
In the selection step (a), the leaf nodes that are not building blocks are colored blue and the selected node is bold.
}
\label{fig:A_star_search}
\end{figure*}
\subsection{A* Search}
As previous methods depend on online value estimation and are not efficient enough, Chen~\textit{et al.}~\cite{chen2020retro} proposed the Retro* algorithm, which combined A*~\cite{hart1968formal} search with the AND/OR tree and models the retrosynthesis problem as a single-player game for focusing on the global value estimation.
Compared with the general heuristic algorithm that only estimates the cost to the target, A* search additionally considers the informed cost of the current position from the starting position, which enables it to find a better solution.
As shown in Fig.~\ref{fig:A_star_search}, A* search repeats the selection, expansion, and update steps to find a solution.
The whole process is very similar to MCTS, but with the rollout step reduced.
This is because that the authors directly predicted the future value of the nodes by an additional neural network during the selection process, which is much more efficient than the iterative rollout strategy in MCTS.
Therefore, the search algorithm can find the desired expansion direction more efficiently and generalize to the unknown reaction data.
For the previously mentioned tree search algorithm, there is a significant flaw: there could be multiple duplicated molecule nodes in a search tree, which significantly require more iterations and limit the success. Therefore, Xie~\textit{et al.}~\cite{xie2022retrograph} proposed a graph-based search method RetroGraph, which is combined with a novel GNN-guided policy to eliminate intra-target duplication and improve the success rate. By representing the search process as a directed graph, different reaction nodes can point to the same molecule nodes, i.e., share the same precursor molecules, dramatically reducing the number of iterations and the number of nodes.
\subsection{Others}
Existing methods tend to optimize single- or multi-step solvers individually, without considering both as a whole. Therefore, Kim~\textit{et al.}~\cite{kim2021self} proposed a universal end-to-end framework that aims to improve the single-step solver for maximizing the success rate of multi-step search algorithms. On top of the original single-step solver and multi-step search algorithm, they introduced two additional models: a reference backward model to discard unrealistic predictions and a forward reaction model to perform data augmentation. To be specific, After gathering a collection of reaction $C$ from the original single-step solver and multi-step search algorithm, the reference model would first discard unrealistic suggestions in $C$, and then the forward model would generate new products by inputting the reactants in $C$ to get a new collection of reaction $C'$. The single-step solver would be trained with $C \cup C'$ to make better predictions. These steps could be repeated until the model achieves optimal results. The authors conducted experiments on the Retro*~\cite{chen2020retro}, where although the accuracy of single-step model did not differ much, the success rate of reaction pathway showed a significant improvement.
To choose the most promising direction of the retrosynthesis, Schwaller~\textit{et al.}~\cite{schwaller2020predicting} proposed an additional metric, round-trip accuracy, to evaluate single-step methods. It refers to the ratio of how many of the predicted products generated by another forward reaction model is the desired product, which verifies the validity of the reaction.
By coupling it with the SCScore~\cite{coley2018scscore} that quantifies the molecular synthetic complexity, they scored the suggestions provided by the single-step solver and chose the most promising suggestion to update the search graph.
Although the vast majority of current multi-step search methods are based on heuristic search, non-heuristic algorithms are still directions worth exploring, one of which is deep reinforcement learning~\cite{li2017deep}.
Reinforcement learning~(RL) is about an agent interacting with the environment by trial and error to learn an optimal policy for sequential decision-making problems, which has made a great impact in chess, shogi, and GO playing~\cite{silver2017mastering,mnih2015human,silver2018general}.
Following this fashion, Schreck~\textit{et al.}~\cite{schreck2019learning} formulated the retrosynthesis as a single-player game. Starting from a random policy, the model is asked to estimate the accurate cost for each molecule and find the better solution by repeating the game many times, resulting in a constantly improved policy. This learned policy through simulated experience can enable the search to explore in more meaningful directions.
\section{Evaluation Methodology}
Here we aims to provide comprehensive quantitative analyses of current representative single-step and multi-step retrosynthesis algorithms on publicly available datasets.
\subsection{Single-step Retrosynthesis}
In order to rigorously determine whether one model is better or worse than another, it is necessary to develop benchmarking metrics that can be evaluated for models trained on publicly available datasets.
In this section, we experimentally make a detailed comparison among different retrosynthesis methods quantitatively and then introduce other evaluation metrics, including round-trip accuracy, class diversity, and invalid rate.
To maximize the fairness of the comparison, we followed the following principles in our experiments: 1) Use the code provided by the authors for the experiments whenever possible. If the authors did not provide the code, we would use the other faithful reimplementation.
2) In our experiments, we used the default parameters suggested by the authors, including the number of network layers, learning rate, and other hyperparameters. Although some methods can get higher performance after a careful selection of hyperparameters, we believe that the robustness of the method is also important.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.48\textwidth]{images/overall_acc_vertical.pdf}
\caption{\textbf{Top-$K$ single-step retrosynthesis results on USPTO-50K (a), USPTO-MIT (b), and USPTO-FULL (c) datasets.}
The black, orange, blue, red, purple, green color bars represent top-1, top-3, top-5, top-10, top-20, top-50 accuracy, respectively.
The black methods are reactant selection ones, the green are template selection ones, the blue are semi-template generation ones, and the red are template-free generation ones.
``Out of Memory'' means that the methods are irreproducible because they require too much computer memory.
}
\label{fig:topk_acc}
\end{figure}
\subsubsection{Datasets}
There are three popular public benchmark datasets for single-step retrosynthesis, USPTO-50K~\cite{schneider2016s}, USPTO-MIT~\cite{jin2017predicting} and USPTO-FULL~\cite{dai2019retrosynthesis}, which are all derived from the United States Patent and Trademark Office~(USPTO)-granted patents database~\cite{lowe2012extraction}.
USPTO-50K is a high-quality dataset containing about 50,000 reactions with accurate atom mappings between products and reactants.
USPTO-MIT contains about 400, 000 reactions as the training set, 30, 000 reactions as the valid set and 40, 000 reactions as the test set. Although it has much more data, the diversity of reactions is not as high as that of USPTO-50K. For example, no chiral molecules are present in this dataset.
USPTO-FULL is a much larger dataset for chemical reactions, consisting of about 1,000,000 reactions. However, as the atom mapping is obtained by the Indigo toolkit, some atom mappings are wrong, which directly leads to incorrect reaction templates and synthons.
\subsubsection{Quantitative Evaluation}
The top-$K$ accuracy is the most valuable and widely used among all the currently available metrics. It means the percentage of the correct answer appearing in the top-$K$ predictions.
This metric is not only intuitive for researchers but also usually the training task for deep models.
It is intuitive to think that the higher the top-1 accuracy, the better method, but sometimes top-10 and top-50 accuracies carry more weight, since in multi-step retrosynthesis it is valuable to provide multiple options for each step to find the global optimal solution.
In Fig.~\ref{fig:topk_acc}, we list the top-$K$ accuracy of the representative methods on different public datasets.
Considering that the reaction type is usually unavailable in realistic scenarios, the performance is all without the reaction type information.
Most of these results are obtained from their published papers, while some are obtained from additional experiments we conducted. The implementation and specific accuracy of each method are available online at https://github.com/otori-bird/DeepRetrosynthesis.
First, we do not dwell on specific methods but focus on comparisons among different categories, drawing the following conclusions: (1) In the case of a small amount of data, namely the USPTO-50K dataset~(Fig.~\ref{fig:topk_acc}a), selection-based methods generally outperform generation-based methods. This is because the chemical priors they introduce successfully guide the model's reaction predictions, i.e., limit the model's search space, while generation-based methods need to find an answer in a larger search space, which may even be an invalid molecule. (2) In the case of a huge amount of data, that is, on the USPTO-MIT and USPTO-FULL datasets~(Fig.~\ref{fig:topk_acc}b, c), we can see the effect of the selection-based method is gradually equaled or even surpassed by the generation-based method. We believe that they are constrained by the introduced chemical knowledge in turn now. Taking template-based methods as an example, for the USPTO-FULL dataset, the reaction templates in the training set can only cover about 77\% of the reactions in the test set, which means that the remaining part can never be predicted correctly. Similarly, if we limit the available reactants to the molecules in the training set, the coverage of reactant-selection methods on the test set of USPTO-50K, USPTO-MIT, and USPTO-FULL datasets would be 17.9\%, 65.2\%, and 72.8\%, respectively.
However, generative-based methods do not suffer from these limitations. (3) As the semi-template generation methods divide the retrosynthesis into two separate steps, which irreversibly makes their overall accuracy limited by the accuracy of the first step, their top-10 and top-50 accuracies are worst among different categories.
Now we are making some comparisons among specific methods. There are some interesting findings: (1) Although GLN~\cite{dai2019retrosynthesis} and LocalRetro~\cite{chen2021deep} outperform Neuralsym~\cite{segler2017neural} obviously on the USPTO-50K, what surprises us greatly is that among template-based methods, these leads are shrinking or even reversing as the amount of the data increases, which may imply that the final performance of template-based methods is less related to the representation of molecules or the model architecture when data is sufficient.
This can be interpreted in two ways:
i.) These complex network architectures accelerate the convergence but also limit the search space.
ii.) The pre-extracted chemical information in the molecular fingerprint is more helpful for the model to pick the suitable template when dealing with a huge amount of data.
(2) The similar phenomenon can also be found in comparing template-free methods. Although GTA~\cite{seo2021gta} and Tied-Transformer~\cite{kim2021valid} are variants of the Transformer, they only perform significantly better than Transformer on the USPTO-50K. Only R-SMILES~\cite{zhong2022root} keeps outperforming Transformer, which is granted as it applies data augmentation and root alignment to the Transformer without changing the model architecture.
Since the performance on large datasets is more convincing and of greater interest, it is uncertain to state that Neuralsym is worse than the later template-based methods.
(3) Compared with another semi-template method RetroPrime~\cite{wang2021retroprime}, GraphRetro~\cite{somnath2021learning} performs poorly in the two large datasets. As mentioned above, GraphRetro~\cite{somnath2021learning} is a combination of semi-template and selection-based, which makes it suffer from the limitation of both the accuracy of the P2S stage and the coverage of the leaving groups.
\subsubsection{Other Evaluation Metrics}
\begin{itemize}
\item Round-trip Accuracy: The round-trip accuracy quantifies what percentage of the retrosynthetic suggestions are considered valid by the forward model.
This indicator was formally proposed by a multi-step retrosynthesis study~\cite{schwaller2020predicting}, and its concept is now widely adopted in multiple single-step methods.
For example, Kim~\textit{et al.}'s Tied-Transformer and Lee~\textit{et al.}'s RetCL both take it as an auxiliary scoring mechanism for model predictions, i.e., their methods consider both backward and forward scores when evaluating model predictions.
We see it as an indication of the multi-step research contributing to the development of single-step research.
\item Class Diversity: Class diversity measures the diversity of the model's predictions when there is no given reaction type.
It is unrealistic for human experts to judge the reaction type for each prediction result, which naturally leads to the question of whether a neural network can be trained to judge it. By adopting the GNN-based or transformer-based reaction type classifier, Chen~\textit{et al.}~\cite{chen2019learning}, Wang~\textit{et al.}~\cite{wang2021retroprime}, and Schwaller~\textit{et al.}~\cite{schwaller2020predicting} illustrated the diversity of their predictions, respectively.
Among them, the discriminator proposed by Schwaller~\textit{et al.}~\cite{schwaller2021mapping} achieved an accuracy of 98.2\%.
\item Top-$K$ invalid Rate: Top-$K$ invalid rate means the percentage of the chemically invalid prediction appearing in the top-$K$ predictions. Since selection-based methods are impossible to provide invalid predictions, this metric is mainly used for judging generation-based methods. Predicting invalid sentences is very common in machine translation methods. For example, the top-10 invalid rate was as high as 22\% for Liu~\textit{et al.}'s seq2seq, which indirectly also determines the upper bound of top-10 accuracy. Although SCROP~\cite{zheng2019predicting} largely reduced the invalid rate by adding a syntax correcting model, it did not provide significant accuracy improvement. This problem is now addressed by the data augmentation on the test set~\cite{tetko2020state}. In our experiments, it was almost impossible for the Transformer to provide an invalid output when the test set was applied to 20$\times$ data augmentation. The only problem with this approach is the additional computation.
\end{itemize}
\begin{table*}[htbp]
\centering
\caption{Performance comparison among the multi-step metthods in the \textit{USPTO} dataset. We compare each method at the success rate of different iteration limit. The numbers of iterations, reaction (Rec.) nodes, and molecule~(Mol.) nodes are showed in average under the limit of 500.~\cite{xie2022retrograph}}
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccccccc}
\toprule
\multirow{2}[4]{*}{\textbf{Method}} & \multicolumn{5}{c}{\textbf{Success Rate of Iteration Limit (\%) ↑}} & \multirow{2}[4]{*}{\textbf{Iteration ↓}} & \multirow{2}[4]{*}{\textbf{Rec. Nodes ↓}} & \multirow{2}[4]{*}{\textbf{Mol. Nodes ↓}} \\
\cmidrule{2-6} & 100 & 200 & 300 & 400 & 500 & & & \\
\midrule
DFPN-E~\cite{kishimoto2019depth} & 50.53 & 58.42 & 64.21 & 68.42 & 75.26 & 208.12 & 3123.33 & 4635.08 \\
MCTS~\cite{segler2018planning} & 43.68 & 47.37 & 54.74 & 58.95 & 62.63 & 254.32 & - & - \\
Retro*~\cite{chen2020retro} & 52.11 & 66.32 & 76.84 & 81.05 & 86.84 & 166.72 & 2927.92 & 4174.52 \\
RetroGraph~\cite{xie2022retrograph} & \textbf{88.42} & \textbf{97.89} & \textbf{98.95} & \textbf{99.47} & \textbf{99.47} & \textbf{45.13} & \textbf{674.22} & \textbf{500.43} \\
\bottomrule
\end{tabular}%
}
\label{tab:multi_uspto}%
\end{table*}%
\begin{table*}[htbp]
\centering
\caption{Performance comparison among the multi-step metthods in the \textit{USPTO-EXT} dataset. We compare each method at the success rate of different iteration limit. The numbers of iterations, reaction (Rec.) nodes, and molecule~(Mol.) nodes are showed in average under the limit of 100.~\cite{xie2022retrograph}}
\resizebox{\textwidth}{!}{
\begin{tabular}{cccccccccc}
\toprule
\multirow{2}[4]{*}{\textbf{Method}} & \multicolumn{6}{c}{\textbf{Success Rate of Iteration Limit (\%) ↑}} & \multirow{2}[4]{*}{\textbf{Iteration ↓}} & \multirow{2}[4]{*}{\textbf{Rec. Nodes ↓}} & \multirow{2}[4]{*}{\textbf{Mol. Nodes ↓}} \\
\cmidrule{2-7} & 10 & 20 & 30 & 40 & 50 & 100 & & & \\
\midrule
Retro*~\cite{chen2020retro} & 42.47 & 48.79 & 51.84 & 53.63 & 55.00 & 57.89 & 48.49 & 790.49 & 1136.51 \\
RetroGraph~\cite{xie2022retrograph} & \textbf{50.84} & \textbf{58.05} & \textbf{62.05} & \textbf{64.26} & \textbf{66.89} & \textbf{72.89} & \textbf{37.25} & \textbf{491.97} & \textbf{373.80} \\
\bottomrule
\end{tabular}%
}
\label{tab:multi_uspto_ext}%
\end{table*}%
\subsection{Multi-step Retrosynthesis}
Due to the ambiguous definition of ``good synthesis route'', evaluating multi-step methods is much more complicated than single-step ones.
Segler~\textit{et al.}~\cite{segler2018planning} have conducted a double-blind AB test where 45 graduate-level organic chemists have to choose a better route from those in the literature and those provided by their algorithm. Although very intuitive and practical, this approach is too time-consuming and expensive to be used for evaluation on large-scale datasets.
In this section, we will present the experimental details of the multi-step retrosynthesis, including the adopted datasets, evaluation metrics, and the quantitative evaluation of several representative methods as well as a public framework for benchmarking multi-step methods.
\subsubsection{Datasets}
The starting material in multi-step retrosynthesis is usually defined as the commercially available compound~\cite{genheden2020aizynthfinder}.
eMolecules~\footnote{https://www.emolecules.com/} that consists of 231 million available molecules, usually serves as the building block library for researchers.
In addition, ZINC~\cite{irwin2020zinc20} is also a reliable database of available materials, which contains about 1.3 billion purchasable molecules.
Heifets~\textit{et al.}~\cite{heifets2012construction} created the first public benchmark dataset that contains 20 synthesis routes derived from undergraduate organic examinations.
Considering the small scale of the test data, Segler~\textit{et al.}~\cite{segler2018planning} and Kishimoto~\textit{et al.}~\cite{kishimoto2019depth} selected 497 and 897 instances for evaluation, respectively.
Based on USPTO~\cite{lowe2012extraction} database and commercially available building blocks from eMolecules,
Chen~\textit{et al.}~\cite{chen2020retro} first created a large benchmark dataset called \textit{USPTO} that contains 299,902 training routes, 65,274 validation routes, and 189 test routes.
After that, Xie~\textit{et al.}~\cite{xie2022retrograph} built an extra test set called \textit{USPTO-EXT}, whose size is ten times the size of \textit{USPTO} test set.
\subsubsection{Quantitative Evaluation}
Some researchers only tested their proposed methods on a few instances to illustrate the validity of their methods, which is contingent and unconvincing.
We believe that a method should be evaluated on a public dataset in terms of success rate and time to perform the convincing evaluation.
Thanks to the work of Xie~\textit{et al.}~\cite{xie2022retrograph}, we can display and compare recent advanced multi-step retrosynthesis search algorithms on the same dataset in Table~\ref{tab:multi_uspto} and Table~\ref{tab:multi_uspto_ext}. These methods all adopted Neuralsym as their single-step solver. The number of iterations in these tables refers to the number of single-step solver calls, which can measure the computational cost as it occupies over 99\% of the running time~\cite{chen2020retro}.
First of all, we can find that RetroGraph~\cite{xie2022retrograph} is the only one of these methods that uses directed graphs instead of search trees, which allows it to eliminate redundant intermediate molecule nodes and speed up the algorithm.
Guided by the more advanced GNN, it also outperforms others at the success rate of any iteration limit. In addition, Retro*~\cite{chen2020retro} and RetroGraph~\cite{xie2022retrograph}, which utilized A* search algorithm, achieved the most comprehensive results. On the other hand, MCTS~\cite{segler2018planning} is not only the worst in terms of success rate, but also slowest due to the four-step update process.
However, considering that the route diversity, i.e., the ability of the method to discover new synthetic routes, is not measured in these experimental data, we cannot directly conclude that MCTS is the worst method.
\subsubsection{Public Benchmark Framework PaRoutes}
As shown in Table~\ref{tab:multi_taxonomy}, most of the multi-step search methods use Neuralsym~\cite{segler2017neural} as their single-step solver for fairness of comparison, but this approach does not allow the academic community to know the power of current state-of-the-art retrosynthesis. For example, combining the best methods in single-step and multi-step retrosynthesis might outperform any previous result.
Genheden~\textit{et al.} proposed a framework called PaRoutes~\cite{genheden2022paroutes} for benchmarking multi-step methods. It examines the performance in several aspects such as route quality, search speed, and route diversity. With implemented DFPN, MCTS, and Retro* algorithms in the open-source retrosynthesis software AiZynthFinder~\cite{genheden2020aizynthfinder}, they found that DFPN is significantly inferior to others and MCTS performs slightly better than Retro* in terms of route quality and route diversity.
Since PaRoutes provides an open and fair evaluation platform for researchers, we encourage the later researchers to evaluate their methods with it.
\section{Database and Platform for Retrosynthesis}
In recent years, many mature and complete data databases have emerged. The vast amount of chemical information and the application of big data technology have provided great convenience for the researchers to develop a fully automatic computer-aided synthesis planning~(CASP) tool. Utilizing the retrosynthesis techniques mentioned above, several research groups have also developed their CASP platforms to assist chemists.
Some of these databases and platforms are open-source and publicly available, while others are supported by industrial companies and require commercial access.
In order to demonstrate the current progress of retrosynthesis, this section introduce popular reaction databases as well as several well-established retrosynthesis platforms.
\subsection{Reaction Database}
\subsubsection{Public Database}
United States Patent and Trademark Office~(USPTO)-granted patents is the predominant current open-source reaction database in the field of machine learning, containing 1,939,253 reactions that were extracted by text-mining from U.S. patents published between 1976 and 2016~\cite{lowe2012extraction}.
However, with a large number of incorrect or duplicate reactions, the data quality is quite uneven. Furthermore, the atom mapping in reactions is automatically generated by the Indigo toolkit and is not guaranteed to be correct.
Researchers usually clean their data before using this database, deriving different datasets for single-step and multi-step retrosynthesis, which will be further elaborated on below.
Cheminformatics Elsevier Melbourne University lab~(ChEMU)~\cite{he2020extended,he2020overview} is a manually annotated database of organic reaction texts in 1500 patents.
In addition to reactants and products, the annotated text also contains the reagent catalyst, solvent, time, temperature, yield, and so on.
\subsubsection{Commercial Database}
Operated by Elsevier, Reaxys~\cite{goodman2009computer} is a major commercial subscription reaction database. It enables users to search for synthetic routes of the target molecule as well as the corresponding reaction conditions like temperature and catalyst. It also supports the addition of a series of user-specified conditions to the search for reactions, such as specific substructures, yields, temperatures, etc. The information for each reaction includes the corresponding patent or literature, which allows users to better understand the reaction principle. The suppliers, prices, and shipping time of the raw materials of the synthesis route are also available.
Similar to Reaxys, Chemical Abstracts Service~(CAS) is a huge database of both organic and inorganic reactions, covering detailed information of over 145 million single-step and multi-step reactions since 1840. It also supports the reaction search with additional user-specified conditions like Reaxys.
Another subscription database is Pistachio. Similar to the USPTO database, it is also built based on a text-mining approach that extracts reaction information from patents. However, in addition to the larger data, it is maintained and kept updating by the NextMove Software. With over 9 million reactions extracted from the patents in United States and European Patent Office, it provides a friendly interface to query and analyze chemical reactions.
\subsection{Retrosynthesis Planning Platform}
\subsubsection{Public Platform}
With template-based single-step solver and the root-parallelized MCTS, Coley~\textit{et al.} developed a open-source software for CASP called ASKCOS~\cite{coley2019robotic}, which was trained with millions of reactions in both USPTO and Reaxys. Its full workflow consists of three main steps:
1) based on the reaction templates extracted from USPTO and Reaxys, ASKCOS propose a promising syntheic pathway for chemists;
2) chemists verify the route quality and determine the necessary actions for robotic arms;
3) the robotic arm completes the entire synthesis process according to the configured operating procedures. The advent of ASKCOS represents a great advancement in CASP, significantly reducing the workload of chemists on manual tasks. Now the authors have deployed the first step of retrosynthesis automatic planning on the website~\footnote{https://askcos.mit.edu/} and allowed free access.
Genheden~\textit{et al.}~\cite{genheden2020aizynthfinder} developed a another open-source CASP software AiZynthFinder, which works similarly to ASKCOS but more efficiently. They aim to provide a robust and transparent platform that not only reduces the learning cost for new users, but also facilitates researchers and integrates more retrosynthesis algorithms in the future. Now it has implemented various multi-step search algorithms, including DFPN, MCTS, and Retro*. The entire software is deployed on a python GUI rather than a web page.
IBM RoboRXN is a reaction prediction tool built on the works of Schwaller~\textit{et al.}~\cite{schwaller2019molecular,schwaller2020predicting}. Deployed on the website~\footnote{https://rxn.res.ibm.com/}, it provides the friendly interface for users to perform both forward reaction and retrosynthesis route prediction. After receiving the molecular structure drawn by the user or SMILES, the website will provide the corresponding prediction results along with the confidence. The attention weight map between the product SMILES and reactant SMILES is also available. Moreover, for retrosynthesis predictions, it allows the user to add some additional constraints and provides multiple optional results at the same time.
\subsubsection{Commercial Platform}
Most of commercial platforms are closed where algorithms, hand-coded reaction rules, or databases are unavailable.
Here we can only give some short descriptions for them.
Synthia~\cite{mikulak2020computational}, aka Chematica, is a hybrid expert-AI system that contains over 100, 000 reaction rules manually coded by chemists and also adopted machine learning algorithms.
As mature commercial platforms, Reaxys synthesis planning~\footnote{https://www.elsevier.com/en-xm/solutions/reaxys/how-reaxys-works/synthesis-planner}, CAS SciFinder$^n$~\footnote{https://www.cas.org/solutions/cas-scifinder-discovery-platform/cas-scifinder}, and ChemAIRS~\footnote{https://www.chemical.ai/} all can simultaneously provide reaction conditions, synthetic routes, and real-time purchase advice for target compounds.
Similar to the double-blind AB tests conducted by Selger~\textit{et al.}~\cite{segler2018planning}, these commercial platforms were also compared with experienced chemists.
The latest research of Synthia~\cite{mikulak2020computational} demonstrates that it has passed the Turing test, where a chemist cannot tell whether a synthetic route was designed by an experienced chemist or an AI system. Apart from that, the developes of ChemAIRS conducted a meaningful competition between ChemAIRS and 16 experienced chemists. They both were asked to propose synthetic routes for 22 molecules without chirality whose synthetic steps range from 8 to 14. The results of the competition shows that not only is the route quality of the ChemAIRS comparable to that of the chemist, but also it outperforms the chemists in terms of design speed and route diversity.
All of the above studies demonstrates that the current CASP method has reached a level comparable to that of experienced chemists.
\section{Challenges and Opportunities}
\subsection{More Reasonable Metric}
In our opinion, the current single-step retrosynthesis tasks, both single-step and multi-step, lack more reasonable and universal metrics in realistic scenarios.
While the top-$K$ accuracy and success rate successfully measure whether the model has mastered the ability to predict precursors or synthetic routes, they are incapable of judging whether the model could predict new feasible ones.
Since it is unknown whether a synthetic proposal is chemically feasible until the experiment is performed in the laboratory, the model should also focus on low cost~\cite{badowski2019selection}, high-yielding, and diversity~\cite{chen2019learning,wang2021retroprime,genheden2022paroutes}.
However, indicators for these aspects are relatively few and are not given enough attention.
Inspired by the reference model used by Kim~\textit{et al.}~\cite{kim2021self} that can eliminate unrealistic reactions, it may be possible to train a new neural network model to determine whether a reaction or reaction route meets the above requirements with existing reaction databases.
There is another problem for multi-step retrosynthesis. The current value function is built with the aim of finding the shortest path in mind. Thus, it naturally loses sight of the factors that make the path longer but necessary~(for example, the protection and deprotection strategies are incompatible). In short, it is difficult for the algorithm to reproduce the complete synthetic route in the literature due to its design principle.
\subsection{More Data}
The currently available data is far from adequate in terms of both volume and diversity.
First of all, the amount of data in the current public datasets is far from sufficient. While the number of reactions in current public datasets is around 1 million, the number of molecules with property annotations in the ZINC~\cite{irwin2020zinc20} database is over 1.4 billion.
In addition, current datasets usually contain only the successful reactions rather than the failed ones. These failed reaction data are also crucial for model training~\cite{kurczab2014influence}.
Finally, reaction conditions such as temperature and catalyst are always ignored for the existing methods, which could be critical to the feasibility of the reaction in some cases. Theoretically, these reaction conditions can guide the model predictions to some extent and serve as the predicted objects to help chemists discover feasible reaction conditions. The collection of reaction databases still has a long way to go.
\subsection{Interpretability}
A significant problem for existing AI-based retrosynthesis is the lack of interpretability.
A recent study has shown that machine learning models for chemistry applications may sometimes simply capture literature popularity trends, as illustrated by the example of predicting reactions for heterocyclic Suzuki–Miyaura coupling~\cite{beker2022machine}.
An outstanding organic synthesis technique should not only solve existing synthesis problems but also inspire chemists to design better drug molecules~\cite{campos2019importance}, which makes it essential to allow chemists to gain insights from models.
Current retrosynthesis work tends to articulate interpretability in the goal of retrosynthesis itself. For instance, some template-based methods consider the reaction template to be interpretable, and semi-template methods consider synthons to be interpretable.
While this approach interprets the retrosynthesis to some extent, it does not explain how the model makes this prediction.
Some studies mentioned in previous sections have already made notable progress in this direction.
In particular, some Transformer-based methods~\cite{sumner2020levenshtein,zhong2022root} use the cross-attention mechanism in Transformer to illustrate the relationship between reactant tokens and resultant tokens, which well explains why most substructures in the product are kept in the reactants. However, it does not account for any new atoms and bonds in the reactants.
The interpretable retrosynthesis method should be able to attribute predictions to specific atoms or substructures.
Some work has been conducted on attributing properties to molecular fragments for molecular property prediction tasks~\cite{xu2017deep,wu2021mining,jia2022explainable}.
If a similar attribution could be achieved in retrosynthesis, it would greatly help researchers to design drug molecules~\cite{murray2009rise}.
\section{Conclusion}
Over the past few years, motivated by both the development of AI and industrial demands, AI-based retrosynthesis has become an inspiring research area. Due to its automatic extraction and learning of reaction principles, it significantly outperforms traditional expert systems based on manual extraction in terms of cost and efficiency. For both single and multi-step, we can see dramatic progress in academia and some mature commercial applications, which fully demonstrate the great promise of this filed.
However, despite the tremendous progress in recent years, the area of retrosynthesis driven by artificial intelligence is far from a mature state.
The ultimate goal of retrosynthesis is to discover synthetic routes for new drugs or to discover new synthetic routes for existing drugs with a lower cost and higher yield.
However, currently we have not seen a study reporting a brand-new synthetic route that was provided entirely by artificial intelligence and successfully synthesized in the laboratory or industrial setting.
To report a plausible and novel synthetic route, it may be necessary for chemists and computer experts to cooperate with each other closely.
As more and more researchers are placed in the field of retrosynthesis, we believe that related studies will come out soon.
|
\section{Introduction}
\subsection{Ergodicity and mixing of unitary frame flows}
Let $(M,g,J)$ be a smooth closed (compact, without boundary) K\"ahler manifold with \emph{negative sectional curvature} and complex dimension $m \geq 2$. Let $SM \to M$ be the unit tangent bundle and let $F_{\mathbb{C}}M \to M$ be the principal $\mathrm{U}(m)$-bundle of \emph{unitary bases} over $M$. A point $w \in F_{\mathbb{C}}M$ over $x \in M$ is the data of an orthonormal basis $(v, \e_2, ..., \e_{m})$ of $(T_xM, h_x)$, where $h_x(\cdot, \cdot) = g_x(\cdot, \cdot) + ig_x(\cdot, J_x \cdot)$ is the canonical Hermitian inner product on the fibres of $TM$. Equivalently, we will see $F_{\mathbb{C}}M$ as a principal $\mathrm{U}(m - 1)$-bundle over $SM$ by the projection map $p : F_{\mathbb{C}}M \to SM$ defined as $p (v,\e_2,...,\e_{m}) = v$.
The \emph{geodesic flow} $(\varphi_t)_{t \in \mathbb{R}}$ on $SM$ is defined as $\varphi_t(v) :=\dot{\gamma}_{v}(t)$, where $t \mapsto \gamma_{v}(t) \in M$ is the geodesic generated by $v\in SM$. The \emph{unitary frame flow} on $F_{\mathbb{C}}M$ is then defined as
\begin{equation}
\label{equation:complex-ff}
\Phi_t(v,\e_2,...,\e_m) := (\varphi_t(v), P_{\gamma_{v}(t)}\e_2, ..., P_{\gamma_{v}(t)}\e_m),
\end{equation}
where $P_{\gamma_{v}(t)} : T_x M \to T_{\gamma_{v}(t)}M$ is the parallel transport along $\gamma_{v}$ with respect to the Levi-Civita connection.
Recall that a flow $(\Phi_t)_{t \in \mathbb{R}}$ on a compact metric space $\mathcal{M}$ is said to be \emph{ergodic} with respect to an invariant probability measure $\mu$ if any flow-invariant function $f \in L^2(\mu)$ is constant. It is said to be \emph{mixing} if for all $f_1, f_2 \in L^2(\mathcal{M},\mu)$,
\[
\lim_{t\to+\infty}\int_{\mathcal{M}} f_1\cdot (f_2 \circ \Phi_t)\, \mathrm{D} \mu = \int_\mathcal{M} f_1\, \mathrm{D} \mu \cdot\int_\mathcal{M} f_2\, \mathrm{D} \mu.
\]
While the geodesic flow $(\varphi_t)_{t \in \mathbb{R}}$ of any negatively curved compact Riemannian manifold is well-known to be ergodic \cite{Hopf-36,Anosov-67} with respect to the Liouville measure on $SM$, on negatively curved compact Kähler manifolds, the ergodicity of the unitary frame flow $(\Phi_t)_{t \in \mathbb{R}}$ with respect to the natural flow-invariant smooth measure $\omega$ induced by the Liouville measure and the Haar measure on the group $\mathrm{U}(m-1)$, is a much more difficult question due to its lack of uniform hyperbolicity. It was proved by Brin-Gromov \cite{Brin-Gromov-80} that this flow is ergodic whenever $m:=\dim_{\mathbb{C}} M$ is odd or $m=2$ but the even-dimensional case $m\ge 4$ has remained open so far. The aim of this paper is to bring a first positive answer when $m\ge 6$ is even and $m \neq 28$, under some pinching hypothesis for the sectional curvature.
Recall that the \emph{holomorphic sectional curvature} of $(M,g,J)$ is defined as
\begin{equation}
\label{equation:holomorphic-intro}
H(X) := R(X,JX,JX,X),
\end{equation}
for all unit vectors $X \in TM$, where $R$ is the Riemann curvature tensor of $(M,g)$. The manifold is said to be \emph{holomorphically $\lambda$-pinched}, for some $\lambda \in (0, 1]$, if there exists a constant $C > 0$ such that
\begin{equation}
\label{equation:pinching-intro}
-C \leq H \leq -C \lambda.
\end{equation}
The manifold is said to be \emph{strictly} holomorphically $\lambda$-pinched if the inequalities in \eqref{equation:pinching-intro} are strict.
In order to state our main result, we introduce the function $m \mapsto \lambda(m)$, defined for even numbers $m \geq 6$ by
\begin{equation}
\label{equation:plot}
\lambda(m): =\dfrac{308m+131}{336m+105}.
\end{equation}
The function $m \mapsto \lambda(m)$ is decreasing, $\lambda(6) = 0.9330...$ and $\lim_{m \to +\infty} \lambda(m) = \tfrac{11}{12} = 0.9166...$.
We will prove that the following holds:
\begin{theorem}
\label{theorem:main}
Let $(M,g,J)$ be a closed connected K\"ahler manifold of complex dimension $m \geq 2$, with negative sectional curvature. The unitary frame flow $(\Phi_t)_{t \in \mathbb{R}}$ on $F_{\mathbb{C}}M$ is \emph{ergodic} and \emph{mixing} with respect to the smooth measure $\omega$ if:
\begin{enumerate}[label=\emph{(\roman*)}]
\item The complex dimension $m$ is odd or $m=2$ \cite{Brin-Gromov-80},
\item The complex dimension $m\ge 6$ is even, $m \neq 28$, and the manifold is strictly holomorphically $\lambda(m)$-pinched.
\end{enumerate}
\end{theorem}
We will actually show that the unitary frame flow is ergodic if and only if it is mixing. We believe that ergodicity should hold \emph{without any pinching condition} but it is clear from the proofs that our method only works with a pinching condition close to $1$.
In the case of constant holomorphic curvature $H = -1$ (that is, on compact quotients $\Gamma \backslash \mathbb{C}\mathbb{H}^m$ of the complex hyperbolic space), the ergodicity of unitary frame flow was shown by Howe-Moore \cite{Howe-Moore-79}. In variable holomorphic curvature, besides \cite{Brin-Gromov-80} in odd complex dimensions and $m=2$, Theorem \ref{theorem:main} seems to be the first result proving ergodicity of unitary frame flows on negatively-curved Kähler manifolds of even complex dimensions $m\ge 6$.
As indicated in Theorem \ref{theorem:main}, it also seems that our technique does not apply in complex dimensions $m=4$ and $m=28$. The former case is related to the fact that $S^7$ is parallelizable, whereas the latter case is connected to an open problem in algebraic topology which is to classify reductions of the structure group of the unitary frame bundle $F_{\mathbb{C}}S^{55}$ over the sphere $S^{55}$. More precisely, we are unable to rule out the possible existence of a special $\mathrm{E}_6$-structure on $S^{55}$ and this eventually turns out to be problematic in order to run our argument, see \S\ref{section:topology} where this is further discussed.
The structure of the argument, described with more details in \S\ref{ssection:structure}, is somewhat similar to our previous article \cite{Cekic-Lefeuvre-Moroianu-Semmelmann-21} proving ergodicity of \emph{real frame flows}\footnote{In the literature, the word ``frame flow'' usually refers to what we call here the ``real frame flow'', that is, the parallel transport of all bases regardless of any almost-complex structure. We added the word ``unitary'' in the K\"ahler case and ``real'' in the Riemannian case in order to make a distinction.} on negatively-curved compact Riemannian manifolds of even real dimensions with nearly $0.25$-pinched (real) sectional curvature, thus almost answering a long-standing conjecture of Brin, see \cite[Conjecture 2.9]{Brin-82}. Nevertheless, the present article is not a mere adaptation of \cite{Cekic-Lefeuvre-Moroianu-Semmelmann-21} as we had to develop new techniques in order to take into account the specificities of the Kähler setting, see Theorem \ref{theorem:topology}, \S\ref{ssection:cntckt} or \S\ref{section:5.3} for instance. Although it is meant to be self-contained, we encourage the reader to consult \cite{Lefeuvre-21,Cekic-Lefeuvre-Moroianu-Semmelmann-21, Cekic-Lefeuvre-Moroianu-Semmelmann-22} as we build here on the framework developed in these articles.
Prior to \cite{Cekic-Lefeuvre-Moroianu-Semmelmann-21}, the real frame flow was known to be ergodic on odd-dimensional negatively-curved Riemannian manifolds (of dimension $\neq 7$) by Brin-Gromov \cite{Brin-75-1,Brin-Gromov-80} and in even dimensions (and dimension $7$) for manifolds with a pinching close to $1$ by Brin-Karcher \cite{Brin-Karcher-84} and Burns-Pollicott \cite{Burns-Pollicott-03}.
The real frame flows are historical examples of \emph{partially hyperbolic flows} studied in the aftermath of Anosov's seminal work on hyperbolic flows \cite{Anosov-67} by Brin and Pesin \cite{Brin-Pesin-74, Brin-75-1,Brin-75-2}. The field of partially hyperbolic dynamical systems is now a well-established and active field of dynamical systems, see \cite{Hasselblatt-Pesin-06} for instance for an introduction to this topic.
Eventually, let us mention that, similarly to the real case where ergodicity of the real frame flow was shown to determine the high-energy behaviour of eigenfunctions of Dirac-type operators \cite{Jakobson-Strohmaier-07}, the ergodicity of the unitary frame flow on Kähler manifolds determines the high-energy behaviour of eigenfunctions of Dolbeault Laplacians and Spin${}^c$ Dirac operators \cite{Jakobson-Strohmaier-Zelditch-08}.
\subsection{Proof ideas} \label{ssection:structure}
Let us summarise the argument which, as mentioned above, is similar to the one developed in \cite{Cekic-Lefeuvre-Moroianu-Semmelmann-21} and consists of three steps:
\begin{enumerate}[label=(\roman*), itemsep=4pt]
\item \textbf{Hyperbolic dynamics:} Following Brin's ideas \cite{Brin-75-1} (see also \cite{Lefeuvre-21} for a more recent approach), the non-ergodicity of the unitary frame flow is described by means of a subgroup $H \lneqq \mathrm{U}(m-1)$ called the \emph{transitivity group}, see \S\ref{ssection:isometry}. In particular, there exists a smooth flow-invariant principal $H$-subbundle $Q \subset F_{\mathbb{C}} M$ over $SM$, such that the restriction of $(\Phi_t)_{t \in \mathbb{R}}$ to $Q$ is ergodic.
\item \textbf{Algebraic topology:} The group $H$ thus provides a reduction of the structure group of $F_{\mathbb{C}}M$ from $\mathrm{U}(m-1)$ to $H$. In particular, restricting to a point $x_0 \in M$ and identifying $S_{x_0}M \simeq S^{2m-1}$, the unitary frame bundle $F_{\mathbb{C}}S^{2m-1} \to S^{2m-1}$ must admit a reduction of its structure group to $H$. In \S\ref{section:topology}, we classify such reductions and show that, for $m \neq 4,28$, $H$ must act reducibly on $\mathbb{C}^{m-1}$.
\item \textbf{Riemannian geometry:} Using the non-Abelian Liv\v sic Theorem of \cite{Cekic-Lefeuvre-21-1}, we then deduce that there exists a smooth flow-invariant complex vector bundle $\mathcal{V} \to SM$ which is a subbundle $\mathcal{V} \subset \pi^*TM$ (where $\pi : SM \to M$ is the projection) satisfying certain algebraic properties. In turn, using the twisted Pestov identity (see Lemma \ref{lemma:pestov}), we rule out the existence of such an object under a certain pinching condition $\lambda > \lambda(m)$ in \S\ref{section:pestov} and \S\ref{section:threshold}.
\end{enumerate}
\subsection{Structure of the article} In \S\ref{section:preliminaries}, we recall standard facts from Riemannian and complex geometry, and (partially) hyperbolic dynamical systems needed in the rest of the article. In \S\ref{section:topology}, we study the possible reductions of the structure group of the unitary frame bundle over the sphere, and deduce the existence of non-zero flow-invariant projectors whenever the frame flow is not ergodic. In \S\ref{section:pestov}, we derive, using the twisted Pestov identity, an inequality that must be satisfied by such an invariant object. In \S\ref{section:threshold}, we complete the proof of Theorem \ref{theorem:main}.
\subsection*{Acknowledgements} We thank Maxime Zavidovique, one of the participants of the \emph{Geometry and Topology seminar} in Jussieu, for pointing out that this problem could be worth studying (although we first thought that it had already been solved a long time ago)! M.C. has received funding from an Ambizione grant (project number
201806) from the Swiss National Science Foundation.
\section{Preliminaries}
\label{section:preliminaries}
\subsection{Riemannian geometry of the unit tangent bundle} Let $(M,g)$ be a closed connected Riemannian manifold of real dimension $n$. Denote by
\[
SM := \left\{ v \in TM ~|~ |v|_g =1\right\}
\]
the unit tangent bundle of $(M,g)$ and by $\pi : SM \to M$ the projection map.
\subsubsection{Tangent bundle of $SM$}
Let $(\varphi_t)_{t \in \mathbb{R}}$ be the geodesic flow on $SM$ with generating vector field $X \in C^\infty(SM,T(SM))$. The tangent bundle $T(SM)$ splits as
\begin{equation}
\label{equation:splitting-tsm}
T(SM) = \mathbb{V} \oplus \mathbb{H} \oplus \mathbb{R} X,
\end{equation}
where $\mathbb{V} := \ker \mathrm{D}\pi$ is the \emph{vertical bundle}, and $\mathbb{H}$ is the \emph{horizontal bundle} defined by means of the Levi-Civita connection, see \cite[Chapter 1]{Paternain-99}. The metric $g$ induces a canonical metric on $T(SM)$ called the \emph{Sasaki metric} such that the splitting \eqref{equation:splitting-tsm} is orthogonal.
If $f \in C^\infty(SM)$ is a smooth function, its gradient $\nabla f \in C^\infty(SM,T(SM))$ computed with respect to the Sasaki metric splits according to \eqref{equation:splitting-tsm} as
\[
\nabla f = \nabla_\mathbb{V} f + \nabla_{\mathbb{H}} f + (Xf) X,
\]
where $\nabla_\mathbb{V} f \in C^\infty(SM, \mathbb{V})$ is the \emph{vertical gradient} and $\nabla_{\mathbb{H}}f \in C^\infty(SM,\mathbb{H})$ is the \emph{horizontal gradient}. The $L^2$-norm on $SM$ is defined using the \emph{Liouville measure} $\mu$ on $SM$ which is the Riemannian measure induced by the Sasaki metric.
Note that the Liouville measure is \emph{invariant} by the geodesic flow.
The \emph{vertical Laplacian} $\Delta_{\mathbb{V}}$ is then defined as $\Delta_{\mathbb{V}} := \nabla_{\mathbb{V}}^* \nabla_{\mathbb{V}}$, where $\nabla_{\mathbb{V}}^*$ denotes the $L^2$-adjoint. Equivalently, given $f \in C^\infty(SM)$ and $x \in M$, denoting the Laplacian of the restriction of $g_x$ to $S_xM$ by $\Delta_{S_xM}$, we hav
\begin{equation}
\label{equation:deltav}
\Delta_{\mathbb{V}}f(v) = \Delta_{S_xM}(f|_{S_xM})(v),\qquad\forall v\in S_xM.
\end{equation}
\subsubsection{Fourier decomposition in the fibers} Since $\pi : SM \to M$ is a sphere bundle, any smooth function $f \in C^\infty(SM)$ can be decomposed into a sum of spherical harmonics on the sphere $S_xM \simeq S^{n-1}$ above each point $x \in M$. In other words, we can write
\[
f = \sum_{k=0}^{+\infty} f_k,
\]
where $f_k \in C^\infty(SM)$ is a spherical harmonic of degree $k \geq 0$, that is, it satisfies the eigenvalue equation
\[
\Delta_{\mathbb{V}} f_k = k(n+k-2) f_k,
\]
where $\Delta_{\mathbb{V}}$ is the vertical Laplacian on each sphere $S_xM$ (for $x \in M$) introduced in \eqref{equation:deltav}. The space of spherical harmonics of degree $k$ over $M$ defines a vector bundle $\Omega_k \to M$ which can be naturally identified with the vector bundle of trace-free symmetric $k$-tensors $S^k_0 TM \to M$ via the map (here we identify $T^*M$ and $TM$ by using the metric $g$)
\begin{equation}
\label{equation:iso}
\pi_k^* : S^k_0 TM \overset{\sim}{\longrightarrow} \Omega_k, \qquad \pi_k^* f (v) := f_x(v,...,v),\qquad\forall v\in S_xM.
\end{equation}
More generally, let $(E, h) \to M$ be a Hermitian (or Euclidean) vector bundle over $M$ equipped with a unitary (or orthogonal) connection $\nabla^E$, by which we mean that
\[Y h(e, f) = h(\nabla^E_Ye, f) + h(e, \nabla^E_Y f), \quad \forall e, f \in C^\infty(M, E), \quad \forall Y \in C^\infty(M, TM).\]
Denoting by $(\mathcal{E},\nabla^{\E}) := (\pi^* E,\pi^*\nabla^{E})$ its pullback to $SM$, any section $f \in C^\infty(SM,\mathcal{E})$ can be uniquely decomposed into a sum of twisted spherical harmonics over each point $x \in M$, that is,
\begin{equation}
\label{equation:sum}
f = \sum_{k=0}^{+\infty} f_k,
\end{equation}
where $f_k \in C^\infty(SM,\mathcal{E})$ is a spherical harmonic of degree $k$ (with values in $\mathcal{E}$). Note that, with respect to an orthonormal basis $(\e_\alpha)$ on $E$ defined locally over $U \subset M$, any section $f \in C^\infty(SM,\mathcal{E})$ can be written as $f|_{U} = \sum_{\alpha} f_\alpha \e_\alpha$, where $f_\alpha \in C^\infty(SM|_U)$. Then, the vertical Laplacian is defined as
\[
\Delta_{\mathbb{V}}^E f = \sum_\alpha (\Delta_{\mathbb{V}} f_\alpha) \e_\alpha,
\]
where $\Delta_{\mathbb{V}}$ was introduced in \eqref{equation:deltav}. The sections $f_k \in C^\infty(SM,\mathcal{E})$ then satisfy the eigenvalue equation
\[
\Delta_{\mathbb{V}}^E f_k = k(n+k-2) f_k.
\]
Equivalently, $f_k$ is a smooth section of the bundle $\Omega_k \otimes E$ over $M$ and this can be identified via \eqref{equation:iso} to an element $S^k_0 TM \otimes E$. We say that a section $f \in C^\infty(SM,\mathcal{E})$ has \emph{even} (resp. \emph{odd}) Fourier degree, if the decomposition \eqref{equation:sum} only involves spherical harmonics of even (resp. odd) degree.
We define the operator $\mathbf{X} := \nabla^{\E}_X$, where $X$ is the geodesic vector field on $SM$. This is the infinitesimal generator of the parallel transport of sections of $E$ along geodesic flow-lines. It has the mapping property
\begin{equation}
\label{equation:split}
\mathbf{X} : C^\infty(M,\Omega_k \otimes E) \to C^\infty(M,\Omega_{k-1} \otimes E) \oplus C^\infty(M,\Omega_{k+1} \otimes E),
\end{equation}
and therefore splits as a sum $\mathbf{X} := \mathbf{X}_- + \mathbf{X}_+$, where each term corresponds to the two summands in \eqref{equation:split}.
There is a natural $L^2$ scalar product on sections $f,f' \in C^\infty(SM, \E)$ given by:
\begin{equation}
\label{equation:l2}
\langle f,f'\rangle_{L^2} :=\int_{SM}h_{\pi(v)}(f(v),f'(v))\,\mathrm{D}\mu,
\end{equation}
where $\mu$ is the Liouville measure on $SM$, and $h$ is the Hermitian (or Euclidean) metric on $E$.
\subsubsection{Twisted Pestov identity} This identity will play a fundamental role in our proof of Theorem \ref{theorem:main}. In the non-twisted case, it was first discovered in some particular cases by Mukhometov \cite{Mukhometov-75,Mukhometov-81} and Amirov \cite{Amirov-86}, and then in its classical shape by Pestov and Sharafutdinov \cite{Pestov-Sharafutdinov-88, Sharafutdinov-94}. Eventually, it was reformulated and described in a general coordinate-free way in \cite{Paternain-Salo-Uhlmann-15, Guillarmou-Paternain-Salo-Uhlmann-16}. It takes the following form:
\begin{lemma}[Localized Pestov identity]
\label{lemma:pestov}
Let $(M,g)$ be a closed $n$-dimensional Riemannian manifold and let $(E,h)$ be a Hermitian (or Euclidean) vector bundle over $M$ equipped with unitary (or orthogonal) connection $\nabla^E$. The following identity holds for all $k \in \mathbb{Z}_{\geq 0}$ and $u \in C^\infty(M,\Omega_k \otimes E)$:
\begin{equation}
\label{equation:pestov}
\begin{split}
\tfrac{(n+k-2)(n+2k-4)}{n+k-3} \|\mathbf{X}_-u\|^2_{L^2}& - \tfrac{k(n+2k)}{k+1} \|\mathbf{X}_+u\|^2_{L^2} + \|Z(u)\|^2_{L^2} \\
&= \langle R\nabla_{\mathbb{V}}^{E}u, \nabla_{\mathbb{V}}^{E}u \rangle_{L^2} + \langle \mathcal{F}^{E}u, \nabla_{\mathbb{V}}^{E}u \rangle_{L^2},
\end{split}
\end{equation}
where:
\begin{itemize}
\item $Z$ is a first order differential operator which we do not make explicit,
\item the term involving $R$ takes the form
\[
\langle R\nabla_{\mathbb{V}}^{E}u, \nabla_{\mathbb{V}}^{E}u \rangle_{L^2} = \int_{M} \int_{S_xM} \sum_\alpha R(v, \nabla_{\mathbb{V}}u_{\alpha},\nabla_{\mathbb{V}}u_{\alpha},v) ~|\mathrm{D} v| |\mathrm{D} x|,
\]
where $R$ is the Riemann curvature tensor, $|dv|$ is the Lebesgue measure on the sphere $S_xM$ (induced by $g_x$) and $|dx|$ is the Riemannian measure, $u = \sum_\alpha u_\alpha \e_\alpha$ with $(\e_\alpha)_{\alpha \in I}$ an orthonormal frame at the point $x \in M$ of $E_x$,
\item the term involving $\mathcal{F}^{E}$ takes the form
\begin{equation}
\label{equation:f}
\langle \mathcal{F}^{E}u, \nabla_{\mathbb{V}}^{E}u \rangle_{L^2} = \int_M \int_{S_xM} \sum_{\alpha} R_E(v,\nabla_{\mathbb{V}} u_\alpha,u,\e_\alpha) ~ |\mathrm{D} v| |\mathrm{D} x|,
\end{equation}
where $R_E$ is the curvature tensor of $E$ and we use the convention:
\[
R_E(X,Y,\omega,\eta) := h(R_E(X,Y)\omega,\eta), \quad \forall X,Y \in TM,\ \forall \omega, \eta \in E,
\]
and $h$ is the Euclidean (or Hermitian) metric on the bundle $E$.
\end{itemize}
\end{lemma}
We refer to \cite[Proposition 3.5]{Guillarmou-Paternain-Salo-Uhlmann-16} for a proof.
\subsection{Complex geometry} We use \cite[Chapter IX]{Kobayashi-Nomizu-96} as basic reference for complex geometry.
\subsubsection{Curvature tensors}\label{sssc:curvature-tensors} Let $(V, g)$ be a Euclidean vector space of dimension $n$. We will usually identify $V$ with its dual $V^*$ and $\Lambda^2 V $ with the space of skew-symmetric endomorphisms of $V$ using the metric $g$. We denote by $S^p V$ the symmetric $p$-tensors on $V$, $S^p_0 V$ the trace-free symmetric tensors and $\Lambda^pV$ the $p$-th exterior power. The space $V^{\otimes 2}$ splits as
\begin{equation}
\label{equation:v2}
V^{\otimes 2} = \mathbb{R} g \oplus S^2_0 V \oplus \Lambda^2 V,
\end{equation}
where each summand is invariant under the $\mathrm{O}(n)$-action, and $S^2 V = \mathbb{R} g \oplus S^2_0 V$. The space $V^{\otimes 2} \simeq \End V$ is equipped with the norm
\begin{equation}
\label{equation:trace}
\langle u, v \rangle = \Tr(u^\top v),
\end{equation}
where ${}^\top$ denotes the transpose operator and \eqref{equation:v2} is orthogonal with respect to \eqref{equation:trace} so that $S^2 V$ and $\Lambda^2 V$ both inherit the metric \eqref{equation:trace}.
A \emph{curvature tensor} $R$ is an element $R \in S^2(\Lambda^2 V)$ satisfying the Bianchi identity
\begin{equation}
\label{equation:bianchi}
R(X,Y,Z,W) + R(Z,X,Y,W)+ R(Y,Z,X,W) = 0, \quad \forall X,Y,Z,W \in V.
\end{equation}
The \emph{sectional curvature} associated to $R$ is the quadratic map $\overline R:S^2V\to \mathbb{R}$ defined by
\[
\overline{R}(X,Y) := R(X,Y,Y,X), \quad X, Y \in V.
\]
Given $X,Y \in V$, we can see $R(X,Y,\cdot,\cdot)$ as a skew-symmetric endomorphism $R(X,Y)$ as follows:
\begin{equation}
\label{equation:equivalent}
\langle R(X,Y)Z,W \rangle:= R(X,Y,Z,W).
\end{equation}
This skew-symmetric endomorphism extends as a derivation to skew-symmetric endomorphisms of the exterior, symmetric and tensor algebras of $V$, denoted respectively by $R_{\Lambda^p}(X,Y)$, $R_{S^p}(X,Y)$ and $R_{V^{\otimes p}}(X,Y)$.
In particular, it can be easily checked that:
\begin{equation}
\label{equation:curv-induced}
R_{V^{\otimes 2}}(X,Y) u = [R(X,Y), u]
\end{equation}
for every $u \in V^{\otimes 2} =\End(V)$.
The action \eqref{equation:curv-induced} is diagonal with respect to the decomposition $V^{\otimes 2} = S^2 V \oplus \Lambda^2 V$. For $X,Y \in V$ and $\omega, \eta \in \Lambda^p V$, we set
\[
R_{\Lambda^p V}(X,Y,\omega,\eta) := \langle R_{\Lambda^p V}(X,Y)\omega, \eta \rangle,
\]
where $\Lambda^p V$ is equipped with the canonical inner product. We use the analogous notation for $S^pV$.
\subsubsection{Curvature and pinching} If $(M,g)$ is a Riemannian manifold, we introduce the $(4,0)$-tensor $g\owedge g$ by:
\begin{equation}
\label{equation:curvature-h}
g \owedge g (X,Y,Z,W) := g( X,Z)g( Y,W) - g( X,W)g( Y,Z),
\end{equation}
for all $X,Y,Z,W \in TM.$
This is precisely the curvature tensor when $(M,g)$ is the real hyperbolic space $\mathbb{H}^n$, whereas if $(M,g,J)$ is the complex hyperbolic space $\mathbb{C} \mathbb{H}^m$, its curvature tensor $G$ takes the form:
\begin{equation}
\label{eq:complexhyp}
\begin{split}
4G(X, Y, Z, W) = &g\owedge g(X, Y, Z, W) + g\owedge g(X, Y, JZ,JW) \\
& \qquad +2g(X, JY)g(Z, JW),
\end{split}
\end{equation}
see \cite[Section 7, Chapter IX]{Kobayashi-Nomizu-96}. Equivalently, \eqref{eq:complexhyp} can be rewritten using \eqref{equation:equivalent} as
\begin{equation}
\label{equation:G}
4G(X,Y) = X \wedge Y + JX \wedge JY - 2 \langle X,JY\rangle J,
\end{equation}
where $X \wedge Y$ is the skew-symmetric endomorphism of $TM$ defined by $(X\wedge Y)(Z):=g(X,Z)Y-g(Y,Z)X$ for all $Z\in TM$.
If $(M,g,J)$ is any Kähler manifold, the \emph{holomorphic sectional curvature} is defined for a unit vector $X \in TM$ by:
\begin{equation}
\label{equation:holomorphic}
H(X) := \overline{R}(X,JX) = R(X,JX,JX,X).
\end{equation}
It can be easily checked that the holomorphic curvature of the complex hyperbolic space is $-1$, that is,
\[
H_{\mathbb{C}\mathbb{H}^m}(X) = G(X,JX,JX,X) = -1,
\]
for $|X|=1$. By analogy with the real case, we introduce the notion of pinching of the holomorphic curvature:
\begin{definition}[Pinched holomorphic sectional curvature]
We say that a Kähler manifold $(M^{2m},g,J)$ is negatively \emph{holomorphically $\lambda$-pinched} (for some $\lambda \in (0,1]$) if there exists a constant $C > 0$ such that for all unit vectors $X \in TM$, we have
\begin{equation}
\label{equation:holomorphic-pinching}
-C \leq H(X) \leq -C \lambda.
\end{equation}
The manifold is said to be \emph{strictly} negatively holomorphically $\lambda$-pinched if the above inequalities are strict.
\end{definition}
Similarly, one can talk about the \emph{real} (or \emph{sectional}) $\delta$-pinching of the manifold $(M^{2m},g)$ by requiring that \eqref{equation:holomorphic-pinching} holds with $\lambda$ being replaced by $\delta$, and $H(X)$ being replaced by the sectional curvature $\overline{R}(X,Y)$ (for all pairs of orthogonal unit vectors $X,Y \in TM$). There exist relations between holomorphic and real pinchings, see \cite{Berger-60-1,Berger-60-2, Bishop-Goldberg-63}.
As in the real case, it is a well known result that negative holomorphic $1$-pinching implies that $(M^{2m},g,J)$ is holomorphically isometric to a compact quotient $\Gamma \backslash \mathbb{C}\mathbb{H}^m$, where $\Gamma$ is a discrete subgroup of $\mathrm{Isom}(\mathbb{C}\mathbb{H}^m)$. In what follows, we will always assume that $(M^{2m},g,J)$ is negatively $\lambda$-holomorphically pinched and, without loss of generality, we rescale the metric such that $C=1$.
The following lemma proved in \cite[Proposition 4.2]{Bishop-Goldberg-63} will be useful:
\begin{lemma}[Bishop-Goldberg '63]
\label{lemma:bishop-goldberg}
Assume $(M^{2m},g,J)$ is a closed Kähler manifold which is negatively holomorphically $\lambda$-pinched. Consider unit vectors $X,Y \in TM$ such that $g( X,Y) = 0$ and $g( X, JY) = \cos \theta$. Then:
\begin{equation}
\label{equation:sect-pinching0}
-( 1- \tfrac{3}{4} \lambda \sin^2\theta) \leq \overline{R}(X,Y) \leq - \tfrac{1}{4}\left( 3(1+\cos^2\theta)\lambda-2\right).
\end{equation}
In particular:
\begin{equation}
\label{equation:sect-pinching}
-1 \leq \overline{R}(X,Y) \leq - \tfrac{3\lambda-2}{4},
\end{equation}
for all orthogonal unit vectors $X, Y \in TM$. If $\lambda \geq 2/3$, then $(M^{2m},g)$ is negatively $\delta$-pinched for the sectional curvature with $\delta = \tfrac{3\lambda-2}{4}$.
\end{lemma}
We now set
\begin{equation}
\label{equation:decomp-r}
R = R_0 + \tfrac{1+\lambda}{2}G,
\end{equation}
where $G$ is the curvature tensor defined in \eqref{eq:complexhyp}. The following holds:
\begin{lemma}
\label{lemma:r0}
Assume that $-1 \leq H(X) \leq -\lambda$ for all unit vectors $X\in TM$. Then, for all unit vectors $X,Y,Z,W \in TM$, one has:
\begin{equation}
\label{equation:bound-r0bar}
|R_0(X,Y,Z,W)| \leq \tfrac{4}{3}(1-\lambda).
\end{equation}
More generally, for all unit vectors $X,Y \in TM$, and for all unit $\omega,\eta \in \Lambda^p TM$ or $S^p TM$, one has:
\begin{equation}
\label{equation:bound-r0bar-general}
|(R_0)_{\Lambda^p TM}(X,Y, \omega, \eta)|, |(R_0)_{S^p TM}(X,Y, \omega, \eta)| \leq \tfrac{4p}{3}(1-\lambda).
\end{equation}
\end{lemma}
\begin{proof}
Let $X,Y$ be unit vectors such that $g( X,Y) = 0$ and $g( X, JY) = \cos \theta$. We have:
\[
\overline{R_0}(X,Y) = \overline{R}(X,Y)-\tfrac{1+\lambda}{2}\overline{G}(X,Y) = \overline{R}(X,Y)+\tfrac{1+\lambda}{2}(1-\tfrac{3}{4}\sin^2\theta).
\]
Inserting \eqref{equation:sect-pinching0} in the previous equation, we get:
\[
-\tfrac{1-\lambda}{2}(1+\tfrac{3}{4} \sin^2\theta) \leq \overline{R_0}(X,Y) \leq \tfrac{1-\lambda}{2}(2 - \tfrac{3}{4}\sin^2\theta).
\]
In particular, the previous inequalities yield:
\begin{equation}
\label{equation:r0bar}
|\overline{R_0}(X,Y)| \leq 1-\lambda.
\end{equation}
Like in the proof of \cite[Lemma 3.7]{Bourguignon-Karcher-78}, \eqref{equation:r0bar} then implies \eqref{equation:bound-r0bar} by writing $R_0(X,Y,Z,W)$ as a sum of terms only involving two vectors in the arguments. The general bound \eqref{equation:bound-r0bar-general} follows immediately from \eqref{equation:bound-r0bar} by diagonalizing over $\mathbb{C}$ the skew-symmetric endomorphism $R_0(X,Y)$.
\end{proof}
\subsection{Isometric extensions of the geodesic flow} \label{ssection:isometry} The unitary frame bundle $\widehat{\pi} : F_{\mathbb{C}}M \to SM$ is a principal $\mathrm{U}(m-1)$-bundle over $SM$. Given $a \in \mathrm{U}(m-1)$, we denote by $R_a : F_{\mathbb{C}}M \to F_{\mathbb{C}}M$ the fiberwise right-action by $a$. The unitary frame flow $(\Phi_t)_{t \in \mathbb{R}}$ is an \emph{extension} of the geodesic flow to a principal bundle in the sense that it satisfies
\[
\pi \circ \Phi_t = \varphi_t \circ \pi, \qquad R_a \circ \Phi_t = \Phi_t \circ R_a,
\]
for all $t \in \mathbb{R}, a \in \mathrm{U}(m-1)$. We will denote by $X_{F_{\mathbb{C}}M}$ its infinitesimal generator.
Initiated by the work of Brin \cite{Brin-75-1,Brin-75-2}, there is now an established theory describing the ergodic components of such an extension flow $(\Phi_t)_{t \in \mathbb{R}}$. This is achieved via the introduction of a closed subgroup $H \leqslant \mathrm{U}(m-1)$, called the \emph{transitivity group}, and defined by means of dynamical holonomies. We refer to \cite{Lefeuvre-21} for a modern construction of the transitivity group $H$. It has the following properties:
\begin{theorem}[Brin '75, \cite{Lefeuvre-21}]
\label{theorem:brin-lefeuvre}
The following holds:
\begin{enumerate}[label=\emph{(\roman*)}]
\item There exists a natural isomorphism
\begin{equation}
\label{equation:ev}
\mathrm{ev} : \ker_{L^2(\omega)} X_{F_{\mathbb{C}}M} \overset{\sim}{\longrightarrow} L^2(H\backslash \mathrm{U}(m-1)).
\end{equation}
\item There exists a principal $H$-subbundle $Q \subset F_{\mathbb{C}}M$ over $SM$ which is invariant by $(\Phi_t)_{t \in \mathbb{R}}$ such that $(\Phi_t|_{Q})_{t \in \mathbb{R}}$ is ergodic (with respect to the induced measure on $Q$).
\end{enumerate}
In particular, the unitary frame flow is ergodic if and only if $H = \mathrm{U}(m-1)$.
\end{theorem}
We refer to \cite[Corollary 3.10]{Lefeuvre-21} for further details. Obviously, by the first item, the only possibility for $\ker_{L^2} X_{F_{\mathbb{C}}M}$ to be reduced to the constants is that $H = \mathrm{U}(m-1)$. The isomorphism in \eqref{equation:ev} is simply defined by taking an arbitrary point $z_\star \in SM$ and setting for $f \in \ker_{L^2} X_{F_{\mathbb{C}}M}$,
\[
\mathrm{ev}(f) := f|_{F_{\mathbb{C}}M_{z_\star}}.
\]
Such a function turns out to be in $L^2(F_{\mathbb{C}}M_{z_\star}) \simeq L^2(\mathrm{U}(m-1))$\footnote{It is not clear \emph{a priori} that such an evaluation map is well-defined, so its definition is also part of Theorem \ref{theorem:brin-lefeuvre}.} and is invariant by the action of $H$ so it yields an element in $L^2(H\backslash \mathrm{U}(m-1))$. The second item in Theorem \ref{theorem:brin-lefeuvre} is already a strong topological constraint on the bundle $F_{\mathbb{C}}M$ and is called a \emph{reduction of the structure group}, see \S\ref{section:topology} where this is further discussed. \\
We introduce $\mathcal{N} \to SM$, the \emph{normal bundle}, to be the Euclidean bundle over $SM$ defined for $v\in S_xM$ as:
\[
\mathcal{N}(v) := \mathrm{Span}(v,Jv)^\perp,
\]
where $\perp$ denotes the orthogonal complement with respect to the Euclidean metric $g_x$. Note that $\mathcal{N}$ is equipped with the complex structure $J$. Observe that
\begin{equation}
\label{equation:n}
\pi^* TM = \mathcal{N} \oplus \mathbb{R} v \oplus \mathbb{R} Jv,
\end{equation}
so that $\mathcal{N}$ can be seen as a subbundle of the pullback bundle $\pi^*TM$. Parallel transport with respect to the Levi-Civita connection $\nabla^{\mathrm{LC}}$ of sections of $\mathcal{N}$ along geodesic flow-lines is well-defined and generated by a first order differential operator
\[
\mathbf{X}:= (\pi^*\nabla^{\mathrm{LC}})_X : C^\infty(SM,\mathcal{N}) \to C^\infty(SM,\mathcal{N}),
\]
which is formally skew-adjoint and commutes with $J$.
Other than describing the ergodic components of the unitary frame flow, the group $H$ allows to construct smooth flow-invariant objects. In what follows, we denote by $\mathrm{Vect}$ the category of finite-dimensional Euclidean vector spaces and call $\mathfrak{o} : \mathrm{Vect} \to \mathrm{Vect}$ an \emph{operation} on this category if $\mathfrak{o}$ is obtained as a finite composition of the following basic operations: tensor powers $V^{\otimes m}$ of a vector space $V$, symmetric $S^m V$ and exterior powers $\Lambda^m V$. Obviously, for any such operation $\mathfrak{o}$, $\mathfrak{o}(\mathcal{N}) \to SM$ is a well-defined Euclidean bundle still equipped with an induced generator $\mathbf{X}$ (for simplicity, we do not introduce any new notation for the generator on this bundle).
The following holds:
\begin{theorem}{\em (Non-Abelian Liv\v sic Theorem, \cite[Theorem 3.5]{Cekic-Lefeuvre-21-1}).}
\label{theorem:non-abelian-livsic}
Let
\[
\mathfrak{o} : \mathrm{Vect} \to \mathrm{Vect},
\]
be any operation on $\mathrm{Vect}$. Then, there exists an isomorphism
\[
\ker \mathbf{X} \cap C^\infty(SM, \mathfrak{o}(\mathcal{N})) \overset{\sim}{\longrightarrow} \left\{ f \in \mathfrak{o}(\mathbb{R}^{2(m-1)}) ~|~ hf = f, \forall h \in H\right\}.
\]
\end{theorem}
The isomorphism map is nothing but evaluation at an arbitrary point of $SM$ (similarly to Theorem \ref{theorem:brin-lefeuvre}, (i)). In other words, flow-invariant smooth sections of tensor products of the normal bundle correspond exactly to algebraic $H$-invariant objects on $\mathbb{R}^{2(m-1)}$. Theorem \ref{theorem:non-abelian-livsic} will allow us to generate smooth flow-invariant sections when the unitary frame flow is not ergodic. For instance, if one can show that $H \leqslant \mathrm{U}(m-1-p) \times \mathrm{U}(p) \lneqq \mathrm{U}(m-1)$, that is, $H$ acts reducibly on $\mathbb{R}^{2(m-1)} \simeq \mathbb{C}^{m-1}$, then $H$ fixes an orthogonal projector $\pi \in S^2 \mathbb{R}^{2(m-1)}$ onto a complex (i.e. $J$-invariant) space $V \subset \mathbb{R}^{2(m-1)}$. In turn, Theorem \ref{theorem:non-abelian-livsic} implies that there exists a flow-invariant complex vector bundle $\mathcal{V} \subset \mathcal{N}$ which is the same as the existence of an orthogonal projector $\pi_{\mathcal{V}} \in C^\infty(SM, S^2 \mathcal{N}) \cap \ker \mathbf{X}$ commuting with $J$.
\section{Topological reductions and flow-invariant sections}
\label{section:topology}
In what follows, we assume that the complex dimension of $M$ is even and larger than $2$, and we write it as $\dim_{\mathbb{C}} M = m = :2p+2$, with $p \geq 1$.
\subsection{Topological reductions}
By Theorem \ref{theorem:brin-lefeuvre}, if the unitary frame flow on $F_{\mathbb{C}}M$ is not ergodic, its transitivity group is a strict subgroup $H \lneqq \mathrm{U}(2p+1)$ and there exists a strict principal $H$-subbundle $Q \subset F_{\mathbb{C}}M$. This is known as a \emph{reduction of the structure group} of $F_{\mathbb{C}}M$ to $H$. Since $F_{\mathbb{C}}M$ admits a reduction to $H$, the same holds true for the restriction of the unitary frame bundle to any sphere $S_{x_0}M$, for $x_0 \in M$. In turn, as $S_{x_0}M \simeq S^{4p+3}$, this implies that the unitary frame bundle $F_\mathbb{C} S^{4p+3}$, admits a reduction of its structure group from $\mathrm{U}(2p+1)$ to $H$.
Note that $\mathrm{U}(2p+2)$ and $\mathrm{SU}(2p+2)$ act transitively on $S^{4p+3}$ with isotropy groups $\mathrm{U}(2p+1)$ and $\mathrm{SU}(2p+1)$ respectively, so we can write:
\begin{equation}
\label{equation:sphere}
S^{4p+3} = \mathrm{SU}(2p+2)/\mathrm{SU}(2p+1) = \mathrm{U}(2p+2)/\mathrm{U}(2p+1).
\end{equation}
The unitary frame bundle $F_{\mathbb{C}}S^{4p+3}$ can be identified with $\mathrm{U}(2p+2)$. Thus the subgroup $\mathrm{SU}(2p+2)$ of $ \mathrm{U}(2p+2)$, seen as
a principal $\mathrm{SU}(2p+1)$-bundle $F_{\mathbb{C},\mathrm{SU}} S^{4p+3}$ over $S^{4p+3}$ is a reduction of $F_{\mathbb{C}}S^{4p+3}$ to $\mathrm{SU}(2p+1)$.
The aim of this section is to examine the possible further reductions of $F_{\mathbb{C},\mathrm{SU}} S^{4p+3}$. Note that as far as the spheres $S^{4p+1}$ are concerned (for $p \geq 1$), it was proved in \cite{Leonard-71} that their special unitary frame bundle $F_{\mathbb{C},\mathrm{SU}} S^{4p+1} \to S^{4p+1}$ does not admit any reduction.
\begin{theorem}
\label{theorem:topology}
Let $p \geq 1$. Assume that the principal $\mathrm{SU}(2p+1)$-bundle $F_{\mathbb{C},\mathrm{SU}} S^{4p+3}$ over $S^{4p+3}$ admits a reduction of its structure group to a strict connected subgroup $H_0 \lneqq \mathrm{SU}(2p+1)$. Then one of the following holds:
\begin{enumerate}[label=\emph{(\roman*)}]
\item \label{item:1} Either the representation of $H_0$ on $\mathbb{C}^{2p+1}$ is reducible;
\item \label{item:1,5} Or $p=1$, and $H_0$ is contained in $\mathrm{SO}(3) \lneqq \mathrm{SU}(3)$;
\item \label{item:2} Or $p=13$, and $H_0$ is contained in $\mathrm{E}_6 \lneqq \mathrm{SU}(27)$.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $\hat H_0$ be a maximal strict subgroup of $\mathrm{SU}(2p+1)$ containing $H_0$. Clearly the principal bundle $P_0=\mathrm{SU}(2p+2)\to S^{4p+3}$ also reduces to $\hat H_0$. Then \cite[Theorem 3]{Leonard-71} applied (with the notation of \cite{Leonard-71}) to $G_{2p + 1} := \mathrm{SU}(2p + 1)$, shows that $\hat H_0$ is a simple Lie group.
If $\hat H_0$ is a classical simple Lie group and $p\ge 2$, by \cite[Theorem 2.1, (D) and (E)]{Cadek-Crabb-06} applied to $G=\hat H_0$, we immediately obtain that the representation of $\hat H_0$ (and thus also the one of $H$) on $\mathbb{C}^{2p+1}$ is reducible. The above result does not hold for $p=1$ (when the corresponding sphere $S^7$ is parallelizable), but it is easy to check that the only simple Lie group strictly contained in $\mathrm{SU}(3)$ whose representation on $\mathbb{C}^3$ is irreducible, is $\mathrm{SO}(3)$, embedded in $\mathrm{SU}(3)$ via the complexification of its standard representation on $\mathbb{R}^3$. This corresponds to case {\em \ref{item:1,5}} of Theorem \ref{theorem:topology}.
It remains to study the case where $\hat H_0$ is (a finite quotient of) one of the 5 exceptional simple compact Lie groups.
First of all, it suffices to look at complex irreducible representations of the exceptional Lie groups of odd dimension $2p+1$. Moreover, by \cite[Proposition 3.1]{Cadek-Crabb-06}, writing $4p+3=\dim \hat H_0 + k + 1$ for some integer $k$, there must exist at least $k$ vector fields on the sphere $S^{4p+3}$, so the Radon-Hurwitz number $\rho(n)$ (defined by the fact that $\rho(n) - 1$ is the maximal number of linearly independent vector fields on $S^{n-1}$) satisfies
\begin{equation}\label{rh1}\rho(4p+4)\ge 4p+3-\dim \hat H_0.\end{equation}
For all $p\ge 3$ we have $\rho(4p+4)\leq 2p+3$. Since no exceptional Lie group has an irreducible complex representation of dimension less than $7$, it follows that $2p+1$ has to be the dimension of an irreducible complex representation of an exceptional Lie group $\hat H_0$, with
\begin{equation}\label{rh}7\leq 2p+1\leq \dim \hat H_0+1.\end{equation}
Denoting by $\widehat{\mathfrak{h}}$ the Lie algebra of $\hat H_0$, it turns out that there is no complex odd-dimensional irreducible representation of an exceptional Lie group $\hat H_0$ satisfying \eqref{rh} except in the following two cases: \\
\textbf{1.} $\widehat{\mathfrak{h}}=\mathfrak{e}_6$, $\dim \widehat{\mathfrak{h}}=78$. There are two 27-dimensional irreducible representations of $\mathfrak{e}_6$ satisfying \eqref{rh}. This case corresponds to a (theoretical) reduction of the structure group $\mathrm{SU}(27)$ of $F_{\mathbb{C},\mathrm{SU}}S^{55}$ to a subgroup of $\mathrm{E}_6$ (case \emph{\ref{item:2}} of Theorem \ref{theorem:topology}). \\
\textbf{2.} $\widehat{\mathfrak{h}}=\mathfrak{g}_2$, $\dim \widehat{\mathfrak{h}}=14$. The only complex odd-dimensional irreducible representation of $\mathfrak{g}_2$ satisfying \eqref{rh} is the complexification of the real $7$-dimensional representation $\rho_7:\mathrm{G}_2\to \mathrm{SO}(7)$ given by the embedding $\mathrm{G}_2\subset \mathrm{SO}(7)$, for $p=3$. However, we will show that $F_{\mathbb{C},\mathrm{SU}}S^{15}$ does not admit any reduction to $\mathrm{G}_2$.
Indeed, if such a reduction $P_{\mathrm{G}_2}$ exists, then the tangent bundle $TS^{15}$ is isomorphic to the direct sum $\mathbb{R}\oplus (\mathbb{C}\otimes F_7)$, where $F_7:=P_{\mathrm{G}_2}\times_{\rho_7}\mathbb{R}^7$ is a real vector bundle of rank $7$ over $S^{15}$.
Now, to each rank $k$ real vector bundle $E$ over $S^{15}$ one can associate an element $\alpha(E)$ in the homotopy group $\pi_{14}(\mathrm{SO}(k))$, namely, the homotopy class of its clutching function at the equator $S^{14} \hookrightarrow S^{15}$. For $k<l$, let $f_{k,l}:\mathrm{SO}(k)\to \mathrm{SO}(l)$ be the standard embedding and denote by $g_{k,l}:\pi_{14}(\mathrm{SO}(k))\to\pi_{14}(\mathrm{SO}(l))$ the group morphisms induced by $f_{k,l}$ in homotopy. If $E$ and $F$ have ranks $k$ and $l$ respectively, we clearly have
\[
\alpha(E\oplus F)=g_{k,k+l}(\alpha(E))+g_{l,k+l}(\alpha(F)).
\]
In our situation, since $\mathbb{C}\otimes F_7$ is topologically isomorphic to $F_7\oplus F_7$, $T S^{15}$ is isomorphic to $\mathbb{R}\oplus F_7\oplus F_7$, so we
can write
\[
\begin{split}
\alpha (T S^{15}) & =g_{14,15}(\alpha(F_7\oplus F_7)) \\
&=g_{14,15}(2g_{7,14}(\alpha(F_7)))=2g_{14,15}(g_{7,14}(\alpha(F_7)))=0,
\end{split}
\]
because $\pi_{14}(\mathrm{SO}(15))=\mathbb{Z}_2$. This is a contradiction since the tangent bundle of $S^{15}$ is non-trivial. Therefore the case $\widehat{\mathfrak{h}}=\mathfrak{g}_2$ is impossible, thus finishing the proof.
\end{proof}
Combined with Theorems \ref{theorem:brin-lefeuvre} and \ref{theorem:non-abelian-livsic}, Theorem \ref{theorem:topology} yields the following:
\begin{corollary}
\label{corollary:invariant-sections}
Let $(M,g,J)$ be a closed connected Kähler manifold with even complex dimension $m$ and non-ergodic unitary frame flow. Then, if $m \neq 4, 28$, there exists a finite cover $(\widehat{M},\widehat{g},\widehat{J})$ of $(M,g,J)$ and a flow-invariant orthogonal projector $\pi_{\mathcal{V}} \in C^\infty(S\widehat{M},S^2 \mathcal{N})$ onto a complex subbundle $\mathcal{V} \subset \mathcal{N}$ of rank $1 \leq r \leq m/2-1$, of even Fourier degree.
\end{corollary}
Note that by \eqref{equation:n}, $\mathcal{N}$ is a subbundle of the pullback bundle $\pi^*TM$ so it makes sense to talk about the decomposition of a section $f \in C^\infty(SM, S^2 \mathcal{N})$ as a sum of spherical harmonics as in \eqref{equation:sum}. The fact that $\mathcal{V}$ is complex is equivalent to the commutation relation $[\pi_{\mathcal{V}},J]=0$.
\begin{proof}[Proof of Corollary \ref{corollary:invariant-sections}]
Up to replacing $M$ by a finite covering if necessary, we can assume that the transitivity group $H \leqslant \mathrm{U}(m-1)$ is connected, see \cite[Lemma 3.3]{Cekic-Lefeuvre-Moroianu-Semmelmann-21}. In what follows, in order to keep notation simple, we will still denote this finite cover by $M$.
The representation $\rho : H \to \mathrm{U}(m-1)$ induces a representation $\det \rho : H \to \mathrm{U}(1)$ whose image is either $\mathrm{U}(1)$ or $\left\{1\right\}$ (by connectedness of $H$). Following an argument of Brin-Gromov \cite{Brin-Gromov-80}, we first show that $(\det \rho)(H)= \mathrm{U}(1)$.
Indeed, assume that $(\det \rho)(H) = \left\{1\right\}$. As $(\det \rho)(H)$ is the transitivity group of the frame flow of the complex line bundle $\Lambda^{m-1,0} \mathcal{N}$, we get by the non-Abelian Liv\v sic Theorem \ref{theorem:non-abelian-livsic} that $\Lambda^{m-1,0} \mathcal{N}$ is trivial. Now, using that
\[
\Lambda^{m-1,0} \mathcal{N} \to \Lambda^{m,0} \pi^*TM, \qquad \omega \mapsto \omega \wedge (v-iJv)
\]
is an isomorphism, the triviality of $\Lambda^{m-1,0} \mathcal{N}$ implies that
\[
c_1(\Lambda^{m,0} \pi^*TM) = \pi^* c_1(\Lambda^{m,0} TM) = 0 \in H^2(SM,\mathbb{Z}).
\]
However, it can be easily checked using the Gysin sequence \cite[Proposition 14.33]{Bott-Tu-82} (and the fact that the dimension of $M$ is $n \geq 4$) that
\[
\pi^* : H^2(M,\mathbb{Z}) \to H^2(SM,\mathbb{Z})
\]
is injective, so $c_1(\Lambda^{m,0} TM) = 0 = -c_1(K_M)$, where $K_M = \Lambda^{m,0} T^*M$ is the canonical line bundle. This is impossible since $(M,g)$ has negative sectional curvature. Hence, $(\det \rho)(H)= \mathrm{U}(1)$.
We will now show that the $H$-representation $\rho$ on $\mathbb{C}^{m-1}$ is reducible. Assume for a contradiction that $\rho$ is irreducible. By the Schur Lemma, the center $C(H)$ of $H$ is contained in the set $\mathrm{U}(1)$ of scalar matrices. On the other hand, the fact that $(\det \rho)(H) = \mathrm{U}(1)$ shows that $H$ is not semi-simple, so its center is at least 1-dimensional. We thus obtain the equality $C(H)=\mathrm{U}(1)$, i.e. $H$ contains the set of scalar matrices.
We now fix an arbitrary point $x_0 \in M$, restrict the unitary frame bundle $F_{\mathbb{C}}M$ to a bundle over $S_{x_0}M$ and identify $S_{x_0}M \simeq S^{4p+3}$. As the structure group of $F_{\mathbb{C}} M$ reduces to $H$ by Theorem \ref{theorem:brin-lefeuvre}, we obtain by restriction to any fiber of $SM\to M$ that the structure group of $F_{\mathbb{C}}S^{4p+3}$ also reduces to $H$, i.e. there exists a principal $H$-bundle $P_H \subset F_{\mathbb{C}}S^{4p+3}$. We claim that $F_{\mathbb{C}}S^{4p+3}$ admits a further reduction to $H_0:=H\cap\mathrm{SU}(2p+1)$. Indeed, the principal $\mathrm{SU}(2p+1)$-bundle $F_{\mathbb{C},\mathrm{SU}}S^{4p+3}$ is already a reduction of $F_{\mathbb{C}}S^{4p+3}$ to $\mathrm{SU}(2p+1)$ and for every $v\in S^{4p+3}$, the intersection of the fibres $F_{\mathbb{C},\mathrm{SU}}S^{4p+3}(v)\cap P_H(v)$ is non-empty: if $u\in P_H(v)\subset F_{\mathbb{C}}S^{4p+3}(v)$, there exists $z\in \mathrm{U}(1)$ such that $uz\in F_{\mathbb{C},\mathrm{SU}}S^{4p+3}(v)$, and since $\mathrm{U}(1)\subset H$ (that is, $H$ contains scalar matrices), we also have $uz\in P_H(v)$, so $uz\in F_{\mathbb{C},\mathrm{SU}}S^{4p+3}(v)\cap P_H(v)$. It is then straightforward to check that $F_{\mathbb{C},\mathrm{SU}}S^{4p+3}\cap P_H$ is a principal bundle over $ S^{4p+3}$ with group $H_0$.
As $m \neq 4, 28$ by assumption, we can apply case \emph{\ref{item:1}} of Theorem \ref{theorem:topology} to deduce that $H_0 \lneqq \mathrm{SU}(m-1)$ acts reducibly on $\mathbb{C}^{m-1}$. However, as $H$ was assumed to act irreducibly on $\mathbb{C}^{m-1}$, $H_0 = H \cap \mathrm{SU}(m-1)$ also acts irreducibly on $\mathbb{C}^{m-1}$ and this is a contradiction. Therefore, $H$ acts reducibly on $\mathbb{C}^{m-1}$.
We can then conclude using the non-Abelian Liv\v sic Theorem \ref{theorem:non-abelian-livsic}: by the remark after Theorem \ref{theorem:non-abelian-livsic}, there exists a smooth (non-zero) flow-invariant orthogonal projector $\pi_{\mathcal{V}'} \in C^\infty(SM,S^2 \mathcal{N})$ onto a flow-invariant smooth complex bundle $\mathcal{V}' \subset \mathcal{N}$. Following \cite[Lemma 3.10]{Cekic-Lefeuvre-Moroianu-Semmelmann-21}, one can find a (possibly different) smooth non-zero flow-invariant orthogonal projector $\pi_{\mathcal{V}} \in C^\infty(SM,S^2 \mathcal{N})$ of \emph{even Fourier degree} onto a flow-invariant smooth complex bundle $\mathcal{V} \subset \mathcal{N}$ of complex rank $1 \leq r \leq m/2-1$.
\end{proof}
\subsection{Complex normal twisted conformal Killing tensors}
\label{ssection:cntckt}
If $m \neq 4,28$ and the unitary frame flow is not ergodic, we know by Corollary \ref{corollary:invariant-sections} that there exists a flow-invariant orthogonal projector $\pi_{\mathcal{V}} \in C^\infty(SM,S^2\mathcal{N})$ of even Fourier degree and commuting with $J$. The flow-invariance condition is equivalent to $\mathbf{X} \pi_{\mathcal{V}} = 0$. A crucial step then, is to show that such a flow-invariant section has \emph{finite Fourier degree}, that is, the decomposition \eqref{equation:sum} only involves a finite number of terms. This is the content of the following:
\begin{lemma}
\label{lemma:degree}
For any operation $\mathfrak{o} : \mathrm{Vect} \to \mathrm{Vect}$, a section $f \in C^\infty(SM,\mathfrak{o}(\mathcal{N}))$ satisfying $\mathbf{X} f = 0$ has finite Fourier degree.
\end{lemma}
The proof of Lemma \ref{lemma:degree} uses the fact that the sectional curvature of $(M,g)$ is negative and relies on the twisted Pestov identity \eqref{equation:pestov}. Lemma \ref{lemma:degree} was first obtained in \cite[Theorem 4.1]{Guillarmou-Paternain-Salo-Uhlmann-16}, see also \cite[Corollary 4.2]{Cekic-Lefeuvre-Moroianu-Semmelmann-22} for a short self-contained proof. As a consequence, we can decompose
\begin{equation}
\label{equation:piv}
\pi_{\mathcal{V}} = u_{k} + u_{k-2} + \ldots + u_2 + u_0,
\end{equation}
where $k \geq 0$ is even, $u_i \in C^\infty(M,\Omega_i \otimes S^2 TM)$ and $u_k \neq 0$. Moreover, since $J$ has degree $0$ (that is, it does not depend on the velocity variable $v$), the commutation relation $[\pi_{\mathcal{V}},J]=0$ yields $[u_i,J] = 0$ for all $i \in \left\{0, \ldots, k\right\}$.
We now set
\[
u := u_k \in C^\infty(M,\Omega_k \otimes S^2 TM),
\]
the spherical harmonic of higher degree in the decomposition \eqref{equation:piv} of $\pi_{\mathcal{V}}$. Using the mapping property \eqref{equation:split} of $\mathbf{X}$, the equation $\mathbf{X} \pi_{\mathcal{V}} = 0$ then gives $\mathbf{X}_+ u = 0$. Such a section $u$ is called a \emph{twisted conformal Killing tensor} in the literature. Moreover, since $\iota_v \pi_{\mathcal{V}} := \pi_{\mathcal{V}} v = 0$ (because $\mathcal{V}$ is orthogonal to the span of $v$ and $Jv$) and $\iota_v$ has the mapping properties
\[
\iota_v : \Omega_k \otimes S^2 TM \to (\Omega_{k-1} \otimes TM) \oplus (\Omega_{k+1} \otimes TM),
\]
we obtain that $\iota_v u \in C^\infty(M,\Omega_{k-1} \otimes TM)$ is of degree $k-1$. The same argument also shows that $\iota_{Jv} u \in C^\infty(M,\Omega_{k-1} \otimes TM)$.
Using similarly that $\iota_v \iota_v \pi_{\mathcal{V}} = \langle \pi_{\mathcal{V}} v, v \rangle = 0$, a refined algebraic argument allows to show that $\iota_v \iota_v u=\iota_{Jv} \iota_{Jv} u \in C^\infty(M,\Omega_{k-2})$ is of degree $k-2$, see \cite[Lemma 4.2]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} for a proof. Furthermore, we have $\iota_v \iota_{Jv} u = \iota_{Jv} \iota_v u = 0$, using that $u$ is symmetric and $J$ is skew-symmetric, and $[u,J]=0$.
A section $u \in C^\infty(M,\Omega_k \otimes S^2 TM)$ satisfying
\begin{equation}
\label{equation:u1}
\mathbf{X}_+ u = 0, \quad \iota_v u \text{ has degree $k-1$}, \quad \iota_v \iota_v u \text{ has degree $k-2$},
\end{equation}
was called in \cite[Section 4.1]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} a \emph{normal twisted conformal Killing tensor}. (The adjective \emph{normal} refers to the conditions on $\iota_v u$ and $\iota_v \iota_v u$.) Here, the section $u$ satisfies the extra condition $[u,J] = 0$. It is thus worth introducing the following terminology:
\begin{definition}
\label{definition:cntckt}
A section $u \in C^\infty(M,\Omega_k \otimes S^2 TM)$ satisfying \eqref{equation:u1} and $[u,J]=0$ is called a \emph{complex normal twisted conformal Killing tensor}.
\end{definition}
By Corollary \ref{corollary:invariant-sections} and the discussion above, we obtain:
\begin{corollary}\label{corollary:tensor}
Let $(M,g,J)$ be a closed connected Kähler manifold with even complex dimension $m$ and non-ergodic unitary frame flow. Then, if $m \neq 4, 28$, there exists a non-zero complex normal twisted conformal Killing tensor $u$ of even degree $k \geq 2$.
\end{corollary}
\begin{proof}
Corollary \ref{corollary:tensor} follows immediately from Corollary \ref{corollary:invariant-sections} and the above discussion, except for the point that $k \geq 2$ which we now prove. If $k = 0$, then $\pi_{\mathcal{V}} = u =u_0$ is of degree $0$ and thus $\pi_{\mathcal{V}}$ can be identified with a section in $C^\infty(M,S^2 TM)$. However, we must also have $\iota_v \pi_{\mathcal{V}} = 0$ for all $v \in TM$ by \eqref{equation:u1}, so $\pi_{\mathcal{V}} = 0$, which contradicts the non-vanishing of $\pi_{\mathcal{V}}$.
\end{proof}
The aim of the remaining sections is now to rule out the existence of such a non-zero complex twisted conformal Killing tensor of even degree $k \geq 2$ under a holomorphic pinching condition $\lambda > \lambda(m)$.
\section{Bounding the terms in the twisted Pestov identity}
\label{section:pestov}
Throughout this section, $(M,g,J)$ is a negatively-curved compact Kähler manifold of real dimension $n=2m$ with $\lambda$-pinched holomorphic curvature, and $u \in C^\infty(M,\Omega_k \otimes S^2 TM)$ is a complex normal twisted conformal Killing tensor of even degree $k \geq 2$. Our aim is to bound from above the terms appearing on the right-hand side of the twisted Pestov identity \eqref{equation:pestov}, namely, the first term $\langle R \nabla_{\V}^{S^2 TM} u, \nabla_{\V}^{S^2 TM} u \rangle_{L^2}$ and the second term $\langle \mathcal{F}^{S^2 TM}u, \nabla_{\mathbb{V}}^{S^2 TM}u \rangle_{L^2}$, and to bound from below the term $\|\mathbf{X}_-u\|^2_{L^2}$ on the left-hand side. Sometimes, it will be convenient to consider general vector bundles $E \to M$ rather than the specific bundle $S^2 TM$.
In order to simplify notation, we will drop the volume forms in the integrands. The reader should keep in mind that integrals over spheres $S_xM$ (for $x \in M$) are always computed with respect to the round measure $|dv|$ on the sphere, while integrals over $M$ are computed with respect to the Riemannian measure $|dx|$ induced by the metric $g$. Moreover, we will often work with expressions involving a local orthonormal basis of a vector bundle $E$, usually denoted by $(\e_\alpha)_\alpha$; for the simplicity of notation, when we write sums over $\alpha$ we will mean that the sums are pointwise (and the basis might change from point to point). We also introduce the following constants:
\begin{equation}
\label{equation:constants}
\begin{array}{ll}
\alpha_{n,k} := k(n+k-2), & \beta_{n,k} := (k(n+k-2)(n-1))^{1/2}, \\
\gamma_{n,k} := \dfrac{(n+k-2)(n+2k-4)k}{(n+k-3)(n+2k-2)(k-1)}, & \delta_{n,k} := n+2k-4.
\end{array}
\end{equation}
\subsection{Bounding the first term in the right-hand side}
Let $E \to M$ be a Euclidean vector bundle equipped with an orthogonal connection $\nabla^E$. Then, the following holds:
\begin{lemma}
\label{lemma:bound-r}
For all $f \in C^\infty(M,\Omega_k \otimes E)$, one has:
\[
\begin{split}
\langle R \nabla_{\V}^{E} f, \nabla_{\V}^{E} f \rangle_{L^2} \leq - \tfrac{3\lambda-2}{4} & \alpha_{n,k}\|f\|^2_{L^2} - \tfrac{3\lambda}{4} \int_{M}\int_{S_xM}\sum_{\alpha} \langle v,J\nabla_{\mathbb{V}}f_\alpha \rangle^2,
\end{split}
\]
where we write locally $f = \sum_{\alpha} f_\alpha \e_\alpha$, for $(\e_\alpha)_{\alpha \in I}$ a local orthonormal basis of $E$.
\end{lemma}
\begin{proof}
Using the upper bound \eqref{equation:sect-pinching0} on the sectional curvature from Lemma \ref{lemma:bishop-goldberg}, we have:
\[
\begin{split}
\langle R \nabla_{\V}^{E} f, \nabla_{\V}^{E} f \rangle_{L^2} & = \int_{M} \int_{S_xM}\sum_\alpha R(v,\nabla_{\V}^{E} f_\alpha,\nabla_{\V}^{E} f_\alpha,v) \\
& \leq - \tfrac{3\lambda-2}{4} \|\nabla_{\mathbb{V}}^{E} f\|^2_{L^2} - \tfrac{3\lambda}{4} \int_{M}\int_{S_xM}\sum_{\alpha} \langle v,J\nabla_{\mathbb{V}}f_\alpha \rangle^2 \\
& = - \tfrac{3\lambda-2}{4} \langle \Delta_{\mathbb{V}}^{E} f, f \rangle_{L^2} - \tfrac{3\lambda}{4} \int_{M}\int_{S_xM} \sum_{\alpha}\langle v,J\nabla_{\mathbb{V}}f_\alpha \rangle^2 \\
& = - \tfrac{3\lambda-2}{4}k(n+k-2)\|f\|^2_{L^2} - \tfrac{3\lambda}{4}\int_{M} \int_{S_xM}\sum_{\alpha} \langle v,J\nabla_{\mathbb{V}}f_\alpha \rangle^2 .
\end{split}
\]
Since $\alpha_{n,k} := k(n+k-2)$, this completes the proof.
\end{proof}
\subsection{Bounding the second term in the right-hand side}
Assume now that $E = \Lambda^p TM$ or $E = S^p TM$ for some $p \geq 1$. Using the decomposition of the Riemannian curvature tensor $R = R_0 + \tfrac{1+\lambda}{2}G$ in \eqref{equation:decomp-r}, we can write the second term on the right-hand side of the Pestov identity \eqref{equation:pestov} as:
\begin{equation}
\label{equation:decomp-f}
\langle \mathcal{F}^{E}f, \nabla_{\mathbb{V}}^{E}f \rangle_{L^2} = \langle \mathcal{F}^{E}_0f, \nabla_{\mathbb{V}}^{E}f \rangle_{L^2} + \tfrac{1+\lambda}{2} \langle \mathcal{G}^{E}f, \nabla_{\mathbb{V}}^{E}f \rangle_{L^2}.
\end{equation}
More precisely, $\mathcal{F}^{E}_0$ and $\mathcal{G}^{E}$ are defined from $R_0$ and $G$, respectively, by extending the latter to $E$ as in \S \S \ref{sssc:curvature-tensors} and using formula \eqref{equation:f}. We now study separately the two terms in \eqref{equation:decomp-f}. We start with:
\begin{lemma}
\label{lemma:f0}
If $E=\Lambda^pTM$ or $E = S^pTM$, then for all $f \in C^\infty(M,\Omega_k \otimes E)$,
\[
|\langle \mathcal{F}^{E}_0f, \nabla_{\mathbb{V}}^{E}f \rangle_{L^2}| \leq \tfrac{4p}{3}(1-\lambda) \beta_{n,k} \|f\|^2_{L^2}.
\]
\end{lemma}
Lemma \ref{lemma:f0} will be applied with $E=TM$ and $E=S^2TM$.
\begin{proof}
The proof is the same as \cite[Lemma 4.5]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} by inserting the bound \eqref{equation:bound-r0bar-general}.
\end{proof}
We now study the second term in \eqref{equation:decomp-f}.
\begin{lemma}
\label{lemma:g-expression}
Let $f \in C^\infty(M,\Omega_k \otimes E)$ such that $\iota_v f, \iota_{Jv} f$ are of degree $k-1$. Then, the following holds:
\begin{enumerate}[label=\emph{(\roman*)}]
\item If $E = TM$, one has:
\[
\begin{split}
\langle \mathcal{G}^{TM}f, \nabla_{\mathbb{V}}^{TM}f \rangle_{L^2} & = \tfrac{1}{4}\delta_{n,k} \left( \|\iota_v f\|^2_{L^2} + \|\iota_{Jv} f\|^2_{L^2} \right) + \tfrac{1}{2}\|f\|^2_{L^2} \\
& \hspace{2cm} + \tfrac{1}{2}\int_M \int_{S_xM} \sum_\alpha\langle v,J\nabla_{\mathbb{V}} f_\alpha\rangle\langle f,J\e_\alpha\rangle.
\end{split}
\]
\item If $E = S^2 TM$, and $[J,f] = 0$, one has:
\[
\begin{split}
\langle \mathcal{G}^{S^2 TM}f, \nabla_{\mathbb{V}}^{S^2 TM}f \rangle_{L^2} = \delta_{n,k} \|\iota_v f\|^2_{L^2} + \|f\|^2_{L^2}.
\end{split}
\]
\end{enumerate}
\end{lemma}
\begin{proof}
We start with the proof for $E = TM$. Note that this equality is an integral equality over $SM$. We will actually prove the integral equality over $S_xM$ for every $x \in M$, and then it suffices to integrate over $x \in M$ to obtain the result. Recall that $G$ is defined in \eqref{eq:complexhyp}. Using the expressions \eqref{equation:f} and \eqref{equation:G}, we have:
\begin{equation}
\label{equation:intermediaire0}
\begin{split}
4 \sum_\alpha G(v, \nabla_{\V} f_\alpha, f, \e_\alpha) & = \sum_\alpha\big( \langle (v \wedge \nabla_{\V} f_\alpha) f, \e_\alpha \rangle \\
& + \langle (Jv \wedge J \nabla_{\V} f_\alpha) f, \e_\alpha \rangle + 2\langle{v, J \nabla_{\V} f_\alpha}\rangle\langle{f, J \e_\alpha}\rangle\big).
\end{split}
\end{equation}
The integral over $S_xM$ of the first term on the right-hand side can be immediately computed using \cite[Lemma 4.6]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} (in the $\Lambda^p$ case with $p=1$)\footnote{We warn the reader that, in the notation of \cite{Cekic-Lefeuvre-Moroianu-Semmelmann-21}, $G$ denotes the curvature tensor of the real hyperbolic space, that is, $G = g \owedge g$. The term $\sum_\alpha \int_{S_xM} \langle v,f\rangle \langle \nabla_{\V} f_\alpha, \e_\alpha\rangle - \langle v,\e_\alpha\rangle \langle \nabla_{\V} f_\alpha, f\rangle$ thus corresponds exactly to the term computed in \cite[Equation after (4.14)]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} with $p=1$.} since $\iota_v f$ is of degree $k-1$ and yields:
\[
\begin{split}
\int_{S_xM} \sum_\alpha \langle (v \wedge \nabla_{\V} f_\alpha) f, \e_\alpha \rangle& = \int_{S_xM} \left(\sum_\alpha \langle v,f\rangle \langle \nabla_{\V} f_\alpha, \e_\alpha\rangle - \langle v,\e_\alpha\rangle \langle \nabla_{\V} f_\alpha, f\rangle\right) \\
&= (n+2k-4)\|\iota_v f\|^2_{L^2(S_xM)} + \|f\|^2_{L^2(S_xM)},
\end{split}
\]
where we use the notation
\[
\|f\|^2_{L^2(S_xM)} = \int_{S_xM} g_{x}(f(v),f(v)), \qquad \|\iota_v f\|^2_{L^2(S_xM)} = \int_{S_xM} |(\iota_v f)(v)|^2.
\]
We claim that the integral over $S_xM$ of the second term in \eqref{equation:intermediaire0} is equal to $(n+2k-4)\|\iota_{Jv} f\|^2_{L^2(S_xM)} + \|f\|^2_{L^2(S_xM)}$.
Indeed, observe first that for each $\alpha$ we have
\[\langle{(Jv \wedge J\nabla_{\V} f_\alpha) f, \e_\alpha}\rangle = \langle{(v \wedge \nabla_{\V} f_\alpha)Jf, J\e_\alpha}\rangle.\]
Now, since $f_\alpha = \langle f,\e_\alpha\rangle = \langle Jf,J\e_\alpha \rangle$, changing the basis $(\e_\alpha)$ by $(J\e_\alpha)$, we see that the second term in \eqref{equation:intermediaire0} is the same as the first term with $f$ replaced by $Jf$. Using $\iota_v Jf = -\iota_{Jv} f$, the result now follows from the previous computation. Inserting the previous equality in \eqref{equation:intermediaire0} gives the desired result (after integration over $M$) since $\delta_{n,k} = n+2k-4$.
We now deal with the case $E = S^2TM$. As before, using the expressions \eqref{equation:f}, \eqref{equation:curv-induced} and \eqref{equation:G}, we have:
\begin{equation}
\label{equation:intermediaire}
\begin{split}
4 \sum_\alpha G_{S^2 TM}&(v, \nabla_{\V} f_\alpha, f, \e_\alpha) = \sum_\alpha \big( \langle [(v \wedge \nabla_{\V} f_\alpha), f], \e_\alpha \rangle \\
& + \langle [(J v \wedge J \nabla_{\V} f_\alpha), f], \e_\alpha \rangle - 2 \langle{v, J \nabla_{\V} f_\alpha}\rangle.\langle{[J,f], \e_\alpha}\rangle\big),
\end{split}
\end{equation}
where $(\e_\alpha)_{\alpha}$ is a local orthonormal basis of $S^2TM$.
The integral over $S_xM$ of the first term can be computed using \cite[Lemma 4.6]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} (case of symmetric $2$-tensors) since $\iota_v f$ is of degree $k-1$ and yields:
\begin{equation}\label{integral}
\int_{S_xM} \sum_\alpha\langle [(v \wedge \nabla_{\V} f_\alpha), f], \e_\alpha \rangle = 2(n+2k-4)\|\iota_v f\|^2_{L^2(S_xM)} + 2\|f\|^2_{L^2(S_xM)}.
\end{equation}
The third term vanishes in \eqref{equation:intermediaire} since $[J,f]=0$ by assumption.
We now claim that the integral over $S_xM$ of the second term in \eqref{equation:intermediaire} is also equal to $2(n+2k-4)\|\iota_v f\|^2_{L^2(S_xM)} + 2\|f\|^2_{L^2(S_xM)}$, which will finish the proof.
Indeed, observe that $JX \wedge JY = -J \circ (X \wedge Y) \circ J$ for any $X, Y \in TM$, and therefore using also $[J, f] = 0$, we get for each $\alpha$
\[[Jv \wedge J \nabla_{\V} f_\alpha, f] = -J \circ [v \wedge \nabla_{\V} f_\alpha, f] \circ J.\]
Consequently, we can rewrite
\[\langle{[Jv \wedge J \nabla_{\V} f_\alpha, f], \e_\alpha}\rangle =\langle -J \circ [v \wedge \nabla_{\V} f_\alpha, f] \circ J, \e_\alpha\rangle= \langle{[v \wedge \nabla_{\V} f_\alpha, f],- J \circ \e_\alpha \circ J}\rangle.\]
Note that $(-J \circ \e_\alpha \circ J)_\alpha$ is also an orthonormal basis of $S^2TM$, and since $f=-J \circ f \circ J =- \sum_\alpha f_\alpha J \circ \e_\alpha \circ J$, we obtain
\[f_\alpha=\langle f,\e_\alpha\rangle=\langle f,-J \circ \e_\alpha \circ J\rangle.\]
The claim thus follows from \eqref{integral}.
\end{proof}
\subsection{Bounding from below the left-hand side}
Going back to the case $E = S^2TM$, we now bound from below the term $\|\mathbf{X}_-u\|^2_{L^2}$ appearing on the left-hand side of the twisted Pestov identity \eqref{equation:pestov}:
\begin{lemma}
\label{lemma:x-}
Let $u \in C^\infty(M,\Omega_k \otimes S^2 TM)$ be a complex normal twisted conformal Killing tensor in the sense of Definition \ref{definition:cntckt}. Then, for $k > 0$ the following inequality holds:
\[
\begin{split}
\tfrac{(k-1)(n+2k-2)}{k} \|\mathbf{X}_-u\|^2_{L^2}
\geq& \big(\tfrac{3\lambda-2}{2}\alpha_{n,k-1}-\tfrac{8(1-\lambda)}{3}\beta_{n,k-1}- \tfrac{29(1+\lambda)}{48} - \tfrac{1+\lambda}{4} \delta_{n,k-1}\big) \|\iota_v u\|^2_{L^2}.
\end{split}
\]
\end{lemma}
The proof of Lemma \ref{lemma:x-} requires an additional step and is postponed to the end of this paragraph.
\begin{lemma}
\label{lemma:x+}
Let $f \in C^\infty(M, \Omega_k \otimes TM)$ such that $\iota_v f, \iota_{Jv} f$ are of degree $k-1$, and assume $\lambda \in [\tfrac{2}{3}, 1]$. Then, the following holds:
\[
\begin{split}
\tfrac{k(n+2k)}{k+1}\|\mathbf{X}_+f\|^2_{L^2} \geq &\left( \tfrac{3\lambda-2}{4} \alpha_{n,k} - \tfrac{4(1-\lambda)}{3} \beta_{n,k} - \tfrac{29(1+\lambda) }{96}\right) \|f\|^2_{L^2} \\
& - \tfrac{1+\lambda}{8}\delta_{n,k}(\|\iota_v f\|^2_{L^2} + \|\iota_{Jv}f\|^2_{L^2}).
\end{split}
\]
\end{lemma}
\begin{proof}
Using the twisted Pestov identity \eqref{equation:pestov} with $E =TM$, and applying the bounds provided by Lemmas \ref{lemma:bound-r}, \ref{lemma:f0} and \ref{lemma:g-expression}, we obtain:
\[
\begin{split}
\tfrac{k(n+2k)}{k+1}\|\mathbf{X}_+f\|^2_{L^2}
\geq& - \langle R \nabla_{\mathbb{V}}^{TM} f, \nabla_{\mathbb{V}}^{TM}f \rangle_{L^2} - \langle \mathcal{F}^{TM}f,\nabla_{\mathbb{V}}^{TM} f\rangle_{L^2} \\
\geq&\, \tfrac{3\lambda-2}{4} \alpha_{n,k}\|f\|^2_{L^2} + \tfrac{3\lambda}{4} \int_M\int_{S_xM} \sum_{\alpha} \langle v,J\nabla_{\mathbb{V}}f_\alpha\rangle^2 -\tfrac{4(1-\lambda)}{3}\beta_{n,k}\|f\|^2_{L^2} \\
& - \tfrac{1+\lambda}{2}\bigg( \dfrac{1}{4}\delta_{n,k}(\|\iota_v f\|^2_{L^2} + \|\iota_{Jv}f\|^2_{L^2}) + \tfrac{1}{2}\|f\|^2_{L^2}\\
&\hskip30pt+ \tfrac{1}{2}\int_M \int_{S_xM} \sum_{\alpha} |\langle v,J\nabla_{\mathbb{V}}f_\alpha\rangle| |\langle f,J\e_\alpha \rangle| \bigg).
\end{split}
\]
We now use the estimate
\[
\tfrac{1}{2} \int_M \int_{S_xM} \sum_{\alpha} \langle v,J\nabla_{\mathbb{V}}f_\alpha\rangle \langle f,J\e_\alpha \rangle \leq \tfrac{1}{4\varepsilon} \|f\|^2_{L^2} + \tfrac{\varepsilon}{4}\int_M \int_{S_xM} \sum_\alpha \langle v,J\nabla_{\mathbb{V}}f_\alpha\rangle^2,
\]
which holds for all $\varepsilon > 0$, to deduce that
\[
\begin{split}
\tfrac{k(n+2k)}{k+1}\|\mathbf{X}_+f\|^2_{L^2}
\geq&\, \tfrac{3\lambda-2}{4} \alpha_{n,k}\|f\|^2_{L^2} + \tfrac{3\lambda}{4}\int_M\int_{S_xM} \sum_{\alpha}\langle v,J\nabla_{\mathbb{V}}f_\alpha\rangle^2 -\tfrac{4(1-\lambda)}{3}\beta_{n,k}\|f\|^2_{L^2} \\
& - \tfrac{1+\lambda}{2}\bigg( \tfrac{1}{4}\delta_{n,k}(\|\iota_v f\|^2_{L^2} + \|\iota_{Jv}f\|^2_{L^2}) + \left(\tfrac{1}{2} +\tfrac{1}{4\varepsilon}\right)\|f\|^2_{L^2}\\
&\hskip30pt+ \tfrac{\varepsilon}{4} \int_M \int_{S_xM}\sum_{\alpha} \langle v,J\nabla_{\mathbb{V}}f_\alpha\rangle^2 \bigg).
\end{split}
\]
Taking the specific value $\varepsilon := 6\lambda/(1+\lambda)$, we see that the coefficients in front of the term $\int_M \sum_{\alpha} \int_{S_xM} \langle v,J\nabla_{\mathbb{V}}f_\alpha\rangle^2$ cancel out. Moreover, since by assumption $\lambda \in [2/3,1]$, we have the lower bound $\varepsilon =\tfrac{6\lambda}{1 + \lambda} \geq \tfrac{12}5$ and this eventually yields the result.
\end{proof}
\begin{remark}
In the last paragraph of the proof of Lemma \ref{lemma:x+}, one could decide to keep using the exact value $\varepsilon = \tfrac{6\lambda}{1 + \lambda}$ in the estimate. However, the benefit of doing so would be minor in the final result so for simplicity we decided to use the trivial lower bound $\varepsilon\geq \tfrac{12}5$.
\end{remark}
We can now prove Lemma \ref{lemma:x-}.
\begin{proof}[Proof of Lemma \ref{lemma:x-}]
Using the equality $\mathbf{X}(\iota_v u) = \iota_v \mathbf{X} u = \iota_v \mathbf{X}_- u$, and the fact that $\iota_{Jv}\mathbf{X}_-u=J(\iota_{v}\mathbf{X}_-u)$, we obtain:
\[
\begin{split}
\|\mathbf{X}_-u\|^2_{L^2} & \geq \|\iota_v \mathbf{X}_- u\|^2_{L^2} + \|\iota_{Jv}\mathbf{X}_-u\|^2_{L^2} =2\|\iota_v \mathbf{X}_- u\|^2_{L^2} \\
& =2 \|\mathbf{X}(\iota_vu)\|^2_{L^2} =2 \|\mathbf{X}_+(\iota_v u)\|^2_{L^2} +2 \|\mathbf{X}_-(\iota_v u)\|^2_{L^2} \\
& \geq 2\|\mathbf{X}_+(\iota_v u)\|^2_{L^2} ,
\end{split}
\]
where in the second line we used that $\iota_v u$ and $\iota_{Jv} u$ are of degree $k - 1$, and the mapping property \eqref{equation:split}. By assumption, $u$ is a complex normal twisted conformal Killing tensor so this implies that $\iota_v \iota_{Jv} u = \iota_{Jv} \iota_v u = 0$ and $\iota_v \iota_v u= \iota_{Jv}\iota_{Jv}u$ is of degree $k-2$. As a consequence, we can apply Lemma \ref{lemma:x+} with $f:=\iota_v u$ (which is of degree $k-1$). Using the fact that $\iota_{Jv} f=0$, together with the fact that
\[\|\iota_v f\|^2_{L^2}=\|\iota_v \iota_v u\|^2_{L^2}\leq\|\iota_v u\|^2_{L^2}, \]
we obtain the announced result.
\end{proof}
\section{Pinching estimates} \label{section:threshold}
\subsection{Computations} We start with the following:
\begin{lemma}\label{lemma:sign}
If $u \in C^\infty(M,\Omega_k \otimes S^2 TM)$ is a complex normal twisted conformal Killing tensor, the following inequality holds:
\begin{equation}
\label{equation:sign}
B_{n,k}(\lambda)\|u\|^2_{L^2} + {C}_{n,k}(\lambda)\|\iota_vu\|^2_{L^2} \leq 0,
\end{equation}
where
\begin{equation}
\label{equation:B}
B_{n,k}(\lambda) := \tfrac{3\lambda-2}{4} \alpha_{n,k} - \tfrac{8}{3}(1-\lambda)\beta_{n,k}- \tfrac{1+\lambda}{2},
\end{equation}
and
\begin{equation}
\label{equation:C}
C_{n,k}(\lambda) := \gamma_{n,k}\left( \tfrac{3\lambda-2}{2}\alpha_{n,k-1}-\tfrac{8(1-\lambda)}{3}\beta_{n,k-1}-\tfrac{29(1+\lambda)}{48} - \tfrac{1+\lambda}{4}\delta_{n,k-1}\right) - \tfrac{1+\lambda}{2}\delta_{n,k}.
\end{equation}
\end{lemma}
\begin{proof} Straightforward computation, inserting in the twisted Pestov identity \eqref{equation:pestov} the lower bound for $\|\mathbf{X}_-u\|^2$ (Lemma \ref{lemma:x-}), and the upper bounds for the terms on the right-hand side (Lemmas \ref{lemma:bound-r} and \ref{lemma:f0} applied to $f=u$, and Lemma \ref{lemma:g-expression} (ii)).
\end{proof}
Recall that $n=2m$ is the real dimension of $M$. The end of the proof of Theorem \ref{theorem:main} then consists in finding a pinching condition $\lambda > \lambda(m)$ for which the left-hand side of \eqref{equation:sign} is nonnegative, thus forcing the complex normal twisted conformal Killing tensor $u$ to be zero, which then contradicts Corollary \ref{corollary:invariant-sections}. More precisely, we have the following:
\begin{lemma}\label{lemma:BC}
If the inequalities
\begin{equation}
\label{equation:conditions}
B_{n,k}(\lambda) > 0 \quad \text{and} \quad B_{n,k}(\lambda) + \frac12{C}_{n,k}(\lambda) > 0
\end{equation}
hold, then \eqref{equation:sign} implies $u \equiv 0$.
\end{lemma}
\begin{proof}
If ${C}_{n,k}(\lambda) \geq 0$, this is immediate from the first part of \eqref{equation:conditions}. If ${C}_{n,k}(\lambda) <0$, using that $ \|u\|^2_{L^2}\geq \|\iota_vu\|^2_{L^2}+\|\iota_{Jv}u\|^2_{L^2}=2\|\iota_vu\|^2_{L^2}$, we get by \eqref{equation:sign}:
\[
(B_{n,k}(\lambda) + \frac12{C}_{n,k}(\lambda))\|u\|^2_{L^2} \leq 0,
\]
so $u \equiv 0$ by the second part of \eqref{equation:conditions}.
\end{proof}
Moreover, the following holds:
\begin{lemma}\label{lemma:l12}
One has: $B_{n,k} > 0 \iff \lambda > \lambda_1(n,k)$ where
\[
\lambda_1(n,k) := \dfrac{6\alpha_{n,k}+32\beta_{n,k}+6}{9\alpha_{n,k}+32\beta_{n,k}-6},
\]
and $B_{n,k} +\frac12 C_{n,k} > 0 \iff \lambda > \lambda_2(n,k)$ where
\small
\[
\lambda_2(n,k) := \dfrac{6\alpha_{n,k}+32\beta_{n,k}+6+\gamma_{n,k}\left(6\alpha_{n,k-1}+16\beta_{n,k-1}+\tfrac{29}{8}+\tfrac{3}{2}\delta_{n,k-1}\right)+3\delta_{n,k}}{9\alpha_{n,k}+32\beta_{n,k} - 6 +\gamma_{n,k}\left(9\alpha_{n,k-1}+16\beta_{n,k-1}-\tfrac{29}{8}-\tfrac{3}{2}\delta_{n,k-1}\right)-3\delta_{n,k}}.
\]
\normalsize
\end{lemma}
\begin{proof}
Follows immediately from \eqref{equation:B} and \eqref{equation:C}, as the expressions of $B_{n,k}$ and $C_{n,k}$ are affine functions in $\lambda$.
\end{proof}
Before proving Theorem \ref{theorem:main}, we need to study the variations of the sequences $k \mapsto \lambda_{1}(n,k)$ and $k \mapsto \lambda_{2}(n,k)$.
\begin{lemma}
\label{lemma:variation}
For $n \geq 4$, the sequences $k \mapsto \lambda_{1}(n,k)$ and $k \mapsto \lambda_{2}(n,k)$ are decreasing for $k \geq 2$.
\end{lemma}
\begin{proof}
It is straightforward to check that both sequences are positive for $k\ge 2$ and $n\ge 4$. Using that $\alpha_{n,k}=\tfrac{\beta_{n,k}^2}{n-1}$,
we can write
$$\tfrac{1}{\lambda_1(n,k)}-1= \tfrac{3\alpha_{n,k}-12}{6\alpha_{n,k}+32\beta_{n,k}+6}=\tfrac{3-\tfrac{12}{\alpha_{n,k}}}{6+\tfrac{32(n-1)}{\beta_{n,k}}+\tfrac{6}{\alpha_{n,k}}}.$$
Since $\alpha_{n,k}$ and $\beta_{n,k}$ are positive and increasing in $k$, the numerator in the right hand term is increasing in $k$, whereas the denominator is positive and decreasing in $k$. Thus $k \mapsto \lambda_{1}(n,k)$ is decreasing.
Consider now the expression $\lambda_{2}(n,k)$. Again it is easy to check that $\lambda_{2}(n,k)>0$ for $k\ge 2$ and $n\ge 4$, and
\small
\begin{eqnarray*}\tfrac{1}{\lambda_2(n,k)}-1&=&\dfrac{3\alpha_{n,k}-12+\gamma_{n,k}\left(3\alpha_{n,k-1}-\tfrac{29}{4}-3\delta_{n,k-1}\right)-6\delta_{n,k}}{6\alpha_{n,k}+32\beta_{n,k}+6+\gamma_{n,k}\left(6\alpha_{n,k-1}+16\beta_{n,k-1}+\tfrac{29}{8}+\tfrac{3}{2}\delta_{n,k-1}\right)+3\delta_{n,k}}\\
&=&\dfrac{3-\tfrac{12}{\alpha_{n,k}}+\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\left(3\alpha_{n,k-1}-\tfrac{29}{4}-3\delta_{n,k-1}\right)-\tfrac{6\delta_{n,k}}{\alpha_{n,k}}}{6+\tfrac{32\beta_{n,k}}{\alpha_{n,k}}+\tfrac{6}{\alpha_{n,k}}+\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\left(6\alpha_{n,k-1}+16\beta_{n,k-1}+\tfrac{29}{8}+\tfrac{3}{2}\delta_{n,k-1}\right)+\tfrac{3\delta_{n,k}}{\alpha_{n,k}}}.
\end{eqnarray*}
\normalsize
Let us denote this last expression by $\tfrac{E_{n,k}}{F_{n,k}}$. We claim that $E_{n,k}$ is increasing in $k$ and $F_{n,k}$ is decreasing in $k$.
Using the fact that $\frac{\gamma_{n,k}}{\alpha_{n,k}}=\frac{n+2k-4}{(n+2k-2)\alpha_{n,k-1}}$ we get
\begin{eqnarray*}E_{n,k}&=&3+3\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\alpha_{n,k-1}-\tfrac{12+6\delta_{n,k}}{\alpha_{n,k}}-\tfrac{29}{4}\tfrac{\gamma_{n,k}}{\alpha_{n,k}}-3\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\delta_{n,k-1}\\
&=&6\tfrac{n+2k-3}{n+2k-2}-6\tfrac{n+2k-2}{\alpha_{n,k}}-\tfrac{29}{4}\tfrac{\gamma_{n,k}}{\alpha_{n,k}}-3\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\delta_{n,k-1}\\
&=&6-\tfrac6{n+2k-2}-\tfrac6k-\tfrac6{n+k-2}-3\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\left((n+2k-4)+\tfrac5{12}\right).
\end{eqnarray*}
In order to express $F_{n,k}$, we remark that
$$\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\beta_{n,k-1}=\tfrac{n+2k-4}{n+2k-2}\tfrac{\beta_{n,k-1}}{\alpha_{n,k-1}}=\tfrac{s_{n,k-1}}{n+2k-2},$$
where we denote by $s_{n,k}:=\frac{(n+2k-2)\beta_{n,k}}{\alpha_{n,k}}.$ A straightforward computation similar to the one for $E_{n,k}$ shows that
\begin{eqnarray*}F_{n,k}&=&6+3\tfrac{n+2k-2}{\alpha_{n,k}}+6\tfrac{n+2k-4}{n+2k-2}+\tfrac{3}{2}\tfrac{\gamma_{n,k}}{\alpha_{n,k}}\left((n+2k-4)+\tfrac5{12}\right)\\
&&+32\tfrac{s_{n,k}}{n+2k-2}+16\tfrac{s_{n,k-1}}{n+2k-2}.\end{eqnarray*}
In order to prove our claim, it is thus enough to remark that:
\begin{eqnarray*}(n+2k-4)\tfrac{\gamma_{n,k}}{\alpha_{n,k}}&=&\tfrac{(n+2k-4)^2}{(k-1)(n+k-3)(n+2k-2)}\\
&=&\tfrac{n-4}{n-2}\cdot\tfrac1{k-1}+\tfrac{n}{n-2}\cdot\tfrac1{n+k-3}+\tfrac{4}{(k-1)(n+k-3)(n+2k-2)}\end{eqnarray*}
is decreasing in $k$ (and thus $\frac{\gamma_{n,k}}{\alpha_{n,k}}$ is decreasing in $k$ too);
$$s_{n,k}=(n+2k-2)\tfrac{\sqrt{(n-1)k(n+k-2)}}{k(n+k-2)}=\sqrt{(n-1)\left(4+\tfrac{(n-2)^2}{k(n+k-2)}\right)}$$ is decreasing in $k$, and
\begin{eqnarray*}\tfrac{n+2k-2}{\alpha_{n,k}}+2\tfrac{n+2k-4}{n+2k-2}=2+\tfrac{n+2k-2}{k(n+k-2)}-\tfrac4{n+2k-2}=2+\tfrac{(n-2)^2}{k(n+k-2)(n+2k-2)}\end{eqnarray*} is decreasing in $k$.
\end{proof}
\subsection{Proof of ergodicity} We can now conclude the proof of the ergodicity statement in Theorem \ref{theorem:main}.
\begin{proof}[Proof of ergodicity in Theorem \ref{theorem:main}]
We need to show that under the pinching condition $\lambda > \lambda(m)$ of Theorem \ref{theorem:main}, the unitary frame flow is ergodic. If the frame flow is not ergodic, and $m \neq 4,28$, we know by Corollary \ref{corollary:invariant-sections} that there exists a flow-invariant orthogonal projector $\pi_{\mathcal{V}} \in C^\infty(SM,S^2 \mathcal{N})$ of even Fourier degree onto a complex vector bundle $\mathcal{V} \subset \mathcal{N}$ of rank $1 \leq r \leq m/2-1$.
By Corollary \ref{corollary:tensor} and Lemma \ref{lemma:sign}, this yields the existence of a non-zero complex normal twisted conformal Killing tensor $u \in C^\infty(M,\Omega_k \otimes S^2 T^*M)$ of even degree $k \geq 2$ which satisfies the inequality \eqref{equation:sign}. We distinguish two cases.
{\em Case 1.} If $k \geq 4$, Lemmas \ref{lemma:l12} and \ref{lemma:variation} show that if the holomorphic pinching $\lambda$ satisfies $\lambda > \max(\lambda_1(n,4),\lambda_2(n,4))$, we then have $B_{n,k} > 0$ and $B_{n,k}+\frac12C_{n,k} > 0$ which by Lemma \ref{lemma:BC} implies that $u \equiv 0$.
{\em Case 2.} If $k=2$, let $\pi_{\mathcal{V}}$ be the projector given by Corollary \ref{corollary:invariant-sections}, whose complex rank $r$ satisfies $1\le r\le m/2-1$. By \cite[Lemma 4.2]{Cekic-Lefeuvre-Moroianu-Semmelmann-21}, one can then show that $\pi_{\mathcal{V}}$ must be of the form
\[
\pi_{\mathcal{V}} = \tfrac{2r}{n} \mathbbm{1}_{TM} + u,
\]
where $u \in C^\infty(M,\Omega_2 \otimes S^2 T^*M)$. In particular, since $\mathbbm{1}_{TM}$ is parallel, this implies $\mathbf{X} u = 0$, that is, $\mathbf{X}_\pm u = 0$. Moreover, a quick algebraic computation (see \cite[p. 38]{Cekic-Lefeuvre-Moroianu-Semmelmann-21}) gives
\begin{equation}
\label{equation:relation-norm}
\|\iota_v u\|^2_{L^2} = \tfrac{2r}{n(n-2r)}\|u\|^2_{L^2},
\end{equation}
and since $2r\le m-2=n/2-2$, we obtain
\begin{equation}
\label{equation:relation-norm1}
\|\iota_v u\|^2_{L^2} \le\tfrac{n-2}{n(n+2)}\|u\|^2_{L^2}.
\end{equation}
Applying the twisted Pestov identity \eqref{equation:pestov} to $u$ for $k=2$, using that $\mathbf{X}_\pm u = 0$ (the bound of Lemma \ref{lemma:x-} is therefore useless), and the upper bounds of Lemmas \ref{lemma:bound-r}, \ref{lemma:f0} and \ref{lemma:g-expression}, we find that:
\[
\left(\tfrac{3\lambda-2}{2}n-\tfrac{8(1-\lambda)}{3}(2n(n-1))^{1/2}-\tfrac{1+\lambda}{2}\right)\|u\|^2_{L^2} - \tfrac{1+\lambda}{2} n \|\iota_vu\|^2_{L^2} \leq 0.
\]
Inserting \eqref{equation:relation-norm} in the previous inequality, we then obtain:
\[
\left(\tfrac{3\lambda-2}{2}n-\tfrac{8}{3}(1-\lambda)(2n(n-1))^{1/2}-\tfrac{1+\lambda}{2} \tfrac{n-2}{n+2} \right)\|u\|^2_{L^2} \leq 0.
\]
This shows that $u \equiv 0$ when $k=2$, as soon as the holomorphic pinching $\lambda$ satisfies
\[
\lambda \geq \lambda_3(n) := \dfrac{6n+ 16(2n(n-1))^{1/2} + \tfrac{6n}{n+2}}{9n+16(2n(n-1))^{1/2}-\tfrac{6n}{n+2}}.
\]
Summarizing the two cases $k\ge 4$ and $k=2$, we have obtained that the the unitary frame flow is ergodic as soon as the holomorphic pinching $\lambda$ satisfies
\[
\lambda> \max(\lambda_1(2m,4),\lambda_2(2m,4),\lambda_3(2m))=:\lambda_0(m).
\]
Using a formal computing tool, it can be easily checked that $ \lambda_0(m)=\lambda_2(2m)$ for $m \geq 6$. However, for practical reasons we will give a bound which is slightly less accurate, but much easier to compute.
Namely, using the obvious inequality $((n+2)(n-1))^{1/2}<n+1/2$, we can write
\begin{eqnarray*}\dfrac{1}{\lambda_1(n,4)}-1&=& \dfrac{3\alpha_{n,4}-12}{6\alpha_{n,4}+32\beta_{n,4}+6}= \dfrac{12(n+2)-12}{24(n+2)+32(4(n+2)(n-1))^{1/2}+6}\\
&>&\dfrac{12n+12}{88n+86}=\dfrac{6n+6}{44n+43}.
\end{eqnarray*}
Similarly, since $16\sqrt2<68/3$, we have
\begin{eqnarray*}\dfrac{1}{\lambda_3(n)}-1&=& \dfrac{3n-\tfrac{12n}{n+2}}{6n+ 16(2n(n-1))^{1/2} + \tfrac{6n}{n+2}}= \dfrac{3-\tfrac{12}{n+2}}{6+ 16(\tfrac{2(n-1)}{n})^{1/2} + \tfrac{6}{n+2}}\\
&>& \dfrac{3-\tfrac{12}{n+2}}{6+ \tfrac{68}3 + \tfrac{6}{n+2}}=\dfrac{9n-18}{86 n+190}.
\end{eqnarray*}
Finally, using the calculation in the proof of Lemma \ref{lemma:variation} together with the obvious inequalities $64/\sqrt3<37$ and
$$\dfrac43>\gamma_{n,4}= \dfrac{4(n+2)(n+4)}{3(n+1)(n+6)}> \dfrac{4n}{3(n+1)},\qquad\forall n\ge3,$$
we obtain
\small
\begin{eqnarray*}&&\hskip-20pt\tfrac{1}{\lambda_2(n,4)}-1=\dfrac{3\alpha_{n,4}-12+\gamma_{n,4}\left(3\alpha_{n,3}-\tfrac{29}{4}-3\delta_{n,3}\right)-6\delta_{n,4}}{6\alpha_{n,4}+32\beta_{n,4}+6+\gamma_{n,4}\left(6\alpha_{n,3}+16\beta_{n,3}+\tfrac{29}{8}+\tfrac{3}{2}\delta_{n,3}\right)+3\delta_{n,4}}\\
&>&\dfrac{12(n+2)-12+\tfrac{4n}{3(n+1)}\left(9(n+1)-\tfrac{29}{4}-3(n+2)\right)-6(n+4)}{24(n+2)+64(n+\tfrac12)+6+\tfrac43\left(18(n+1)+16\sqrt3n+\tfrac{29}{8}+\tfrac{3}{2}(n+2)\right)+3(n+4)}\\
&>&\dfrac{14n-\tfrac{77}3}{(117+\tfrac{64}{\sqrt3})n+126+\tfrac{29}{6}}>\dfrac{14n-26}{154n+131}.
\end{eqnarray*}
\normalsize
It is straightforward to check that
$$\dfrac{6n+6}{44n+43}>\dfrac{9n-18}{86 n+190}>\dfrac{14n-26}{154n+131},\qquad\forall n\ge 10,$$
which implies $\tfrac1{\lambda_0(m)}-1>\dfrac{14n-26}{154n+131}$, so eventually
$$\lambda_0(m)<\dfrac{154n+131}{168n+105}=\dfrac{308m+131}{336m+105}=:\lambda(m),$$
thus proving the ergodicity statement of Theorem \ref{theorem:main}.
\end{proof}
We conclude this paragraph by a remark on the remaining cases $m=4,28$:
\begin{remark} In the case $m=28$, one can check\footnote{Using the LiE program for instance.} that $\mathrm{E}_6$ fixes an element of $S^3 \mathbb{C}^{27}$. As we saw in the proof of Corollary \ref{corollary:invariant-sections}, the transitivity group is \emph{never} semi-simple as it always contains the scalar matrices: Theorem \ref{theorem:topology} therefore does not exclude that subgroups of $\mathrm{E}_6 \times \mathrm{U}(1) \lneqq \mathrm{U}(27)$ occur as transitivity groups in complex dimension $m=28$. In this case, the transitivity group would fix an orthogonal projector of $\mathrm{End}(S^3 \mathbb{C}^{27})$. Nevertheless, the argument given below involving the twisted Pestov identity does not carry over to $\mathrm{End}(S^3 \mathbb{C}^{27})$ because this vector space involves tensorial powers of too high degree. In other words, the argument would result in a pinching condition $\lambda > \lambda(28)$ for some $\lambda(28) > 1$ so the statement would be empty. A similar remark holds for the case $m=4$. This should be compared with \cite[Theorem 3.8]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} where we only deal with elements of $\mathfrak{o}(\mathcal{N})$ with $\mathfrak{o} \in \left\{\mathrm{id},S^2,\Lambda^2,\Lambda^3\right\}$. The worse pinching estimate in \cite[Theorem 1.2]{Cekic-Lefeuvre-Moroianu-Semmelmann-21} comes from the exterior power $\Lambda^3 \mathcal{N}$.
\end{remark}
\subsection{Proof of mixing} \label{section:5.3}
It now remains to show the mixing property. We will actually show the following:
\begin{proposition}
\label{proposition:mixing}
The unitary frame flow on a negatively-curved Kähler manifold $(M,g)$ is ergodic if and only if it is mixing.
\end{proposition}
\begin{proof}
Mixing implies ergodicity so it remains to show that ergodicity implies mixing in this setting. By \cite[Proof of Lemma 3.7]{Lefeuvre-21}, it suffices to show that the equation
\begin{equation}
\label{equation:mixing}
X_{F_{\mathbb{C}}M} u = i \lambda u, \qquad \lambda \in \mathbb{R} \setminus \left\{0\right\},\ u \in L^2(F_{\mathbb{C}}M),
\end{equation}
implies that $u \equiv 0$. Let $u \in L^2(F_{\mathbb{C}}M)$ be a solution to \eqref{equation:mixing}. Since $F_{\mathbb{C}}M \to SM$ is a principal $\mathrm{U}(m-1)$-bundle over $SM$, the space $L^2(F_{\mathbb{C}}M)$ splits as
\begin{equation}
\label{equation:splitting}
L^2(F_{\mathbb{C}}M) = \bigoplus_{k=0}^{+\infty} L^2(SM, \Theta_k),
\end{equation}
where $\Theta_k$ is the vector bundle over $SM$ whose fibre over $v\in SM$ is the eigenspace of the Casimir operator of $\mathrm{U}(m-1)$ acting on functions of the fibre $(F_{\mathbb{C}}M)_v$, associated to the eigenvalue $\mu_k \geq 0$ (and $\mu_k \neq \mu_j$ for $k \neq j$). Since $X_{F_{\mathbb{C}}M}$ preserves the splitting \eqref{equation:splitting} (because the frame flow $(\Phi_t)_{t \in \mathbb{R}}$ commutes with the right-action of $\mathrm{U}(m-1)$), we deduce that for each $k\ge 0$, the component $u_k\in L^2(SM,\Theta_k)$ of some function $u = \sum_{k} u_k \in \oplus_{k=0}^{+\infty}L^2(SM,\Theta_k)$ satisfying \eqref{equation:mixing} must satisfy
\[
X_{F_{\mathbb{C}}M} u_k = i \lambda u_k.
\]
Since $\Theta_0 = \mathbb{C}$ is trivial over $SM$, the equation $X_{F_{\mathbb{C}}M} u_0 = i\lambda u_0$ reads $X u_0 = i \lambda u_0$, where $u_0 \in L^2(SM)$ and $X$ is the geodesic vector field. But the geodesic flow is mixing as $(M,g)$ has negative curvature, so we deduce that $u_0 = 0$.
For $k \neq 0$, \cite[Proposition 2.6]{Lefeuvre-21} based on microlocal analysis shows that $u_k \in L^2(SM,\Theta_k)$ is actually smooth, that is, $u_k \in C^\infty(SM,\Theta_k)$. Moreover, $X_{F_{\mathbb{C}}M}|u_k|^2 = 0$, so using the ergodicity assumption on the frame flow $(\Phi_t)_{t \in \mathbb{R}}$, we get that $|u_k|$ is constant on $F_{\mathbb{C}}M$. If $u_k \neq 0$, then up to rescaling, we obtain that $u_k : F_{\mathbb{C}} M \to \mathrm{U}(1)$ is a well-defined smooth map. In order to simplify the notation, we will drop the index $k$ from now on and simply write $u$ instead of $u_k$, and $\mu$ instead of $\mu_k$.
Denote by $P$ the unit circle bundle of the complex line bundle $\Lambda^{m-1,0}\mathcal{N} \to SM$. Now, observe that there is a natural surjective bundle map $\psi : F_{\mathbb{C}}M \to P$ given by
\[
\psi : (v, \e_2, \ldots,\e_m) \mapsto (v, (\e_2-i J\e_2) \wedge \ldots\wedge (\e_m-iJ\e_m)).
\]
There is a natural unitary parallel transport along geodesic flow-lines of sections of $\Lambda^{m-1,0}\mathcal{N}$, so there is a flow $(\Phi_t^P)_{t \in \mathbb{R}}$ on $P$ with generator $X_P$ extending the geodesic flow $(\varphi_t)_{t\in\mathbb{R}}$ as in \S\ref{ssection:isometry}, that is, writing $\pi : P \to SM$ for the projection map, one has $\pi \circ \Phi_t^P = \varphi_t \circ \pi$ for all $t \in \mathbb{R}$. Moreover, $\psi$ intertwines the frame flow on $F_{\mathbb{C}}M$ and the flow on $P$, that is
\begin{equation}
\label{equation:relation}
\Phi_t^P \circ \psi = \psi \circ \Phi_t.
\end{equation}
We claim that the following holds:
\begin{claim}
There exists a smooth function $w \in C^\infty(P)$ such that $u = \psi^* w$, and $X_P w = i \lambda w$.
\end{claim}
Note that, once we know that $u = \psi^*w$ for some function $w$, the relation $X_P w = i \lambda w$ is immediate using \eqref{equation:relation} and $X_{F_{\mathbb{C}}M} u = i \lambda u$.
\begin{proof}
We show that $u = \psi^*w$ for some $w \in C^\infty(P)$. We fix an arbitrary point $v_0 \in SM$ and $w_0 \in (F_{\mathbb{C}}M)_{v_0}$. There is then a commutative diagram
\begin{equation}
\label{equation:diagram}
\xymatrix{
(F_{\mathbb{C}}M)_{v_0} \ar[r]^\psi \ar[d] & P_{v_0} \ar[d] \\
\mathrm{U}(m-1) \ar[r]^{\mathrm{det}} & \mathrm{U}(1)
}
\end{equation}
where the downward arrows are isometries. Hence, by restricting $u$ to the fiber $(F_{\mathbb{C}}M)_{v_0}$ and identifying isometrically $(F_{\mathbb{C}}M)_{v_0} \simeq \mathrm{U}(m-1)$, we get a map $F := u(v_0)$ such that
\begin{equation}
\label{equation:submersion}
F : F_{\mathbb{C}}M \simeq \mathrm{U}(m-1) \to \mathrm{U}(1),
\end{equation}
and $\Delta_{\mathrm{U}(m-1)}F = \mu F$ with $\mu \neq 0$ since $u$ takes values in $\Theta_k$ for $k \neq 0$. In other words, $F$ is an eigenfunction of the Casimir operator on $\mathrm{U}(m-1)$ associated to the eigenvalue $\mu \neq 0$ and of constant modulus.
Now, recall that $\mathrm{U}(m-1)$ is a split group extension of the circle group $\mathrm{U}(1)$ by the special unitary group $\mathrm{SU}(m-1)$, that is, $\mathrm{det} : \mathrm{U}(m-1) \to \mathrm{U}(1)$ is a fiber bundle with fibers isometric to $\mathrm{SU}(m-1)$. We claim that such a function $F$ must then necessarily be constant on the $\mathrm{SU}(m-1)$-fibers of the bundle $\mathrm{U}(m-1)$, that is, $F = \mathrm{det}^* f$ for some $f \in C^\infty(\mathrm{U}(1))$, $\mu = j^2$ for some integer $j \geq 0$ and $f$ is an eigenfunction of $\Delta_{\mathrm{U}(1)}$ with eigenvalue $j^2$.
Indeed, since $\mathrm{U}(m-1) = \mathrm{SU}(m-1) \times_{\mathbb{Z}_m} \mathrm{U}(1)$, there is a $\mathbb{Z}_m$-bundle map\footnote{The $\mathbb{Z}_m$-action on $\mathrm{SU}(m-1) \times \mathrm{U}(1)$ is simply given by $(w,z) \mapsto (w \omega^{-k}, \omega^k z)$ for $(w,z) \in \mathrm{SU}(m-1) \times \mathrm{U}(1)$, $k \in \left\{0,...,m-2\right\}$ and $\omega := e^{2i\pi/{(m-1)}}$.} $p : \mathrm{SU}(m-1) \times \mathrm{U}(1) \mapsto \mathrm{U}(m-1)$ and this map is locally a Riemannian isometry. As a consequence, we get that $\Delta_{\mathrm{SU}(m-1) \times \mathrm{U}(1)} p^*F = \mu~ p^* F$. As $\mathrm{SU}(m-1) \times \mathrm{U}(1)$ is a Riemannian product, its eigenfunctions are obtained as sums of products of eigenfunctions on each factor and the eigenvalues are sums of the eigenvalues on each factor. Hence, we can write
\[
p^*F(w,z) = \sum_{j=1}^N a_j(w) b_j(z) = \sum_{j=1}^N a_j(w) z^{k_j},
\]
for some finite number $N > 0$, where $(w,z) \in \mathrm{SU}(m-1) \times \mathrm{U}(1)$, $k_j \in \mathbb{Z}$ and $k_{j_1} \neq k_{j_2}$ for $j_1 \neq j_2$, $a_j \neq 0$ is an eigenfunction of $\Delta_{\mathrm{SU}(m-1)}$ associated to the eigenvalue $\lambda_j$ and $\lambda_j + k^2_j = \mu > 0$. However, we also know that $p^*F$ has constant modulus (equal to $1$). Freezing the point $w = \mathbf{1}_{\mathbb{C}^{m-1}} \in \mathrm{SU}(m-1)$ and moving $z \in \mathrm{U}(1)$, we easily get that $N$ must be equal to $1$, that is, $p^*F(w,z) = a(w) z^{k}$ for some $k\in\mathbb{Z}$. Hence $a \in C^\infty(\mathrm{SU}(m-1))$ satisfies $|a|=1$ and $\Delta_{\mathrm{SU}(m-1)} a = \lambda a$ for some $\lambda \geq 0$ (with $k^2+\lambda=\mu$).
We claim that $\lambda = 0$, that is, $a$ is constant, and $F$ is thus constant along the $\mathrm{SU}(m-1)$-fibers. Indeed, if $\lambda \neq 0$, using
\[
\begin{split}
0&=\Delta_{\mathrm{SU}(m-1)} |a|^2\\
& = (\Delta_{\mathrm{SU}(m-1)} a) \overline{a} + a (\Delta_{\mathrm{SU}(m-1)} \overline{a}) - 2 \nabla a \cdot \nabla \overline{a} \\
& = 2 \lambda - 2 |\nabla a|^2,
\end{split}
\]
we get that $|\nabla a|$ is a non-zero constant. Since $\mathrm{SU}(m-1)$ is simply connected, the map $a$ lifts to the universal cover of $\mathrm{U}(1)$, so there exists a map $\theta:\mathrm{SU}(m-1)\to \mathbb{R}$ such that $a=e^{i\theta}$. At a critical point of $\theta$ we thus get $\nabla a=0$, which is absurd.
As a consequence, $\lambda = 0$ and this shows that $F$ is constant along $\mathrm{SU}(m-1)$-fibers of the bundle map $\mathrm{det} : \mathrm{U}(m-1) \to \mathrm{U}(1)$. In turn, as \eqref{equation:diagram} commutes, we get that $u$ is constant on the preimages of the map $\psi : F_{\mathbb{C}} M \to P$ so $u = \psi^* w$ for some smooth $w \in C^\infty(P)$.
\end{proof}
As a consequence, we have just shown that if there is a non-zero solution to \eqref{equation:mixing}, then there is also a non-zero solution to
\begin{equation}
\label{equation:mixing2}
X_P w = i \lambda w, \qquad \lambda \in \mathbb{R} \setminus \left\{0\right\}, w \in C^\infty(P).
\end{equation}
Hence, it remains to show that \eqref{equation:mixing2} has no non-zero solutions, and we will actually show it for $w \in L^2(P)$.
Assume that $w \in L^2(P)$ is a solution to \eqref{equation:mixing2}. As above, using that $P$ is a principal $\mathrm{U}(1)$-bundle, we can decompose
\begin{equation}
\label{equation:l2split}
L^2(P) = \bigoplus_{k \in \mathbb{Z}} L^2(SM, L_k),
\end{equation}
where $L_k$ is the complex line bundle given by functions $f \in L^2(P)$ satisfying $V f = ik f$, where $V$ is the infinitesimal generator of the $\mathrm{U}(1)$-action on the fibers of $P$. Observe that $L_k = L^{\otimes k}$ for $k \in \mathbb{Z}$ where $L := L_1$ and $L_1 \simeq \Lambda^{m-1,0} \mathcal{N}$. Now, using the splitting \eqref{equation:l2split}, the equation $Xw = i \lambda w$ reads $X w_k = i \lambda w_k$ for all $k \in \mathbb{Z}$, where $w_k \in L^2(SM,L_k)$. By \cite[Proposition 2.6]{Lefeuvre-21}, we also have that $w_k$ is smooth, that is, $w_k \in C^\infty(SM,L_k)$. Moreover, $X|w_k|^2 = 0$, so $|w_k|$ is constant by ergodicity of the flow $(\varphi_t)_{t \in \mathbb{R}}$. If $k=0$, $L_0 = \mathbb{C}$ is trivial over $SM$ so the mixing of the geodesic flow $(\varphi_t)_{t \in \mathbb{R}}$ then implies that $w_0 \equiv 0$. If $k \neq 0$, and $w_k \neq 0$, then we obtain a smooth nowhere vanishing section $w_k \in C^\infty(SM,L_k)$ so $L_k$ is trivial. But in turn, it implies that $L_1 = \Lambda^{m-1,0}\mathcal{N}$ is trivial, which is a contradiction (see the proof of Corollary \ref{corollary:invariant-sections}). Hence, any solution to \eqref{equation:mixing2} is trivial, and thus, so is any solution to \eqref{equation:mixing}. This finishes the proof of Proposition \ref{proposition:mixing}.
\end{proof}
\bibliographystyle{alpha}
|
\section*{Acknowledgments}
This work is supported by the Australian Research Council (ARC) DECRA Project No. DE200100964.
\fi
\bibliographystyle{IEEEtran}
\section{Introduction}
Research into graph-structured data started with the Konigsberg bridge problem in the 18 century, that is:``\textit{How can we design a path among seven bridges in Konigsberg city that crosses each bridge only once}?'' In 1741, Euler modeled seven bridges into a graph and proved that this graph cannot be fully traversed without duplicate paths \cite{euler1741solutio}. In the graph, the nodes represent the junctions between bridges, while the edges represent the bridges themselves.
Since then, graph-structured data has become an indispensable part of modeling and exploring the world.
\IEEEpubidadjcol
Within the vast literature pertaining to graphs, graph-level learning has been drawing considerable attention.
Graph-level learning is the analysis of a dataset consisting of graphs.
For example, the graph isomorphism problem, i.e., the question of whether two pair-wise graphs are isomorphic\footnote{Two graphs $\mathcal G_1$ and $\mathcal G_2$ are isomorphic if the following two conditions are met: (1) There exists matching between nodes in $\mathcal G_1$ and $\mathcal G_2$; (2) Two nodes are linked by an edge in $\mathcal G_1$ \textit{iff} the corresponding nodes are linked by an edge in $\mathcal G_2$.} has attracted an enormous number of studies \cite{harary1969four, weisfeiler1968reduction,mckay1981practical} since the problem was first raised in 1942 \cite{kelly1942isometric}.
Further, before a quasi-polynomial-time solution was proposed in 2016 \cite{babai2016graph}, graph isomorphism was considered a candidate for NP-immediate.
Other applications of great value have also been studied.
As an example, Wang \textit{et al.} \cite{wang2022molecular} took graphs of molecules, where the nodes denote atoms and the edges represent chemical bonds, and performed graph regression as a way of predicting molecular proprieties to help discover more economical crystals.
In another study, a graph generation task based on a series of protein graphs was used to produce graphs of proteins with specific functions to support drug discovery \cite{vamathevan2019applications}.
Likewise, graph classification with brain graphs has the potential to distinguish between brain structures with neurological disorders from those of healthy individuals \cite{lanciano2020explainable}.
Traditional graph-level learning mostly relies on hand-crafted features as a way of distinguishing a graph from the others.
However, although hand-crafted features give the traditional methods good interpretability, they typically have high computation complexity, which is a major obstacle to their deployment.
The technique are complex because: (1) hand-crafted features in large-scale graph data are usually too sparse; and (2) most graph-level learning algorithms relying on these hand-crafted features usually involve graph isomorphism problems.
For instance, some traditional methods classify graphs based on a series of substructures decomposed from a graph dataset - the primary problem in the decomposition being to decide whether there are isomorphic substructures.
Deep learning techniques, however, have seen the dawn of a new era for graph-level learning.
With these methods, the non-linear and task-specific features of graphs can be learned automatically and in an end-to-end fashion, which has resulted in new benchmark for state-of-the-art performance. Additionally, the high-dimensional representations learned by deep neural networks works to support the ever-increasing size of graph data.
The fly in the ointment is that the black-box nature of deep learning leads to compromised trustworthiness, so a new trend is to develop reliable neural networks through traditional techniques.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{Intro_graph-level.pdf}
\caption{Toy examples of graph learning on a single graph and graph datasets.}
\label{intro_graphlevel}
\end{figure*}
Additionally, understanding the various graph-level learning techniques spanning both traditional and deep learning requires a comprehensive review of the vast amount of graph-level learning literature.
There are surveys on learning graph-structured data.
However, these reviews suffer from two main disadvantages.
First, most existing surveys concentrate on articles that explore the node/edge/substructures in a single graph, such as network embedding \cite{cui2018survey}, community detection \cite{su2022comprehensive,10.5555/3491440.3492133}, anomaly detection \cite{ma2021comprehensive}, and graph neural networks \cite{wu2020comprehensive,zhang2020deep}. Graph-level learning is treated as a by-product taking up a subsection or less. The differences between graph learning on a single graph and graph-level learning are illustrated in Fig. \ref{intro_graphlevel}.
Second, graph-level learning is only investigated from a single perspective, such as graph kernels \cite{kriege2020survey} or graph pooling \cite{liu2022graph}.
As such, the surveys are not comprehensive as they overlook the interactions between different graph-level learning techniques.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{timeline.pdf}
\caption{The timeline of graph-level learning in terms of four mainstream techniques.}
\label{timeline}
\end{figure*}
To the best of our knowledge, this is the first comprehensive survey of graph-level learning that, spans both traditional methods and deep learning-based techniques.
This article exhaustively depicts the mainstream techniques in different periods of graph-level learning (see Fig. \ref{timeline}), and further discusses the evolutions, interactions, and communications between them.
Thus, the contributions of this survey include:
\begin{itemize}
\item \textbf{A comprehensive taxonomy:} We propose a comprehensive taxonomy for graph-level learning techniques. Specifically, our taxonomy covers graph-level learning through both traditional and deep learning methods.
\item \textbf{An in-depth review:} Over four categories, we summarize the representative algorithms, make comparisons, and discuss the contributions and limitations of existing methods.
\item \textbf{Abundant resources:} This survey provides readers with abundant resources of graph-level learning, including information on the state-of-the-art algorithms, the benchmark datasets for different domains, fair evaluation metrics for different graph-level learning tasks, and practical downstream applications. The repository of this article is available at \href{https://github.com/ZhenyuYangMQ/Awesome-Graph-Level-Learning}{https://github.com/ZhenyuYangMQ/Awesome-Graph-Level-Learning}.
\item \textbf{Future directions:} We identify 13 important future directions in the graph-level learning area.
\end{itemize}
\section{Future Directions}
Although graph-level learning has gone through a long journey, there are still open issues that have been less explored. In this section, we spotlight 13 future directions on the topic of graph-level learning for readers to refer to.
\subsection{Graph-level Imbalance Learning}
A machine learning model trained on the data with an imbalanced label distribution might be biased towards the majority classes. That is, with many samples and the minority classes consisting only of a small number of samples, the model may be under-fit.
Representative tasks that need imbalance learning and must distinguish between samples from the majority and minority classes include anomaly detection \cite{zhang2022dual} and long-tail event detection \cite{agarwal2012catching}.
\emph{Opportunities}:
Although imbalanced learning has been a long-standing issue in deep learning, graph-level imbalance learning, especially with deep models, is underexplored.
Wang \textit{et al.} \cite{wang2021imbalanced} over-sampled graphs in the minority class to relieve imbalance distributions between the majority and minority classes.
They also appended a self-consistency between the original and the augmented graphs.
Over-sampling the minority samples is a traditional solution to imbalanced learning. However, this approach has been criticized for some shortcomings, such as over-fitting and changing the original distribution of the dataset.
Additionally, minority graphs generally contain special substructures that are different from those in the majority graphs.
Strengthening the structural awareness of the current graph-level learning tools could be a feasible way of overcoming this problem.
\subsection{Graph-level Anomaly Detection}
The aim of anomaly detection is to identify objects that significantly deviate from the majority of other objects.
However, when it comes to graph-structured data, almost all graph anomaly detection research focuses on detecting anomalous nodes in a single graph \cite{ma2021comprehensive}.
\emph{Opportunities:} Graph-level anomaly detection that identifies anomalous graphs in a graph dataset is a research topic of great value application-wise.
For example, such a method could help to detect proteins with special functions from a large number of common protein structures.
Some pioneering studies \cite{ma2022deep, qiu2022raising, zhao2021using} combine state-of-the-art GL-GNNs with traditional anomaly detection methods (e.g., one-class classification \cite{ruff2018deep}) to detect anomalous graphs in a graph dataset.
However, these graph convolution operations were not specifically designed to detect anomalous graphs.
Most graph convolution works like a low-pass filter \cite{balcilar2020analyzing} that smooths the anomalous information in a graph \cite{tang2022rethinking}. Hence, more analysis of the reasons behind anomalous graphs is needed and specific graph convolutions need to be proposed that are purposefully designed to detect anomalous graph information manifested in graph structures and/or attributes.
\subsection{Out-of-Distribution Generalization}
Out-of-distribution (OOD) learning improves a model's generalization ability.
It applies to scenarios where the test data does not have the same distribution as the training data.
OOD settings can have two types of distribution shift, concept shift and covariate shift.
Concept shift refers to situations where the conditional distribution between the inputs and outputs differs from the training data to the test data.
Covariate shift means that the test data has some certain features not shown in the training data.
\emph{Opportunities:} Almost all the graph-level learning algorithms assume that the training and the test data will have the same distribution. However, this I.I.D. (independent, identically distributed) assumption may be violated in some scenarios.
For example, molecules with the same function may contain some different scaffolds.
When the test data have a scaffold that has never appeared in training data, graph-level learning methods models will not perform nearly as well. The graph-level learning community has recently noticed this issue and has embarked on related research in response.
Gui \textit{et al.} \cite{gui2022good} proposed a graph OOD learning benchmark. Inspired by invariant learning, Wu \textit{et al.} \cite{wu2022discovering} identified the casual subgraphs that are invariant across different distributions to improve the OOD generalization ability of GL-GNNs. Similarly, Bevilacqua \textit{et al.} \cite{bevilacqua2021size} employed an inference model to capture approximately invariant causal graphs to improve the extrapolation abilities of GL-GNNs.
In addition to invariant learning, many techniques such as meta-learning, data augmentation, and disentanglement learning are feasible for OOD learning. Combining these techniques with GL-GNNs is likely to be the future of achieving graph-level learning models with a strong OOD generalization capacity.
\subsection{Informative Graph Pooling}
We categorized the existing pooling techniques into two families, i.e., global and hierarchical pooling (see Section \ref{Pooling}).
The aim of the top-$k$ approaches \cite{zhang2018end,zhang2020structure}, which are among the most representative global pooling methods, is to select some nodes for the pooled graph.
However, one cannot ensure that the redundancy of the selected nodes will be low.
Further, the mechanism of the hierarchical family tends to smooth the node representations, which means the uninformative nodes tend to be selected for the pooled graphs \cite{mesquita2020rethinking}.
\emph{Opportunities:} Existing state-of-the-art graph pooling methods are not able to coarsen the original graph into a pooled graph with nodes of low redundancy. However, a pooled graph consisting of dissimilar nodes is critical for graph-level learning. For example, an atomic pair composed of different atoms can empower different proprieties to molecules.
Traditional subgraph mining methods \cite{wu2014multi,wu2014bag} can then be used to identify the discriminative subgraphs of low redundancy as representative graphs.
Hence, referring to the ideas of traditional subgraph mining for graph pooling methods that can identify informative nodes and/or subgraphs might yield feasible solutions to this problem.
\subsection{Brain Graphs Analytics}
Brain networks, also known as connectomes, are maps of the brain where the nodes denote the brain regions of interest (ROIs) in the brain and the edges denote the neural connections between these ROIs.
An important application of machine learning models pertaining to brain networks is to distinguish brains with neurological disorders from normal individuals and identify those regions of the brain that are the cause of brain disease.
\emph{Opportunities:} Existing graph-level learning algorithms especially GL-DNNs and GL-GNNs, tend to be over-parameterized for learning brain networks, which are usually sparse. Further, obtaining a brain network usually comes at a high cost, because it involves scanning an individual's brain and converting the neuro-image into a brain network.
In addition, existing GL-DNNs and GL-GNNs cannot handle the correspondence of nodes between different graphs. However, different brain networks have the same ROIs, and node identities and ROIs are one-to-one correspondence \cite{sporns2022graph}.
In summary, graph-level learning with brain networks requires models that are lightweight and can identify corresponding nodes between different graphs.
\subsection{Graph-level Federated Learning}
Graph data are generally sourced from information collected by institutions.
However, due to privacy considerations, graph data from different institutions is generally not used to jointly train graph-level learning models.
In practice, numerous graph-level learning techniques are data-hungry, especially the currently mainstream GL-GNNs.
Therefore, it is a practical topic to promote the joint training of graph-level learning models by different institutions using their respective graph data.
\emph{Opportunities:} Federated learning solves the data isolation problem, feeding data-driven machine learning models from different sources with rich amounts of data while maintaining privacy.
For example, Xie \textit{et al.} \cite{xie2021federated} proposed a federated learning framework specifically for GL-GNNs, where different GL-GNNs are trained based on different graph sets and sharing weights are learned by the GL-GNNs.
Graph-level federated learning is an emerging topic with great challenges.
In fact, a benchmark for this task has recently been released \cite{he2021fedgraphnn}.
\subsection{Graph-level Interaction Learning}
Almost all the literature on graph-level learning treats each graph in a dataset as an independent sample.
However, considering the interactions between graphs should lead to challenging and highly novel research.
For example, learning the interactions between graphs might be used to predict the chemical reactions when two compounds meet or to explore the effect of taking two or more drugs at the same time.
\emph{Opportunities:} Although this topic has strong practical implications for graph-level learning applications in biochemistry, it is still understudied.
So far, only a few GL-GNNs have been designed to tackle this topic and its related tasks.
DSS-GNN \cite{bevilacqua2021equivariant}, for instance, predicts the interactions between subgraphs located in a single graph, while Graph of Graphs Neural Network (GoGNN) \cite{10.5555/3491440.3491623} predicts chemical-chemical and drug-drug interactions.
These two tasks own the off-the-shelf datasets, DDI \cite{Zitnik2018}, CCI \cite{kuhn2007stitch}, and SE \cite{Zitnik2018}.
\subsection{Graph-level Fairness Learning}
The bias in data can easily lead to issues with fairness, where machine learning models make discriminatory predictions towards certain demographic groups based on sensitive attributes such as race.
One feasible solution to debiasing the data is to conduct a competitive game between a biased and a debiased encoder.
The game is won when the fairness-aware debiaser is able to cheat its competitor \cite{bose2019compositional,masrour2020bursting}.
Other algorithms add constraints to the loss function to counterbalance model performance with fairness \cite{kang2020inform,li2020dyadic}.
\emph{Opportunities:}
Most work on improving the fairness of models have involved node-level tasks and single graphs \cite{dong2022edits}.
However, injecting an awareness of fairness into graph-level learning algorithms is also critical work.
Some graph-level learning tasks, such as disease prediction and fraud detection, demand fair results if they are to accurately guide people's decision-making.
One challenge to be overcome in attempting to make graph-level learning fair is that the representative GL-GNNs, i.e., MPNNs, will tend to produce unfair predictions in the face of data bias because the message passing mechanisms actually spread the bias via neighborhood structures \cite{dai2021say}.
We refer readers who are interested in this topic to \cite{dai2022comprehensive}, which gives an exhaustive introduction to fairness learning with graph-structured data.
\subsection{Graph-level Learning on Complex Graphs}
In this survey, almost all the investigated graph-level learning methods are assumed to work on fundamental graphs (i.e., unweighted and undirected graphs and their nodes and edges are homogeneous).
This is because fundamental graphs are easy to understand and easy for models to handle. However, realistic graphs are usually complex. For example, the edges between actors and movies have a different meaning to the edges between two movies in multi-relational graphs. Collaborators on a paper can be linked together by a hyperedge (i.e., hypergraphs), while authors, papers, and venues can all be nodes in a citation network, even though they are distinct taxonomic entities (i.e., heterogeneous graphs), etc.
\emph{Opportunities:} Compared to highly developed graph-level learning on fundamental graphs, mining complex graphs still requires further development.
For instance, most GL-GNNs for heterogeneous graphs rely on manually-defined meta-paths (i.e., a sequence of relations between nodes or edges) that are based on domain knowledge.
However, defining meta-paths is not only expensive, it will not capture comprehensive semantic relationships \cite{hussein2018meta,yang2021interpretable}.
Lv \textit{et al.} \cite{lv2021we} also raised the issue that, empirically, some heterogeneous GL-GNNs do not perform as well as simple GL-GNNs.
In addition, it is hard to fairly evaluate hypergraph GL-GNNs since the hypergraphs are acquired from a range of different sources and built by a range of different construction approaches.
In conclusion, there are numerous worthwhile directions to explore when it comes to graph-level learning with complex graphs, such as benchmarking evaluation \cite{lv2021we} and datasets.
\subsection{Neural Architecture Search (NAS) for GL-GNNs}
Existing GL-GNNs often have a complex architecture, consisting of a number of different components, e.g., multiple graph convolutions and graph pooling layers.
GL-GNNs require careful parameter tuning to achieve optimal performance since most of them are non-convex.
Hence, it is expensive to search for a well-performing architecture from among the bulk of optional components and their numerous parameters.
\emph{Opportunities:} Developing effective NAS methods to free researchers from the task of repeatedly searching for good architectures manually and, in turn, tuning the parameters is an urgent goal.
As a pioneering work in NAS, Knyazev \textit{et al.} \cite{knyazev2021parameter} modeled the search for an architecture as a graph in which each node represents a neural network layer or operation (e.g., a convolution layer) and each edge represents the connectivity between a pair of operations. Subsequently, GNNs can work on the constructed graphs to seek the optimal architecture.
Also, Pooled Architecture Search (PAS) \cite{wei2021pooling} is a preliminary study on searching for an adaptive graph pooling component for graph classification tasks.
We argue that constructing an optimization goal based on knowledge of deep learning might be a practical way of providing an automatic NAS for various GL-GNNs.
\subsection{Geometrically Equivariant GL-GNNs}
In geometric graphs \cite{bronstein2021geometric}, each node is described by two vectors, i.e., a feature vector and a geometric vector. For example, in 3D molecule graphs, atoms are assigned geometric information such as speeds, coordinates, and spins which together comprise the geometric vector.
Constructing GL-GNNs that can learn geometric graphs is an important part of modeling in chemistry and physics.
\emph{Opportunities:} GL-GNNs that can predict a set of geometric graphs need to be equivariant. For example, when inputting a geometric graph with a specific rotation into a GL-GNN, the corresponding output should reflect the same rotation. There are some algorithms about geometrically equivariant GL-GNNs. For example, Satorras \textit{et al.}'s \cite{satorras2021n} Equivariant Graph Neural Networks (EGNN) expands MPNNs aggregating both feature vectors and geometric vectors, while GemNet \cite{gasteiger2021gemnet} infuses more geometric information into the message passing mechanism, like dihedral angles.
Both of these methods achieve state-of-the-art performance with 3D molecule prediction tasks.
For more details on this topic, we refer readers to \cite{han2022geometrically}.
\subsection{Self-explainable GL-GNNs}
Most algorithms for explaining the predictions of GL-GNNs are post-hoc (e.g., PGExplainer \cite{luo2020parameterized}), where the aim is to train a model to interpret a pre-trained GL-GNN.
In other words, the training and explaining processes in GL-GNNs are independent.
\emph{Opportunities:} Miao \textit{et al.} \cite{miao2022interpretable} proposed that the separate prediction and explanation processes will inevitably lead to sub-optimal model performance. For example, the explanation model may detect substructures that have spurious correlations to the graph labels when interpreting predictions \cite{wu2022discovering}. Designing self-explaining GL-GNNs where the prediction and explanation components enhance each other should therefore be a fruitful future direction of research for the graph-level learning community.
\subsection{Multi-Graph-Level Learning} Standard graph-level learning views each graph as an instance, which can be restrictive in practical applications.
Considering a product that has multiple reviews on an online shopping page.
Each review can be represented as a graph of the textual semantics among the words.
To predict any properties of that online product, one needs to learn from review-based multi-graphs \textemdash that is, multi-graph-level learning.
\emph{Opportunities:} To the best of our knowledge, the current multi-graph-level learning algorithms are all traditional.
For example, Boosting based Multi-graph Graph Classification (bMGC) \cite{wu2014boosting} and Multi-Instance Learning Discriminative Mapping (MILDM) \cite{wu2018multigraph} are both subgraph mining methods that classify multi-graph objects by extracting informative subgraphs from a set of objects.
However, both two methods cannot use label information to guide the feature selection process due to the limitations of traditional learning.
Developing deep learning models can better extract features for multi-graph-level learning through the label information.
\section{Conclusions}
This survey paper provides a comprehensive review of graph-level learning methods.
Due to the irregular structure of graphs, graph-level learning has long been a non-trivial task with related research Spanning the traditional to the deep learning era.
However, the community is eager for a comprehensive taxonomy of this complex field.
In this paper, we framed the representative graph-level learning methods into four categories based on different technical directions.
In each category, we provided a detailed discussion on, and comparison of, the representative methods.
We also discussed open-source materials to support research in this field, including datasets, algorithm implementations, and benchmarks, along with the most graph-level learning tasks and their potential industrial applications.
Lastly, we raised 13 future directions based on currently open issues that would make valuable contributions to the graph-level learning community.
\section{Definitions}
\newtheorem{definition}{Definition}
This section, provides some definitions that are essential to understanding this paper. Bold lowercase characters (e.g., $\mathbf x$) are used to denote vectors.
Bold uppercase characters (e.g., $
\mathbf X$) are used to denote matrices.
Plain uppercase characters (e.g., $\mathcal V$) are used to denote mathematical sets, and lowercase -italic characters (e.g., $n$) are used to denote constants.
\begin{definition}
\textbf{(Graph)}: A graph can be denoted as $\mathcal G = \{\mathcal V, \mathcal E\}$, where the node set $\mathcal V$ having $n$ nodes (also known as vertices) and the edge set $\mathcal E$ having $m$ edges. $\mathcal E_{u,v} \in \mathcal E$ represents that there is an edge connecting nodes $u$ and $v$, where $u\in\mathcal V$ and $v\in \mathcal V$. If $\mathcal G$ is unweighted, we use an adjacency matrix $\mathbf A \in \{0,1\}^{n \times n}$ to describe its topological structure, where $\mathbf {A} _{u,v} = 1$ if $\mathcal E_{u,v} \in \mathcal E$, otherwise, 0. If $\mathcal G$ is weighted, the value of $\mathbf {A} _{u,v}$ refers to the weight value of $\mathcal E_{u,v}$. $\mathbf X\in \mathbb R^{n\times f}$ is the node attribute matrix and a node $u\in \mathcal V$ can be described by an attribute vector $\mathbf x_{u} \in \mathbb R^f$. Similarly, the edge feature matrix is denoted as $\mathbf S\in \mathbb R^{m\times d}$, where $\mathbf s_{u, v}\in \mathbb R^{d}$ describes the edge $\mathcal E_{u, v}\in \mathcal E$.
\end{definition}
\begin{definition}
\textbf{(Graph Dataset)}: A graph dataset $\mathbb G$ is composed of $N$ graphs, where $\mathbb G=\{\mathcal G_1, ..., \mathcal G_N\}$.
\end{definition}
\begin{definition}
\textbf{(Subgraph/Substructure)}: A graph $g_m = \{\mathcal V_{g_m}, \mathcal E_{g_m}\}$ can be regarded as the subgraph/substructure of $\mathcal G=\{\mathcal V, \mathcal E\}$ \textit{iff} $\mathcal V_{g_m}\subseteq \mathcal V$ and $\mathcal E_{g_m}\subseteq \mathcal E$.
\end{definition}
\begin{definition}
\textbf{(Graph-level Learning)}:
Graph-level learning takes a graph dataset $\mathbb G = \{\mathcal G_1,...,\mathcal G_N\}$ consisting of $N$ graphs as inputs and takes each graph $\mathcal G_i \in \mathbb G$ as the object of study. The graph structure is permutation-invariant \textemdash any changes in the ordering of nodes and/or edges do not change the structure itself. Hence, the graph-level learning algorithms should also be permutation invariant. That is to say, the prediction results of a graph-level learning algorithm are invariant to any permutations of the order of nodes and/or edges of each input graph.
\end{definition}
\section{Taxonomy of Graph-level Learning Techniques}
\begin{figure*}[htbp!]
\centering
\includegraphics[width=0.8\textwidth]{tree.pdf}
\caption{The taxonomy tree of graph-level learning techniques.}
\label{tree}
\end{figure*}
This section provides a taxonomy of graph-level learning techniques.
Its categories include traditional learning, graph-level deep neural networks (GL-DNNs), graph-level graph neural networks (GL-GNNs), and graph pooling.
The taxonomy tree in Fig. \ref{tree} depicts these four branches of graph-level learning with selected algorithms highlighted.
Each category is briefly introduced next.
\subsection{Traditional Learning:}
Traditional forms of learning used to be the mainstream in graph-level learning before the great success of deep learning techniques emerged.
The majority of traditional learning methods rely on handcrafted features to describe the graphs, such as random walk sequences \cite{kriege2020survey}.
Given deterministic graph descriptions, off-the-shelf machine learning models were used to perform downstream tasks, like graph classification, in a non-end-to-end fashion.
Even today, traditional methods have some advantages over deep learning techniques, such as better interpretability and a better ability to model irregular structures.
For these reasons, traditional methods are still providing the graph-level learning community with valuable insights.
\subsection{Graph-Level Deep Neural Networks (GL-DNNs):} Traditional approaches do not just include classic methods.
They also include tentative explorations into deep learning techniques, such as RNNs, CNNs, and CapsNet.
These three types of deep neural networks were not initially designed to learn non-Euclidean data like graphs. Hence, one of the important issues with GL-DNNs is how to enable these deep neural networks to learn graph-structured data that varies in size and has irregular neighborhood structures.
\subsection{Graph-Level Graph Neural Networks (GL-GNNs):} GL-GNNs use graph convolution operations specifically proposed for graph-structured data as the backbone for performing graph-level learning.
Most GL-GNNs use the graph convolutions MPNNs frameworks because they are simple, easy to understand, and have linear complexity. GL-GNNs condense the most fruitful achievements of graph-level learning.
In addition, some practitioners integrate the advantages of MPNN-based GL-GNNs with other techniques, particularly traditional learning techniques, to improve graph-level learning.
\subsection{Graph Pooling:} Graph pooling is a kind of graph downsizing technology where compact representations of a graph are produced by compressing a series of nodes into a super node.
For example, graph pooling can globally aggregate all of a graph's node embeddings into one ultimate super node via a summation or averaging operation.
At the same time, graph pooling can reduce the size of a graph layer by layer.
This kind of multi-layer downsizing process tends to aggregate nodes in the same hierarchy (e.g., a community) at each layer.
\section{Traditional Learning}
\begin{table*}[]
\caption{Summary of Traditional Graph-level Learning Methods.}\label{table_traditional}
\begin{tabular}{cclllll}
\hline
Subsection & Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{17}{*}{\begin{tabular}[c]{@{}c@{}}Graph \\ Kernels\end{tabular}} & \multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Message\\ Passing\\ Kernels\end{tabular}} & 2009 & NHK\cite{hido2009linear} & ICDM & Python & https://github.com/ysig/GraKeL \\ \cline{3-7}
& & 2011 & WL\cite{shervashidze2011weisfeiler} & JMLR & C++ & https://github.com/BorgwardtLab/graph-kernels \\ \cline{3-7}
& & 2016 & PK\cite{neumann2016propagation} & ML & MATLAB & https://github.com/marionmari/propagation\_kernels \\ \cline{3-7}
& & 2017 & Global-WL\cite{morris2017glocalized} & ICDM & C++ & https://github.com/chrsmrrs/glocalwl \\ \cline{3-7}
& & 2019 & P-WL\cite{rieck2019persistent} & ICML & Python & https://github.com/BorgwardtLab/P-WL \\ \cline{2-7}
& \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}ShortestPath\\ Kernels\end{tabular}} & 2005 & SPK\cite{borgwardt2005shortest} & ICDM & Python & https://github.com/ysig/GraKeL \\ \cline{3-7}
& & 2017 & SPK-DS\cite{nikolentzos2017shortest} & EMNLP & - & - \\ \cline{2-7}
& \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Random\\ Walk\\ Kernels\end{tabular}} & 2003 & RWK\cite{gartner2003graph} & LNAI & Python & https://github.com/jajupmochi/graphkit-learn \\ \cline{3-7}
& & 2004 & ERWK\cite{mahe2004extensions} & ICML & Python & https://github.com/jajupmochi/graphkit-learn \\ \cline{3-7}
& & 2010 & SOMRWK\cite{vishwanathan2010graph} & JMLR & Python & https://github.com/ysig/GraKeL \\ \cline{2-7}
& \multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}Optimal\\ Assignment\\ Kernels\end{tabular}} & 2005 & OAK\cite{frohlich2005optimal} & ICML & - & - \\ \cline{3-7}
& & 2013 & PS-OAK\cite{pachauri2013solving} & NeurIPS & Python & https://github.com/zju-3dv/multiway \\ \cline{3-7}
& & 2015 & GE-OAK\cite{johansson2015learning} & KDD & - & - \\ \cline{3-7}
& & 2015 & TAK\cite{schiavinato2015transitive} & SIMBAD & - & - \\ \cline{2-7}
& \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Subgraph\\ Kernels\end{tabular}} & 2009 & Graphlet\cite{shervashidze2009efficient} & AISTATS & Python & https://github.com/ysig/GraKeL \\ \cline{3-7}
& & 2010 & NSPDK\cite{costa2010fast} & ICML & Python & https://github.com/fabriziocosta/EDeN \\ \cline{3-7}
& & 2012 & SMK\cite{10.5555/3042573.3042614} & ICML & C++ & https://github.com/fapaul/GraphKernelBenchmark \\ \hline
\multirow{9}{*}{\begin{tabular}[c]{@{}c@{}}Subgraph \\ Mining\end{tabular}} & \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Frequent\\ Subgraph\\ Mining\end{tabular}} & 2000 & AGM\cite{inokuchi2000apriori} & ECML PKDD & C++ & https://github.com/Aditi-Singla/Data-Mining \\ \cline{3-7}
& & 2001 & FSG\cite{kuramochi2001frequent} & ICDM & C++ & https://github.com/NikhilGupta1997/Data-Mining-Algorithms \\ \cline{3-7}
& & 2002 & gSpan\cite{yan2002gspan} & ICDM & Python & https://github.com/betterenvi/gSpan \\ \cline{2-7}
& \multirow{6}{*}{\begin{tabular}[c]{@{}c@{}}Discrimina\\ -tive\\ Subgraph\\ Mining\end{tabular}} & 2008 & LEAP\cite{yan2008mining} & SIGMOD & - & - \\ \cline{3-7}
& & 2009 & CORK\cite{thoma2009near} & SDM & - & - \\ \cline{3-7}
& & 2010 & gMLC\cite{kong2010multi} & ICDM & - & - \\ \cline{3-7}
& & 2010 & gSSC\cite{kong2010semi} & KDD & - & - \\ \cline{3-7}
& & 2011 & gPU\cite{zhao2011positive} & ICDM & - & - \\ \cline{3-7}
& & 2014 & gCGVFL\cite{wu2014multi} & ICDM & - & - \\ \hline
\multirow{9}{*}{\begin{tabular}[c]{@{}c@{}}Graph \\ Embedding\end{tabular}} & \multirow{6}{*}{\begin{tabular}[c]{@{}c@{}}Determin\\ -istic\end{tabular}} & 2017 & FGSD\cite{verma2017hunt} & NeurIPS & Python & https://github.com/vermaMachineLearning/FGSD \\ \cline{3-7}
& & 2018 & AWE\cite{ivanov2018anonymous} & ICML & Python & https://github.com/nd7141/AWE \\ \cline{3-7}
& & 2019 & LDP\cite{cai2018simple} & ICLR RLGM & Python & https://github.com/Chen-Cai-OSU/LDP \\ \cline{3-7}
& & 2020 & SLAQ\cite{tsitsulin2020just} & WWW & Python & \begin{tabular}[c]{@{}l@{}}https://github.com/google-research/google-research/tree/\\ master/graph\_embedding/slaq\end{tabular} \\ \cline{3-7}
& & 2021 & VNGE\cite{liu2021bridging} & WWW & Python & https://github.com/xuecheng27/WWW21-Structural-Information \\ \cline{3-7}
& & 2021 & A-DOGE\cite{sawlani2021fast} & ICDM & Python & https://github.com/sawlani/A-DOGE \\ \cline{2-7}
& \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Learn\\ -able\end{tabular}} & 2016 & Subgraph2vec\cite{narayanan2016subgraph2vec} & KDD MLG & Python & https://github.com/MLDroid/subgraph2vec\_tf \\ \cline{3-7}
& & 2017 & Graph2vec\cite{narayanan2017graph2vec} & KDD MLG & Python & https://github.com/MLDroid/graph2vec\_tf \\ \cline{3-7}
& & 2018 & GE-FSG\cite{nguyen2018learning} & SDM & Python & https://github.com/nphdang/GE-FSG \\ \hline
\end{tabular}
\end{table*}
Traditional graph-level learning algorithms work in a deterministic way, encoding graphs using handcrafted features.
Traditional graph-level learning methods can be divided into three main types: i.e., those based on graph kernels (GKs, Section \ref{41}), subgraph mining (Section \ref{42}), and graph embedding (Section \ref{43}).
All traditional graph-level learning publications discussed in this section are summarized in Table \ref{table_traditional}.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1\textwidth]{GraphKernels.pdf}
\caption{Different mechanisms of four graph kernels in decomposing and comparing pairwise graphs.}
\label{fgraphkernel}
\end{figure*}
\subsection{Graph Kernels (GKs)}\label{41}
GKs perform graph-level learning based on kernel values (i.e., pair-wise graph similarities).
Given a graph dataset $\mathbb G$, GKs decompose each graph $\mathcal G$ into a bag-of-graphs $S^{\mathcal G}=\{g_1,...,g_I\}$, where $g_i\subseteq \mathcal G$ and $g_i$ can be a node or a subgraph.
Most GKs are based on the paradigm of an $R$-Convolution kernel that obtains the kernel value $K_{R-conv} \left(\mathcal G,\mathcal G'\right)$ of two graphs $\mathcal G$ and $\mathcal G'$ by:
\begin{equation}
K_{R-conv} \left(\mathcal G,\mathcal G'\right) = \sum_{i=1}^{I}\sum_{j=1}^{J} K_{parts}\left(g_i,g'_j\right),
\end{equation}
where $K_{parts}\left(g_i,g'_j\right)$ is the kernel function that defines how to measure the similarity between $g_i$ and $g'_j$.
A kernel matrix that packages all kernel values is then fed into an off-the-shelf machine learning model, such as a support vector machine (SVM), to classify the graphs.
\subsubsection{Message Passing Kernels (MPKs)}\label{MPKs_formal} MPKs perform message passing on neighborhood structures to obtain graph representations.
The 1-dimensional Weisfeiler-Lehman (1-WL) algorithm\footnote{1-WL is also a well-known algorithm for graph isomorphism test.} \cite{weisfeiler1968reduction,shervashidze2011weisfeiler} is one of the most representative MPKs. 1-WL updates a node's label (or color) iteratively. An illustration of $1$-th iteration is shown in Fig. \ref{fgraphkernel} (D).
At the $h$-th iteration, 1-WL aggregates node $v$'s label $l^{(h-1)}(v)$ and its neighbor's labels $l^{(h-1)}(u), u\in\mathcal N(v)$ to form a multi-set\footnote{In a multiset, multiple elements are allowed to be the same instance.} of labels $\{l^{(h-1)}(v),sort(l^{(h-1)}(u):u\in\mathcal N(v))\}$.
Subsequently, 1-WL employs an injective hash function $\phi(\cdot)$ to map the $\{l^{(h-1)}(v),sort(l^{(h-1)}(u):u\in\mathcal N(v))\}$ into a new label $l^{(h)}(v)$. Formally:
\begin{equation}\label{MPK}
l^{(h)}(v) = \phi\left(l^{(h-1)}(v),sort(l^{(h-1)}(u):u\in\mathcal N(v))\right).
\end{equation}
When $\phi(\cdot)$ no longer changes the labels of any nodes, 1-WL stops iterating and generates a vector $\phi_{wl}(\mathcal G)$ that describes $\mathcal G$. That is,
\begin{equation}
\phi_{wl}\left(\mathcal G\right) = [c^{(0)}(l^{(0)}_1),.., c^{(0)}(l^{(0)}_{I_0});...;c^{(H)}(l^{(H)}_1),...,c^{(H)}(l^{(H)}_{I_H})],
\end{equation}
where $l^{(h)}_i$ is the $i$-th label generated at the $h$-th iteration, and $c^{(h)}(l^{(h)}_i)$ counts the occurrences of nodes labeled with $l^{(h)}_i$ in the $h$-th iteration.
The kernel value of 1-WL between $\mathcal G$ and $\mathcal G'$ is the inner product of $\phi_{wl}\left(\mathcal G\right)$ and $\phi_{wl}\left(\mathcal G'\right)$:
\begin{equation}
K_{WL}\left(\mathcal G,\mathcal G'\right) = <\phi_{wl}\left(\mathcal G\right),\phi_{wl}\left(\mathcal G'\right)>.
\end{equation}
The followed upgrading of 1-WL mainly focuses on aggregation and relabeling steps.
Hido and Kashima \cite{hido2009linear} replaced the hash function with a binary arithmetic giving rise to a faster $\phi(\cdot)$.
Morris \textit{et al.} \cite{morris2017glocalized} used the idea of $k$-WL to relabel node groups consisting of $k$ nodes that could form a connected graph.
Theoretically, $k$-WL is more powerful than $1$-WL for distinguishing between graph structures.
Further, Neumann \textit{et al.} \cite{neumann2016propagation} proposed a random label aggregation process based on node label distributions that only considers labels of part of neighbors.
Random label aggregation saves time and computational resources making work on large-scale graphs more efficient.
Persistent Weisfeiler–Lehman (P-WL) \cite{rieck2019persistent} is the recent enhancement to MPKs that adds weighted edges into the aggregation process. To calculate the edge weight, P-WL measures the distance between the continuous iterative updated labels of two end nodes.
Additionally, P-WL can track changes in substructures that cannot be identified by 1-WL, such as cycles.
\subsubsection{Shortest-path Kernels (SPKs)}
SPKs denote the kernel value as a comparison between pair-wise node sequences (see Fig. \ref{fgraphkernel} B). For example, the shortest-path kernel \cite{borgwardt2005shortest} determines the shortest path between the vertices $v$ and $u$ via the Floyd-Warshall \cite{floyd1962algorithm} or Dijkstra's \cite{dijkstra1959note} algorithms. The distance between the pairwise shortest paths from $\mathcal G$ and $\mathcal G'$ is defined as the kernel value between them. Formally,
\begin{equation}
\begin{aligned}
&K_{SP}\left(\mathcal G,\mathcal G'\right) = \sum\limits_{\substack{v,u \in V\\ v \neq u}} \sum\limits_{\substack{v',u' \in V'\\ v' \neq u' }} K_{Parts}\left(\left(v,u\right),\left(v',u'\right)\right), \\
&\text{where} \quad K_{Parts}\left(\left(v,u\right),\left(v',u'\right)\right) := \\
&\begin{cases}
K_D\left(P\left(v,u\right),P\left(v',u'\right)\right) & \text{if} \ l\left(v\right) \equiv l\left(v'\right) \land l\left(u\right) \equiv l\left(u'\right),\\
0 & \text{otherwise},
\end{cases}
\end{aligned}
\end{equation}
where $l(v)$ is the label of node $v$, $P\left(v,u\right)$ is the length of shortest path between vertices $v$ and $u$, and $K_D(\cdot,
\cdot) $ measures the distance by computing the absolute difference. As shown in Eq. (5), the shortest-path kernel only considers the difference in length of the two shortest paths.
Nikolentzos \cite{nikolentzos2017shortest} proposed a variant of SPKs that draws on more information in a shortest path, such as node and edge labels, to calculate the distance of any two paths.
\subsubsection{Random Walk Kernels (RWKs)}\label{RWKs_formal}
RWKs are another kernel method guided by node sequences. G{\"a}rtner \textit{et al.} \cite{gartner2003graph} was the first to propose a random walk kernel. This technique counts the same random walk sequences that pair-wise graphs both own. Performing random walks on $\mathcal G = \{\mathcal V, \mathcal E\}$ and $\mathcal G' = \{\mathcal V', \mathcal E'\}$ simultaneously is the same as conducting random walks on a direct product graph $\mathcal G_{\times} = \{\mathcal V_{\times}, \mathcal E_{\times}\}$, where
\iffalse
Performing the random walks on $\mathcal G_{\times}$ is equivalent to do random walks on $\mathcal G$ and $\mathcal G'$ simultaneously.
RWKs are another node sequence-guided kernel methods, which count the same random walk sequences that pair-wise graphs both own. G{\"a}rtner \textit{et al.} \cite{gartner2003graph} firstly introduced random walks into GKs.
The direct product graph $\mathcal G_{\times} = \{\mathcal V_{\times}, \mathcal E_{\times}\}$ about two given graphs $\mathcal G = \{\mathcal V, \mathcal E\}$ and $\mathcal G' = \{\mathcal V', \mathcal E'\}$ is denoted as:
Another analogous method with the shortest path is the RWKs, which counts the number of occurrences of walks paths (i.e. node label sequences) that two graphs both own. G{\"a}rtner \textit{et al.} \cite{gartner2003graph} introduced random walks into the graph kernel method and proposed the direct product graph for walking randomly on two graphs simultaneously.
For two given graphs $\mathcal G = \{\mathcal V, \mathcal E\}$ and $\mathcal G' = \{\mathcal V', \mathcal E'\}$, the direct product graph $\mathcal G_x = \{\mathcal V_x, \mathcal E_x\}$ is denoted as:
\fi
\begin{equation}
\begin{aligned}
&\mathcal V_{\times} = \{\left(v,v'\right):v\in\mathcal V \land v'\in\mathcal V' \land l(v) \equiv l(v')\}, \\
&\mathcal E_{\times} = \{\{\left(v,v'\right),\left(u,u'\right) \in \mathcal V_{\times}\}:\mathcal E_{v,u} \in \mathcal E \land \mathcal E'_{v',u'} \in \mathcal E' \}.
\end{aligned}
\end{equation}
Given $\mathcal G_{\times}$, the kernel function is defined as:
\begin{equation}\label{krw}
K_{RW}\left(\mathcal G,\mathcal G'\right) = \sum\limits_{i=1}^{|\mathcal V_{\times}|}\sum\limits_{j=1}^{|\mathcal V_{\times}|}\left[\sum\limits_{p=0}^{P}\lambda_p \mathbf A_{\times}^p\right]_{ij},
\end{equation}
where $\mathbf A_{\times}$ is the adjacency matrix of $\mathcal G_{\times}$,
$P$ is the predefined max length of random walking sequences, and $\lambda_p$ are the weights given to different $P$.
$K_{RW}\left(\mathcal G,\mathcal G'\right)$ counts the occurrences of common walk paths in $\mathcal G$ and $\mathcal G'$ with lengths equal to or less than $P$.
The random walk kernel in Eq. (\ref{krw}) assumes a uniform distribution for the beginning and ending probabilities of the walks across two graphs.
However, Vishwanathan \textit{et al.} \cite{vishwanathan2010graph} proposed a generalized version of RWKs. Specifically, they defined $\mathbf p$ and $\mathbf q$ as the beginning and ending probability vectors in $\mathcal G$, respectively.
In addition, they used the Kronecker product operation $\otimes$ to derive $\mathbf A_{\times}$, that is $\mathbf A_{\times}= \mathbf A \otimes \mathbf A'$. Formally, the kernel value is:
\begin{equation}
K_{RW}\left(\mathcal G,\mathcal G'\right) = \sum_{l=0}^{\infty} \mu_{l} (\mathbf q \otimes \mathbf q' \mathbf)^{\top} (\mathbf A_{\times})^l (\mathbf p \otimes \mathbf p'),
\end{equation}
where $\mu_{l}$ is the convergence coefficient.
RWKs suffer from a problem called tottering, where a random walk sequence traverses $v$ to $u$ and immediately returns to $v$ via the same edge.
To address tottering, Mah{\'e} \textit{et al.} \cite{mahe2004extensions} employed a second-order Markov random walk that considers the last two steps in the current random walk sequence when deciding the next step.
\subsubsection{Optimal Assignment Kernels (OAKs)}
Fr{\"o}hlich \textit{et al.} \cite{frohlich2005optimal} was the first to propose OAKs. OAKs consider nodes as a basic unit for measuring kernel values. Of all the GKs introduced in this paper, OAKs are the only family of GKs that do not belong to $R$-Convolution paradigm. Specifically, given a fixed $i$ in Eq. (1), OAKs only add in the maximum similarity value between $g_i$ and $g'_j$ where $j\in\{1,...,J\}$. Formally, OAKs are defined as:
\begin{equation}\label{OA}
K_{OA} \left(\mathcal G,\mathcal G'\right) = \begin{cases}
\max\limits_{\pi \in \prod_J} \sum_{i=1}^{I} K_{parts}\left(g_i,g'_{\pi[i]}\right), & \text{if} \ J \geq I\\
\max\limits_{\pi \in \prod_I} \sum_{j=1}^{J} K_{parts}\left(g_{\pi[j]},g'_j\right),& \text{otherwise}
\end{cases}
\end{equation}
where $\prod_I$ represents all permutations of the indexes of a bag-of-graphs $\{1,...,I\}$, and $\pi$ is the optimal node permutation to reach maximum similarity value between two graphs.
Searching for a pair-wise element with the maximum similarity tends to be a highly time-consuming process.
Hence, to reduce the time requirement of this task, Johansson et al. \cite{johansson2015learning} mapped the graphs in geometric space and then calculated the Euclidean distance between pair-wise nodes.
This method enables OAKs to use approximate nearest neighbors algorithms in Euclidean space as a way to speed up the process.
Transitive Assignment Kernels (TAKs) \cite{schiavinato2015transitive,pachauri2013solving} are variants of OAKs.
Unlike OAKs that search for the optimal assignment among pair-wise graphs, TAKs identify node permutations that with the most similar node pairs among three or more graphs.
OAKs have been confined to node similarity measurement, although they can be extended to measure subgraph similarities so as to capture a graph's topological information \cite{woznica2010adaptive}. As discussed next, we introduce the GKs with subgraph information.
\subsubsection{Subgraph Kernels (SGKs)}
SGKs calculate the similarity between two graphs by comparing their subgraphs.
For example, the representative SGK \textemdash Graphlet Kernel \cite{shervashidze2009efficient} uses either depth-first search (DFS) or sampling to identify the subgraphs.
With these subgraphs, the vector $\phi_{SG}\left(\mathcal G\right) = [c^{\left(\mathcal G\right)}_{\mathcal T_1},...,c^{\left(\mathcal G\right)}_{\mathcal T_N}]$ is then used to describe the graph $\mathcal G$, where $\mathcal T_i$ means the $i$-th isomorphism type of subgraphs, $N$ is the total number of subgraphs' types, and $c^{\left(\mathcal G\right)}_{\mathcal T_i}$ counts the occurrences of the $\mathcal T_i$ category subgraphs in graph $\mathcal G$.
Graphlet's kernel value is then defined as the inner product of $\phi_{SG}\left(\mathcal G\right)$ and $\phi_{SG}\left(\mathcal G'\right)$:
\begin{equation}
K_{SG}\left(\mathcal G,\mathcal G'\right) = <\phi_{SG}\left(\mathcal G\right),\phi_{SG}\left(\mathcal G'\right)>.
\end{equation}
There are several different implementations of SGKs kernel functions.
For instance, Wale \textit{et al.} \cite{wale2008comparison} employed a min-max kernel $\displaystyle{ \tfrac{\sum_{i=1}^N min(c^{\left(\mathcal G\right)}_{\mathcal T_i},c^{\left(\mathcal G'\right)}_{\mathcal T_i})}{\sum_{i=1}^N max(c^{\left(\mathcal G\right)}_{\mathcal T_i},c^{\left(\mathcal G'\right)}_{\mathcal T_i})}}$ to measure the distance between two graphs.
Subgraph Matching Kernels (SMKs) \cite{10.5555/3042573.3042614} calculate the similarity between two subgraphs by counting the number of nodes with the same labels.
Then the similarities between all pairwise subgraphs sourced from the two graphs are summed as the kernel value of the SMKs.
Methods of identifying the subgraphs in SGKs have also been explored.
For example, Neighborhood Subgraph Pairwise Distance Kernels (NSPDK) \cite{costa2010fast} denotes the subgraphs as the first-, second-, and third-hop neighborhoods of pairwise vertices with the shortest path of a predefined length.
However, the main contributions of SGKs lie in assessing the similarity of graphs in terms of a set of selected subgraphs, not how the subgraphs are chosen.
More detailed and sophisticated subgraph mining methods are demonstrated next.
\subsection{Subgraph Mining}\label{42}
\begin{figure*}[htbp!]
\begin{minipage}{0.48\linewidth}
\vspace{3pt}
\centerline{\includegraphics[width=\textwidth]{SubgraphFrequent.pdf}}
\centerline{(A) Frequent Subgraph Mining (FSM).}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\vspace{3pt}
\centerline{\includegraphics[width=\textwidth]{SubgraphDiscriminative.pdf}}
\centerline{(B) Discriminative Subgraph Mining (DSM).}
\end{minipage}
\caption{Different subgraph extraction methods of FSM and DSM.} \label{fsubgraph}
\end{figure*}
Subgraph mining is similar to SGKs, where the vector $\textbf x_i = [x_i^1,...,x_i^M]^\top$ is taken as a graph-level representation of the graph $\mathcal G_i$.
Here, $x_i^m\in\{0,1\}$, $x_i^m=1$ if $g_m \subseteq \mathcal G_i$, otherwise, $x_i^m=0$. The established graph-level representation is then directly input into an off-the-shelf machine learning model, such as SVM classifier, for downstream tasks.
What is different about subgraph mining algorithms is that they place particular emphasis on how to extract the optimal subgraph set $\mathcal S^* = \{g_1,...g_T\}$ from the subgraph set $\{g_1,...g_M\}$, where $g_1, ...,g_M$ denote all possible subgraphs of $\mathbb G =\{\mathcal G_1,...,\mathcal G_N\}$. Techniques for extracting subgraphs can be divided into two branches depending on how the supervision information is used.
Frequent subgraph mining is the unsupervised method, as illustrated in Fig. \ref{fsubgraph} A, while discriminative subgraph mining is the supervised or semi-supervised method (see Fig. \ref{fsubgraph} B).
\subsubsection{Frequent Subgraph Mining (FSM)}\label{FSM} FSM identifies the subgraphs whose frequency of occurence in $\mathbb G$ sits over a predefined threshold $\delta$.
These subgraphs are then added to $\mathcal S^*$.
Apriori-like algorithms, such as AGM \cite{inokuchi2000apriori} and FSG \cite{kuramochi2001frequent}, enumerate subgraphs from size one to a predefined largest size as candidates for $\mathcal S^*$.
In the enumeration, these apriori-like algorithms pick up the candidates that occur more frequently than $\delta$ and add them to $\mathcal S^*$.
Others subgraphs are dropped and expansions based on those subgraphs are no longer considered.
Testing for subgraph isomorphism with vast numbers of candidate subgraphs can mean apriori-like algorithms suffer from computation bottlenecks.
To address this issue, gSpan \cite{yan2002gspan} employs a depth-first-search (DFS) strategy to search subgraphs, while assigning a unique DFS code of minimum length for each subgraph searched.
gSpan can then do a quick check for isomorphism by simply comparing the DFS codes of pairwise subgraphs.
\subsubsection {Discriminative Subgraph Mining (DSM)} DSM extracts discriminative subgraphs from a set of all possible subgraphs of $\mathbb G$ based on the label information.
Given a binary graph classification task, Thoma \textit{et al.} \cite{thoma2009near} defined an evaluation criterion called CORK which describes the discriminative score of a subgraph set $\mathcal S$, $\mathcal S \subseteq \{g_1,...,g_M\}$. Formally,
\begin{equation}
\begin{aligned}
&CORK(\mathcal S) = -1 \times num (\mathcal G_i, \mathcal G_j), \\ &s.t. \ \mathcal G_i \subset \mathbb G_{+} \ \land \ \mathcal G_j \subset \mathbb G_{-} \ \land \ \forall g_m \in \mathcal S : x_m^i = x_m^j,
\end{aligned}
\end{equation}
where $num(\cdot)$ counts the number of pairs of graphs $(\mathcal G_i, \mathcal G_j)$ satisfying the specific conditions.
$\mathbb G_{+}$ is the set of graphs with positive labels, while $\mathbb G_{-}$ is the set of graphs with negative labels.
The optimal subgraph set $\mathcal S^*$ has the highest CORK score among all possible subgraph sets $\mathcal S$ containing $T$ subgraphs, denoted as:
\begin{equation}
\mathcal S^* = \mathop{argmax}\limits_{\mathcal S \subseteq \{g_1,...,g_M\}} CORK(\mathcal S) \quad s.t. \ |\mathcal S| \leq T.
\end{equation}
The CORK score can also be used to prune the subgraph search space of gSpan, mentioned in Section \ref{FSM}.
More specifically, if replacing any existing element of $\mathcal S^*$ with a subgraph $g_m$ does not $\mathcal S^*$'s CORK score, gSpan will no longer perform DFS along $g_m$.
To speed up DSM based on discriminative scores and gSpan, Yan \textit{et al.} \cite{yan2008mining} proposed LEAP, which initializes an optimal subgraph set $\mathcal S^*$ with frequent subgraphs.
In this way, LEAP prunes gSpan's search space right at the beginning. In addition, Kong \textit{et al.} \cite{kong2010multi} and Wu \textit{et al.} \cite{wu2014multi} expanded DSM to the multi-label\footnote{Each graph owns more than one label, such as a drug molecular can own different labels to represent anti-cancer effects for various cancers, e.g., breast cancer (+) and lung cancer (-).} and multi-view\footnote{An object has different views, where each view can represent a separate graph, e.g., a scientific publication network is shown as two graphs, an abstract graph demonstrating the keywords correlations in the abstract of papers, and a reference citation graph about citation relationships.} scenarios, respectively.
Note, however, that all the DSM methods discussed are supervised methods.
In terms of semi-supervised subgraph mining, Kong and Yu \cite{kong2010semi} proposed gSSC which maps each graph into a new feature space by $\mathcal S^*$. Unlabeled graphs are separated from each other in the new feature space.
In the labeled group, graphs with the same label are close, whereas graphs with different labels remain distant. In addition, Zhao \textit{et al.} \cite{zhao2011positive} only used the positively labeled graphs and unlabeled graphs to select $\mathcal S^*$ when performing binary graph classification tasks. This is because sometimes the real-world data is composed of an incomplete set of positive instances and unlabeled graphs.
\subsection{Graph Embedding}\label{43}
Both the GKs and subgraph mining methods are deterministic, which is a hallmark of traditional learning.
Some graph embedding methods following a deterministic approach make hand-crafted features that reflect a graph's' inherent proprieties.
In addition, a few graph embedding methods are not strictly traditional. These methods begin to embrace deep learning, encoding graphs via learnable neural networks.
\subsubsection{Deterministic Graph Embedding} These methods extract graph-level representations from the inherent properties of graphs, e.g., their topologies and eigenspectrums.
Local Degree Profile (LDP) \cite{cai2018simple} summarizes the degree information of each node and its 1-hop neighbors as node features.
LDP constructs graph representations by building an empirical distribution or histogram of some hand-crafted node features.
In addition to node degree, deterministic graph embedding can also leverage anonymous random walk sequences to describe a graph's topological information.
Specifically, anonymous random walks record the status change of node labels. Two anonymous random walk sequences $A \to B \to A$ and $B \to A \to B$ can be both written as $1 \to 2 \to 1$. Anonymous Walk Embeddings (AWE) \cite{ivanov2018anonymous} encodes a graph via an $n$-dimensional vector in which each element represents the occurrence frequency of a specific anonymous random walk sequence.
In spectral graph theory \cite{chung1997spectral}, the spectrum of a graph is determined by its topology. Based on this theory, the Family of Graph Spectral Distances (FGSD) method \cite{verma2017hunt} proposes that the distance between the spectrums of two graphs can be used to test whether the graphs are isomorphic.
Thus, the histogram of the spectrum is used to construct a graph-level representation.
Analogously, A-DOGE \cite{sawlani2021fast} depicts a graph by computing the spectral density across its eigenspectrum.
However, these methods are limited to use with small graphs given the prohibitive costs of computing eigenspectrum decompositions with large-scale graphs.
As a possible solution to this limitation, SlaQ \cite{tsitsulin2020just} uses stochastic approximations as a way of quickly calculating the distance between two graphs' spectral densities.
More specifically, these authors employed von Neumann graph entropy (VNGE) \cite{braunstein2006laplacian,chen2019fast} as a way of approximately representing the spectral properties of the graphs.
In turn, this approximation supports fast computation by tracing a Laplacian matrix of the graph.
Liu \textit{et al.} \cite{liu2021bridging} proposed another fast approximation method involving VNGE, which is based on deriving the error bound of the approximation estimation.
\subsubsection{Learnable Graph Embedding}
Inspired by Skip-gram \cite{mikolov2013efficient}, Narayanan \textit{et al.} \cite{narayanan2016subgraph2vec} presented a subgraph representation method named subgraph2vec.
Subgraph2vec first takes the ($d$-$1$)-, $d$-, ($d$+1)-hop neighborhoods of the $v$th selected node in the graph $\mathcal G_i$ as three subgraphs $g^i_{v-1}$, $g^i_{v}$, $g^i_{v+1}$, respectively, where $ d\ge1$ is a predefined value. $\{\mathbf w^1_{1-1},...,\mathbf w^1_{V+1};...;\mathbf w^N_{1-1},...,\mathbf w^N_{V+1}\}$ are the randomly initialized embeddings of all sampled subgraphs $\{g^1_{1-1},...,g^1_{V+1};...;g^N_{1-1},...,g^N_{V+1}\}$ respectively, where $N$ represents the total number of graphs, and $V$ is the number of selected nodes in each graph. Then, the Skip-gram model is used to update the subgraph embeddings.
The Skip-gram model takes $\mathbf w^i_{v}$ as its input, and predicts the context of $\mathbf w^i_{v}$ (i.e., $\mathbf w^i_{v-1}$ and $\mathbf w^i_{v+1}$).
Then the prediction results are back-propagated to update $\mathbf w^i_{v}$.
To summarize, subgraph2vec's learning objective is to maximize the following log-likelihood:
\begin{equation}
\sum_{i=1}^{N} \sum_{v=1}^{V}\log \operatorname{Pr}\left(\mathbf w^i_{v-1}, \ldots, \mathbf w^i_{v+1} \mid \mathbf w^i_{v}\right).
\end{equation}
Another method, Graph2vec \cite{narayanan2017graph2vec} was designed to tackle graph representation tasks.
By establishing a semantic association between a graph and its sampled subgraphs, Graph2Vec employs the idea of Skip-gram to learn a graph embedding.
Following this work, Dang \textit{et al.} \cite{nguyen2018learning} replaced the sampled subgraphs in Graph2vec with frequent subgraphs that have more discriminative features for graph classification tasks.
\iffalse
\subsection{Discussion}
We must admit that traditional learning techniques have some drawbacks, such as high computational cost and poorer performance than deep learning on high-dimensional data. These weaknesses have led to a gradual cooling of research on traditional technologies. However, these traditional techniques have had a profound impact on graph-level learning, even still affecting state-of-the-art GL-GNNs. For instance, the neural network version of message passing kernel has become the most typical GL-GNN in section \ref{MPNN_INTRO}, and the WL algorithm has been used for depict the expressivity of various GL-GNNs \cite{xu2018powerful}. In addition, the graphlet kernel has been borrowed to design the state-of-art subgraph-based GL-GNNs \cite{bouritsas2022improving} in section \ref{subgraph-based GL-GNNs}.
\fi
\section{Graph-Level Deep Neural Networks (GL-DNNs)}
\begin{table*}[]
\centering
\caption{Summary of Graph-Level Deep Neural Networks (GL-DNNs).}\label{table_dnn}
\begin{tabular}{clllll}
\hline
Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}RNN-\\ Based\end{tabular}} & 2016 & GGNN\cite{li2016gated} & ICLR & Python-Tensorflow & https://github.com/Microsoft/gated-graph-neural-network-samples \\ \cline{2-6}
& 2018 & GAM\cite{lee2018graph} & KDD & Python-Pytorch & https://github.com/benedekrozemberczki/GAM \\ \cline{2-6}
& 2018 & SAN\cite{zhao2018substructure} & AAAI & - & - \\ \cline{2-6}
& 2018 & NetGAN\cite{bojchevski2018netgan} & ICML & Python-Tensorflow & https://github.com/danielzuegner/netgan \\ \cline{2-6}
& 2018 & GraphRNN\cite{you2018graphrnn} & ICML & Python-Pytorch & https://github.com/snap-stanford/GraphRNN \\ \hline
\multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}CNN-\\ Based\end{tabular}} & 2016 & PATCHYSAN\cite{niepert2016learning} & ICML & Python & https://github.com/tvayer/PSCN \\ \cline{2-6}
& 2016 & DCNN\cite{NIPS2016_390e9825} & NIPS & Python & https://github.com/jcatw/dcnn \\ \cline{2-6}
& 2017 & ECC\cite{simonovsky2017dynamic} & CVPR & Python-Pytorch & https://github.com/mys007/ecc \\ \cline{2-6}
& 2018 & KCNN\cite{nikolentzos2018kernel} & ICANN & Python-Pytorch & https://github.com/giannisnik/cnn-graph-classification \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}CapsNet-\\ Based\end{tabular}} & 2018 & GCAPSCNN\cite{verma2018graph} & WCB & Python & https://github.com/vermaMachineLearning/Graph-Capsule-CNN-Networks \\ \cline{2-6}
& 2019 & CapsGNN\cite{xinyi2018capsule} & ICLR & Python-Pytorch & https://github.com/benedekrozemberczki/CapsGNN \\ \cline{2-6}
& 2019 & PATCHYCaps\cite{mallea2019capsule} & Arxiv & Python & https://github.com/BraintreeLtd/PatchyCapsules \\ \hline
\end{tabular}
\end{table*}
GL-DNNs form the basis of a pioneering set of works that employ deep learning techniques to achieve graph-level learning. Researchers have explored graph-level learning techniques based on classic deep neural networks including recurrent neural networks (RNNs), convolution neural networks (CNNs), and capsule neural networks (CapsNets) to achieve RNN-based (see Section \ref{5.1}), CNN-based (see Section \ref{5.2}), and CapsNets-based (see Section \ref{5.3}) GL-DNNs, respectively. The representative GL-DNNs mentioned in this section are summarized in Table \ref{table_dnn}.
\subsection{RNN-based GL-DNNs}\label{5.1}
RNNs are particularly good at learning sequential data, such as text and speech.
There are two main types of algorithms that apply RNNs to graph-level learning.
One type transforms graphs into sequential-structured data.
The other aggregates neighborhood information about the target node and relabels those aggregated features through an RNN.
This is similar to Message Passing Kernels (MPKs, Section \ref{MPKs_formal}).
A natural way to capture the sequential information in graphs is to use a series of random walk paths to represent a graph.
For example, GAM \cite{lee2018graph} employs a long short-term memory (LSTM) model to guide a random walk on graphs.
Meanwhile, the LSTM model generates a representation for the walk sequence to describe the graph.
In addition, Zhao \textit{et al.} \cite{zhao2018substructure} proposed an RNN-based graph classification algorithm called SAN.
Starting from a node, SAN employs an RNN model that adds nodes and edges to form an informative substructure whose representation is progressively generated by the RNN model.
A graph-level representation that can be used for graph classification tasks is then generated by summing all the representations of the formed substructures.
Given a graph generation task, NetGAN \cite{bojchevski2018netgan} uses an LSTM model as a generator to yield fake walk sequences,
while a discriminator disambiguates the graph's real walk sequences from the generated fake ones to reverse-train the generator.
Another graph generation model Graphrnn \cite{you2018graphrnn} creates various permutations of graphs, with various combinations of nodes and edges as sequential data to be input into an RNN model.
The second category of RNN-based GL-DNNs implements neural networks version of MPKs through RNN models.
As such, the algorithms in this category can be viewed as the predecessors of MPNNs.
For instance, Li \textit{et al.} \cite{li2016gated} proposed the idea of using a gated recurrent unit (GRU) to relabel the aggregated information from the 1-hop neighborhoods of the center node.
Referring to the MPKs shown in Eq. \ref{MPK}, this approach can be defined as:
\begin{equation}
\mathbf h_v^{(k)} = \text{GRU}\left(\mathbf h_v^{(k-1)}, \text{AGG}^{(k)} \left(\mathbf h_u^{(k-1)}: u\in \mathcal N(v)\right) \right),
\end{equation}
where $\mathbf h_v^{(k)}$ represents the node representation of $v$ at the $k$-th iteration, and $\mathbf h_v^{(0)}$ is the node feature $\mathbf x_v$.
This algorithm continues the recurrent process until it hits the predefined $K$ number of iterations needed to form the node representations.
A graph-level representation is then produced via:
\begin{equation}\label{GRUPooling}
\mathbf{h}_{\mathcal{G}}=\tanh \left(\sum_{v \in \mathcal{V}} f_{t}\left(\mathbf{h}_{v}^{(K)}, \mathbf{h}_{v}^{(0)}\right) \odot \tanh \left(\mathbf{h}_{v}^{(K)}\right)\right),
\end{equation}
where $f_{t}\left(\cdot\right)$ is a softmax function guided by an attention mechanism, that preserves and aggregates valuable node representations for specific graph-level tasks. $\tanh\left(\cdot\right)$ is an activation function, and $\odot$ is element-wise multiplication.
\subsection{CNN-based GL-DNNs}\label{5.2}
\begin{figure*}[htbp!]
\begin{minipage}{1\linewidth}
\vspace{3pt}
\centerline{\includegraphics[width=\textwidth]{DNN1.pdf}}
\centerline{(A)}
\end{minipage}
\begin{minipage}{1\linewidth}
\vspace{3pt}
\centerline{\includegraphics[width=\textwidth]{DNN2.pdf}}
\centerline{(B)}
\end{minipage}
\caption{This figure shows two ways of tentative exploration by CNN-based GL-DNNs on graph-structured data.} \label{fGLDNN}
\end{figure*}
Another significant deep learning technique that works in the Euclidean domain is CNN.
Here, grid-structured data, such as images, are studied.
Similar to RNN-based GL-DNNs, there are two main branches of CNN-based graph-level learning.
The first branch is illustrated in Fig. \ref{fGLDNN} (A), which sorts nodes and arranges the node features to form a concentration matrix as the grid-structured data for training the CNNs.
As a second branch, researchers have developed a CNN-guided neural network version of an MPK, which is shown in Fig. \ref{fGLDNN} (B).
The first branch sorts nodes and arranges the node features to form a concentration matrix, of grid-structured data, to train the CNNs.
PATCHY-SAN \cite{niepert2016learning} selects a fixed number of neighbors of a central node and sorts neighbors to concatenate their features as the grid-structured feature matrix. By choosing a series of central nodes, PATCHY-SAN constructs some matched feature matrices.
Finally, a graph-level representation is produced by the CNN model from the concatenation matrix of all built feature matrices.
In addition, Kernel Convolutional Neural Network (KCNN) \cite{nikolentzos2018kernel} sorts all vertices in a graph to form grid-structured data.
A neighborhood graph is built for each vertex and a kernel matrix is constructed by implementing the kernel function (i.e., an SPK or an MPK) between all pairwise neighborhood graphs. In this work, the grid-structured data for feeding up CNN is the kernel matrix, where each row is a vector describing the similarities between the neighborhood graph of the matched index vertex and the other neighborhood graphs.
The second branch, involves CNN-guided neural network versions of MPKs.
These methods have two main steps: aggregating neighborhood information to the central node, and using the CNN model to relabel the aggregated features.
For example, ECC \cite{simonovsky2017dynamic} concatenates 1-hop neighbor embeddings $\left(\mathbf h_u^{(k-1)}: u\in \mathcal N(v)\right)$ around the central node $v$ to construct a feature matrix by the $k$-th iteration.
Subsequently, a convolution and average operation is executed on the aggregated neighbor feature matrix to obtain a representation for the central node.
Then a graph-level representation is produced via max-pooling the node representations.
\begin{equation}
\begin{aligned}
\mathbf h_v^{(k)} = \frac{1}{|\mathcal N(v)|}& \left(\mathbf W \odot \mathbf H\right) + b^{(k)}, \quad \mathbf H = [\mathbf h_u^{(k-1)}: u\in \mathcal N(v)], \\
&\mathbf{h}_{\mathcal{G}} = \text{MaxPooling}\left(\mathbf h_v^{(K)} :v\in\mathcal V \right).
\end{aligned}
\end{equation}
Another related work, Diffusion CNN (DCNN) \cite{NIPS2016_390e9825} aggregates multi-hop neighborhood features to the central nodes through a matrix multiplication $\mathbf P \mathbf X$, where $\mathbf P = [\mathbf A, \mathbf A^2, ..., \mathbf A^h] \in \mathbb R^{h \times n\times n}$ is a three-dimensional tensor containing multi-hop (i.e., 1-, 2-, ..., h-hops) adjacent matrices and $\mathbf X \in \mathbb R^{n\times f}$ is the node features matrix. $\mathbf P \mathbf X \in \mathbb R^{h \times n\times f}$ represents the updated node features after multi-hop aggregation.
For graph classification tasks, DCNN permutes the dimensions giving $\mathbf P \mathbf X \in \mathbb R^{n \times h\times f}$ and all node representations are averaged as $\mathbf P^{*} \in \mathbb R^{h\times f}$.
Subsequently, a convolution operation is implemented on $\mathbf P^{*}$ to produce a graph-level representation. Formally,
\begin{equation}
\mathbf{h}_{\mathcal{G}} = f\left(\mathbf W \odot \mathbf P^{*}\right),
\end{equation}
where $f\left(\cdot\right)$ is a nonlinear activation function, and $\mathbf W$ is a trainable weight matrix for convolution and summation.
\subsection{CapsNet-based GL-DNNs}\label{5.3}
CapsNets \cite{hinton2011transforming} were originally designed to capture more spatial relationships between the partitions of an entity than CNNs. CapsNets are available to assemble vectorized representations of different features (e.g., colors, textures) to a capsule dealt with by a specific network.
Thus, applying a CapsNet to a graph preserves rich features and/or structure information at the graph level.
Graph Capsule Convolutional Neural Networks (GCAPS-CNN) \cite{verma2018graph} iteratively aggregates neighbor information under different statistical moments (e.g., mean, standard deviation) to form a capsule representation of the central node, formulated as:
\begin{equation}
\mathbf h_v^{(k)} = \frac{1}{\left|\mathcal N(v)\right|}
\left[\begin{array}{cc}
\left(\sum\limits_{u\in \mathcal N(v)} \mathbf h_u^{(k-1)}\right) \mathbf W_1 \text { (mean) } \\
\left(\sum\limits_{u\in \mathcal N(v)} \left(\mathbf h_u^{(k-1)}-\mu\right)^{2}\right) \mathbf W_2 \text { (std) } \\
\left( \sum\limits_{u\in \mathcal N(v)} \left(\frac{\mathbf h_u^{(k-1)}-\mu}{\sigma}\right)^{3}\right) \mathbf W_3 \text { (skewness) } \\
\vdots
\end{array}\right],
\end{equation}
where $\left(\mathbf W_1,\mathbf W_2,...\right)$ are the learnable weight matrices for mapping the aggregated features into a uniform hidden space with a dimensionality of $h$.
If the number of statistical moments is $p$ and the final iteration number is $K$, each node will be represented as $\mathbf h_v^{(K)} \in \mathbb R^{p \times h}$, and the matrix of all $n$ node embeddings will be $H^{(K)} \in \mathbb R^{n \times p \times h}$.
This approach employs a covariance function as the permutation-invariance layer to output a graph-level representation, defined as:
\begin{equation}
\mathbf{h}_{\mathcal{G}} = \frac{1}{n} (H^{(K)}-\mu)^{\top}(H^{(K)}-\mu).
\end{equation}
CapsGNN \cite{xinyi2018capsule} iteratively aggregates node features to a center node, and, in turn, adds the aggregation results of each iteration to a capsule representation of the central node.
An attention mechanism is then applied to all node capsules so as to generate a graph capsule that can be plugged into a capsule network for graph classification.
Mallea \textit{et al.} \cite{mallea2019capsule} employs the same approach as PATCHY-SAN \cite{niepert2016learning} to find substructures in graphs, while the feature matrices of searched substructures are assembled in a capsule network for graph classification.
\section{Graph-Level Graph Neural Networks (GL-GNNs)}
\begin{table*}[htbp!]
\centering
\caption{Summary of Graph-Level Graph Neural Networks (GL-GNNs).}\label{table_gnn}
\begin{tabular}{clllll}
\hline
Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{12}{*}{\begin{tabular}[c]{@{}c@{}}Message\\ Passing\\ Neural\\ Networks\end{tabular}} & 2015 & Fingerprint\cite{duvenaud2015convolutional} & NeurIPS & Python-Tensorflow & https://github.com/HIPS/neural-fingerprint \\ \cline{2-6}
& 2016 & GraphSim\cite{battaglia2016interaction} & NeurIPS & Python-Tensorflow & https://github.com/clvrai/Relation-Network-Tensorflow \\ \cline{2-6}
& 2017 & MPNN\cite{gilmer2017neural} & ICML & Python-Pytorch & https://github.com/priba/nmp\_qc \\ \cline{2-6}
& 2017 & DTNN\cite{schutt2017quantum} & NC & Python-Tensorflow & https://github.com/atomistic-machine-learning/dtnn \\ \cline{2-6}
& 2019 & GIN\cite{xu2018powerful} & ICLR & Python-Pytorch & https://github.com/weihua916/powerful-gnns \\ \cline{2-6}
& 2019 & K-GNNs\cite{morris2019weisfeiler} & AAAI & Python-Pytorch & https://github.com/chrsmrrs/k-gnn \\ \cline{2-6}
& 2019 & PPGN\cite{maron2019provably} & NeurIPS & Python-Tensorflow & https://github.com/hadarser/ProvablyPowerfulGraphNetworks \\ \cline{2-6}
& 2019 & RP\cite{murphy2019relational} & ICML & Python-Pytorch & https://github.com/PurdueMINDS/RelationalPooling \\ \cline{2-6}
& 2021 & FGNN\cite{azizian2021expressive} & ICLR & Python-Pytorch & https://github.com/mlelarge/graph\_neural\_net \\ \cline{2-6}
& 2021 & SWL\cite{bodnar2021weisfeiler_icml} & ICML & Python-Pytorch & https://github.com/twitter-research/cwn \\ \cline{2-6}
& 2021 & CWN\cite{bodnar2021weisfeiler_nips} & NeurIPS & Python-Pytorch & https://github.com/twitter-research/cwn \\ \cline{2-6}
& 2021 & RNI\cite{DBLP:conf/ijcai/AbboudCGL21} & IJCAL & - & - \\ \hline
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Subgraph\\ -Based\end{tabular}} & 2020 & GSN\cite{bouritsas2022improving} & TPAMI & Python-Pytorch & https://github.com/gbouritsas/GSN \\ \cline{2-6}
& 2020 & SubGNN\cite{AlsentzerFLZ20subgraph} & NeurIPS & Python-Pytorch & https://github.com/mims-harvard/SubGNN \\ \cline{2-6}
& 2021 & GNN-AK\cite{zhao2022from} & ICLR & Python-Pytorch & https://github.com/LingxiaoShawn/GNNAsKernel \\ \cline{2-6}
& 2021 & NGNN\cite{zhang2021nested} & NeurIPS & Python-Pytorch & https://github.com/muhanzhang/nestedgnn \\ \cline{2-6}
& 2021 & GraphSNN\cite{wijesinghe2021new} & ICLR & Python-Pytorch & https://github.com/wokas36/GraphSNN \\ \cline{2-6}
& 2021 & SUGAR\cite{sun2021sugar} & WWW & Python-Tensorflow & https://github.com/RingBDStack/SUGAR \\ \cline{2-6}
& 2021 & ESAN\cite{bevilacqua2021equivariant} & ICLR & Python-Pytorch & https://github.com/beabevi/esan \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Kernel\\ -Based\end{tabular}} & 2019 & GNTK\cite{du2019graph} & NeurIPS & Python & https://github.com/KangchengHou/gntk \\ \cline{2-6}
& 2019 & DDGK\cite{al2019ddgk} & WWW & Python-Tensorflow & \begin{tabular}[c]{@{}l@{}}https://github.com/google-research/google-research/\\ tree/master/graph\_embedding/ddgk\end{tabular} \\ \cline{2-6}
& 2020 & GCKN\cite{chen2020convolutional} & ICML & Python-Pytorch & https://github.com/claying/GCKN \\ \cline{2-6}
& 2020 & RWNN\cite{NEURIPS2020_ba95d78a} & NeurIPS & Python-Pytorch & https://github.com/giannisnik/rwgnn \\ \cline{2-6}
& 2021 & GSKN\cite{long2021theoretically} & WWW & Python-Pytorch & https://github.com/YimiAChack/GSKN \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Contrastive\\ -Based\end{tabular}} & 2020 & GraphCL\cite{NEURIPS2020_3fe23034} & NeurIPS & Python-Pytorch & https://github.com/Shen-Lab/GraphCL \\ \cline{2-6}
& 2020 & InfoGraph\cite{sun2019infograph} & ICLR & Python-Pytorch & https://github.com/fanyun-sun/InfoGraph \\ \cline{2-6}
& 2020 & GCC\cite{qiu2020gcc} & KDD & Python-Pytorch & https://github.com/THUDM/GCC \\ \cline{2-6}
& 2020 & MVGRL\cite{hassani2020contrastive} & ICML & Python-Pytorch & https://github.com/kavehhassani/mvgrl \\ \cline{2-6}
& 2021 & JOAO\cite{you2021graph} & ICML & Python-Pytorch & https://github.com/Shen-Lab/GraphCL\_Automated \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Spectral\\ -Based\end{tabular}} & 2016 & ChebNet\cite{defferrard2016convolutional} & NeurIPS & Python-Tensorflow & https://github.com/mdeff/cnn\_graph \\ \cline{2-6}
& 2021 & GNNTFS\cite{levie2021transferability} & JMLR & - & - \\ \cline{2-6}
& 2021 & GNNMatlang\cite{balcilar2021breaking} & ICML & Python-Tensorflow & https://github.com/balcilar/gnn-matlang \\ \cline{2-6}
& 2021 & ARMA\cite{bianchi2021graph} & TPAMI & Python-Pytorch & https://github.com/dmlc/dgl/tree/master/examples/pytorch/arma \\ \cline{2-6}
& 2021 & UFG\cite{zheng2021framelets} & ICML & Python-Pytorch & https://github.com/YuGuangWang/UFG \\ \hline
\end{tabular}
\end{table*}
This section focuses on GL-GNNs, which are the most influential graph-level learning techniques at present.
The cornerstone branch of GL-GNNs \textemdash Message Passing Neural Networks (MPNNs) (see Section \ref{MPNN_INTRO}) \textemdash are introduced first, followed by.
Some emerging methods in GL-GNNs, such as subgraph-based methods (see Section \ref{subgraph-based GL-GNNs}), graph kernel-based methods (see Section \ref{kernel-based GL-GNNs}), and contrastive learning-based approaches (see Section \ref{contrastive GL-GNNs}).
Notably, these emerging approaches take advantage of some of the insights from traditional graph-level learning methods.
In addition, we review progress in spectral GL-GNNs (see Section \ref{spectralgnn}), which push graph-level learning forward through spectrum properties.
Please refer to Table \ref{table_gnn} for the GL-GNNs discussed in this section.
In the end, we investigate the expressivity (see Section \ref{expressivity}), generalizability (see Section \ref{Generalizability}), and explainability (see Section \ref{Explanation-GNN}) of GL-GNNs.
\subsection{Message Passing Neural Networks (MPNNs)}\label{MPNN_INTRO}
As mentioned, researchers have developed RNN- and CNN- based versions of MPKs.
However, as the influence of deep learning has expanded, researchers have also developed various feedforward versions of Message Passing Kernels (MPKs, refer to Section \ref{MPKs_formal}).
Collectively, these are called MPNNs.
Gilmer \textit{et al.} \cite{gilmer2017neural} summarizes a collection of MPNNs \cite{battaglia2016interaction,schutt2017quantum,duvenaud2015convolutional} and further proposes a unified framework for this branch of techniques, as shown in Fig. \ref{fmpnn} (A) and denoted as:
\begin{equation}
\mathbf h_v^{(k)} = U^{(k)}\left(\mathbf h_v^{(k-1)},\sum\limits_{u \in \mathcal{N}(v)}\text{AGG}^{(k)}\left(\mathbf h_v^{(k-1)},\mathbf h_u^{(k-1)},\mathcal E_{v,u}\right)\right),
\end{equation}
where $\mathbf h_v^{(0)} = \mathbf x_v$, and $U^{(k)}\left(\cdot\right)$ is a vertex update function that outputs the embedding of the target node based on itself and its neighbors' information.
After multiple iterations, the node embeddings $\mathbf h_v^{(k)}$ learn the local structure information and the graph-level topology has distributed in all nodes.
A readout function reads all node embeddings and outputs a graph-level representation, that is:
\begin{equation}
\mathbf h_{\mathcal G} = \text{readout}\left(\mathbf h_v^{(k)}:v\in\mathcal V\right).
\end{equation}
MPNNs have become the mainstream of graph-level studies.
They are also representative of spatial-based GL-GNNs since they are easy to use through matrix operations.
Lastly, the time and memory complexity of MPNNs only grows linearly with the graph size, making this a very practical approach for large sparse graphs.
In recent years, practitioners have developed numerous enhanced versions of MPNNs, including subgraph-enhanced MPNNs (see Section \ref{subgraph-based GL-GNNs}), and kernel-enhanced MPNNs (see Section \ref{kernel-based GL-GNNs}).
\subsection{Subgraph-Based GL-GNNs}\label{subgraph-based GL-GNNs}
In recent years, investigating GL-GNNs that are capable of capturing more topological information has been a crucial stream of study. This is especially, since a number of works have uncovered structure-aware flaws in MPNNs.
To this end, practitioners have devised subgraph-based GL-GNNs, which leverage the rich structural information in subgraphs.
These subgraph-based GL-GNNs can be divided into two branches.
The first branch enhances an MPNN by injecting the subgraph information into the aggregation process, as outlined in Fig. \ref{fmpnn} (B).
The other branch borrows the graphlet idea and decomposes the graph into a few subgraphs, merging multiple subgraph embeddings to produce an embedding of the entire graph.
\subsubsection{Subgraph Enhanced MPNNs} As mentioned, MPNNs learn topological information via a neighborhood aggregation process.
However, standard MPNNs only aggregate node features, not structural information.
Therefore, a straightforward way of strengthening an MPNNs is to enrich the features of the nodes or edges with subgraph information.
Graph Substructure Network (GSN) \cite{bouritsas2022improving}, for example, counts the number of occurrences of a predefined subgraph pattern $g_1,...,g_M$ (e.g., a cycle or a triangle) that involves the target node $v$ or edge $\mathcal E_{v,u}$.
From these, subgraph feature vectors are constructed for $v$ as $\mathbf x_{v}^g$ or for $\mathcal E_{v,u}$ as $\mathbf S_{v,u}^g$, denoted as:
\begin{equation}
\begin{aligned}
&x_v^{g_m} = |\{g_s \simeq g_m: v \in \mathcal V, v \in \mathcal V_{g_s}, g_s \subseteq \mathcal G\}|,\\
&S_{v,u}^{g_m} = |\{g_s \simeq g_m: \mathcal E_{v,u} \in \mathcal E, \mathcal E_{v,u} \in \mathcal E_{g_s},g_s \subseteq \mathcal G\}|,\\
&\mathbf x_{v}^g = [x_v^{g_1},...,x_v^{g_M}]^{\top}, \quad \mathbf S_{v,u}^g = [S_{v,u}^{g_1},...,S_{v,u}^{g_M}]^{\top},
\end{aligned}
\end{equation}
where $g_m$ is a predefined subgraph pattern, and $g_s \simeq g_m$ means $g_s$ is isomorphic to $g_m$, $x_v^{g_m}$ counts the number of isomorphic subgraphs $g_s$ containing the node $v$, and $S_{v,u}^{g_m}$ indicates the number of isomorphic subgraphs $g_s$ containing the edge $\mathcal E_{v,u}$.
As a last step, the subgraph feature vectors for the node $\mathbf x_{v}^g$ and the edge $\mathbf S_{v,u}^g$ are injected into the aggregation layer, which is defined as:
\begin{equation}
\begin{aligned}
&\mathbf h_v^{(k)} = U^{(k)}\left(\mathbf h_v^{(k-1)},\mathbf m_v^{(k)}\right), \quad \mathbf m_v^{(k)} = \\
&\begin{cases}
\sum\limits_{u \in \mathcal{N}(v)}\text{AGG}^{(k)}\left(\mathbf h_v^{(k-1)},\mathbf h_u^{(k-1)},\mathbf x_{v}^g, \mathbf x_{u}^g, \mathcal E_{v,u}\right)\textbf{(Node)},\\
\sum\limits_{u \in \mathcal{N}(v)}\text{AGG}^{(k)}\left(\mathbf h_v^{(k-1)},\mathbf h_u^{(k-1)},\mathbf S_{v,u}^g, \mathcal E_{v,u}\right)\textbf{(Edge)}.
\end{cases}
\end{aligned}
\end{equation}
GSN is a promising start for subgraph-enhanced MPNNs.
However, they have one fatal drawback in that searching for and testing subgraphs for isomorphism is computationally prohibitive.
To avoid this high computational bottleneck, GNN-AK \cite{zhao2022from} samples subgraphs and swift encodes them into node embeddings.
Specifically, GNN-AK extracts the neighborhoods of each node as subgraphs (i.e., the neighborhood of node $v$ is a subgraph $g_v$), and applies a base MPNN to each neighborhood subgraph to obtain the final node embeddings, i.e.:
\begin{equation}
\begin{aligned}
&\mathbf x_v^g = \\
&[Emb\left(v|g_v\right) \mid \sum_{u\in \mathcal V \land u \neq v} Emb\left(u|g_v\right) \mid \sum_{u\in \mathcal V \land u \neq v} Emb\left(v|g_u\right)],
\end{aligned}
\end{equation}
where $Emb\left(v|g_v\right)$ is the embedding of node $v$ produced by running the base MPNN on subgraph $g_v$, $Emb\left(v|g_u\right) == 0$ if subgraph $g_u$ does not contain node $v$ (i.e., $v \not\subset \mathcal V_{g_u}$), and $\mathbf x_v^g$ is the subgraph feature of node $v$ for MPNN's aggregation.
Analogously, Nested Graph Neural Networks (NGNN) \cite{zhang2021nested} extracts nodes (i.e., $\mathcal N(v) \cup v$) and edges (i.e., $\mathcal E_{v_1,v_2} \in \mathcal E \And v_1,v_2 \in \mathcal N(v) \cup v$) in the 1-hop neighborhood of node $v$, as a neighborhood subgraph $g_v$, to be encoded by a GNN.
The subgraph $g_v$ is then encoded as the embedding $h_{g_v}$, which denotes the subgraph feature of node $v$.
One thing common to all the above methods is that they dilute or replace the node features. But such feature properties are essential for graph-level learning.
Thus, GraphSNN \cite{wijesinghe2021new} incorporates the idea of encoding the subgraph features into the edge's weight for aggregation without changing the node features.
This approach defines the formula for calculating the degree of isomorphism between two subgraphs.
The weight of $\mathcal E_{v,u}$ is equal to the degree of isomorphism between two specific subgraphs, where one of the subgraphs is the node $v$'s neighborhood subgraph, and the other subgraph is the overlap between the neighborhood subgraphs of nodes $u$ and $v$.
By normalizing the computed weights at the end, GraphSNN builds a subgraph-guided attention mechanism partaking in the MPNN's aggregation.
\subsubsection{Graphlet} In addition to empowering MPNNs through subgraph information, researchers have directly used the embeddings of subgraphs to form a graph-level representation.
SUGAR \cite{sun2021sugar}, for example, uses GNNs to embed discriminative subgraphs selected through reinforcement learning.
A readout function over all learned subgraph embeddings is then used to build a graph-level representation for classification, which can be used for classification.
Correspondingly, the graph classification results are back-propagated to train the GNNs that embed selected subgraphs.
Similarly, Subgraph Neural Networks (SubGNN) \cite{AlsentzerFLZ20subgraph} views the subgraphs of a graph as instances with independent labels.
For each instance, SubGNN samples a few finer local structures, and forms embeddings through the GNN.
A representation of each instance is generated by aggregating all the embeddings of the sampled local structures.
Another approach, Equal Subgraph Aggregation Network (ESAN) \cite{bevilacqua2021equivariant}, enhances this branch by applying two GNNs, one for learning individual embeddings for sampled subgraphs and the other for learning message passing among them. Finally, a universal set encoder \cite{qi2017pointnet} compresses all the subgraph embeddings into one graph-level representation.
\subsection{Graph Kernel-Based GL-GNNs}\label{kernel-based GL-GNNs}
Like the revival of the subgraph idea in the deep learning field, graph kernels that incorporate deep learning techniques have also attracted attention.
Similar to subgraph-based GL-GNNs, there are generally two branches of graph kernel-based GL-GNNs.
As Fig. \ref{fmpnn} (C) shows, one branch replaces the 1-hop neighbor aggregation and vertex update functions in MPNNs with a kernel function.
This group is, called the kernel-enhanced MPNNs.
In the other branch, differentiable and parameterizable kernels are designed to plug kernels into the neural networks so as to form learnable and fast deep graph kernels.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1\textwidth]{MPNNs.pdf}
\caption{Different mechanisms of MPNNs, Subgraph Enhanced MPNNs, and Kernel Enhanced MPNNs. In Subgraph Enhanced MPNN, we used 1-hop neighborhoods as the subgraph for easy understanding, but the specific subgraph extraction method is subject to the article.}
\label{fmpnn}
\end{figure*}
\subsubsection{Kernel-enhanced MPNNs} This type of method often uses a graph kernel to update the node embeddings, which in turn are used to recalculate the graph kernel.
Kernel-enhanced MPNNs break the local update of the MPNN (i.e., the node features are aggregated via adjacent neighbors) so as to capture more structure information.
For example, Graph Convolutional Kernel Networks (GCKN) \cite{chen2020convolutional} and Graph Structured Kernel Networks (GSKN) \cite{long2021theoretically} employ walk-based kernels to iteratively generate node embeddings.
Specifically, these methods generate $q$ walking sequences starting from the target node, where each sequence records all node embeddings in the walk.
As an example, in the $k$-th iteration, the one-step walk sequence $P_{i}$ from node $v$ to $u$ would be represented as $R(P_{i}) = [\mathbf h_v^{(k-1)},\mathbf h_u^{(k-1)}]^\top$.
By building a kernel function $K\left(R(P_{i}),R(P_{j})\right)$ (e.g., a random walk kernel) as the similarity measurement for any two walking sequences $P_{i}$ and $P_{j}$, GCKN and GSKN aggregate the kernel values as the updated node embeddings, that is:
\begin{equation}
h_v^{(k)} = \sum_{1 \leq i \leq q} [K\left(R(P_{1}),R(P_{i})\right), \cdots, K\left(R(P_{q}),R(P_{i})\right)]^\top.
\end{equation}
To follow up, the node embeddings updated by the graph kernel are used to obtain the kernel value in the next iteration.
Du \textit{et al.} \cite{du2019graph} combined a Neural tangent kernel (NTK) \cite{jacot2018neural} with an MPNN, summarizing the advantages of this category of approach.
Overall, the technique gives better theoretical explanations, brought about by the graph kernel, and the convex optimized tasks are easy to train.
Thus, kernel-enhanced MPNNs use a kernel function to replace the aggregation and vertex update functions in MPNNs.
The walk-based kernels do particularly well at capturing local structures to encode into the node embeddings.
\subsubsection{Deep Graph Kernel} Hand-crafted feature extraction by traditional graph kernels is far slower than parameterized automatic feature learning by neural networks.
Recently, Lei \textit{et al.} \cite{lei2017deriving} discussed deep graph kernels as parameterized learnable graph kernels for deriving neural operations.
These deep graph kernels can be optimized for specific tasks with a fast computation speeds and good interpretability.
Deep Divergence Graph Kernels (DDGK) \cite{al2019ddgk} takes $M$ base graphs $\{\mathcal G_1, \mathcal G_2, \cdots, \mathcal G_M\}$ to represent a target graph $\mathcal G_t$ as $M$-dimensional vectors $\mathbf h_{\mathcal G_t} = [K_D \left(\mathcal G_1, \mathcal G_t\right), \cdots, K_D \left(\mathcal G_M, \mathcal G_t\right)]^\top$, where $K_D \left(\mathcal G_m, \mathcal G_t\right)$ is a trainable kernel for measuring the distance between $\mathcal G_m$ and $\mathcal G_t$.
First, DDGK uses each base graph $\{\mathcal G_1, \mathcal G_2, \cdots, \mathcal G_M\}$ to train an encoder $\{\mathcal Z_1, \mathcal Z_2, \cdots, \mathcal Z_M\}$.
The encoder $\mathcal Z_m$ takes the one-hot encoding of nodes (e.g., the first node's encoding is $[1, 0, 0, \cdots]^\top$) in $\mathcal G_m$ as the input and tries to predict their neighbors (e.g., if a node only links to the second and third nodes, the correct output should be $[0, 1, 1, 0, \cdots]^\top$).
Then, the trained encoder $\mathcal Z_m$ is used by the kernel function $K_D \left(\mathcal G_m, \mathcal G_t\right)$ to predict the node's neighbors in $\mathcal G_t$, as a measurement of the distance between two graphs.
That is:
\begin{equation}
K_D \left(\mathcal G_m, \mathcal G_t\right) = \sum\limits_{v_i,v_j \in \mathcal V_t, \mathcal E_{i,j} \in \mathcal E_t} -\log\left(v_j | v_i, \mathcal Z_m\right).
\end{equation}
Random Walk graph Neural Networks (RWNN) \cite{NEURIPS2020_ba95d78a} also derives a trainable random walk kernel $K_{RW}\left(\cdot, \cdot\right)$ through a series of learnable graph patterns $\{\mathcal G_1, \mathcal G_2, \cdots, \mathcal G_M\}$.
A learnable graph $\mathcal G_m$ has a fixed node set $\mathcal V_{m}$ but a changeable edge set $\mathcal E_{m}$.
RWNN produces graph-level embeddings $\mathbf h_{\mathcal G_t} = [K_{RW} \left(\mathcal G_1, \mathcal G_t\right), \cdots, K_{RW} \left(\mathcal G_M, \mathcal G_t\right)]^\top$ of the target graph $\mathcal G_t$ for graph classification tasks.
Correspondingly, the classification results are backpropagated to change the adjacency matrix of learnable graph patterns.
That is, RWNN uses the prediction results to train the input of the kernel function (i.e., graph patterns) so that the kernel values can be learned according to the downstream task.
\subsection{Contrastive Learning-Based GL-GNNs}\label{contrastive GL-GNNs}
Contrastive learning \cite{hjelm2018learning} is a data augmentation method that creates new, plausible instances by transposing existing data without affecting the semantics.
Investigating contrastive learning for GL-GNNs is significant since GL-GNNs are data-driven models that will always encounter bottlenecks given insufficiently labeled graphs.
Graph Contrastive Learning (GraphCL) \cite{NEURIPS2020_3fe23034} defines four approaches to creating new instances as augmentation data: (1) node dropping, which randomly removes a proportion of nodes from the graph; (2) edge perturbation, which randomly adds or removes a certain percentage of edges from the graph; (3) feature masking, where some of the features of some nodes are randomly masked; and (4) subgraphs, where subgraphs are taken from the graph.
To be noticed, the newly-produced instances must be labeled as the same class as the source graph.
InfoGraph \cite{sun2019infograph}, for example, samples subgraphs $g_m$ from a source graph $\mathcal G$ as new instances.
A GL-GNN encoder $\mathcal H^{\phi}$ with some parameters $\phi$ is then used to generate graph-level representations of $g_m$ and $\mathcal G$, denoted as $\mathbf h^{\phi}_{g_m}$ and $\mathbf h^{\phi}_{\mathcal G}$.
InfoGraph's learning objective is to maximize the mutual information between $\mathbf h^{\phi}_{\mathcal G}$ and all $\mathbf h^{\phi}_{g_m}, g_m \in \mathcal G$.
This can be brought of as an evaluation of the statistical dependencies between two variables. Formally:
\begin{equation}
\mathcal H^{\phi}, \mathcal H^{\psi}=\underset{\phi, \psi}{\operatorname{argmax}} \sum_{\mathcal G \in \mathbb G} \frac{1}{|\{g_m\}|} \sum_{g_m \in \mathcal G} I_{\phi, \psi}\left(\mathbf h^{\phi}_{g_m} ; \mathbf h^{\phi}_{\mathcal G}\right),
\end{equation}
where $\mathcal H^{\psi}$ is the mutual information estimator with the parameters $\psi$, and $I_{\phi, \psi}\left(\cdot, \cdot\right)$ measures the mutual information.
Similarly, Graph Contrastive Coding (GCC) \cite{qiu2020gcc} samples subgraphs $g_1,...,g_M$ from the graph dataset $\mathbb G = \{\mathcal G_1,...,\mathcal G_N\}$ as new instances.
The embeddings of the subgraphs $g_m$ and graph $\mathcal G_n$ produced by the GL-GNN are denoted as $\mathbf h_{g_m}$ and $\mathbf h_{\mathcal G_n}$, respectively.
In the GL-GNN, a GCC employs InfoNCE loss \cite{oord2018representation} as the learning objective, that is:
\begin{equation}\label{infonce}
\mathcal L = \sum_{\mathcal G_n \in \mathbb G} - \log \frac{\sum_{g_m \in \mathcal G_n} \exp\left(\mathbf h_{\mathcal G_n}^{\top} \mathbf h_{g_m} / \tau \right)}
{\sum_{i=0}^{M} \exp\left( \mathbf h_{\mathcal G_n}^{\top} \mathbf h_{g_i} / \tau \right)},
\end{equation}
where $\tau$ is the temperature hyper-parameter. If $g_m \in \mathcal G_n$, the InfoNCE aims to maximize the similarity between $\mathbf h_{g_m}$ and $\mathbf h_{\mathcal G_n}$.
Otherwise, it separates $\mathbf h_{g_m}$ and $\mathbf h_{\mathcal G_n}$ as far away in the semantic space as possible.
Hassani \textit{et al.} \cite{hassani2020contrastive} extended graph contrastive learning to the multi-view scenario. Here, each graph view is regarded as an independent instance. This work maximizes the mutual information between a graph view and other views of the same graph.
There are several ways to generate new instances for graph contrastive learning, which raises the question of how to choose the most suitable method for the dataset one is working with.
Joint augmentation optimization (JOAO) \cite{you2021graph} was developed to address this challenge by automating the search for a proper graph data augmentation method.
JOAO trains a probability matrix that can be iteratively updated to select the optimal data augmentation approach. Its performance is competitive.
\subsection{Spectral-Based GL-GNNs} \label{spectralgnn}
Balcilar et al. \cite{balcilar2020analyzing} described spectral and spatial graph convolution in a unified way and performed spectral analysis on convolution kernels.
The results of the analysis demonstrate that a vast majority of MPNNs are low-pass filters in which only smooth graph signals are retained. Graph signals with a low-frequency profile are useful for node-level tasks on assortative networks where nodes have similar features to their neighborhoods \cite{bo2021beyond}. However, with graph-level tasks, graph signals beyond the low frequency may be critical since they can highlight the differences between different graphs \cite{balcilar2020bridging}, and, although MPNNs have been widely used, they overlook the signal frequency of graph data.
In terms of a feature $\mathbf x\in \mathbb R^n$ (a column vector of $\mathbf X\in \mathbb R^{n\times f}$) as a graph signal on a graph with $n$ nodes, spectral graph convolution performs graph signal filtering after transforming the graph signals $\mathbf x$ in spatial space into the frequency domain. According to spectral graph theory \cite{chung1997spectral}, the frequency domain generally takes the eigenvectors of the graph Laplacian $\mathbf L=\mathbf D-\mathbf A$ where $\mathbf D$ is the degree matrix (or the normalized version $\mathbf L = \mathbf I- \mathbf D^{-\frac{1}{2}}\mathbf A \mathbf D^{-\frac{1}{2}}$) of a set of space bases.
Note, though, that other bases can also be used, such as graph wavelet bases \cite{xu2019graph,hammond2011wavelets}. Specifically, $\{\lambda_1, ..., \lambda_n\}$ where $0 \leq \lambda_1 \leq ... \leq \lambda_n \leq 2$, and $\mathbf U=(\mathbf u_1, ..., \mathbf u_n)$ are the $n$ eigenvalues and $n$ orthogonal eigenvectors of $\mathbf L$, respectively. $\lambda_i$ represents the smoothness degree of $\mathbf u_i$ about $\mathbf L$. Based on the graph Fourier transformation $\hat{\mathbf x} =\mathbf U^\mathrm{T}\mathbf x$, the graph signal $\mathbf x$ is mapped to the frequency domain. And $\mathbf x=\mathbf U\hat{\mathbf x}$ is the graph Fourier inverse transformation that can restore the graph signal in spectral domain to the spatial domain. The polynomial filter is adopted by most of spectral graph convolution methods, for example, ChebNet \cite{defferrard2016convolutional} defines the spectral graph convolution as $\mathbf U \text{diag}(\Phi(\mathbf \Lambda))\mathbf U^\mathrm{T}\mathbf x$, where $\Lambda=\text{diag}(\{\lambda_i\}_{i=1}^{i=n})$, $\Phi(\mathbf \Lambda)=\sum_{k=0}^{K}\theta_k\mathbf \Lambda^{k}$ is the polynomial filtering function, $K$ are the hyper-parameters that realize the localized spectral graph convolution, and $\theta_k$ is the polynomial coefficient.
Spectral graph convolution can be task-agnostic when graph signals with any frequency profiles are filtered.
Conversely, they can also be task-specific \textemdash for example, a band-pass filter can highlight graph signals that are strongly relate to downstream tasks \cite{balcilar2020analyzing}.
However, only a few practitioners have designed graph-level neural networks from the perspective of spectral graph convolution \cite{balcilar2021breaking, bianchi2021graph}.
The main problem with applying spectral convolution in graph-level tasks is the transferability of the spectral filter coefficients from the training graph set to the unseen graphs. The spectral filters depend on the graph Laplacian decomposition, but different graph structures have different graph Laplacian decomposition results.
Most recently, Levie \textit{et al.} \cite{levie2021transferability} theoretically proved the transferability of spectral filters on multigraphs.
Balcilar \textit{et al.} \cite{balcilar2021breaking} proposed a custom filter function that could output frequency components from low to high to better distinguish between graphs.
Due to the limitation of polynomial filters in modeling sharp changes in the frequency response, Bianchi \textit{et al.} \cite{bianchi2021graph} employed an auto-regressive moving average (ARMA) filter to perform graph-level tasks. The ARMA filter is more robust to the changes or perturbations on graph structures as it does not depend on the eigen-decomposition of the graph Laplacian explicitly.
In addition, Zheng \textit{et al.} \cite{zheng2021framelets} proposed a graph convolution based on graph Framelet transforms instead of graph Fourier transform with a shrinkage activation to decompose graphs into both low-pass and high-pass frequencies. However, there is no theoretical proof of the transferability of framelet decomposition.
\subsection{The expressivity of GL-GNNs}\label{expressivity}
As the cutting-edge technology for graph-level learning, people want to explore the power of GL-GNNs for distinguishing graphs \textemdash namely, they want to investigate the expressivity of GL-GNNs.
Practitioners generally employ a representative MPK of the 1-WL algorithm \cite{weisfeiler1968reduction,shervashidze2011weisfeiler} to evaluate the expressivity of standard GL-GNNs (i.e., MPNNs) since MPNNs are the neural network versions of MPKs.
The intimate connection between GL-GNNs and 1-WL is exploited in the Graph Isomorphism Network (GIN) \cite{xu2018powerful}.
This framework shows the upper expressivity bound of an MPNN equals the 1-WL algorithm.
Several research teams have subsequently proved that MPNNs equivalent to 1-WL can not distinguish some substructures in graphs (e.g., cycles, triangles, and Circulant Skip Links) \cite{arvind2020weisfeiler,vignac2020building,murphy2019relational}.
However, these indistinguishable substructures play a significant role in learning social network and chemical compounds graphs \cite{chen2020can}.
To break the 1-WL expressivity limitation, the expressivity of GL-GNNs has been empowered through $K$-WL, convolution enhancement, and feature enrichment.
\subsubsection{$K$-WL} A complex variant of 1-WL is the $K$-WL algorithm, which identifies more substructures in graphs by relabeling a set of $K$ vertices.
Morris \textit{et al.} \cite{morris2019weisfeiler} employed MPNNs dealing with $K$-dimensional tensors to apply $K$-WL by neural networks, that is $K$-GNN.
$K$-GNN achieved the expressivity approximately near but slightly weaker than the $K$-WL, but its computational cost increases exponentially with $K$ since it needs to calculate $K$-ranked tensors.
To avoid processing high-dimensional tensors, Provably Powerful Graph Networks (PPGN) \cite{maron2019provably} adopts a variant of the 2-WL algorithm (i.e., 2-FWL \cite{cai1992optimal}) for designing GL-GNNs and achieves the expressivity over 3-WL.
Further, PPGN replaces the relabel function in 2-FWL with a matrix multiplication based on a single quadratic operation.
Similarly, Folklore Graph Neural Networks (FGNN) \cite{azizian2021expressive} implements 2-FWL through matrix operations on tensors, pursuing the expressive power as 3-WL.
Despite these common efforts on $K$-WL equivalence GL-GNNs, the majority of them theoretically exceed 1-WL but do not empirically exceed 1-WL \cite{dwivedi2020benchmarking}.
This weak performance by $K$-WL equivalent GL-GNNs is due to two main reasons which are explained next.
\subsubsection{Convolution Enhancement} One reason for the failure of the $K$-WL approaches is that they break the local updates of MPNNs \cite{battaglia2018relational}, i.e., they no longer update vertices based on neighborhood information.
In practice, GL-GNNs require local updates to preserve the inductive bias property of the graph convolutions \cite{battaglia2018relational}.
Therefore, some researchers have explored more powerful GL-GNNs by upscaling the graph convolutions yet preserving the local updates.
Alon and Yahav \cite{alon2021on} noticed that the majority of GL-GNNs do not capture the long-range interactions between nodes because the number of convolutional layers is limited by over-smoothing issues \textemdash that is, the node embeddings tend to be similar after multiple aggregations.
However, long-range interactions can influence the discriminativeness of graphs.
For example, methylnonane is identified by the atoms posited in the compound's two end sides.
To address this issue, these researchers appended a fully linked adjacency matrix to the convolutional layer which aggregates the long-range information without violating any local updates.
Another powerful tool for enhancing the convolution layer is the matrix query language (MATLANG) \cite{brijder2019expressive,geerts2021expressive}.
MATLANG strengthens a GL-GNN so that it can recognize more special substructures through its matrix operations, thereby reaching 3-WL expressivity.
Inspired by this work, Balcilar \textit{et al.} \cite{balcilar2021breaking} added MATLANG to the convolutional layer, while Greets and Reutter \cite{geerts2022expressiveness} evaluated the expressiveness of GL-GNNs through MATLANG instead of the 1-WL algorithm.
\subsubsection{Feature Enrichment} Another reason that $K$-WL methods outperform 1-WL in theory but do not achieve superior performance in experiments is that they ignore the role of node features.
As complementary information for graph structures, node features allow almost all graphs to be discriminated by 1-WL GL-GNNs.
Some practitioners have emphasized that considering node features can also improve the expressiveness of GL-GNNs, rather than just focusing on graph structures.
Murphy \textit{et al.} \cite{murphy2019relational} annotated a unique position descriptor for each node, that is, sorting all nodes.
Adopting these position descriptors as node features can help a 1-WL GL-GNN to better handle featureless graphs and identify more structures.
To maintain permutation-invariant of graph-level learning, all permutations of node order should be enumerated and the average results should be taken.
Similarly, Colored Local Iterative Procedure (CLIP) \cite{10.5555/3491440.3491734} sorts the nodes in a substructure and gives them a local position descriptor for feature enrichment.
In addition, both Sato \textit{et al.} \cite{sato2020survey} and Abboud \textit{et al.} \cite{DBLP:conf/ijcai/AbboudCGL21} insert random features into nodes giving rise to stable and powerful GL-GNNs.
\subsubsection{High-order Neural Networks} Recently, researchers have tried to improve the expressivity of GL-GNNs through algebraic topology.
This is because equipping graphs as a geometric structure can preserve more valuable properties. Cellular GL-GNNs \cite{bodnar2021weisfeiler_icml,bodnar2021weisfeiler_nips} perform MPNNs on cell complexes, an object including hierarchical structures (e.g., vertices, edges, triangles, tetrahedra). By replacing graphs with cell complexes, cellular GL-GNNs benefits from the better computational fabric for larger expressivity. Furthermore, sheaf neural networks \cite{bodnar2022neural,barbero2022sheaf} decorate a graph with a geometrical structure, sheaf, which constructs vector space for each node and edge and applies linear transformations among these spaces.
A correct sheaf setting will allow an MPNN to pass messages along a richer structure.
Thus, linearly separate embeddings can be learned, which will enhance the expressive power of GL-GNNs.
\subsection{The generalizability of GL-GNNs}\label{Generalizability}
Real-world applications with graph data tend to involve complex scenarios, such as needing to train a model with only a small amount of labeled data that can ultimately perform well with a large-scale unlabeled test (i.e., size shift) or using only a few labeled training graphs to fit the bulk of unlabeled test graphs. The ability to generalize GL-GNNs is hence a crucial aspect of dealing with these challenges.
\subsubsection{Size Generalization}
Sinha \textit{et al.} \cite{sinha2020evaluating} stress the importance of generalizing GL-GNNs and presented evaluation criteria for this.
Xu \textit{et al.} \cite{xu2021how} theoretically explain that GL-GNNs have better size generalization capabilities than MLPs and can extrapolate trained models to test data that is different from the training set.
To this end, they presented a trick for MPNNs where the graph's vertices are updated by minimizing the aggregated information instead of through summation.
This trick improves generalization ability by altering the learning process from one that is non-linear to one that is linear.
Yehudai \textit{et al.} \cite{yehudai2021local} theoretically and empirically found the generalization ability of GL-GNNs as the discrepancies in substructures between large and small graphs grows.
To solve this problem, they forced the GL-GNN to pay more attention to the substructures that are hidden in large unlabeled graphs but rarely appear in small labeled graphs.
SizeShiftReg \cite{buffelli2022sizeshiftreg} constrains GNNs to be robust to size-shift through a regularization approach.
SizeShiftReg coarsens the input graph and minimizes the discrepancy between the distribution of the original and coarsened graph embeddings.
\subsubsection{Few-shot Learning} In considering few-shot learning scenarios, Ma \textit{et al.} \cite{ma2020adaptive} found that there are also differences in the substructures between a few labeled graphs and a large number of unlabeled graphs.
This is because a statistical sample of the training data is too small to represent the substructural distributions of the whole dataset.
Thus, they paid more attention to capturing substructures in unseen unlabeled graphs.
Chauhan \textit{et al.} \cite{Chauhan2020FEW-SHOT} clusters graphs based on their spectral properties, to produce super-class graphs.
Graph-level representations can be learned from super-class graphs as they have excellent generalization.
\subsection{The explainability of GL-GNNs}\label{Explanation-GNN}
The black-box nature of deep neural networks limits the applicability of GL-GNNs to situations where trust is not an absolutely crucial requirement.
Making GL-GNNs explain their predictions in a way that is more interpretable to humans is therefore of great significance to extending the research of GL-GNNs.
Studies on GL-GNNs need to shed insights into how they handle node features and topologies when it comes to predictions.
They also need to more clearly demonstrate how the models identify significant subgraphs and features.
Methods to explain GL-GNNs can be roughly divided into two categories.
One group involves methods that explain the prediction of each input graph;
the other group of methods captures common patterns in the predictions of a set of graphs as explanations.
\subsubsection{A Single Graph} There are three ways to understand GL-GNNs predictions based on a single graph: they can be perturbation-based, model-proxy-based, or gradient-based.
Perturbation-based methods mask nodes, edges, or substructures in the input graph to generate new predictions.
These are then compared to the original input prediction to highlight the important features or structures influencing the GL-GNNs.
For example, GNNExplainer \cite{ying2019gnnexplainer} masks nodes and edges by changing the feature and adjacency matrices, to form masked graphs.
An input graph and its masked graphs are predicted by a trained GL-GNN, while GNNExplainer aims to find the masked graphs with maximized mutual information between its' prediction and the input's prediction.
This found masked graph is the one that preserves the most significant substructures to the GL-GNN's given prediction.
Alternatively, SubgraphX \cite{yuan2021explainability} samples a group of nodes' neighborhoods as subgraphs.
A trained GL-GNN is then used to compute Shapley values \cite{kuhn1953contributions} for all the sampled subgraphs. These values represent each subgraph's contribution to the GL-GNN's prediction.
PGExplainer \cite{luo2020parameterized} trains an MLP to determine which edges are valuable to a GNN's prediction and then removes any irrelevant edges to form a new graph. Subsequently, the original and the newly-formed graph are fed into a trained GL-GNN so as to optimize an MLP by maximizing the mutual information between their predictions.
Model-proxy-based methods utilize a simpler surrogate model to approximate the predictions of GL-GNNs.
PGM-Explainer \cite{vu2020pgm} adopts an explainable Bayesian network \cite{pearl1988markov} to calculate the relationship dependencies between nodes, so as to generate a probability graph that describes the input graph.
Gradient-based approaches measure the importance of different input features by back-propagating the gradients of the neural networks.
Gradient-weighted Class Activation Mapping (Grad-CAM) \cite{pope2019explainability}, for example, takes the gradient value of each node embedding in a graph classification task as a measurement of the nodes' significance to the GL-GNN's prediction.
Grad-CAM then measures this subgraph's importance to the prediction by averaging the gradient values of all node embeddings from the subgraph.
\subsubsection{A Set of Graphs} What is common to all the above methods is that they can only learn independent explanations for each instance of a graph \cite{ying2019gnnexplainer,yuan2020explainability}.
However, often the predictions of GL-GNNs made by GL-GNNs are based on a set of graphs.
Thus, understanding the rules or graph patterns that a GL-GNN mine from a set of graphs can provide high-level and generic insights into the explainability of GL-GNNs.
XGNN \cite{yuan2020xgnn} employs a reinforcement learning guided graph generator that generates a graph pattern for different graphs in the same class.
The graph generator is trained via policy gradient to maximize the certain label prediction \cite{sutton1999policy}.
\section{Graph Pooling}\label{Pooling}
\begin{table*}[]
\centering
\caption{Summary of Graph Pooling.}\label{table_graphpooling}
\begin{tabular}{clllll}
\hline
Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Global-\\ Numeric\end{tabular}} & 2020 & PNA\cite{NEURIPS2020_99cad265} & NeurIPS & Python-Pytorch & https://github.com/lukecavabarrett/pna \\ \cline{2-6}
& 2020 & TextING\cite{zhang2020every} & ACL & Python-Tensorflow & https://github.com/CRIPAC-DIG/TextING \\ \cline{2-6}
& 2020 & SOPOOL\cite{wang2020second} & TPAMI & Python-Pytorch & https://github.com/divelab/sopool \\ \hline
\begin{tabular}[c]{@{}c@{}}Global-\\ Attention\end{tabular} & \multicolumn{1}{c}{2016} & Set2Set\cite{set2set} & ICLR & Python-Pytorch & https://github.com/pyg-team/pytorch\_geometric \\ \hline
\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Global-\\ CNN\end{tabular}} & 2016 & PATCHYSAN\cite{niepert2016learning} & ICML & Python & https://github.com/tvayer/PSCN \\ \cline{2-6}
& 2018 & KCNN\cite{nikolentzos2018kernel} & ICANN & Python-Pytorch & \begin{tabular}[c]{@{}l@{}}https://github.com/giannisnik/\\ cnn-graph-classification\end{tabular} \\ \hline
\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Global-\\ Top K\end{tabular}} & 2018 & SortPool\cite{zhang2018end} & AAAI & Python-Pytorch & https://github.com/muhanzhang/pytorch\_DGCNN \\ \cline{2-6}
& 2020 & GSAPool\cite{zhang2020structure} & WWW & Python-Pytorch & https://github.com/psp3dcg/gsapool \\ \hline
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Hierarchical-\\ Clustering\end{tabular}} & 2014 & DLCN\cite{ICLR2014BRUNA} & ICLR & - & - \\ \cline{2-6}
& 2015 & GraphCNN\cite{henaff2015deep} & Arxiv & Python-Tensorflow & https://github.com/mdeff/cnn\_graph \\ \cline{2-6}
& 2018 & DiffPool\cite{ying2018hierarchical} & NeurIPS & Python-Pytorch & https://github.com/RexYing/diffpool \\ \cline{2-6}
& 2019 & EigenPool\cite{ma2019graph} & KDD & Python-Pytorch & https://github.com/alge24/eigenpooling \\ \cline{2-6}
& 2020 & StructPool\cite{yuan2020structpool} & ICLR & Python-Pytorch & https://github.com/Nate1874/StructPool \\ \cline{2-6}
& 2020 & MinCutPool\cite{bianchi2020spectral} & ICML & Python-Tensorflow & \begin{tabular}[c]{@{}l@{}}https://github.com/FilippoMB/Spectral-Clustering\\ -with-Graph-Neural-Networks-for-Graph-Pooling\end{tabular} \\ \cline{2-6}
& 2020 & GMT\cite{baek2020accurate} & ICLR & Python-Pytorch & https://github.com/JinheonBaek/GMT \\ \hline
\multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}Hierarchical-\\ Top K\end{tabular}} & 2018 & SHGC\cite{cangea2018towards} & Arxiv & Python-Tensorflow & https://github.com/HeapHop30/hierarchical-pooling \\ \cline{2-6}
& 2019 & U-Nets\cite{gao2019graph} & ICML & Python-Pytorch & https://github.com/HongyangGao/Graph-U-Nets \\ \cline{2-6}
& 2019 & SAGPool\cite{lee2019self} & ICML & Python-Pytorch & https://github.com/inyeoplee77/SAGPool \\ \cline{2-6}
& 2020 & ASAP\cite{ranjan2020asap} & AAAI & Python-Pytorch & https://github.com/malllabiisc/ASAP \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Hierarchical-\\ Tree\end{tabular}} & 2017 & MoNet\cite{monti2017geometric} & CVPR & Python-Pytorch & https://github.com/dmlc/dgl/tree/master/examples/mxnet/monet \\ \cline{2-6}
& 2019 & EdgePool\cite{diehl2019edge} & Arxiv & Python-Pytorch & \begin{tabular}[c]{@{}l@{}}https://github.com/pyg-team/pytorch\_geometric\\ /blob/master/torch\_geometric/nn/pool/edge\_pool.py\end{tabular} \\ \cline{2-6}
& 2022 & HRN\cite{wu2021structural} & IJCAL & Python & https://github.com/Wu-Junran/HierarchicalReporting \\ \hline
\end{tabular}
\end{table*}
Generally, deep graph-level learning methods encode graphs based on node representations. Graph pooling is a technique that integrates node embeddings into a graph embedding. In this section, we introduce two mainstream types of graph pooling techniques, i.e., global and hierarchical graph pooling (see Section \ref{global} and \ref{hierarchical}).
Moreover, we summarize some recent investigations about the efficacy of this newly-emerging technique (see Section \ref{effectiveness}).
Table \ref{table_graphpooling} summarizes the graph pooling approaches introduced in this section.
There are four different types of graph pooling \textemdash numeric operation, attention-based, CNN-based, and global top-$K$ \textemdash all of which aggregate all node embeddings at once to build a graph-level representation.
\subsubsection{Numeric Operation} Adopting a simple numeric operation for all node embeddings is a common graph pooling method \cite{xu2018powerful,duvenaud2015convolutional}, since it is easy to use and obeys the permutation invariant.
An illustration of a type of numeric operation (i.e., a summation) for all node embeddings is shown in Fig. \ref{fglobalpool} (A).
It is common to see practitioners aggregating node embeddings via summation, maximization, minimization, mean, and concatenation functions. For example:
\begin{equation}
\mathbf h_{\mathcal G} = \sum_{v \in \mathcal V} \mathbf h_v \mathopen\biggg / \mathop{\text{max/min}}\limits_{v \in \mathcal V} \left(\mathbf h_v\right) \mathopen\biggg / \frac{1}{|\mathcal V|}\sum_{v \in \mathcal V} \mathbf h_v \mathopen\biggg / [\mathbf h_{v_1}|...|\mathbf h_{v_{|\mathcal V|}}].
\end{equation}
Duvenaud \textit{et al.} \cite{duvenaud2015convolutional} empirically proved that, in graph-level learning, summation has no weaker an outcome than a hash function.
Similarly, GIN \cite{xu2018powerful} shows us that the injective relabeling function in the WL algorithm can be replaced with a simple numeric operation.
Further, GIN also allows us to analyze the efficacy of different functions: summation, maximization, and mean functions.
Summation comprehensively summarizes the full features and structure of a graph.
Maximization emphasizes significant node embeddings, and mean learns the distribution of labels.
Inspired by GIN, Principal Neighbourhood Aggregation (PNA) \cite{NEURIPS2020_99cad265} employs all three of these functions to pool the node embeddings, while TextING \cite{zhang2020every} includes both mean and maximization pooling to capture the label distribution and strengthen the keyword features.
A few variants of graph pooling have also been developed.
For example, Deep Tensor Neural Network (DTNN) \cite{schutt2017quantum} applies a neural layer that processes the node embeddings before the summation function and second-order pooling (SOPOOL) \cite{wang2020second} is executed as $\mathbf h_{\mathcal G} = [\mathbf h^T_{v_1}\mathbf h_{v_1}|...|\mathbf h^T_{v_{|\mathcal V|}}\mathbf h_{v_{|\mathcal V|}}]$.
\subsubsection{Attention-based} The contributions of node embeddings to graph-level representations may not be equal, as some of them contain may more important information than others.
Hence, some researchers have tried using an attention mechanism to aggregate the node embeddings based on their particular contribution, as outlined in Fig. \ref{fglobalpool} (C).
Li \textit{et al.} \cite{li2016gated} and Duvenaud \textit{et al.} \cite{duvenaud2015convolutional}, for example, both employ a softmax function as an attention-based global pooling for aggregation.
This can be written as:
\begin{equation}
\mathbf h_{\mathcal G} = \sum_{v,k} \text{softmax} \left(w_v^k, \mathbf h_v^k\right),
\end{equation}
where $w_v^k$ is a trainable weight for the embedding $h_v^k$ of node $v$ in iteration $k$.
Note that $w_v^k$ will be large if $h_v^k$ is important to the downstream task.
Set2Set \cite{set2set} is a more complicated attention-based graph pooling model. It learns the attention coefficients of all node embeddings from an ordered sequence generated by LSTM. Although Set2Set handles sequential node embeddings, the order of nodes is determined by an LSTM model without affecting permutation invariance.
\subsubsection{CNN-based} PATCHY-SAN \cite{niepert2016learning} and KCNN \cite{nikolentzos2018kernel} are based on the idea of ordering vertices and applying a 1-D convolutional layer to pool the ordered vertices features.
These two models are permutation invariant because they order vertices according to certain rules regardless of the input order.
\subsubsection{Global Top-$K$} Global top-$K$ graph pooling sorts all nodes and selects the first $K$ node embeddings for aggregation, as shown in Fig. \ref{fglobalpool} (B).
In this way, the pooling layer only preserves $K$ significant vertices and drops out others.
SortPool \cite{zhang2018end} employs graph convolution operations to project each node into a one-dimensional vector as the ranking score for selecting the $K$ vertices with the highest scores.
Subsequently, a GL-GNN is used to produce the node embeddings of the selected $K$ nodes, which come together to form the graph-level representation.
Graph Self-Adaptive Pooling (GSAPool) \cite{zhang2020structure} is another global top-$K$ graph pooling model that ranks nodes based on the summing of feature and structure scores.
The node structure scores are 1-dimensional vectors projected by the graph convolution operations as same as SortPool, while the feature scores are learned by feeding the node features into an MLP.
\subsection{Hierarchical Graph Pooling}\label{hierarchical}
\subsection{Global Graph Pooling}\label{global}
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1\textwidth]{PoolingGlobal.pdf}
\caption{Toy examples of Global Pooling methods.}
\label{fglobalpool}
\end{figure*}
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1\textwidth]{PoolingHierarchical.pdf}
\caption{Toy examples of Hierarchical Pooling methods.}
\label{fhierpool}
\end{figure*}
Global graph pooling ignores the hierarchical structures in graphs.
The evolution of a graph is to collect nodes into hierarchical structures (e.g., communities), then to form the graph.
Hence, researchers tend to capture hierarchical information through an aggregation process that has multiple parses, which coarsens the graph each time.
We have divided hierarchical graph pooling techniques into three branches: clustering-based, hierarchical top-$K$, and tree-based.
\subsubsection{Clustering-based} Clustering methods were originally designed to capture the hidden hierarchical structures in graphs, but these techniques can be incorporated into the pooling process.
Fig. \ref{fhierpool} (A) demonstrates clustering-based graph pooling, which has been the focus of many studies.
For instance, Henaff \textit{et al.} \cite{henaff2015deep} implemented multi-resolution spectral clustering \cite{von2007tutorial} which assigns each node to a matched cluster.
Subsequently, the clusters in the input graph are treated as new nodes of the new coarsened graph.
The embedding of the new node is obtained by averaging all node embeddings in the cluster.
This coarsening process is iterative and operates until only one or very few vertices remain in the most recent coarsened graph.
Similarly, Bruna \textit{et al.} \cite{ICLR2014BRUNA} adopted hierarchical agglomerative clustering \cite{hastie2009elements} to coarsen graphs, while StructPool \cite{yuan2020structpool} employs conditional random fields \cite{lafferty2001conditional} to cluster each node by considering the assignments of other vertices.
However, clustering-based graph pooling cannot optimize the clustering process for downstream tasks given just any old off-the-shelf clustering method.
Rather, the clustering method must be designed to consider downstream tasks.
For example, Graph Multiset Transformer (GMT) \cite{baek2020accurate} uses a multi-head self-attention mechanism to cluster nodes into different sets according to the final task and a graph-level representation is therefore derived through these sets.
MinCutPool \cite{bianchi2020spectral} assigns each node to a cluster via an MLP, which is optimized by two goals: first that the clusters are similar in size, and, second, that the clusters' embeddings are separable.
Finally, the graph-level representation is obtained by pooling the substructure-level embeddings.
EigenPool \cite{ma2019graph} involves a spectral clustering method that coarsens graphs and pools node embeddings into cluster-level embeddings by converting spectral-domain signals.
These clustering-based algorithms assume that each node belongs to a certain cluster, yet DiffPool \cite{ying2018hierarchical} assigns each node to multiple clusters through a trainable soft assignment matrix $\mathbf S^{(k)} \in \mathbb R^{n^{(k)} \times n^{(k+1)}}$, where $n^{(k)}$ is the number of vertices in the input graph at the $k$-th layer, and $n^{(k+1)}$ represents the cluster's number in the input graph or the node's number in the coarsened graph.
To be specific, at the $k$-th layer, each row of $\mathbf S^{(k)}$ corresponds to a node in the input graph, and each column of $\mathbf S^{(k)}$ corresponds to a new node in the coarsened graph (i.e., a cluster in the input graph).
The assignment matrix $\mathbf S^{(k)}$ is trained by a graph convolutional layer, which is defined as:
\begin{equation}
\mathbf S^{(k)} = \text{softmax} \left(\text{Conv}^{(k)} \left(\mathbf A^{(k)},\mathbf H^{(k)},\mathbf W^{(k)}\right) \right),
\end{equation}
where $\mathbf A^{(k)} \in \mathbb R^{n^{(k)} \times n^{(k)}}$ and $\mathbf H^{(k)} \in \mathbb R^{n^{(k)} \times f}$ are the adjacent matrix and node embedding matrix of the input graph at the $k$-th layer, respectively. $\mathbf W^{(k)} \in \mathbb R^{f \times n^{(k+1)}} $ is the trainable weight matrix, and $\text{softmax}\left(\cdot\right)$ function is applied to each row.
\subsubsection{Hierarchical Top-$k$} The high complexity of the clustering process exacerbates the computational cost burden of cluster-based hierarchical graph pooling.
For example, the DiffPool \cite{ying2018hierarchical} is extremely costly in terms of time and memory because the assignment matrices need to be trained.
So, to speed up the process of hierarchical graph pooling, researchers have looked to replace the clustering process with a scheme that coarsens the graph according to the top-$K$ idea, as shown in Fig. \ref{fhierpool} (B).
Graph U-nets \cite{gao2019graph}, for example, projects each node feature into a 1-dimensional vector $\mathbf Y$, as the rank score. Subsequently, the $K$ nodes with the highest score are selected to form the new coarsened graph, which is defined as:
\begin{equation}
\begin{aligned}
\mathbf Y &= \frac{\mathbf Z^{(l)}\mathbf P^{(l)}}{||\mathbf P^{(l)}||}, &\text{idx} = \text{Top}\ K\left(\mathbf Y\right),\\
\mathbf Z^{(l+1)} &= \left(\mathbf Z^{(l)} \odot \left(\text{sigmoid}\left(\mathbf Y_{\text{idx}} \right) \mathbf 1^\top_Z\right) \right), &\mathbf A^{(l+1)} = \mathbf A^{(l)}_{\text{idx},\text{idx}},
\end{aligned}
\end{equation}
where $Z^{(l)}$ is the input node features at the layer $l$, $\mathbf P^{(l)}$ is a learnable projection matrix, $\text{Top}\ K\left(\cdot \right)$ is a function that returns the index of the top-$K$ nodes, and all the elements are 1 in the vector $\mathbf 1_Z$ which has the same dimension as the node feature.
Cangea \textit{et al.} \cite{cangea2018towards} employed the Graph U-nets to coarsen graphs and concatenated the mean and maximum values of node embeddings on the coarsened graphs as graph-level representations.
Further, SAGPool \cite{lee2019self} chooses the top-$K$ nodes to generate the coarsened graph by adopting a graph convolution operation to project node features as scores.
All these methods generate a coarsened graph by preserving the top-$K$ nodes.
However, Ranjan \textit{et al.} \cite{ranjan2020asap} presented the novel idea of ranking the clusters and preserving on the top-$K$ of them.
The clusters were ranked by employing a self-attention algorithm called Master2Token \cite{shen2018disan} that scores each cluster based on the node embeddings within it.
\subsubsection{Tree-based} Tree-based hierarchical graph pooling implements the coarsening process via an encoding tree, where the input graph is coarsened layer by layer to the ultimate node from the leaf layer to the root layer, as shown in Fig. \ref{fhierpool} (C).
ChebNet \cite{defferrard2016convolutional} and MoNet \cite{monti2017geometric} use the Graclus \cite{dhillon2007weighted} algorithm to pair nodes in the graph based on the graph spectrum and merge the pair-wise nodes as a new node in the coarsened graph.
That is to say, these two methods build a balanced binary tree to coarsen the graph, and each father node on the tree is obtained by coarsening its two child nodes.
Wu \textit{et al.} \cite{wu2021structural} uses a structure encoding tree \cite{li2016structural} for tree-based hierarchical graph pooling.
Structural coding trees compress the hierarchy of a graph into a tree. Here, the leaves are the nodes, the root represents the whole graph, and the other non-leaf nodes are the hierarchical structures (e.g., the communities).
An MLP merges the features of the child nodes in the structure encoding tree, to generate an embedding of the father node.
The result is an embedding of the root node, which serves as a graph-level representation.
Alternatively, EdgePool \cite{diehl2019edge} scores edges based on the features of the nodes the edges link, eliminating the highest ranked edge by merging its two end nodes.
The features of the newly generated node, which maintains all the neighbors of the original two nodes, are obtained by summing the features of the two merged nodes.
EdgePool falls into the category of being a tree-based hierarchical graph pooling method because it merges two child nodes in a tree into a father node.
\subsection{The effectivity of Graph Pooling}\label{effectiveness}
As a downstream summarization component of GNNs, graph pooling has attracted a surge of research interest.
However, since graph pooling is so new, much work is required to investigate the effectiveness of all the various graph pooling algorithms.
Mesquita \textit{et al.} \cite{mesquita2020rethinking} conducted controlled experiments to empirically evaluate the effectiveness of clustering-based hierarchical graph pooling.
First, they adopted two opposite strategies for guiding some clustering-based hierarchical graph pooling processes \textemdash specifically, clustering each of non-adjacent and adjacent nodes.
The final results not only show that the two strategies are comparable, they also indicate that off-the-shelf clustering algorithms, which tend to cluster adjacent nodes, fail to improve graph pooling.
As part of the experiments, Mesquita and colleagues also replaced the learnable assignment matrix in DiffPool \cite{ying2018hierarchical} with an immutable probability assignment matrix: uniform, normal and Bernoulli distributions were selected.
The experimental results verify that the performance of fixed-probability-assignment-matrix-guided graph pooling is not weaker than that of DiffPool.
Overall, they concluded that the current clustering-based hierarchical pooling may not be particularly effective and matched this will a call for more sanity checks and ablation studies of the current graph pooling algorithms to fairly evaluate their contributions.
Another study on Pooled Architecture Search (PAS) \cite{wei2021pooling} was dedicated to investigating the effectiveness of graph pooling \textemdash this time with different datasets.
The results of the study show that the effectiveness of graph pooling algorithms is data-specific, that is to say, different input data needs to be handled by a suitable graph pooling algorithm.
For this reason, PAS includes a differentiable search method to select the most appropriate graph pooling algorithm for the given input data.
\section{Benchmarks}
\begin{table*}[]
\centering
\caption{Summary of Selected Benchmark Datasets}\label{table_data}
\begin{threeparttable}
\resizebox{\linewidth}{!}{
\begin{tabular}{llllllllcl}
\hline
Category & Dataset & Size & \#Graphs & \begin{tabular}[c]{@{}l@{}}Average\\ \#Nodes\end{tabular} & \begin{tabular}[c]{@{}l@{}}Average\\ \#Edges\end{tabular} & \begin{tabular}[c]{@{}l@{}}Node\\ Attr.\end{tabular} & \begin{tabular}[c]{@{}l@{}}Edge\\ Attr.\end{tabular} & \#Classes & Source \\ \hline
\multirow{6}{*}{Biology} & ENZYMES & Small & 600 & 32.6 & 62.1 & \checkmark & - & 6 & \cite{borgwardt2005protein,morris2020tudataset} \\
& PROTEINS & Small & 1113 & 39.1 & 72.8 & \checkmark & - & 2 & \cite{borgwardt2005protein,morris2020tudataset} \\
& D\&D & Small & 1178 & 284.3 & 715.7 & \checkmark & - & 2 & \cite{dobson2003distinguishing,morris2020tudataset} \\
& BACE & Small & 1513 & 34.1 & 36.9 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,subramanian2016computational} \\
& MUV & Medium & 93087 & 24.2 & 26.3 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,rohrer2009maximum} \\
& ppa & Medium & 158100 & 243.4 & 2266.1 & - & \checkmark & 37 & \cite{zitnik2019evolution,hu2020open} \\ \hline
\multirow{14}{*}{Chemistry} & MUTAG & Small & 188 & 17.9 & 19.8 & \checkmark & \checkmark & 2 & \cite{10.5555/3042573.3042614,morris2020tudataset} \\
& SIDER & Small & 1427 & 33.6 & 35.4 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,altae2017low} \\
& ClinTox & Small & 1477 & 26.2 & 27.9 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,gayvert2016data} \\
& BBBP & Small & 2039 & 24.1 & 26.0 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,martins2012bayesian} \\
& Tox21 & Small & 7831 & 18.6 & 19.3 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,challengetox21} \\
& ToxCast & Small & 8576 & 18.8 & 19.3 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,richard2016toxcast} \\
& molhiv & Small & 41127 & 25.5 & 27.5 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,hu2020open} \\
& molpcba & Medium & 437929 & 26.0 & 28.1 & \checkmark & \checkmark & 2 & \cite{wu2018moleculenet,hu2020open} \\
& FreeSolv & Small & 642 & 8.7 & 8.4 & \checkmark & \checkmark & - & \cite{wu2018moleculenet,mobley2014freesolv} \\
& ESOL & Small & 1128 & 13.3 & 13.7 & \checkmark & \checkmark & - & \cite{wu2018moleculenet,delaney2004esol} \\
& Lipophilicity & Small & 4200 & 27.0 & 29.5 & \checkmark & \checkmark & - & \cite{wu2018moleculenet,wenlock2015experimental} \\
& AQSOL & Small & 9823 & 17.6 & 35.8 & \checkmark & \checkmark & - & \cite{dwivedi2020benchmarking,sorkun2019aqsoldb} \\
& ZINC & Small & 12000 & 23.2 & 49.8 & \checkmark & \checkmark & - & \cite{dwivedi2020benchmarking,irwin2012zinc} \\
& QM9 & Medium & 129433 & 18.0 & 18.6 & \checkmark & \checkmark & - & \cite{wu2018moleculenet,morris2020tudataset} \\ \hline
\multirow{7}{*}{\begin{tabular}[c]{@{}l@{}}Social \\ Networks\end{tabular}} & IMDB-BINARY & Small & 1000 & 19.8 & 96.5 & - & - & 2 & \cite{Yanardag2015kdd,morris2020tudataset} \\
& IMDB-MULTI & Small & 1500 & 13.0 & 65.9 & - & - & 3 & \cite{Yanardag2015kdd,morris2020tudataset} \\
& DBLP\_v1 & Small & 19456 & 10.5 & 19.7 & \checkmark & \checkmark & 2 & \cite{morris2020tudataset} \\
& COLLAB & Medium & 5000 & 74.5 & 2457.8 & - & - & 3 & \cite{Yanardag2015kdd,morris2020tudataset} \\
& REDDIT-BINARY & Small & 2000 & 429.6 & 497.8 & - & - & 2 & \cite{Yanardag2015kdd,morris2020tudataset} \\
& REDDIT-MULTI-5K & Medium & 4999 & 508.5 & 594.9 & - & - & 5 & \cite{Yanardag2015kdd,morris2020tudataset} \\
& REDDIT-MULTI-12K & Medium & 11929 & 11.0 & 391.4 & - & - & 11 & \cite{Yanardag2015kdd,morris2020tudataset} \\ \hline
\multirow{4}{*}{\begin{tabular}[c]{@{}l@{}}Computer\\ Science\end{tabular}} & CIFAR10 & Medium & 60000 & 117.63 & 941.1 & \checkmark & - & 10 & \cite{dwivedi2020benchmarking,krizhevsky2009learning} \\
& MNIST & Medium & 70000 & 70.57 & 564.53 & \checkmark & - & 10 & \cite{dwivedi2020benchmarking,lecun1998mnist} \\
& code2 & Medium & 452741 & 125.2 & 124.2 & \checkmark & \checkmark & - & \cite{husain2019codesearchnet,hu2020open} \\
& MALNET & Large & 1262024 & 15378 & 35167 & - & - & 696 & \cite{freitas2021a} \\ \hline
\end{tabular}}
\vspace{+1mm}
\begin{tablenotes}
\small
\item[*] The category of computer science includes computer vision, cybersecurity, and program coding datasets.
\item[*] Node Attr. and Edge Attr. indicates the labels or features of nodes and edges, respectively.
\item[*] The size of datasets follows the setting of OGB \cite{hu2020open}, medium datasets have more than 1 million nodes or more than 10 million edges, and large datasets own over 100 million nodes or 1 billion edges.
\end{tablenotes}
\end{threeparttable}
\end{table*}
\subsection{Datasets}
Table \ref{table_data} summarizes a selection of benchmark graph-level datasets, including TUDateset \cite{morris2020tudataset}, Open Graph Benchmark (OGB) \cite{hu2020open}, MOLECULENET \cite{wu2018moleculenet}, MALNET \cite{freitas2021a}, and others \cite{dwivedi2020benchmarking}.
The graph datasets collected by the group at TUDateset \cite{morris2020tudataset} have been widely used to evaluate graph-level learning approaches.
These graph datasets consist of molecules, proteins, images, social networks, synthetic graphs, and data from many other domains.
However, despite their wide use, they have attracted criticism from some practitioners.
For example, Ivanov \textit{et al}. \cite{ivanov2019understanding} contends that the sets suffer from isomorphism bias, i.e., they contain isomorphic graphs with different labels, which may hinder model training \textemdash a claim based on the analysis of 54 widely-used graph datasets.
They also note that some of the datasets are too small to train a data-hungry deep learning model.
For example, Dwivedi \textit{et al.} \cite{dwivedi2020benchmarking} presented that most GL-GNNs have a close performance to others in the small dataset.
Further, some topology-agnostic baselines yield a performance that is competitive to GL-GNNs.
Developing practical and large-scale benchmark datasets has become an important issue for the graph-level learning community.
To this eend, Wu et al. \cite{wu2018moleculenet} proposed a benchmark named MOLECULENET that contains a set of large-scale graph datasets of molecules.
The dataset is designed to be used for graph regression and classification tasks. Dwivedi \textit{et al.} \cite{dwivedi2020benchmarking} transformed images into graphs for classification, in which a group of pixels is clustered as a node. Based on real-world cybersecurity scenarios, Freitas \textit{et al.} \cite{freitas2021a} proposed a large-scale graph dataset of over 1.2 million graphs with imbalanced labels.
Furthermore, OGB \cite{hu2020open} has published application-oriented large-scale graph datasets of molecules, proteins, and source code cooperation networks.
\subsection{Evaluations}
The development of graph-level learning has been impeded by unfair evaluations.
For example, Ruffinelli \textit{et al.} \cite{ruffinelli2019you} argue that some graph-level learning models only produce state-of-the-art performance because of tricks with the model's training, not because of the novel ideas proposed in the articles. However, there is no consensus on which evaluation to use with the most widely used graph datasets, such as TUDatasets, nor is there even a universally-accepted data split \cite{Errica2020A}.
Hence, to evaluate the graph-level learning models in a unified and fair way, some researchers have attempted to establish a standard model evaluation protocol.
For example, Dwivedi \textit{et al.} \cite{dwivedi2020benchmarking} built a benchmark framework based on PyTorch and DGL\footnote{DEEP GRAPH LIBRARY: \href{https://www.dgl.ai}{https://www.dgl.ai}}
that evaluates models on graph classification and graph regression tasks with an unified model evaluation protocol.
They do apply training tricks, such as batch normalization, residual connections, and graph size normalization, to GL-GNNs to measure their effects.
But all models being evaluated with the protocol are subject to the same training regime.
Similarly, in addition to the large-scale graph datasets, OGB \cite{hu2020open} provides a standard model evaluation protocol that includes a unified way to load and split data, the model evaluation itself, plus the cross-validations. Recently, Zhu \textit{et al.} \cite{zhu2021an} provides a benchmark framework for graph contrastive learning.
\section{Downstream tasks and Applications}
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1\textwidth]{application.pdf}
\caption{An overview of graph-level learning downstream tasks and their practical applications.}
\label{fapplication}
\end{figure*}
This section introduces the mainstream downstream tasks of graph-level learning and their corresponding applications. Fig. \ref{fapplication} summarizes some common downstream tasks and applications in graph-level learning.
\noindent{\textbf{Graph Generation.}} This task aims to generate new graphs that have specific proprieties based on a series of graphs.
Graph generation has a broad application in the field of biochemistry. For instance, \emph{drug development} involves experimenting with a tremendous number of molecule arrangements, but, through graph generation, the overall time and investment required to do this work can be reduced \cite{vamathevan2019applications}. Similarly, \emph{molecule generation} \cite{Shi*2020GraphAF:,morgan1965generation} has been used to explore new catalysts \cite{chanussot2021open}.
Sanchez \textit{et al.} \cite{sanchez2020learning} applied graph generation into \emph{physical systems modeling} to simulate real-world particle motions. \emph{Scene graph generation} \cite{xu2017scene,ren2015faster} can be used to understand the scene of images and generate abstraction for images to summarize the relationship among objects in an image. Most recently, a few works \cite{allamanis2018learning,dinella2020hoppity} have employed graph generation for \emph{program debugging}, which modifies the nodes (i.e., variables or functions) and links in the program flow graph to fix bugs.
\noindent\textbf{Subgraph Discovery.}
This is the task of detecting discriminative substructures in a graph dataset. Subgraph Discovery can be applied to \emph{molecular structure search} \cite{ralaivola2005graph,duvenaud2015convolutional}, which explores the functional structures in chemical compounds, or to \emph{social event detection} \cite{shao2017efficient}, where subgraph discovery can be used to detect the substructures that represent great events in a series of social networks.
\noindent \textbf{Graph Classification.}
The goal of graph classification is to learn the mapping relationship between graphs and corresponding class labels and predict the labels of unseen graphs. Graph classification is a critical graph-level learning task with a range of applications.
For example, classifying molecular graphs \cite{wang2022molecular,fang2022geometry} can be used to determine anti-cancer activity, toxicity, or the mutagenicity of molecules. Classifying protein graphs \cite{borgwardt2005protein} can help to identify proteins with specific functions, such as enzymes. By converting texts to graphs in which nodes denote words and edges are the relationships between words, \emph{text categorization} \cite{Rousseau2015textgraphclass,peng2018large} can distinguish documents with different topics. By the same token, pixels in images can be regarded as nodes and adjacent pixels are linked to yield graphs for \emph{image recognition} \cite{chen2019multi,wu2015multi}. This task can be extended to \emph{medical diagnosis} to deal with computed tomography scans \cite{hao2022uncertainty} and clinical images \cite{wu2020learning}. In addition, graph classification can also be used for \emph{online product recommendation} \cite{wu2014bag} and \emph{fake news detection} \cite{dou2021user,silva2021embracing}.
Recently, it has been impressive to see that graph-level learning can deliver \emph{IQ tests} \cite{Wang2020Abstract} that select graphs with a specific style from a group of graphs based on the style learning on the other group of graphs.
\noindent \textbf{Graph Regression.} This task aims to predict the continuous proprieties of graphs. Taking molecules as examples, graph regression can predict different molecular proprieties related to the tightness of chemical bounds, fundamental vibrations, the state of electrons in molecules, the spatial distribution of electrons in molecules, and so on \cite{gilmer2017neural, jiang2021could, stokes2020deep}. Hence, the most promising application of graph regression is \emph{drug discovery}. In addition, employing graph regression to predict ratings or avenues of films is feasible.
\noindent \textbf{Graph Comparison.}
This task involves measuring the distance or similarity of pair-wise graphs in a graph dataset. The applications of this task include: \emph{semantic inference}, which infers text-document affiliations \cite{haghighi2005robust}; matching images with texts describing the same thing \cite{liu2020graph}; \emph{semantic metrics}, which measures the semantic similarity between texts \cite{ramage2009random,hughes2007lexical}; and \emph{cross-language information retrieval}, which seeks information in a language context that is different from the query \cite{xu-etal-2019-cross-lingual,monz2005iterative}.
\noindent \textbf{Applying Complex Scenarios.} In addition to simple downstream tasks, researchers have extended graph-level learning to some complex scenarios.
For instance, \emph{multi-view GL} targets learning in scenarios where an object is described by multiple graphs (i.e., multi-graph-views).
In multi-view GL, practitioners mine information from each single-graph-view and then strategically fuse information from all graph-views \cite{wu2016multi,wu2017multiple}.
\emph{Multi-task GL} \cite{pan2015joint,pan2016task} is generally used to optimize multiple related tasks; hence, it focuses on detecting the discriminative features across all the different tasks.
In real-world scenarios, there are a vast number of unlabeled graphs that go unused since most GL techniques require learning from labeled information.
Consequently, \emph{semi-supervised GL} \cite{kong2010semi} was developed, which can learn from a dataset containing only a few labeled graphs and very many unlabeled graphs. Likewise, \emph{positive and unlabeled GL} \cite{wu2014PU,wu2016positive} only requires a few labeled graphs in one class along with other unlabeled graphs.
In terms of dynamic scenarios, there are also applications that record changing graphs over time as graph streams.
For example, a paper and its references can be regarded as a citation graph, and a graph stream can be produced of citations in chronological order of the corresponding papers.
\emph{Graph stream GL} \cite{pan2014graphstream,aggarwal2011classification}, for example, is specifically designed for graph stream data and mines valuable patterns from dynamic graph records.
\section{Taxonomy}
\begin{figure*}[hbp!]
\centering
\includegraphics[width=0.8\textwidth]{tree.pdf}
\caption{The taxonomy tree of graph-level learning techniques.}
\label{tree}
\end{figure*}
This section provides a taxonomy of graph-level learning techniques.
Its categorizes, include traditional learning, graph-level deep neural networks (GL-DNNs), graph-level graph neural networks (GL-GNNs), and graph pooling.
The taxonomy tree in Fig. \ref{tree} depicts these four branches of graph-level learning with selected algorithms highlighted.
Each category is briefly introduced next.
\subsection{Traditional Learning:}
Traditional forms of learning used to be the mainstream in graph-level learning before the great success of deep learning techniques emerged.
The majority of traditional learning methods rely on handcrafted features to describe the graphs, such as random walk sequences \cite{kriege2020survey}.
Given deterministic graph descriptions, off-the-shelf machine learning models were used to perform downstream tasks, like graph classification, in a non-end-to-end fashion.
Even today, traditional methods have some advantages over deep learning techniques, such as better interpretability and a better ability to model irregular structures.
For these reasons, traditional methods are still providing the graph-level learning community with valuable insights.
\subsection{Graph-Level Deep Neural Networks (GL-DNNs):} Traditional approaches do not just include classic methods.
They also include tentative explorations into deep learning techniques, such as RNNs, CNNs, and CapsNet.
These three types of deep neural networks were not initially designed to learn non-Euclidean data like graphs. Hence, one of the important issues with GL-DNNs is how to enable these deep neural networks to learn graph-structured data that varies in size and has irregular neighborhood structures.
\subsection{Graph-Level Graph Neural Networks (GL-GNNs):} GL-GNNs use graph convolution operations specifically proposed for graph-structured data as the backbone for performing graph-level learning.
Most GL-GNNs use the graph convolutions MPNNs frameworks because they are simple, easy to understand, and have linear complexity. GL-GNNs condense the most fruitful achievements of graph-level learning.
In addition, some practitioners integrate the advantages of MPNN-based GL-GNNs with other techniques, particularly traditional learning techniques, to improve graph-level learning.
\subsection{Graph Pooling:} Graph pooling is a kind of graph downsizing technology where compact representations of a graph are produced by compressing a series of nodes into a super node.
For example, graph pooling can globally aggregate all of a graph's node embeddings into one ultimate super node via a summation or averaging operation.
At the same time, graph pooling can reduce the size of a graph layer by layer.
This kind of multi-layer downsizing processes tends to aggregate nodes in the same hierarchy (e.g., a community) at each layer.
\section{Traditional Learning}
\begin{table*}[]
\caption{Summary of Traditional Graph-level Learning Methods.}\label{table_traditional}
\begin{tabular}{cclllll}
\hline
Subsection & Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{17}{*}{\begin{tabular}[c]{@{}c@{}}Graph \\ Kernels\end{tabular}} & \multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Message\\ Passing\\ Kernels\end{tabular}} & 2009 & NHK\cite{hido2009lineara} & ICDM & Python & https://github.com/ysig/GraKeL \\ \cline{3-7}
& & 2011 & WL\cite{shervashidze2011weisfeiler} & JMLR & C++ & https://github.com/BorgwardtLab/graph-kernels \\ \cline{3-7}
& & 2016 & PK\cite{neumann2016propagationa} & ML & MATLAB & https://github.com/marionmari/propagation\_kernels \\ \cline{3-7}
& & 2017 & Global-WL\cite{morris2017glocalizeda} & ICDM & C++ & https://github.com/chrsmrrs/glocalwl \\ \cline{3-7}
& & 2019 & P-WL\cite{rieck2019persistenta} & ICML & Python & https://github.com/BorgwardtLab/P-WL \\ \cline{2-7}
& \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}ShortestPath\\ Kernels\end{tabular}} & 2005 & SPK\cite{borgwardt2005shortest} & ICDM & Python & https://github.com/ysig/GraKeL \\ \cline{3-7}
& & 2017 & SPK-DS\cite{nikolentzos2017shortesta} & EMNLP & - & - \\ \cline{2-7}
& \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Random\\ Walk\\ Kernels\end{tabular}} & 2003 & RWK\cite{gartner2003graph} & LNAI & Python & https://github.com/jajupmochi/graphkit-learn \\ \cline{3-7}
& & 2004 & ERWK\cite{mahe2004extensionsa} & ICML & Python & https://github.com/jajupmochi/graphkit-learn \\ \cline{3-7}
& & 2010 & SOMRWK\cite{vishwanathan2010graph} & JMLR & Python & https://github.com/ysig/GraKeL \\ \cline{2-7}
& \multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}Optimal\\ Assignment\\ Kernels\end{tabular}} & 2005 & OAK\cite{frohlich2005optimal} & ICML & - & - \\ \cline{3-7}
& & 2013 & PS-OAK\cite{pachauri2013solvinga} & NeurIPS & Python & https://github.com/zju-3dv/multiway \\ \cline{3-7}
& & 2015 & GE-OAK\cite{johansson2015learninga} & KDD & - & - \\ \cline{3-7}
& & 2015 & TAK\cite{schiavinato2015transitivea} & SIMBAD & - & - \\ \cline{2-7}
& \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Subgraph\\ Kernels\end{tabular}} & 2009 & Graphlet\cite{shervashidze2009efficient} & AISTATS & Python & https://github.com/ysig/GraKeL \\ \cline{3-7}
& & 2010 & NSPDK\cite{costa2010fasta} & ICML & Python & https://github.com/fabriziocosta/EDeN \\ \cline{3-7}
& & 2012 & SMK\cite{10.5555/3042573.3042614} & ICML & C++ & https://github.com/fapaul/GraphKernelBenchmark \\ \hline
\multirow{9}{*}{\begin{tabular}[c]{@{}c@{}}Subgraph \\ Mining\end{tabular}} & \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Frequent\\ Subgraph\\ Mining\end{tabular}} & 2000 & AGM\cite{inokuchi2000apriori} & ECML PKDD & C++ & https://github.com/Aditi-Singla/Data-Mining \\ \cline{3-7}
& & 2001 & FSG\cite{kuramochi2001frequent} & ICDM & C++ & https://github.com/NikhilGupta1997/Data-Mining-Algorithms \\ \cline{3-7}
& & 2002 & gSpan\cite{yan2002gspana} & ICDM & Python & https://github.com/betterenvi/gSpan \\ \cline{2-7}
& \multirow{6}{*}{\begin{tabular}[c]{@{}c@{}}Discrimina\\ -tive\\ Subgraph\\ Mining\end{tabular}} & 2008 & LEAP\cite{yan2008mininga} & SIGMOD & - & - \\ \cline{3-7}
& & 2009 & CORK\cite{thoma2009near} & SDM & - & - \\ \cline{3-7}
& & 2010 & gMLC\cite{kong2010multia} & ICDM & - & - \\ \cline{3-7}
& & 2010 & gSSC\cite{kong2010semi} & KDD & - & - \\ \cline{3-7}
& & 2011 & gPU\cite{zhao2011positivea} & ICDM & - & - \\ \cline{3-7}
& & 2014 & gCGVFL\cite{wu2014multi} & ICDM & - & - \\ \hline
\multirow{9}{*}{\begin{tabular}[c]{@{}c@{}}Graph \\ Embedding\end{tabular}} & \multirow{6}{*}{\begin{tabular}[c]{@{}c@{}}Determin\\ -istic\end{tabular}} & 2017 & FGSD\cite{verma2017hunt} & NeurIPS & Python & https://github.com/vermaMachineLearning/FGSD \\ \cline{3-7}
& & 2018 & AWE\cite{ivanov2018anonymous} & ICML & Python & https://github.com/nd7141/AWE \\ \cline{3-7}
& & 2019 & LDP\cite{cai2018simplea} & ICLR RLGM & Python & https://github.com/Chen-Cai-OSU/LDP \\ \cline{3-7}
& & 2020 & SLAQ\cite{tsitsulin2020justa} & WWW & Python & \begin{tabular}[c]{@{}l@{}}https://github.com/google-research/google-research/tree/\\ master/graph\_embedding/slaq\end{tabular} \\ \cline{3-7}
& & 2021 & VNGE\cite{liu2021bridginga} & WWW & Python & https://github.com/xuecheng27/WWW21-Structural-Information \\ \cline{3-7}
& & 2021 & A-DOGE\cite{sawlani2021fasta} & ICDM & Python & https://github.com/sawlani/A-DOGE \\ \cline{2-7}
& \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Learn\\ -able\end{tabular}} & 2016 & Subgraph2vec\cite{narayanan2016subgraph2vec} & KDD MLG & Python & https://github.com/MLDroid/subgraph2vec\_tf \\ \cline{3-7}
& & 2017 & Graph2vec\cite{narayanan2017graph2veca} & KDD MLG & Python & https://github.com/MLDroid/graph2vec\_tf \\ \cline{3-7}
& & 2018 & GE-FSG\cite{nguyen2018learninga} & SDM & Python & https://github.com/nphdang/GE-FSG \\ \hline
\end{tabular}
\end{table*}
All traditional graph-level learning publications discussed in this section are summarized in Table \ref{table_traditional}.
\subsection{Graph Kernels (GKs)}
\subsubsection{Message Passing Kernels (MPKs)}
The followed upgrading of 1-WL mainly focuses on aggregation and relabeling steps.
Hido and Kashima \cite{hido2009lineara} replaced the hash function with a binary arithmetic giving rise to a faster $\phi(\cdot)$.
Morris \textit{et al.} \cite{morris2017glocalizeda} used the idea of $k$-WL to relabel node groups consisting of $k$ nodes that could form a connected graph.
Theoretically, $k$-WL is more powerful than $1$-WL for distinguishing between graph structures.
Further, Neumann \textit{et al.} \cite{neumann2016propagationa} proposed a random label aggregation process based on node label distributions that only considers labels of part of neighbors.
Random label aggregation saves time and computational resources making work on large-scale graphs more efficient.
Persistent Weisfeiler–Lehman (P-WL) \cite{rieck2019persistenta} is the recent enhancement to MPKs that adds weighted edges into the aggregation process. To calculate the edge weight, P-WL measures the distance between the continuous iterative updated labels of two end nodes.
Additionally, P-WL can track changes in substructures that cannot be identified by 1-WL, such as cycles.
\subsubsection{Shortest-path Kernels (SPKs)}
Nikolentzos \cite{nikolentzos2017shortesta} proposed a variant of SPKs that draws on more information in a shortest path, such as node and edge labels, to calculate the distance of any two paths.
\subsubsection{Random Walk Kernels (RWKs)}
RWKs suffer from a problem called tottering, where a random walk sequence traverses $v$ to $u$ and immediately returns to $v$ via the same edge.
To address tottering, Mah{\'e} \textit{et al.} \cite{mahe2004extensionsa} employed a second-order Markov random walk that considers the last two steps in the current random walk sequence when deciding the next step.
\subsubsection{Optimal Assignment Kernels (OAKs)}
Searching for a pair-wise element with the maximum similarity tends to be a highly time-consuming process.
Hence, to reduce the time requirement of this task, Johansson et al. \cite{johansson2015learninga} mapped the graphs in geometric space and then calculated the Euclidean distance between pair-wise nodes.
This method enables OAKs to use approximate nearest neighbors algorithms in Euclidean space as a way to speed up the process.
Transitive Assignment Kernels (TAKs) \cite{schiavinato2015transitivea,pachauri2013solvinga} are variants of OAKs.
Unlike OAKs that search for the optimal assignment among pair-wise graphs, TAKs identify node permutations that with the most similar node pairs among three or more graphs.
OAKs have been confined to node similarity measurement, although they can be extended to measure subgraph similarities so as to capture a graph's topological information \cite{woznica2010adaptivea}. As discussed next, we introduce the GKs with subgraph information.
\subsubsection{Subgraph Kernels (SGKs)}
There are several different implementations of SGKs kernel functions.
For instance, Wale \textit{et al.} \cite{wale2008comparisona} employed a min-max kernel $\displaystyle{ \tfrac{\sum_{i=1}^N min(c^{\left(\mathcal G\right)}_{\mathcal T_i},c^{\left(\mathcal G'\right)}_{\mathcal T_i})}{\sum_{i=1}^N max(c^{\left(\mathcal G\right)}_{\mathcal T_i},c^{\left(\mathcal G'\right)}_{\mathcal T_i})}}$ to measure the distance between two graphs.
Subgraph Matching Kernels (SMKs) \cite{10.5555/3042573.3042614} calculate the similarity between two subgraphs by counting the number of nodes with the same labels.
Then the similarities between all pairwise subgraphs sourced from the two graphs are summed as the kernel value of the SMKs.
Methods of identifying the subgraphs in SGKs have also been explored.
For example, Neighborhood Subgraph Pairwise Distance Kernels (NSPDK) \cite{costa2010fasta} denotes the subgraphs as the first-, second-, and third-hop neighborhoods of pairwise vertices with the shortest path of a predefined length.
However, the main contributions of SGKs lie in assessing the similarity of graphs in terms of a set of selected subgraphs, not how the subgraphs are chosen.
More detailed and sophisticated subgraph mining methods are demonstrated next.
\subsection{Subgraph Mining}
\iffalse
\subsubsection{Frequent Subgraph Mining (FSM)}
Testing for subgraph isomorphism with vast numbers of candidate subgraphs can mean apriori-like algorithms suffer from computation bottlenecks.
To address this issue, gSpan \cite{yan2002gspana} employs a depth-first-search (DFS) strategy to search subgraphs, while assigning a unique DFS code of minimum length for each subgraph searched.
gSpan can then do a quick check for isomorphism by simply comparing the DFS codes of pairwise subgraphs.
\fi
\subsubsection{Discriminative Subgraph Mining (DSM)}
To speed up DSM based on discriminative scores and gSpan, Yan \textit{et al.} \cite{yan2008mininga} proposed LEAP, which initializes an optimal subgraph set $\mathcal S^*$ with frequent subgraphs.
In this way, LEAP prunes gSpan's search space right at the beginning. In addition, Kong \textit{et al.} \cite{kong2010multia} and Wu \textit{et al.} \cite{wu2014multi} expanded DSM to the multi-label\footnote{Each graph owns more than one label, such as a drug molecular can own different labels to represent anti-cancer effects for various cancers, e.g., breast cancer (+) and lung cancer (-).} and multi-view\footnote{An object has different views, where each view can represent a separate graph, e.g., a scientific publication network is shown as two graphs, an abstract graph demonstrating the keywords correlations in the abstract of papers, and a reference citation graph about citation relationships.} scenarios, respectively.
Note, however, that all the DSM methods discussed are supervised methods.
In terms of semi-supervised subgraph mining, Kong and Yu \cite{kong2010semi} proposed gSSC which maps each graph into a new feature space by $\mathcal S^*$. Unlabeled graphs are separated from each other in the new feature space.
In the labeled group, graphs with the same label are close, whereas graphs with different labels remain distant. In addition, Zhao \textit{et al.} \cite{zhao2011positivea} only used the positively labeled graphs and unlabeled graphs to select $\mathcal S^*$ when performing binary graph classification tasks. This is because sometimes the real-world data is composed of an incomplete set of positive instances and unlabeled graphs.
\subsection{Graph Embedding}
\subsubsection{Deterministic Graph Embedding}
Local Degree Profile (LDP) \cite{cai2018simplea} summarizes the degree information of each node and its 1-hop neighbors as node features.
LDP constructs graph representations by building an empirical distribution or histogram of some hand-crafted node features.
In addition to node degree, deterministic graph embedding can also leverage anonymous random walk sequences to describe a graph's topological information.
A-DOGE \cite{sawlani2021fasta}, for instance, depicts a graph by computing the spectral density across its eigenspectrum.
However, these methods are limited to use with small graphs given the prohibitive costs of computing eigenspectrum decompositions with large-scale graphs.
As a possible solution to this limitation, SlaQ \cite{tsitsulin2020justa} uses stochastic approximations as a way of quickly calculating the distance between two graphs' spectral densities.
More specifically, these authors employed von Neumann graph entropy (VNGE) \cite{braunstein2006laplaciana,chen2019fasta} as a way of approximately representing the spectral properties of the graphs.
In turn, this approximation supports fast computation by tracing a Laplacian matrix of the graph.
Liu \textit{et al.} \cite{liu2021bridginga} proposed another fast approximation method involving VNGE, which is based on deriving the error bound of the approximation estimation.
\subsubsection{Learnable Graph Embedding}
Graph2vec \cite{narayanan2017graph2veca} was designed to tackle graph representation tasks.
By establishing a semantic association between a graph and its sampled subgraphs, Graph2Vec employs the idea of Skip-gram to learn a graph embedding.
Following this work, Dang \textit{et al.} \cite{nguyen2018learninga} replaced the sampled subgraphs in Graph2vec with frequent subgraphs that have more discriminative features for graph classification tasks.
\section{Graph-Level Deep Neural Networks (GL-DNNs)}
The representative GL-DNNs mentioned in this section are summarized in Table \ref{table_dnn}.
\begin{table*}[]
\centering
\caption{Summary of Graph-Level Deep Neural Networks (GL-DNNs).}\label{table_dnn}
\begin{tabular}{clllll}
\hline
Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}RNN-\\ Based\end{tabular}} & 2016 & GGNN\cite{li2016gated} & ICLR & Python-Tensorflow & https://github.com/Microsoft/gated-graph-neural-network-samples \\ \cline{2-6}
& 2018 & GAM\cite{lee2018graph} & KDD & Python-Pytorch & https://github.com/benedekrozemberczki/GAM \\ \cline{2-6}
& 2018 & SAN\cite{zhao2018substructurea} & AAAI & - & - \\ \cline{2-6}
& 2018 & NetGAN\cite{bojchevski2018netgana} & ICML & Python-Tensorflow & https://github.com/danielzuegner/netgan \\ \cline{2-6}
& 2018 & GraphRNN\cite{you2018graphrnna} & ICML & Python-Pytorch & https://github.com/snap-stanford/GraphRNN \\ \hline
\multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}CNN-\\ Based\end{tabular}} & 2016 & PATCHYSAN\cite{niepert2016learning} & ICML & Python & https://github.com/tvayer/PSCN \\ \cline{2-6}
& 2016 & DCNN\cite{NIPS2016_390e9825a} & NeurIPS & Python & https://github.com/jcatw/dcnn \\ \cline{2-6}
& 2017 & ECC\cite{simonovsky2017dynamic} & CVPR & Python-Pytorch & https://github.com/mys007/ecc \\ \cline{2-6}
& 2018 & KCNN\cite{nikolentzos2018kernel} & ICANN & Python-Pytorch & https://github.com/giannisnik/cnn-graph-classification \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}CapsNet-\\ Based\end{tabular}} & 2018 & GCAPSCNN\cite{verma2018graph} & WCB & Python & https://github.com/vermaMachineLearning/Graph-Capsule-CNN-Networks \\ \cline{2-6}
& 2019 & CapsGNN\cite{xinyi2018capsulea} & ICLR & Python-Pytorch & https://github.com/benedekrozemberczki/CapsGNN \\ \cline{2-6}
& 2019 & PATCHYCaps\cite{mallea2019capsulea} & Arxiv & Python & https://github.com/BraintreeLtd/PatchyCapsules \\ \hline
\end{tabular}
\end{table*}
\subsection{RNN-based GL-DNNs}
Zhao \textit{et al.} \cite{zhao2018substructurea} proposed an RNN-based graph classification algorithm called SAN.
Starting from a node, SAN employs an RNN model that adds nodes and edges to form an informative substructure whose representation is progressively generated by the RNN model.
A graph-level representation that can be used for graph classification tasks is then generated by summing all the representations of the formed substructures.
Given a graph generation task, NetGAN \cite{bojchevski2018netgana} uses an LSTM model as a generator to yield fake walk sequences,
while a discriminator disambiguates the graph's real walk sequences from the generated fake ones to reverse train the generator.
Another graph generation model Graphrnn \cite{you2018graphrnna} creates various permutations of graphs, with various combinations of nodes and edges as sequential data to be input into an RNN model.
\subsection{CNN-based GL-DNNs}
There are two main branches of CNN-based graph-level learning.
The first branch is illustrated in Fig. \ref{fGLDNN} (A), which sorts nodes and arranges the node features to form a concentration matrix as the grid-structured data for training the CNNs.
As a second branch, researchers have developed a CNN-guided neural network version of an MPK, which is shown in Fig. \ref{fGLDNN} (B).
\begin{figure*}[h]
\begin{minipage}{1\linewidth}
\vspace{3pt}
\centerline{\includegraphics[width=\textwidth]{DNN1.pdf}}
\centerline{(A)}
\end{minipage}
\begin{minipage}{1\linewidth}
\vspace{3pt}
\centerline{\includegraphics[width=\textwidth]{DNN2.pdf}}
\centerline{(B)}
\end{minipage}
\caption{Two ways of tentative exploration by CNN-based GL-DNNs on graph-structured data.} \label{fGLDNN}
\end{figure*}
Kernel Convolutional Neural Network (KCNN) \cite{nikolentzos2018kernel} sorts all vertices in a graph to form grid-structured data.
A neighborhood graph is built for each vertex and a kernel matrix is constructed by implementing the kernel function (i.e., an SPK or an MPK) between all pairwise neighborhood graphs. In this work, the grid-structured data for feeding up CNN is the kernel matrix, where each row is a vector describing the similarities between the neighborhood graph of the matched index vertex and the other neighborhood graphs.
Another related work, Diffusion CNN (DCNN) \cite{NIPS2016_390e9825a} aggregates multi-hop neighborhood features to the central nodes through a matrix multiplication $\mathbf P \mathbf X$, where $\mathbf P = [\mathbf A, \mathbf A^2, ..., \mathbf A^h] \in \mathbb R^{h \times n\times n}$ is a three-dimensional tensor containing multi-hop (i.e., 1-, 2-, ..., h-hops) adjacent matrices and $\mathbf X \in \mathbb R^{n\times f}$ is the node features matrix. $\mathbf P \mathbf X \in \mathbb R^{h \times n\times f}$ represents the updated node features after multi-hop aggregation.
For graph classification tasks, DCNN permutes the dimensions giving $\mathbf P \mathbf X \in \mathbb R^{n \times h\times f}$ and all node representations are averaged as $\mathbf P^{*} \in \mathbb R^{h\times f}$.
Subsequently, a convolution operation is implemented on $\mathbf P^{*}$ to produce a graph-level representation. Formally,
\begin{equation}
\mathbf{h}_{\mathcal{G}} = f\left(\mathbf W \odot \mathbf P^{*}\right),
\end{equation}
where $f\left(\cdot\right)$ is a nonlinear activation function, and $\mathbf W$ is a trainable weight matrix for convolution and summation.
\subsection{CapsNet-based GL-DNNs}
CapsGNN \cite{xinyi2018capsulea} iteratively aggregates node features to a center node, and, in turn, adds the aggregation results of each iteration to a capsule representation of the central node.
An attention mechanism is then applied to all node capsules so as to generate a graph capsule that can be plugged into a capsule network for graph classification.
Mallea \textit{et al.} \cite{mallea2019capsulea} employs the same approach as PATCHY-SAN \cite{niepert2016learning} to find substructures in graphs, while the feature matrices of searched substructures are assembled in a capsule network for graph classification.
\section{Graph-Level Graph Neural Networks (GL-GNNs)}
This section discusses contrastive learning as it pertains to GL-GNNs (Section \ref{contrastive GL-GNNs}), the expressivity (see Section \ref{expressivity}), generalizability (see Section \ref{Generalizability}), and explainability (see Section \ref{Explanation-GNN}) of GL-GNNs.
The GL-GNNs referred to are summarized in Table \ref{table_gnn}.
\begin{table*}[]
\centering
\caption{Summary of Graph-Level Graph Neural Networks (GL-GNNs).}\label{table_gnn}
\begin{tabular}{clllll}
\hline
Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{12}{*}{\begin{tabular}[c]{@{}c@{}}Message\\ Passing\\ Neural\\ Networks\end{tabular}} & 2015 & Fingerprint\cite{duvenaud2015convolutional} & NeurIPS & Python-Tensorflow & https://github.com/HIPS/neural-fingerprint \\ \cline{2-6}
& 2016 & GraphSim\cite{battaglia2016interaction} & NeurIPS & Python-Tensorflow & https://github.com/clvrai/Relation-Network-Tensorflow \\ \cline{2-6}
& 2017 & MPNN\cite{gilmer2017neural} & ICML & Python-Pytorch & https://github.com/priba/nmp\_qc \\ \cline{2-6}
& 2017 & DTNN\cite{schutt2017quantum} & NC & Python-Tensorflow & https://github.com/atomistic-machine-learning/dtnn \\ \cline{2-6}
& 2019 & GIN\cite{xu2018powerful} & ICLR & Python-Pytorch & https://github.com/weihua916/powerful-gnns \\ \cline{2-6}
& 2019 & K-GNNs\cite{morris2019weisfeilera} & AAAI & Python-Pytorch & https://github.com/chrsmrrs/k-gnn \\ \cline{2-6}
& 2019 & PPGN\cite{maron2019provablya} & NeurIPS & Python-Tensorflow & https://github.com/hadarser/ProvablyPowerfulGraphNetworks \\ \cline{2-6}
& 2019 & RP\cite{murphy2019relationala} & ICML & Python-Pytorch & https://github.com/PurdueMINDS/RelationalPooling \\ \cline{2-6}
& 2021 & FGNN\cite{azizian2021expressivea} & ICLR & Python-Pytorch & https://github.com/mlelarge/graph\_neural\_net \\ \cline{2-6}
& 2021 & SWL\cite{bodnar2021weisfeiler_icmla} & ICML & Python-Pytorch & https://github.com/twitter-research/cwn \\ \cline{2-6}
& 2021 & CWN\cite{bodnar2021weisfeiler_nipsa} & NeurIPS & Python-Pytorch & https://github.com/twitter-research/cwn \\ \cline{2-6}
& 2021 & RNI\cite{DBLP:conf/ijcai/AbboudCGL21a} & IJCAI & - & - \\ \hline
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Subgraph\\ -Based\end{tabular}} & 2020 & GSN\cite{bouritsas2022improving} & TPAMI & Python-Pytorch & https://github.com/gbouritsas/GSN \\ \cline{2-6}
& 2020 & SubGNN\cite{AlsentzerFLZ20subgraph} & NeurIPS & Python-Pytorch & https://github.com/mims-harvard/SubGNN \\ \cline{2-6}
& 2021 & GNN-AK\cite{zhao2022from} & ICLR & Python-Pytorch & https://github.com/LingxiaoShawn/GNNAsKernel \\ \cline{2-6}
& 2021 & NGNN\cite{zhang2021nested} & NeurIPS & Python-Pytorch & https://github.com/muhanzhang/nestedgnn \\ \cline{2-6}
& 2021 & GraphSNN\cite{wijesinghe2021new} & ICLR & Python-Pytorch & https://github.com/wokas36/GraphSNN \\ \cline{2-6}
& 2021 & SUGAR\cite{sun2021sugar} & WWW & Python-Tensorflow & https://github.com/RingBDStack/SUGAR \\ \cline{2-6}
& 2021 & ESAN\cite{bevilacqua2021equivariant} & ICLR & Python-Pytorch & https://github.com/beabevi/esan \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Kernel\\ -Based\end{tabular}} & 2019 & GNTK\cite{du2019graph} & NeurIPS & Python & https://github.com/KangchengHou/gntk \\ \cline{2-6}
& 2019 & DDGK\cite{al2019ddgk} & WWW & Python-Tensorflow & \begin{tabular}[c]{@{}l@{}}https://github.com/google-research/google-research/\\ tree/master/graph\_embedding/ddgk\end{tabular} \\ \cline{2-6}
& 2020 & GCKN\cite{chen2020convolutional} & ICML & Python-Pytorch & https://github.com/claying/GCKN \\ \cline{2-6}
& 2020 & RWNN\cite{NEURIPS2020_ba95d78a} & NeurIPS & Python-Pytorch & https://github.com/giannisnik/rwgnn \\ \cline{2-6}
& 2021 & GSKN\cite{long2021theoretically} & WWW & Python-Pytorch & https://github.com/YimiAChack/GSKN \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Contrastive\\ -Based\end{tabular}} & 2020 & GraphCL\cite{NEURIPS2020_3fe23034a} & NeurIPS & Python-Pytorch & https://github.com/Shen-Lab/GraphCL \\ \cline{2-6}
& 2020 & InfoGraph\cite{sun2019infographa} & ICLR & Python-Pytorch & https://github.com/fanyun-sun/InfoGraph \\ \cline{2-6}
& 2020 & GCC\cite{qiu2020gcca} & KDD & Python-Pytorch & https://github.com/THUDM/GCC \\ \cline{2-6}
& 2020 & MVGRL\cite{hassani2020contrastivea} & ICML & Python-Pytorch & https://github.com/kavehhassani/mvgrl \\ \cline{2-6}
& 2021 & JOAO\cite{you2021grapha} & ICML & Python-Pytorch & https://github.com/Shen-Lab/GraphCL\_Automated \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Spectral\\ -Based\end{tabular}} & 2016 & ChebNet\cite{defferrard2016convolutional} & NeurIPS & Python-Tensorflow & https://github.com/mdeff/cnn\_graph \\ \cline{2-6}
& 2021 & GNNTFS\cite{levie2021transferability} & JMLR & - & - \\ \cline{2-6}
& 2021 & GNNMatlang\cite{balcilar2021breaking} & ICML & Python-Tensorflow & https://github.com/balcilar/gnn-matlang \\ \cline{2-6}
& 2021 & ARMA\cite{bianchi2021graph} & TPAMI & Python-Pytorch & https://github.com/dmlc/dgl/tree/master/examples/pytorch/arma \\ \cline{2-6}
& 2021 & UFG\cite{zheng2021framelets} & ICML & Python-Pytorch & https://github.com/YuGuangWang/UFG \\ \hline
\end{tabular}
\end{table*}
\subsection{Contrastive Learning-Based GL-GNNs}\label{contrastive GL-GNNs}
Contrastive learning \cite{hjelm2018learninga} is a data augmentation method that creates new, plausible instances by transposing existing data without affecting the semantics.
Investigating contrastive learning for GL-GNNs is significant since GL-GNNs are data-driven models that will always encounter bottlenecks given insufficiently labeled graphs.
Graph Contrastive Learning (GraphCL) \cite{NEURIPS2020_3fe23034a} defines four approaches to creating new instances as augmentation data: (1) node dropping, which randomly removes a proportion of nodes from the graph; (2) edge perturbation, which randomly adds or removes a certain percentage of edges from the graph; (3) feature masking, where some of the features of some nodes are randomly masked; and (4) subgraphs, where subgraphs are taken from the graph.
To be noticed, the newly-produced instances must be labeled as the same class as the source graph.
InfoGraph \cite{sun2019infographa}, for example, samples subgraphs $g_m$ from a source graph $\mathcal G$ as new instances.
A GL-GNN encoder $\mathcal H^{\phi}$ with some parameters $\phi$ is then used to generate graph-level representations of $g_m$ and $\mathcal G$, denoted as $\mathbf h^{\phi}_{g_m}$ and $\mathbf h^{\phi}_{\mathcal G}$.
InfoGraph's learning objective is to maximize the mutual information between $\mathbf h^{\phi}_{\mathcal G}$ and all $\mathbf h^{\phi}_{g_m}, g_m \in \mathcal G$.
This can be brought of as an evaluation of the statistical dependencies between two variables. Formally:
\begin{equation}
\mathcal H^{\phi}, \mathcal H^{\psi}=\underset{\phi, \psi}{\operatorname{argmax}} \sum_{\mathcal G \in \mathbb G} \frac{1}{|\{g_m\}|} \sum_{g_m \in \mathcal G} I_{\phi, \psi}\left(\mathbf h^{\phi}_{g_m} ; \mathbf h^{\phi}_{\mathcal G}\right),
\end{equation}
where $\mathcal H^{\psi}$ is the mutual information estimator with the parameters $\psi$, and $I_{\phi, \psi}\left(\cdot, \cdot\right)$ measures the mutual information.
Similarly, Graph Contrastive Coding (GCC) \cite{qiu2020gcca} samples subgraphs $g_1,...,g_M$ from the graph dataset $\mathbb G = \{\mathcal G_1,...,\mathcal G_N\}$ as new instances.
The embeddings of the subgraphs $g_m$ and graph $\mathcal G_n$ produced by the GL-GNN are denoted as $\mathbf h_{g_m}$ and $\mathbf h_{\mathcal G_n}$, respectively.
In the GL-GNN, a GCC employs InfoNCE loss \cite{oord2018representationa} as the learning objective, that is:
\begin{equation}\label{infonce}
\mathcal L = \sum_{\mathcal G_n \in \mathbb G} - \log \frac{\sum_{g_m \in \mathcal G_n} \exp\left(\mathbf h_{\mathcal G_n}^{\top} \mathbf h_{g_m} / \tau \right)}
{\sum_{i=0}^{M} \exp\left( \mathbf h_{\mathcal G_n}^{\top} \mathbf h_{g_i} / \tau \right)},
\end{equation}
where $\tau$ is the temperature hyper-parameter. If $g_m \in \mathcal G_n$, the InfoNCE aims to maximize the similarity between $\mathbf h_{g_m}$ and $\mathbf h_{\mathcal G_n}$.
Otherwise, it separates $\mathbf h_{g_m}$ and $\mathbf h_{\mathcal G_n}$ as far away in the semantic space as possible.
Hassani \textit{et al.} \cite{hassani2020contrastivea} extended graph contrastive learning to the multi-view scenario. Here, each graph view is regarded as an independent instance. This work maximizes the mutual information between a graph view and other views of the same graph.
There are several ways to generate new instances for graph contrastive learning, which raises the question of how to choose the most suitable method for the dataset one is working with.
Joint augmentation optimization (JOAO) \cite{you2021grapha} was developed to address this challenge by automating the search for a proper graph data augmentation method.
JOAO trains a probability matrix that can be iteratively updated to select the optimal data augmentation approach. Its performance is competitive.
\subsection{The expressivity of GL-GNNs}\label{expressivity}
As the cutting-edge technology for graph-level learning, people want to explore the power of GL-GNNs for distinguishing graphs \textemdash namely, they want to investigate the expressivity of GL-GNNs.
Practitioners generally employ a representative MPK of the 1-WL algorithm \cite{weisfeiler1968reduction,shervashidze2011weisfeiler} to evaluate the expressivity of standard GL-GNNs (i.e., MPNNs) since MPNNs are the neural network versions of MPKs.
The intimate connection between GL-GNNs and 1-WL is exploited in the Graph Isomorphism Network (GIN) \cite{xu2018powerful}.
This framework shows the upper expressivity bound of an MPNN equals the 1-WL algorithm.
Several research teams have subsequently proved that MPNNs equivalent to 1-WL can not distinguish some substructures in graphs (e.g., cycles, triangles, and Circulant Skip Links) \cite{arvind2020weisfeilera,vignac2020buildinga,murphy2019relationala}.
However, these indistinguishable substructures play a significant role in learning social network and chemical compounds graphs \cite{chen2020cana}.
To break the 1-WL expressivity limitation, the expressivity of GL-GNNs has been empowered through $K$-WL, convolution enhancement, and feature enrichment.
\subsubsection{$K$-WL} A complex variant of 1-WL is the $K$-WL algorithm, which identifies more substructures in graphs by relabeling a set of $K$ vertices.
Morris \textit{et al.} \cite{morris2019weisfeilera} employed MPNNs dealing with $K$-dimensional tensors to apply $K$-WL by neural networks, that is $K$-GNN.
$K$-GNN achieved the expressivity approximately near but slightly weaker than the $K$-WL, but its computational cost increases exponentially with $K$ since it needs to calculate $K$-ranked tensors.
To avoid processing high-dimensional tensors, Provably Powerful Graph Networks (PPGN) \cite{maron2019provablya} adopts a variant of the 2-WL algorithm (i.e., 2-FWL \cite{cai1992optimala}) for designing GL-GNNs and achieves the expressivity over 3-WL.
Further, PPGN replaces the relabel function in 2-FWL with a matrix multiplication based on a single quadratic operation.
Similarly, Folklore Graph Neural Networks (FGNN) \cite{azizian2021expressivea} implements 2-FWL through matrix operations on tensors, pursuing the expressive power as 3-WL.
Despite these common efforts on $K$-WL equivalence GL-GNNs, the majority of them theoretically exceed 1-WL but do not empirically exceed 1-WL \cite{dwivedi2020benchmarking}.
This weak performance by $K$-WL equivalent GL-GNNs is due to two main reasons which are explained next.
\subsubsection{Convolution Enhancement} One reason for the failure of the $K$-WL approaches is that they break the local updates of MPNNs \cite{battaglia2018relationala}, i.e., they no longer update vertices based on neighborhood information.
In practice, GL-GNNs require local updates to preserve the inductive bias property of the graph convolutions \cite{battaglia2018relationala}.
Therefore, some researchers have explored more powerful GL-GNNs by upscaling the graph convolutions yet preserving the local updates.
Alon and Yahav \cite{alon2021ona} noticed that the majority of GL-GNNs do not capture the long-range interactions between nodes because the number of convolutional layers is limited by over-smoothing issues \textemdash that is, the node embeddings tend to be similar after multiple aggregations.
However, long-range interactions can influence the discriminativeness of graphs.
For example, methylnonane is identified by the atoms posited in the compound's two end sides.
To address this issue, these researchers appended a fully linked adjacency matrix to the convolutional layer which aggregates the long-range information without violating any local updates.
Another powerful tool for enhancing the convolution layer is the matrix query language (MATLANG) \cite{brijder2019expressivea,geerts2021expressivea}.
MATLANG strengthens a GL-GNN so that it can recognize more special substructures through its matrix operations, thereby reaching 3-WL expressivity.
Inspired by this work, Balcilar \textit{et al.} \cite{balcilar2021breaking} added MATLANG to the convolutional layer, while Greets and Reutter \cite{geerts2022expressivenessa} evaluated the expressiveness of GL-GNNs through MATLANG instead of the 1-WL algorithm.
\subsubsection{Feature Enrichment} Another reason that $K$-WL methods outperform 1-WL in theory but do not achieve superior performance in experiments is that they ignore the role of node features.
As complementary information for graph structures, node features allow almost all graphs to be discriminated by 1-WL GL-GNNs.
Some practitioners have emphasized that considering node features can also improve the expressiveness of GL-GNNs, rather than just focusing on graph structures.
Murphy \textit{et al.} \cite{murphy2019relationala} annotated a unique position descriptor for each node, that is, sorting all nodes.
Adopting these position descriptors as node features can help a 1-WL GL-GNN to better handle featureless graphs and identify more structures.
To maintain permutation-invariant of graph-level learning, all permutations of node order should be enumerated and the average results should be taken.
Similarly, Colored Local Iterative Procedure (CLIP) \cite{10.5555/3491440.3491734a} sorts the nodes in a substructure and gives them a local position descriptor for feature enrichment.
In addition, both Sato \textit{et al.} \cite{sato2020surveya} and Abboud \textit{et al.} \cite{DBLP:conf/ijcai/AbboudCGL21a} insert random features into nodes giving rise to stable and powerful GL-GNNs.
\subsubsection{High-order Neural Networks} Recently, researchers have tried to improve the expressivity of GL-GNNs through algebraic topology.
This is because equipping graphs as a geometric structure can preserve more valuable properties. Cellular GL-GNNs \cite{bodnar2021weisfeiler_icmla,bodnar2021weisfeiler_nipsa} perform MPNNs on cell complexes, an object including hierarchical structures (e.g., vertices, edges, triangles, tetrahedra). By replacing graphs with cell complexes, cellular GL-GNNs benefits from the better computational fabric for larger expressivity. Furthermore, sheaf neural networks \cite{bodnar2022neurala,barbero2022sheafa} decorate a graph with a geometrical structure, sheaf, which constructs vector space for each node and edge and applies linear transformations among these spaces.
A correct sheaf setting will allow an MPNN to pass messages along a richer structure.
Thus, linearly separate embeddings can be learned, which will enhance the expressive power of GL-GNNs.
\subsection{The generalizability of GL-GNNs}\label{Generalizability}
Real-world applications with graph data tend to involve complex scenarios, such as needing to train a model with only a small amount of labeled data that can ultimately perform well with a large-scale unlabeled test (i.e., size shift) or using only a few labeled training graphs to fit the bulk of unlabeled test graphs. The ability to generalize GL-GNNs is hence a crucial aspect of dealing with these challenges.
\subsubsection{Size Generalization}
Sinha \textit{et al.} \cite{sinha2020evaluatinga} stress the importance of generalizing GL-GNNs and presented evaluation criteria for this.
Xu \textit{et al.} \cite{xu2021howa} theoretically explain that GL-GNNs have better size generalization capabilities than MLPs and can extrapolate trained models to test data that is different from the training set.
To this end, they presented a trick for MPNNs where the graph's vertices are updated by minimizing the aggregated information instead of through summation.
This trick improves generalization ability by altering the learning process from one that is non-linear to one that is linear.
Yehudai \textit{et al.} \cite{yehudai2021locala} theoretically and empirically found the generalization ability of GL-GNNs as the discrepancies in substructures between large and small graphs grows.
To solve this problem, they forced the GL-GNN to pay more attention to the substructures that are hidden in large unlabeled graphs but rarely appear in small labeled graphs.
SizeShiftReg \cite{buffelli2022sizeshiftrega} constrains GNNs to be robust to size-shift through a regularization approach.
SizeShiftReg coarsens the input graph and minimizes the discrepancy between the distribution of the original and coarsened graph embeddings.
\subsubsection{Few-shot Learning} In considering few-shot learning scenarios, Ma \textit{et al.} \cite{ma2020adaptivea} found that there are also differences in the substructures between a few labeled graphs and a large number of unlabeled graphs.
This is because a statistical sample of the training data is too small to represent the substructural distributions of the whole dataset.
Thus, they paid more attention to capturing substructures in unseen unlabeled graphs.
Chauhan \textit{et al.} \cite{Chauhan2020FEW-SHOTa} clusters graphs based on their spectral properties, to produce super-class graphs.
Graph-level representations can be learned from super-class graphs as they have excellent generalization.
\subsection{The explainability of GL-GNNs}\label{Explanation-GNN}
The black-box nature of deep neural networks limits applicability of GL-GNNs to situations where trust is not an absolutely crucial requirement.
Making GL-GNNs explain their predictions in a way that is more interpretable to humans is therefore of great significance to extending the research of GL-GNNs.
Studies on GL-GNNs need to shed insights into how they handle node features and topologies when it comes to predictions.
They also need to more clearly demonstrate how the models identify significant subgraphs and features.
Methods to explain GL-GNNs can be roughly divided into two categories.
One group involves methods that explain the prediction of each input graph;
the other group of methods captures common patterns in the predictions of a set of graphs as explanations.
\subsubsection{A Single Graph} There are three ways to understand GL-GNNs predictions based on a single graph: they can be perturbation-based, model-proxy-based, or gradient-based.
Perturbation-based methods mask nodes, edges, or substructures in the input graph to generate new predictions.
These are then compared to the original input prediction to highlight the important features or structures influencing the GL-GNNs.
For example, GNNExplainer \cite{ying2019gnnexplainera} masks nodes and edges by changing the feature and adjacency matrices, to form masked graphs.
An input graph and its masked graphs are predicted by a trained GL-GNN, while GNNExplainer aims to find the masked graphs with maximized mutual information between its' prediction and the input's prediction.
This found masked graph is the one that preserves the most significant substructures to the GL-GNN's given prediction.
Alternatively, SubgraphX \cite{yuan2021explainabilitya} samples a group of nodes' neighborhoods as subgraphs.
A trained GL-GNN is then used to compute Shapley values \cite{kuhn1953contributionsa} for all the sampled subgraphs. These values represent each subgraph's contribution to the GL-GNN's prediction.
PGExplainer \cite{luo2020parameterized} trains an MLP to determine which edges are valuable to a GNN's prediction and then removes any irrelevant edges to form a new graph. Subsequently, the original and the newly-formed graph are fed into a trained GL-GNN so as to optimize an MLP by maximizing the mutual information between their predictions.
Model-proxy-based methods utilize a simpler surrogate model to approximate the predictions of GL-GNNs.
PGM-Explainer \cite{vu2020pgma} adopts an explainable Bayesian network \cite{pearl1988markova} to calculate the relationship dependencies between nodes, so as to generate a probability graph which describes the input graph.
Gradient-based approaches measure the importance of different input features by back-propagating the gradients of the neural networks.
Gradient-weighted Class Activation Mapping (Grad-CAM) \cite{pope2019explainabilitya}, for example, takes the gradient value of each node embedding in a graph classification task as a measurement of the nodes' significance to the GL-GNN's prediction.
Grad-CAM then measures this subgraph's importance to the prediction by averaging the gradient values of all node embeddings from the subgraph.
\subsubsection{A Set of Graphs} What is common to all the above methods is that they can only learn independent explanations for each instance of a graph \cite{ying2019gnnexplainera,yuan2020explainabilitya}.
However, often the predictions of GL-GNNs made by GL-GNNs are based on a set of graphs.
Thus, understanding the rules or graph patterns that a GL-GNN mines from a set of graphs can provide high-level and generic insights into the explainability of GL-GNNs.
XGNN \cite{yuan2020xgnn} employs a reinforcement learning guided graph generator that generates a graph pattern for different graphs in the same class.
The graph generator is trained via policy gradient to maximize the certain label prediction \cite{sutton1999policya}.
\section{Graph Pooling}
Table \ref{table_graphpooling} summarizes the graph pooling approaches introduced in this section.
\begin{table*}[]
\centering
\caption{Summary of Graph Pooling.}\label{table_graphpooling}
\begin{tabular}{clllll}
\hline
Model & Year & Method & Venue & Language & Code Repository \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Global-\\ Numeric\end{tabular}} & 2020 & PNA\cite{NEURIPS2020_99cad265a} & NeurIPS & Python-Pytorch & https://github.com/lukecavabarrett/pna \\ \cline{2-6}
& 2020 & TextING\cite{zhang2020everya} & ACL & Python-Tensorflow & https://github.com/CRIPAC-DIG/TextING \\ \cline{2-6}
& 2020 & SOPOOL\cite{wang2020seconda} & TPAMI & Python-Pytorch & https://github.com/divelab/sopool \\ \hline
\begin{tabular}[c]{@{}c@{}}Global-\\ Attention\end{tabular} & \multicolumn{1}{c}{2016} & Set2Set\cite{set2seta} & ICLR & Python-Pytorch & https://github.com/pyg-team/pytorch\_geometric \\ \hline
\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Global-\\ CNN\end{tabular}} & 2016 & PATCHYSAN\cite{niepert2016learning} & ICML & Python & https://github.com/tvayer/PSCN \\ \cline{2-6}
& 2018 & KCNN\cite{nikolentzos2018kernel} & ICANN & Python-Pytorch & \begin{tabular}[c]{@{}l@{}}https://github.com/giannisnik/cnn-graph-classification\end{tabular} \\ \hline
\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Global-\\ Top K\end{tabular}} & 2018 & SortPool\cite{zhang2018end} & AAAI & Python-Pytorch & https://github.com/muhanzhang/pytorch\_DGCNN \\ \cline{2-6}
& 2020 & GSAPool\cite{zhang2020structure} & WWW & Python-Pytorch & https://github.com/psp3dcg/gsapool \\ \hline
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Hierarchical-\\ Clustering\end{tabular}} & 2014 & DLCN\cite{ICLR2014BRUNA} & ICLR & - & - \\ \cline{2-6}
& 2015 & GraphCNN\cite{henaff2015deep} & Arxiv & Python-Tensorflow & https://github.com/mdeff/cnn\_graph \\ \cline{2-6}
& 2018 & DiffPool\cite{ying2018hierarchical} & NeurIPS & Python-Pytorch & https://github.com/RexYing/diffpool \\ \cline{2-6}
& 2019 & EigenPool\cite{ma2019grapha} & KDD & Python-Pytorch & https://github.com/alge24/eigenpooling \\ \cline{2-6}
& 2020 & StructPool\cite{yuan2020structpool} & ICLR & Python-Pytorch & https://github.com/Nate1874/StructPool \\ \cline{2-6}
& 2020 & MinCutPool\cite{bianchi2020spectral} & ICML & Python-Tensorflow & \begin{tabular}[c]{@{}l@{}}https://github.com/FilippoMB/Spectral-Clustering\\ -with-Graph-Neural-Networks-for-Graph-Pooling\end{tabular} \\ \cline{2-6}
& 2020 & GMT\cite{baek2020accurate} & ICLR & Python-Pytorch & https://github.com/JinheonBaek/GMT \\ \hline
\multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}Hierarchical-\\ Top K\end{tabular}} & 2018 & SHGC\cite{cangea2018towardsa} & Arxiv & Python-Tensorflow & https://github.com/HeapHop30/hierarchical-pooling \\ \cline{2-6}
& 2019 & U-Nets\cite{gao2019graph} & ICML & Python-Pytorch & https://github.com/HongyangGao/Graph-U-Nets \\ \cline{2-6}
& 2019 & SAGPool\cite{lee2019self} & ICML & Python-Pytorch & https://github.com/inyeoplee77/SAGPool \\ \cline{2-6}
& 2020 & ASAP\cite{ranjan2020asap} & AAAI & Python-Pytorch & https://github.com/malllabiisc/ASAP \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Hierarchical-\\ Tree\end{tabular}} & 2017 & MoNet\cite{monti2017geometric} & CVPR & Python-Pytorch & https://github.com/dmlc/dgl/tree/master/examples/mxnet/monet \\ \cline{2-6}
& 2019 & EdgePool\cite{diehl2019edgea} & Arxiv & Python-Pytorch & \begin{tabular}[c]{@{}l@{}}https://github.com/pyg-team/pytorch\_geometric\\ /blob/master/torch\_geometric/nn/pool/edge\_pool.py\end{tabular} \\ \cline{2-6}
& 2022 & HRN\cite{wu2021structurala} & IJCAI & Python & https://github.com/Wu-Junran/HierarchicalReporting \\ \hline
\end{tabular}
\end{table*}
\subsection{Global Graph Pooling}
\subsubsection{Numeric Operation}
Duvenaud \textit{et al.} \cite{duvenaud2015convolutional} empirically proved that, in graph-level learning, summation has no weaker an outcome than a hash function.
Similarly, GIN \cite{xu2018powerful} shows us that the injective relabeling function in the WL algorithm can be replaced with a simple numeric operation.
Further, GIN also allows us to analyze the efficacy of different functions: summation, maximization, and mean functions.
Summation comprehensively summarizes the full features and structure of a graph.
Maximization emphasizes significant node embeddings, and mean learns the distribution of labels.
Inspired by GIN, Principal Neighbourhood Aggregation (PNA) \cite{NEURIPS2020_99cad265a} employs all three of these functions to pool the node embeddings, while TextING \cite{zhang2020everya} includes both mean and maximization pooling to capture the label distribution and strengthen the keyword features.
A few variants of graph pooling have also been developed.
For example, Deep Tensor Neural Network (DTNN) \cite{schutt2017quantum} applies a neural layer that processes the node embeddings before the summation function and second-order pooling (SOPOOL) \cite{wang2020seconda} is executed as $\mathbf h_{\mathcal G} = [\mathbf h^T_{v_1}\mathbf h_{v_1}|...|\mathbf h^T_{v_{|\mathcal V|}}\mathbf h_{v_{|\mathcal V|}}]$.
\subsubsection{Attention-based}
Set2Set \cite{set2seta} is a more complicated attention-based graph pooling model. It learns the attention coefficients of all node embeddings from an ordered sequence generated by LSTM. Although Set2Set handles sequential node embeddings, the order of nodes is determined by an LSTM model without affecting permutation invariance.
\subsubsection{Global Top-$K$}
Graph Self-Adaptive Pooling (GSAPool) \cite{zhang2020structure} is a global top-$K$ graph pooling model that ranks nodes based on summing of feature and structure scores.
The node structure scores are 1-dimensional vectors projected by the graph convolution operations as same as SortPool, while the feature scores are learned by feeding the node features into an MLP.
\subsection{Hierarchical Graph Pooling}
\subsubsection{Clustering-Based}
EigenPool \cite{ma2019grapha} involves a spectral clustering method that coarsens graphs and pools node embeddings into cluster-level embeddings by converting spectral-domain signals.
\subsubsection{Hierarchical Top-$k$}
Cangea \textit{et al.} \cite{cangea2018towardsa} employed the Graph U-nets to coarsen graphs and concatenated the mean and maximum values of node embeddings on the coarsened graphs as graph-level representations.
\subsubsection{Tree-based}
Wu \textit{et al.} \cite{wu2021structurala} uses a structure encoding tree \cite{li2016structurala} for tree-based hierarchical graph pooling.
Structural coding trees compress the hierarchy of a graph into a tree. Here, the leaves are the nodes, the root represents the whole graph, and the other non-leaf nodes are the hierarchical structures (e.g., the communities).
An MLP merges the features of the child nodes in the structure encoding tree, to generate an embedding of the father node.
The result is an embedding of the root node, which serves as a graph-level representation.
Alternatively, EdgePool \cite{diehl2019edgea} scores edges based on the features of the nodes the edges link, eliminating the highest ranked edge by merging its two end nodes.
The features of the newly generated node, which maintains all the neighbors of the original two nodes, are obtained by summing the features of the two merged nodes.
EdgePool falls into the category of being a tree-based hierarchical graph pooling method because it merges two child nodes in a tree into a father node.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1\textwidth]{application.pdf}
\caption{An overview of graph-level learning downstream tasks and their practical applications.}
\label{fapplication}
\end{figure*}
\subsection{The effectivity of Graph Pooling}\label{effectiveness}
As a downstream summarization component of GNNs, graph pooling has attracted a surge of research interest.
However, since graph pooling is so new, much work is required to investigate the effectiveness of all the various graph pooling algorithms.
Mesquita \textit{et al.} \cite{mesquita2020rethinking} conducted controlled experiments to empirically evaluate the effectiveness of clustering-based hierarchical graph pooling.
First, they adopted two opposite strategies for guiding some clustering-based hierarchical graph pooling processes \textemdash specifically, clustering each of non-adjacent and adjacent nodes.
The final results not only show that the two strategies are comparable, they also indicate that off-the-shelf clustering algorithms, which tend to cluster adjacent nodes, fail to improve graph pooling.
As part of the experiments, Mesquita and colleagues also replaced the learnable assignment matrix in DiffPool \cite{ying2018hierarchical} with an immutable probability assignment matrix: uniform, normal and bernoulli distributions were selected.
The experimental results verify that the performance of fixed-probability-assignment-matrix-guided graph pooling is not weaker than that of DiffPool.
Overall, they concluded that the current clustering-based hierarchical pooling may not be particularly effective and matched this will a call for more sanity checks and ablation studies of the current graph pooling algorithms to fairly evaluate their contributions.
Another study on Pooled Architecture Search (PAS) \cite{wei2021pooling} was dedicated to investigating the effectiveness of graph pooling \textemdash this time with different datasets.
The results of the study show that the effectiveness of graph pooling algorithms is data-specific, that is to say, different input data needs to be handled by a suitable graph pooling algorithm.
For this reason, PAS includes a differentiable search method to select the most appropriate graph pooling algorithm for the given input data.
\section{Benchmarks}
\subsection{Datasets}
The graph datasets collected by the group at TUDateset \cite{morris2020tudataset} have been widely used to evaluate graph-level learning approaches.
These graph datasets consist of molecules, proteins, images, social networks, synthetic graphs, and data from many other domains.
However, despite their wide use, they have attracted criticism from some practitioners.
For example, Ivanov \textit{et al}. \cite{ivanov2019understandinga} contends that the sets suffer from isomorphism bias, i.e., they contain isomorphic graphs with different labels, which may hinder model training \textemdash a claim based on the analysis of 54 widely-used graph datasets.
They also note that some of the datasets are too small to train a data-hungry deep learning model.
For example, Dwivedi \textit{et al.} \cite{dwivedi2020benchmarking} presented that most GL-GNNs have a close performance to others in the small dataset.
Further, some topology-agnostic baselines yield a performance that is competitive to GL-GNNs.
Developing practical and large-scale benchmark datasets has become an important issue for the graph-level learning community.
To this eend, Wu et al. \cite{wu2018moleculenet} proposed a benchmark named MOLECULENET that contains a set of large-scale graph datasets of molecules.
The dataset is designed to be used for graph regression and classification tasks. Dwivedi \textit{et al.} \cite{dwivedi2020benchmarking} transformed images into graphs for classification, in which a group of pixels is clustered as a node. Based on real-world cybersecurity scenarios, Freitas \textit{et al.} \cite{freitas2021a} proposed a large-scale graph dataset of over 1.2 million graphs with imbalanced labels.
Furthermore, OGB \cite{hu2020open} has published application-oriented large-scale graph datasets of molecules, proteins, and source code cooperation networks.
\section{Downstream Tasks and Applications}
Fig. \ref{fapplication} summarizes some common downstream tasks and applications in graph-level learning.
\section{Future Directions}
\subsection{Graph-level Fairness Learning}
The bias in data can easily lead to issues with fairness, where machine learning models make discriminatory predictions towards certain demographic groups based on sensitive attributes such as race.
One feasible solution to debiasing the data is to conduct a competitive game between a biased and a debiased encoder.
The game is won when the fairness-aware debiaser is able to cheat its competitor \cite{bose2019compositionala,masrour2020burstinga}.
Other algorithms add constraints to the loss function to counterbalance model performance with fairness \cite{kang2020informa,li2020dyadica}.
\emph{Opportunities:}
Most work on improving the fairness of models have involved node-level tasks and single graphs \cite{dong2022editsa}.
However, injecting an awareness of fairness into graph-level learning algorithms is also critical work.
Some graph-level learning tasks, such as disease prediction and fraud detection, demand fair results if they are to accurately guide people's decision-making.
One challenge to be overcome in attempting to make graph-level learning fair is that the representative GL-GNNs, i.e., MPNNs, will tend to produce unfair predictions in the face of data bias because the message passing mechanisms actually spread the bias via neighborhood structures \cite{dai2021saya}.
We refer readers who are interested in this topic to \cite{dai2022comprehensivea}, which gives an exhaustive introduction to fairness learning with graph-structured data.
\end{appendices}
|
\section{Introduction}
The fundamental role of center vortices in the ground-state vacuum fields governing the confinement of quarks and chiral symmetry breaking has been shown in Lattice QCD in the past decades~\cite{tHooft:1977nqb, tHooft:1979rtg, CORNWALL, DelDebbio:1996lih, Faber:1997rp, DelDebbio:1998luz, Bertle:1999tw, Faber:1999gu, Engelhardt:1999xw, Bertle:2000qv, Greensite:2003bk, Engelhardt:2003wm, Greensite:2016pfc}. The \textit{center vortex model} for quark-confinement assumes that closed quantised color-magnetic flux tubes percolating the QCD vacuum are responsible for a non-vanishing string tension. It explains confinement, topological charge and chiral symmetry breaking~\cite{deForcrand:1999ms,Alexandrou:1999vx,Reinhardt:2000ck,Engelhardt:2002qs,Bornyakov:2007fz,hollwieser:2008tq,Hollwieser:2011uj,Schweigler:2012ae,Hollwieser:2012kb,Hollwieser:2013xja,Hoellwieser:2014isa,Hollwieser:2014mxa,Hollwieser:2014lxa,Greensite:2014gra,Hollwieser:2015koa,Trewartha:2015nna,Hollwieser:2015qea,Altarawneh:2015bya,Altarawneh:2016ped,Trewartha:2017ive,Faber:2017alm,Biddle:2018dtc} and reproduces the expected behavior of the string tension throughout a wide range of $\beta$-values, although the detection of center vortices in field configurations can proof difficult.
The main methods to detect center vortices are based on Maximal Center
Gauge (MCG)~\cite{DelDebbio:1996mh,Langfeld:1997jx,Langfeld:2003ev} via maximizing an appropriate gauge functional $R_\text{MCG}$, usually restricted by an upper and lower limit which we can consider without loss of generality as $0 \leq R_\text{MCG} \leq 1$. It is not possible to numerically identify a global maximum and one can only consider an ensemble of gauge copies $\{ \mathcal{C}_1, \mathcal{C}_2, \mathcal{C}_3, ..., \mathcal{C}_n \}$ which can be seen as a partially ordered set $ \mathcal{C}_1 \leq \mathcal{C}_2 \leq \mathcal{C}_3 \leq ... \leq \mathcal{C}_n $ with respect to $R_\text{MCG}$. During an optimization procedure the physical interpretation of consecutive gauge copies can vary, but differences should decrease with the gauge functional approaching its maximal value. If they do not, one speaks of \textit{Gribov ambiguities} or \textit{Gribov problems} and care has to be taken when interpreting such gauge configurations.
We were able to solve the problems by usage of non-trivial center regions \cite{Rudolf_Golubich_93027173,Rudolf_Golubich_89891800,Golubich_2021,Golubich_2020,Rudolf_Golubich_68347507,Rudolf_Golubich_65935846,golubich2022improvement}. Our solution does not perform the gauge fixing solely on a microscopic level (taking only single lattice points and the links attached to them into account) but requires the non-trivial center regions, macroscopic structures, to guide the gauge fixing procedure.
In this work we present a different approach that extracts the quark-antiquark potential from an ensemble of gauge copies, the \textit{Ensemble averaged Maximal Center Gauge} (EaMCG). Motivated by our findings in \cite{Rudolf_Golubich_93501217} we look at a reasonable large ensemble of MCG copies with respect to the value of the gauge functional and the string tension. The method starts with the production of the ensemble and executes on each single gauge copy a gradient climb towards the nearest local maximum of the gauge functional. After center projection, average values of Wilson loops $\langle W_\textrm{CP}\rangle$ and the quark-antiquark potentials are determined. This new method is compared in SU(2) gauge theory to other center gauges such as \textit{adjoint Laplacian Landau gauge} (ALLG) and \textit{direct Laplacian Landau gauge} (DLCG) and to the potentials extracted from the unprojected (original) field configurations. The expected scaling behavior as predicted by renormalization group theory is reproduced.
\section{Methods and Formalism: From Center gauges to string tension in lattice-QCD}
We work on periodic lattices with $U_\mu(x) \in \text{SU(2)}$ denoting a gluonic link at lattice site $x$ pointing in direction $\mu$, where Wilson action
\begin{equation}
S_\text{gluons} = \beta \sum_{x,\mu<\nu}\left( 1- \frac{1}{2}\mathfrak{Re}\text{Tr}(U_{\mu\nu}(x)) \right)
\end{equation}
with $U_{\mu\nu}(x)$ corresponding to a plaquette in the plane spanned by the directions $\mu$ and $\nu$ is used in a Monte Carlo procedure. For the inverse coupling $\beta = \frac{4}{g^2}$ we choose $\beta=2.3, 2.4, 2.5, 2.6, 2.7$.
Center vortices are usually detected in \textit{direct maximal center gauge} (DMCG) that uses a gradient climb to identify the gauge transformation $g(x) \in \text{SU(2)}$ maximizing the gauge functional
\begin{equation}\label{GaugeFunctR}
R_\text{MCG}=\sum_x\sum_\mu\mid\text{Tr}[{}^gU_\mu(x)]\mid^2
\end{equation}
with ${}^gU_\mu(x)=g(x+\hat\mu)U_\mu(x)g^\dagger(x)$. Starting on multiple random gauge copies this procedure is executed and the gauge copy with the largest value of the gauge functional is taken for further processing: The link variables ${}^gU_{\mu}(x)$ are projected to the center degrees of freedom, that is $\pm1$ for SU(2),
\begin{equation}\label{CentrProj}
{}^gU_\mu(x)\rightarrow Z_\mu(x)\equiv \mathrm{sign Tr}[{}^gU_\mu(x)].
\end{equation}
The resulting non-trivial center projected plaquettes are known as \textit{P-plaquettes}, $U_\Box=-1$, which contain one or three non-trivial links. The duals of P-plaquettes form closed surfaces in the four dimensional dual lattice and the dual P-vortices correspond to the closed flux lines evolving in time. Center projection is the common ground of center detection methods, only the specific gauge fixing procedures vary.
Since $\text{Tr}U\,\text{Tr}U^\dagger=\text{Tr}U_A+1$, where
\begin{equation}\label{UA}
[U^A]_{ij} = \frac{1}{2}\text{Tr}[\sigma_i U\sigma_j U^\dagger]
\end{equation}
is the adjoint representation, MCG also maximizes
\begin{equation}\label{RAL}
R_\text{AL}=\sum_{x,\mu}\text{Tr}_A[{}^gU_\mu^A(x)],
\end{equation}
the gauge functional of the Landau gauge in the adjoint representation, which is blind to center elements. Landau gauge finds the configuration closest to a trivial field configuration. Hence, the functionals $R_\text{MCG}$ and $R_\text{AL}$ may fail to detect vortices in smooth field configurations, especially in the continuum limit. As was raised by Engelhardt and Reinhardt \cite{Engelhardt:1999xw} in the context of continuum Yang-Mills theory a thin vortex gets singular in this limit and must fail to approach a smooth configuration.
Gauges not directly suffering from Gribov copies are Laplacian gauges which require to solve the lattice Laplacian eigenvalue problem. They were first applied in Laplacian Landau gauge by Vink and Wiese~\cite{Vink:1992ys} for the detection of Abelian monopoles and modified to Laplacian center gauge in \cite{Alexandrou99,deForcrand:2000pg} for center vortices which requires solving the eigenvalue problem of the lattice Laplacian in the adjoint representation
\begin{equation}\label{lapoperator}
D_{ij}(x,y) = \sum_{\mu}
\left(2\delta_{xy}\delta_{ij}-[U^A_\mu(x)]_{ij}\,\delta_{y,x+\hat{\mu}}
-[U^A_\mu(x-\hat\mu)]_{ji}\,\delta_{y,x-\hat\mu}\right).
\end{equation}
The three lowest eigenvectors of this Laplacian are combined to matrices $M(x)$ in \textit{adjoint Laplacian Landau gauge} (ALLG) which maximizes the functional
\begin{equation}\label{GaugeFunct}
R_\text{ALLG}= \sum_x\sum_\mu\text{Tr}[M^\dagger(x)U^A_\mu(x)M(x+\hat\mu)]
\end{equation}
with $M(x) \in SO(3)$ "on average", that is, with weakened orthogonality constrained $\langle M^\dagger \cdot M \rangle = 1$. Even if the solution of the eigenvalue problem is a unique procedure, small modifications of the configuration may lead to different eigenfunctions corresponding to different positions of vortices. The combination of ALLG followed by DMCG~\cite{Faber_2001,Faber_2002} is referred to as \textit{direct Laplacian center gauge} (DLCG).
In the Monte-Carlo runs on a $24^4$ lattice we did for each $\beta\in\{2.3, 2.4, 2.5, 2.6, 2.7\}$ 20 random starts, 3000 initial sweeps and produced 1000 configurations with a distance of 100 sweeps. For these $20\,000$ configurations we determined the averages of unprojected Wilson loops. Due to the large computational needs we did center projection by ALLG, DLCG and EaMCG for every 100th of these $20\,000$ configurations only. For EaMCG we did 100 unbiased gauge copies for each of the 200 configurations. In Fig.~\ref{fig:stringtension2} we compare the $24^4$ data for unprojected configurations with the same number of analogously produced $32^4$ configurations.
\subsection*{The problem of the gauge functional and its resolution}
The gauge functional~(\ref{GaugeFunctR}) has a huge number of local maxima and it is practically impossible to find the absolute maximum. The numeric maximization of the gauge functional suffers from a severe problem as was shown by Bornyakov, Komarov, and Polikarpov (BKP) in ref.~\cite{Bornyakov_2001} with simulated annealing: The largest maxima of the functional underestimate the density of vortices and therefore also the string tension. Since there are only relatively few of these maxima it is rather tiresome to find them by simulated annealing.
Random gauge copies are usually determined by random gauge matrices $g(x)$ with a homogeneous density on the SU(2) group manifold, as produced by the Box-Muller method~\cite{BoxMuller}. One can get directly to consecutively increasing local maxima of the gauge functional by a bias in these random gauge transformations by restricting the gauge transformation to the hemisphere of SU(2)$\cong\mathbb S^3$ with positive trace. In the two diagrams in Fig.~\ref{fig:biasedgauge} we show the maximizing histories for two independently created field configurations (out of 200) and for ten of their random gauge copies. For each of these 20 field configurations we perform 100 times the following two steps: First, we maximize the functional by gradient climb. Secondly, we apply the biased random gauge transformations. For center projected field configurations of sufficient large lattice size already one configuration can result in reasonable values of Creutz-ratios
\begin{equation}\label{Eq:Creutz}
\sigma\approx\chi_\textrm{CP}(R)=-\ln\frac{W_\textrm{CP}(R+1,R+1)\;W_\textrm{CP}(R,R)}{W_\textrm{CP}(R,R+1)\;W_\textrm{CP}(R+1,R)},
\end{equation}
for small Wilson loops $W_\textrm{CP}(R,T)$ of size $R \times T$, which are averages over the center projected configurations of the ensemble. In Fig.~\ref{fig:biasedgauge} we show $\chi_\textrm{CP}-R_\text{MCG}$ pairs for the evolution through the biased gauge fixing iterations. A nearly linear relation between $\chi_\textrm{CP}$ and $R_\text{MCG}$ can be observed. With increasing gauge functional we observe a linear decrease of the string tension, underestimating the expected value of $\sigma=0.0350(12)$~\cite{Bali1994}.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.45\linewidth]{img/verglCreutz2L24b25r0H2}
\includegraphics[width=0.45\linewidth]{img/verglCreutz2L24b25r1H2}
\caption{In the two diagrams we show the maximizing histories of two independent field configurations on a $24^4$ lattice for ten random gauge copies each. Interchanging hundred times gradient climbs and biased random gauge transformations we get hundred points in the $\chi_\textrm{CP}-R_\text{MCG}$ plane connected by a line. Observe the nearly linear relation between Creutz ratio $\chi_\textrm{CP}$ and gauge functional $R_\text{MCG}$ and especially the final low value of $\chi_\textrm{CP}$, much lower than the expected string tension $\sigma=0.035$.}
\label{fig:biasedgauge}
\end{figure}
With this method we did not find any configuration where the maximum of the gauge functional did not dramatically underbid the string tension. On the other hand, DMCG~\cite{DelDebbio:1998luz,Hollwieser_2008} predicts the string tension very successfully using a few gauge copies and choosing the one with the highest value of the functional for further analysis. This demonstrates the physical relevance of the gauge functional and implies that it is the prescription for global maximization which is problematic. In DLCG~\cite{Faber_2001,Faber_2002} the mentioned problems were circumvented using the eigenfunctions of the adjoint Laplacian operator to select a smooth gauge field before maximizing the gauge functional in a gradient climb. In this article we investigate a different approach: \textit{ensemble averaged maximal center gauge} (EaMCG). We produce random gauge copies with the correct $\mathbb S^3$-weight, approach the next local maximum by the gradient method and take the average of an ensemble of such random gauge copies. The idea is that not the best local maximum alone carries the physical information but the average over all local maxima does: maxima with a higher value of the gauge functional result in a reduced string tension, but they do not dominate the ensemble. The same holds for lower valued maxima, possibly overestimating the string tension. This method is strengthened by density plots of 20\,000 $R_\text{MCG}/\chi_\textrm{CP}(2)$ pairs in Fig.~\ref{fig:densityPlots} for $\beta\in\{2.3,2.4,2.5,2.6,2.7\}$. For all $\beta$-values we see a clear linear correlation as expected from Fig.~\ref{fig:biasedgauge}. The average slope of the distributions agreees in both figures. The $\beta$-dependence of the elliptic distributions is a running coupling effect. With increasing $\beta$, higher Creutz ratios can be resolved, but give broader distributions due to larger statistical errors.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{img/allPltsR2}
\caption{Correlations between Creutz ratios $\chi_\textrm{CP}(2)$ and the values of the maximized gauge functionals $R_\text{MCG}$ for center projected configurations for $\beta\in\{2.3,2.4,2.5,2.6,2.7\}$. On a $24^4$-lattice we produced the density plots with 20 random starts and 10 configurations with a distance of 10000 MC iterations and 100 gauge copies each. For each $\beta$ the average values are marked by a star.}
\label{fig:densityPlots}
\end{figure}
The average values of the distributions are marked by a ``$\star$'' and give first estimates of the string tension. From the type of distributions shown in Fig.~\ref{fig:densityPlots} it gets understandable that the gauge configurations with extremely high functional~(\ref{GaugeFunctR}) do not really contribute to the average. According to Eq.~(\ref{Eq:Creutz}) Creutz ratios are determined from almost symmetric Wilson loops which are polluted by higher excitations and give therefore less confidential values for the string tension than the potential, extracted from maximally asymmetric Wilson loops with good signals as we will describe in the next section.
\subsection*{Extracting the string tension from Wilson loop data with the 1-exp-fit}
We extract the static quark-antiquark potential from asymmetric $R\times T$ Wilson loops of an ensemble of configurations
\begin{equation}\label{Wexp}
\langle W(R,T)\rangle=\sum^\infty_{i=0}c_i\mathrm e^{-\varepsilon_i(R)T}.
\end{equation}
Only for the lowest states the amplitudes $c_i$ and energies $\varepsilon_i$ are resolvable. In Fig. \ref{fig:asymwils} the logarithms of Wilson loops are shown at $\beta=2.5$ comparing a 3-exp fit of original Wilson loops, a 2-exp fit of ALLG center projected loops, a 1-exp fit of DLCG based loops and 1-exp fit of EaMCG averaged loops. The curvature of original Wilson loops at small $T$ is due to excited states that are almost eliminated in DLCG and EaMCG.
\begin{figure}[h!]
\centering
\parbox{0.48\linewidth}{
\centering
\includegraphics[width=0.98\linewidth]{img/asymfull}}
\hspace{1mm}
\parbox{0.48\linewidth}{
\centering
\includegraphics[width=0.97\linewidth]{img/asymallg}}
\parbox{0.48\linewidth}{
\centering
\includegraphics[width=0.97\linewidth]{img/asymdlcg}}
\hspace{1mm}
\parbox{0.48\linewidth}{
\centering
\includegraphics[width=0.97\linewidth]{img/asymeamcg}}
\caption{\small
A comparison of logarithms of Wilson loop data is depicted at $\beta=2.5$ for original and projected configurations using different gauge fixing procedures. Observe that excited states contribute for large $R$ essentially for unprojected configurations, less for ALLG and very little for DLCG and EaMCG.}
\label{fig:asymwils}
\end{figure}
The larger the excitation energy, the more are excited states exponentially suppressed with increasing $T$. Finally, when the Wilson loops remain resolvable, only the ground state survives. To determine the ground state energy $\varepsilon_0(R)$ we introduce for each $R$ an initial $T_i$ and define a one-exponential fitting function (\textit{1-exp fit}) for Wilson loops with $T \geqslant T_i$ as
\begin{equation}\label{groundeps}
\langle W(R,T)\rangle = c(R,T_i) e^{-\varepsilon_0 (R,T_i) T}.
\end{equation}
By removing small $T$s we observe that $\varepsilon_0 (R,T_i)$ gets nearly stable in a fit range $T_i\approx T_\text{opt}$, the plateau region. $T_\text{opt}$ is a function of $R$ and therefore the plateau region has to be investigated for each $R$ separately.
Keeping numerical issues under control proofs quite difficult: The Wilson loop averages from a single configuration are disturbed by noise, hence multiple configurations are averaged into $\langle W(R,T) \rangle$. The data is still disturbed by systematic errors resulting from unwanted higher excitations and statistical errors with increasing loop sizes: a decrease of the systematic error is accompanied by an increase of the statistical error. Therefore, the optimal $T_i$ value, $T_\text{opt}$, we determine increasing $T_i$ from the first local minimum of an error quantifier \cite{Rudolf_Golubich_93501217},
\begin{equation}\label{err}
\text{Err} = \frac{2}{3} \langle \Delta_{\delta i} \rangle + \frac{1}{3} \langle \Delta_\text{err} \rangle
\end{equation}
with $\langle \Delta_{\delta i} \rangle$ being the average change of the fit variable $\varepsilon_0 (R,T_i)$ from previous to present $T_\text{opt}$ and from present to next $T_\text{opt}$ and $\langle \Delta_\text{err} \rangle$ being the average error of the respective three data points $\varepsilon(R,T_{i-1})$, $\varepsilon(R,T_{i})$ and $\varepsilon(R,T_{i+1})$. Fig.~\ref{fig:1ExpFitsb25} depicts some examples for the $T_i$ dependence of $\varepsilon_0(R,T_i)$ and the chosen $T_\text{opt}$ for $\beta=2.5$ and some $R$ for unprojected SU(2) configurations, ALLG, DLCG and EaMCG. For unprojected configurations and also for ALLG the excited states contribute at small distances and $\varepsilon_0 (R,T_\text{opt}(R))$ is therefore approached from above. We observed in Fig.~\ref{fig:asymwils} that center projection for DLCG and EaMCG suppresses the excited states, the variations of $\varepsilon_0$ are therefor much smaller.
\begin{figure}[!htb]
\centering\includegraphics[width=0.45\linewidth]{img/compareFixRb25r2full}
\centering\includegraphics[width=0.45\linewidth]{img/compareFixRb25r2allg}
\centering\includegraphics[width=0.45\linewidth]{img/compareFixRb25r5allg}
\centering\includegraphics[width=0.45\linewidth]{img/compareFixRb25r2dlcg}
\centering\includegraphics[width=0.45\linewidth]{img/compareFixRb25r2eamcg}
\centering\includegraphics[width=0.45\linewidth]{img/compareFixRb25r5eamcg}
\caption{\small
Some examples of the $T_i$ dependence of $\varepsilon_0(R,T_i)$ are shown. The final values at $T_\text{opt}$ are depicted by horizontal lines. For original Wilson loops and ALLG, $\varepsilon_0 (R,T_\text{opt}(R))$ is approached from above for $T_i\rightarrow T_\text{opt}$ due to the decaying contributions of excited states. Observe that for DLCG and EaMCG the variations of $\varepsilon_0(R,T_i)$ are much smaller since the lines in the lower part of Fig.~\ref{fig:asymwils} are remarkably straight. }
\label{fig:1ExpFitsb25}
\end{figure}
Once, $\varepsilon_0(R,T_\text{opt})$ is determined, we can identify it with the potential
\begin{equation}\label{groundpot}
V_W(R) = \varepsilon_0 (R,T_\text{opt}(R))
\end{equation}
extracted from Wilson loops.
\section{Results and Discussion}
\subsection{Center Dominance and Precocious Linearity}
In this section, we compare the potentials extracted from original configurations with those which we get after center projections, especially with the main subject of this article with ensemble averaged maximal center gauge, EaMCG. To extract the physical parameters of the gluonic quark-antiquark string we fit the potentials $V_W(R)$ of Eq.~(\ref{groundpot}) with a Cornell type of potential
\begin{equation}\label{eq:potfitallg}
V(R) = v_0 + \sigma R - \dfrac{c}{R}
\end{equation}
with three parameters, self energy $v_0$, string tension $\sigma$ and Coulomb coefficient $c$. For original and ALLG configurations we need all three parameters. For DLCG and EaMCG we observe the feature of precocious linearity; i.e.\ the fact that the Coulombic term is strongly suppressed. Therefore we choose $c=0$ and start the fit at some $R_i$.
\begin{figure}[h!]
\centering
\includegraphics[scale=.75]{img/potfullallgdlcgeamcgb25}
\caption{\small For $\beta=2.5$ on a $24^4$ lattice, potentials determined with different methods are shown and compared to Ref~\cite{Bali1994}. To ease comparisons, the potentials of EaMCG and DLCG are shifted by the self energy $v_0$ of the potential extracted from unprojected configurations.}
\label{fig:fullshproj2}
\end{figure}
For $\beta=2.5$ we show in Fig. \ref{fig:fullshproj2} the potential $V_W(R)$ of unprojected configurations, the fit potential $V(R)$, its contributions $v_0-\dfrac{c}{R}$ and $v_0+\sigma R$ together with the data of Ref.~\cite{Bali1994}. We compare them with the potentials and their fits resulting from the three center projection methods, ALLG, DLCG and EaMCG. ALLG has a sizeable Coulomb contribution, differing in this respect from DLCG and EaMCG. ALLG leads to confidential values in the same region as $V_W(R)$. The potential is more difficult to resolve than for DLCG and EaMCG. ALLG has a smaller linear term than DLCG, where this term agrees nicely with the linear term of the unprojected configurations. Over the whole $R$ region the EaMCG values are very well resolvable and start at $R_i=3$ to predict a stable value for the string tension $\sigma$. It is interesting to observe in Fig. \ref{fig:fullshproj2} that, despite the linear term is larger than for DLCG and the unprojected configurations, the potential runs at large $R$ nicely parallel to the unprojected potential. Since the Coulomb factor of the unprojected configurations is nearly proportional to $\pi/12$ and thus mainly due to string shape fluctuations this may be an indication that on the finite periodic lattice, in average, vortices do follow the shape fluctuations of the gluonic string. This behavior is reproduced for all tested values of $\beta$ as can be seen in Fig. \ref{fig:fullshproj3}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.47\linewidth]{img/potfullallgdlcgeamcgb23}
\includegraphics[width=0.47\linewidth]{img/potfullallgdlcgeamcgb24}
\includegraphics[width=0.47\linewidth]{img/potfullallgdlcgeamcgb26}
\includegraphics[width=0.47\linewidth]{img/potfullallgdlcgeamcgb27}
\caption{\small For $\beta\in\{2.3, 2.4, 2.6, 2.7\}$ on a $24^4$ lattice, the different methods to extract potentials and string tension are compared. For better visual comparison, the potentials of EaMCG and DLCG are shifted by the self energies of potentials extracted from original configurations.}
\label{fig:fullshproj3}
\end{figure}
The parameters of the potentials are summarized in table \ref{table:potpar}. The indicated errors are the statistical errors only. The systematic errors are difficult to quantify, but they are at least of the same size as the statistical errors as we can see comparing the data.
\begin{table}[!htb]
\centering
\begin{tabular}{l*{6}{c}r}
\hline
\textbf{method} &$v_0$&$\sigma$& $c$ \\
\hline
$\beta=2.3$ \\
\hline
unprojected, $24^4$ &$0.492(8)$ & $0.1496(27)$ & $0.214(6)$ \\
unprojected, $24^4$, fix $c$ & $0.5289(43)$ & $0.1431(21)$ & $\pi/12$ \\
unprojected, $32^4$ & $0.495(5)$ & $0.1490(16)$ & $0.216(3)$ \\
$Z_2$, ALLG & $0.557(66)$ & $0.1508(193)$ & $0.053(48)$ \\
$Z_2$, DLCG &$0.0496(18)$&$0.1247(6)$ & -- \\
$Z_2$, EaMCG & $0.0237(13)$ & $0.1412(4)$ & -- \\
\hline
$\beta=2.4$ \\
\hline
unprojected, $24^4$ &$0.557(15)$ & $0.0730(25)$ & $0.284(20)$ \\
unprojected, $24^4$, fix $c$ & $0.5472(41)$ & $0.0741(12)$ & $\pi/12$ \\
unprojected, $32^4$ & $0.552(10)$ & $0.0728(17)$ & $0.271(13)$ \\
$Z_2$, ALLG & $0.547(22)$ & $0.0608(56)$ & $0.112(16)$ \\
$Z_2$, DLCG & $0.0233(17)$ & $0.0676(3)$ & -- \\
$Z_2$, EaMCG & $-0.0070(21)$ & $0.0811(4)$ & -- \\
\hline
$\beta=2.5$ \\
\hline
unprojected, $24^4$ &$0.543(5)$ & $0.0371(8)$ & $0.266(6)$ \\
unprojected, $24^4$, fix $c$ & $0.5410(16)$ & $0.0373(5)$ & $\pi/12$ \\
unprojected, $32^4$ & $0.552(5)$ & $0.0350(8)$ & $0.276(6)$ \\
$Z_2$, ALLG & $0.413(3)$ & $0.0299(7)$ & $0.081(2)$ \\
$Z_2$, DLCG & $0.01423(9)$ & $0.03550(2)$ & -- \\
$Z_2$, EaMCG & $0.0040(14)$ & $0.0427(3)$ & -- \\
\hline
$\beta=2.6$ \\
\hline
unprojected, $24^4$ &$0.524(4)$ & $0.0191(6)$ & $0.258(5)$ \\
unprojected, $24^4$, fix $c$ & $0.5256(12)$ & $0.0189(3)$ & $\pi/12$ \\
unprojected, $32^4$ & $0.518(3)$ & $0.0200(5)$ & $0.251(4)$ \\
$Z_2$, ALLG & $0.302(2)$ & $0.0197(5)$ & $0.054(2)$ \\
$Z_2$, DLCG & $0.00655(11)$ & $0.01946(2)$ & -- \\
$Z_2$, EaMCG & $0.0013(8)$ & $0.0242(2)$ & -- \\
\hline
$\beta=2.7$ \\
\hline
unprojected, $24^4$ &$0.479(3)$ & $0.0133(5)$ & $0.219(5)$ \\
unprojected, $24^4$, fix $c$ & $0.4951(31)$ & $0.0118(7)$ & $\pi/12$ \\
unprojected, $32^4$ & $0.489(2)$ & $0.0116(3)$ & $0.234(3)$ \\
$Z_2$, ALLG & $0.211(1)$ & $0.0097(2)$ & $0.034(2)$ \\
$Z_2$, DLCG & $0.00822(34)$ & $0.00988(6)$ & -- \\
$Z_2$, EaMCG & $0.0042(2)$ & $0.01406(6)$ & -- \\
\hline
\end{tabular}
\caption{\small The parameters, string tension $\sigma$, self energy $v_0$ and Coulomb factor $c$, of the fitting potential~(\ref{eq:potfitallg}) are compared for the various methods.}
\label{table:potpar}
\end{table}
\subsection{Scaling behavior}
The string tension is expected to be proportional to the square of the lattice spacing $a$. For the SU(2) gauge group in the quenched approximation~\cite{rothe1992} asymptotic freedom predicts in two-loop order
\begin{equation}\label{eq:eq26prd}
a^2 = \dfrac{1}{\Lambda_L^2} (\dfrac{6 \pi^2 \beta}{11})^{\dfrac{102}{121}} \exp({-\dfrac{6 \pi^2}{11}\beta}),
\end{equation}
where $\Lambda_L$ is the standard lattice scale parameter. In Fig.~\ref{fig:stringtension2} we compare the string tensions extracted from unprojected and center projected configurations of Table~\ref{table:potpar} with the asymptotic freedom prediction. The main diagram shows the results of our calculations on a $24^4$ lattice. In the insets we show fits to the asymptotic behavior $a^2\propto\exp({-\dfrac{6 \pi^2}{11}\beta})$, in the upper inset for EaMCG and in the lower inset for unprojected configurations on a $32^4$ lattice which scale much better than $24^4$ data. We can clearly see in the insets that EaMCG and unprojected show the same scaling behavior and that EaMCG scales better than ALLG and DLCG.
\begin{figure}[!htb]
\centering
\includegraphics[scale=.7]{img/stringfunbeta2}
\caption{\small The dependence of string tensions extracted from unprojected and center projected configurations on $\beta$ are compared with the predictions of asymptotic freedom. For each of the methods the lattice scale paramter $\Lambda_L$ is adjusted separately to the data with $\beta \geqslant 2.5$.
}
\label{fig:stringtension2}
\end{figure}
\section{Conclusion}
Center gauge methods suffer from the large differences between the center projected and unprojected configurations. These differences appear especially in the region of center vortices. Maximizing the center blind gauge functional~(\ref{GaugeFunctR}) favors therefore to overlook vortices and to underestimate the string tension. This property is the basis of an approximate linear relation between the gauge functional and the string tension as shown in Figs.~\ref{fig:biasedgauge} and \ref{fig:densityPlots}. This is a clear indication that the request for a global maximization of the gauge functional is not only difficult but erroneous.
Previously, there were two suggestions to overcome this failure of the global maximization condition. The first idea~\cite{Faber_2001,Faber_2002} was to preselect a region of smooth gauge configurations where we maximize the gauge functional. With the three lowest eigenfunctions of the center blind lattice Laplacian in the adjoint representation we defined in ALLG a smooth gauged transformed configuration to start the maximization procedure by a gradient climb and to arrive at DLCG. In~\cite{Rudolf_Golubich_93027173,Golubich_2021,Rudolf_Golubich_68347507} we suggested a completely different, but computationally rather involved approach, where we proposed to use global structures, non-trivial center regions, to guide the gauge fixing procedure. These two successful methods indicate that the failure of Direct Maximal Center Gauge was in the prescription to determine the global maximum of the gauge functional. In the new method suggested in this article, ensemble averaged maximal center gauge (EaMCG), we are using the physical information contained in the huge ensemble of local maxima of the functional. The gluon string between a static quark-antiquark pair with large separation fluctuates leading to a Coulombic contribution proportional to $\pi/12$~\cite{Luscher1980fr}. Projected vortices follow these fluctuations leading to an increase of the P-vortex area compared to a straight string and result in a slightly larger string tension than DLCG. The potential extracted from these areas follows nicely the large distance behavior of the string on a finite periodic lattice.
Recently, an improved method for computing the static quark-anti-quark potential in lattice QCD was investigated in \cite{Hollwieser:2022doy, Hollwieser:2022pov, Hollwieser:2022bqp}, which is not based on Wilson loops, but formulated in terms of (temporal) {\it Laplace trial state correlators}, formed by eigenvector components of the covariant lattice Laplace operator. This approach seems very promising and an application to center vortex configurations as well as a generalization of the methods introduced in this article to $SU(3)$ is left for future investigations.
\section*{Acknowledgements} The authors gratefully acknowledge the Vienna Scientific Cluster (VSC) for providing supercomputer resources. R.H. was supported by the MKW NRW under the funding code NW21-024-A. We thank Sedigheh Deldar for valuable discussions.
|
\section{Introduction}
Given a hypergraph $G$ and a positive integer $r$, let $\mathrm{mc}_r(G)$ be the largest integer $t$ such that every $r$-coloring of the edges of $G$ contains a monochromatic component of order at least $t$. Let $K_n^k$ denote the complete $k$-uniform hypergraph on $n$ vertices (and $K_n = K_n^2$ as usual).
The starting point of this paper is the problem of determining the value of $\mathrm{mc}_r(K_n^k)$ for all $n,r,k\geq 2$. While this problem has been studied since the 70's, it is still open in most cases. For example, the exact value (even asymptotically) of $\mathrm{mc}_r(K_n^k)$ is not known when $k=2$, $r=7$; $k=3$, $r=8$; or $k=4$, $r=6$ to name a few small cases.
As for positive results, Gy\'arf\'as proved the following.
\begin{theorem}[Gy\'arf\'as \cite{Gy}]~\label{thm:Gy}
\begin{enumerate}
\item For all $n\geq r\geq 2$, $\mathrm{mc}_r(K_n)\geq \frac{n}{r-1}$. This is best possible when $(r-1)^2$ divides $n$ and there exists an affine plane of order $r-1$.
\item For all $n\geq r\geq 2$, $\mathrm{mc}_r(K_n^r)=n$.
\item For all $n\geq r\geq 3$, $\mathrm{mc}_{r+1}(K_n^{r})\geq \frac{rn}{r+1}$. This is best possible\footnote{Simply partition a set of $n$ vertices as equally as possible into $r+1$ sets $A_1, \dots, A_{r+1}$. Now consider a complete $r$-uniform hypergraph $K$ with vertex set $A_1\cup \dots A_{r+1}$ where for every edge $e\in E(K)$ we assign it a color $i\in [r+1]$ such that $e\cap A_{i}=\emptyset$ (which must exist since $|e|=r$). Note that for all $i\in [r+1]$, every component of color $i$ has order at most $n-|A_i|\leq n-\floor{\frac{rn}{r+1}}$.} for all $n$.
\end{enumerate}
\end{theorem}
We will mention a few other cases in which the value of $\mathrm{mc}_r(K_n^k)$ is known in the conclusion.
A natural question which has received attention lately has been to determine conditions under which a $k$-uniform hypergraph $G$ on $n$ vertices satisfies $\mathrm{mc}_r(G)= \mathrm{mc}_r(K_n^k)$.
Note that for $1\leq r\leq k$, $\mathrm{mc}_r(G)=n= \mathrm{mc}_r(K_n^k)$ if and only if the $r$-shadow of $G$ is complete (see Section \ref{sec:not} for the definition of $r$-shadow). Indeed, if every $r$-set of $G$ is contained an edge, then since $\mathrm{mc}_r(K_n^r)=n$, we have $\mathrm{mc}_r(G)=n$. Furthermore, if some $r$-set $\{x_1, \dots, x_r\}$ is not contained in an edge, then we can color the edges of $G$ with $r$-colors such that color $i$ is never used on $x_i$ and thus $\mathrm{mc}_r(G)<n$.
On the other hand, as first noted by Gy\'arf\'as and S\'ark\"ozy \cite{GSc}, when $r>k=2$ it is surprisingly possible for $\mathrm{mc}_r(G)= \mathrm{mc}_r(K_n)$ provided $G$ has large enough minimum degree. Improving on results in \cite{GS2} and \cite{DKS}, Guggiari and Scott \cite{GSc} and independently Rahimi \cite{R} proved that if $\delta(G)\geq 5n/6-1$, then $\mathrm{mc}_3(G)\geq \frac{n}{2}$. Then F\"uredi and Luo \cite{FL} proved that for $r\geq 4$, if $\delta(G)\geq (1-\frac{1}{6(r-1)^3})n$, then $\mathrm{mc}_r(G)\geq \frac{n}{r-1}$ (and thus $\mathrm{mc}_r(G)= \mathrm{mc}_r(K_n^k)$ when $(r-1)^2$ divides $n$ and there exists an affine plane of order $r-1$). It is still an open problem to determine the exact minimum degree threshold which guarantees $\mathrm{mc}_r(G)\geq \frac{n}{r-1}$ for $r\geq 4$, but it is known to be at least $(1-\frac{r-2}{r^2-r})n-1$ (see \cite{DK} and \cite{GSc}).
The main purpose of the paper is to consider a generalization of the results mentioned in the previous paragraph to $k$-uniform hypergraphs with $k\geq 3$. Our main result is a strengthening of Theorem \ref{thm:Gy}(iii) for $r$-uniform hypergraphs with sufficiently large $(r-1)$-degree.
\begin{theorem}\label{thm:mindeg}
For all $n\geq r\geq 3$, if $G$ is an $r$-uniform hypergraph on $n$ vertices with
\begin{equation}\label{codeg}
\delta_{r-1}(G)\geq \frac{rn}{r+1}-(r-1),
\end{equation}
then $\mathrm{mc}_{r+1}(G)\geq \frac{rn}{r+1}$. Furthermore, the degree condition is best possible.
\end{theorem}
We prove Theorem \ref{thm:mindeg} in Section \ref{sec:main}.
\subsection{Notation}\label{sec:not}
Given $0\leq \ell\leq k$ and a $k$-uniform hypergraph $G$, we let $$\delta_{\ell}(G)=\min_{S\in \binom{V(G)}{\ell}}|\{e\in E(G): S\subseteq e\}|.$$ When $\ell=k-1$ it is customary to refer to $\delta_{k-1}(G)$ as the minimum \emph{codegree} of $G$.
Given $2\leq r\leq k$ and a $k$-uniform hypergraph $G$, the \emph{$r$-shadow} of $G$ is the $r$-uniform hypergraph on $V(G)$ with edge set $\{f\in \binom{V(G)}{r}: \exists e\in E(G), f\subseteq e\}$.
\section{Large minimum codegree implies large monochromatic components}\label{sec:main}
We begin by giving the example which shows that the degree condition in Theorem \ref{thm:mindeg} is best possible.
\begin{example}
For all $r\geq 3$ and $n\geq 3r+1$, there exists an $r$-uniform hypergraph $G$ on $n$ vertices with $\delta_{r-1}(G)= \ceiling{\frac{rn}{r+1}}-r$ for which $\mathrm{mc}_{r+1}(G)= \ceiling{\frac{rn}{r+1}}-1$.
\end{example}
\begin{proof}
Let $n=k(r+1)+s$ where $0\leq s\leq r$ and note that $$\ceiling{\frac{rn}{r+1}}=\ceiling{\frac{r(k(r+1)+s)}{r+1}}=rk+\ceiling{\frac{rs}{r+1}}=rk+s.$$
Let $A_1, \dots, A_{r+1}$ be a collection of $r+1$ sets such that $|A_i|=k-1$ for all $i\in [r]$ and $|A_{r+1}|=k-(r-s)$. Let $X_{1}, X_{2}, \dots, X_{r}$ be a collection of $r$ sets each of order $2$. Let $V=A\cup X$ where $A:=A_1\cup \dots \cup A_{r+1}$ and $X := X_{1} \cup \dots \cup X_{r}$ and note that $|V|=n$. We begin with a complete $r$-uniform hypergraph on $V$ and we delete all edges which intersect all of $(X_{i})_{i\in P}$, $(A_j)_{j\in Q}$ for some partition $\{P, Q\}$ of $[r]$ (where $P, Q\neq \emptyset$)\footnote{For example, if $r=4$, we would delete all edges which intersect all of $X_1, X_2, A_3, X_4$, or all of $X_1, A_2, A_3, A_4$, and so on}. This is our hypergraph $G$. Note that for all $S\in \binom{V}{r-1}$, $S$ has $(r-1)$-degree at least $$n-(r-1)-(k+1)=kr+s-r=\ceiling{\frac{rn}{r+1}}-r.$$
Now, for all edges $e$ which lie entirely inside $A$, we assign to $e$ any color $j\in [r+1]$ such that $e\cap A_j=\emptyset$ (which must exist since $|e|=r$). For all edges which lie entirely inside $X$, we assign color $r+1$. For all other remaining edges $e$ which intersect both $A$ and $X$, there will -- by construction -- exist $i\in [r]$ such that $e\cap (A_i\cup X_i)=\emptyset$ and we assign $e$ such a color $i$. Now for all $i\in [r]$, we have that there are no edges of color $i$ touching $X_i\cup A_i$ and since $|A_i\cup X_i|=k+1$, the largest component of color $i$ has order at most $n-(k+1)=kr+s-1= \ceiling{\frac{rn}{r+1}}-1$. Finally, the largest component of color $r+1$ has order at most $r(k-1)< \ceiling{\frac{rn}{r+1}}-1$.
\end{proof}
Now we move on to the main part of the proof.
\begin{proof}[Proof of Theorem \ref{thm:mindeg}]
Let $G$ be an $r$-uniform hypergraph on $n$ vertices satisfying \eqref{codeg} and suppose the edges of $G$ have been colored with $r+1$ colors. Let $x\in V(G)$ and for all $i\in [r+1]$, let $C_i$ be the component of color $i$ which contains $x$. Note that even if $x$ is not incident with any edges of color $i$, there is still a trivial component of color $i$ (consisting only of $x$) which contains $x$. Now for all $i\in [r+1]$, let $F_i=V(G)\setminus V(C_i)$. If $|F_i|\leq \frac{n}{r+1}$ for some $i\in [r+1]$, then we are done, so suppose that $|F_i|>\frac{n}{r+1}$ for all $i\in [r+1]$. Let $F=F_1\cup \dots \cup F_{r+1}$.
\begin{claim}\label{clm:3way*}
For all distinct $h,i,j\in [r+1]$, $F_h\cap F_i\cap F_j=\emptyset$.
\end{claim}
\begin{proofclaim}
Suppose for contradiction that $F_h\cap F_i\cap F_j\neq \emptyset$ for some distinct $h,i,j\in [r]$ and without loss of generality suppose $\{h,i,j\}=[3]$. Let $x_{123}\in F_1\cap F_2\cap F_3$ and let $x_i\in F_i$ for all $4\leq i\leq r$. By the choice of $x,x_{123}, x_4, \dots, x_{r}$, we have that every edge containing all of $x,x_{123}, x_4, \dots, x_{r}$ must have color $r+1$ and thus $|V(C_{r+1})|\geq r-1+\delta_{r-1}(G)\geq \frac{rn}{r+1}$, contradicting the fact that $|F_{r+1}|>\frac{n}{r+1}$.
\end{proofclaim}
Note that the previous claim implies that for all distinct $i,j\in [r+1]$, $F_i\cap F_j=(F_i\cap F_j)\setminus (F\setminus (F_i\cap F_j))$.
\begin{claim}\label{clm:2way2*}
For all distinct $h,i,j,k\in [r+1]$, $F_h\cap F_i=\emptyset$ or $F_j\cap F_k=\emptyset$.
\end{claim}
\begin{proofclaim}
Suppose for contradiction that $F_h\cap F_i\neq \emptyset$ and $F_j\cap F_k\neq \emptyset$ for some distinct $h,i,j,k\in [r+1]$ and without loss of generality suppose $\{h,i\}=\{1,2\}$ and $\{j,k\}=\{3,4\}$. Let $x_{12}\in F_1\cap F_2$ and $x_{34}\in F_3\cap F_4$ and let $x_i\in F_i$ for all $5\leq i\leq r$. By the choice of $x,x_{12}, x_{34}, x_5 \dots, x_{r}$, we have that every edge containing all of $x,x_{12}, x_{34}, x_5 \dots, x_{r}$ must have color $r+1$ and thus $|V(C_{r+1})|\geq r-1+\delta_{r-1}(G)\geq \frac{rn}{r+1}$, contradicting the fact that $|F_{r+1}|>\frac{n}{r+1}$. Note that when $r=3$, we immediately have a contradiction since $x_{34}$ and $x$ cannot be contained together in an edge of color $4$.
\end{proofclaim}
For all $i\in [r+1]$, set $F_i^*=F_i\setminus (F\setminus F_i)$; i.e. the set of vertices which are in $F_i$ but not in $F_j$ for any $j\in [r+1]\setminus \{i\}$.
\begin{claim}\label{clm:1way2way*}
For all distinct $h,i,j\in [r+1]$, $F_h\cap F_i=\emptyset$ or $F_j^*=\emptyset$.
\end{claim}
\begin{proofclaim}
Suppose for contradiction that $F_h\cap F_i\neq \emptyset$ and $F_j^*\neq \emptyset$ for some distinct $h,i,j\in [r+1]$ and without loss of generality suppose $\{h,i\}=\{1,2\}$ and $j=3$. Let $x_{12}\in F_1\cap F_2$ and $x_3\in F_3^*$ and let $x_i\in F_i$ for all $4\leq i\leq r$. By the choice of $x_{12}, x_{3}, x_4 \dots, x_{r}$, we have that $x_{12}\in C_j$ for all $j\in [r+1]\setminus\{1,2\}$ and $x_3\in C_j$ for all $j\in [r+1]\setminus\{3\}$ and thus every edge containing all of $x_{12}, x_{3}, x_4 \dots, x_{r}$ must have color $r+1$. Since $x_{12}\in C_{r+1}$, this means that $|V(C_{r+1})|\geq r-1+\delta_{r-1}(G)\geq \frac{rn}{r+1}$, contradicting the fact that $|F_{r+1}|>\frac{n}{r+1}$.
\end{proofclaim}
Since $|F_i|> \frac{n}{r+1}$ for all $i\in [r+1]$, we have (by pigeonhole) that $F_i\cap F_j\neq \emptyset$ for some distinct $i,j\in [r+1]$; without loss of generality, say $i=r$ and $j=r+1$. Furthermore, by the comment after Claim \ref{clm:3way*} we must have that $F_{r}\cap F_{r+1}=(F_{r}\cap F_{r+1})\setminus (F_1\cup\dots \cup F_{r-1})$. Thus by Claim \ref{clm:2way2*}, we have that for all distinct $i,j\in [r-1]$, $F_i\cap F_j=\emptyset$, and by Claim \ref{clm:1way2way*} we have that for all $i\in [r-1]$, $F_i^*=\emptyset$.
So for all $i\in [r-1]$, we have $F_i^*=\emptyset$, $F_i\cap F_{r}\cap F_{r+1}=\emptyset$, and for all $j\in [r-1]\setminus\{i\}$, $F_i\cap F_j=\emptyset$, thus $$|F_i\cap F_{r}|+|F_i\cap F_{r+1}|= |F_i|> \frac{n}{r+1}.$$ Let $i=r-1$ say, and suppose without loss of generality that $|F_{r-1}\cap F_{r+1}|> \frac{n}{2(r+1)}>0$. Thus by Claim \ref{clm:2way2*} we have that for all $j\in [r-2]$, $F_j\cap F_{r}=\emptyset$ which in turn implies that for all $j\in [r-2]$, $|F_j\cap F_{r+1}|= |F_j|>\frac{n}{r+1}$. This in turn implies that $|F_{r-1}\cap F_{r+1}|= |F_{r-1}|>\frac{n}{r+1}$. Now by Claim \ref{clm:1way2way*}, this implies that $F_{r}^*=\emptyset$. Thus $|F_{r}\cap F_{r+1}|= |F_{r}|>\frac{n}{r+1}$. So all together we have $r$ sets, $F_1\cap F_{r+1}, \dots, F_{r-1}\cap F_{r+1}, F_{r}\cap F_{r+1}$ all with more than $\frac{n}{r+1}$ vertices.
We claim that there is a component of color $r+1$ which contains every vertex in $\bigcup_{i\in [r]}(F_i\cap F_{r+1})$ and thus this component has order greater than $\frac{rn}{r+1}$. To see this, let $j\in [r]$ and let $x_i\in F_i\cap F_{r+1}$ for all $i\in [r]\setminus \{j\}$. By \eqref{codeg} and the fact that $|F_j\cap F_{r+1}|>\frac{n}{r+1}$, we have that there exists $x_j\in F_j\cap F_{r+1}$ such that $e:=\{x_1, \dots, x_r\}\in E(G)$. By the definition of the $x_i$'s, it must be the case that $e$ has color $r+1$. Thus we have just shown that for all $j\in [r]$ and all $x_i\in F_i\cap F_{r+1}$ for all $i\in [r]\setminus \{j\}$, there is a component of color $r+1$ which contains all of $x_1, \dots, x_{j-1}, x_{j+1}, \dots x_r$ and thus there is a single component of color $r+1$ which contains all of $\bigcup_{i\in [r]}(F_i\cap F_{r+1})$ and thus this component has order greater than $\frac{rn}{r+1}$.
\end{proof}
\section{Conclusion}
For all $1\leq \ell\leq k-1$ and $n\geq r\geq k+1\geq 4$, let $f(n,r,k,\ell)$ be the smallest integer for which it holds that if $G$ is a $k$-uniform hypergraph on $n$ vertices with minimum $\ell$-degree at least $f(n,r,k,\ell)$, then $\mathrm{mc}_{r}(G)=\mathrm{mc}_r(K_n^k)$.
While we gave a best possible strengthening of Theorem \ref{thm:Gy}(iii) in terms of minimum $(r-1)$-degree, it would be interesting to consider minimum $\ell$-degree for $1\leq \ell\leq r-2$; in particular $\ell=1$.
We note that for $r\geq k+2$, determining the (asymptotic) value of $\mathrm{mc}_r(K_n^k)$ itself is mostly open and thus the general problem of determining $f(n,r,k,k-1)$ is likely out of the question at the moment. That being said, there is one general result of
F\"uredi and Gy\'arf\'as which for all $k\geq 3$, determines the exact value of $\mathrm{mc}_r(K_n^k)$ for infinitely many $r$ and $n$.
\begin{theorem}[F\"uredi and Gy\'arf\'as~\cite{FG}]\label{fgthm}
Let $k,r\geq 2$ and let $q$ be the smallest integer such that $r\leq q^{k-1}+q^{k-2}+\dots+q+1$. Then $\mathrm{mc}_r(K^k_n)\geq \frac{n}{q}$.
This is sharp when $q^k$ divides $n$, $r=q^{k-1}+q^{k-2}+\dots+q+1$, and an affine space of dimension $k$ and order $q$ exists.
\end{theorem}
As for the case $k=3$ and small values of $r$ not covered by any of the results mentioned thus far, Gy\'arf\'as and Haxell proved the following.
\begin{theorem}[Gy\'arf\'as and Haxell~\cite{GH}]\label{ghthm}
$\mathrm{mc}_{5}(K^3_n)\geq 5n/7$ and $\mathrm{mc}_{6}(K^3_n)\geq 2n/3$. Furthermore these are tight when $n$ is divisible by $7$ and $6$ respectively.
\end{theorem}
So we conclude with the following problems which may be within reach.
\begin{problem}~
\begin{enumerate}
\item Determine the value of $f(n, r+1, r, 1)$ for all $r\geq 3$.
\item Determine the value of $f(n, r, k, k-1)$ for those $r$ and $k$ mentioned in Theorem \ref{fgthm} and Theorem \ref{ghthm}.
\end{enumerate}
\end{problem}
|
\section{Introduction}
Multimedia material, such as images and videos, has grown in popularity on different social media platforms during the last several decades. One may now create a realistic-looking face that doesn't happen in reality or do a high-level realism face swap in a movie because of recent breakthroughs in deep learning-based image and video generation approaches, such as generative adversarial networks (GAN). The community refers to the latter as the Deepfake\cite{zhou2017two}. Prior until today, doing such a face swap needed expertise in areas like theatrical visual effects (VFX) and/or fast-tracking using indicators.~\cite{wodajo2021deepfake}
\vspace{3mm}
\\Therefore, the fight against Deepfakes is now in dire need of resources. There have been various attempts to recognize Deepfake movies, although Deepfake detection is a relatively newer research area ~\cite{huang2022fakelocator}. Although certain methods extensively depend on categorizing images and videos as true or fraudulent using raw image Deepfake data, others employ more traditional digital forensics methods ~\cite{chen2022pulseedit}. Moreover, if deep image synthesis technologies progress soon, such detection algorithms that just use raw pixel-domain input may develop to be less operative as Deepfake images as well as videos convert to be further convincing ~\cite{khalil2021icaps}. As a result, a completely new Deepfake detection approach is required~\cite{zhang2022cascaded}.
\vspace{3mm}
\\Deepfakes, or fake films and pictures, have caused a slew of societal problems in recent years ~\cite{ciftci2020hearts}. Deepfakes are made by iterating a real data-based creation and verification job via two opposing deep learning models utilizing the generative adversarial network (GANs) methodology~\cite{du2020towards}. Owing to the pixel collapsing phenomenon which produces false visual irregularities in the skin tone or the shape of the face or often occurring visual aberrations, Deepfake videos may be recognised with the naked eye in their early stages. While, in recent years, as technology has developed, Deepfakes have become practically impossible to tell apart from actual photographs ~\cite{yasrab2021fighting}.
\vspace{3mm}
There has been an upsurge in the incidence of indecorous use of technology as it develops. Deepfake, which produces a large number of pornographic photos of politicians and celebrities to spread propaganda, has caused a wide variety of social problems~\cite{fernandes2019predicting}. As a consequence, a lot of research has been conducted in creating a system for confirming the reliability of Deepfakes. A method for detecting the collapsing of pixels and other visible abnormalities in Deepfake has become one of the consistent assessment approaches that have become the subject of most research ~\cite{cozzolino2021id}.
\vspace{3mm}
\\However, because the creator as well as the classifier in the GANs paradigm gradually evolved to dodge such validation, using this technique has become more difficult ~\cite{ciftci2020fakecatcher}. Therefore, the usage of this approach has grown challenging because the generator as well as the discriminator in the GANs models have developed to evade such confirmation~\cite{vinay2022afmb}.
\section*{GAN (Generative Adversarial Networks)}
GAN is one of the most difficult deep learning techniques to train and apply on computational hands\cite{dong2022restricted}. There, we see two different neural networks: the generator and the discriminator. Although the generator net seems to be quite similar to the autoencoder net, we are able to produce much better outcomes because of the discriminator net's ability to exclude certain poor samples\cite{masood2022deepfakes}. Therefore, the GANs approach of Deepfake production assumes that the generator should be able to deceive the discriminator, which is another computer\cite{falahkheirkhah2022deepfake}. This is what makes these fakes more comparable to actual movies and also makes it harder for one’s eyes to recognise them\cite{almars2021deepfakes}. This method is being used by a few of the open-source projects, like Face swap-GAN.
\begin{figure}[H]
\centering
\includegraphics{{Figure1.jpg}}
\caption{Example of training GAN\cite{maksutov2020methods}}
\label{fig:figure1}
\end{figure}
\section*{Background : An overview of Deepfake}
A politician, actress, and perhaps other celebrity's face may be switched by any other artist or people's face using the tampering or manipulation technique known as "Deepfake." This may be done with any photo or video\cite{akhtar2019comparative}. It is possible to create fake films, audio recordings or photographs that appear and sound genuine. Using massive datasets that include audio, videos or photographs as building blocks, the system constructs a model of a person speaking or doing something\cite{zendran2021swapping}. The manipulation is accomplished by using a dataset consisting of hundreds or thousands of photographs of the individual who is the subject of the investigation. Deepfake is a video and image manipulation technology that uses artificial intelligence techniques. Machine learning, an implementation of artificial intelligence, is used as it is a technique which enables a computer to adapt from the inputs of easily available information \cite{zhao2021learning}.
\vspace{3mm}
\\An application known as FakeApp was made available on Reddit, and its primary function was to walk users through the fundamental phases of the Deepfake algorithm. It is possible to make a Deepfake picture or video even with a little understanding of machine learning and programming. Deepfakes are often created for one of three reasons: to exact vengeance on another person; to publish a pornographic movie featuring a celebrity, or to blackmail a person by displaying a video or picture that has been changed or manipulated. Deepfakes are also used in the production of fake films depicting politicians for the purpose of fabricating news. In a nutshell, Deepfake has developed into a significant concern in today's society\cite{chen2021magdr}.
\section{ STATEMENT OF PROBLEM }
Deepfakes are becoming an urgent and visible danger to the integrity of the multimedia information we have at our disposal. Deepfakes, for example, when used on politicians and fed with targeted disinformation, might have a detrimental influence on people's opinions and have unintended consequences, including corrupted and tampered votes.
\vspace{2mm}
\\As per the Washington Post, face photos, as well as pornographic images, are now masterfully combined and transmitted through social media without the accord or understanding of the involved entities. The crime victims of these Deepfakes pictures are also spread to include the wider populace. Even some companies specialize in shipping Deepfakes. Deepfakes include pornographic images and videos where a person's face may be replaced with another person's using neural networks. Since Deepfakes are an issue for the general population, procedures for spotting them must be created \cite{dolhansky2020deepfake}. In the creation of Deepfake, tampering is not completely successful leading to certain anomalies in face structure like an anomaly in eye movement like blinking or eyebrow movement. With this review paper, we aim to study these Deepfake detection methods that employ biological features in detection. Broadly, we study five biological features: Eyebrow, eye blinking, eye movement, ear and mouth movement and heartbeat detection.
\section{DEEPFAKE}
\subsection{Deepfake Creation}
\begin{figure}[H]
\centering
\includegraphics{Figure2.png}
\caption{Deepfake Creation\cite{nguyen2022deep}}
\label{fig:figure2}
\end{figure}
Due to the high calibre of the altered videos and the ease of various implementations for a broad variety of users with different computer abilities, from experts to novices, Deepfakes have grown in popularity. These applications are usually made using deep learning methods. The ability of deep learning to represent complicated and multidimensional data has a long history\cite{jung2020deepvision}. Deep autoencoders, a type of convolutional model having this feature, have been extensively used for image compression and picture compression. A Reddit user created Fakeapp utilized an autoencoder-decoder coupling architecture as the first effort to build a Deepfake\cite{abu2022analysis}. Using the latent features that the auto-encoder has gleaned from the face pictures, the decode recreates the facial image in this manner. To switch features among the target and source photos, two encoder-decoder pairings are required, each of which is learned on a collection of images, and the settings of the encoding are exchanged by the two networking pairs\cite{qi2020deeprhythm}. So, two pairs make use of the identical encoder. The common encoder can identify and understand the similarities among multiple groups of face pictures using this approach since faces have common characteristics including the locations of the eyebrows, lips, or mouths\cite{lee2022bznet}.
\vspace{3mm}
\\Large number of photos of the two subjects are used to create Deepfakes, which are created using the following steps:
\vspace{3mm}
\\Step 1: All of the photographs are first taken and encoded using a deep learning CNN by an encoder. The encoded data is then decoded using a decoder to recreate the picture. The encoder and the decoder struggle to store all of the parameters in their separate memory areas due to the enormous number of parameters \cite{sun2022faketracer}. The encoder will only remove major components that are necessary for reconstructing the original input in order to get around this issue. The decoder will continue to decode those features when the feature extraction operation is complete. Backpropagation will be used throughout the rest of the training procedure until the output and the input are identical. Graphical processing units, sometimes known as GPUs, are used since the procedure requires a significant amount of time.
\begin{figure}[H]
\centering
\includegraphics{{Figure3.jpg}}
\caption{Using two separate decoders\cite{nguyen2022deep}}
\label{fig:figure3}
\end{figure}
Step 2: It involves replacing the person's visage with another's on a frame-by-frame basis once the training procedure has been completed\cite{nguyen2022deep}. Face recognition software is used to isolate person A's visage, which is then supplied into the encoder. Instead of giving it to its original decoder, the picture is fed into the decoder of person B, who then uses it to reconstruct the image. As a result of this, characteristics of individual A from the original drawings from the video are done on person B. As can be seen in the figure that follows, after this step has been completed, the freshly fabricated face is merged with the initial picture.
\begin{figure}[H]
\centering
\includegraphics{{Figure4.jpg}}
\caption{Merging of Newly Created Face\cite{nguyen2022deep}}
\label{fig:figure4}
\end{figure}
Step 3: Prior to beginning the training process, it is necessary to have hundreds or even thousands of photographs of both individuals\cite{guarnera2020deepfake}. The quality of these face photographs may be highly increased, that allows for much superior outcomes. It is essential to get rid of any pictures that have poor lighting or quality, as well as any others that include other people.
\begin{figure}[H]
\centering
\includegraphics{{Figure5.jpg}}
\caption{Similar resemblance such as face shape \cite{nguyen2022deep}}
\label{fig:figure5}
\end{figure}
Step 4: If the resolution of the finished image is different from the resolution of the source photo, the Deepfake will seem unnatural. This problem may be solved by cropping and rearranging the picture such that it has dimensions of 256 by 256. The training procedure is the only time the 160x160 pixel section in the middle is used; after training, it is scaled down even more to 64x64 pixels. As a consequence of this, the rebuilt faces have a resolution of 64 by 64 pixels, and they are included into a movie\cite{frank2021wavefake}. The freshly produced photos are then changed so that their dimensions are brought back in line with the originals, as seen in the figure below. However, as a result of this transition, the face will take on a hazy appearance. Both training a neural network that is able to function with pictures that have higher pixel sizes and decreasing the resolution of the video or image that is to be switched are two options that might be examined in order to circumvent this challenge.
\begin{figure}[H]
\centering
\includegraphics{{Figure6.jpg}}
\caption{Resolution morphed back to match original size\cite{nguyen2022deep}}
\label{fig:figure6}
\end{figure}
After the Deepfake video has been made, there are two technical considerations that need to be taken into consideration as the fifth step. Both of these technical characteristics are part of artificial intelligence systems; the first of these is referred to as a discriminator, while the second of these is called a generator. Both of these systems operate in tandem with one another; the generator is the component that is responsible for producing phoney films or photographs. After the phoney video has been prepared, the discriminator will evaluate it to decide whether or not it is a fake or a real video. While it is determined that the video is a fake, the discriminator provides a hint to the generator about the preventative measures that it needs to take when it is making the subsequent clip. This results in the formation of a network known as a Generative Adversarial Network, or GAN. In order for a GAN to function, the user must first provide the desired kind of output to the system, which then generates a dataset for the generator to learn from. The generator will begin producing a variety of movies, and as soon as the necessary output has been reached, the videos will be sent to the discriminator to be analyzed. The more errors that are discovered in phoney videos, the more improvements that may be made to the generator for the following iteration. Therefore, improving the discriminator's ability to recognise false movies will also improve the generator's ability to create convincing fake videos, as the figure below demonstrates.
\begin{figure}[H]
\centering
\includegraphics{{Figure7.jpg}}
\caption{Discriminator detecting whether the image is real or fake\cite{nguyen2022deep}}
\label{fig:figure7}
\end{figure}
\subsection{Deepfake Detection}
The threat of Deepfakes to democracy, social security, and privacy is growing. Early attempts relied on fabricated traits that were the result of mistakes and shortcomings in the fake video generation process\cite{hernandez2020deepfakeson}. On the other extreme, deep learning is currently becoming used to extract features with significant and distinctive qualities in order to identify Deepfakes.
\vspace{3mm}
\\Classifiers are used to differentiate between authentic movies and ones that have been manipulated in the process of Deepfake detection. This issue is also referred to as a binary classification problem. To train categorization algorithms using this strategy, a sizable collection of real and fraudulent films is needed\cite{lyu2020deepfake}. False movies are more accessible than ever, but their application is still restricted in terms of providing a standard for comparing different detection algorithms\cite{nguyen2022deep}. Employing clips from the searchable database, low and high quality Deepfake videos were created that accurately mimicked facial emotions, lips movement, especially eye blinking. Then, several Deepfake detection algorithms were tested using these videos.
\begin{figure}[H]
\centering
\includegraphics{{Figure8.jpg}}
\caption{Deepfake detection image}
\label{fig:figure8}
\cite{https://doi.org/10.48550/arxiv.1905.08233}
\end{figure}
The experiments' findings show that popular face recognition algorithms based on VGG and Face net are ineffective at effectively spotting Deepfakes\cite{lyu2020deepfake}. Whenever applied to the same freshly collected data set, additional approaches, such as the lip-syncing approach as well as SVM picture quality evaluations, focus on providing extremely significant error rates. This elevates questions regarding the urgent requirement for the creation of future, higher reliable techniques for identifying real Deepfakes versus Deepfakes.
\subsection{Fake Image Detection}
Face swapping offers a wide range of exciting uses, such as portrait transformation, video compositing, and, most significantly, individuality defense since it could substitute people's faces in photos with stock photos. To get unauthorized access to identification or authentication systems, cybercriminals also utilize this tactic. Since deep learning models can preserve the location, facial expression, and lighting of the shots, switched face photos are becoming more challenging for forensics algorithms. Deep learning models include CNN and GAN.
\vspace{3mm}
\\SVM, random forest (RF), and multi-layer perceptrons were a few of the classifiers used to distinguish swapped face photos from real ones. A collection of condensed features was extracted using the bag of words approach (MLP). Considering that GAN-generated pictures are the most realistic and of the highest quality among deep learning-generated images, this is likely because GAN is capable of understanding the distribution of complicated input data and producing new outcomes featuring a similar feature distribution. The generality capability of the detection methods also isn't considered in the bulk of works on the detection of GAN-produced images, notwithstanding the reality that GAN research is ongoing and several novel GAN extensions are consistently created an image pre-processing procedures to eliminate low-level, high-frequency suggestions in GAN pictures, including Gaussian blur as well as Gaussian noise. In comparison to earlier image forensics techniques or image steganalysis networks, this increases the statistical resemblance between actual and false pictures at the pixel level and drives the forensic classifier to understand additional intrinsic as well as relevant characteristics.
\subsection{Current Deepfake Detection Methods}
Face-swapping movies, which make up the overwhelming bulk of Deepfake films published online, are primarily targeted by Deepfake detection algorithms now in use. Numerous approaches now in use are founded on issues concerning binary classification at the segmented stage. Various techniques are categorized into three main groups depending on the traits they use. The initial individual categories strategies are centered on variations in the physiological as well as physical properties of the Deepfake videos. The method in use is based on the discovery that a lot of Deepfake videos lack comforting eye blinking as a result of the use of online portraits as training data, which often do not contain closed eyes for aesthetic reasons. Incoherent head movements in Deepfake films are used to uncover Deepfake movies. The second group of Deepfake detection techniques takes the use of synthesis-related artifacts at the signal level. Many methods currently in use are based on problems with binary classification at the subcellular level. According to the characteristics they use, different approaches are divided into three primary groupings. The first category-specific tactics are based on differences in the physiological and physical characteristics of Deepfake films.
\subsection{Deepfake Detection: Current challenges and Next steps}
The current issues that are now being faced as well as those that will be faced in the future with Deepfake detection Fundamentally speaking, there are three primary categories of Deepfake videos.
\begin{itemize}
\item Head puppetry involves mixing a video of a source person's head with a video of a target person's whole head as well as upper shoulders such that the synthesizing targeting seems to work in the same method as the source. This creates the illusion that the source is the target.
\item Lip syncing is the method of making a fake video by simply employing the target's lip area such that the target seems to say approximately that they do not pronounce in real life.
\end{itemize}
\subsection*{Existing challenges:}
In the meantime, the subject of identifying Deepfake films is often presented, addressed, and estimated as a binary classification problem. As a result, every video is classified as either authentic or a Deepfake. When developing and testing Deepfake detection algorithms employing movies that are neither original nor made employing Deepfake production methods, it is feasible to establish a contradiction of this kind. However, when the detection approach is employed in real-world scenarios, the issue becomes more challenging.
The circumstance that a video is not recognized as a Deepfake does not always imply that the video is a legitimate one. For instance, movies may be produced or altered using techniques other than Deepfakes. A single label would not be able to adequately represent the fact that a Deepfake clip may have undergone different alterations. Additionally, Deepfake will only create one or a small number of those faces for a fraction of the frames in the movie if a film comprises the faces of many different persons. Therefore, to fully handle the complexity of real-world media forgeries, the binary classification approach has to be enhanced to incorporate multiclass, multilabel, and local classification and detection.
\section{RELATED WORK}
In the research that has been done so far, a variety of methods for detecting Deepfake photos and videos have been suggested. a comparison of the most important techniques in the field, with a focus on the ones that are based on physiological monitoring and paying particular attention to the bogus detectors. We provide information on the previous study's methodology, as well as its classifiers, highest achievable level of performance, and research databases.
\subsection{Current Directions in Deepfake Studies}
The first proposed GANs model \cite{goodfellow2014advances} is notable because it developed a brand-new technique for learning by gathering information with the Generator and authorizing it using the Discriminator. The saddles' difficulty and the value function problem \cite{radford2015unsupervised}, which cause strange ranges in the shape and shade of produced images, are two defects that it does have. A research team from the University of Washington built upon this improvement in 2017. They created complex false movies that replicated the voice and mouth of a speaker in a video and generated the exact form of his mouth at every instant \cite{suwajanakorn2017synthesizing}. They achieved this by using university-developed software. This allowed the adoption of methods like jaw correction\cite{bayar2016deep} to greatly improve upon the prior limits of pixel crush, jaw form, wrinkles, and so forth.
\vspace{3mm}
\\Alec Radford and colleagues presented (DCGAN) Deep Convolutional Generative Adversarial Networks in 2016, which allowed for the computation of arithmetic operations with filtering among pictures employing latent vectors. This was done by using CNN models on GANs, which produced more advanced fakes \cite{radford2015unsupervised}.
\subsection{Research on Detection of Deepfakes}
For generalized fake video detection \cite{cozzolino2017recasting}, face matching \cite{rossler2018faceforensics}\cite{hsu2018learning}, and eye blinking [42], they looked at a variety of forensic techniques. The most popular Deepfake detection techniques include training deep neural networks on a dataset of forged faces or spotting an aberrant pixel \cite{guera2018deepfake}\cite{koopman2018detection}.
Many Deepfakes faces, according to a study \cite{guera2018deepfake}, do not blink their eyes. However, a few more recent instances have modified the discriminator to examine blinking to circumvent these detection techniques\cite{rossler2018faceforensics}.
\vspace{3mm}
\\To develop deep learning-based algorithms, Face Forensics++ \cite{rossler2018faceforensics} is a helpful dataset of face forgeries. Additionally, research \cite{hsu2018learning} describes a method for spotting fake videos using a trained CNN. The techniques produce reliable findings, but they necessitate a large amount of input and frequent updates. As a consequence, we focused on research, which requires less information but is more effective. It will probably be used more widely\cite{lawrence1997face}.
\vspace{3mm}
\\The literature has also looked at methods based on the identification of face-warping artifacts. Among the best ways to spot concealed face modifications is through CNN-based identification algorithms, which can recognize the presence of such abnormalities from the face as well as its surroundings\cite{li2020sharp}.
\vspace{3mm}
\\The most widely used fake detectors are probably those that use just deep learning features, putting as many actual and fake videos into the networks as feasible and letting the networks automatically extract the distinguishing traits. These false detectors have generally proven effective. outstanding results when using standard network architectures.\cite{cozzolino2021id}
\vspace{3mm}
\\Using cutting-edge network topologies, the scientists assessed each face region's capacity for discrimination and obtained intriguing findings on Deepfake datasets from the first and second generations\cite{tolosana2001beyond}. This is in contrast to the majority of methods, which rely on the complete face to identify fraudulent movies.
\vspace{3mm}
\\Table 1 highlights the research publications on Deepfake detection and Super-resolution, as well as the various strategies utilized for Deepfake detection. The following table analyzes the different techniques and characteristics utilized for Deepfake detection. It comprises approaches based on Machine Learning and Deep Learning.
\begin{table}
\caption{Analysis Table}
\centering
\begin{tabular}{|c|l|l|l|c|}
\toprule
\textbf{Sr. No.}& \textbf{Title of Paper}& \textbf{Techniques used}&\textbf{Dataset used} &\textbf{Accuracy}\\
\midrule
\hline &&&&\\
\cite{yadav2019deepfake} & Deepfake: A Survey on & 1. Convolutional Neural &Face2Face, Reddi & {95\%} \\
& Facial Forgery Technique & \hspace{0.3cm} Networks (CNN) & user Deepfakes& \\
& Using Generative Adversarial & 2. Long Short-Term & &\\
& Network &\hspace{0.3cm}Memory (LSTM)& &\\
&&&&\\\hline &&&&\\
\cite{amerini2019deepfake} & Deepfake Video Detection & Convolutional Neural &Face2Face & VGG16 {81.61\%}, \\
& through Optical Flow & Networks (CNN) & & ResNet50 \\
& based CNN & & &{75.46\%} \\
&&&&\\\hline &&&&\\
\cite{guera2018deepfake} & Deepfake Video Detection & 1. Convolution Neural &HOHA dataset & Conv-LSTM \\
& Using Recurrent Neural & \hspace{0.3cm} Networks (CNN) & &(20 frames) {96.7\%}, \\
& Network & 2. Long Short-Term & &Conv-LSTM\\
& &\hspace{0.3cm} Memory (LSTM)& &(40 frames)\\
&&&&\\\hline &&&&\\
\cite{ranjan2020improved} & Improved Generalizability of & 1. Convolutional Neural &1. FaceForensics++ & With Transfer \\
& Deep-Fakes Detection &\hspace{0.3cm}Networks (CNN) & 2. Celeb-DF &Learning {84\%}, \\
& Using Transfer Learning & 2. Long Short-Term &3. Deepfake Detection& Without Transfer \\
& Based CNN Framework &\hspace{0.3cm}Memory (LSTM)& Challenge&Learning {75\%}\\
&&&&\\\hline &&&&\\
\cite{luo2018multi} & Multi-scale face & 1.Convolution Neural &1. Celeb A & Discrete- {95\%} and \\
& detection based on &\hspace{0.3cm}Networks (CNN) & 2. AFW & for continuous, it \\
& convolutional neural & & 3. FDDB & is {74\%}\\
& networks.& & & \\
&&&&\\\hline &&&&\\
\cite{matern2019exploiting} & Exploiting Visual Artifacts & The neural network &1.CelebA & MLP {84\%(Eyes)},\\
& to Expose Deepfakes & classifier as MLP and & 2. ProGAN & LogReg \\
& and Face Manipulations & the logistic regression & 3.Glow & {83\%(Eyes)}\\
& & model as LogReg & &\\
&&&&\\\hline &&&&\\
\cite{mccloskey2019detecting} & Detecting Gan-generated & SVM classifier & Image net dataset & {92\%,} \\
& imagery using saturation cues &&&\\
&&&&\\\hline &&&&\\
\cite{kharbat2019image} & Image Feature Detectors & 1. SVM classifier & Unnamed with & HOG {94.5\%},\\
& for Deepfake Video Detection & 2. Feature extraction & 98 videos & SURF {90\%},\\
&&\hspace{0.3cm}algorithms & KAZE {76.5\%}&\\
&&&&\\\hline &&&&\\
\cite{malolan2020explainable} & Explainable Deep-Fake & 1.Xception net (CNN) & Face Forensics++ & {90.17\%}\\
& Detection Using Visual & 2.LRP and LIME &&\\
& Interpretability Method &&&\\
&&&&\\
\bottomrule
\end{tabular}
\label{tab:table}
\end{table}
\section{PERFORMANCE ANALYSIS}
Currently, Deepfake video identification is frequently discussed, analyzed, and presented as a binary classification issue. Each video is thus categorized as either legitimate or a Deepfake. In randomized studies wherein we design as well as test Deepfake detection techniques employing movies that are neither authentic nor made utilizing Deepfake production techniques, it is simple to set up a dichotomy similar to the ones indicated above. When the detecting approach is used in the actual world, the situation becomes murkier. Different than using Deepfakes, videos might well be created or edited in rather ways\cite{korshunov2019vulnerability}, thus just because a video isn't classified as a Deepfake doesn't always mean it's a genuine one. A single label may not sufficiently convey this if it occurs that a Deepfake video has been subjected to numerous forms of manipulation. Additionally, Deepfake will only create one or a limited number of faces in a video with many different people's faces for a tiny portion of the frames. The development of the binary classification method to include multi-class, multi-label, including local categorization as well as identification is necessary due to the complexity of real-world media forgeries.
\vspace{3mm}
\\The capacity of deep learning techniques, such as CNN as well as GAN, to maintain picture position, facial expression, as well as illumination, has made it more demanding for forensics algorithms to distinguish amongst swapped face shots \cite{zhang2017automated}. Zhang et al\cite{pu2020noisescope}'s bag of words approach was utilized to extract a set of condensed characteristics, that were then passed into a variety of classifications, including SVM, RF, and MLP, which are two techniques that can distinguish between images of swapped faces and real faces. Because they are the most realistic and high-quality pictures produced by deep learning, those produced by GAN models are arguably the most difficult to distinguish. This is so that fresh outputs produced by GAN models will have a distribution that is comparable to the learned distribution of the complicated input data.
\vspace{3mm}
\\Regardless of the fact because GAN is continuously developed and several new extensions are often published, the majority of research on the detection of GAN-generated images does not examine the generalization capacity of the detection models. Even if GAN is often extended in novel ways, this is still true. For instance, Xuan et al. \cite{yang2021defending} used Gaussian blur and Gaussian noise to GAN pictures to filter out low-level, high-frequency information. The forensic classifier must acquire more intrinsic and relevant traits due to the rising pixel-level statistical similarity between real and fake pictures.
\vspace{3mm}
\\On the other side, a problem with hypothesis testing is the GAN-based Deepfake detection. They did this by introducing a statistical framework by making use of an information-theoretic analysis of authentication. Oracle error is the term used to refer to the smallest acceptable gap that can be established between the distributions of real pictures and those of images produced by a certain GAN.
\section{EXPLAINABILITY}
Explainability is another difficult aspect. In many different situations, a numerical score indicates the possibility that the picture or video was fake. In situations like this, it is an extremely common practice to inquire about the reason behind the numerical score that must be acceptable for the analysis. Moreover, several data-driven DF detection technologies are black boxes. Therefore it is difficult for these algorithms to explain their results. This is especially true about techniques that rely on deep neural networks model DNN. Additionally, because of the "black box" character of DNN algorithms, many data-driven DF detection systems, particularly many which rely on their use, cannot usually understand the results.
\vspace{3mm}
\\\textbf{Temporal Aggregation:}The majority of the currently available techniques for detecting Deepfakes at the segment level are predicated on binary identification. In other words, they estimate the likelihood that a certain frame is Deepfake or real. Temporal aggregation, meanwhile, goes one step further. Although it seems simple and uncomplicated, this approach has two significant drawbacks. Firstly, even though several Deepfake films include behavioral abnormalities, real or Deepfake images generally appear at constant intervals. The chronological integrity across pictures also isn't specifically checked. And secondly, it requires an additional step to be taken whenever a score for the video's integrity is needed. To calculate such a score, we must first aggregate the scores obtained from each frame.
\vspace{3mm}
\\textbf{Protection measures:} One of the many approaches they have recently researched is to include adversarial perturbations, which are carefully created patterns that are undetectable to human eyes but may result in detection failures. This is one of the numerous methods they have recently explored. The second element, which involves making a genuine video that seems like a Deepfake video by incorporating approximated transmitted signal characteristics required by existing detection algorithms, could also be employed to construct anti-forensic techniques. This situation is what we call phoney Deepfake. In this regard, anti-forensic methods may also be devised. It is necessary to enhance further Deepfake technology detection to counter both harmful and intentional attacks.
\vspace{3mm}
\\\textbf{Social Media Laundering:} The dissemination of a sizable amount of internet videos is rapidly being done via social media sites like Facebook, Instagram, and Twitter. These videos often undergo extensive compression, meta-data removal, and file size reduction before being uploaded to social media. This is carried out to protect viewers' privacy as well as to reduce network traffic. These actions, more often known as social media laundering, make it harder to identify underlying manipulation and increase the chance of false positive detections, which is the practice of categorizing a real video as a Deepfake.
\vspace{3mm}
\\\textbf{Anti-forensics:} The second element, entails introducing simulated signal characteristics needed by current detection techniques to make real footage appear to be Deepfake footage, which may also be used to construct anti-forensic techniques. We refer to this circumstance as a false Deepfake. Anti-forensics: The second element allows for the development of anti-forensic measures. Additional Deepfake detection algorithms need to be improved to deal with these purposeful and antagonistic attacks.
\vspace{3mm}
\\However, biological features such as eyebrow recognition, eye blinking detection, eye movement detection, ear and mouth detection, and heartbeat detection are clearly described. A comparison of each biological feature with different classifiers/techniques along with its key findings is discussed.
\section{BIOLOGICAL FEATURES}
\subsection{Eyebrow Recognition}
In \cite{nguyen2020eyebrow}, they used eyebrows to identify Deepfakes via biometric comparison. Deepfake detection method has used a biometric comparison of the brow region. One of the most affected areas in the synthesized images is the brow region. When examined by a biometric comparison pipeline, eyebrow alterations become more recognizable, particularly in high-resolution and high-quality Deepfakes. The model must first discover the participants' identity for this method to be effective (biometric enrolment is needed). Additionally, this will hold if the targets are well-known individuals like politicians or celebrities. Examples include LightCNN, Resnet, DenseNet, and SqueezeNet. In the publication of biometric research, they are often used. They believe they will do well in the brow-matching task as a consequence.
\vspace{3mm}
\\A large and excellent forensic dataset was created for Celeb-DF. This dataset contains 5639 Deepfake films in addition to 590 original footage from 59 celebrities. The paper has analyzed the highest quality Deepfake dataset, which provides the most convincing fake video following image synthesis visual artifacts. Although this is a challenging task, it could be useful in outlining the advantages of the recommended technique. Celeb-DF, in contrast to the other datasets, mostly lacks apparent Deepfake artifacts such as splicing boundaries, color mismatches, and inconsistent face orientation. As a result, several studies on this dataset's Deepfake detection have claimed low accuracy ratings.
\subsection{Eye Blinking Detection}
In \cite{li2018ictu} , the aim is to find facial videos created by AI by looking for the lack of eye blinking. To represent the phenomenology and chronological recurring patterns in the operation of eye blinking, CNN and RNN are combined in a unique deep-learning algorithm. To classify between open-eye and closed-eye states in a frame, current methods use Convolutional Neural Networks (CNN) as a binary classifier. CNN, on the other hand, bases its predictions exclusively on a single frame and ignores temporal domain data.
\vspace{3mm}
\\Long-term Recurrent Convolutional Neural Networks (LRCN) are employed to classify between open and closed eye states while taking past temporal information into account since human eye blinking has a considerable temporal relationship with earlier states. The proposed method in the literature performs well in identifying Deepfake videos when tested against benchmark datasets for eye-blinking detection.
\vspace{3mm}
\\A method is suggested for spotting eye blinking in clips and the general pipeline of the method discussed is given as follows. To discount head movements and variations in alignment utilising facial landmark point recognition, the recommended approach first detects faces in each video frame. The detected faces are then aligned into the same coordinate system. The sections matching each eye are eliminated to create a stable sequence. After completing these pre-processing steps, the LRCN (Long-term Recurrent Convolutional Networks) model is utilized to detect eye blinking by determining how wide an eye is open during each video frame.
\subsection{Eye Movement Detection}
FakeET \cite{gupta2020eyes} is an eye-tracking dataset used to study how Deepfake videos are perceived by viewers. Considering that Deepfakes' primary objective is to deceive human observers, FakeET was developed to understand and gauge viewers' propensity to spot artificial video artifacts. It is possible to measure the differences between exploratory and explanation by analyzing the waveform using a variety of eye track-based measurement techniques, including preoccupation intensity, scan-path length, reaffirmed, and fixation entropy \cite{li2021exposing}.
\vspace{3mm}
\\FakeET is the name of the initial set of user behavioral responses for Deepfake detection. Via extensive experimental investigation, they show the value of both the eye-gaze and EEG modalities in enabling automated Deepfake recognition. Both the eye gaze and the EEG in \cite{gupta2020eyes} modalities reveal distinct patterns suggestive of information problems. Google/Jigsaw Deepfake Detection is the database. 811 films—331 genuine and 480 fake—were included in our study out of the 3068 false and 363 real videos in the collection. Furthermore, based on the visual inspection of the fake videos, they were categorized into easy and difficult forgeries; this categorization yielded 244 simple and 236 difficult-to-detect phoney videos\cite{masood2022deepfakes}.
\vspace{3mm}
\\Based on the reported findings in \cite{gupta2020eyes}, the Dec-Tree and T-CNN algorithms outperform random in terms of F1-score, the k-NN classification obtains the second-highest F1 metric of 0.54, and the Naive Bayes classifier produces the best EEG-based FD having AUC and F1 ratings of 0.55. A random classifier performs similarly to regression models and linear SVM. While a small increase in FD results was achieved in this study, stronger EEG descriptors (such as Spectral Density characteristics and EEG spectrograms) and better classification techniques may be investigated for potential advancements\cite{startsev2022evaluating}. Generally, the results of EEG-based categorization show that Better-than-chance FD can only be achieved using PCA-ed EEG information and incorrect classifying techniques.
\subsection{Ear and Mouth Movement}
In a lip-sync Deepfake, a person’s mouth area is swapped with a fabricated or artificially manufactured audio recording. The human ear was not taken into consideration while making these deep-fake films. The ears belong to the original self even if the face identification in a face swap Deepfake might exactly resemble the co-opted identity. Even if the lips may be perfectly synchronized to the music in a lip-sync Deepfake, the ear's motion characteristics will be separated from the mouth as well as jaw movements \cite{agarwal2021detecting}.
\vspace{3mm}
\\For the above research, a total of 64 films of Mark Zuckerberg, Donald Trump, Angela Merkel, and Joe Biden were downloaded from YouTube. The duration of these films ranged from 12 to 59 seconds. In each video, the left or right ear was audible. They calculated the audio RMSE, the vertical separation between the lips, and the auditory motion for each frame of each video. In a lip-sync Deepfake, an existing video’s mouth movement is altered to relate to a new soundtrack \cite{gu2022delving}.
\vspace{3mm}
\\The correlations of face, aural and auditory signals across all Biden, Merkel , Trump and Zuckerberg- starring clips are determined between the facial (lip vertical) or auditory (RMSE) signal and the helix, tragus, or lobules horizontal and vertical motions. The connections are mostly characterized by a person-specific nature. Trump, but not the other candidates, displays a strong negative correlation, similar to the vertical lobule motion. The vertical aural motion follows this core structure consistently. For Trump, but not for remaining candidates, the tragus movements are substantially associated with audio. A linear classifier is trained in the manner described below to determine how effectively these dynamical acoustic parameters can detect lip-sync Deepfakes. For the accessible films, every person's instruction and assessment sets are classified into non {80\% / 20\%} groups. Using 12 aural/oral associations between the real videos and the simulated false ones, a multinomial logit model is developed. Then, utilizing real movies together with simulation, Generative adversarial network, and in-the-wild false films, this model was assessed\cite{elhassan2022dft}.
\vspace{3mm}
\\The mean evaluation accuracy rises from 0.84 to 0.87 with this individual training as well as the mean mentoring AUC rises from 0.91 to 0.98. Despite just looking at brief (10-second) portions, accuracy is often rather good. A complete film may be integrated with a simple majority rule to increase this accuracy\cite{gerstner2022detecting}.
\subsection{Heartbeat Detection}
For deepfake detection that used heartbeat, a research paper, \cite{boccignone2022deepfakes} that leverages rPPG and heartbeat was reviewed. The suggested deepfake detection method based on the physiological heart rate estimate postulates that when calculating the localized pulse rate from a picture of a genuine head vs. a counterfeit one, rPPG (remote photoplethysmography) approaches should yield noticeably different findings. In their proposed method, after detecting the face of the (susceptible) manipulated subject, a group of 100 patches is automatically tracked on it. The pixels color intensities within time are averaged, thus forming 100 RGB traces. Further, for each patch, blood Volume Pulse is quoted. Disruption is anticipated in each patch with BVP signals. Thus, certain features analyzed pertain to have a clear meaning as to such physiological behavior which lead to DeepFake detection otherwise . In \cite{fernandes2019predicting}, it is believed that the majority of the current Deepfake altering algorithms do not yet account for the physiological characteristics of a human being since the color changes and lighting induced by oxygen content are slight and imperceptible to the human eye. The early design of Deepfakes ON Phys was based on the Big Phys model, whose goal was to estimate the actual heart rate using face video sequences\cite{fernandes2019predicting}. The deep learning-based model's objective was to collect temporal data from movies in a manner resembling the behavior of classic manual rPPG procedures. It is possible to extract facial traits by observing the change in color in individuals' faces, that are brought on by variations in blood oxygen levels. Additionally, signal processing techniques are employed to differentiate between color changes brought on by plasma and other alterations that may be brought on by noise and ambient light.
\vspace{3mm}
\\A thorough experimental evaluation of DeepFakesON-Phys uses Celeb-DF v2 and DFDC Preview, two of the most current available datasets from the second Deepfake generation. Google created one of the biggest public databases accessible via partnerships with other corporations and major universities including Microsoft, Amazon, as well as MIT. The name DFDC appears in the database. The 1,131 genuine recordings made by 66 performers who were contracted for the DFDC Preview collection provide realistic diversity in genders, skin color, and age. Celeb-DF v2 is one of the most difficult Deepfake datasets to use at the present, and it is important to emphasize that, unlike other well-known databases, this collection was not produced using either publicly accessible information or data from social media platforms. The Celeb-DF v2 database's goal was to produce fake videos with superior graphics than those in the UADFV database. This collection includes 590 actual YouTube videos on well-known individuals who identify with various racial, ethnic, or gender identities.
\vspace{3mm}
\\The recommended method analyses both short-term video level and frame-level fake detection while assessing DeepFakesON-Phys. In both tests, DeepFakesON-Phys outperforms the state-of-the-art and produces amazingly accurate answers.
\vspace{3mm}
\\Table 2 represents the biological feature and classifiers / techniques, and used data sets, key findings of biological features.
\begin{table}
\caption{Comparison of biological features}
\centering
\begin{tabular}{|c|l|l|l|l|l|}
\toprule
\textbf{Ref} &\textbf{ Biological} & \textbf{Classifiers/ Techniques} & \textbf{Datasets Used} & \textbf{Dealing} & \textbf{Key Findings}\\
&\textbf{ feature}& & \textbf{ with}& &\\
\midrule
\hline&&&&&\\
.\cite{mao2021exposing} & Heart Rate &Photoplethysmography & Face & Images & The plan increases overall\\
& & and autoregressive & Forensic ++ & & accuracy for the Face\\
& & (AR) model & dataset, Celeb-DF & & Forensic ++ dataset by \\
& & & dataset & & {25.66\%} when associated with\\
& & & & & the baseline network and by\\
& & & & & {1.48\%} when compared to the\\
& & & & & baseline approach.\\
&&&&&\\\hline &&&&&\\
.\cite{li2018ictu} & Eye &LRCN method&CEW Dataset&Images/&This approach is evaluated\\
&movement&&&Videos& using benchmark datasets\\
&&&&& for eye-blinking detection,\\
&&&&& and it exhibits good results\\
&&&&& when used to identify films\\
&&&&& produced using Deepfake.\\
&&&&&\\\hline &&&&&\\
.\cite{nguyen2020eyebrow} & Eyebrow &LightCNN, ResNet, & Celeb-DF & Images & Because once implemented to \\
& movement & DenseNet, SqueezeNet & [Li20] & & the best quality Deepfake\\
&&&&& dataset, our technique offered\\
&&&&& a 0.88 AUC and 20.7\% EER\\
&&&&& for Deepfake identification.\\
&&&&&\\\hline &&&&&\\
.\cite{agarwal2021detecting}&Ear and&Aural dynamics&64 videos were&Videos&The measured signal unfolds\\
&mouth& &downloaded& & across hundreds of frames\\
&movement& &from YouTube of& & in dynamic auditory\\
& & &Joe Biden,&& and oral analysis, in contrast\\
& & &Angela Merkel,& & to existing synthesis approaches,\\
& & &Donald Trump,& & which generally work on\\
& & &and Mark & & one or a small number\\
& & &Zuckerberg.& & of video frames. \\
&&&&&\\\hline &&&&&\\
.\cite{gazi2021deepfake}& Eyelid & LRCN method, & UBIRIS, & Images & In terms of eyelid aperture\\
.\cite{fuhl2016eyes} & & computer vision-based & CASIA & & estimate and eyelid outline\\
& & approach & & & similarity, the suggested \\
& & & & & technique performs better than\\
& & & & & the state-of-the-art.\\
&&&&&\\\hline &&&&&\\
.\cite{jung2020deepvision}& Eyeblink & adversarial network & Face & Images & The recommended method\\
& & (GANs) model & Forensics & & called as Deep Vision\\
& & & ++& & is utilized as a measurement\\
& & & & & to verify an abnormality\\
& & & & & whenever eye blinks are\\
& & & & & frequently replayed in a very\\
& & & & &short period of time depending\\
& & & & & on the period, the repeating\\
& & & & & frequency, as well as expiring \\
& & & & & eye blink length.\\
&&&&&\\\hline &&&&&\\
.\cite{lin2017pulse} & Pulse rate & PPG sensors & ECG signal & Images & According to the findings of \\
&&&&& the experiment, the suggested\\
&&&&& strategy may improve the\\
&&&&& usefulness and likelihood of\\
&&&&& PPG signal identification on\\
&&&&& palms.\\
\bottomrule
\end{tabular}
\label{tab:table}
\end{table}
\section{DISCUSSION}
In this paper, we mainly concentrated on both the generation and detection of Deepfakes and also existing Deepfake detection techniques and their difficulties. However, biological aspects like recognizing an eyebrow, detecting an eye blink, detecting an eye movement, detecting an ear or mouth, and detecting a pulse are precisely defined. The important findings from a comparison of each biological trait using various classifiers and methodologies are highlighted. Finding the integrity verification index values in the GANs method that are extremely challenging to evaluate requires utilizing the discriminator, producing fake pictures, a single recurring and subconscious activity, which provides a new technique. The similarity of the data generated by GANs to actual data is one of its most alluring features. They, therefore, have many different real-world uses. They can create text, pictures, audio, and video that are identical to real data. One pertinent scope was identified by studying these deepfake detection methods that use biological features for identification - can these biological features be altered by using high quality filters in video processing or by applying make-up to alter these features namely fine lines in face or eye-brow movement ? Use of these smoothing techniques can cause hindrance in identifying distinguishing features of facial regions that are leveraged by these identification methods. Thus, these methods may result in giving false positives to the video that are actually not deepfakes.
\bibliographystyle{plain}
|
\section{Introduction}
The first detection of gravitational waves (GWs) \cite{LIGOScientific:2016aoc} by advanced Laser Interferometer Gravitational-Wave Observatory (LIGO) Scientific Collaboration \cite{Harry:2010zz, LIGOScientific:2014pky} and Virgo Collaboration \cite{VIRGO:2014yos} opened a new window to test Einstein's general relativity and understand the nature of the gravity \cite{LIGOScientific:2016lio, LIGOScientific:2018dkp, LIGOScientific:2019fpa, LIGOScientific:2020tif}.
However, the ground-based GW observatories are only sensitive to GWs in the high frequency bands $10-10^3$ Hz due to seismic noise and gravity gradient noise.
The proposed space-based GW detectors such as Laser Interferometer Space Antenna (LISA) \cite{Danzmann:1997hm, LISA:2017pwj}, Taiji \cite{Hu:2017mde}, and TianQin \cite{TianQin:2015yph} will detect GWs in the low frequency regimes $10^{-4}-10^{-1}$ Hz.
Accurate sky localization is an important scientific object for space-based GW detectors because
the source position is strongly correlated with the other parameters of the binary source.
The accurate information about the source
localization is essential for the
follow-up observations of counterparts and the statistical
identification of the host galaxy when no counterpart is
present,
so that GWs can be used as standard sirens \cite{Schutz:1986gp, Holz:2005df} to probe the evolution of the Universe and study the problem of the Hubble tension \cite{Riess:2019cxk}.
Moreover, accurate knowledge about the GW source position
may provide important information about the environments where the binary source resides.
Because space-based GW detectors can measure GWs for months to years and
the detector's center moves in the orbit around the Sun, the Doppler modulation on the amplitude and phase of GWs carries the position information of the sources.
Therefore, a single space-based GW detector is able to locate the source position.
In contrast to ground-based detectors,
space-based GW detectors can detect long-duration GWs
and the motion of spacecrafts in space makes it hard to maintain the exact equality between the arm lengths.
Because of the unequal arm lengths of space-based GW detectors,
the laser frequency noise cannot cancel out.
To reduce the laser frequency noise, properly chosen time shifted and linear combinations were proposed to synthesize virtual equal arm interferometric measurements \cite{Tinto:1994kg,Tinto:1999yr,Armstrong_1999}.
This technique is known as time-delay interferometry (TDI) \cite{Tinto:1999yr,Armstrong_1999,Estabrook:2000ef,Tinto:2003vj,Vallisneri:2004bn}.
The first-generation TDI combinations can cancel out the laser frequency noise in a static unequal-arm configuration,
and the second-generation TDI further cancels the frequency noise in time-dependent arm interferometry.
The first-generation TDI configurations were applied for space-based GW detectors such as LISA, Taiji and TianQin in \cite{Estabrook:2000ef,Tinto:2001ii, Tinto:2001ui,Hogan:2001jn,Armstrong:2001uh,Prince:2002hp,Tinto:2002de,Shaddock:2003dj,Tinto:2003uk,Nayak:2003na,Tinto:2004nz,Romano:2006rj,Zhang:2020khm},
and the applications of second-generation TDI configurations were discussed in \cite{Tinto:2003vj,Cornish:2003tz,Vallisneri:2004bn,Krolak:2004xp,Vallisneri:2005ji,Wang:2017aqq,Wang:2020fwa,RajeshNayak:2004jzp,Nayak:2005un}.
For more discussion on TDI algorithm and its application to space-based GW detectors, please see \cite{Tinto:2004wu,Tinto:2014lxa,Tinto:2020fcc} and references therein.
There are two different constellations for space-based GW detectors.
For LISA/Taiji, the spacecrafts are in the heliocentric orbit behind/ahead of the Earth by about $20^\circ$,
the inclination angle between the plane of constellations and ecliptic plane is $60^\circ$, and they keep the geometry of an equilateral triangle with the arm length $L=2.5\times10^9$ m$/3\times10^9$ m.
For TianQin, the spacecrafts are in the geocentric orbit around the Earth and rotate around the Sun,
the arm length is $L=1.73\times10^8$ m, and
the normal vector of the detector plane points to the source RX J0806.3+1527 at ($\theta_{tq}=-4.7^{\circ}$, $\phi_{tq}=120.5^{\circ}$).
The main differences between different constellations of space-based GW detectors are the time-changing orientation of the detector plane,
the arm length, the orbital period of spacecrafts and the noise curve \cite{Zhang:2020hyx}.
It is necessary to discuss the effects of
these factors on the accuracy of source localization.
As discussed in \cite{Zhang:2020hyx,Zhang:2020drf,Gong:2021gvw} for monochromatic sources,
the amplitude modulation due to the rotation of the detector plane not only helps LISA and Taiji get better accuracy in the sky localization of GW source but also enlarges the sky coverage at frequencies below 1 mHz.
The ability in the sky localization of TianQin is better than LISA and Taiji at frequencies above 30 mHz
and TianQin has blind spots for sources from the directions with $\phi_s$ around $30^\circ$ or $-150^\circ$.
At higher frequencies when the wavelength of GWs is larger
than the detector's arm length,
the frequency-dependent transfer function
deteriorates the signal-to-noise ratio (SNR) registered in the detector.
For the network of LISA and TianQin, it has a better ability in sky localization for the sources with frequencies in the range 1-100 mHz, and the network has larger sky coverage for the angular resolution than the individual detector.
The discussion was based on the equal-arm Michelson interferometric detector.
But for space-based GW detectors,
we should apply the TDI method in the analysis of sky localization of GW sources.
In this paper, we extend the previous work on the analysis of the main factors that influence the accuracy of source localization by considering the TDI configuration.
In Sec. \ref{meth}, we provide a brief overview of TDI response and Fisher information matrix (FIM) method.
In Sec. \ref{result}, we consider fiducial detectors to discuss the effects of different factors on the sky localization for monochromatic GWs,
and compare the results of angular resolution between LISA, Taiji, TianQin, and their network.
Finally, we present the conclusion in Sec. \ref{discussion}.
\section{Methodology}\label{meth}
In this section, we provide a brief overview on TDI responses of space-based GW detectors to monochromatic GWs and parameter estimations with the method of FIM.
\subsection{The GW signal}
\label{wave}
For GWs propagating along the direction $\hat{\Omega}(\theta,\phi)$, we can define two perpendicular unit vectors $\hat{p}$ and $\hat{q}$ that satisfy the orthogonal relation $\hat{\Omega}=\hat{p}\times \hat{q}$.
To account for the rotational degree of freedom around $\hat{\Omega}$, we introduce the polarization angle $\psi$ to form two new orthonormal vectors
\begin{equation}
\hat{r}=\cos(\psi) \hat{p}+\sin(\psi)\hat{q},\qquad\hat{s}=-\sin(\psi) \hat{p}+\cos(\psi)\hat{q}.
\end{equation}
With these two orthonormal bases, the plus $(+)$ and cross $(\times)$ polarization tensors are
\begin{equation}
e^+_{ij}=\hat{r}_i\hat{r}_j-\hat{s}_i\hat{s}_j, \qquad e^\times_{ij}=\hat{r}_i\hat{s}_j+\hat{r}_j\hat{s}_i,
\end{equation}
and the GW signal can be expressed as
\begin{equation}
h_{ij}(t)=\sum_{A=+,\times} e_{ij}^A h_A(t).
\end{equation}
For monochromatic GWs with the frequency $f_0$ emitted from sources such as compact binaries containing stellar or intermediate-mass black holes, white dwarfs or neutron stars, in the lowest order of quadrupole approximation the tensor modes are
\begin{equation}
\label{monowave}
\begin{split}
&h_+(t) = \mathcal{A}[1+ \cos^2(\iota)]e^{(2\pi i f_0 t + i \phi_0)},\\
&h_\times(t) = 2i \mathcal{A} \cos^2(\iota) e^{(2\pi i f_0 t + i \phi_0)},
\end{split}
\end{equation}
where the amplitude $\mathcal{A}=2m_1 m_2(\pi f_0)^{2/3}/[(m_1+m_2)^{1/3} d_L]$, $m_1$ and $m_2$ are masses of the binary components, $d_L$ is the luminosity distance between the source and the observer, $\phi_0$ is the initial phase of the orbital plane and $\iota$ is the inclination angle of the detector plane.
The signal registered in the detector $\alpha$ is
\begin{equation}
\label{signal}
z_{\alpha}(t)=\sum_A F^A_{\alpha}(f,\theta,\phi,\psi) h_A(t) e^{i\phi_D(t)} + n_\alpha(t),
\end{equation}
where $n_\alpha(t)$ is the detector noise, $\phi_D(t)$ is the Doppler phase, and the $F^A_{\alpha}(f,\theta,\phi,\psi)$ is the response function in the detector $\alpha$ \cite{Estabrook:2000ef}.
For GW signals with the frequency $f_0$, the Doppler phase is
\begin{equation}
\label{dphase}
\phi_{D}(t)=2 \pi f_0 R_s \sin \theta \cos \left(\frac{2 \pi t}{P}-\phi-\phi_{\alpha}\right),
\end{equation}
where $R_s=1$ AU is the distance between the sun and the earth, $\theta$ and $\phi$ are the angular position of the GW source in the ecliptic coordinate, $P$ is the period of rotation for the earth around the sun, and $\phi_{\alpha}$ is the detector's ecliptic longitude at $t=0$, so the source parameters are $\bm{\Lambda} = \left(\theta,\ \phi,\ \mathcal{A},\ \iota,\ \psi,\ \phi_0\right)$.
\subsection{The TDI configurations and fiducial detectors}\label{instrument}
For space-based GW detectors, it is difficult to maintain the exact equality of the arm lengths due to the motion of spacecrafts (SCs),
so laser frequency noise cannot be effectively removed when two beams are recombined directly
at the photo-detector.
Fortunately, the multiple readouts of space-based GW detectors can provide observables that are insensitive to laser fluctuation and optical-bench motions by properly choosing time shifted and linear combinations.
This technique is known as TDI.
There are six configurations for the first-generation TDI combinations as shown in Fig. \ref{schematic},
Sagnac $\zeta$, six-pulse $(\alpha, \beta, \gamma)$, unequal-arm Michelson $(X, Y, Z)$,
Relay $(U, V, W)$, Beacon $(P, Q, R)$, and monitor $(E, F, G)$.
The spacecrafts are labeled as 1, 2, 3,
the optical paths between two SCs are denoted by $L_a$ with the unit vector $n_a$,
and the index $a$ corresponds to the opposite SC.
The three channels of each configuration are obtained by cyclical permutation of the spacecraft indices.
For the delayed data stream, following \cite{Estabrook:2000ef, Prince:2002hp},
we donate the $y_{ab,c}$ as the relative frequency fluctuations time series measured from reception at $\text{SC}_b$ with transmission from $\text{SC}_d$ ($d\equiv a$ and $d\equiv b$) along $L_a$ by delayed $L_c$, and the notation, for example, are $y_{31,2}=y_{31}(t-L_2)$ and $y_{31,32}=y_{31}(t-L_3-L_2)$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\columnwidth]{schematic.pdf}
\caption{The configurations of six first-generation TDI combinations.}
\label{schematic}
\end{figure}
For the response of the one-way transmission, as shown in Fig. \ref{schematic2}, between $\text{SC}_1$ and $\text{SC}_2$ with the distance $L$ along the unit direction $\hat{n}$ is \cite{Larson:2002xr}
\begin{equation}
\label{resp1}
\begin{split}
\delta L&=\sum_{A,i,j}\hat{n}_i\hat{n}_j e^A_{ij}\frac{\sin[\omega L(1-\hat{n}\cdot\hat{\Omega})/2]}{\omega(1-\hat{n}\cdot\hat{\Omega})}e^{-i\omega L(1-\hat{n}\cdot\hat{\Omega}+2\Omega\cdot\vec{r}_2/L)/2}h^A(f)\\
&=\sum_{A,i,j}L \hat{n}_i\hat{n}_j e^A_{ij}T(\omega,\hat{n}\cdot\hat{\Omega})h^A(f),
\end{split}
\end{equation}
where $\omega=2\pi f$ is the angular frequency of GWs, $\hat{\Omega}$ is the propagating direction of GWs.
and we take the speed of light $c=1$.
The transfer function $T(\omega,\hat{n}\cdot\hat{\Omega})$ is \cite{Cornish:2001qi}
\begin{equation}
\label{transf1}
\begin{split}
T(\omega,\hat{n}\cdot\hat{\Omega})&=\frac{\sin[\omega L(1-\hat{n}\cdot\hat{\Omega})/2]}{\omega L(1-\hat{n}\cdot\hat{\Omega})}e^{-i\omega L(1-\hat{n}\cdot\hat{\Omega}+2\Omega\cdot \vec{r}_2/L)/2}\\
&=\frac{1}{2}\text{sinc}\left[\omega L(1-\hat{n}\cdot\hat{\Omega})/2\right]e^{-i\omega L(1-\hat{n}\cdot\hat{\Omega}+2\Omega\cdot \vec{r}_2/L)/2},
\end{split}
\end{equation}
where $\text{sinc}(x)=\sin{x}/x$ and $\vec{r}_2$ is the location of $\text{SC}_2$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.6\columnwidth]{schematic2.pdf}
\caption{The one-way Doppler tracking.
We put $\text{SC}_1$ at the origin of the coordinate,
the other spacecrafts $\text{SC}_2$ and $\text{SC}_3$ are located at $\vec{r}_2$ and $\vec{r}_3$.
GWs propagate along the direction $\hat{\Omega}$, the laser beam transmits along $\hat{n}$, and the length between $\text{SC}_2$ and $\text{SC}_3$ is $L$.}
\label{schematic2}
\end{figure}
By using Eqs. (\ref{resp1}) and (\ref{transf1}), the six TDI signal responses of GW are
\begin{equation}
\begin{split}
y_{ab}\left(T_D\right)&=-\frac{\delta L_a}{L_e}=-\sum_{A, i, j} g_a \hat{n}_a^i \hat{n}_a^j e_{i j}^A T_{a b}\left(T_D\right) h^A(f),\\
T_{12}\left(T_D\right)&=\frac{1}{2} \operatorname{sinc}\left[\frac{1}{2} u_1\left(1-\mu_1\right)\right] e^{-i u_1\left(1-\mu_1\right) / 2-i\left(\mu_3 u_3+T_D\right)},\\
T_{23}\left(T_D\right)&=\frac{1}{2} \operatorname{sinc}\left[\frac{1}{2} u_2\left(1-\mu_2\right)\right] e^{-i u_2\left(1-\mu_2\right) / 2-i\left(\mu_2 u_2+T_D\right)},\\
T_{32}\left(T_D\right)&=\frac{1}{2} \operatorname{sinc}\left[\frac{1}{2} u_3\left(1-\mu_3\right)\right] e^{-i u_3\left(1-\mu_3\right) / 2-i\left(\mu_3 u_3+T_D\right)},\\
T_{31}\left(T_D\right)&=\frac{1}{2} \operatorname{sinc}\left[\frac{1}{2} u_3\left(1+\mu_3\right)\right] e^{-i u_3\left(1+\mu_3\right) / 2-i T_D},\\
T_{13}\left(T_D\right)&=\frac{1}{2} \operatorname{sinc}\left[\frac{1}{2} u_1\left(1+\mu_1\right)\right] e^{-i u_1\left(1+\mu_1\right) / 2-i\left(\mu_2 u_2+T_D\right)},\\
T_{21}\left(T_D\right)&=\frac{1}{2} \operatorname{sinc}\left[\frac{1}{2} u_2\left(1+\mu_2\right)\right] e^{-i u_2\left(1+\mu_2\right) / 2-i T_D},
\end{split}
\end{equation}
where $g_a=L_a/L_e$, $L_e$ is the expectation length, $u=(2\pi f) L_e$, $u_a=(2\pi f) L_a$, $\mu=\hat{n}_a \cdot \hat{\Omega}$, $T_D$ is the corresponding time delay, for example, $T_D=u_1+u_2$ for $y_{ab,12}$, and the response function $F^A_{\alpha}(f,\theta,\phi,\psi)=g_a \hat{n}_a^i \hat{n}_a^j e_{i j}^A T_{a b}\left(T_D\right)$.
Due to the cancellation of frequency noises in the stationary unequal-arm interferometry by the first-generation TDI combinations,
the dominant detector noises are the acceleration noise and the single-link optical metrology noise.
The acceleration noises for LISA \cite{Robson:2018ifk} and Taiji \cite{Wang:2020fwa} are
\begin{equation}
\begin{split}
&S^{a}_\text{LISA} = S^{a}_\text{TJ}=(3\times 10^{-15}\ \text{m s}^{-2})^2 \bigg[1+\big(\frac{0.4 \ \text{mHz}}{f}\big)^2\bigg]\bigg[1+\big(\frac{f}{8 \ \text{mHz}}\big)^4\bigg]\ \text{Hz}^{-1},
\end{split}
\end{equation}
the single-link optical metrology noises for LISA and Taiji are
\begin{equation}
\begin{split}
&S^{x}_\text{LISA} = (1.5\times 10^{-11}\ \text{m})^2 \bigg[1+\big(\frac{2 \ \text{mHz}}{f}\big)^4\bigg]\ \text{Hz}^{-1},\\
&S^{x}_\text{TJ} = (8\times 10^{-12}\ \text{m})^2 \bigg[1+\big(\frac{2 \ \text{mHz}}{f}\big)^4\bigg]\ \text{Hz}^{-1},
\end{split}
\end{equation}
respectively.
The acceleration noise and the single-link optical metrology noise
for TianQin are \cite{TianQin:2015yph}
\begin{equation}
\begin{split}
&S^a_\text{TQ}=10^{-30}\text{m}^2\ \text{s}^{-4}/\text{Hz},\\
&S^x_\text{TQ}=10^{-24}\text{m}^2/\text{Hz}.
\end{split}
\end{equation}
The shot and proof mass noises are related to the single-link optical metrology and acceleration noises as
\begin{equation}
\label{tdinoise1}
\begin{split}
S_y^{\rm shot}=S_{y}^{x}/L^2,\\
S_y^{\rm proof\ mass}=S_y^{a}/(L^2(2\pi f)^4),
\end{split}
\end{equation}
where $L$ is the arm length of the detector $y$. For LISA, $L=2.5\times10^9$ m; for Taiji, $L=3\times10^9$ m; and for TianQin, $L=1.78\times10^8$ m.
The noises in different TDI combinations are \cite{Estabrook:2000ef}
\begin{equation}
\label{pnalpha}
P_n^\alpha=[8\sin^2(3\pi f L)+16\sin^2(\pi fL)]S_y^{\rm proof\ mass}+6 S_y^{\rm shot},
\end{equation}
\begin{equation}
\label{pnzeta}
P_n^\zeta=24\sin^2(\pi fL) S_y^{\rm proof\ mass}+6S_y^{\rm shot},
\end{equation}
\begin{equation}
\label{pnx}
P_n^X=[8\sin^2(4\pi f L)+32\sin^2(2\pi fL)]S_y^{\rm proof\ mass}+16\sin^2(2\pi fL)S_y^{\rm shot},
\end{equation}
\begin{equation}
\label{pnp}
P_n^P=[8\sin^2(2\pi f L)+32\sin^2(\pi fL)]S_y^{\rm proof\ mass}+[8\sin^2(2\pi fL)+8\sin^2(\pi fL)]S_y^{\rm shot},
\end{equation}
\begin{equation}
\label{pne}
P_n^E=[32\sin^2(\pi f L)+8\sin^2(2\pi fL)]S_y^{\rm proof\ mass}+[8\sin^2(\pi fL)+8\sin^2(2\pi fL)]S_y^{\rm shot},
\end{equation}
\begin{equation}
\label{pnu}
\begin{split}
P_n^U=&[16\sin^2(\pi f L)+8\sin^2(2\pi fL)+16\sin^2(3\pi fL)]S_y^{\rm proof\ mass}\\
&+[4\sin^2(\pi fL)+8\sin^2(2\pi fL)+4\sin^2(3\pi fL)]S_y^{\rm shot}.
\end{split}
\end{equation}
Since GWs come from all directions,
we use the average response function $R(f)$
by taking the average of the response function $F^A$ overall source directions and the polarization angle \cite{Zhang:2019oet}.
The sensitivity curve is $S_n(f)=P_n(f)/R(f)$.
In the top panel of Fig. \ref{sensitivity} we plot the sensitivity curve $\sqrt{S_n(f)}$ of LISA for the tensor mode for the six TDI configurations.
Since the $(X, Y, Z)$ configuration has the lowest sensitivity, we consider the $(X, Y, Z)$ configuration in particular, and the corresponding sensitivity curves for Taiji and TianQin are shown in the bottom panel of Fig. \ref{sensitivity}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\columnwidth]{sensitivity.pdf}
\caption{The sensitivity curves $\sqrt{S_n(f)}$ for the tensor mode.
The top panel shows the sensitivity curves $\sqrt{S_n(f)}$ of LISA for the six TDI configurations, and the bottom panel shows the sensitivity curves $\sqrt{S_n(f)}$ for LISA, Taiji, and TianQin for the $(X, Y, Z)$ combinations.}
\label{sensitivity}
\end{figure}
For different constellations of space-based GW detectors, three dominant factors affect the accuracy of source localization:
the rotation, the arm length, and the rotation period of the spacecraft \cite{Zhang:2020hyx}.
To examine the effects of these factors individually,
we use the control variable method and construct five fiducial GW detectors as shown in Table \ref{fdetector} \cite{Zhang:2020hyx}.
The LISA-like fiducial detectors R and R1 are in the heliocentric orbit with the orbital period of one year and the normal vector of the detector plane rotates around the normal vector of the ecliptic plane with a period of one year.
The TianQin-like fiducial detectors C, C1 and C2 are in the geocentric orbit and the normal vector of the detector plane points to the calibration source RX J0806.3+1527.
The orbital period of the fiducial detectors C and C1 is one year while the orbital period of the fiducial detector C2 is 3.65 days.
The detailed orbit equations for these fiducial detectors were given in Ref. \cite{Zhang:2020hyx}.
The fiducial detector R/R1 and C/C1 are constructed to study the effect of rotation,
the fiducial detector R/C and R1/C1 are constructed to study the effect of arm length, and
the fiducial detector C1 and C2 are constructed to study the effect of rotation period of the spacecrafts.
\begin{table}
\begin{tabular}{ |p{3cm}||c| c| c| c| c| c|}
\hline
Detector & $\text{R}$ & $\text{C}$ &$\text{R1}$ & $\text{C1}$ &$\text{C2}$\\
\hline
Arm Length (m) & $3.7\times10^9$ & $3.7\times10^9$ &$1.73\times10^8$ & $1.73\times10^8$ &$1.73\times10^8$\\
\hline
Period & $1\ \text{year}$ & $1\ \text{year}$ &$1\ \text{year}$ & $1\ \text{year}$ &$3.65\ \text{days}$\\
\hline
\end{tabular}
\caption{The parameters of the fiducial detectors. The LISA-like fiducial detectors R and R1 are in the heliocentric orbit,
the TianQin-like fiducial detectors C, C1 and C2 are in the geocentric orbit and their normal vector of the detector plane points to the calibration source RX J0806.3+1527.
}
\label{fdetector}
\end{table}
\subsection{FIM method}
\label{analysis}
Given two frequency-domain signals $h_1(f)$ and $h_2(f)$, the inner product $\langle h_1|h_2\rangle$ is defined as
\begin{equation}\label{overlap}
\langle h_1|h_2\rangle=2\int_0^{+\infty } \frac{h_1(f) h_2^*(f)+ h_2(f) h_1^*(f)}{ P_n(f)}\,df,
\end{equation}
where $h^*$ denotes the complex conjugate of $h$.
The SNR of GW signal $z_\alpha$ is
\begin{equation}
\label{snr}
\begin{split}
\rho^2=& \left< z_\alpha | z_\alpha \right>\\
=&4\ \int_{0}^{\infty}\frac{ z_\alpha(f) z_\alpha^*(f)}{P_{n}^\alpha(f)}df.
\end{split}
\end{equation}
In this paper, we choose the threshold of detecting a signal as $\rho\ge 10$.
To estimate the uncertainty of source parameters, we apply the FIM method.
The FIM is defined as
\begin{equation}
\label{tgamma}
\begin{split}
\Gamma_{ij}=& \left<\frac{\partial z_\alpha}{\partial \bm{\Lambda}_i}\left|\frac{\partial z_\alpha^*}{\partial \bm{\Lambda}_j}\right.\right>\\
=&4\ \text{Re} \int_{0}^{\infty}\frac{\partial_i z_\alpha(f)\partial_j z_\alpha^*(f)}{P_{n}^\alpha(f)}df\\
=&\frac{2}{P_{n}^\alpha(f)}\ \text{Re}\int_{-\infty}^{\infty}\partial_i z_\alpha(t)\partial_j z_\alpha^*(t)dt,
\end{split}
\end{equation}
where $\alpha$ represents the TDI channel, $\partial_i z_\alpha=\partial z_\alpha/\partial \bm{\Lambda}_i$ and $\bm{\Lambda}_i$ is the $i$th parameter of GW source.
For monochromatic sources with the frequency $f$, because there is almost no frequency evolution, by using Parseval's theorem \cite{Cutler:1997ta, Vecchio:2004ec} the noise function $P_{n}^\alpha(f)$ can be taken out from the integration.
The covariance matrix $\sigma_{ij}$ between the parameters errors $\Delta\bm{\Lambda}_i=\bm{\Lambda}_i-\langle\bm{\Lambda}_i\rangle$ and $\Delta\bm{\Lambda}_j$ can be approximated by the inverse of the Fisher matrix in the large SNR limit,
\begin{equation}
\sigma_{ij}=\left\langle\Delta\bm{\Lambda}_i\Delta\bm{\Lambda}_j\right\rangle\approx (\Gamma^{-1})_{ij}.
\end{equation}
The angular uncertainty of the sky localization is defined as
\begin{equation}
\Delta \Omega_s = 2\pi\left|\sin\theta\right|
\sqrt{\sigma_{\theta\theta}\sigma_{\phi\phi}-\sigma^2_{\theta\phi}}.
\end{equation}
For a network of $n$ detectors, the SNR and FIM are defined as $\rho^2=\Sigma^{n}_{\alpha=1}\rho_{\alpha}^2$ and $\Gamma_{ij}=\Sigma_{\alpha=1}^{n}\Gamma^{\alpha}_{ij}$,
respectively.
\section{Results}\label{result}
In this section, we show the sky map of the angular resolution with different fiducial detectors for monochromatic sources with the TDI $(X, Y, Z)$ combination
and make the comparison with the results without TDI.
We simulate 3600 GW sources with six parameters $\bm{\Lambda} = \left(\theta,\ \phi,\ \mathcal{A},\ \iota,\ \psi,\ \phi_0\right)$, $\theta$ and $\phi$ are chosen randomly in $\left[-\pi/2,\pi/2 \right]$ and $\left[-\pi,\pi \right]$, $\iota=1$, $\psi=\phi_0=0$.
Following Ref. \cite{Zhang:2020hyx}, we fix the amplitude $\mathcal{A}$ by considering sources with the same masses and distance randomly distributed in the sky.
We take the sources to be equal
mass binary system with the total mass $(6,3\times10^2,10^4)M_\odot$ at the distance $(2.3,1.3\times10^3,10^4)$ Mpc so that the minimum SNR $\rho>10$, the corresponding GW frequencies are $(10^{-1},10^{-2},10^{-3})$ Hz, respectively.
The frequency $f_0=1$ mHz, $f_0=0.01$ Hz and $f_0=0.1$ Hz represent the low, medium and high frequency.
The mean and median values of the angular resolutions are shown in Table \ref{tdivalue}, the sky maps of the angular resolutions are shown in Fig. \ref{map1}
and the ratios of the angular resolutions between different detectors are shown in
Figs. \ref{crratio}, \ref{cc1ratio} and \ref{cc2ratio}.
In Table \ref{tdivalue}, we also give the results without the TDI and the results show that the angular resolutions are almost the same with and without the TDI.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{omegaf.pdf}
\caption{The sky map of angular resolutions in the unit of steradian ($1\ \text{sr}=(180/\pi)^2\ \text{deg}^2$) for different fiducial detectors with the TDI $(X, Y, Z)$ combination. For left to right, the frequencies are $0.1$ Hz, $0.01$ Hz and $1$ mHz.}
\label{map1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\columnwidth]{cr.pdf}
\caption{The ratios of the mean and median values of angular resolutions for the detectors C and R, and the detectors C1 and R1.}
\label{crratio}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\columnwidth]{cc1.pdf}
\caption{The ratios of the mean and median values of angular resolutions for the detectors C and C1, and the detectors R and R1.}
\label{cc1ratio}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\columnwidth]{c1c2.pdf}
\caption{The ratios of the mean and median values of angular resolutions for the detectors C1 and C2.}
\label{cc2ratio}
\end{figure}
\subsection{The effect of detector configuration on sky localization}
To see the rotation effect of the detector plane on the uncertainty of sky localization,
we compare the results of angular resolutions for the fiducial detectors R and C, and for R1 and C1.
Note that the fiducial detectors have the same arm length and noise curve.
For the detector R, the angular resolutions $\Delta\Omega_s$ are from $9.49\times10^{-4}$ to $3.35\times10^{-3}$ steradians at $f_0=10^{-3}$ Hz, from $1.30\times10^{-5}$ to $2.90\times10^{-4}$ steradians at $f_0=10^{-2}$ Hz, and from $8.97\times10^{-8}$ to $7.93\times10^{-6}$ steradians at $f_0=10^{-1}$ Hz.
For the detector C, the range of angular resolutions $\Delta\Omega_s$ is from $1.01\times10^{-3}$ to $1.86\times10^{-1}$ steradians at $f_0=10^{-3}$ Hz,
from $8.86\times10^{-6}$ to $1.61\times10^{-3}$ steradians at $f_0=10^{-2}$ Hz,
and from $9.51\times10^{-8}$ to $1.58\times10^{-5}$ steradians at $f_0=10^{-1}$ Hz.
The effect of detector's changing orientation at frequency $f_0$ is to spread the measured power over a range $f_0\pm 2/P$.
For the orbital period $P=1$ year, $\delta f=2/P=6.3\times 10^{-8}$.
This amplitude modulation is independent of the sky location of the source,
henceforth can help the detector to localize sources across the sky.
On the other hand, the effect of Doppler shift is to spread the power over a range $f_0(1\pm v/c)$, where $v/c\sim10^{-4}$,
so the amplitude modulation is dominated at low frequency bands ($\lesssim 10^{-3}$ Hz) and the effects of the amplitude and Doppler modulations get roughly equal at $f_0 \sim 1$ mHz.
At higher frequency bands ($>1$ mHz), the Doppler modulation effect becomes the dominant
and the effect of detector's changing orientation can be neglected.
For the Doppler phase modulation, the angular resolutions for monochromatic sources at the frequency $f_0$ depend on the sky location of the source and is given as \cite{Blaut:2011zz}
\begin{equation}
\label{omegad}
\delta\Omega\approx 0.02(1{\rm mHz}/f_0)^2(10/\rho)^2\left|\sin(\theta)\right|^{-1}.
\end{equation}
The uncertainty of the sky localization is largest at the equatorial plane with the latitude $\theta=0$, and it is independent of $\phi$.
Except for R and R1 at the low frequency of 1 mHz because of the amplitude modulation, the equatorial patterns can be seen in Fig. \ref{map1}.
As the frequency increases, the error of the angular resolution decreases.
Since the amplitude modulation is effective at low frequency bands,
the angular resolutions of R and R1 are a little more than two times better than C and C1 at the frequency $f_0=1$ mHz,
and there is no equatorial pattern in the sky map for R and R1.
At the frequency $f_0=0.01$ Hz and $f_0=0.1$ Hz,
the effect of the amplitude modulation is negligible,
so the angular resolutions of R and R1 are almost the same as C and C1 and
equatorial patterns exist in the sky map.
We also plot the ratios of the mean and median values of the angular resolutions between R and C, and R1 and C1 in Fig. \ref{crratio}.
To study the effect of arm length on angular resolutions, we compare the results between the detectors R and R1, and C and C1.
As shown in Table \ref{fdetector}, the fiducial detectors R and R1 are in the heliocentric orbit and they have different arm length,
the fiducial detectors C and C1 are in the geocentric orbit and their main difference is in the arm length.
For the detector C1, the range of angular resolutions $\Delta\Omega_s$ is from $1.01\times 10^{-3}$ to $1.25\times10^{-1}$ steradians at $f_0=1$ mHz, from $7.18\times10^{-6}$ to $8.89\times10^{-4}$ steradians at $f_0=10^{-2}$ Hz, and from $4.96\times10^{-9}$ to $6.11\times10^{-7}$ steradians at $f_0=10^{-1}$ Hz.
For the detector R1, the range of angular resolutions $\Delta\Omega_s$ is from $9.55\times10^{-4}$ to $3.29\times10^{-3}$ steradians at $f_0=1$ mHz,
from $8.27\times10^{-6}$ to $2.15\times10^{-4}$ steradians at $f_0=10^{-2}$ mHz, and from $6.21\times10^{-9}$ to $2.95\times10^{-7}$ steradians at $f_0=10^{-1}$ Hz.
The arm length of the detectors R and C is $3.7\times 10^9$ m and the corresponding transfer frequency $f^*=c/2\pi L=0.013$ Hz.
The arm length of the detectors R1 and C1 is $1.73\times 10^8$ m and the corresponding transfer frequency $f^*=c/2\pi L=0.28$ Hz.
At frequencies $f_0=1$ mHz and $f_0=0.01$ Hz, the wavelength of GWs is larger than the arm length, so the results for R and R1, C and C1 are similar.
However, the frequency $f_0=0.1$ Hz is larger than the transfer frequency of R and C, but smaller than the transfer frequency of R1 and C1,
so the uncertainties of the sky localization for the detectors R1 and C1 are smaller than those from detectors R and C at the frequency $f_0=0.1$ Hz.
The ratios of the mean and median values of the angular resolutions between the detectors R and R1, and C and C1 are shown in Fig. \ref{cc1ratio}.
From the results in Table \ref{tdivalue}, Figs. \ref{map1} and \ref{cc1ratio},
we see that the effect of the arm length on the angular resolutions mainly comes from the noise curve
and it is almost the same for both heliocentric and geocentric constellations.
Finally, we compare the results of angular resolutions for the fiducial detectors C1 and C2 to study the effect of the orbital period of spacecrafts.
The period is always 1 year in the heliocentric orbit,
so we construct the fiducial detectors C1 and C2 in the geocentric orbit with different orbital periods.
For the detector C2, the range of angular resolutions $\Delta\Omega_s$ is from $1.01\times10^{-3}$ to $1.23\times10^{-1}$ steradians at $f_0=1$ mHz,
from $7.18\times10^{-6}$ to $8.74\times10^{-4}$ steradians at $f_0=10^{-2}$ Hz,
and from $4.95\times10^{-9}$ to $5.94\times10^{-7}$ steradians at $f_0=10^{-1}$ Hz.
The ratio of the mean and median values of angular resolutions for the detectors C1 and C2 is plotted in Fig. \ref{cc2ratio}.
From the results in Table \ref{tdivalue}, Figs. \ref{map1} and \ref{cc2ratio},
we see that the orbital period of spacecrafts has little effect on angular resolutions.
\begin{table}
\centering
\begin{tabular}{ p{1.2cm}|c c c c c }
\hline\hline
&\multicolumn{5}{c}{The mean value of $\Delta\Omega_s$}\\
\hline
$f\ (\text{Hz})$ & $\text{R}$ & $\text{C}$ &$\text{R1}$ & $\text{C1}$ &$\text{C2}$\\
\hline
$0.001$ & $2.3857\times10^{-3}$ & $5.3375\times10^{-3}$ &$2.3872\times10^{-3}$ & $5.2051\times10^{-3}$ &$5.2218\times10^{-3}$\\
& $(2.3920\times10^{-3})$ & $(5.3516\times10^{-3})$ &($2.3872\times10^{-3})$ & $(5.2051\times10^{-3})$ &$(5.2218\times10^{-3})$\\
\hline
$0.01$ & $4.1597\times10^{-5}$ & $4.6816\times10^{-5}$ &$3.4595\times10^{-5}$ & $3.7093\times10^{-5}$ &$3.7210\times10^{-5}$\\
& $(4.1947\times10^{-5})$ & $(4.7211\times10^{-5})$ &$(3.4596\times10^{-5})$ & $(3.7094\times10^{-5})$ &$(3.7211\times10^{-5})$\\
\hline
$0.1$ & $5.6327\times10^{-7}$ & $5.9101\times10^{-7}$ &$2.7335\times10^{-8}$ & $ 2.5636\times10^{-8}$ &$2.5563\times10^{-8}$\\
& $(5.7017\times10^{-7})$ & $(5.9825\times10^{-7})$ &$(2.7376\times10^{-8})$ & $( 2.5598\times10^{-8})$ &$(2.5525\times10^{-8})$\\
\hline\hline
&\multicolumn{5}{c}{The median value of $\Delta\Omega_s$}\\
\hline
$f\ (\text{Hz})$ & $\text{R}$ & $\text{C}$ &$\text{R1}$ & $\text{C1}$ &$\text{C2}$\\
\hline
$0.001$ & $2.4022\times10^{-3}$ & $2.9751\times10^{-3}$ &$2.4020\times10^{-3}$ & $2.9449\times10^{-3}$ &$2.9548\times10^{-3}$\\
& $(2.4085\times10^{-3})$ & $(2.9829\times10^{-3})$ &$(2.4020\times10^{-3})$ & $(2.9449\times10^{-3})$ &$(2.9548\times10^{-3})$\\
\hline
$0.01$ & $2.9310\times10^{-5}$ & $2.6108\times10^{-5}$ &$2.5348\times10^{-5}$ & $2.0986\times10^{-5}$ &$2.1055\times10^{-5}$\\
& $(2.9557\times10^{-5})$ & $(2.6108\times10^{-5})$ &$(2.5349\times10^{-5})$ & $(2.0986\times10^{-5})$ &$2.1056\times10^{-5})$\\
\hline
$0.1$ & $3.0053\times10^{-7}$ & $3.0788\times10^{-7}$ &$1.7338\times10^{-8}$ & $1.4467\times10^{-8}$ &$1.4422\times10^{-8}$\\
& $(3.0421\times10^{-7})$ & $(3.1165\times10^{-7})$ &$(1.7364\times10^{-8})$ & $(1.4489\times10^{-8})$ &$(1.4400\times10^{-8})$\\
\hline\hline
\end{tabular}
\caption{The mean and median values of angular resolutions $\Delta\Omega_s$ for different fiducial detectors with and without the TDI method. The results in brackets are for those without TDI.}
\label{tdivalue}
\end{table}
\subsection{The effect of network}
In this subsection, we apply the TDI method to space-based GW detectors LISA, Taiji, TianQin and the LISA-Taiji-TianQin network.
The sky map of angular resolutions is shown in Fig. \ref{mapnetwork}.
The mean and median values of angular resolutions are shown in Table \ref{netvalue}.
In Table \ref{netvalue} we also show the mean and median values of angular resolutions for space-based GW detectors and their network without applying the TDI method.
The results show that the effect of the TDI is negligible.
\begin{table}
\centering
\begin{tabular}{ p{1.2cm}|c c c c}
\hline\hline
&\multicolumn{4}{c}{The mean value of $\Delta\Omega_s$}\\
\hline
$f\ (\text{Hz})$ & $\text{LISA}$ & $\text{Taiji}$ &$\text{TianQin}$ & $\text{Network}$ \\
\hline
$0.001$ & $2.6849\times10^{-3}$ & $1.3030\times10^{-3}$ &$9.1255\times10^{-2}$ & $ 6.1312\times10^{-4}$ \\
& $(2.6881\times10^{-3})$ & $(1.3055\times10^{-3})$ &$(9.1256\times10^{-2})$ & $(6.1410\times10^{-4})$ \\
\hline
$0.01$ & $3.6733\times10^{-5}$ & $7.8631\times10^{-6}$ &$4.1719\times10^{-5}$ & $4.3272\times10^{-6}$ \\
& $(3.6890\times10^{-5})$ & $(8.0162\times10^{-6})$ &$(4.1725\times10^{-5})$ & $(4.3838\times10^{-6})$ \\
\hline
$0.1$ & $3.0966\times10^{-7}$ & $1.5947\times10^{-7}$ &$2.3068\times10^{-8}$ & $1.6795\times10^{-8}$ \\
& $(3.1250\times10^{-7})$ & $(1.5947\times10^{-7})$ &($2.3068\times10^{-8})$ & $(1.6808\times10^{-8})$ \\
\hline\hline
&\multicolumn{4}{c}{The median value of $\Delta\Omega_s$}\\
\hline
$f\ (\text{Hz})$ & $\text{LISA}$ & $\text{Taiji}$ &$\text{TianQin}$ & $\text{Network}$ \\
\hline
$0.001$ & $2.5871\times10^{-3}$ & $1.3051\times10^{-3}$ &$5.1638\times10^{-2}$ & $6.0841\times10^{-4}$ \\
& $(2.59021\times10^{-3})$ & $(1.3076\times10^{-3})$ &$(5.1638\times10^{-2})$ & $(6.0939\times10^{-4})$ \\
\hline
$0.01$ & $2.7778\times10^{-5}$ & $5.7658\times10^{-6}$ &$2.3607\times10^{-5}$ & $2.8500\times10^{-6}$ \\
& $(2.79001\times10^{-5})$ & $(5.8981\times10^{-6})$ &$(2.3610\times10^{-5})$ & $(2.8856\times10^{-6})$ \\
\hline
$0.1$ & $1.6370\times10^{-7}$ & $7.1032\times10^{-8}$ &$1.3014\times10^{-8}$ & $8.8234\times10^{-9}$ \\
& $(1.6521\times10^{-7})$ & $(7.1032\times10^{-8})$ &$(1.3014\times10^{-8})$ & $(8.8313\times10^{-9})$ \\
\hline\hline
\end{tabular}
\caption{The mean and median values of angular resolutions $\Delta\Omega_s$ for space-based GW detectors and their network with and without the TDI method. The results in brackets are for those without TDI.}
\label{netvalue}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{omeganet.pdf}
\caption{The sky map of angular resolutions in the unit of steradian for LISA, Taiji, TianQin, and the LISA-Taiji-TianQin network with the TDI $(X, Y, Z)$ combination. For left to right, the frequencies are $0.1$ Hz, $0.01$ Hz and $1$ mHz.}
\label{mapnetwork}
\end{figure}
At the frequency $f_0=1$ mHz, the amplitude modulation helps LISA and Taiji on improving the sky localization and there is no equatorial pattern in the sky map;
for the sky localization, LISA is better than TianQin by more than two orders of magnitude with the help of the noise curve,
Taiji is two times better than LISA,
and the LISA-Taiji-TianQin network is $2.5$ times better than Taiji.
At the relative medium and high frequencies $f_0=0.01$ Hz and $f_0=0.1$ Hz,
the amplitude modulation is negligible and the Doppler modulation dominates,
the equatorial pattern exists in the sky map as shown in Fig. \ref{mapnetwork}.
At the medium frequency $f_0=0.01$ Hz,
the GW wavelength is still bigger than the arm length of LISA and Taiji;
for the sky localization, LISA is less than two times better than TianQin,
Taiji is almost five times better than LISA,
and the LISA-Taiji-TianQin network is less than two times better than Taiji.
At the relatively high frequency $f_0=0.1$ Hz,
the GW wavelength is smaller than the arm length of LISA and Taiji,
but bigger than the arm length of TianQin,
so TianQin has a better ability of sky localization than LISA and Taiji.
For the sky localization,
Taiji is two times better than LISA,
TianQin is seven times better than Taiji,
and the LISA-Taiji-TianQin network is around ten times better than Taiji.
For monochromatic sources, the detector in different orbital positions can be thought as an independent detector,
so the improvement on the angular resolutions by the network of combined detectors is small.
\section{Conclusion}
\label{discussion}
For space-based GW detectors, the technique of TDI is needed to cancel out laser frequency noises due to the difficulty of maintaining the exact equality of the arm length in space.
We extend the previous work on the analysis of the main factors that influence the accuracy of source localization by considering the TDI $(X, Y, Z)$ configuration.
In particular, we study the effects of the time-changing orientation of the detector plane, the arm length, the orbital period of spacecrafts, and the noise curve on the accuracy of source localization.
Due to the translational motion of the detector's center around the Sun,
the Doppler modulation on the amplitude and phase of GWs carries the position information of the sources
and it depends on the sky location of the source.
However, the amplitude modulation due to the rotation of the detector plane is independent of the sky location of the source.
The effects of Doppler and amplitude modulations get roughly equal at $f_0\sim 1$ mHz.
At low frequencies ($f_0\lesssim 1$ mHz),
the effect of the amplitude modulation dominates,
so it not only improves the sky localization but also enlarges the sky coverage for LISA and Taiji,
and helps LISA and Taiji to eliminate the equatorial pattern in the sky map.
As the frequency of monochromatic GWs increases ($f_0\gtrsim 0.01$ Hz),
the Doppler modulation becomes dominate
and the equatorial pattern appears on the sky map.
So equatorial patterns exist in the sky map at medium and high frequencies for LISA and Taiji.
For TianQin, it has no amplitude modulation to help eliminate the equatorial pattern in the sky map and its angular resolution has blind spots for sources from the direction with $\phi$ around $30^\circ$ or $-150^\circ$.
When the wavelength of GWs is larger than the detector's arm length, the frequency dependency on the transfer function deteriorates the SNR registered in the detector.
We find that the effect of arm length on the angular resolution mainly comes from the noise curve,
and it is almost the same for both heliocentric and geocentric constellations;
the orbital period of the spacecrafts has little effect on the angular resolutions.
Comparing the results with and without applying the TDI method, the angular resolutions are almost the same.
At the low frequency $f_0=1$ mHz,
the angular resolution with LISA is better than that with TianQin by more than two orders of magnitude with the help of the noise curve,
Taiji is two times better than LISA,
and the LISA-Taiji-TianQin network is $2.5$ times better than Taiji.
At the medium frequency $f_0=0.01$ Hz,
the angular resolution with LISA is less than two times better than that with TianQin,
Taiji is almost five times better than LISA,
and the LISA-Taiji-TianQin network is less than two times better than Taiji.
At the relatively high frequency $f_0=0.1$ Hz,
the angular resolution with Taiji is two times better than that with LISA,
TianQin is seven times better than Taiji,
and the LISA-Taiji-TianQin network is around ten times better than Taiji.
For monochromatic sources, the detector in different orbital positions can be thought as an independent detector,
so the improvement on the angular resolutions by the network of combined detectors is small.
The mean values of angular resolutions with the LISA-Taiji-TianQin network are about $6.1\times10^{-4}$ steradian at 1 mHz, $4.3\times10^{-6}$ steradian at 10 mHz and $1.7\times10^{-8}$ steradian at 100 mHz.
The angular resolutions are almost the same with and without the TDI combination.
\begin{acknowledgments}
This research is supported in part by the National Key Research and Development Program of China under Grant No. 2020YFC2201504.
\end{acknowledgments}
\begin{comment}
|
\section{Introduction}
The problem of unconstrained continuous convex optimization consists in finding an element of the set $\argmin_{q\in\mathbb{R}^d} f(q)$, for some lower bounded convex $f:\mathbb{R}^d\to\mathbb{R}$, generally assumed to be regular, e.g., twice continuously differentiable: $f\in C^2(\mathbb{R}^d,\mathbb{R})$.
\subsection{Acceleration in discrete- and continuous-time}
In 1979 Nemirovsky and Yudin \cite{nemirovskii1979problem} showed that, if $f$ is convex and $L$-smooth\footnote{A differentiable function $f:\mathbb{R}^d\to\mathbb{R}$ is said to be $L$-smooth if it has $L$-Lipschitz gradients.}, no gradient-based optimizer can converge to a solution faster than $O(1/k^2)$, where $k$ is the number of gradient evaluations\footnote{This lower bound holds just for $k<d$ hence it is only interesting in the high-dimensional setting.}. While Gradient Descent~(\texttt{GD}) converges like $O(1/k)$, the optimal rate $O(1/k^2)$ is achieved by the celebrated Accelerated Gradient Descent~(\texttt{AGD}) method, proposed by Nesterov in 1982~\cite{nesterov1983method}: starting from $p_0=0$ and a random $q_0$, the approximation $q_k$ to a problem solution $q^*$ is computed iteratively as\footnote{Many similar writings are possible. Here, we consider the particular version studied in~\cite{su2016} and a physicist notation, where $p_k$ is a velocity variable. This makes the connection to continuous-time cleaner and consistent with recent work on the geometry of momentum methods~\cite{francca2020dissipative}.}
\begin{equation*}
\tag{\texttt{AGD}}
\begin{cases}
q_{k+1} = q_k + \beta_k h p_k - h^2 \nabla f(q_k + \beta_k h p_k)\\
p_{k+1} = (q_{k+1}-q_k)/h
\end{cases}.
\end{equation*}
where $\beta_k = \frac{k-1}{k+2}$ and $h^2$ is the step-size~(we use the notation $h^2$ instead of the standard $\eta$ for a reason which will become apparent in the next sections). Interestingly, the different behaviour of \texttt{GD} and \texttt{AGD} is retained in the continuous-time limit~(as the step-size vanishes), recently studied by Su, Boyd and Candes \cite{su2016}, but already present in the seminal works of Polyak~\cite{polyak1964some} and Gavurin~\cite{gavurin1958nonlinear}:
\begin{equation*}
\tag{\texttt{GD-ODE}}
\dot q + \nabla f(q) = 0;
\end{equation*}
\begin{equation*}
\tag{\texttt{AGD-ODE}}
\ddot q + \frac{r}{t} \dot q + \nabla f(q) = 0.
\end{equation*}
Namely, we have that \texttt{GD-ODE} converges like $O(1/t)$ and \texttt{AGD-ODE}~(with $r\ge3$) like $O(1/t^2)$, where $t>0$ is the time variable. This result gave researchers a new tool to grasp the baffling essence~(see discussion in \cite{allen2014linear,su2016}) of accelerated optimizers, and led to the design of many novel fast interpretable algorithms~\cite{alimisis2019continuous,krichene2015,xu2018accelerated,wilson2019accelerating}.
\subsection{Evaluating gradients at a shifted position}
\label{sec:preliminary_experiments}
There are two modifications of \texttt{GD} that bring \texttt{AGD} about:
\begin{enumerate}[leftmargin=0.68cm]
\item inclusion of the momentum term (i.e. using $\beta_k\ne 0$);
\item change in gradient extrapolation point:
$\textcolor{blue}{\nabla f(q_k)}\to\textcolor{magenta}{\nabla f(q_k + \beta_k h p_k)}.$
\end{enumerate}
Questions arise immediately:
\begin{center}
\textit{Are both these modifications necessary for acceleration?\\
In particular, is evaluating the gradient at non-iterate points\\ crucial or even necessary for acceleration?}
\end{center}
To put these questions in the right historical context, one has to go back to Polyak's 1964 seminal paper~\cite{polyak1964some}, where the very first momentum method was proposed for $C^2$ and $\mu$-strongly-convex problems\footnote{In the strongly-convex case, $\beta_k$ is not monotonically increasing, but is instead chosen to be a constant dependent on the strong-convexity modulus $\mu$, that is $\beta = \left(\frac{\sqrt{L}-\sqrt{\mu}}{\sqrt{L}+\sqrt{\mu}}\right)^2$.}. Using an elegant functional-analytic argument on multistep methods, Polyak proved that momentum alone --- without shifted gradient evaluation (a.k.a. Heavy-ball (\texttt{HB}), see equation below) --- is able to achieve acceleration\footnote{Here to be intended as a dependency of the rate on the square root of the condition number $L/\mu$.} in a neighborhood of the solution. This local argument becomes of course global in the quadratic case (for a simplified proof, see Proposition 1 in~\cite{lessard2016analysis}).
\begin{equation*}
\tag{\texttt{HB}}
\begin{cases}
q_{k+1} = q_k + \beta_k h p_k - h^2 \nabla f(q_k)\\
p_{k+1} = (q_{k+1}-q_k)/h
\end{cases}.
\end{equation*}
Despite the many attempts, nobody in the last 56 years has been able to show that \texttt{HB} has a global (i.e. for any initialization) accelerated rate --- neither in the strongly-convex case~(using a fixed momentum) nor in the non-strongly-convex case~(using an increasing $\frac{k-1}{k+2}$ momentum). Beyond the technical difficulty, another plausible reason may also be lack of interest, as the introduction of Nesterov's globally accelerated method in 1982, that overshadowed the conceptually simpler method from Polyak.
However, many researchers in the last decade, supported by numerical evidence and by the success of Heavy-ball in deep learning~\cite{kingma2014adam}, expressed their belief that \texttt{HB} is accelerated:
\vspace{3mm}
\begin{chapquote}{Ghadimi et al. \cite{ghadimi2015global}, 2015}
[\dots] supported by the numerical
simulations we envisage that the convergence factor could be strengthened even further. This
is indeed left as a future work.\vspace{-1mm}
\end{chapquote}
\begin{chapquote}{Gorbunov et al. \cite{gorbunov2019stochastic}, 2019}
Despite the long history of this approach,
there is still an open question whether the heavy ball method converges to the optimum globally with
accelerated rate when the objective function is twice continuous differentiable.\vspace{-3mm}
\end{chapquote}
\begin{chapquote}{Muehlebach and Jordan \cite{muehlebach2020optimization}, 2020}
Neither the evaluation of the gradient at a
shifted position, nor a specifically engineered damping parameter, as for example proposed
in Nesterov (2004, Sec. 2.2), \textit{seem}\footnote{After talking to the first author, we decided to replace ``\textit{are}'' (as in the original preprint) with ``\textit{seem}'': indeed, the argument in~\cite{muehlebach2020optimization} is asymptotic and therefore somewhat equivalent to the one of Polyak~\cite{polyak1964some}.} necessary.
\vspace{-3mm}
\end{chapquote}
Other researcher believe \texttt{HB} is not accelerated:
\vspace{3mm}
\begin{chapquote}{Shi et al. \cite{shi2018understanding}, 2018}
If we can translate this argument to the discrete case we can understand why \texttt{AGD} achieves
acceleration globally for strongly-convex functions but the Heavy-ball method does not.
\end{chapquote}
While on the theoretical side the opinion is mixed, on the experimental side no numerical simulation\footnote{In \cite{lessard2016analysis}, the authors show that there exist a strongly-convex smooth function such that Heavy-ball does not converge. However, as also pointed out by Ghadimi et al.~\cite{ghadimi2015global}, such function is not $C^2$, and that a big step-size is used --- which violates the convergence conditions of Thm.~4 in~\cite{ghadimi2015global}. As such, this function does not constitute a proper counterexample.} has been able to show that \texttt{HB} is not accelerated. In Figure~\ref{fig:conjecture}, we provide two examples for the non-strongly-convex case (i.e. $\mu$ very small, such that an increasing momentum is preferable, leading $1/k^2$ convergence as opposed to $(1-\sqrt{\mu/L})^k$). In particular, we show that \texttt{HB} is comparable to \texttt{AGD} through the lens of the pathological lower-bounding quadratic example introduced by~\cite{nemirovskii1979problem} and used to construct the $O(1/k^2)$ bound in convex optimization --- at least until the effect of non-trivial strong-convexity becomes dominant (at around $f(q_k)=10^{-6}$).
\begin{figure}[ht]
\centering
\begin{subfigure}{0.45\textwidth}\centering\includegraphics[width=1\columnwidth]{img/lower_bound}
\caption{\texttt{HB} and \texttt{AGD} on the worst-case (lower bound) quadratic objective from Nesterov~\cite{nesterov2018lectures}.}\end{subfigure}
\hspace{1cm}
\begin{subfigure}{0.45\textwidth}\centering\includegraphics[width=1\columnwidth]{img/regression}\caption{\texttt{HB} and \texttt{AGD} on ill-conditioned linear regression. The optimal step-size $1/L$ was used.}\end{subfigure}
\caption{For both examples \texttt{HB} with momentum $\frac{k-1}{k+2}$ exhibits an accelerated $1/k^2$ convergence rate, even though\texttt{ AGD} with momentum $\frac{k-1}{k+2}$ is faster in a neighborhood of the optimizer due to strong-convexity. Instead, \texttt{GD} violates the Nesterov $O(1/k^2)$ upper bound. We recall that, while Nesterov's upper bound holds for all $k>0$, the $O(1/k^2)$ lower bound~(originally discovered by Nemirovski and Yudin~\cite{nemirovskii1979problem}) only holds at $k = d/2$ (for more details, check the discussion in~\cite{nesterov2018lectures}).}
\label{fig:conjecture}
\vspace{-3mm}
\end{figure}
\subsection{Contributions}
The purpose of the manuscript at hand is to study the effect of shifts in gradient extrapolation points on acceleration in convex optimization (i.e. to study the difference between Heavy-ball and Nesterov's method). In particular, the next pages are organized as follows:
\vspace{-1mm}
\begin{enumerate}[leftmargin=0.48cm]
\item We start from a continuous-time argument: inspired by a recent idea from Flammarion and Bach~\cite{flammarion2015averaging}, in Section~\ref{sec:quadratic_invariants_cont} we show how \texttt{AGD-ODE} with damping $2/t$ can be derived from the equation of a simple harmonic oscillator: $\ddot u = - Au$. By using Lyapunov equations and a simple change of variables, we retrieve the Lyapunov function proposed by Su, Boyd and Candes~\cite{su2016} to prove a rate $O(1/t^2)$ for \texttt{AGD-ODE}. This procedure is principled and leads to many insights on Lyapunov function design.
\item In Section~\ref{sec:quadratic_invariants_disc}, we apply the same methodology in discrete time, and show that \texttt{HB} with momentum $\frac{k-1}{k+1}$ can be derived from the St{\"o}rmer--Verlet discretization of the simple harmonic oscillator. Solving again Lyapunov's equations, we are able to show an $O(1/k^2)$ rate for a Heavy-ball argorithm for convex quadratics. While this rate is already present in~\cite{flammarion2015averaging}, our proof technique is different as it relies on a Lyapunov function as opposed to an eigenvalue analysis.
\item In Section \ref{sec:gener_quadratic}, by generalizing the discrete-time Lyapunov function found in Section~\ref{sec:quadratic_invariants_disc} we derive a modified Heavy-ball method
$$q_{k+1} = q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1}) - h^2\frac{k+\frac{r-2}{2}}{k+r-1} \nabla f(q_k),$$
with a rate of convergence $O(1/k^2)$ for any $k\ge2$ and $r\ge2$. Our result not only generalizes the theory in~\cite{flammarion2015averaging}, but also provides an interesting connection between the continuous and the discrete --- as the used Lyapunov function converges, in the limit $h\to 0$, to the one used in~\cite{su2016} for $r\ge2$.
\end{enumerate}
\paragraph{Recent related works.} Very recently, Wang et al.~\cite{wang2022provable} proved that Heavy-ball is accelerated for a class of functions satisfying the Polyak-\L{}ojasiewicz condition. Instead, here we provide a Lyapunov function for the non-strongly-convex setting, where the Polyak-\L{}ojasiewicz constant vanishes. We remark that, for strongly-convex quadratic potentials, Heavy-ball is already known to achieve acceleration~\cite{lessard2016analysis}. However, the eigenvalue argument used in~\cite{lessard2016analysis} cannot be leveraged in the non-strongly-convex setting, where the minimum eigenvalue can be arbitrarily low. As such, our work provides insights on how to construct effective Lyapunov functions in the non-quadratic case, where Lyanonov arguments are often the go-to option.
\section{From quadratic invariants of oscillators to accelerated rates}
\label{sec:quadratic_invariants}
Our procedure in this section is inspired by a beautiful idea presented by Flammarion and Bach~\cite{flammarion2015averaging}: it is sometimes possible to translate a time-dependent convergence rate problem into a time-independent stability problem. Here we go one step further, and show how, with an additional step (computation of quadratic invariants), it is possible to derive Lyapunov functions and rates for the corresponding algorithms. We first illustrate the idea in continuous-time and then proceed with the discrete-time analysis.
Our starting point is the following ODE:
\begin{equation*}
\tag{\texttt{AGD-ODE}2}
\ddot q + \frac{2}{t} \dot q + \nabla f(q) = 0.
\end{equation*}
From the analysis in~\cite{su2016}, we know that on a quadratic $f(q) = f^*+\frac{1}{2}\langle (q-q^*),A(q-q^*)\rangle $, with $A$ positive semidefinite and $f^*\in\mathbb{R}$, the solution converges to $q^*\in\argmin_{x\in\mathbb{R}^d} f(q)$ at the rate $O(1/t^2)$. To prove this rate, the authors in~\cite{su2016} use the following Lyapunov function:
\begin{equation}
V(q,t) = 2t^2 (f(q)-f^*) + \|t\dot q + (q-q^*)\|^2.
\label{eq:ly_su_boyd_quadratics}
\end{equation}
We show here a \textit{constructive way} to derive $V$~(Section~\ref{sec:quadratic_invariants_cont}) and then~(Section~\ref{sec:quadratic_invariants_disc}) we apply the same procedure to get a Lyapunov function for Heavy-ball (i.e., the discretization). For simplicity, we consider here $f^*=0$ and $q^* =0$.
\subsection{Lyapunov functions from continuous-time invariants}
\label{sec:quadratic_invariants_cont}
Consider an harmonic oscillator on the potential $f(u) = \frac{1}{2}u^\top A u$, i.e. $\ddot u = -A u$. From basic physics, we know that such a system is marginally stable (bounded dynamics). By choosing $u = tq$ we get $\dot u = q + t\dot q $ and $\ddot u = \dot q + \dot q + t\ddot q$. This implies
\begin{equation*}
\dot q + \dot q + t\ddot q = \ddot u = -A t q \ \quad
\Rightarrow \quad \ \ddot q +\frac{2}{t} \dot q + Aq =0.
\end{equation*}
That is, \texttt{AGD-ODE} can be reconstructed from a simple linearized pendulum. By introducing the variable $v = \dot u$, we can write the pendulum in phase space as a linear dynamical system
$$\begin{pmatrix}
\dot u\\\dot v
\end{pmatrix} = \begin{pmatrix}
0 & I \\ -A& 0
\end{pmatrix}\begin{pmatrix}
u\\ v
\end{pmatrix}.$$
Hence, the pendulum has the form $\dot y = F y$, where $y = (u,v)$. We would now like to get a Lyapunov function for this system. To do this, we recall a fundamental proposition~(check Thm. 4.6. in \cite{khalil2002nonlinear}).
\begin{proposition}[Continuous-time Lyapunov equations]
The linear system $\dot y = Fy$ is Lyapunov stable if and only if for all positive semidefinite matrices $Q$, there exists a symmetric matrix $P$ such that
\begin{equation}
PF+F^TP = -Q.
\label{eq:lyapunov_eq_cont}
\end{equation}
Moreover, $V(y) = y^T P y$ is a Lyapunov function and $\dot V (y) = -y^T Q y$.
\end{proposition}
Since we know that a pendulum is only marginally stable (i.e., not asymptotically stable), we can limit ourselves to the choice of a null matrix $Q$. Hence, we need to solve the Lyapunov equation $PF = -F^TP$ for $P$. A solution to this equation~(many exist) is
$P = \begin{pmatrix}
A & 0 \\ 0& I
\end{pmatrix}$, which implies that
\begin{equation}
V(u) = \langle u,A u\rangle + \|v\|^2
\label{eq:energy_continuous}
\end{equation} is a quadratic invariant, i.e. $\dot V(u)=0$. This is well known, since $V$ is actually twice the total energy (Hamiltonian) of the pendulum. Finally, we can change variables and get that
\begin{equation}
V(q) = t^2\langle q,A q\rangle + \left\|\frac{\d}{\d t}(tq)\right\|^2 = 2t^2 f(q) + \|t\dot q + (q-q^*)\|^2,
\label{eq:ly_quad_cont}
\end{equation}
is a Lyapunov function for \texttt{AGD-ODE}2, with $\dot V(q)=0$. This is precisely equation~\ref{eq:ly_su_boyd_quadratics}.
\paragraph{From quadratic to convex.} With a small modification (using a factor $r-1$ instead of $1$), it is possible to get a Lyapunov function that works for \texttt{AGD-ODE} in the more general convex case.
\begin{proposition}[Theorem 3 from~\cite{su2016}] For convex $L$-smooth objectives, \texttt{AGD-ODE} converges at a rate $O(1/t^2)$. This follows from the fact that
\begin{equation}
V(q,t) = 2t^2 (f(q)-f^*) +\|t\dot q + (r-1)(q-q^*)\|^2.
\label{eq:lyapunov_su_general}
\end{equation}
is a Lyapunov function, for $r\ge3$.
\end{proposition}
\subsection{Discrete-time invariants}
\label{sec:quadratic_invariants_disc}
We apply the construction from the last subsection to the discrete case. Inspired by Flammarion and Bach~\cite{flammarion2015averaging}, we consider at first a slightly modified \texttt{HB}:
\begin{equation*}
\tag{\texttt{HB}$2$}
q_{k+1} = q_k + \frac{k-1}{k+1} (q_k-q_{k-1}) - h^2 \frac{k}{k+1} \nabla f(q_k).
\end{equation*}
This algorithm is the discrete-time equivalent of $\ddot q + \frac{2}{t} \dot q +\nabla f(q) = 0$. As for the continuous-time case, we start from $f(q) = \frac{1}{2}\langle q,A q \rangle$. In this case, \texttt{HB}$2$ can be written as
$$(k+1) q_{k+1} = 2 k q_k + (k-1) q_{k-1} - h^2 A (kq_k).$$
That is, if we set $u_k = kq_k$, we get
\begin{equation}
u_{k+1} - 2u_k + u_{k-1} = -h^2 A u_k.
\label{eq:stormer-verlet}
\end{equation}
With surprise, we recognize that this is the St{\"o}rmer--Verlet method~\cite{hairer2003geometric} on $\ddot u = -A u$, with step-size $h$ (that's why had $h^2$ from the very beginning).
It would be natural, as for the continuous-time case, to consider the total energy as a quadratic invariant to derive a Lyapunov function. However, it turns out that, interestingly, \textit{the St{\"o}rmer--Verlet method does not precisely conserve the total energy}: there are small oscillations~(see Section 3 in~\cite{hairer2014challenges})! Taking into account such small oscillations (Figure~\ref{fig:verlet}) is of fundamental importance --- since they lead to a crucial modification of the invariants we have to use.
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\textwidth]{img/verlet_energy.pdf}
\caption{The St{\"o}rmer--Verlet method on a one-dimensional quadratic potential (i.e., a simplified pendulum) does not conserve the total energy. Details on this phenomenon can be found in~\cite{hairer2016long,hairer2014challenges}.}
\label{fig:verlet}
\end{figure}
\begin{proposition}[Discrete-time Lyapunov equations]
The system $ y_{k+1} = Fy_k$ is Lyapunov stable if an only if for all positive semidefinite matrices $Q$, there exists a symmetric matrix $P$ such that
\begin{equation}
F^T P F - P = -Q.
\label{eq:lyapunov_eq_discrete}
\end{equation}
Moreover, $V(y) = y^T P y$ is a Lyapunov function and $V (y_{k+1})-V(y_k) = -y_k^T Q y_k$ for all $k$.
\end{proposition}
We apply the theorem above~(for $Q=0$) to the linear system
$$\begin{pmatrix}
u_{k+1}\\ v_{k+1}
\end{pmatrix} = \begin{pmatrix}
I-h^2 A & h I \\ -hA& I
\end{pmatrix}\begin{pmatrix}
u_k\\ v_k
\end{pmatrix}.$$
Under the choice $v_k = (u_{k}-u_{k-1})/h$, this system is equivalent to equation~\ref{eq:stormer-verlet}, i.e., the discretized pendulum we want to find a quadratic invariant for. Solving the discrete Lyapunov equation gives us
$$P = \begin{pmatrix}
A & -hA/2 \\ -hA/2& I
\end{pmatrix},$$
and the associated modified total energy:
\begin{equation}
V(u_k,v_k) = \underbrace{\langle u_k, A u_k\rangle +\|v_k\|^2}_{\text{continuous-time invariant (energy)}}- \underbrace{h\langle v_k, A u_k\rangle}_{\text{vanishing cross-term}}.
\label{eq:invariant_stormer}
\end{equation}
We make the following comments:
\begin{itemize}[leftmargin=0.48cm]
\item As $h\to 0$, the modified energy approaches the total energy in equation~\ref{eq:energy_continuous}. The purpose of the additional cross-term is to eliminate the small energy oscillations we see in Figure~\ref{fig:verlet}.
\item Assuming without loss of generality that $A$ does not have zero eigenvalues, $P$ is positive semidefinite (i.e., yields a valid Lyapunov function) if and only if the Schur complement of $I$ (i.e., the block $P_{22}$) in $P$ is positive semidefinite. That is, we need
\begin{equation}
B:=A-\frac{h^2}{4}A^2 = A\left(I-\frac{h^2}{4}A\right)\ge0.
\label{eq:B_matrix}
\end{equation}
Since $A$ and $(I-h^2A/4)$ are co-diagonalizable, the product is positive semidefinite if and only if both $A\ge 0$ and $(I-h^2A/4)\ge0$. This requires $0\le A\le \frac{4}{h^2}I$, which in turns implies an upper bound on the step-size $h^2$:
$$h^2\le\frac{4}{\lambda_{\text{max}}(A)} = \frac{4}{L}.$$
The same condition~(note that our step-size is $h^2$, not $h$) can be deduced from the analysis of \texttt{HB}$2$ in~\cite{flammarion2015averaging}.
\end{itemize}
Now it's time to change variables back: $u_k = kq_k$. If we set $p_k := (q_k-q_{k-1})/h$, as also done in the introduction, we get
\begin{align*}
hv_{k} &= h(u_{k}-u_{k-1})/h\\ &= k q_k - (k-1) q_{k-1}+ (k-1) q_{k}- (k-1) q_{k}\\
&= (k-1)(q_k-q_{k-1})+q_k\\
&= h(k-1)p_k+q_k.
\end{align*}
By substituting these formulas in equation~\ref{eq:invariant_stormer}, we get the following final form for an effective Lyapunov function for \texttt{HB}$2$ --- for the quadratic case:
\begin{align*}
&V_k=\langle u_k, A u_k\rangle + \|v_k\|^2-\langle h v_k, A u_k\rangle\\ \Longrightarrow \ & V_k = k^2\langle q_k,Aq_k\rangle + \frac{1}{h^2}\|h(k-1)p_k+q_k\|^2-k\langle h(k-1)p_k+q_k, A q_k\rangle.
\end{align*}
To better understand this Lyapunov function, we multiply everything by $h^2$ and get
$$V_k = (kh)^2\langle q_k,Aq_k\rangle + \|h(k-1)p_k+q_k\|^2 - h^2k\langle h(k-1)p_k+q_k, A q_k\rangle.$$
Recalling that the ``time'' variable $t$ is defined to be $t_k = hk$, this cost becomes
$$V_k = t_k^2\langle q_k,Aq_k\rangle + \|t_{k-1}p_k+q_k\|^2 - h t_k\langle t_{k-1}p_k+q_k, A q_k\rangle.$$
This Lyapunov function can be easily generalized by noting that $\langle q_k,Aq_k\rangle = 2(f(q_k)-f^*)$ and $A q_k = \nabla f(q_k)$:
\begin{equation}
V_k = 2t_k^2 (f(q_k)-f^*) + \|t_{k-1}p_k+q_k\|^2 - t_k\langle\nabla f(q_k), t_{k-1}p_k+q_k\rangle.
\label{eq:lyapunov_HB_first}
\end{equation}
Finally, note that
\begin{itemize}[leftmargin=0.48cm]
\item From equation~\ref{eq:lyapunov_HB_first} as $h\to0$ we get equation~\ref{eq:ly_quad_cont}: the continuous-time Lyapunov function of~\cite{su2016}.
\item The mixing term is necessary and makes the positive definiteness~(see equation~\ref{eq:B_matrix}) of $V_k$ non-trivial.
\end{itemize}
All in all, in this subsection, we proved the following result.
\begin{tcolorbox}
\begin{proposition}
Let $A\in\mathbb{R}^{d\times d}$ be positive semidefinite, $f^*\in\mathbb{R}$ and $f(q)= f^* + \frac{1}{2}(q-q^*)^T A (q-q^*)$. Let $(q_k)_{k\ge0}$ be the iterates of \texttt{HB}$2$ and $p_k:=(q_k-q_{k-1})/h$. If the step-size $h^2<\frac{4}{\lambda_\text{max}(A)}$, then equation~\ref{eq:lyapunov_HB_first} is non-negative and such that $V_{k+1}-V_k =0$ along the \texttt{HB}$2$ trajectory, for all $k$. From this, one can deduce an accelerated rate of $O(1/k^2)$ in suboptimality.
\label{prop:proof_quadratic_bach}
\end{proposition}
\end{tcolorbox}
Details of the proof are given in the proof of Theorem~\ref{thm:proof_quadratic}~(more general).
\section{Accelerated Heavy-ball methods for convex quadratics}
\label{sec:gener_quadratic}
In this section, we start to lift the discussion to the convex non-quadratic setting, by providing a generalization of \texttt{HB}2. Indeed, we know from the continuous-time analysis in~\cite{su2016} that $\ddot q + \frac{2}{t}\dot q + \nabla f(q)=0$ may not have an accelerated rate for functions which are convex but not necessarily quadratic. In this case, a rate of $O(1/t^2)$ only holds~\footnote{The case $0<r\le3$ was studied by Attouch et al.~\cite{attouch2019rate}: a convergence rate of $O(1/t^p)$ with $p<2r/3$ is shown in this case. The same result also holds in discrete time.} for $$\ddot q + \frac{r}{t}\dot q + \nabla f(q)=0,$$ with $r\ge 3$. In the same way, we expect that \texttt{HB}$2$~(which is the discretization for $r=2$) may not have an accelerated rate in the convex non-quadratic setting and a \textit{generalization corresponding to high friction is therefore necessary}.
Our objective in this chapter is to construct such a generalization of \texttt{HB}$2$, which we name \texttt{HB}$r$.
\subsection{A generalized Heavy-ball with high friction and guarantees on quadratics}
After a few weeks of intense calculations, we found that this algorithm gives the desired result~(Thm.~\ref{thm:proof_quadratic}).
\begin{equation*}
\tag{\texttt{HB}$r$}
q_{k+1} = \underbrace{q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})}_{\text{iterate + momentum}} - h^2\underbrace{ \frac{k+\frac{r-2}{2}}{k+r-1} \nabla f(q_k)}_{\text{scaled gradient of iterate}}.
\end{equation*}
First, note that $r=2$ recovers \texttt{HB}$2$ --- which we proved to be accelerated in the last subsection using a novel Lyapunov argument. The second, and perhaps the most crucial, thing to note is that \texttt{HB}$r$ recalls the high friction generalization of \texttt{AGD} proposed by~\cite{su2016}~(see Theorem 6 in their paper):
\begin{equation*}
\tag{\texttt{AGD}$r$}
q_{k+1} = \underbrace{q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})}_{\text{iterate + momentum}} - h^2 \underbrace{\nabla f\left(q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})\right)}_{\text{gradient of [iterate + momentum]}}.
\end{equation*}
Between \texttt{HB}$r$ and \texttt{AGD}$r$ there are a few important differences:
\begin{itemize}[leftmargin=0.48cm]
\item In \texttt{AGD}$r$ the gradient is evaluated at $q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})$, while in \texttt{HB}$r$ it is evaluated at $q_k$.
\item in \texttt{HB}$r$ the effective step-size~(i.e. what multiplies the gradient) is iteration-dependent, and goes from $h^2/2$ to $h^2$ as $k\to\infty$. We believe this has \textit{not to be regarded as part of the acceleration mechanism}: it is just a small modification needed to make the analysis easier.
\item Arguably \texttt{HB}$r$~(neglecting the small correction) is conceptually simpler that \texttt{AGD}$r$: compared to \texttt{GD}, only a momentum term is added at each iteration --- and this can be thought of as the source of acceleration.
\end{itemize}
We proceed in proving that \texttt{HB}$r$ is accelerated in the quadratic case.
\begin{tcolorbox}
\begin{theorem}Let $A\in\mathbb{R}^{d\times d}$ be positive semidefinite, $f^*\in\mathbb{R}$ and $f(q)= f^* + \frac{1}{2}(q-q^*)^T A (q-q^*)$. Let $(q_k)_{k\ge0}$ be the iterates of \texttt{HB}$r$~($r\ge 2$) and $p_k:=(q_k-q_{k-1})/h$. If $h^2\le\frac{4}{\lambda_\text{max}(A)}$, then
\begin{multline}
V_k = 2(k+r-2)^2 h^2 (f(q_k)-f^*) + \|h(k-1)p_k+(r-1)(q_k-q^*)\|^2\\ - h^2 (k+r-2)\langle\nabla f(q_k), h(k-1)p_k+(r-1)(q_k-q^*)\rangle
\label{eq:lyapunov_HB}
\end{multline}
is non-negative and such that $V_{k+1}-V_k \le0$ along the \texttt{HB}$r$ trajectory, for all $k$. Moreover, for any $h^2<\frac{4}{\lambda_\text{max}(A)}$, \texttt{HB}$r$ is accelerated. In particular, if $h^2=\frac{2}{\lambda_\text{max}(A)}$, we have the rate
$$f(q_k)-f^*\le \frac{\lambda_\text{max}(A)V_0}{2(k+r-2)^2}.$$
\label{thm:proof_quadratic}
\end{theorem}
\end{tcolorbox}
We note a couple of facts about the Lyapunov function $V_k$ in equation~\ref{eq:lyapunov_HB}.
\begin{itemize}[leftmargin=0.48cm]
\item It reduces to equation~\ref{eq:lyapunov_HB_first} in the case $r=2$. For $r>2$, it is a graspable generalization of equation~\ref{eq:lyapunov_HB_first} --- which we instead derived in a systematic way using Lyapunov equations. The term $(r-1)$ is inspired by the continuous-time limit in equation~\ref{eq:lyapunov_su_general}.
\item Consider the Lyapunov function above, but without cross term i.e.
$$2(k+r-2)^2 h^2 (f(q_k)-f^*) + \|h(k-1)p_k+(r-1)(q_k-q^*)\|^2.$$
This function works for proving an $O(1/k^2)$ rate for \texttt{AGDr} (it's a Lyapunov function, see Thm. 6 from~\cite{su2016}). Therefore, higher complexity~(i.e., an additional cross term) is needed to study the acceleration of Heavy-ball, when compared to Nesterov's method.
\item As $h\to0$, the cross term vanishes $V_k$ converges to equation~\ref{eq:lyapunov_su_general} --- its continuous-time equivalent. Indeed, both \texttt{HBr} and \texttt{AGDr} converge to \texttt{AGD-ODE} as $h\to 0$.
\end{itemize}
\subsection{Proof of the theorem}
It is useful to simplify equation~\ref{eq:lyapunov_HB} and to work with variables $q_k$ and $q_{k-1}$ --- a natural choice in the discrete setting. We split the Lyapunov function into two parts: $V_k = V^1_k + V^2_k$.
\begin{align}
&V^1_k := 2(k+r-2)^2 h^2 (f(q_k)-f^*)\label{eq:V1}\\ &\quad \quad\quad - h^2 (k+r-2)\langle\nabla f(q_k),(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\rangle.\nonumber\\
& V^2_k := \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.\label{eq:V2}
\end{align}
First, we are going to study $V_k^2$ in the non-quadratic case, and then $V_k^1$ in the quadratic case. Theorem~\ref{thm:proof_quadratic} will follow from a combination of the two corresponding lemmata.
The first lemma shares many similarities with the proof of Theorem 1 in~\cite{ghadimi2015global}.
\begin{lemma}
For any differentiable function $f:\mathbb{R}^d\to\mathbb{R}$ (not necessarily convex or $L$-smooth) and any sequence of iterates $(q_k)_{k\ge0}$ returned by \texttt{HB}$r$, we have:
\begin{align*}
V_{k+1}^2-V_k^2\ = \ &-h^2(r-1)(2k+r-2)\langle\nabla f(q_k),q_k-q^*\rangle\\
&-h^2(k-1)(2k+r-2)\langle\nabla f(q_k),q_{k}-q_{k-1}\rangle\\
&+\frac{h^4}{4}(2k+r-2)^2\|\nabla f(q_k)\|^2,
\end{align*}
where $V^2_k$ is defined in equation~\ref{eq:V2}.
\label{lemma:V2_quad}
\end{lemma}
\begin{proof}
Let $g_{k} := (k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)$, then,
$$V^2_{k+1}-V^2_k = \|g_{k+1}\|^2-\|g_{k}\|^2 = \langle g_{k+1}+g_{k},g_{k+1}-g_k\rangle.$$
We proceed in computing $g_{k+1}-g_k$. The algorithm symmetric structure here is fundamental:
\begin{align*}
g_{k+1}-g_k = \ & k(q_{k+1}-q_{k})+(r-1)(q_{k+1}-q^*)-(k-1)(q_k-q_{k-1})-(r-1)(q_{k}-q^*)\\
= \ & (k+r-1)q_{k+1}-(k+r-1)q_k-(k-1)(q_k-q_{k-1})\\
\overset{\text{(\texttt{HB}$r$)}}{=} & - h^2\left(k+\frac{r-2}{2}\right) \nabla f(q_k).
\end{align*}
Instead, $g_{k+1}+g_k$ is slightly more complex.
\begin{align*}
g_{k+1}+g_k = \ & k(q_{k+1}-q_{k})+(r-1)(q_{k+1}-q^*)+(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\\
= \ & (k+r-1)q_{k+1}+(-k+r-1)q_k+(k-1)(q_k-q_{k-1})-2(r-1)q^*\\
\overset{\text{(\texttt{HB}$r$)}}{=} &(k+r-1)q_{k} + (k-1)(q_k-q_{k-1})- h^2\left(k+\frac{r-2}{2}\right) \nabla f(q_k)\\&+(-k+r-1)q_k+(k-1)(q_k-q_{k-1})-2(r-1)q^*\\
= \ &2(r-1)(q_{k}-q^*)+2(k-1)(q_k-q_{k-1})- h^2\left(k+\frac{r-2}{2}\right) \nabla f(q_k).
\end{align*}
The proof is concluded by taking the inner product.
\end{proof}
We proceed by computing the difference $V^1_{k+1}-V^1_k$. Our calculations will be very quick, since we can leverage, in the quadratic case, on a simplified expression for $V^1_k$.
\begin{lemma}
Let $V^1_k$ be defined as equation~\ref{eq:V1}. In the context of Theorem~\ref{thm:proof_quadratic}, we have
$$V^1_k = h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle$$
and
\begin{align*}
V_{k+1}^1-V_k^1\ = & \ h^2(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle\\
&+h^2(k-1)(2k+r-2)\langle q_k-q^*, A(q_k-q_{k-1})\rangle\\
&-\frac{h^4}{2} (2k+r-2)k\|A(q_k-q^*)\|^2.
\end{align*}
\vspace{-3mm}
\label{lemma:V1_quad}
\end{lemma}
\begin{proof}
From equation~\ref{eq:V1}, we get
\begin{align*}
V^1_k = \ & \ (k+r-2)^2 h^2 \langle q_k-q^*, A(q_k-q^*)\rangle\\ &- h^2 (k+r-2)\langle A(q_k-q^*),(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\rangle\\
= \ & \ (k+r-2)^2 h^2 \langle q_k-q^*, A(q_k-q^*)\rangle\\
&\ - h^2 (k+r-2)\langle A(q_k-q^*),(k+r-2)(q_k-q^*)-(k-1)(q_{k-1}-q^*)\rangle\\
= \ & \ h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle.
\end{align*}
We proceed computing $V^1_{k+1}-V^1_{k}$ using this simplified form:
\begin{align*}
V^1_{k+1} - V^1_k = \ & \ h^2(k+r-1)k\langle q_{k}-q^*,A(q_{k+1}-q^*)\rangle\\&-h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle\\
= \ & h^2\langle q_{k}-q^*,A \Delta_k\rangle,
\end{align*}
where
\begin{align*}
\Delta_k := \ & \ (k+r-1)k(q_{k+1}-q^*)-(k+r-2)(k-1)(q_{k-1}-q^*).
\end{align*}
Now, recall the definition of \texttt{HB}$r$:
$$(q_{k+1}-q^*) = (q_k-q^*) + \frac{k-1}{k+r-1} (q_k-q_{k-1}) - h^2 \frac{k+\frac{r-2}{2}}{k+r-1}A(q_k-q*),$$
where we subtracted $q^*$ from both sides. By plugging this into $\Delta_k$, we get
\begin{align*}
\Delta_k = \ & \ (k+r-1)k(q_k-q^*) + (k-1)k (q_k-q_{k-1}) - \frac{h^2}{2} (2k+r-2)k A(q_k-q*)\\
&-(k+r-2)(k-1)(q_{k-1}-q^*)\\
=\ & \ (2k+r-2)k(q_k-q^*)-(2k+r-2)(k-1)(q_{k-1}-q^*)\\&- \frac{h^2}{2} (2k+r-2)k A(q_k-q^*)\\
=\ & \ (2k+r-2)(q_k-q^*)+(2k+r-2)(k-1)(q_k-q_{k-1})- \frac{h^2}{2} (2k+r-2)k A(q_k-q^*).
\end{align*}
The result follows after taking the inner product $h^2\langle q_{k}-q^*,A \Delta_k\rangle$.
\end{proof}
We are finally ready to prove the result.
\begin{proof}[Proof of Theorem~\ref{thm:proof_quadratic}]
First, we compute $V_{k+1}^1-V_k = (V_{k+1}^1-V_k^1)+(V_{k+1}^2-V_k^2)$ using Lemma~\ref{lemma:V1_quad} and Lemma~\ref{lemma:V2_quad}~(written for quadratic $f$). Next, we show that a certain condition on the step-size implies positivity of $V_k$ and a convergence rate.
\begin{align*}
(V_{k+1}^1-V_k^1)+(V_{k+1}^2-V_k^2)\ = & \ h^2(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle\\
&\hcancel{+h^2(k-1)(2k+r-2)\langle q_k-q^*, A(q_k-q_{k-1})\rangle}\\
&-\frac{h^4}{2} (2k+r-2)k\|A(q_k-q^*)\|^2\\ &-h^2(r-1)(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle\\
&\hcancel{-h^2(k-1)(2k+r-2)\langle q_k-q^*,A(q_{k}-q_{k-1})\rangle}\\
&+\frac{h^4}{4}(2k+r-2)^2\|A(q_k-q^*)\|^2.
\end{align*}
Crucially, note that the terms including $\langle q_k-q^*,A(q_{k}-q_{k-1})\rangle$ cancel. This is necessary to make our proof~(or, probably, any proof) work, since such inner product between the gradient and the momentum changes sign (infinitely) many times along the trajectory, and therefore cannot be easily compared to other quantities. For the same reason, in the corresponding continuous-time proof from~\cite{su2016}, the terms including $\langle \nabla f(q), p\rangle$ also perfectly cancel out.
All in all, by collecting some terms, we get
\begin{align*}
V_{k+1}-V_{k} &= -h^2(r-2)(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle \\& \quad \quad+ \frac{h^4}{4}(2k+r-2)(r-2)\|A(q_k-q^*)\|^2\\
&=-h^2(r-2)(2k+r-2)\langle q_k-q^*,B(q_k-q^*)\rangle,
\end{align*}
where $B:=A-\frac{h^2}{4}A^2$ is the matrix that we already studied in the context of Lyapunov equations~(see equation~\ref{eq:B_matrix}). Since $r\ge2$, a sufficient condition for $V_{k+1}-V_{k}\le 0$ is $B\ge0$, which holds under $h^2\le\frac{4}{\lambda_\text{max}(A)}$. As a sanity check, the reader can appreciate the fact that, if $r=2$, then $V_{k+1}=V_k$ --- as we already proved in Proposition~\ref{prop:proof_quadratic_bach}~(follows from the fact that $V_k$ solves the Lyapunov equations).
Last, we have to translate the fact that $V_k$ is non-increasing to a convergence rate. This is not trivial in our case, since $V_k$ also contains a cross term which is not necessarily positive. Actually, we do not even know that $V_k\ge0$ yet! Hence, we have to come up with some tricks. We start from rewriting the~(simplified) Lyapunov function:
$$V_k =
h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle + \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.$$
Now, let us add and subtract a term $c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle$, with $c>0$. We have:
$$V_k = c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle + \tilde V_k,$$
with
\begin{align*}
\tilde V_k :=& -c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle \\&\quad + h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle\\&\quad + \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.
\end{align*}
Now, if we show that $\tilde V_k$ is always positive, then $V_{k+1}\le V_k$ for all $k$ implies:
$$c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle \le c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle + \tilde V_k= V_k\le V_0,$$
which gives the desired rate:
$$ f(q_k)-f^*=\frac{1}{2}\langle q_{k}-q^*,A(q_k-q^*)\rangle \le \frac{V_0}{2c h^2(k+r-2)^2}.$$
Therefore, we only need to show $\tilde V_k\ge0$. To do this, we introduce two new variables:
\begin{equation*}
u_{k} := (k+r-2)(q_k-q^*),\quad\quad w_{k} := (k-1)(q_{k-1}-q^*),
\end{equation*}
and get a simplified form for $\tilde V_k$
\begin{align*}
\tilde V_k &= -c h^2 \langle u_k,Au_k\rangle + h^2\langle v_k,A w_k\rangle + \|u_k-w_k\|^2\\
&=\langle u_k, (I-ch^2A)u_k\rangle + \|w_k\|^2 - 2\langle u_k,\left(I-\frac{h^2}{2}A\right)w_k\rangle.
\end{align*}
Hence, we just need to show that
$$\tilde P =\begin{pmatrix}
I-ch^2A & -\left(I-\frac{h^2}{2}A\right)\\
-\left(I-\frac{h^2}{2}A\right) & I
\end{pmatrix}$$
is positive definite, for some $c$ and $h^2$. Using the Schur characterization for positive semidefinite matrices, $\tilde P\ge0$ if and only if
\begin{align*}
0\le\tilde B(c) &:=I-ch^2A-\left(I-\frac{h^2}{2}A\right)^2\\
&=I-ch^2A-I-\frac{h^4}{4}A^2+h^2A\\
&=h^2 A\left(1-c-\frac{h^2}{4}A\right).
\end{align*}
It is clear that $\tilde B(c)$ is positive semidefinite if and only if
$1-c-\frac{h^2}{4}\lambda_{\text{max}}(A)\ge 0$. That is,
$$h^2\le\frac{4(1-c)}{\lambda_\text{max}(A)}.$$
Hence, for any $c\in (0,1)$ we get an acceleration. In particular, in the theorem, we chose $c=\frac{1}{2}$.
\end{proof}
\subsection{Numerical verification of our Lyapunov function} We verify numerically that the Lyapunov function for \texttt{HB}r proposed in equation~\ref{eq:lyapunov_HB} works on quadratics. To more clearly show the effect the inner product correction term, which originated from the quadratic invariant of the St{\"o}rmer--Verlet method, we use here a slightly different notation: $V_k = V^{11}_k+V^{12}_k+V^{2}_k$, with $V^{1}_k = V^{11}_k+V^{12}_k$.
\begin{align*}
V^{11}_k &:= 2(k+r-2)^2 h^2 (f(q_k)-f^*);\\
V^{12}_k&:= - h^2 (k+r-2)\langle\nabla f(q_k),(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\rangle;\\
V^{2}_k &:= \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.
\end{align*}
We recall that, the term $V^{12}_k$ (a.k.a. the \textit{cross-term}) vanishes as $h\to0$, and is indeed not present in the continuous-time limit. We show that this term, which we derived using Lyapunov equations in Sec.~\ref{sec:quadratic_invariants}, plays a fundamental role in ensuring $V_{k+1}-V_k\le0$. In Figure~\ref{fig:lyapunov_HBr_linreg_1} we verify numerically Thm.~\ref{thm:proof_quadratic}. In Figure~\ref{fig:lyapunov_HBr_linreg_3} we show the essential role of $V^{12}$. Here we used $h^2=1/L$, but \texttt{HB}$r$ can take larger steps~(up to $4/L$), while the other algorithms become unstable~(Figure~\ref{fig:HBr_big_step}).
\clearpage
\begin{figure}[ht!]
\centering
\begin{subfigure}{\textwidth}\centering\includegraphics[width=0.8\columnwidth]{img/regression_ly_r2.png}
\end{subfigure}
\vspace{-4mm}
\begin{subfigure}{\textwidth}\centering\includegraphics[width=0.8\columnwidth]{img/regression_ly_r3.png}
\end{subfigure}
\vspace{-2mm}
\caption{Dynamics of the Lyapunov function for \texttt{HB}$r$ on linear regression~(ill-conditioned Hessian, with condition number $\kappa$). Shown is the behavior for $r=2,3$ with step-size $1/L$. For $r=2$, $V_{k}$ is constant, as predicted by Prop.~\ref{prop:proof_quadratic_bach}. For $r=3$, $V_{k}$ is decreasing as predicted by Thm.~\ref{thm:proof_quadratic}. }
\label{fig:lyapunov_HBr_linreg_1}
\end{figure}
\vspace{-2mm}
\begin{figure}[ht!]\centering\includegraphics[width=0.8\columnwidth]{img/regression_ly_no.png}
\vspace{-4mm}
\caption{Same setting of the second example in Figure~\ref{fig:lyapunov_HBr_linreg_1}, but different candidate Lyapunov function (no cross term). This confirms the cross-term is necessary.}
\label{fig:lyapunov_HBr_linreg_3}
\end{figure}
\vspace{-2mm}
\begin{figure}[ht!]\centering\includegraphics[width=0.35\columnwidth]{img/regression_big_step.pdf}
\vspace{-2mm}
\caption{\texttt{HB}r also works for big step-sizes~(see conditions in Theorem~\ref{thm:proof_quadratic}). Here, used is $h^2=3.9/L$.}
\label{fig:HBr_big_step}
\end{figure}
\section{Conclusion}
\vspace{-1mm}
In conclusion, the question of whether the Heavy-ball method is globally accelerated for non-strongly-convex quadratic problems has yet to be fully answered, and has attracted the attention of recent research~\cite{wang2022provable}. Our study takes a novel approach by examining momentum through the lens of quadratic invariants of simple harmonic oscillators, and by utilizing the modified Hamiltonian of Stormer-Verlet integrators we were able to construct a Lyapunov function that demonstrates an $O(1/k^2)$ rate for Heavy-ball in the case of convex quadratic problems, where eigenvalues can vanish. This is a promising first step towards potentially proving the acceleration of Polyak's momentum method through Lyapunov function arguments.
\vspace{-1mm}
\section{Acknowledgements}
\vspace{-1mm}
I would like to extend my deepest gratitude to Prof. Boris Polyak, Prof. Christian Lubich, and Konstantin Mishchenko for the stimulating discussions. My appreciation goes to Prof. Aurelien Lucchi and Prof. Thomas Hofmann for their unwavering support and motivation, which helped me to develop the project idea in Spring 2020. Lastly, I cannot express enough my gratitude to Johannes Brahms for his Violinkonzert D-Dur op. 77, which provided the perfect soundtrack to my late-night calculations, igniting my passion and drive to push through the toughest moments.
\printbibliography
\bibstyle{apalike}
\end{document}
\section{Introduction}
The problem of unconstrained continuous convex optimization consists in finding an element of the set $\argmin_{q\in\mathbb{R}^d} f(q)$, for some lower bounded convex $f:\mathbb{R}^d\to\mathbb{R}$, generally assumed to be regular, e.g., twice continuously differentiable: $f\in C^2(\mathbb{R}^d,\mathbb{R})$.
\subsection{Acceleration in discrete- and continuous-time}
In 1979 Nemirovsky and Yudin \cite{nemirovskii1979problem} showed that, if $f$ is convex and $L$-smooth\footnote{A differentiable function $f:\mathbb{R}^d\to\mathbb{R}$ is said to be $L$-smooth if it has $L$-Lipschitz gradients.}, no gradient-based optimizer can converge to a solution faster than $O(1/k^2)$, where $k$ is the number of gradient evaluations\footnote{This lower bound holds just for $k<d$ hence it is only interesting in the high-dimensional setting.}. While Gradient Descent~(\texttt{GD}) converges like $O(1/k)$, the optimal rate $O(1/k^2)$ is achieved by the celebrated Accelerated Gradient Descent~(\texttt{AGD}) method, proposed by Nesterov in 1982~\cite{nesterov1983method}: starting from $p_0=0$ and a random $q_0$, the approximation $q_k$ to a problem solution $q^*$ is computed iteratively as\footnote{Many similar writings are possible. Here, we consider the particular version studied in~\cite{su2016} and a physicist notation, where $p_k$ is a velocity variable. This makes the connection to continuous-time cleaner and consistent with recent work on the geometry of momentum methods~\cite{francca2020dissipative}.}
\begin{equation*}
\tag{\texttt{AGD}}
\begin{cases}
q_{k+1} = q_k + \beta_k h p_k - h^2 \nabla f(q_k + \beta_k h p_k)\\
p_{k+1} = (q_{k+1}-q_k)/h
\end{cases}.
\end{equation*}
where $\beta_k = \frac{k-1}{k+2}$ and $h^2$ is the step-size~(we use the notation $h^2$ instead of the standard $\eta$ for a reason which will become apparent in the next sections). Interestingly, the different behaviour of \texttt{GD} and \texttt{AGD} is retained in the continuous-time limit~(as the step-size vanishes), recently studied by Su, Boyd and Candes \cite{su2016}, but already present in the seminal works of Polyak~\cite{polyak1964some} and Gavurin~\cite{gavurin1958nonlinear}:
\begin{equation*}
\tag{\texttt{GD-ODE}}
\dot q + \nabla f(q) = 0;
\end{equation*}
\begin{equation*}
\tag{\texttt{AGD-ODE}}
\ddot q + \frac{r}{t} \dot q + \nabla f(q) = 0.
\end{equation*}
Namely, we have that \texttt{GD-ODE} converges like $O(1/t)$ and \texttt{AGD-ODE}~(with $r\ge3$) like $O(1/t^2)$, where $t>0$ is the time variable. This result gave researchers a new tool to grasp the baffling essence~(see discussion in \cite{allen2014linear,su2016}) of accelerated optimizers, and led to the design of many novel fast interpretable algorithms~\cite{alimisis2019continuous,krichene2015,xu2018accelerated,wilson2019accelerating}.
\subsection{Evaluating gradients at a shifted position}
\label{sec:preliminary_experiments}
There are two modifications of \texttt{GD} that bring \texttt{AGD} about:
\begin{enumerate}[leftmargin=0.68cm]
\item inclusion of the momentum term (i.e. using $\beta_k\ne 0$);
\item change in gradient extrapolation point:
$\textcolor{blue}{\nabla f(q_k)}\to\textcolor{magenta}{\nabla f(q_k + \beta_k h p_k)}.$
\end{enumerate}
Questions arise immediately:
\begin{center}
\textit{Are both these modifications necessary for acceleration?\\
In particular, is evaluating the gradient at non-iterate points\\ crucial or even necessary for acceleration?}
\end{center}
To put these questions in the right historical context, one has to go back to Polyak's 1964 seminal paper~\cite{polyak1964some}, where the very first momentum method was proposed for $C^2$ and $\mu$-strongly-convex problems\footnote{In the strongly-convex case, $\beta_k$ is not monotonically increasing, but is instead chosen to be a constant dependent on the strong-convexity modulus $\mu$, that is $\beta = \left(\frac{\sqrt{L}-\sqrt{\mu}}{\sqrt{L}+\sqrt{\mu}}\right)^2$.}. Using an elegant functional-analytic argument on multistep methods, Polyak proved that momentum alone --- without shifted gradient evaluation (a.k.a. Heavy-ball (\texttt{HB}), see equation below) --- is able to achieve acceleration\footnote{Here to be intended as a dependency of the rate on the square root of the condition number $L/\mu$.} in a neighborhood of the solution. This local argument becomes of course global in the quadratic case (for a simplified proof, see Proposition 1 in~\cite{lessard2016analysis}).
\begin{equation*}
\tag{\texttt{HB}}
\begin{cases}
q_{k+1} = q_k + \beta_k h p_k - h^2 \nabla f(q_k)\\
p_{k+1} = (q_{k+1}-q_k)/h
\end{cases}.
\end{equation*}
Despite the many attempts, nobody in the last 56 years has been able to show that \texttt{HB} has a global (i.e. for any initialization) accelerated rate --- neither in the strongly-convex case~(using a fixed momentum) nor in the non-strongly-convex case~(using an increasing $\frac{k-1}{k+2}$ momentum). Beyond the technical difficulty, another plausible reason may also be lack of interest, as the introduction of Nesterov's globally accelerated method in 1982, that overshadowed the conceptually simpler method from Polyak.
However, many researchers in the last decade, supported by numerical evidence and by the success of Heavy-ball in deep learning~\cite{kingma2014adam}, expressed their belief that \texttt{HB} is accelerated:
\vspace{3mm}
\begin{chapquote}{Ghadimi et al. \cite{ghadimi2015global}, 2015}
[\dots] supported by the numerical
simulations we envisage that the convergence factor could be strengthened even further. This
is indeed left as a future work.\vspace{-1mm}
\end{chapquote}
\begin{chapquote}{Gorbunov et al. \cite{gorbunov2019stochastic}, 2019}
Despite the long history of this approach,
there is still an open question whether the heavy ball method converges to the optimum globally with
accelerated rate when the objective function is twice continuous differentiable.\vspace{-3mm}
\end{chapquote}
\begin{chapquote}{Muehlebach and Jordan \cite{muehlebach2020optimization}, 2020}
Neither the evaluation of the gradient at a
shifted position, nor a specifically engineered damping parameter, as for example proposed
in Nesterov (2004, Sec. 2.2), \textit{seem}\footnote{After talking to the first author, we decided to replace ``\textit{are}'' (as in the original preprint) with ``\textit{seem}'': indeed, the argument in~\cite{muehlebach2020optimization} is asymptotic and therefore somewhat equivalent to the one of Polyak~\cite{polyak1964some}.} necessary.
\vspace{-3mm}
\end{chapquote}
Other researcher believe \texttt{HB} is not accelerated:
\vspace{3mm}
\begin{chapquote}{Shi et al. \cite{shi2018understanding}, 2018}
If we can translate this argument to the discrete case we can understand why \texttt{AGD} achieves
acceleration globally for strongly-convex functions but the Heavy-ball method does not.
\end{chapquote}
While on the theoretical side the opinion is mixed, on the experimental side no numerical simulation\footnote{In \cite{lessard2016analysis}, the authors show that there exist a strongly-convex smooth function such that Heavy-ball does not converge. However, as also pointed out by Ghadimi et al.~\cite{ghadimi2015global}, such function is not $C^2$, and that a big step-size is used --- which violates the convergence conditions of Thm.~4 in~\cite{ghadimi2015global}. As such, this function does not constitute a proper counterexample.} has been able to show that \texttt{HB} is not accelerated. In Figure~\ref{fig:conjecture}, we provide two examples for the non-strongly-convex case (i.e. $\mu$ very small, such that an increasing momentum is preferable, leading $1/k^2$ convergence as opposed to $(1-\sqrt{\mu/L})^k$). In particular, we show that \texttt{HB} is comparable to \texttt{AGD} through the lens of the pathological lower-bounding quadratic example introduced by~\cite{nemirovskii1979problem} and used to construct the $O(1/k^2)$ bound in convex optimization --- at least until the effect of non-trivial strong-convexity becomes dominant (at around $f(q_k)=10^{-6}$).
\begin{figure}[ht]
\centering
\begin{subfigure}{0.45\textwidth}\centering\includegraphics[width=1\columnwidth]{img/lower_bound}
\caption{\texttt{HB} and \texttt{AGD} on the worst-case (lower bound) quadratic objective from Nesterov~\cite{nesterov2018lectures}.}\end{subfigure}
\hspace{1cm}
\begin{subfigure}{0.45\textwidth}\centering\includegraphics[width=1\columnwidth]{img/regression}\caption{\texttt{HB} and \texttt{AGD} on ill-conditioned linear regression. The optimal step-size $1/L$ was used.}\end{subfigure}
\caption{For both examples \texttt{HB} with momentum $\frac{k-1}{k+2}$ exhibits an accelerated $1/k^2$ convergence rate, even though\texttt{ AGD} with momentum $\frac{k-1}{k+2}$ is faster in a neighborhood of the optimizer due to strong-convexity. Instead, \texttt{GD} violates the Nesterov $O(1/k^2)$ upper bound. We recall that, while Nesterov's upper bound holds for all $k>0$, the $O(1/k^2)$ lower bound~(originally discovered by Nemirovski and Yudin~\cite{nemirovskii1979problem}) only holds at $k = d/2$ (for more details, check the discussion in~\cite{nesterov2018lectures}).}
\label{fig:conjecture}
\vspace{-3mm}
\end{figure}
\subsection{Contributions}
The purpose of the manuscript at hand is to study the effect of shifts in gradient extrapolation points on acceleration in convex optimization (i.e. to study the difference between Heavy-ball and Nesterov's method). In particular, the next pages are organized as follows:
\vspace{-1mm}
\begin{enumerate}[leftmargin=0.48cm]
\item We start from a continuous-time argument: inspired by a recent idea from Flammarion and Bach~\cite{flammarion2015averaging}, in Section~\ref{sec:quadratic_invariants_cont} we show how \texttt{AGD-ODE} with damping $2/t$ can be derived from the equation of a simple harmonic oscillator: $\ddot u = - Au$. By using Lyapunov equations and a simple change of variables, we retrieve the Lyapunov function proposed by Su, Boyd and Candes~\cite{su2016} to prove a rate $O(1/t^2)$ for \texttt{AGD-ODE}. This procedure is principled and leads to many insights on Lyapunov function design.
\item In Section~\ref{sec:quadratic_invariants_disc}, we apply the same methodology in discrete time, and show that \texttt{HB} with momentum $\frac{k-1}{k+1}$ can be derived from the St{\"o}rmer--Verlet discretization of the simple harmonic oscillator. Solving again Lyapunov's equations, we are able to show an $O(1/k^2)$ rate for a Heavy-ball argorithm for convex quadratics. While this rate is already present in~\cite{flammarion2015averaging}, our proof technique is different as it relies on a Lyapunov function as opposed to an eigenvalue analysis.
\item In Section \ref{sec:gener_quadratic}, by generalizing the discrete-time Lyapunov function found in Section~\ref{sec:quadratic_invariants_disc} we derive a modified Heavy-ball method
$$q_{k+1} = q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1}) - h^2\frac{k+\frac{r-2}{2}}{k+r-1} \nabla f(q_k),$$
with a rate of convergence $O(1/k^2)$ for any $k\ge2$ and $r\ge2$. Our result not only generalizes the theory in~\cite{flammarion2015averaging}, but also provides an interesting connection between the continuous and the discrete --- as the used Lyapunov function converges, in the limit $h\to 0$, to the one used in~\cite{su2016} for $r\ge2$.
\end{enumerate}
\paragraph{Recent related works.} Very recently, Wang et al.~\cite{wang2022provable} proved that Heavy-ball is accelerated for a class of functions satisfying the Polyak-\L{}ojasiewicz condition. Instead, here we provide a Lyapunov function for the non-strongly-convex setting, where the Polyak-\L{}ojasiewicz constant vanishes. We remark that, for strongly-convex quadratic potentials, Heavy-ball is already known to achieve acceleration~\cite{lessard2016analysis}. However, the eigenvalue argument used in~\cite{lessard2016analysis} cannot be leveraged in the non-strongly-convex setting, where the minimum eigenvalue can be arbitrarily low. As such, our work provides insights on how to construct effective Lyapunov functions in the non-quadratic case, where Lyanonov arguments are often the go-to option.
\section{From quadratic invariants of oscillators to accelerated rates}
\label{sec:quadratic_invariants}
Our procedure in this section is inspired by a beautiful idea presented by Flammarion and Bach~\cite{flammarion2015averaging}: it is sometimes possible to translate a time-dependent convergence rate problem into a time-independent stability problem. Here we go one step further, and show how, with an additional step (computation of quadratic invariants), it is possible to derive Lyapunov functions and rates for the corresponding algorithms. We first illustrate the idea in continuous-time and then proceed with the discrete-time analysis.
Our starting point is the following ODE:
\begin{equation*}
\tag{\texttt{AGD-ODE}2}
\ddot q + \frac{2}{t} \dot q + \nabla f(q) = 0.
\end{equation*}
From the analysis in~\cite{su2016}, we know that on a quadratic $f(q) = f^*+\frac{1}{2}\langle (q-q^*),A(q-q^*)\rangle $, with $A$ positive semidefinite and $f^*\in\mathbb{R}$, the solution converges to $q^*\in\argmin_{x\in\mathbb{R}^d} f(q)$ at the rate $O(1/t^2)$. To prove this rate, the authors in~\cite{su2016} use the following Lyapunov function:
\begin{equation}
V(q,t) = 2t^2 (f(q)-f^*) + \|t\dot q + (q-q^*)\|^2.
\label{eq:ly_su_boyd_quadratics}
\end{equation}
We show here a \textit{constructive way} to derive $V$~(Section~\ref{sec:quadratic_invariants_cont}) and then~(Section~\ref{sec:quadratic_invariants_disc}) we apply the same procedure to get a Lyapunov function for Heavy-ball (i.e., the discretization). For simplicity, we consider here $f^*=0$ and $q^* =0$.
\subsection{Lyapunov functions from continuous-time invariants}
\label{sec:quadratic_invariants_cont}
Consider an harmonic oscillator on the potential $f(u) = \frac{1}{2}u^\top A u$, i.e. $\ddot u = -A u$. From basic physics, we know that such a system is marginally stable (bounded dynamics). By choosing $u = tq$ we get $\dot u = q + t\dot q $ and $\ddot u = \dot q + \dot q + t\ddot q$. This implies
\begin{equation*}
\dot q + \dot q + t\ddot q = \ddot u = -A t q \ \quad
\Rightarrow \quad \ \ddot q +\frac{2}{t} \dot q + Aq =0.
\end{equation*}
That is, \texttt{AGD-ODE} can be reconstructed from a simple linearized pendulum. By introducing the variable $v = \dot u$, we can write the pendulum in phase space as a linear dynamical system
$$\begin{pmatrix}
\dot u\\\dot v
\end{pmatrix} = \begin{pmatrix}
0 & I \\ -A& 0
\end{pmatrix}\begin{pmatrix}
u\\ v
\end{pmatrix}.$$
Hence, the pendulum has the form $\dot y = F y$, where $y = (u,v)$. We would now like to get a Lyapunov function for this system. To do this, we recall a fundamental proposition~(check Thm. 4.6. in \cite{khalil2002nonlinear}).
\begin{proposition}[Continuous-time Lyapunov equations]
The linear system $\dot y = Fy$ is Lyapunov stable if and only if for all positive semidefinite matrices $Q$, there exists a symmetric matrix $P$ such that
\begin{equation}
PF+F^TP = -Q.
\label{eq:lyapunov_eq_cont}
\end{equation}
Moreover, $V(y) = y^T P y$ is a Lyapunov function and $\dot V (y) = -y^T Q y$.
\end{proposition}
Since we know that a pendulum is only marginally stable (i.e., not asymptotically stable), we can limit ourselves to the choice of a null matrix $Q$. Hence, we need to solve the Lyapunov equation $PF = -F^TP$ for $P$. A solution to this equation~(many exist) is
$P = \begin{pmatrix}
A & 0 \\ 0& I
\end{pmatrix}$, which implies that
\begin{equation}
V(u) = \langle u,A u\rangle + \|v\|^2
\label{eq:energy_continuous}
\end{equation} is a quadratic invariant, i.e. $\dot V(u)=0$. This is well known, since $V$ is actually twice the total energy (Hamiltonian) of the pendulum. Finally, we can change variables and get that
\begin{equation}
V(q) = t^2\langle q,A q\rangle + \left\|\frac{\d}{\d t}(tq)\right\|^2 = 2t^2 f(q) + \|t\dot q + (q-q^*)\|^2,
\label{eq:ly_quad_cont}
\end{equation}
is a Lyapunov function for \texttt{AGD-ODE}2, with $\dot V(q)=0$. This is precisely equation~\ref{eq:ly_su_boyd_quadratics}.
\paragraph{From quadratic to convex.} With a small modification (using a factor $r-1$ instead of $1$), it is possible to get a Lyapunov function that works for \texttt{AGD-ODE} in the more general convex case.
\begin{proposition}[Theorem 3 from~\cite{su2016}] For convex $L$-smooth objectives, \texttt{AGD-ODE} converges at a rate $O(1/t^2)$. This follows from the fact that
\begin{equation}
V(q,t) = 2t^2 (f(q)-f^*) +\|t\dot q + (r-1)(q-q^*)\|^2.
\label{eq:lyapunov_su_general}
\end{equation}
is a Lyapunov function, for $r\ge3$.
\end{proposition}
\subsection{Discrete-time invariants}
\label{sec:quadratic_invariants_disc}
We apply the construction from the last subsection to the discrete case. Inspired by Flammarion and Bach~\cite{flammarion2015averaging}, we consider at first a slightly modified \texttt{HB}:
\begin{equation*}
\tag{\texttt{HB}$2$}
q_{k+1} = q_k + \frac{k-1}{k+1} (q_k-q_{k-1}) - h^2 \frac{k}{k+1} \nabla f(q_k).
\end{equation*}
This algorithm is the discrete-time equivalent of $\ddot q + \frac{2}{t} \dot q +\nabla f(q) = 0$. As for the continuous-time case, we start from $f(q) = \frac{1}{2}\langle q,A q \rangle$. In this case, \texttt{HB}$2$ can be written as
$$(k+1) q_{k+1} = 2 k q_k + (k-1) q_{k-1} - h^2 A (kq_k).$$
That is, if we set $u_k = kq_k$, we get
\begin{equation}
u_{k+1} - 2u_k + u_{k-1} = -h^2 A u_k.
\label{eq:stormer-verlet}
\end{equation}
With surprise, we recognize that this is the St{\"o}rmer--Verlet method~\cite{hairer2003geometric} on $\ddot u = -A u$, with step-size $h$ (that's why had $h^2$ from the very beginning).
It would be natural, as for the continuous-time case, to consider the total energy as a quadratic invariant to derive a Lyapunov function. However, it turns out that, interestingly, \textit{the St{\"o}rmer--Verlet method does not precisely conserve the total energy}: there are small oscillations~(see Section 3 in~\cite{hairer2014challenges})! Taking into account such small oscillations (Figure~\ref{fig:verlet}) is of fundamental importance --- since they lead to a crucial modification of the invariants we have to use.
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\textwidth]{img/verlet_energy.pdf}
\caption{The St{\"o}rmer--Verlet method on a one-dimensional quadratic potential (i.e., a simplified pendulum) does not conserve the total energy. Details on this phenomenon can be found in~\cite{hairer2016long,hairer2014challenges}.}
\label{fig:verlet}
\end{figure}
\begin{proposition}[Discrete-time Lyapunov equations]
The system $ y_{k+1} = Fy_k$ is Lyapunov stable if an only if for all positive semidefinite matrices $Q$, there exists a symmetric matrix $P$ such that
\begin{equation}
F^T P F - P = -Q.
\label{eq:lyapunov_eq_discrete}
\end{equation}
Moreover, $V(y) = y^T P y$ is a Lyapunov function and $V (y_{k+1})-V(y_k) = -y_k^T Q y_k$ for all $k$.
\end{proposition}
We apply the theorem above~(for $Q=0$) to the linear system
$$\begin{pmatrix}
u_{k+1}\\ v_{k+1}
\end{pmatrix} = \begin{pmatrix}
I-h^2 A & h I \\ -hA& I
\end{pmatrix}\begin{pmatrix}
u_k\\ v_k
\end{pmatrix}.$$
Under the choice $v_k = (u_{k}-u_{k-1})/h$, this system is equivalent to equation~\ref{eq:stormer-verlet}, i.e., the discretized pendulum we want to find a quadratic invariant for. Solving the discrete Lyapunov equation gives us
$$P = \begin{pmatrix}
A & -hA/2 \\ -hA/2& I
\end{pmatrix},$$
and the associated modified total energy:
\begin{equation}
V(u_k,v_k) = \underbrace{\langle u_k, A u_k\rangle +\|v_k\|^2}_{\text{continuous-time invariant (energy)}}- \underbrace{h\langle v_k, A u_k\rangle}_{\text{vanishing cross-term}}.
\label{eq:invariant_stormer}
\end{equation}
We make the following comments:
\begin{itemize}[leftmargin=0.48cm]
\item As $h\to 0$, the modified energy approaches the total energy in equation~\ref{eq:energy_continuous}. The purpose of the additional cross-term is to eliminate the small energy oscillations we see in Figure~\ref{fig:verlet}.
\item Assuming without loss of generality that $A$ does not have zero eigenvalues, $P$ is positive semidefinite (i.e., yields a valid Lyapunov function) if and only if the Schur complement of $I$ (i.e., the block $P_{22}$) in $P$ is positive semidefinite. That is, we need
\begin{equation}
B:=A-\frac{h^2}{4}A^2 = A\left(I-\frac{h^2}{4}A\right)\ge0.
\label{eq:B_matrix}
\end{equation}
Since $A$ and $(I-h^2A/4)$ are co-diagonalizable, the product is positive semidefinite if and only if both $A\ge 0$ and $(I-h^2A/4)\ge0$. This requires $0\le A\le \frac{4}{h^2}I$, which in turns implies an upper bound on the step-size $h^2$:
$$h^2\le\frac{4}{\lambda_{\text{max}}(A)} = \frac{4}{L}.$$
The same condition~(note that our step-size is $h^2$, not $h$) can be deduced from the analysis of \texttt{HB}$2$ in~\cite{flammarion2015averaging}.
\end{itemize}
Now it's time to change variables back: $u_k = kq_k$. If we set $p_k := (q_k-q_{k-1})/h$, as also done in the introduction, we get
\begin{align*}
hv_{k} &= h(u_{k}-u_{k-1})/h\\ &= k q_k - (k-1) q_{k-1}+ (k-1) q_{k}- (k-1) q_{k}\\
&= (k-1)(q_k-q_{k-1})+q_k\\
&= h(k-1)p_k+q_k.
\end{align*}
By substituting these formulas in equation~\ref{eq:invariant_stormer}, we get the following final form for an effective Lyapunov function for \texttt{HB}$2$ --- for the quadratic case:
\begin{align*}
&V_k=\langle u_k, A u_k\rangle + \|v_k\|^2-\langle h v_k, A u_k\rangle\\ \Longrightarrow \ & V_k = k^2\langle q_k,Aq_k\rangle + \frac{1}{h^2}\|h(k-1)p_k+q_k\|^2-k\langle h(k-1)p_k+q_k, A q_k\rangle.
\end{align*}
To better understand this Lyapunov function, we multiply everything by $h^2$ and get
$$V_k = (kh)^2\langle q_k,Aq_k\rangle + \|h(k-1)p_k+q_k\|^2 - h^2k\langle h(k-1)p_k+q_k, A q_k\rangle.$$
Recalling that the ``time'' variable $t$ is defined to be $t_k = hk$, this cost becomes
$$V_k = t_k^2\langle q_k,Aq_k\rangle + \|t_{k-1}p_k+q_k\|^2 - h t_k\langle t_{k-1}p_k+q_k, A q_k\rangle.$$
This Lyapunov function can be easily generalized by noting that $\langle q_k,Aq_k\rangle = 2(f(q_k)-f^*)$ and $A q_k = \nabla f(q_k)$:
\begin{equation}
V_k = 2t_k^2 (f(q_k)-f^*) + \|t_{k-1}p_k+q_k\|^2 - t_k\langle\nabla f(q_k), t_{k-1}p_k+q_k\rangle.
\label{eq:lyapunov_HB_first}
\end{equation}
Finally, note that
\begin{itemize}[leftmargin=0.48cm]
\item From equation~\ref{eq:lyapunov_HB_first} as $h\to0$ we get equation~\ref{eq:ly_quad_cont}: the continuous-time Lyapunov function of~\cite{su2016}.
\item The mixing term is necessary and makes the positive definiteness~(see equation~\ref{eq:B_matrix}) of $V_k$ non-trivial.
\end{itemize}
All in all, in this subsection, we proved the following result.
\begin{tcolorbox}
\begin{proposition}
Let $A\in\mathbb{R}^{d\times d}$ be positive semidefinite, $f^*\in\mathbb{R}$ and $f(q)= f^* + \frac{1}{2}(q-q^*)^T A (q-q^*)$. Let $(q_k)_{k\ge0}$ be the iterates of \texttt{HB}$2$ and $p_k:=(q_k-q_{k-1})/h$. If the step-size $h^2<\frac{4}{\lambda_\text{max}(A)}$, then equation~\ref{eq:lyapunov_HB_first} is non-negative and such that $V_{k+1}-V_k =0$ along the \texttt{HB}$2$ trajectory, for all $k$. From this, one can deduce an accelerated rate of $O(1/k^2)$ in suboptimality.
\label{prop:proof_quadratic_bach}
\end{proposition}
\end{tcolorbox}
Details of the proof are given in the proof of Theorem~\ref{thm:proof_quadratic}~(more general).
\section{Accelerated Heavy-ball methods for convex quadratics}
\label{sec:gener_quadratic}
In this section, we start to lift the discussion to the convex non-quadratic setting, by providing a generalization of \texttt{HB}2. Indeed, we know from the continuous-time analysis in~\cite{su2016} that $\ddot q + \frac{2}{t}\dot q + \nabla f(q)=0$ may not have an accelerated rate for functions which are convex but not necessarily quadratic. In this case, a rate of $O(1/t^2)$ only holds~\footnote{The case $0<r\le3$ was studied by Attouch et al.~\cite{attouch2019rate}: a convergence rate of $O(1/t^p)$ with $p<2r/3$ is shown in this case. The same result also holds in discrete time.} for $$\ddot q + \frac{r}{t}\dot q + \nabla f(q)=0,$$ with $r\ge 3$. In the same way, we expect that \texttt{HB}$2$~(which is the discretization for $r=2$) may not have an accelerated rate in the convex non-quadratic setting and a \textit{generalization corresponding to high friction is therefore necessary}.
Our objective in this chapter is to construct such a generalization of \texttt{HB}$2$, which we name \texttt{HB}$r$.
\subsection{A generalized Heavy-ball with high friction and guarantees on quadratics}
After a few weeks of intense calculations, we found that this algorithm gives the desired result~(Thm.~\ref{thm:proof_quadratic}).
\begin{equation*}
\tag{\texttt{HB}$r$}
q_{k+1} = \underbrace{q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})}_{\text{iterate + momentum}} - h^2\underbrace{ \frac{k+\frac{r-2}{2}}{k+r-1} \nabla f(q_k)}_{\text{scaled gradient of iterate}}.
\end{equation*}
First, note that $r=2$ recovers \texttt{HB}$2$ --- which we proved to be accelerated in the last subsection using a novel Lyapunov argument. The second, and perhaps the most crucial, thing to note is that \texttt{HB}$r$ recalls the high friction generalization of \texttt{AGD} proposed by~\cite{su2016}~(see Theorem 6 in their paper):
\begin{equation*}
\tag{\texttt{AGD}$r$}
q_{k+1} = \underbrace{q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})}_{\text{iterate + momentum}} - h^2 \underbrace{\nabla f\left(q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})\right)}_{\text{gradient of [iterate + momentum]}}.
\end{equation*}
Between \texttt{HB}$r$ and \texttt{AGD}$r$ there are a few important differences:
\begin{itemize}[leftmargin=0.48cm]
\item In \texttt{AGD}$r$ the gradient is evaluated at $q_k + \frac{k-1}{k+r-1} (q_k-q_{k-1})$, while in \texttt{HB}$r$ it is evaluated at $q_k$.
\item in \texttt{HB}$r$ the effective step-size~(i.e. what multiplies the gradient) is iteration-dependent, and goes from $h^2/2$ to $h^2$ as $k\to\infty$. We believe this has \textit{not to be regarded as part of the acceleration mechanism}: it is just a small modification needed to make the analysis easier.
\item Arguably \texttt{HB}$r$~(neglecting the small correction) is conceptually simpler that \texttt{AGD}$r$: compared to \texttt{GD}, only a momentum term is added at each iteration --- and this can be thought of as the source of acceleration.
\end{itemize}
We proceed in proving that \texttt{HB}$r$ is accelerated in the quadratic case.
\begin{tcolorbox}
\begin{theorem}Let $A\in\mathbb{R}^{d\times d}$ be positive semidefinite, $f^*\in\mathbb{R}$ and $f(q)= f^* + \frac{1}{2}(q-q^*)^T A (q-q^*)$. Let $(q_k)_{k\ge0}$ be the iterates of \texttt{HB}$r$~($r\ge 2$) and $p_k:=(q_k-q_{k-1})/h$. If $h^2\le\frac{4}{\lambda_\text{max}(A)}$, then
\begin{multline}
V_k = 2(k+r-2)^2 h^2 (f(q_k)-f^*) + \|h(k-1)p_k+(r-1)(q_k-q^*)\|^2\\ - h^2 (k+r-2)\langle\nabla f(q_k), h(k-1)p_k+(r-1)(q_k-q^*)\rangle
\label{eq:lyapunov_HB}
\end{multline}
is non-negative and such that $V_{k+1}-V_k \le0$ along the \texttt{HB}$r$ trajectory, for all $k$. Moreover, for any $h^2<\frac{4}{\lambda_\text{max}(A)}$, \texttt{HB}$r$ is accelerated. In particular, if $h^2=\frac{2}{\lambda_\text{max}(A)}$, we have the rate
$$f(q_k)-f^*\le \frac{\lambda_\text{max}(A)V_0}{2(k+r-2)^2}.$$
\label{thm:proof_quadratic}
\end{theorem}
\end{tcolorbox}
We note a couple of facts about the Lyapunov function $V_k$ in equation~\ref{eq:lyapunov_HB}.
\begin{itemize}[leftmargin=0.48cm]
\item It reduces to equation~\ref{eq:lyapunov_HB_first} in the case $r=2$. For $r>2$, it is a graspable generalization of equation~\ref{eq:lyapunov_HB_first} --- which we instead derived in a systematic way using Lyapunov equations. The term $(r-1)$ is inspired by the continuous-time limit in equation~\ref{eq:lyapunov_su_general}.
\item Consider the Lyapunov function above, but without cross term i.e.
$$2(k+r-2)^2 h^2 (f(q_k)-f^*) + \|h(k-1)p_k+(r-1)(q_k-q^*)\|^2.$$
This function works for proving an $O(1/k^2)$ rate for \texttt{AGDr} (it's a Lyapunov function, see Thm. 6 from~\cite{su2016}). Therefore, higher complexity~(i.e., an additional cross term) is needed to study the acceleration of Heavy-ball, when compared to Nesterov's method.
\item As $h\to0$, the cross term vanishes $V_k$ converges to equation~\ref{eq:lyapunov_su_general} --- its continuous-time equivalent. Indeed, both \texttt{HBr} and \texttt{AGDr} converge to \texttt{AGD-ODE} as $h\to 0$.
\end{itemize}
\subsection{Proof of the theorem}
It is useful to simplify equation~\ref{eq:lyapunov_HB} and to work with variables $q_k$ and $q_{k-1}$ --- a natural choice in the discrete setting. We split the Lyapunov function into two parts: $V_k = V^1_k + V^2_k$.
\begin{align}
&V^1_k := 2(k+r-2)^2 h^2 (f(q_k)-f^*)\label{eq:V1}\\ &\quad \quad\quad - h^2 (k+r-2)\langle\nabla f(q_k),(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\rangle.\nonumber\\
& V^2_k := \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.\label{eq:V2}
\end{align}
First, we are going to study $V_k^2$ in the non-quadratic case, and then $V_k^1$ in the quadratic case. Theorem~\ref{thm:proof_quadratic} will follow from a combination of the two corresponding lemmata.
The first lemma shares many similarities with the proof of Theorem 1 in~\cite{ghadimi2015global}.
\begin{lemma}
For any differentiable function $f:\mathbb{R}^d\to\mathbb{R}$ (not necessarily convex or $L$-smooth) and any sequence of iterates $(q_k)_{k\ge0}$ returned by \texttt{HB}$r$, we have:
\begin{align*}
V_{k+1}^2-V_k^2\ = \ &-h^2(r-1)(2k+r-2)\langle\nabla f(q_k),q_k-q^*\rangle\\
&-h^2(k-1)(2k+r-2)\langle\nabla f(q_k),q_{k}-q_{k-1}\rangle\\
&+\frac{h^4}{4}(2k+r-2)^2\|\nabla f(q_k)\|^2,
\end{align*}
where $V^2_k$ is defined in equation~\ref{eq:V2}.
\label{lemma:V2_quad}
\end{lemma}
\begin{proof}
Let $g_{k} := (k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)$, then,
$$V^2_{k+1}-V^2_k = \|g_{k+1}\|^2-\|g_{k}\|^2 = \langle g_{k+1}+g_{k},g_{k+1}-g_k\rangle.$$
We proceed in computing $g_{k+1}-g_k$. The algorithm symmetric structure here is fundamental:
\begin{align*}
g_{k+1}-g_k = \ & k(q_{k+1}-q_{k})+(r-1)(q_{k+1}-q^*)-(k-1)(q_k-q_{k-1})-(r-1)(q_{k}-q^*)\\
= \ & (k+r-1)q_{k+1}-(k+r-1)q_k-(k-1)(q_k-q_{k-1})\\
\overset{\text{(\texttt{HB}$r$)}}{=} & - h^2\left(k+\frac{r-2}{2}\right) \nabla f(q_k).
\end{align*}
Instead, $g_{k+1}+g_k$ is slightly more complex.
\begin{align*}
g_{k+1}+g_k = \ & k(q_{k+1}-q_{k})+(r-1)(q_{k+1}-q^*)+(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\\
= \ & (k+r-1)q_{k+1}+(-k+r-1)q_k+(k-1)(q_k-q_{k-1})-2(r-1)q^*\\
\overset{\text{(\texttt{HB}$r$)}}{=} &(k+r-1)q_{k} + (k-1)(q_k-q_{k-1})- h^2\left(k+\frac{r-2}{2}\right) \nabla f(q_k)\\&+(-k+r-1)q_k+(k-1)(q_k-q_{k-1})-2(r-1)q^*\\
= \ &2(r-1)(q_{k}-q^*)+2(k-1)(q_k-q_{k-1})- h^2\left(k+\frac{r-2}{2}\right) \nabla f(q_k).
\end{align*}
The proof is concluded by taking the inner product.
\end{proof}
We proceed by computing the difference $V^1_{k+1}-V^1_k$. Our calculations will be very quick, since we can leverage, in the quadratic case, on a simplified expression for $V^1_k$.
\begin{lemma}
Let $V^1_k$ be defined as equation~\ref{eq:V1}. In the context of Theorem~\ref{thm:proof_quadratic}, we have
$$V^1_k = h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle$$
and
\begin{align*}
V_{k+1}^1-V_k^1\ = & \ h^2(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle\\
&+h^2(k-1)(2k+r-2)\langle q_k-q^*, A(q_k-q_{k-1})\rangle\\
&-\frac{h^4}{2} (2k+r-2)k\|A(q_k-q^*)\|^2.
\end{align*}
\vspace{-3mm}
\label{lemma:V1_quad}
\end{lemma}
\begin{proof}
From equation~\ref{eq:V1}, we get
\begin{align*}
V^1_k = \ & \ (k+r-2)^2 h^2 \langle q_k-q^*, A(q_k-q^*)\rangle\\ &- h^2 (k+r-2)\langle A(q_k-q^*),(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\rangle\\
= \ & \ (k+r-2)^2 h^2 \langle q_k-q^*, A(q_k-q^*)\rangle\\
&\ - h^2 (k+r-2)\langle A(q_k-q^*),(k+r-2)(q_k-q^*)-(k-1)(q_{k-1}-q^*)\rangle\\
= \ & \ h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle.
\end{align*}
We proceed computing $V^1_{k+1}-V^1_{k}$ using this simplified form:
\begin{align*}
V^1_{k+1} - V^1_k = \ & \ h^2(k+r-1)k\langle q_{k}-q^*,A(q_{k+1}-q^*)\rangle\\&-h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle\\
= \ & h^2\langle q_{k}-q^*,A \Delta_k\rangle,
\end{align*}
where
\begin{align*}
\Delta_k := \ & \ (k+r-1)k(q_{k+1}-q^*)-(k+r-2)(k-1)(q_{k-1}-q^*).
\end{align*}
Now, recall the definition of \texttt{HB}$r$:
$$(q_{k+1}-q^*) = (q_k-q^*) + \frac{k-1}{k+r-1} (q_k-q_{k-1}) - h^2 \frac{k+\frac{r-2}{2}}{k+r-1}A(q_k-q*),$$
where we subtracted $q^*$ from both sides. By plugging this into $\Delta_k$, we get
\begin{align*}
\Delta_k = \ & \ (k+r-1)k(q_k-q^*) + (k-1)k (q_k-q_{k-1}) - \frac{h^2}{2} (2k+r-2)k A(q_k-q*)\\
&-(k+r-2)(k-1)(q_{k-1}-q^*)\\
=\ & \ (2k+r-2)k(q_k-q^*)-(2k+r-2)(k-1)(q_{k-1}-q^*)\\&- \frac{h^2}{2} (2k+r-2)k A(q_k-q^*)\\
=\ & \ (2k+r-2)(q_k-q^*)+(2k+r-2)(k-1)(q_k-q_{k-1})- \frac{h^2}{2} (2k+r-2)k A(q_k-q^*).
\end{align*}
The result follows after taking the inner product $h^2\langle q_{k}-q^*,A \Delta_k\rangle$.
\end{proof}
We are finally ready to prove the result.
\begin{proof}[Proof of Theorem~\ref{thm:proof_quadratic}]
First, we compute $V_{k+1}^1-V_k = (V_{k+1}^1-V_k^1)+(V_{k+1}^2-V_k^2)$ using Lemma~\ref{lemma:V1_quad} and Lemma~\ref{lemma:V2_quad}~(written for quadratic $f$). Next, we show that a certain condition on the step-size implies positivity of $V_k$ and a convergence rate.
\begin{align*}
(V_{k+1}^1-V_k^1)+(V_{k+1}^2-V_k^2)\ = & \ h^2(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle\\
&\hcancel{+h^2(k-1)(2k+r-2)\langle q_k-q^*, A(q_k-q_{k-1})\rangle}\\
&-\frac{h^4}{2} (2k+r-2)k\|A(q_k-q^*)\|^2\\ &-h^2(r-1)(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle\\
&\hcancel{-h^2(k-1)(2k+r-2)\langle q_k-q^*,A(q_{k}-q_{k-1})\rangle}\\
&+\frac{h^4}{4}(2k+r-2)^2\|A(q_k-q^*)\|^2.
\end{align*}
Crucially, note that the terms including $\langle q_k-q^*,A(q_{k}-q_{k-1})\rangle$ cancel. This is necessary to make our proof~(or, probably, any proof) work, since such inner product between the gradient and the momentum changes sign (infinitely) many times along the trajectory, and therefore cannot be easily compared to other quantities. For the same reason, in the corresponding continuous-time proof from~\cite{su2016}, the terms including $\langle \nabla f(q), p\rangle$ also perfectly cancel out.
All in all, by collecting some terms, we get
\begin{align*}
V_{k+1}-V_{k} &= -h^2(r-2)(2k+r-2)\langle q_k-q^*, A(q_k-q^*)\rangle \\& \quad \quad+ \frac{h^4}{4}(2k+r-2)(r-2)\|A(q_k-q^*)\|^2\\
&=-h^2(r-2)(2k+r-2)\langle q_k-q^*,B(q_k-q^*)\rangle,
\end{align*}
where $B:=A-\frac{h^2}{4}A^2$ is the matrix that we already studied in the context of Lyapunov equations~(see equation~\ref{eq:B_matrix}). Since $r\ge2$, a sufficient condition for $V_{k+1}-V_{k}\le 0$ is $B\ge0$, which holds under $h^2\le\frac{4}{\lambda_\text{max}(A)}$. As a sanity check, the reader can appreciate the fact that, if $r=2$, then $V_{k+1}=V_k$ --- as we already proved in Proposition~\ref{prop:proof_quadratic_bach}~(follows from the fact that $V_k$ solves the Lyapunov equations).
Last, we have to translate the fact that $V_k$ is non-increasing to a convergence rate. This is not trivial in our case, since $V_k$ also contains a cross term which is not necessarily positive. Actually, we do not even know that $V_k\ge0$ yet! Hence, we have to come up with some tricks. We start from rewriting the~(simplified) Lyapunov function:
$$V_k =
h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle + \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.$$
Now, let us add and subtract a term $c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle$, with $c>0$. We have:
$$V_k = c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle + \tilde V_k,$$
with
\begin{align*}
\tilde V_k :=& -c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle \\&\quad + h^2(k+r-2)(k-1)\langle q_{k-1}-q^*,A(q_k-q^*)\rangle\\&\quad + \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.
\end{align*}
Now, if we show that $\tilde V_k$ is always positive, then $V_{k+1}\le V_k$ for all $k$ implies:
$$c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle \le c h^2(k+r-2)^2\langle q_{k}-q^*,A(q_k-q^*)\rangle + \tilde V_k= V_k\le V_0,$$
which gives the desired rate:
$$ f(q_k)-f^*=\frac{1}{2}\langle q_{k}-q^*,A(q_k-q^*)\rangle \le \frac{V_0}{2c h^2(k+r-2)^2}.$$
Therefore, we only need to show $\tilde V_k\ge0$. To do this, we introduce two new variables:
\begin{equation*}
u_{k} := (k+r-2)(q_k-q^*),\quad\quad w_{k} := (k-1)(q_{k-1}-q^*),
\end{equation*}
and get a simplified form for $\tilde V_k$
\begin{align*}
\tilde V_k &= -c h^2 \langle u_k,Au_k\rangle + h^2\langle v_k,A w_k\rangle + \|u_k-w_k\|^2\\
&=\langle u_k, (I-ch^2A)u_k\rangle + \|w_k\|^2 - 2\langle u_k,\left(I-\frac{h^2}{2}A\right)w_k\rangle.
\end{align*}
Hence, we just need to show that
$$\tilde P =\begin{pmatrix}
I-ch^2A & -\left(I-\frac{h^2}{2}A\right)\\
-\left(I-\frac{h^2}{2}A\right) & I
\end{pmatrix}$$
is positive definite, for some $c$ and $h^2$. Using the Schur characterization for positive semidefinite matrices, $\tilde P\ge0$ if and only if
\begin{align*}
0\le\tilde B(c) &:=I-ch^2A-\left(I-\frac{h^2}{2}A\right)^2\\
&=I-ch^2A-I-\frac{h^4}{4}A^2+h^2A\\
&=h^2 A\left(1-c-\frac{h^2}{4}A\right).
\end{align*}
It is clear that $\tilde B(c)$ is positive semidefinite if and only if
$1-c-\frac{h^2}{4}\lambda_{\text{max}}(A)\ge 0$. That is,
$$h^2\le\frac{4(1-c)}{\lambda_\text{max}(A)}.$$
Hence, for any $c\in (0,1)$ we get an acceleration. In particular, in the theorem, we chose $c=\frac{1}{2}$.
\end{proof}
\subsection{Numerical verification of our Lyapunov function} We verify numerically that the Lyapunov function for \texttt{HB}r proposed in equation~\ref{eq:lyapunov_HB} works on quadratics. To more clearly show the effect the inner product correction term, which originated from the quadratic invariant of the St{\"o}rmer--Verlet method, we use here a slightly different notation: $V_k = V^{11}_k+V^{12}_k+V^{2}_k$, with $V^{1}_k = V^{11}_k+V^{12}_k$.
\begin{align*}
V^{11}_k &:= 2(k+r-2)^2 h^2 (f(q_k)-f^*);\\
V^{12}_k&:= - h^2 (k+r-2)\langle\nabla f(q_k),(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\rangle;\\
V^{2}_k &:= \|(k-1)(q_k-q_{k-1})+(r-1)(q_{k}-q^*)\|^2.
\end{align*}
We recall that, the term $V^{12}_k$ (a.k.a. the \textit{cross-term}) vanishes as $h\to0$, and is indeed not present in the continuous-time limit. We show that this term, which we derived using Lyapunov equations in Sec.~\ref{sec:quadratic_invariants}, plays a fundamental role in ensuring $V_{k+1}-V_k\le0$. In Figure~\ref{fig:lyapunov_HBr_linreg_1} we verify numerically Thm.~\ref{thm:proof_quadratic}. In Figure~\ref{fig:lyapunov_HBr_linreg_3} we show the essential role of $V^{12}$. Here we used $h^2=1/L$, but \texttt{HB}$r$ can take larger steps~(up to $4/L$), while the other algorithms become unstable~(Figure~\ref{fig:HBr_big_step}).
\clearpage
\begin{figure}[ht!]
\centering
\begin{subfigure}{\textwidth}\centering\includegraphics[width=0.8\columnwidth]{img/regression_ly_r2.png}
\end{subfigure}
\vspace{-4mm}
\begin{subfigure}{\textwidth}\centering\includegraphics[width=0.8\columnwidth]{img/regression_ly_r3.png}
\end{subfigure}
\vspace{-2mm}
\caption{Dynamics of the Lyapunov function for \texttt{HB}$r$ on linear regression~(ill-conditioned Hessian, with condition number $\kappa$). Shown is the behavior for $r=2,3$ with step-size $1/L$. For $r=2$, $V_{k}$ is constant, as predicted by Prop.~\ref{prop:proof_quadratic_bach}. For $r=3$, $V_{k}$ is decreasing as predicted by Thm.~\ref{thm:proof_quadratic}. }
\label{fig:lyapunov_HBr_linreg_1}
\end{figure}
\vspace{-2mm}
\begin{figure}[ht!]\centering\includegraphics[width=0.8\columnwidth]{img/regression_ly_no.png}
\vspace{-4mm}
\caption{Same setting of the second example in Figure~\ref{fig:lyapunov_HBr_linreg_1}, but different candidate Lyapunov function (no cross term). This confirms the cross-term is necessary.}
\label{fig:lyapunov_HBr_linreg_3}
\end{figure}
\vspace{-2mm}
\begin{figure}[ht!]\centering\includegraphics[width=0.35\columnwidth]{img/regression_big_step.pdf}
\vspace{-2mm}
\caption{\texttt{HB}r also works for big step-sizes~(see conditions in Theorem~\ref{thm:proof_quadratic}). Here, used is $h^2=3.9/L$.}
\label{fig:HBr_big_step}
\end{figure}
\section{Conclusion}
\vspace{-1mm}
In conclusion, the question of whether the Heavy-ball method is globally accelerated for non-strongly-convex quadratic problems has yet to be fully answered, and has attracted the attention of recent research~\cite{wang2022provable}. Our study takes a novel approach by examining momentum through the lens of quadratic invariants of simple harmonic oscillators, and by utilizing the modified Hamiltonian of Stormer-Verlet integrators we were able to construct a Lyapunov function that demonstrates an $O(1/k^2)$ rate for Heavy-ball in the case of convex quadratic problems, where eigenvalues can vanish. This is a promising first step towards potentially proving the acceleration of Polyak's momentum method through Lyapunov function arguments.
\vspace{-1mm}
\section{Acknowledgements}
\vspace{-1mm}
I would like to extend my deepest gratitude to Prof. Boris Polyak, Prof. Christian Lubich, and Konstantin Mishchenko for the stimulating discussions. My appreciation goes to Prof. Aurelien Lucchi and Prof. Thomas Hofmann for their unwavering support and motivation, which helped me to develop the project idea in Spring 2020. Lastly, I cannot express enough my gratitude to Johannes Brahms for his Violinkonzert D-Dur op. 77, which provided the perfect soundtrack to my late-night calculations, igniting my passion and drive to push through the toughest moments.
\printbibliography
\bibstyle{apalike}
\end{document}
|
\section{REFERENCES}
\bibliographystyle{IEEEbib}
\section{Introduction}
\label{sec:intro}
Intelligent systems are at a crossroad between perception and prediction. There has been significant advance on tackling the challenges of perception using learning-based algorithms. For instance, robust models \cite{prabhushankar2021extracting, lee2021open, kwon2020novelty, lehman2020structures} have been developed for image perception-based applications under challenging conditions. In addition, the success of learning-based models on natural images has fostered its usage on computational image analysis including seismic \cite{benkert2021explaining, mustafa2021man} and medical \cite{temel2019relative} fields. While perception algorithms endows an intelligent system with the ability to see the current surroundings, prediction on future states informs intent and suggests safe maneuver.
\begin{figure}[th]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=8.5cm]{figures/Intro_interact.pdf}}
\end{minipage}
\caption{Example frames from the PIE dataset where a pedestrian intends to cross while interacting with traffic objects. The left figure shows a pedestrian 1) passes the vehicle b at further distance 2)
slows down as the closer vehicle a brakes. The observed trajectory is shown in green dashed curve.}
\label{fig:intro-interact}
\end{figure}
Predicting future states of traffic agents, especially pedestrians, in complex urban scenes is safety-critical for intelligent transportation systems. The future states of pedestrians can be in the form of upcoming actions \cite{rasouli2020pedestrian}, e.g., road crossing. In this paper, we consider predicting future crossing as pedestrian intent prediction.
Prediction can be challenging when pedestrian's future crossing depends on other traffic agents and environment. Traffic agents constantly avoid collisions by interacting within their local neighborhood as illustrated in Fig. \ref{fig:intro-interact}. Looking into local interaction at different regions, e.g., pedestrian $\leftrightarrow$ car a and pedestrian $\leftrightarrow$ car b together informs how their behavior change from one neighborhood to another across time from a global respective. Hence, we hypothesize that learning relation from spatiotemporal region-wise interaction with respect to a target agent informs intent. We develop a framework that learns relation from pedestrian visual interaction and motion dynamics to estimate the probability of binary crossing intent. The contributions of this paper include:
\begin{enumerate}
\item We develop a relation extraction module that infers relation between pedestrian and surroundings from region-wise local interaction.
\item We show that integrating region-wise relation learning into prediction framework achieves competitive performance against existing methods. The improvement in F1-score and recall suggests lower ratio of missed crossing predictions.
\end{enumerate}
\section{Related Work}
\label{sec:format}
\textbf{Pedestrian Intent prediction}
The objective of pedestrian intent prediction, specifically crossing prediction, is to anticipate whether a target pedestrian will cross the road at some time in the future. \cite{saleh2019real} extends DenseNet to extract spatio-temporal features from image sequences to estimate pedestrians’ crossing actions. Many algorithms rely on various input modalities and architectures \cite{rasouli2017they, rasouli2020pedestrian, rasouli2020multi}. For instance, the framework in \cite{rasouli2017they} utilizes traffic scene features and pedestrians looking and walking attributes by convolutional networks to forecast crossing action. Some recent methods utilize recurrent modeling \cite{rasouli2020pedestrian, bouhsain2020pedestrian}. The authors in \cite{rasouli2020pedestrian} develop a framework with different input modalities including pedestrian trajectories, poses, traffic scenes, and fuse them hierarchically within a stacked recurrent model.
\textbf{Relation Modeling}
Some studies model the dependency between pedestrians and traffic surroundings with object-level representations. For instance, the authors in \cite{yao2021coupling} design a module to extract object-wise interaction between various traffic objects based on their pre-detected location and visual states. Different from their method, we develop a relation extraction module that aggregates region-wise joint interaction to inform prediction.
\section{Technical Approach}
\label{sec:pagestyle}
In this section, we introduce our framework to predict the probability of pedestrian crossing intent with region-wise relation representations and motion dynamics. We formalize the intent prediction problem in {subsection {3}.1}. We describe how relation representations with respect to motion dynamics are learned through our framework in {subsection {3}.2}.
\subsection{Problem Definition}
We define the pedestrian intent prediction as binary classification that predicts whether a target pedestrian will cross the road in the future based on the observation data. Without loss of generality, we assume that the observed trajectory of a target pedestrian is obtained from an object detection and tracking framework in advance. Let $X_i = [X_i^{t-\tau+1},..., X_i^{t}]$ denote the past trajectory of $i^{th}$ pedestrian from time $t-\tau+1$ to time $t$, where $X_i^{t}=(u_i^t, v_i^t, w_i^t, h_i^t)\in{\rm I\!R}^4 $ denotes the bounding box center, width, height of $i^{th}$ pedestrian at time $t$. Given the past trajectory $X_i$ along with the observed frames $I=[I^{t-\tau+1},..., I^{t}]$, we estimate the $i^{th}$ pedestrian's crossing probability $\hat{y}_i$.
\begin{figure}[th]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=7.5cm]{figures/RATP_visual_encoder.pdf}}
\end{minipage}
\caption{Joint information between object and surroundings is encoded within feature maps by a visual encoder.}
\label{fig:ratp-visual-encoder}
\end{figure}
\subsection{Framework Overview}
Our prediction framework consists of three major components: ({1}) a visual encoder to extract region-wise interaction evolving along time; ({2}) a relation extraction module that jointly relate local interaction from different regions to obtain relation from a global perspective; ({3}) a trajectory recurrent encoder that encodes pedestrian's observed motion dynamics. We elaborate each component in more details in the following.
\textbf{Visual Encoder}
We employ a visual convolutional encoder that extracts local joint interaction between object and surroundings from observation frame sequence $I$ as shown in Fig.\ref{fig:ratp-visual-encoder}. The visual encoder can be divided into two parts: ({1}) a {spatial module} that encodes frame-level joint local interaction within feature-map vectors; ({2}) a {temporal module} that aggregates frame-level feature-map vectors to extract spatiotemporal local interaction.
The spatial module takes $\tau$ observed frames $I=[I^{t-\tau+1},..., I^{t}]$ as inputs and passes them through a 2D ResNet34 \cite{he2016deep} backbone $G_{sf}$ to generate frame-level features
$f_s\in {\rm I\!R}^{\tau\times h\times w\times c} $. The feature vectors (cells) in $f_s$ encode joint interaction within local regions, i.e, the feature vector in orange encodes pedestrian-car information. To capture temporal changes of joint information locally, we pass $f_s$ through a temporal module $G_{tf}$ to generate spatiotemporal local interaction representations $f_{st}$.
Specifically, we employ $\tau\times1\times1$ convolutional 3D filters with $c=$512 channels to aggregate cells $f_s^{(k)}\in {\rm I\!R}^{\tau\times 1\times 1\times c}$ in $f_s$ along temporal axis to generate $f_{st}\in {\rm I\!R}^{h\times w\times c}$, within which a cell is a $c$-dimensional spatiotemporal feature vector $f_{st}^{(k)}\in {\rm I\!R}^{c}$.
\begin{figure*}[t]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=14.5cm]{figures/RATP_detailed_arch.pdf}}
\end{minipage}
\caption{{The detailed architecture of our intent prediction framework. Each pairwise relation between $f_{st}^m$ and $f_{st}^n$ is conditioned on past agent-specific motion states $q_i$.}}
\label{fig:RATP-detailed-arch}
\end{figure*}
\textbf{Relation Extraction}
For a target pedestrian, relating local interaction from different cells $f_{st}^{(k)}$ (e.g., car-car and pedestrian-car) informs how behavioral interaction changes from one local region to another when objects moving around in the scene. While the visual encoder encodes spatiotemporal local interaction representations $f_{st}^{(k)}$, it less explicitly relate region-wise representations to jointly learn relation. Hence, we employ a relation extraction module to extract global-wise relation from
feature vectors $f_{st}^{(1)},...f_{st}^{(k)},...,f_{st}^{(h\times w)}$ in $f_{st}$. Inspired by \cite{santoro2017simple}, we extend the notion of entity to a spatiotemporal feature representation encoded from a local region across time, which is $f_{st}^{(k)}$ in our case.
Hence, we define an entity set as $f_{st} = \{f_{st}^{(1)},...f_{st}^{(k)},...,f_{st}^{(h\times w)}\}$.
Following such definition, we learn the relation between entities in $f_{st}$ via a composite function $R(*)$ as following:
\begin{equation}
R(f_{st}) = f_\phi \left(\sum_{m,n}g_\theta \left( f_{st}^m, f_{st}^n \right)\right)
\end{equation}
where $g_\theta(*)$ denotes a pair-wise relation function that takes a pair of feature vectors $f_{st}^m, f_{st}^n$ in $f_{st}$ as input to learn their relation. The outcomes of $g_\theta(*)$ from all feature vector pairs are then aggregated to be passed through $f_\phi(*)$ to infer relations globally. In general, $f_\phi(*)$ and $g_\theta(*)$ can be simple neural network functions, e.g., multi-layer perceptrons (MLPs). We use one linear layer as $g_\theta$ and an identity function as $f_\phi$. In the high level, $R(*)$ receives pair-wise features that represents region-wise local interaction, if any, and integrate them to form global relation.
We introduced how to model global relation from region-wise interactions via a relation module. We then elaborate on how we incorporate extracted relation representations $R(f_{st})$ into our prediction framework.
\textbf{Trajectory Encoding} The past observation of pedestrian's bounding box implicitly captures the motion dynamics in relative distance to the surroundings, which is informative to relation learning. Thus, We employ a recurrent encoder to extract motion dynamics of a target pedestrian. The encoder consists of gated recurrent union (GRU) with hidden size 256 that generates motion hidden states $q_i$ from i-th pedestrian's trajectory $X_i$, as shown in the lower part of Fig. \ref{fig:RATP-detailed-arch}.
The relation features $R(f_{st})$ are agent-agnostic representations. Given the same observation frames $I$ with multiple pedestrians and traffic objects, the intentions of different pedestrians can be influenced by different relation context. To learn agent-specific relation context for each target pedestrian, $R(f_{st})$ should be conditioned on pedestrian's historical motion dynamics. Hence, we condition relation learning on $q_i$ for $i$-th pedestrian and use the trajectory-conditioned relation features $R(f_{st}, q_i)$ to predict intent, as shown in the relation extraction block in {Fig.~\ref{fig:RATP-detailed-arch}}.
Specifically, we concatenate $q_i$ to every concatenated pair of $f_{st}^m, f_{st}^n$ and pass triplets $(f_{st}^m, f_{st}^n, q_i)$ to $g_\theta$ and $f_\phi$ to obtain $R(f_{st}, q_i)$. We append a MLP as classifier to estimate the intent probability $\hat{y}_i$ given relation features $R(f_{st}, q_i)$.
\textbf{Learning Objectives} We employ the binary cross-entropy objective for intent estimation $\hat{y}_i$ and ground-truth intent $y_i$ for each pedestrian. The loss function can be expressed as following:
\begin{equation}
\mathcal{L} = - \frac{1}{N} \sum_{i=1}^{N} (y_i\log{\hat{y}_i} + (1-y_i)\log({1-\hat{y}_i}))
\end{equation}
where $N$ denotes the total number of training pedestrians.
\section{Experiments}
\label{sec:typestyle}
In this section, we empirically evaluate our intent prediction framework on two public datasets: Joint Attention for Autonomous Driving (JAAD) \cite{kotseruba2016joint} and Pedestrian Intention Estimation (PIE) \cite{rasouli2019pie} that contain scenarios where pedestrians approaching the road and might cross in the future. We provide a comparative study to demonstrate the efficacy of our approach and an ablation discussion on relation extraction.
\def1{1}
\begin{table*}[th]
\caption{Results comparison of our approach with other methods on JAAD and PIE datasets.}
\begin{center}
\begin{tabular}{|c|ccccc|ccccc|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{5}{c|}{PIE} & \multicolumn{5}{c|}{JAAD} \\ \cline{2-11}
& Acc & AUC &F1 &P & R & Acc & AUC &F1 &P & R
\\ \hline
SF-GRU\cite{rasouli2020pedestrian} & 0.82 & 0.79 & 0.69 & 0.67 & 0.70 & 0.84 & 0.84 & 0.65 & 0.54 & \textbf{0.84}\\ \cline{1-11}
ARN\cite{yao2021coupling} & \textbf{0.84} & \textbf{0.88} & \textbf{0.90} & \textbf{0.96} & 0.85 & \textbf{0.87} & \textbf{0.92} & 0.70 & 0.66 & 0.75\\ \cline{1-11}
Ours (w/o relation) & 0.82 & 0.81 & 0.84 & 0.89 & 0.80 & 0.81 & 0.84 & 0.70 & 0.63 & 0.82\\ \cline{1-11}
Ours & \textbf{0.84} & 0.85 & \textbf{0.90} & 0.94 & \textbf{0.86} & 0.84 & 0.89 & \textbf{0.74} & \textbf{0.67} & \textbf{0.84}\\ \hline
\end{tabular}
\end{center}
\label{Tab:T1}
\end{table*}
\subsection{Datasets and Experimental Setting}
\textbf{Datasets}
The two datasets contain video sequences collected in urban driving scenes. JAAD contains 323 video clips with total 2786 pedestrians. PIE contains long videos from 6 sets covering 6 hours of driving footage with total 1842 pedestrians. There are 2100 not-crossing pedestrians and 686 crossing pedestrians in the JAAD dataset. PIE contains 1322 non-crossing pedestrians and 512 crossing ones. The pedestrian tracks in both datasets are annotated at 30Hz.
\textbf{Experimental Setting}
Throughout our experiments, we follow the same data preprocessing procedure and evaluation protocol as previous work \cite{rasouli2019pie, rasouli2020pedestrian}. We use the standard data splits of JAAD as in \cite{rasouli2018s} where the dataset is divided into 177 training clips, 29 validation clips and 117 test clips, respectively. The pedestrian tracks are divided into 1355, 202 and 1023 for train/val/test splits in the JAAD data. For the PIE dataset, we follow the same data splits defined in \cite{rasouli2019pie}: videos from set01, set02 and set06 are used for training, set04 and set05 for validation and set03 for test. The number of pedestrian tracks in PIE is 880, 243 and 719 in training, validation and test sets.
To predict the crossing probability of a pedestrian, we subsample each pedestrian track into multiple observation sequences with fixed length of $\tau=$16 frames (0.5s observation). The sample overlap ratio is set to 0.6 for PIE and 0.8 for JAAD dataset. The last frame of each subsampled observation sequence falls between 30 and 60 frames (or 1s -- 2s) prior to the annotated starting point of a crossing event, as illustrated in Fig. \ref{fig:Setting-subsample}. Such time-to-event (TTE) setting allows prediction beforehand as some time should be allotted for emergency maneuver.
\begin{figure}[thb]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=4.5cm]{figures/Setting_subsample.pdf}}
\end{minipage}
\caption{Subsampling for pedestrian crossing prediction: sample observation frames prior to the starting point of crossing. We subsample the observation data by varying time-to-event.}
\label{fig:Setting-subsample}
\end{figure}
\textbf{Evaluation Metrics.}
To evaluate the performance of our prediction framework, we use several binary classification metrics: accuracy, area under the curve (AUC), precision, recall and F1-score.
\subsection{Results}
We compare our method with two other state-of-the-art prediction methods as following:
\begin{enumerate}
\setlength{\itemsep}{0pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{0pt}
\item \textit{SF-GRU}\cite{rasouli2020pedestrian}: a framework that hierarchically fuses features from trajectories, poses and traffic scenes within a stacked recurrent model.
\item \textit{ARN}\cite{yao2021coupling}: a framework that extracts object-wise interaction from object location and visual states.
\end{enumerate}
The evaluation results of state-of-the-art methods and ours are shown in Table \ref{Tab:T1}. Our method outperforms SF-GRU model in several metrics, and achieves similar or superior performance in recall and F1-score on both datasets. Specifically, our model results in 0.04 increase in F1-score on JAAD dataset and 0.01 increase in recall on PIE dataset.
The improvement in recall and F1-score implies lower ratio of missed crossing predictions. Similar trend can be observed when comparing the fusion model SF-GRU and the interaction model ARN. This indicates that incorporating relation learning into prediction framework helps to alarm the intelligent system about potential risks. Our method yields lower acc and AUC on JAAD. The pedestrians in JAAD are close to the curbside with similar trajectories while less interaction with traffic infrastructures. The relation representations $R(f_{st}, q_i)$ on JAAD is majorly influenced by the less discriminative $q_i$. Hence, there are fewer True Negatives, leading to the drop of acc and AUC.
To investigate the effect of relation extraction within our framework, we conduct an ablation experiment by removing the relation extraction. We employ a model that consists of the same visual encoder, the same trajectory encoder and the same final classifier on the same inputs. However, we replace the relation extraction ($g_\theta$ and $f_\phi$) with an MLP with same number of layers. This MLP connects to the concatenated features of full feature maps $f_{st}$ and $q_i$, instead of triplets of vectors $(f_{st}^m, f_{st}^n, q_i)$. Thus, there are more parameters in this model. We include the results of this non-relation model in the third row of the Table\ref{Tab:T1} as Ours (w/o relation). The improved performance by using relation extraction confirms its contribution in the estimation of pedestrian crossing intent.
\section{Conclusion}
In this work, we develop a framework with trajectory-conditioned relation learning to predict pedestrian crossing behavior. A relation extraction module is developed to learn relations between a target pedestrian and surroundings from region-wise interaction through our framework. We evaluate our method on two public intent prediction datasets and compare against state-of-the-art approaches. The experimental results demonstrate that our framework helps to inform potential risks during crossing events with 0.04 increase in F1-score on JAAD dataset and 0.01 increase in recall on PIE dataset. We further conduct ablation experiments to verify the contribution of the relation extraction in our framework.
|
\section{Introduction}
Problems of nonlinear approximation of functions are in high demand in computational practice for many applied problems. Rational approximations are widely used in various variants \cite{braess1986nonlinear}.
Remez algorithm \cite{iske2018approximation} and recently developed algorithms for rational approximation \cite{nakatsukasa2018aaa,hofreither2021algorithm} are used to find parameters of the approximating function.
Much attention is paid to approximation by the sum of exponents in function approximation.
In particular, achievements in this area are reflected by the works of \cite{holmstrom2002review,hristov2022prony}.
In many cases, the approximation of the function $f(x)$ is
\[
f(x) \approx \sum_{i=1}^{m} u_i \varphi(x, v_i)
\]
with the known functional dependence $\varphi(x, v)$.
Some restrictions may be imposed on the approximation parameters to be sought.
A typical example with non-negative coefficients $u_i, \ i = 1,2, \dots, m$ is interesting for many applications.
The theory \cite{mhaskar2000fundamentals} and computational practice \cite{meinardus2012approximation,trefethen2019approximation} of function approximation are well developed in the linear case.
In this the function $f(x)$ is approximated by a given set of trial functions $\varphi_i(x), \ i = 1,2, \dots, m$:
\[
f(x) \approx \sum_{i=1}^{m} u_i \varphi_i(x) .
\]
Optimal approximations are constructed in Hilbert spaces using the least squares method \cite{lawson1995solving,bjorck1996numerical}.
Special note that computational algorithms have long been well developed to consider the constraints $u_i > 0, \ i = 1,2, \dots, m$.
We propose a heuristic algorithm for solving the nonlinear function approximation problem.
It is based on expanding the set of test functions and their subsequent selection during the iterative solution of the nonlinear least squares problem.
The paper is organized as follows.
In Section 2, the problem of nonlinear approximation of functions is posed.
The proposed computational algorithm is described in Section 3.
Section 4 presents results on rational approximation and approximation by the sum of exponents of functions $x^{-\alpha}, \ x \geq 1$ and $\exp(- x^{\alpha} ), \ x \geq 0$ at $0 < \alpha < 1$.
The work results are summarized in Section 5.
\section{Problem formulation}
We consider the problem of nonlinear approximation of one-dimensional function $f(x)$ on the interval $[a, b]$.
In the Hilbert space $L_2([a,b], \varrho (x))$ with weight $\varrho (x) > 0$ the scalar product and norm are defined as follows
\[
(g, q) = \int_{a}^{b} \varrho (x) \big (g(x) - q(x) \big )^2 d x ,
\quad \| g \| = (g , g )^{1/2} .
\]
The function $f(x)$ is approximated by the function $r(x, \bm u, \bm v)$ with two numerical parameter sets $\bm u = \{u_1, u_2, \ldots, u_m\}$, $\bm v = \{v_1, v_2, \ldots, v_m\}$.
Let's assume that the approximating function has the form
\begin{equation}\label{1}
r(x, \bm u, \bm v) = \sum_{i=1}^{m} u_i \varphi(x, v_i) .
\end{equation}
In the representation (\ref{1}) we have isolated the linear coefficients $u_i, \ i =1,2, \ldots, m$, and the dependence on $v_i, \ i =1,2, \ldots, m,$ is determined by the given function $\varphi(x,v)$.
Among the methods of nonlinear approximation of functions, rational approximation and approximation by the sum of exponents are the most widespread.
In particular, in the case of rational approximation at $a \geq 0$ we use the parametric function
\begin{equation}\label{2}
\varphi(x, v) = \frac{1}{1 + v x} .
\end{equation}
When approximated by the sum of the exponents, we have
\begin{equation}\label{3}
\varphi(x, v) = \exp(- v x) .
\end{equation}
In the nonlinear approximation problem we consider, $\bm u, \bm v$ are subject to the following restrictions: the parameters $u_i, \ i =1,2, \ldots, m,$ are non-negative, and $v_i, \ i =1,2, \ldots, m,$ are chosen from the interval $[c, d]$. We come to the problem
\begin{equation}\label{4}
J(\bm u, \bm v) \rightarrow \min,
\quad (\bm u, \bm v) \in K,
\end{equation}
where
\begin{equation}\label{5}
J(\bm u, \bm v) = \|f(x) - r(x, \bm u, \bm v)\|^2,
\quad K = \{ (\bm u, \bm v) \ | \ u_i > 0, \ v_i \in [c, d], \ i =1,2, \ldots, m \} .
\end{equation}
An approximate solution to the functional minimization problem (\ref{1}), (\ref{4}), (\ref{5}) is constructed by setting the functions $f(x), \ r(x, \bm u, \bm v)$ on a sufficiently detailed set of points on the interval $[a,b]$.
\section{Nonlinear approximation algorithm}
We begin by dividing the interval $[a,b]$ into $n$ partial intervals of length $h_j, \ j = 1,2, \ldots, n$, so that
\[
b - a = \sum_{j=1}^{n} h_j.
\]
We denote the centers of the intervals $h_j$ by $x_j, \ j = 1,2, \ldots, n$.
Using the quadrature formula of rectangles, we compare the original problem (\ref{4}), (\ref{5}) with the minimization problem
\begin{equation}\label{6}
J^h(\bm u, \bm v) \rightarrow \min,
\quad (\bm u, \bm v) \in K,
\end{equation}
\begin{equation}\label{7}
J^h(\bm u, \bm v) = \sum_{j=1}^{n} \varrho(x_j) \big (f(x_j) - r(x_j, \bm u, \bm v) \big )^2 h_j .
\end{equation}
To simplify the approximation problem, the coefficients $v_i, \ i =1,2, \ldots, m,$ will not be evaluated over the whole interval $[c, d]$, but only at given points
\[
\widetilde{v}_k \in V_l,
\quad V_l = \{c \leq \widetilde{v_1} \leq \widetilde{v_2} \leq \ldots \leq \widetilde{v_l} \leq d \}
\]
for a sufficiently fine partition ($l \gg m$).
Thus we take instead of $K$ the set of constraints in the form
\[
\widetilde{K} = \{ (\bm u, \bm v) \ | \ u_i > 0, \ v_i \in V_l, \ i =1,2, \ldots, m \} .
\]
Given (\ref{1}), we will come from the problem (\ref{6}), (\ref{7}) to the problem
\begin{equation}\label{8}
J^h(\bm u, \bm v) \rightarrow \min,
\quad (\bm u, \bm v) \in \widetilde{K},
\end{equation}
\begin{equation}\label{9}
J^h(\bm u, \bm v) = \sum_{j=1}^{n} \varrho(x_j) h_j \Big ( \sum_{i=1}^{m} u_i \varphi(x_j, v_i) - f(x_j) \Big )^2 .
\end{equation}
Instead of the vector $\bm u$ with components $u_i, \ i = 1,2, \ldots, m,$ we introduce a vector $\widetilde{\bm u}$ of larger dimension with components $\widetilde{u}_k, \ k = 1,2, \ldots, l$. The components of the vector $\widetilde{\bm u}$ we define from the approximation condition
\begin{equation}\label{10}
\sum_{i=1}^{m} u_i \varphi(x_j, v_i) = \sum_{k=1}^{l} \widetilde{u}_k \varphi(x_j, \widetilde{v}_k) .
\end{equation}
By doing so, we put $\widetilde{u}_k = u_i$ if $\widetilde{v}_k = v_i$ and $\widetilde{u}_k = 0$ if $\widetilde{v}_k \neq v_i$.
Considering (\ref{10}), we proceed from the minimization problem (\ref{8}), (\ref{9}) to the problem
\begin{equation}\label{11}
\widetilde{J} (\widetilde{\bm u}) \rightarrow \min,
\quad \widetilde{u}_k \geq 0,
\quad k = 1,2, \ldots, l ,
\end{equation}
\begin{equation}\label{12}
\widetilde{J} (\widetilde{\bm u}) = \sum_{j=1}^{n} \varrho(x_j) h_j \Big ( \sum_{k=1}^{l} \widetilde{u}_k \varphi(x_j, \widetilde{v}_k) - f(x_j) \Big )^2 .
\end{equation}
As a result, we obtained a constrained (non-negative) least squares problem of a larger dimension to determine the parameters $\widetilde{u}_k, \ k = 1,2, \ldots, l,$ in the linear representation of the approximating function.
Computational algorithms for the minimization problem (\ref{11}), (\ref{12}) are well studied \cite{bjorck1996numerical,lawson1995solving}.
In computational practice, the most widely used is the NNLS (Non-Negative Least Squares) algorithm described in detail in \cite{lawson1995solving}.
Given the specifics of the problem (\ref{11}), (\ref{12}), we will separately specify variants of NNLS algorithms for large-scale problems (see, for example, \cite{kim2013non,myre2017tnt}).
The standard NNLS algorithm is a two-step iterative method with main and inner loop iterations.
The number of positive coefficients is initially set to zero.
As the number of iterations increases, the residual decreases, though not monotonically.
First, decreasing the residual is provided by increasing the number of positive coefficients.
Taking into account the above features of the iterative process of NNLS algorithm, we can propose the following strategy of coefficient selection in a nonlinear approximation of functions based on the minimization problem (\ref{11}), (\ref{12}).
We perform a sufficiently large number of iterations of the NNLS algorithm.
At each iteration, we control the residual and the number $\widetilde{m}$ of positive coefficients $\widetilde{u}_k, \ k = 1,2, \ldots, l$.
The number of iterations of the NNLS algorithm is chosen such that $\widetilde{m} = m$ and the residual is minimal.
No additional modifications of the standard algorithm are performed.
The computational implementation is based on the non-negative least squares solver from the SciPy library \cite{2020SciPy-NMeth} (module \textsf{optimize}, function \textsf{nnls}).
\section{Numerical experiments}
Let us illustrate the possibility of constructing nonlinear approximations of functions with two examples.
First, we construct rational approximations of the function $x^{-\alpha}, 0 < \alpha < 1$ at $x \geq 1$.
In the last decade, such a problem has been actively discussed in the literature in connection with solving boundary value problems with fractional power elliptic operators (see, for example, \cite{bonito2018numerical,harizanov2020rev}), and also when considering more general problems with \cite{vabishchevich2022some} operator functions.
The second example concerns the approximation of the function $\exp(-x^{-\alpha}), 0 < \alpha < 1$ at $x \geq 0$.
In this case, we use approximations by the sum of exponents.
Such problems are typical in approximate solutions of nonstationary problems with memory when the difference kernel of the integral term \cite{vabMemory} is approximated.
\subsection{Approximation of $x^{-\alpha}$}
We will approximate the function $f(x) = x^{-\alpha}$ when $a = 1$ and $b$ is large enough.
We used $b = 10^{15}$ in the following calculations.
In applied problems, it is often important to impose an additional restriction on the approximating function
\begin{equation}\label{13}
r(a, \bm u, \bm v) = f(a) .
\end{equation}
In the class of rational approximations of the type (\ref{2}), given (\ref{13}), for the approximation $x^{-\alpha}$, we obtain the representation
\[
x^{-\alpha} \approx 1 + \sum_{i=1}^{m} u_i \varphi(x, v_i) ,
\]
when
\[
\varphi(x, v_i) = \frac{1}{1 + v_i x} - \frac{1}{1 + v_i},
\quad i = 1,2, \ldots, m .
\]
The choice of the weight function $\varrho (x)$, which allows controlling the approximation accuracy in some parts of the interval $[a,b]$, requires special attention.
In our case, the approximated function decreases to zero as $x$ increases.
A more significant influence of points at small $x$ is provided by setting a decreasing function $\varrho (x)$.
To partition the interval $[a,b]$ we use partial intervals of increasing length ($h_{i+1} > h_i, \ i = 1,2, \ldots, m-1$).
This approach is formalized by introducing a new variable $\theta$ instead of $x$.
Put $x = \exp(\theta)$, so that $\theta \in [0, \beta]$ ($\exp(\beta) = b$).
For the residual functional, we get
\[
J(\bm u, \bm v) = \int_{0}^{\beta }\varrho \big(\exp(\theta) \big) \exp(\theta) \Big (\exp(-\alpha\theta) - 1
- \sum_{i=1}^{m} u_i \varphi \big(\exp(\theta), v_i \big) \Big )^2 d \theta .
\]
The following is the computational data when setting $\varrho \big(\exp(\theta) \big) = \exp(-\theta)$ ($\varrho(x) = x^{-1}$).
At sufficiently large values of the number of points $n$ on the interval $[a,b]$, the approximation accuracy changes insignificantly.
In the presented results of calculations, we limited ourselves to the case $n=5000$.
Of greater interest are the calculated data at different partitioning of the interval $[c,d]$ of permissible values of the parameter $v$.
Figure~\ref{fig-1} illustrates the effect of $l$ in (\ref{12}) when we approximate the function $x^{-\alpha}$ with $\alpha = 0.5$.
The residual equal to $\widetilde{J}^{1/2} (\widetilde{\bm u})$ using different numbers of iterations of the non-negative least squares method is shown in the left-hand side of the figure. We observe a reasonably fast, generally speaking, non-monotonic decrease in the residual with the increasing number of iterations.
This residual is achieved with varying numbers of non-zero coefficients $m$ with a general tendency of $m$ increasing as the number of iterations of the non-negative least squares method increases. From these data, we estimate the number of iterations to achieve the minimum norm of residual for a given number $m$ of terms of the rational approximation (\ref{10}).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{1-1}
\includegraphics[width=0.49\linewidth]{1-2}
\caption{Residual (left) and the number of non-zero elements $m$ (right) for $\widetilde{u}_k, \ k = 1,2, \ldots, l,$ in individual iterations of the non-negative least squares method.}
\label{fig-1}
\end{figure}
The effect of interval partitioning $[c,d]$ on the approximation accuracy is shown in Fig.\ref{fig-2}.
Here is the data for $m=10$ at $l = 500, 1000, 2000$.
The approximation accuracy is estimated by the value
\[
\varepsilon (x_j) = \Big | \sum_{i=1}^{m} u_i \varphi(x_j, v_i) - f(x_j) \Big |,
\quad j = 1,2, \ldots, n .
\]
We observe a remarkable similarity in determining the approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right-hand side of the figure) and in the accuracy of the approximation function $x^{-\alpha}$ with $\alpha=0.5$ at $x \in [1, 10^{15}]$ (left-hand side of the figure).
With this in mind, we fixed the number of partitions $l=1000$ in our calculations.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{2-1}
\includegraphics[width=0.49\linewidth]{2-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) at different partitioning of the interval $[c,d]$.}
\label{fig-2}
\end{figure}
Figure~\ref{fig-3} shows the approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ when given $m = 5, 10, 20$.
There is a significant increase in the approximation accuracy with increasing $m$.
For this variant, the number of iterations is $13, 38$, and $110$, respectively.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{3-1}
\includegraphics[width=0.49\linewidth]{3-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) for the function $x^{-\alpha}$ with $\alpha = 0.5$ at various $m$.}
\label{fig-3}
\end{figure}
When approximating the function $x^{-\alpha}$, special attention is paid to the influence of the parameter $\alpha$.
The approximation accuracy for $m = 5, 10, 20$ at $\alpha = 0.25$ is shown in Figure \ref{fig-4}.
Similar results at $\alpha = 0.75$ are presented in Figure \ref{fig-5}.
As with other computational algorithms of rational approximation \cite{bonito2018numerical,harizanov2020rev}, increasing $\alpha$ increases the accuracy.
Note also that the calculation of the coefficient $u_i, v_i, \ i = 1,2, \ldots, m,$ at larger $\alpha$ is performed with a more significant number of iterations of the non-negative least squares method.
Table 1 shows the calculated values of the approximation parameter $u_i, v_i, \ i = 1,2, \ldots, 10,$ ($m = 10$) when approximating the function $x^{-\alpha}$ for $\alpha = 0.25, 0.5, 0.75$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{4-1}
\includegraphics[width=0.49\linewidth]{4-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) for the function $x^{-\alpha}$ with $\alpha = 0.25$ at various $m$.}
\label{fig-4}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{5-1}
\includegraphics[width=0.49\linewidth]{5-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) for the function $x^{-\alpha}$ with $\alpha = 0.75$ at various $m$.}
\label{fig-5}
\end{figure}
\begin{center}
\begin{table}[htp]
\label{tab-1}
\caption{Parameters approximation with $m=10$ for $x^{-\alpha}$ }
\centering
\begin{tabular}{l c ll c ll c ll}
\hline
&& \multicolumn{2}{l}{$\alpha = $ 0.25} && \multicolumn{2}{l}{$\alpha = $ 0.5} && \multicolumn{2}{l}{$\alpha = $ 0.75}\\
\cline{3-4} \cline{6-7} \cline{9-10}
$i$ && $u_i$ & $v_i$ && $u_i$ & $v_i$ && $u_i$ & $v_i$ \\
\hline
1 && 1.060084e-03 & 2.115485e-13 && 1.263660e-04 & 5.816049e-09 && 1.653295e-06 & 1.135126e-08 \\
2 && 2.778250e-03 & 6.526663e-11 && 1.318851e-04 & 6.336196e-08 && 1.664949e-05 & 6.273950e-07 \\
3 && 7.184790e-03 & 3.607348e-09 && 2.177478e-03 & 2.389865e-06 && 2.008706e-04 & 1.954833e-05 \\
4 && 1.608844e-02 & 1.812161e-07 && 1.423375e-02 & 1.453310e-04 && 9.792299e-04 & 2.343140e-04 \\
5 && 2.879614e-02 & 3.853128e-06 && 3.605113e-02 & 3.090116e-03 && 7.011612e-03 & 2.808580e-03 \\
6 && 6.751752e-02 & 8.192757e-05 && 4.002657e-02 & 9.723689e-03 && 3.444878e-02 & 3.059759e-02 \\
7 && 1.117978e-01 & 1.439033e-03 && 7.561481e-02 & 4.933185e-02 && 9.142663e-03 & 6.570394e-02 \\
8 && 2.518764e-01 & 2.088023e-02 && 2.411886e-01 & 1.552328e-01 && 2.280614e-01 & 3.333403e-01 \\
9 && 3.723954e-01 & 3.667548e-01 && 3.672604e-01 & 1.397038e+00 && 6.238136e+00 & 8.579865e+00 \\
10 && 6.229275e-01 & 1.537079e+00 && 2.193928e+00 & 3.631519e+00 && 2.478274e+00 & 1.842403e+01 \\
\hline
\end{tabular}
\end{table}
\end{center}
\clearpage
\subsection{Approximation of $\exp(- x^{\alpha} )$}
The second example concerns the approximation of the stretching exponential function:
\[
f(x) = \exp(- x^{\alpha} ),
\quad 0 < \alpha < 1 .
\]
We will consider the case where the approximation is made with $a = 0, \ b = 10^3$, and an additional constraint (\ref{13}).
When approximating by the sum of exponents, we have
\[
\exp(- t^{\alpha}) \approx 1 + \sum_{i=1}^{m} u_i \varphi(x, v_i) ,
\]
where
\[
\varphi(x, v_i) = \exp( - v_i x) - 1,
\quad i = 1,2, \ldots, m .
\]
The choice of the weight function $\varrho(x)$ is done similarly to how it was when approximating the function $x^{-\alpha}$.
A uniform grid on the new variable $\theta \in [0, \beta]$ is introduced, with $x = \exp(\theta)-1$ ($\exp(\beta) - 1 = b$).
The calculations are performed using $\varrho(x) = (1+x)^{-1}$.
For the base variant the following parameters are chosen: $\alpha = 0.5$, $c = 10^{-4}$, $d = 10^{4}$, $n = 5000$, $l = 1000$, $m = 10$.
Figure~\ref{fig-6} shows the residual and the number of non-zero elements $m$ when using different numbers of iterations of the NNLS method.
The no significant influence of the interval partition detail of $[c, d]$ on the approximation accuracy and values of approximation parameters is illustrated by Fig.\ref{fig-7}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{6-1}
\includegraphics[width=0.49\linewidth]{6-2}
\caption{Residual (left) and number of nonzero elements $m$ (right) for $\widetilde{u}_k, \ k = 1,2, \ldots, l,$ in separate iterations of the NNLS method when approximating the function $\exp(- x^{\alpha} )$.}
\label{fig-6}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{7-1}
\includegraphics[width=0.49\linewidth]{7-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) at different partitioning of the interval $[c,d]$ when approximating $\exp(- x^{\alpha} )$.}
\label{fig-7}
\end{figure}
The increase in approximation accuracy with increasing $m$ is illustrated by Fig.\ref{fig-8}; the figure also shows the approximation parameters
$u_i, v_i, \ i = 1,2, \ldots, m$. Similar data when setting $m = 5, 10, 20$ for the approximation of the function $\exp(- x^{\alpha} )$ at
$\alpha = 0.25$ and $\alpha = 0.75$ are shown in Figures \ref{fig-9},\ref{fig-10}, respectively.
When the parameter $\alpha$ is reduced, the approximation accuracy decreases.
The obtained approximation parameters $u_i, v_i, \ i = 1,2, \ldots, 10,$ ($m = 10$) for approximation of function $\exp(- x^{\alpha} )$ for $\alpha = 0.25, 0.5, 0.75$ are presented in Table 2.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{8-1}
\includegraphics[width=0.49\linewidth]{8-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) for the function $\exp(- x^{\alpha} )$ with $\alpha = 0.5$ at various $m$.}
\label{fig-8}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{9-1}
\includegraphics[width=0.49\linewidth]{9-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) for the function $\exp(- x^{\alpha} )$ with $\alpha = 0.25$ at various $m$.}
\label{fig-9}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.49\linewidth]{10-1}
\includegraphics[width=0.49\linewidth]{10-2}
\caption{Approximation accuracy $\varepsilon$ (left) and approximation parameters $u_i, v_i, \ i = 1,2, \ldots, m,$ (right) for the function $\exp(- x^{\alpha} )$ with $\alpha = 0.75$ at various $m$.}
\label{fig-10}
\end{figure}
\begin{center}
\begin{table}[htp]
\label{tab-2}
\caption{Parameters approximation with $m=10$ for $\exp(- x^{\alpha} )$}
\centering
\begin{tabular}{l c ll c ll c ll}
\hline
&& \multicolumn{2}{l}{$\alpha = $ 0.25} && \multicolumn{2}{l}{$\alpha = $ 0.5} && \multicolumn{2}{l}{$\alpha = $ 0.75}\\
\cline{3-4} \cline{6-7} \cline{9-10}
$i$ && $u_i$ & $v_i$ && $u_i$ & $v_i$ && $u_i$ & $v_i$ \\
\hline
1 && 3.684368e-02 & 4.361538e-03 && 8.918599e-03 & 4.939622e-02 && 1.204240e-01 & 3.772042e-01 \\
2 && 4.849511e-02 & 1.282650e-02 && 6.127055e-02 & 1.064209e-01 && 3.399225e-01 & 6.078323e-01 \\
3 && 1.017103e-01 & 4.546295e-02 && 2.567263e-01 & 2.821308e-01 && 2.922859e-01 & 1.180517e+00 \\
4 && 1.241971e-01 & 1.714882e-01 && 2.731277e-01 & 9.794697e-01 && 5.133306e-02 & 2.024447e+00 \\
5 && 1.319054e-01 & 6.742622e-01 && 1.406392e-01 & 2.821308e+00 && 1.093081e-01 & 3.400412e+00 \\
6 && 6.933268e-02 & 1.942175e+00 && 1.200802e-01 & 8.296959e+00 && 4.550720e-02 & 8.648423e+00 \\
7 && 1.337784e-01 & 5.831305e+00 && 4.969377e-02 & 2.491130e+01 && 2.014393e-02 & 2.066880e+01 \\
8 && 1.251127e-01 & 4.098384e+01 && 4.906249e-02 & 7.959777e+01 && 1.329286e-02 & 5.479472e+01 \\
9 && 8.343032e-02 & 2.543346e+02 && 2.565896e-02 & 4.452959e+02 && 6.492292e-03 & 2.940820e+02 \\
10 && 1.409798e-01 & 4.184289e+04 && 1.487512e-02 & 3.471687e+04 && 1.283043e-03 & 3.400412e+04 \\
\hline
\end{tabular}
\end{table}
\end{center}
\clearpage
\section{Conclusions}
\begin{enumerate}
\item The problem of nonlinear approximation of functions based on two sets of parameters is considered.
A particular term of the approximating function is the product of the first non-negative unknown parameter by a given nonlinear function that depends on the second unknown parameter.
\item We present a heuristic computational algorithm for the nonlinear approximation of functions based on the minimization of the residual functional by values of the approximated function at separate points.
The unknown nonlinear approximation parameters are set on many points of the interval of permissible values.
Linear non-negative parameters are determined using the non-negative least squares method.
We find the solution of the nonlinear function approximation problem based on the estimation of the residual and the number of the first non-negative unknown parameters at each iteration of the non-negative least squares method.
\item The main points of practical use of the computational algorithm are illustrated by two typical problems of nonlinear approximation of functions. In the first example, we construct a rational approximation of the function $x^{-\alpha}, \ 0 < \alpha < 1$ at $x \geq 1$.
The second example is the approximation of the function $\exp(- x^{\alpha} ), \ \quad 0 < \alpha < 1$ at $ x \geq 0$ by the sum of exponents.
\end{enumerate}
|
\section{Introduction}
In transonic flow on wings of transport aircraft a local supersonic region is formed that is terminated by a shock wave. The buffet phenomenon is characterized by a self-sustained large-scale oscillation of this shock wave at low frequencies in the range of $Sr \sim \mathcal{O}(10^{-2}) - \mathcal{O}(10^{-1})$ where $Sr =\frac{fc}{u_\infty}$ is the non-dimensional frequency denoted as Strouhal number. The spatial oscillation of the shock wave leads to periodically alternating forces on the wing structure. The structural response to those dynamic loads, generally referred to as buffeting, have detrimental effects on the structural integrity and overall flight safety. Therefore, it defines a strict boundary to an aircraft's flight envelope. Economical and ecological challenges require aviation to make ever more efficient use of the available resources. A thorough knowledge of the aeroelastic behavior of aircraft is crucial in the endeavor to further reduce structural weight and optimize fuel consumption. Therefore, a comprehensive understanding of the buffet phenomenon is substantially important in the design process of future aircraft.
Although the buffet phenomenon has been intensively investigated for decades a comprehensive description of the underlying mechanisms is still subject to debate. The most widely accepted model explaining the self-sustained nature of the low-frequency shock oscillation has been introduced by Lee~\cite{lee2001}. The proposed mechanism is based on a closed feedback loop between the shock wave and the trailing edge flow. The interaction of the incoming boundary layer with the shock wave leads to the generation of vortices which convect downstream and eventually pass over the trailing edge. The transition from a wall-bounded flow to a free-shear layer causes the generation of pressure waves which propagate back upstream. The pressure waves interact with the shock wave and push the shock upstream leading to a thickening and ultimately a separation of the boundary layer downstream of the shock. The separation of the boundary layer mitigates or massively reduces the local wall-shear stress near the trailing edge flow. Thus, also the generation of pressure waves at the trailing edge is attenuated due to the decreased Lamb vector, i.e., the outer product of vorticity and the velocity vectors, and the shock moves back downstream closing the feedback loop. Good agreement with the buffet model proposed by Lee was found, among others, by Deck~\cite{deck2005}, Xiao et al.~\cite{xiao2006}, Hartmann et al.~\cite{hartmann2012, hartmann2013a, hartmann2013b} and Feldhusen-Hoffmann et al.~\cite{feldhusen2017, feldhusen2021}.
Lee's model was further developed by Hartmann et al.~\cite{hartmann2013b}, who proposed that the relevant interaction region of the shock wave and the pressure waves originating from the trailing edge is at the upper part of the shock wave instead of the shock foot. This formulation led to a significantly better agreement between the predicted and the measured buffet frequency.
In the pursuit to design ever more efficient aircraft, engine development is clearly moving toward larger and larger bypass ratios, so called ultra-high bypass ratio (UHBR) turbofan engines. The large diameter of such engines requires them to be mounted closely under the wing to ensure sufficient ground clearance without having to massively redesign the landing gear much more generously. As a result, the flow around the wing is significantly affected by the engine flow. In transonic flight, shock waves with all their detrimental effects, i.e., shock-induced separation, shock unsteadiness, and total pressure loss, can occur on the engine nacelle~\cite{dietz2008, spinner2021, spinner2022}. The buffet phenomenon is, however, highly sensitive to the upstream flow conditions, which are strongly altered by an upstream shock wave. Furthermore, the shear layer developing on the engine nacelle merges into the boundary layer on the pressure side of the airfoil. As a result, also the flow in the vicinity of the trailing edge is influenced by the integration of UHBR engines. Since the trailing edge flow is a central element to trigger shock wave oscillations~\cite{lee2001, hartmann2013b} the quality of the buffet phenomenon can be significantly affected.
In recent years, the buffet phenomenon has been studied extensively ~\cite{lee2001, xiao2006, deck2005, hartmann2012, hartmann2013a, hartmann2013b, feldhusen2017, feldhusen2021, giannelis2017, jacquin2009}. Yet, little is known about the interaction of engine induced flow disturbances and the involved dynamics. A recent experimental study by Spinner \& Rudnik \cite{spinner2022} investigates the phenomenon of shock buffet on the lower surface of the wing, which occurs at negative angles of attack due to the integration of the UHBR. To the best of the authors' knowledge, studies on the impact of engine nacelle integration on the buffet phenomenon on the upper wing surface do not exist in the archival literature. Yet, engine nacelle integration is an important aspect of the aerodynamic design of commercial aircraft. This means, it is of great interest to complement the current knowledge by the effect of engine induced flow disturbances on the buffet phenomenon.
In this paper, first studies are made to better understand the interaction between engine-induced disturbances of the flow field and the buffet phenomenon on the suction side of the wing. Using wall-modeled large-eddy simulations (WM-LES) the flow field around the OAT15A airfoil with a generic UHBR engine nacelle geometry and a baseline configuration without the nacelle is computed at buffet conditions. To allow the analysis of the upstream shock and the free-shear layer interacting with the airfoil boundary layer at reasonable computational cost the nacelle is modeled as a 2D-periodic flow-through nacelle. The flow field is analyzed focusing on the disturbance of the flow field due to the integration of the nacelle geometry. The resulting changes of the occurring shock dynamics are examined using the sparsity-promoting dynamic mode decomposition (SP-DMD).
The paper is organized as follows. In section~\ref{sec:methods}, the simulation framework and the numerical methods are briefly presented. In particular, the wall-modeling approach is discussed in subsection~\ref{ssec:wallmodel} and the SP-DMD algorithm is introduced in subsection~\ref{ssec:dmd}. The computational setup including the numerical mesh and the mesh refinement strategy is discussed in section~\ref{sec:setup}. The results are presented in section~\ref{sec:results} with subsection~\ref{ssec:flowfield} focusing on direct insights from the WM-LES. The results obtained performing SP-DMD on the simulation data are discussed in subsection~\ref{ssec:dmdanalysis}. Finally, conclusions are drawn in section~\ref{sec:conclusion}.
\section{Numerical Methods}\label{sec:methods}
All simulations are performed using the Cartesian solver of the m-AIA framework which was formerly denoted zonal flow solver (ZFS)~\cite{lintermann2020}. The governing equations are the compressible, unsteady Navier-Stokes equations. The equations are spatially filtered and solved on an unstructured, hierarchical Cartesian mesh using the finite-volume method. The advective upstream splitting method (AUSM) is used for the convective fluxes. A central difference scheme is employed for the discretization of the viscous fluxes. For the temporal integration, a second-order accurate 5-stage Runge-Kutta method is used. Small scale turbulence is accounted for by the approach of monotone integrated LES (MILES)~\cite{boris1992} such that no additional turbulence model is necessary~\cite{meinke2002}. Boundaries of embedded bodies are realized using a cut-cell method~\cite{schneiders2016}. Further details on the numerical methods used by the m-AIA simulation framework are given in~\cite{meinke2002, schneiders2016}.
\subsection{Wall-stress model}
\label{ssec:wallmodel}
\begin{figure}[htb]
\centering
\begin{tikzpicture}[scale=1.0]
\path [use as bounding box] (0.0,0.0) rectangle (8.0, 4.75);
\coordinate (S) at (0.0, 1.0);
\coordinate (E) at (8.0, 0.0);
\coordinate (VE) at (1.9, 3.375);
\coordinate (VS) at (1.9, 0.5);
\coordinate (VM) at (3.4, 3.375);
\draw[step = 0.75, thin, lightgray] (0.01, 0.01) grid (8.0, 4.75);
\draw[name path = vert, black, thick] (VS) -- (VE);
\draw[fill, white] (S) to [bend left = 10] (E);
\draw[name path = body, black] (S) to[bend left = 10] (E);
\draw[fill, white] (0,0) -- (S) -- (E);
\path[name intersections={of=body and vert,by=VS}];
\draw[black, thick, name path = profile] (VS) to[out=0 , in=-100] (VM);
\foreach \i in {1,...,4} {
\coordinate (AI) at ($(VS) + ($0.25*\i*($(VE) - (VS)$)$)$);
\path[name path = hor] (AI) -- ($(AI) + (10.0, 0.0)$);
\path[name intersections={of=hor and profile, by=AE}];
\draw[->, black, very thick] (AI) -- (AE);
}
\def \offset {0.4};
\draw[decoration={brace, amplitude = 5pt}, decorate, very thick] ($(VS)-(\offset, 0.0)$) -- node[left=8pt, rectangle, draw = black, fill=white, rounded corners = 0.1cm, inner sep = 0.2cm] {$h_{wm}$} ($(VE)-(\offset, 0.0)$);
\node[draw, circle, fill=white] (sp) at (AI){$\bm u$};
\node[draw, circle, fill=white] (sf) at (VS){$\bm \tau$};
\node[draw, rounded corners = 0.2cm, fill=white, inner sep = 0.4cm] at (6,2.5) (wm) {\large Wall Model};
\draw[->, very thick] (sp) to[out=45 , in=90] (wm);
\draw[->, very thick] (wm) to[out= -90, in=-45] (sf);
\end{tikzpicture}
\caption{Graphic representation of the wall-model mechanism, by courtesy of~\cite{luerkens2022}.}
\label{fig:wallmnodel}
\end{figure}
Traditional wall-resolved LES of boundary layer flows are restricted by the resolution of the viscous sublayer which requires extremely small cells at high Reynolds numbers. This results in a large amount of mesh cells and thus in a enormous demands on the available computer hardware that often can not be fulfilled. At the same time, the computational time step is restricted by the CFL condition leading to a very slow progress of the simulation in terms of convective time units $CTU = c/u_{\infty}$ making wall-resolved LES of high Reynolds number flows computationally very expensive.
Wall-modeling approaches are a convenient method to circumvent the massive constraints on mesh resolution for traditional LES of boundary layer flows while maintaining a high spatial and temporal resolution. Omitting the resolution of the viscous sublayer and only resolving large scale structures in the boundary layer WM-LES enables not only a significant reduction in the total number of cells but does also allow a substantially larger computational time step. Thereby, LES of dynamic phenomena at low Strouhal numbers $Sr = \frac{f\cdot c}{u_\infty}$ such as the buffet phenomenon can be performed even at high Reynolds numbers with reasonable computational effort. The applicability of such a wall-stress model to buffet flows has been demonstrated by Fukushima \& Kawai~\cite{fukushima2018}.
In this study, an analytical wall-stress model is used to compute the wall-shear stress based on flow information obtained from the outer boundary layer. The function is an implicit single equation expression for the law of the wall introduced by Spalding~\cite{spalding1961}:
\begin{equation}\label{eq:spalding}
y^+ = u^+ + e^{-\kappa B} \left\{ e^{\kappa u^+} - \sum_{n=0}^{4} \frac{(\kappa u^+)^n}{n!} \right\}
\end{equation}
with $u^+ = \frac{u_{\|}}{u_\tau}$, $y^+=\frac{h_{wm}u_\tau}{\nu}$, and $u_\tau = \sqrt{\tau_w / \rho_w}$. The subscript $(\bullet)_{\|}$ denotes projection onto the wall-tangential plane. The von Kármán constant $\kappa$ is set to $0.4$ and $B$ is $5.0$. Equation~\ref{eq:spalding} is evaluated at a sampling distance $y=h_{wm}$ normal to the surface which is illustrated in figure~\ref{fig:wallmnodel}. The implicit expression is solved iteratively for $u_\tau$ using Newton's method. To apply the wall-shear stress to the boundary surface an additional correction loop within the viscous flux computation is performed adding an artificial viscosity $\mu_{wm}$ such that the wall-shear stress given by
\begin{equation}
\left( \mu + \mu_{wm} \right) \frac{\partial u_{\|}}{\partial n} = \tau_{wm}
\end{equation}
is imposed. To prevent an obvious misapplication of the model in regions of separated flow the velocity gradient at the wall is checked for $\partial u_{\|}/\partial{n} < 0$. If flow separation is detected, the application of the wall-stress model is omitted by setting $\mu_{wm} = 0$.
\subsection{Dynamic mode decomposition}
\label{ssec:dmd}
The dynamic mode decomposition~\cite{schmid2010} is a data-driven technique that allows the decomposition of a given data field $\bm{q}(\bm{x},t)$ into a series representation of spatio-temporal modes $\phi_n$. Every mode is characterized by an individual, complex frequency $\lambda_n$, and amplitude $a_n$. The resulting series representation of the data field reads
\begin{align}
\bm{q}(\bm{x},t) = \sum_n a_n e^{\lambda_n t} \bm{\phi}_n(\bm{x}).
\end{align}
For these modes characteristic dynamic phenomena of distinct frequencies can be determined and analyzed in detail. The buffet phenomenon is typically restricted to narrow frequency bands such that DMD can be an effective tool for the in-depth analysis of the isolated dynamics and the underlying physical mechanisms, which is reflected by its frequent use in the literature~\cite{kou2017, kou2018, gao2017, poplingher2019, feldhusen2021}. In general, a DMD performed on a set of $N$ snapshots results in $N-1$ modes, which are given as complex conjugate pairs. For a large number of snapshots this results in an inconveniently large number of modes to assess. The sparsity-promoting DMD algorithm proposed by Jovanović et al. \cite{jovanovic2014} addresses this problem by introducing a penalty term $\gamma\sum_{n=1}^{N-1}\lvert a_n\rvert$ into the minimization problem determining the modes, which reads
\begin{align}
\label{eq:spdmd}
\underset{a}{\text{minimize}} \, \Vert \bm{V}^{N-1}_1 - \bm{\Phi}\bm{D}_a \bm{V}_{and} \Vert_F^2 + \gamma \sum_{n=1}^{N-1} \lvert a_n \rvert \,.
\end{align}
Here, $\bm{V}_1^{N-1}$ is the discrete snapshot sequence and $\bm{\Phi}$ contains the dynamic modes given in matrix form by
\begin{align}
\bm{V}_1^{N-1} = \left[ \bm{v}_1, \bm{v}_2, \ldots, \bm{v}_{N-1}\right], \quad \bm{\Phi} = \left[ \bm{\phi}_1, \bm{\phi}_2, \ldots, \bm{\phi}_{N-1}\right]\, .
\end{align}
The matrix $\bm{D}_a = \text{diag}(a)$ is the diagonal matrix of all optimized amplitudes $a_n$. The Vandermonde matrix $\bm{V}_{and}$ contains the so-called Ritz eigenvalues $\mu_n = e^{\lambda_n \Delta t}$ and reads
\begin{align}
\bm{V}_{and} = \left(
\begin{matrix}
1 & \mu_1 & \ldots & \mu_1^{N-1} \\
1 & \mu_2 & & \mu_2^{N_1} \\
\vdots & \vdots & & \vdots \\
1 & \mu_{N-1} & & \mu_{N-1}^{N-1}
\end{matrix}
\right).
\end{align}
When solving the minimization problem given by expression \ref{eq:spdmd}, increasing values of the penalty parameter $\gamma$ imply a stronger penalization of non-zero amplitudes, and thus result in an overall lower number of non-zero amplitude modes. More details on the SP-DMD algorithm are given in \cite{jovanovic2014}. The application of SP-DMD to transonic buffet is thoroughly discussed in \cite{feldhusen2021}.
\section{Computational Setup}
\label{sec:setup}
\begin{figure}
\begin{subfigure}[t]{0.49\textwidth}
\vskip 0pt
\includegraphics[width=1\linewidth]{pictures/gridRawBlack.png}
\vspace{16.5pt}
\caption{Computational mesh of the nacelle configuration on the coarsest level.}
\label{fig:grid_raw}
\end{subfigure}
\begin{subfigure}[t]{0.49\textwidth}
\vskip 0pt
\begin{overpic}[width=1\linewidth]{pictures/gridlevels_jet.png}
\put(86.5,7){\footnotesize $17$}
\put(71,7){\footnotesize $16$}
\put(55.5,7){\footnotesize $15$}
\put(39.7,7){\footnotesize $14$}
\put(24.4,7){\footnotesize $13$}
\put(8.7,7){\footnotesize $12$}
\put(30, 1){\footnotesize $\text{cell refinement level}$}
\end{overpic}
\caption{Regions of grid refinement around the OAT15A airfoil and the nacelle.}
\label{fig:grid_ref}
\end{subfigure}
\caption{Overview of the computational mesh of the nacelle configuration, by courtesy of~\cite{luerkens2022}.}
\end{figure}
In the following, the computational setup of the nacelle configuration is presented. Note that except for the integration of the nacelle geometry
the baseline configuration, i.e., no engine is taken into account, and the UHBR-airfoil configuration, and the corresponding mesh refinement both setups are identical and therefore no further differentiation is made unless otherwise necessary. A setup with only mild flow separation while still providing low-frequency shock oscillations is considered. Based on the work by Fukushima \& Kawai~\cite{fukushima2018}, Jacquin et al.~\cite{jacquin2009} and Deck~\cite{deck2005} the angle of attack was set to $\alpha = 3.5^\circ$ and the freestream Mach number to $Ma_\infty = 0.73$. In agreement with the associated experimental campaign by Schauerte \& Schreyer~\cite{schauerte2022} the chord based Reynolds number was set to $Re_c = 2\cdot10^6$.
For the nacelle, a generic configuration based on a NACA 0012 airfoil was generated. The outer geometry of the nacelle geometry with respect to the local chord length is roughly based on the UHBR flow-through nacelle by Spinner \& Rudnik~\cite{spinner2021}, which was designed for the Airbus XRF-1 research model. That is, the inlet and the outlet diameter were set to ${d_{nac} = 0.45c}$ and the nacelle length to ${l_{nac} = 0.75c}$.
The grid dimensions in the x-, y-, and z-coordinate directions $L_x \times L_y \times L_z = 25.4c \times 24.9c \times 0.05c$, which proved to be sufficient for comparable flows by Zauner \& Sandham~\cite{zauner2020} and Moise et al.~\cite{moise2021}. The grid is locally refined within the boundary layer satisfying the mesh requirements for WM-LES proposed by Kawai \& Larsson~\cite{kawai2012}, i.e., at least 20 cells are located within the boundary layer at $x_{oat}/c = 0.2$ and $x_{nac}/c = 0.2$. Since the acoustics in the vicinity of the trailing edge is a key element of established buffet models additional a-priori refinement is applied on the suction side of the airfoil and in the wake region of the airfoil and the nacelle. The Cartesian mesh of the nacelle configuration on the coarsest refinement level is shown in figure~\ref{fig:grid_raw}. Regions of additional mesh refinement are highlighted by a color distribution in figure~\ref{fig:grid_ref}. The local cell length is given by
\begin{align}
L_{lvl} = L_0\cdot\left( \frac{1}{2} \right)^{lvl}
\end{align} where $lvl$ is the cell refinement level and $L_0$ is the mesh base length, which in this case corresponds to the longest mesh dimension of $L_x = 25.4c$.
The resulting number of Cartesian cells for both configurations is given in table~\ref{tab:meshsize}.
{\def\arraystretch{1.5}\tabcolsep=5pt
\begin{table}[htb]
\centering
\begin{tabular}{|c|c|c|}
\hline
& baseline & with nacelle \\
\hline
number of cells & $0.73\cdot 10^9$ & $1.24\cdot 10^9$ \\
\hline
\end{tabular}
\caption{Mesh size of the baseline configuration without nacelle and the configuration including the nacelle.}
\label{tab:meshsize}
\end{table}
}
The wall model is applied to all solid wall boundaries. Following the requirements proposed by Kawai \& Larsson~\cite{kawai2012} the evaluation distance $h_{wm}$ of the wall model is set to a distance of approximately three cell lengths from the wall. A characteristic outflow boundary condition is used for the main outflow plane to suppress spurious reflections. Standard in- and outflow boundary conditions with sponge layers are applied to the remaining in- and outflow planes in the x- and y-direction. Periodicity is applied to spanwise boundaries, which is a great simplification of the in general highly three-dimensional flow field around the engine nacelle. This, however, allows the study of the effect of the upstream shock wave and the free-shear layer introduced by the nacelle geometry onto the buffet phenomenon while keeping the total number of grid points at a manageable level. At the nacelle inlet, a standard outflow boundary condition is applied. For the engine outlet, a $1/7$th power law velocity profile is prescribed while maintaining mass conservation with the engine inlet. Hence, the setup represents a simple flow-through nacelle.
To enforce a controlled transition of the boundary layer into a turbulent state the tripping method by Schlatter and \"Orl\"u~\cite{schlatter2012} is employed at $x/c = 0.07$ on both sides of the airfoil.
\section{Results}\label{sec:results}
For both setups, data were collected for about 70 convective time units $CTU~=~c/u_\infty$. Spanwise averaged surface data of the OAT15A airfoil and the nacelle were sampled at a sampling interval of $4.1\cdot10^{-3}\,CTU$ at every $1\%$ of the airfoil's chord length. Additionally, wall normal velocity and pressure profiles were determined at every $5\%$ of the airfoil's chord length at the same sampling interval of $4.1\cdot10^{-3}\,CTU$. For the analysis of the shock dynamics with DMD, volumetric data of the flow field were collected at a sampling interval of $0.1025\,CTU$.
\subsection{General flow field}\label{ssec:flowfield}
In figure~\ref{fig:cp}, the airfoil's mean and instantaneous pressure coefficients at the most upstream and most downstream shock position are shown for both configurations. It is remarkable that the modeling of the engince nacelle as a flow-through nacelle affects the pressure side only marginally. It is, however, obvious that the integration of the nacelle leads to a significantly weaker shock and an upstream displacement of the mean shock location on the suction side. The mean shock position is determined at approximately $x_s = 0.53$ for the baseline configuration and $x_s = 0.47$ for the configuration including the nacelle. Considering the most upstream and most downstream shock locations, the nacelle configuration exhibits a noticeably lower amplitude in the shock dynamics compared to the baseline configuration. This finding is supported by the time-resolved pressure signal at the mean shock position shown in figure~\ref{fig:psignal}. After a transient phase the baseline configuration is characterized by a stable -- almost harmonic -- oscillation. The pressure signal of the nacelle configuration, however, exhibits highly variable dynamics of noticeably lower amplitude. The frequency spectra of these pressure signals are shown in figures~\ref{fig:cpnj_fft} and \ref{fig:cpj_fft}. The consistent low-frequency shock oscillation of the baseline configuration is evidenced in a clear peak at $Sr= 0.072$. The spectrum of the nacelle configuration, however, is characterized by a much more broadband peak in the range of $0.03 < Sr < 0.04$ with the expected lower amplitude. An additional peak at $Sr = 2.0$ is prominent, which is not observed in the baseline configuration.
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{pictures/cp.pdf}
\caption{Mean pressure coefficient and instantaneous pressure coefficient on the airfoil at the most upstream and downstream shock position of both configurations.}
\label{fig:cp}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{pictures/p_signal_at_shock.pdf}
\caption{Time history of the surface pressure at the mean shock position for both configurations.}
\label{fig:psignal}
\end{figure}
\begin{figure}[htb]
\centering
\begin{subfigure}[t]{0.49\textwidth}
\includegraphics[width=\linewidth]{pictures/cpnj_fft_53.pdf}
\caption{baseline: $x_{s,mean} = 0.53c$}
\label{fig:cpnj_fft}
\end{subfigure}
\begin{subfigure}[t]{0.49\textwidth}
\includegraphics[width=\linewidth]{pictures/cpj_fft_47.pdf}
\caption{with nacelle $x_{s,mean} = 0.47c$}
\label{fig:cpj_fft}
\end{subfigure}
\caption{FFT of the pressure signal at the mean shock position for both configurations.}
\end{figure}
To understand this alteration to the shock dynamics, the disturbance of the flow onto the airfoil due to the integration of the engine nacelle will be considered next. In figure~\ref{fig:diff_alpha} the deviation of the local flow angle of the nacelle configuration from the baseline configuration given by
\begin{align}
\Delta \alpha = \arctan \left( \frac{\bar{u}}{\bar{v}} \right)_{nac} - \arctan \left( \frac{\bar{u}}{\bar{v}} \right)_{base}
\end{align}
is shown. It is evident that immediately upstream of the airfoil the flow is deflected to smaller angles of attack by the contour of the engine nacelle. In figure~\ref{fig:diff_mach} the deviation of the local Mach number of the nacelle configuration from the baseline configuration defined by
\begin{align}
\Delta Ma = \left( \frac{\lvert \bm{\bar{u}} \rvert}{\sqrt{\gamma \bar{p}/\bar{\rho}}} \right)_{nac} - \left(\frac{\lvert \bm{\bar{u}} \rvert}{\sqrt{\gamma \bar{p}/\bar{\rho}}} \right)_{base}
\end{align}
is shown. As a consequence of the shock and the boundary layer on the upper part of the nacelle the Mach number of the flow onto the airfoil is significantly reduced.
The buffet phenomenon is, however, highly sensitive to the angle of attack and the Mach number. Considering the numerical and experimental studies on the OAT15A airfoil at comparable conditions \cite{deck2005, jacquin2009, schauerte2022}, it is reasonable that the reduction of the angle of attack and the Mach number due to the integration of the engine nacelle results in a shift of the dynamics from fully developed buffet to pre-developed buffet.
Furthermore, the Mach number in the free-shear layer downstream of the trailing edge is evidently affected by the integration of the nacelle. The wake flow coming from the upper part of the nacelle interacts with the boundary layer on the pressure side of the airfoil leading to a reduced local Mach number. On the suction side, the weaker shock wave leads to a reduced flow deceleration and to a weaker boundary layer separation and thus, to a higher local Mach number.
Contours of $p'_{rms}$ of both configurations are depicted in figures~\ref{fig:prms_nj} and \ref{fig:prms_j}. The more upstream mean shock position and the lower amplitude of the shock oscillation indicated by figure \ref{fig:cp} are confirmed. Additionally, the pressure fluctuations of the nacelle configuration are massively reduced. Therefore, the integration of the engine nacelle can not merely be considered as a perturbation of the global flow topology, but also has a notable impact on the acoustics between trailing edge and shock wave. The shock-induced separation on the suction side and the noise generated by the trailing edge shear flow, however, are essential to established buffet models \cite{lee2001, hartmann2013b}. This means the alteration of the shock dynamics due to the integration of the nacelle does have an important impact on the overall aerodynamics of the system.
\begin{figure}[htb]
\centering
\begin{subfigure}[t]{0.49\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dAlpha.png}
\put(90,2.4){\footnotesize $-20$}
\put(90,9.3){\footnotesize $-15$}
\put(90,16.2){\footnotesize $-10$}
\put(91,23.0){\footnotesize $-5$}
\put(94.5,29.8){\footnotesize $0$}
\put(94.5,36.6){\footnotesize $5$}
\put(93,43.4){\footnotesize $10$}
\put(93,50.3){\footnotesize $15$}
\put(93,57.2){\footnotesize $20$}
\put(87,62.5){\footnotesize $\Delta\alpha\,[^\circ]$}
\end{overpic}
\caption{Flow angle deviation $\Delta\alpha$. Black lines show iso-contours at $\Delta\alpha = -3.5^\circ.$}
\label{fig:diff_alpha}
\end{subfigure}
\begin{subfigure}[t]{0.49\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dMa.png}
\put(90,2.4){\footnotesize $-0.8$}
\put(90,9.3){\footnotesize $-0.6$}
\put(90,16.2){\footnotesize $-0.4$}
\put(90,23.0){\footnotesize $-0.2$}
\put(93.8,29.8){\footnotesize $0.0$}
\put(93.8,36.6){\footnotesize $0.2$}
\put(93.8,43.4){\footnotesize $0.4$}
\put(93.8,50.3){\footnotesize $0.6$}
\put(93.8,57.2){\footnotesize $0.8$}
\put(87,62.5){\footnotesize $\Delta Ma$}
\end{overpic}
\caption{Mach number deviation $\Delta Ma$. Black lines show iso-contours at $\Delta Ma=-0.1$.}
\label{fig:diff_mach}
\end{subfigure}
\caption{Deviation of the flow field of the nacelle configuration from the baseline configuration.}
\end{figure}
\begin{figure}[htb]
\centering
\begin{subfigure}[t]{0.49\textwidth}
\begin{overpic}[width=\linewidth]{pictures/pfluc_nj.png}
\put(91, 1.0){\footnotesize $0.0$}
\put(91, 8.5){\footnotesize $0.05$}
\put(91, 16.2){\footnotesize $0.1$}
\put(91, 23.8){\footnotesize $0.15$}
\put(91, 31.3){\footnotesize $0.2$}
\put(91, 38.8){\footnotesize $0.25$}
\put(87, 44.5){\footnotesize $p'_{rms}$}
\end{overpic}
\caption{baseline configuration}
\label{fig:prms_nj}
\end{subfigure}
\begin{subfigure}[t]{0.49\textwidth}
\begin{overpic}[width=\linewidth]{pictures/pfluc_j.png}
\put(91, 1.0){\footnotesize $0.0$}
\put(91, 8.5){\footnotesize $0.05$}
\put(91, 16.2){\footnotesize $0.1$}
\put(91, 23.8){\footnotesize $0.15$}
\put(91, 31.3){\footnotesize $0.2$}
\put(91, 38.8){\footnotesize $0.25$}
\put(87, 44.5){\footnotesize $p'_{rms}$}
\end{overpic}
\caption{nacelle configuration}
\label{fig:prms_j}
\end{subfigure}
\caption{Contours of $p'_{rms}$ of the baseline configuration and the nacelle configuration.}
\end{figure}
\subsection{DMD analysis}\label{ssec:dmdanalysis}
\begin{figure}[htb]
\centering
\begin{subfigure}[t]{0.495\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dmd_nj_p_0072.png}
\put(86.5, 0.6){\footnotesize $-8.0$}
\put(86.5, 6.0){\footnotesize $-6.0$}
\put(86.5, 11.5){\footnotesize $-4.0$}
\put(86.5, 16.75){\footnotesize $-2.0$}
\put(90.3, 22.1){\footnotesize $0.0$}
\put(90.3, 27.5){\footnotesize $2.0$}
\put(90.3, 32.8){\footnotesize $4.0$}
\put(90.3, 38.2){\footnotesize $6.0$}
\put(90.3, 43.7){\footnotesize $8.0$}
\put(81.8, 46.7){\footnotesize $\times 10^{-4}$}
\end{overpic}
\caption{pressure mode}
\label{fig:dmd_nj_p_buffet}
\end{subfigure}
\begin{subfigure}[t]{0.495\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dmd_nj_u_0072.png}
\put(86.5, 0.6){\footnotesize $-2.0$}
\put(86.5, 6.0){\footnotesize $-1.5$}
\put(86.5, 11.5){\footnotesize $-1.0$}
\put(86.5, 16.75){\footnotesize $-0.5$}
\put(90.3, 22.1){\footnotesize $0.0$}
\put(90.3, 27.5){\footnotesize $0.5$}
\put(90.3, 32.8){\footnotesize $1.0$}
\put(90.3, 38.2){\footnotesize $1.5$}
\put(90.3, 43.7){\footnotesize $2.0$}
\put(81.8, 46.7){\footnotesize $\times 10^{-4}$}
\end{overpic}
\caption{streamwise velocity mode}
\label{fig:dmd_nj_u_buffet}
\end{subfigure}
\caption{Buffet modes at $Sr=0.072$ of the pressure and the streamwise velocity fields of the baseline configuration.}
\end{figure}
\begin{figure}[htb]
\centering
\begin{subfigure}[t]{0.495\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dmd_j_p_201.png}
\put(86.5, 0.6){\footnotesize $-4.0$}
\put(86.5, 6.0){\footnotesize $-3.0$}
\put(86.5, 11.5){\footnotesize $-2.0$}
\put(86.5, 16.75){\footnotesize $-1.0$}
\put(90.3, 22.1){\footnotesize $0.0$}
\put(90.3, 27.5){\footnotesize $1.0$}
\put(90.3, 32.8){\footnotesize $2.0$}
\put(90.3, 38.2){\footnotesize $3.0$}
\put(90.3, 43.7){\footnotesize $4.0$}
\put(81.8, 46.7){\footnotesize $\times 10^{-4}$}
\end{overpic}
\caption{pressure mode}
\label{fig:dmd_j_p_nac}
\end{subfigure}
\begin{subfigure}[t]{0.495\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dmd_j_u_201.png}
\put(86.5, 0.6){\footnotesize $-2.0$}
\put(86.5, 6.0){\footnotesize $-1.5$}
\put(86.5, 11.5){\footnotesize $-1.0$}
\put(86.5, 16.75){\footnotesize $-0.5$}
\put(90.3, 22.1){\footnotesize $0.0$}
\put(90.3, 27.5){\footnotesize $0.5$}
\put(90.3, 32.8){\footnotesize $1.0$}
\put(90.3, 38.2){\footnotesize $1.5$}
\put(90.3, 43.7){\footnotesize $2.0$}
\put(81.8, 46.7){\footnotesize $\times 10^{-4}$}
\end{overpic}
\caption{streamwise velocity mode}
\label{fig:dmd_j_u_nac}
\end{subfigure}
\caption{Modes of the shock-wave/boundary-layer interaction at $Sr=2.01$ of the pressure and the streamwise velocity fields of the nacelle configuration.}
\end{figure}
\begin{figure}[htb]
\centering
\begin{subfigure}[t]{0.495\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dmd_j_p_0043.png}
\put(86.5, 0.6){\footnotesize $-4.0$}
\put(86.5, 6.0){\footnotesize $-3.0$}
\put(86.5, 11.5){\footnotesize $-2.0$}
\put(86.5, 16.75){\footnotesize $-1.0$}
\put(90.3, 22.1){\footnotesize $0.0$}
\put(90.3, 27.5){\footnotesize $1.0$}
\put(90.3, 32.8){\footnotesize $2.0$}
\put(90.3, 38.2){\footnotesize $3.0$}
\put(90.3, 43.7){\footnotesize $4.0$}
\put(81.8, 46.7){\footnotesize $\times 10^{-4}$}
\end{overpic}
\caption{pressure mode}
\label{fig:dmd_j_p_buffet}
\end{subfigure}
\begin{subfigure}[t]{0.495\textwidth}
\begin{overpic}[width=\linewidth]{pictures/dmd_j_u_0047.png}
\put(86.5, 0.6){\footnotesize $-2.0$}
\put(86.5, 6.0){\footnotesize $-1.5$}
\put(86.5, 11.5){\footnotesize $-1.0$}
\put(86.5, 16.75){\footnotesize $-0.5$}
\put(90.3, 22.1){\footnotesize $0.0$}
\put(90.3, 27.5){\footnotesize $0.5$}
\put(90.3, 32.8){\footnotesize $1.0$}
\put(90.3, 38.2){\footnotesize $1.5$}
\put(90.3, 43.7){\footnotesize $2.0$}
\put(81.8, 46.7){\footnotesize $\times 10^{-4}$}
\end{overpic}
\caption{streamwise velocity mode}
\label{fig:dmd_j_u_buffet}
\end{subfigure}
\caption{Buffet modes of the pressure field at $Sr=0.043$ and the streamwise velocity field at $Sr=0.047$ of the nacelle configuration.}
\end{figure}
To further analyze the dynamics and investigate the underlying physical mechanisms with respect to established buffet models, SP-DMD was performed on data of the pressure and the streamwise velocity field. The penalty parameter $\gamma$ of the SP-DMD algorithm was varied to obtain only the most relevant modes. The performance loss defined by
\begin{align}
\Pi_{loss} = \frac{\|\bm{V}_1^N - \bm{\phi} \bm{D}_a\bm{V}_{and}\|_F}{\|\bm{V}_1^N\|_F} \cdot 100 \%
\end{align}
gives the relative deviation of the reconstructed field from the original snapshot sequence and allows the evaluation of the quality of the flow field reconstruction from the selected modes. More details on the SP-DMD algorithm are given in \cite{jovanovic2014, feldhusen2021}.
Performing SP-DMD on the pressure field data of the baseline configuration with $\gamma = 31257.44$ leaves only one relevant mode at a frequency of $Sr = 0.072$ which agrees well with the peak found in the FFT results of the surface pressure signal at the mean shock position given in figure \ref{fig:cpnj_fft}. Contours of the mode's amplitude are given in figure \ref{fig:dmd_nj_p_buffet}. They clearly show the association with the shock buffet. The performance loss at the given $\gamma$ is $1.7\%$ such that in return $98.3\%$ of the dynamics of the pressure field can be reconstructed using only this mode and its complex conjugate. Lowering the penalty parameter $\gamma$ merely introduces upper or lower harmonics of the buffet frequency into the DMD spectrum. From that it is obvious that the dynamics of the pressure field is solely dominated by the buffet mode. The corresponding mode of the streamwise velocity field at the same frequency of $Sr = 0.072$ is given in figure \ref{fig:dmd_nj_u_buffet}. Apart from the shock movement, this mode also reveals the periodic thickening and shrinking of the boundary layer past the shock wave, which is essential for established buffet models \cite{lee2001, hartmann2013b}. It is remarkable that there are two regions of changing sign in the vicinity of the trailing edge. These regions coincide with the peaks in the pressure fluctuations shown in figure \ref{fig:prms_nj}. The change of sign in the streamwise velocity mode implies a temporal variation of the local shear rate. Fluctuations in the local shear rate are, however, part of the driving acoustic source term in free-shear flows, which is the perturbed Lamb vector given by
\begin{align}
\bm{L'} = (\bm{\omega} \times \bm{u})'
\end{align}
where $\bm{\omega}$ is the vorticity vector \cite{ewert2003}.
One can conclude that the origin of the pressure fluctuations in the vicinity of the trailing edge is due to acoustic perturbations which agrees well with the established buffet models \cite{lee2001, hartmann2013b}. It should also be noted that both the pressure and velocity modes have a Ritz eigenvalue of close to $1$ such that the transient component of these modes is negligible.
Performing SP-DMD on the pressure field data of the nacelle configuration with $\gamma = 24770.98$ leaves a single remaining mode at a frequency of $Sr = 2.01$. A peak at this very frequency was already found in the FFT results in figure~\ref{fig:cpj_fft}. The real parts of the corresponding pressure and streamwise velocity modes are shown in figures \ref{fig:dmd_j_p_nac} and \ref{fig:dmd_j_u_nac}. It is obvious that this dynamic feature is restricted almost completely to the flow on the upper surface of the nacelle and can be associated with the shock-induced variation of the boundary layer on the nacelle and the subsequent vortex generation. The order of magnitude of the frequency generally agrees well with the findings of Moise et al. \cite{moise2022} who investigated the effect of boundary layer tripping on the buffet characteristics. The vortices passing over the trailing edge of the nacelle primarily convect along the pressure side of the airfoil and are subject to substantial dissipation. On the suction side, these structures are only poorly established. In particular, the shock wave on the airfoil suction side exhibits no considerable dynamic features and appears to be negligibly affected at this frequency.
Lowering the penalty parameter to $\gamma = 12328.57$ includes the low-frequency mode at $Sr = 0.043$ which agrees well with the plateau-like low-frequency region in figure \ref{fig:cpj_fft}. The real part of the corresponding pressure mode is given in figure \ref{fig:dmd_j_p_buffet}. The corresponding streamwise velocity mode is detected at a slightly higher frequency of $Sr = 0.047$ and is shown in figure \ref{fig:dmd_j_u_buffet}. It is remarkable that both the nacelle and the airfoil shock display dynamic features at the respective frequency. The spatial amplitude of the shock motion, however, is clearly smaller compared to the buffet found for the baseline configuration. Considering the flow field disturbances by the nacelle geometry discussed in section \ref{ssec:flowfield}, this supports the suggestion that the nacelle suppresses developed buffet at the given flow parameters and a pre-developed state of buffet is observed. The shared dynamic features, however, suggest that a coupling mechanism between the dynamics of the nacelle shock and the airfoil shock exists. Note that similar to the baseline configuration, the streamwise velocity mode exhibits a periodic behavior of the nacelle boundary layer downstream of the shock. The wake flow of the nacelle merges into the boundary layer on the airfoil pressure side, such that eventually also the trailing edge flow of the airfoil is affected by the shock dynamics on the nacelle. Thereby, a connection between the dynamics of the nacelle shock and the mechanisms of existing buffet models \cite{lee2001, hartmann2013b} can be expected. Yet, as discussed in section \ref{ssec:flowfield} the acoustic field at the airfoil trailing edge is considerably weaker for the nacelle configuration such that determining the exact mechanism will require further scrutiny including detailed investigation of the acoustic source terms in the vicinity of the trailing edge.
\section{Conclusion}\label{sec:conclusion}
WM-LES of the OAT15A complemented by a generic 2D nacelle geometry and a corresponding baseline configuration without the nacelle have been performed. The baseline case exhibits a well developed low-frequency shock oscillation at $Sr=0.072$. Investigation of the pressure fluctuations in the vicinity of the trailing edge and the flow field dynamics extracted by means of SP-DMD agree well with established buffet models.
The nacelle configuration is characterized by a system of two shocks, one on the upper part of the engine nacelle and one on the airfoil suction side. Both shocks are subject to a low-frequency oscillation at the same frequency of $Sr=0.043$ which, however, has a smaller spatial amplitude than the buffet found for baseline configuration. The occurrence of a shared dynamic mode of the nacelle and the airfoil shock suggests the existence of a coupling mechanism, the exact determination of which requires further investigation. The analysis of the pressure field reveals a considerable attenuation of the pressure fluctuation in the vicinity of the trailing edge.
Comparison of the mean flow fields of both configurations shows a significant disturbance of the flow conditions onto the airfoil. That is, both Mach number and local angle of attack are reduced by the introduction of the nacelle. The buffet phenomenon is, however, highly sensitive to variations of angle of attack and Mach number. Considering the mitigated and irregular shock dynamics, it is therefore stated that the altered flow conditions lead to a shift from fully developed buffet to pre-developed buffet.
While the reduction of the UHBR nacelle to a 2D-periodic geometry obviously is a huge simplification it allowed the controlled study of the altered flow field and the occurring dynamics at reasonable computational cost. Even for such a generic case, the multitude of disturbances present in the flow makes the identification of a single cause for the altered shock dynamics and the coupling between the nacelle and the airfoil shock a challenging task that requires further investigation.
\backmatter
\bmhead{Acknowledgments} The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V. (www.gauss-centre.eu) for funding this project by providing computing time on the GCS Supercomputer HAWK at Höchstleistungsrechenzentrum Stuttgart (www.hlrs.de).
\\
\\
The authors greatfully acknowledge the Deutsche Forschungsgemeinschaft DFG (German Research Foundation) for funding this work in the framework of the research unit FOR 2895. \\
\\
The authors also wish to thank ONERA for providing the OAT15A airfoil geometry.
\begin{appendices}
\end{appendices}
\newpage
|
\section{Introduction and results}
\subsection{Introduction}
We study the graph Laplacian on a sparse tree and its Hausdorff dimension.
The Hausdorff dimension is defined for sets or measures.
We estimate the Hausdorff dimension of the spectral measure for the graph Laplacian on a sparse tree.
Note that the Hausdorff dimension of a measure and that of the support of the measure are different from each other in general.
If the spectra are purely point spectra, then the Hausdorff dimension of the spectral measure is zero.
If the spectrum is purely absolutely continuous, then the Hausdorff dimension of the spectral measure is one.
In this paper, we show the Hausdorff dimension of a sparse tree exactly.
This paper is organized as follows:
In the rest of Section $1$, we give the main result.
In Section $2$, we give a decomposition of the graph Laplacian.
From this, we can identify the graph Laplacian with one-dimensional discrete Schr\"{o}dinger operators with a sparse potential.
In the Section $3$, we prove that the intermittency function gives the upper bound of the upper Hausdorff dimension.
In Section $4$, we prepare to estimate the intermittency function. Here, we estimate the operator kernel, by using a quadratic form theory and Helffer-Sj\"{o}strand formula.
In Section $5$, we estimate the intermittency function and prove the main theorem.
We define a sparse tree.
We say that $G=(V,E)$ is a graph, if $V$ is a countable set and $E\subset \{e \in 2^V \mid \#e=2 \}$. An element of $V$ $($resp. $E$$)$ is called the vertex $($resp. the edge$)$.
Vertices $a,b\in V$ are said to be adjacent, if $\{a,b\} \in E$. We denote by $a\sim b$, if $a,b \in V$ are adjacent.
Note that this definition implies that there are no egdes which are adjacent to itself.
Vertices $a,b \in V$ are said to be linked, if
there exist
$a_i \in V$, $i=1,2,...,n-1$ such that $a_i \notin \{a,b\}$, $a \sim a_1$, $a_i \sim a_{i+1}$, $i=1,2,...,n-2$, and $a_{n-1}\sim b$.
Here $\{a, a_1,...,a_{n-1},b\}\in 2^V$ is called a path from $a$ to $b$.
Let the degree ${\rm deg}(a)$ of the vertex $a$ be defined by
${\rm deg}(a)=\# \{b \in V \mid a\sim b \}$.
A graph is said to be locally finite, if ${\rm deg}(a)<\infty$ for any vertex $a \in V$.
A graph $G$ is said to be connected if any vertices $a,b \in V$ are linked.
We say that a graph $G=(V,E)$ is a tree, if
$G$ is connected and for any vertices $a,b \in V$, there exists a unique path from $a$ to $b$.
We fix a vertex $o$ of the tree $G$, and $o$ is called the root of $G$.
A tree $G$ with a fixed root $o$ is called a rooted tree.
Let $p(a,b) \in 2^V$ be the unique path from the vertex $a$ to the vertex $b$.
The metric $d(\cdot,\cdot)$ on $V$ is defined by
\begin{displaymath}
d(a,b)=
\begin{cases}
0,
&a=b,
\\
\#p(a,b)-1,
&a \neq b.
\end{cases}
\end{displaymath}
Let $G$ be a rooted tree with a root $o$, and let
$S_n=\{a\in V \mid d(o,a)=n\}$ for $n=0,1,...$.
We say that a rooted tree $G=(V,E)$ is a spherically homogeneous tree if
any vertices in $S_n$ have the same degree $d_n$.
A locally finite spherically homogeneous tree $G$ is uniquely determined by the sequence $(g_n)_{n=0}^{\infty}$,
\[
g_n=
\begin{cases}
d_0,& n=0,\\
d_n-1,& n\geq1.
\end{cases}
\]
\begin{Definition}\label{sparse}
Let $L_n=2^{n^n}$, $n=1,2,... ,$ and $\Gamma \in (0,1)$.
We say that a locally finite spherically homogeneous tree $G=(V,E)$ is a $\Gamma$- sparse tree, if for any $n \geq 0$,
\begin{displaymath}
g_n=
\begin{cases}
[n^{\frac{1-\Gamma}{\Gamma}}],&n\in \{L_m \mid m\in \mathbb{N}\},\\
1,&n \notin \{L_m \mid m \in \mathbb{N}\}.
\end{cases}
\end{displaymath}
\end{Definition}
We define the graph Laplacian for the locally finite graph.
Let $G=(V,E)$ be a locally finite graph.
Let $l^2(V)$ be the set of square summable functions on $V$, and this is the Hilbert space with the inner product given by
\begin{eqnarray}
\left(f,g\right)=
\sum_{u\in V}\overline{f(u)}g(u).\nonumber
\end{eqnarray}
Let $\mathcal{D}\subset l^2(V)$ be defined by
$\begin{displaystyle}
\mathcal{D}=
\left\{
f:V\rightarrow \mathbb{C}\mid \#{\rm supp}(f)<\infty
\right\}
\end{displaystyle}$.
Let $L$, $A$, and $D$ be operators with its domain $\mathcal{D}$ , and defined by
\begin{eqnarray}
Lf(u)&=&
\sum_{v\sim u}(f(v)-f(u)),\nonumber\\
Af(u)&=&
\sum_{v\sim u}f(v),\nonumber\\
Df(u)&=&
\sum_{v\sim u}f(u)={\rm deg}(u)f(u).\nonumber
\end{eqnarray}
These are called graph Laplacian, adjacency matrix, and degree matrix, respectively. The graph Laplacian $L$ is essentially self-adjoint, if the graph is connected \cite[Threom3.1]{Jorgensen}.
\begin{comment}
\begin{Lemma}\label{ess self-adjoint}
$L$ is essentially self-adjoint.
\end{Lemma}
\begin{proof}
\cite[Lemma 8]{self-adjoint}
\end{proof}
\begin{Proof}
\rm
Let $f\in\mathcal{D}(L^*)$ and $g\in\mathcal{D}$.
Since the support of $g$ is a finite set,
\begin{eqnarray*}
( f, L g )&=&\sum_{u\in V}\overline{f(u)}\sum_{v:u \sim v}(g(v)-g(u))\\
&=&\sum_{u\in V}\sum_{v:u \sim v}\overline{f(u)}g(v)-\sum_{u\in V}\sum_{v:u \sim v}\overline{f(u)}g(u)\\
&=&\sum_{u\in V}\sum_{v:u \sim v}\overline{f(v)}g(u)-\sum_{u\in V}\sum_{v:u \sim v}\overline{f(u)}g(u)\\
&=&\sum_{u\in V}\overline{\sum_{u \sim v}(f(v)-f(u))}g(u).\\
\end{eqnarray*}
This implies that for any $f\in \mathcal{D}(L^*)$
\[
L^*f(u)=\sum_{v \in V;u\sim v}(f(v)-f(u)).
\]
Assume $f\in\rm{Ker}\it{(L^*\pm i)}$. Then
\begin{eqnarray*}
( f, L^* f ) &=&\sum_{u\in V}\overline{f(u)}\sum_{v;u\sim v}(f(v)-f(u))\\
&=&\sum_{(u,v)\in V\times V:u\sim v}(\overline{f(u)}f(v)-|f(u)|^2).
\end{eqnarray*}
This implies that $( f, L^* f )$ is a real number, but
$( f, L^* f )=\mp i \|f\|^2$ by the assumption.
Therefore we have $f=\bm{o}$, and $L$ is essentially self-adjoint, if $G$ is connected or
\qed
\end{Proof}
\end{comment}
Let $(X,d_X)$ be a metric space, and $\mathcal{B}(X)$ be the Borel $\sigma$-field of $(X,d_X)$. Let $A$ be a subset of $X$ and the diameter $d_X(A)$ of $A$ be defined by $d_X(A)= \sup \{d_X(x,y) \mid x,y \in A\}$.
Let $\delta>0$ and a family $\{A_i\}_{i=1}^{\infty}$ of subsets of $X$
is called a $\delta$-{\rm cover} of $A$, if
$
A\subset \bigcup_{i=1}^{\infty}A_i
$ and
$
\sup_{1\leq i < \infty}d_X(A_i)\leq \delta
$.
\begin{Definition}
Let $\alpha \in [0,\infty)$ and $\delta>0$.
Let $h^{\alpha}_{\delta}, h^{\alpha}:2^X \rightarrow [0,\infty]$
be defined by
\begin{eqnarray}
h^{\alpha}_{\delta}(A)&=& \inf \left\{ \sum_{i=1}^{\infty}d_X(A_i)^{\alpha} \mid \{A_i\}_{i=1}^{\infty}\text{is a $\delta$-{\rm cover} of $A$}
\right\}, \nonumber \\
h^{\alpha}(A)&=&\lim_{\delta \rightarrow 0}h^{\alpha}_{\delta}(A). \nonumber
\end{eqnarray}
\end{Definition}
Here $h^{\alpha}$ is called $\alpha$-dimensional Hausdorff measure of $X$.
Actually, the restriction of $h^{\alpha}$ to $\mathcal{B}(X)$ is a measure.
Let ${\rm dim}A$ be defined for $A\in 2^{X}$ by
\[
{\rm dim}A= \sup \left \{\alpha \mid h^{\alpha}(A) \neq 0\right\}.
\]
This is called the Hausdorff dimension of $A$.
It follows that
if $0\leq \alpha < {\rm dim}A$, then $h^{\alpha}(A)=\infty$ and that
if ${\rm dim}A <\alpha $, then $h^{\alpha}(A)=0$.
Let $\mu : \mathcal{B}(X)\rightarrow [0,\infty]$ be a measure, and let the lower Hausdorff dimension
${\rm dim}_*\mu$ and
the upper Hausdorff dimension ${\rm dim}^*\mu$
of $\mu$
be defined by
\begin{eqnarray}
{\rm dim}_*\mu&=& \inf \left\{
{\rm dim}A \mid \text{$A\in \mathcal{B}(X)$ such that }\mu(A) \neq 0
\right\}, \nonumber \\
{\rm dim}^*\mu &=& \inf \left\{
{\rm dim}A \mid \text{$A\in \mathcal{B}(X)$ such that } \mu(\mathbb{R}\setminus A)=0
\right\}.\nonumber
\end{eqnarray}
If $\alpha={\rm dim}_*\mu={\rm dim}^*\mu$, then $\mu$ is said to have the exact $\alpha$-Hausdorff dimension.
Let $L$ be the graph Laplacian of the $\Gamma$-sparse tree. The following lemma is proved by Breuer \cite{Breuer}.
\begin{Lemma}\label{former-result}
Let $H=-\overline{L}$, and
let $E$ be the spectral measure of $H$ and
$\tilde{E}$ be the restriction of $E$ to the interval $(0,4)$,
where $\overline{L}$ is the clousre of $L$.
Then it follows that
\begin{enumerate}[$(1)$]
\item
$\sigma_{\rm ac}(H)=\emptyset$,
$\sigma_{\rm pp}(H)\cap(0,4)=\emptyset$,
$\sigma_{\rm sc}(H)\cap(0,4)=(0,4)$,
\item
$\Gamma \leq {\rm dim}_*\tilde{E} \leq {\rm dim}^* \tilde{E} \leq \frac{2\Gamma}{1+\Gamma}$.
\end{enumerate}
\end{Lemma}
We obtain the main theorem below.
\begin{Theorem}\label{main}
We suppose the same assumptions as Lemma \ref{former-result}.
Then
$\Gamma={\rm dim}_*\tilde{E}={\rm dim}^*\tilde{E}$, and
$\tilde{E}$ has the exact $\Gamma$-Hausdorff dimension,
\end{Theorem}
This theorem implies the corollary below.
\begin{_corollary}
For any $\Gamma\in(0,1)$, the restriction of the spectral measure for the graph Laplacian on the $\Gamma$-sparse tree to the interval $(0,4)$ has the exact $\Gamma$-Hausdorff dimension.
\end{_corollary}
\subsection{Preceding results}
The spectral analysis of a sparse tree bears some similarities to the theory of one-dimensional discrete Schr\"{o}dinger operators with a sparse potential.
In Simon-Stolz \cite{sparsepotential}, Schr\"{o}dinger operators with a sparse potential have singular continuous spectrum.
Gilbert-Pearson \cite{Gilbert} finds a relationship between the behavior of subordinate solutions and the spectrum of one-dimensional Schr\"{o}dinger operators.
Jitomirskaya and Last \cite{Jitomirskaya} show a relationship between the Hausdorff dimension of the spectral measure and the behavior of subordinate and non-subordinate solutions. Moreover, they estimate the Hausdorff dimension of the spectral measure by calculating the $L$-norm of non-subordinate solutions. This subordinate solution method is also used in \cite{Breuer}.
On the other hand, the relationship between the type of spectra and the time-averaged behavior of Schr\"{o}dinger operators is also studied.
RAGE theorem implies that if the initial state is singular continuous, then the time-averaged evolution goes to infinity.
Barbaroux, Combes and Montcho \cite{Bar} give a lower bound of the time-averaged momentum of one-dimensional discrete Schr\"{o}dinger operators, by using the upper Hausdorff dimension.
This also shows an inequality between the upper Hausdorff dimension and an intermittency function.
It is, however, crucial to estimate the intermittency function exactly.
Tcheremchantsev \cite{Tcheremchantsev} gives the intermittency function explicitly in the case of sparse potentials.
We will apply \cite{Tcheremchantsev} to the graph Laplacian on a sparse tree.
\section{Preliminaries}
Threre are some decomposition methods for Schr\"{o}dinger operators on some trees.
These methods stem from Naimark and Solomyak \cite{Naimark}.
Breuer \cite{BreuerA}, Kostenko, and Nicolussi \cite{Kostenko} developped this method recently.
They study the case of the continuum Kirchhoff Laplacian.
Allard, Froese \cite{Allard} and Breuer \cite{Breuer} study the case of the graph Laplacian.
We introduce their results as Lemma $\ref{jacobi identification}$.
Their results imply that the graph Laplacian on the spherically homogeneous tree is identified with the direct sum of Jacobi matrices.
Hence, it is sufficient to study Jacobi matrices instead of the graph Laplacian $H$.
Let $G=(V,E)$ be a spherically homogeneous tree determined by the sequence $\{ g_n\}_{n=0}^{\infty}$ and $H=-\overline{L}$, where
$\overline{L}$ is the closure of $L$.
Let $\alpha_n=\#S_n$ for $n=0,1,...$ and $\alpha_{-1}=0$.
Since
$\{\alpha_n\}_{n=0}^{\infty}$ is non-decreasing,
there exists a unique $N(k)\in \mathbb{N}\cup \{0\}$ such that
$\alpha_{N(k)-1}< k \leq \alpha_{N(k)}$ for every $k \in \mathbb{N}$.
Let $k,n\in \mathbb{N}$ and let
$d_{k}=(d_k(n))_{n=1}^{\infty}$ and $a_{k}=(a_{k}(n))_{n=1}^{\infty}$ be defined by the following: in the case of $k=1$,
\begin{eqnarray}
d_1(n)&=&
=
\begin{cases}
g_0&(n=1),\\
g_{n-1}+1&(n \geq 2),
\end{cases}
,\nonumber\\
a_{1}(n)&=&=-\sqrt{g_{n-1}}.\nonumber
\end{eqnarray}
In the case of $k\geq2$,
\begin{eqnarray}
d_k(n)&=&g_{n+N(k)-1}+1,\nonumber\\
a_{k}(n)&=&-\sqrt{g_{n+N(k)-1}}.\nonumber
\end{eqnarray}
By calculatig straightfowardly, we see that for any $k,n \in \mathbb{N}$
\begin{eqnarray}\label{4.1}
d_k(n)=a_k(n)^2+1-\delta_1(k)\delta_1(n),
\end{eqnarray}
where
$\delta_j(k)=
\begin{cases}
1&(k=j),\\
0&(k \neq j).
\end{cases}
$
Let Jacobi matrices $H^{(k)},A^{(k)},D^{(k)} :l^2(\mathbb{N})\rightarrow l^2(\mathbb{N})$ be defined by
\begin{displaymath}
H^{(k)}=
\renewcommand{\arraystretch}{1.6}
\left(
\begin{array}{ccccccc}
d_{k}({1})&-a_{k}({1})\\
-a_{k}({1})&d_{k}({2})&-a_{k}({2})\\
&-a_{k}({2})&d_{k}({3})&-a_{k}({3})\\
&&-a_{k}({3})&\ddots&\ddots\\
&&&\ddots\\
\end{array}\right),
\renewcommand{\arraystretch}{1}
\end{displaymath}
\begin{displaymath}
A^{(k)}=
\renewcommand{\arraystretch}{1.6}
\left(
\begin{array}{ccccccc}
0&a_{k}({1})\\
a_{k}({1})&0&a_{k}({2})\\
&a_{k}({2})&0&a_{k}({3})\\
&&a_{k}({3})&\ddots&\ddots\\
&&&\ddots\\
\end{array}\right)
\renewcommand{\arraystretch}{1},
D^{(k)}=
\renewcommand{\arraystretch}{1.6}
\left(
\begin{array}{ccccccc}
d_{k}({1})&\\
&d_{k}({2})&\\
&&d_{k}({3})&\\
&&&\ddots&\\
&&&\\
\end{array}\right).
\renewcommand{\arraystretch}{1}
\end{displaymath}
Note that $H^{(k)}=D^{(k)}-A^{(k)}$. The next lemma shows the decomposition of graph Laplacian.
\begin{Lemma}\label{jacobi identification}
$H$ and $\begin{displaystyle}
\bigoplus_{k=1}^{\infty}H^{(k)}
\end{displaystyle}$
are unitarly equivalent.
\end{Lemma}
\begin{Proof}\rm
See Appendix \ref{Decomposition of the graph Laplacian}.
\qed
\end{Proof}
\begin{comment}
We introduce some lemmas about the fractal maesure which is used to prove Lemma $\ref{H}.$
Let $\mu: \mathcal{B}^1\rightarrow [0,\infty]$ be a measure, $\mathcal{L}$ be the Lebeasugue measure on $\mathbb{R}$, and $\alpha \in (0,1)$.
We say that $\mu$ is uniformly $\alpha$-H\"{o}lder continuous,
if there exists
$C>0$ such that $\mu(I)<C\mathcal{L}(I)^{\alpha}$
for any interval $I\subset \mathbb{R}$ with $\mathcal{L}(I)<1$.
We consider the Fourier transformation of the fractal measure.
Let $f \in L^1(\mathbb{R},d\mu)$ and
let $\widehat{f\mu}(\xi)$, $\xi \in \mathbb{R}$, be defined by
\[
\widehat{f\mu}(\xi)=\int_{\mathbb{R}}f(x)e^{-i\xi x} \mu(dx).
\]
\begin{Lemma}\label{cor1}
Let $\mu$ be a uniformly $\alpha$-H\"{o}lder continuous and finite measure.
Then there exists $C_1=C_1(\alpha,\mu)>0$ such that for any $f \in L^2(\mathbb{R}, d\mu)$,
\begin{equation}
\sup_{T\geq 1}
T^{\alpha-1}
\int_{0}^{\infty}
e^{-\frac{t}{T}}
|\widehat{f\mu}(t)|^2dt
\leq
C_1\|f\|^2.
\nonumber
\end{equation}
\end{Lemma}
\begin{Proof}
\qed
\end{Proof}
We review the relationship between the local Hausdorff dimension and the Hausdorff dimension for the measure.
The local Hausdorff dimension $\gamma:\mathbb{R}\rightarrow \mathbb{R}$ of the measure $\nu$ be defined by
\begin{equation}
\gamma(x)=
\liminf_{\delta \rightarrow 0}
\cfrac{{\rm log}(\mu([x-\delta, x+\delta]))}{{\rm log}\delta}.
\nonumber
\end{equation}
\begin{Lemma}\label{lem1}
We have
$\begin{displaystyle}\mu {\rm \mathchar`-}
\esssup\displaylimits_{x}\gamma(x)={\rm dim}^*\mu
\end{displaystyle}$, and
$\begin{displaystyle}\mu {\rm \mathchar`-}
\essinf\displaylimits_{x}\gamma(x)={\rm dim}_*\mu.
\end{displaystyle}$
\end{Lemma}
\begin{Proof}
\rm
\cite[Chapter 10, Lemma10.2, 10.3]{Haus}
\qed
\end{Proof}
\end{comment}
\begin{comment}
We prove by contradiction.
Let $\alpha= {\rm dim}^*\mu$ and assume that
$\mu {\rm \mathchar`-} \esssup\displaylimits\gamma=\alpha-\epsilon<\alpha$.
There exists a sequence $\{ \delta_n >0\}$ such that
$\begin{displaystyle}
\lim_{n\rightarrow \infty}\delta_n=0
\end{displaystyle}$ and
\begin{eqnarray}
\lim_{n \rightarrow \infty}\cfrac{{\rm log}(\mu([x-\delta_n,x+\delta_n]))}{{\rm log}\delta_n}=\gamma(x)\leq \alpha-\epsilon
\qquad \text{ \rm a.e. } x
\nonumber
\end{eqnarray}
with respect to $\nu$.
Let $n\in \mathbb{N}$ be large enough, then we see that
\begin{equation}\label{P}
\cfrac{{\rm log}(\mu([x-\delta_n,x+\delta_n]))}{{\rm log}\delta_n}\leq \alpha-
\cfrac{\epsilon}{2}.
\nonumber
\end{equation}
Therefore, we have
\begin{eqnarray}
\lim_{n\rightarrow \infty}
\cfrac{\mu([x-\delta_n,x+\delta_n])}{\delta_n^{\alpha-\frac{1}{4}\epsilon}}
&=&
\lim_{n\rightarrow \infty}
{\rm exp}
\left(
\left(
\cfrac{{\rm log}\mu([x-\delta_n,x+\delta_n])}{{\rm log}\delta_n}
-\alpha+\cfrac{1}{4}\epsilon
\right)
{\rm log}\delta_n
\right)
\nonumber
\\
&\geq&
\lim_{n\rightarrow \infty}
{\rm exp}
\left(
\left(
\alpha-
\cfrac{\epsilon}{2}
-\alpha+\cfrac{1}{4}\epsilon
\right)
{\rm log}\delta_n
\right)
=\infty.
\nonumber
\end{eqnarray}
This implies that
$D^{\alpha-\frac{1}{4}\epsilon}\mu(x)=\infty$
for a.e. $x\in \mathbb{R}$
with respect to $\nu$.
By Lemma $\ref{rogers}$ and Lemma $\ref{E}$ $(2)$, we see that
$\alpha={\rm dim}^*\mu\leq \alpha -\frac{1}{4}\epsilon$.
By the contradiction, we see that $\nu {\rm \mathchar`-} \esssup\displaylimits\gamma \geq \alpha$.
Assume that $\mu {\rm \mathchar`-}\esssup\displaylimits\gamma=\alpha+\epsilon > \alpha$.
By Lemma $\ref{rogers}$ and Lemma $\ref{E}$ $(2)$,
${\rm dim}^*\mu=\alpha$ implies that
\begin{equation}
D^{\alpha+\frac{1}{4}\epsilon}\mu(x)=\infty \qquad \text{ \rm a.e. } x
\nonumber
\end{equation}
with respect to $\mu$. By the assumption,
there exists $S\in \mathcal{B}^1$ such that $\mu(S)>0$ and
for $x \in S$
\begin{equation}
\alpha+\frac{3}{4}\epsilon\leq \gamma(x) \leq \alpha +\epsilon.
\label{A}
\nonumber
\end{equation}
Let $x\in S$, then there exists $\delta^{\prime}>0$ such that
\begin{equation}\label{S}
\inf_{0<\delta<\delta^{\prime}}
\cfrac{{\rm log}\mu([x-\delta,x+\delta])}{{\rm log}\delta}\geq \alpha
+\cfrac{1}{2}\; \epsilon.
\nonumber
\end{equation}
Therefore, we see that for any $x\in S$,
\begin{eqnarray}
D^{\alpha+\frac{1}{4}\epsilon}\mu(x)&=&
\lim_{\delta^{\prime}\rightarrow 0}\sup_{0<\delta<\delta^{\prime}}
{\rm exp}\left(
{\rm log}\delta
\left(
\cfrac{{\rm log}\mu([x-\delta,x+\delta])}{{\rm log}\delta}
-\alpha-\cfrac{1}{4}\;\epsilon
\right)
\right)
\nonumber\\
&\leq&
\lim_{\delta^{\prime}\rightarrow 0}\sup_{0<\delta<\delta^{\prime}}
{\rm exp}\left(
{\rm log}\delta
\left(
\inf_{0<\delta<\delta^{\prime}}
\cfrac{{\rm log}\mu([x-\delta,x+\delta])}{{\rm log}\delta}
-\alpha-\cfrac{1}{4}\;\epsilon
\right)
\right)
\nonumber\\
&\leq&
\lim_{\delta^{\prime}\rightarrow 0}\sup_{0<\delta<\delta^{\prime}}
{\rm exp}\left(
{\rm log}\delta
\left(
\alpha+\cfrac{1}{2}\; \epsilon
-\alpha-\cfrac{1}{4}\; \epsilon
\right)
\right)
\nonumber\\
&\leq&
\lim_{\delta^{\prime}\rightarrow 0}
{\delta^{\prime}}^{\frac{1}{4}\epsilon}=0.
\nonumber
\end{eqnarray}
By the contradiction, we see that
$\mu {\rm \mathchar`-}\esssup\displaylimits\gamma \leq \alpha$.
\end{comment}
\section{Intermittency function and Hausdorff dimension}
In this section, we introduce an intermittency function and give an important inequality in Lemma $\ref{H}$ which shows that the intermittency function is the upper bound of Hausdorff dimension.
\begin{comment}
Let $\mu: \mathcal{B}^1\rightarrow [0,\infty]$ be a measure, and
$\alpha \in (0,1)$.
Let $D^{\alpha}\mu:\mathbb{R}\rightarrow [0,\infty]$ be defined by
\[
D^{\alpha}\mu(E)= \limsup_{\delta \rightarrow 0}\cfrac{\mu((E-\delta, E+\delta))}{\delta^{\alpha}}.
\]
Let $\alpha \in [0,\infty)$.
The measure $\mu$ is said to be $\alpha$-continuous, if
$\mu(A)=0$ for any $A\in \mathcal{B}(X)$ with $h^{\alpha}(A)=0$,
$\mu$ is said to be $\alpha$-{\rm singular}, if
there exists $A\in \mathcal{B}(X)$ such that $\mu(X\setminus A)=0$ and
$h^{\alpha}(A)=0$, and $\mu$ is said to be $\alpha$-absolutely continuous, if
there exists a measureble function $f$ such that $d\mu=fdh^{\alpha}$.
\end{comment}
Let $\psi \in l^2(\mathbb{N})$ and
$E^{(k)}$ be the spectral measure of $H^{(k)}$.
We consider the time-averaged dynamics of $\text{exp}(-itH^{(k)})\psi$.
Let a finite measure $\mu_{\psi}^{(k)} : \mathcal{B}^1 \rightarrow [0,\infty]$ be defined by
\begin{eqnarray}
\mu_{\psi}^{(k)}(A)= (\psi,E^{(k)}(A)\psi).
\nonumber
\end{eqnarray}
\begin{Definition}
Let $\psi_k(t)= e^{-itH^{(k)}}\psi$ and $\psi_k(t,n)= (\delta_n, \psi_k(t))$
for $t \in \mathbb{R}$ and $n\in \mathbb{N}$.
Let $a_{\psi}^{(k)}(n,T)$, $\langle |X|^p\rangle_{\psi}^{(k)}(T)$, and $\beta_{\psi}^{(k)}(p)$ be defined by, for $T >0$ and $p>0$,
\begin{eqnarray}
a_{\psi}^{(k)}(n,T)&=&
\cfrac{1}{T}\int_{0}^{\infty} e^{-\frac{t}{T}}|\psi_k(t,n)|^2 dt,
\nonumber
\\
\langle |X|^p\rangle_{\psi}^{(k)}(T)&=&
\sum_{n=1}^{\infty}n^pa_{\psi}^{(k)}(n,T),
\nonumber
\\
\beta_{\psi}^{(k)}(p)&=&
\cfrac{1}{p}\liminf_{T \rightarrow \infty}\cfrac{{\rm log}\langle |X|^p\rangle_{\psi}^{(k)}(T)}{{\rm log}T}.
\nonumber
\end{eqnarray}
\end{Definition}
We call $\beta_{\psi}^{(k)}$ the intermittency function.
The closed subspace $\mathcal{H}_{\psi}$ of $l^2(\mathbb{N})$ is defined by
\begin{eqnarray}
\mathcal{H}_{\psi}=
\overline{
\left\{
p(H^{(k)})\psi \in l^2(\mathbb{N})\mid \text{$p$ is a polynomial}
\right\}},
\nonumber
\end{eqnarray}
and let $U_{\psi}:\mathcal{H}_{\psi}\rightarrow L^2(\mathbb{R}, d\mu_{\psi}^{(k)})$
be defined by
\[
U_{\psi}(p(H^{(k)})\psi)(x)= p(x).
\]
\begin{Lemma}\label{H}
Let $\alpha = {\rm dim}^*(\mu_\psi^{(k)})$ and $\epsilon>0$.
Then there exists
$C_1=C_1(\epsilon, \psi)>0$ such that for any $T>0$,
\[
\langle |X|^p\rangle_{\psi}^{(k)}(T) \geq C_1T^{p(\alpha-\epsilon)}.
\]
In particular, for any $p>0$,
\[
{\rm dim}^*(\mu_{\psi}^{(k)}) \leq \beta_{\psi}^{(k)}(p).
\]
\end{Lemma}
\begin{Proof}\rm
We denote
$\mu_{\psi}^{(k)}$, $a_{\psi}^{(k)}$, $\langle |X|^p \rangle^{(k)}_{\psi}$, and $\psi_{k}(t,n)$ by
$\mu_{\psi}$, $a_{\psi}$, $\langle |X|^p \rangle_{\psi}$, and $\psi(t,n)$ for simplicity of notation.
Let $\epsilon>0$ and $\gamma:\mathbb{R}\rightarrow \mathbb{R}$ be the local Hausdorff dimension of $\mu_{\psi}$:
\[
\gamma(x)=
\liminf_{\delta \rightarrow 0}
\cfrac{{\rm log}(\mu_{\psi}([x-\delta, x+\delta]))}{{\rm log}\delta}.
\nonumber
\]
By \cite[Chapter 10, Proposition 10.1]{Haus}, we see that $\begin{displaystyle}\mu {\rm \mathchar`-}
\esssup\displaylimits_{x}\gamma(x)={\rm dim}^*\mu_{\psi}=\alpha
\end{displaystyle}$.
Thus there exists $S_{\epsilon}\in \mathcal{B}^1$ such that
$\mu_\psi(S_{\epsilon})>0$ and
$\gamma(x)>\alpha-\epsilon$ for $x \in S_{\epsilon}$.
Let
$\begin{displaystyle}
\gamma_{\delta}(x)=
\inf_{\delta^{\prime}<\delta}\cfrac{{\rm log}\mu_{\psi}([x-\delta^{\prime},x+\delta^{\prime}])}{{\rm log}\delta^{\prime}}
\end{displaystyle}$.
By Egorov's theorem, there exists
$S_{\epsilon}^{\prime}\subset S_{\epsilon}$ such that
$\mu_{\psi}(S_{\epsilon}^{\prime})>0$ and
$\gamma_{\delta}$ converges uniformly to $\gamma$ on $S_{\epsilon}^{\prime}$.
Let $\psi^{\prime} = E(S_{\epsilon}^{\prime})\psi$. We see that
$\|\psi^{\prime}\|^2=\mu_{\psi}(S_{\epsilon}^{\prime})>0$ and
$\mu_{\psi^{\prime}}$ is uniformly $(\alpha-\epsilon)$-H\"{o}lder continuous.
Let $\chi= \psi-\psi^{\prime}$. Then we see that
\begin{eqnarray}
&&\sum_{n=1}^{N}
a_{\psi}(n,T)
\nonumber\\
&&=
\cfrac{1}{T}
\int_{\mathbb{R}}e^{-\frac{t}{T}}
\sum_{n=1}^{N}|\psi^{\prime}(t,n)+\chi(t,n)|^2dt
\nonumber
\\
&&\leq
\sum_{n=1}^{N}
\cfrac{1}{T}
\int_{\mathbb{R}}e^{-\frac{t}{T}}
|\psi^{\prime}(t,n)|^2dt
+2
\left(
\sum_{n=1}^{N}
\cfrac{1}{T}
\int_{\mathbb{R}}e^{-\frac{t}{T}}
|\psi^{\prime}(t,n)|^2dt
\right)^{\frac{1}{2}}\|\chi\|
+\|\chi\|^2.
\label{eq}
\end{eqnarray}
We assume that $c >0$ and $N\in \mathbb{N}$ satisfy
\begin{eqnarray}
\left(
\sum_{n=1}^{N-1}
\cfrac{1}{T}
\int_{\mathbb{R}}e^{-\frac{t}{T}}
|\psi^{\prime}(t,n)|^2dt
\right)^{\frac{1}{2}}
\leq c\|\psi^\prime\|.
\label{q2e}
\end{eqnarray}
By $(\ref{eq})$ and $(\ref{q2e})$, we see that
\begin{eqnarray}
\sum_{n=1}^{N-1}
a_{\psi}(n,T)\leq \left(
c\|\psi^\prime\|+\|\chi\|
\right)^2.
\nonumber
\end{eqnarray}
Taking $c=-\frac{\|\chi\|}{\|\psi^{\prime}\|}+
\sqrt{\left(\frac{\|\chi\|}{\|\psi^{\prime}\|}\right)^2+\cfrac{1}{2}}$\:,
we have
\begin{equation}
\sum_{n=1}^{N-1}a_{\psi}(n,T)
\leq
\frac{1}{2}\|\psi^{\prime}\|^2+\|\chi\|^2.
\label{LK}
\end{equation}
On the other hand, let $C_1^{\prime}=-\frac{\|\chi\|}{\|\psi^{\prime}\|}+
\sqrt{\left(\frac{\|\chi\|}{\|\psi^{\prime}\|}\right)^2+\cfrac{1}{2}}$ and
\begin{equation}
N(T)=
{\rm max}
\left\{
N \in \mathbb{N} \mid
\left(
\sum_{n=1}^{N-1}
\cfrac{1}{T}
\int_{\mathbb{R}}e^{-\frac{t}{T}}
|\psi^{\prime}(t,n)|^2dt
\right)^{\frac{1}{2}}
\leq C_1^{\prime}\|\psi^{\prime}\|
\right\}.
\label{KO8}
\end{equation}
Then $(\ref{q2e})$ holds for $c=C_1^{\prime}$ and $N=N(T)$.
By $(\ref{LK})$, we have
\begin{equation}
\sum_{n=N(T)}^{\infty}a_{\psi}(n,T)\geq \cfrac{1}{2}\|\psi^{\prime}\|^2.
\label{MNBV}
\end{equation}
Note that $\mu_{\psi^\prime}$ is uniformly $(\alpha-\epsilon)$-H\"{o}lder continuous.
Hence, by Lemma \ref{cor1}, there exists $\widetilde{C}=\widetilde{C}(\alpha-\epsilon, \mu_{\psi^\prime})>0$ such that for any $T>0$,
\begin{equation}
\sum_{n=1}^{N}
\cfrac{1}{T}
\int_{\mathbb{R}}e^{-\frac{t}{T}}
|\psi^{\prime}(t,n)|^2dt
=
\sum_{n=1}^{N}\cfrac{1}{T}\int_{0}^{\infty} e^{-\frac{t}{T}}|\widehat{U_{\psi^\prime}\delta_n \mu_{\psi^\prime}}|^2 dt
\leq
\widetilde{C} NT^{-(\alpha-\epsilon)}.
\label{hint}
\end{equation}
By $(\ref{KO8})$ and $(\ref{hint})$, we have
\begin{equation}
N(T)\geq \frac{(C_1^{\prime} \|\psi^{\prime}\|)^2}{\widetilde{C}}
T^{\alpha-\epsilon}.
\label{MNBV1}
\end{equation}
$(\ref{MNBV})$ and $(\ref{MNBV1})$ show that there exists $C_1=C_1(\epsilon, \psi)>0$ such that for any $T>0$,
\begin{eqnarray}
\langle |X|^p\rangle_{\psi}(T)\geq\sum_{n=N(T)}^{\infty}n^pa_{\psi}(n,T)
\geq\cfrac{1}{2}\|\psi^{\prime}\|^2N(T)^p
\geq
C_1T^{p(\alpha-\epsilon)}.
\nonumber
\end{eqnarray}
Moreover, we see that for $\epsilon>0$ and $T^{\prime}>1$,
\begin{eqnarray}
\cfrac{1}{p}
\inf_{T>T^{\prime}}
\cfrac{{\rm log}\langle |X|^p\rangle_{\psi}(T)}{{\rm log}T}\geq
\alpha-\epsilon+
\cfrac{1}{p}
\inf_{T>T^{\prime}}
\cfrac{{\rm log}C_1}{{\rm log}T}
=\alpha-\epsilon.
\nonumber
\end{eqnarray}
This implies our assertion.
\qed
\end{Proof}
\section{Estimates of operator kernel}
In this section we prepare some lemmas to estimate the intermittency function.
We estimate the operator kernel in Lemma $\ref{6.2}$ and $\ref{6.2.4}$ by using a quadratic form theory and Hellfer-Sj\"{o}strand formula.
\begin{comment}
Let $\mathcal{H}$ ba a complex Hilbert space.
Let $\mathfrak{t}:\mathcal{H}\times \mathcal{H}\rightarrow \mathbb{C}$ be
a densely defined, closed, and symmetric form bounded from below, and
let $T$ be the self-adjoint operator associated with $\mathfrak{t}$.
Suppose that $\mathfrak{s}:\mathcal{H}\times \mathcal{H}\rightarrow \mathbb{C}$ is
a relatively bounded sesquilinear form with respect to $\mathfrak{t}$ such that
for any $f \in \mathcal{D}(\mathfrak{t}) \subset \mathcal{D}(\mathfrak{s})$,
\[
|\mathfrak{s}[f]|\leq a\mathfrak{t}[f]+b\|f\|^2, \qquad 0<a<1,\;b\geq0.
\]
Then it is known that $\mathfrak{t}^{\prime}= \mathfrak{s}+\mathfrak{t}$ is sectorial and closed.
Let $T^{\prime}$ be the m-sectorial operators associated with $\mathfrak{t}^{\prime}$.
If $z\in \rho(T)$ and
\[
2\|(aT+b)(T-z)^{-1}\|\leq \gamma <1,
\]
then
$z\in \rho (T^{\prime})$ and
\[
\|(T^{\prime}-z)^{-1}-(T-z)^{-1}\|\leq \cfrac{4\gamma}{(1-\gamma)^2}\|(T-z)^{-1}\|.
\]
\end{comment}
We denote $H^{(k)}$, $a_{k}(n)$, and $d_k(n)$ by
$H$, $a(n)$, and $d(n)$, respectively for simplicity of notation.
Let $\beta>0$ and let $\mathscr{D}=\{f:\mathbb{N}\rightarrow \mathbb{C}\mid \#{\rm supp}(f)<\infty\}$.
Let $P$, $\Delta$, and $M_{\beta}:l^2(\mathbb{N})\rightarrow l^2(\mathbb{N})$ with its domain $\mathscr{D}$ be defined by
\begin{eqnarray}
P f(n)&=& a(n)f(n+1),
\nonumber
\\
\Delta f&=& (P-I)f,
\nonumber
\\
M_{\beta} f(n)&=& \beta^nf(n),
\nonumber
\end{eqnarray}
and let $T_{\beta}= M_{\beta}^{-1}TM_{\beta}$ for an operator $T:l^2(\mathbb{N})\rightarrow l^2(\mathbb{N})$.
\begin{Lemma}
Let $f\in \mathscr{D}$. It follows that
\begin{enumerate}[$(1)$]
\item
$P^*f(n)=\begin{cases}
0&(n=1),
\label{I}
\\
a(n-1)f(n-1)&(n \geq 2),
\end{cases}
$
\item
\label{II}
$
H^{(k)}f=
\begin{cases}
(\Delta\Delta^*-\delta_1) f&(k=1),\\
\Delta\Delta^* f&(k\geq2),
\end{cases}
$
\item
$\Delta_{\beta}f=(\beta P-I)f$,
\label{III}
\item
$(\Delta^*)_{\beta}f=(\beta^{-1} P^*-I)f$.
\label{IV}
\end{enumerate}
\end{Lemma}
\begin{Proof}\rm
Let $f,g \in \mathscr{D}$. Then we see that
\begin{eqnarray}
(Pg,f)=\sum_{n=1}^{\infty}a(n)\overline{g(n+1)}f(n)=\sum_{n=2}^{\infty}\overline{g(n)}a(n-1)f(n-1).
\nonumber
\end{eqnarray}
This implies $(\ref{I})$. We see that
\begin{eqnarray}
\Delta\Delta^* f(n)&=&a(n)\Delta^* f(n+1)-\Delta^* f(n)
\nonumber
\\
&=&
\begin{cases}
a(1)(a(1)f(1)-f(2))+f(1)&(n=1)\\
a(n)(a(n)f(n)-f(n+1))-a(n-1)f(n-1)+f(n)&(n\geq 2)
\end{cases}
\nonumber
\\
&=&
\begin{cases}
-a(1)f(2)+\{a(1)^2+1\}f(1)&(n=1)\\
-a(n)f({n+1})+\{a(n)^2+1\}f(n)-a({n-1})f({n-1})&(n\geq 2).
\end{cases}
\label{098}
\end{eqnarray}
$(\ref{II})$ follows from $(\ref{098})$ and (\ref{4.1}).
We can prove $(\ref{III})$ and $(\ref{IV})$ straightforwardly.
\qed
\end{Proof}
Let $\beta>0$ and the sesquilinear form $\mathfrak{h}_{\beta}:l^2(\mathbb{N})\times l^2(\mathbb{N})\rightarrow \mathbb{C}$ with its domain $\mathscr{D}$ be defined by
\begin{equation}
\mathfrak{h}_{\beta}(f,g)
=
((\Delta_{\beta})^*f,(\Delta^*)_{\beta}g)
=(f,H_{\beta} \; g).
\nonumber
\end{equation}
\begin{Lemma}
For any $t>0$ and $f\in \mathscr{D}$,
\begin{eqnarray}
|\mathfrak{h}_{\beta}[f]-\mathfrak{h}_1[f]|
\leq C(\beta)\;\cfrac{t}{2}\;
\mathfrak{h}_1[f]
+C(\beta)(1+\cfrac{1}{2t}\:)\|f\|^2,
\label{5.2.2}
\end{eqnarray}
where $C(\beta)=|\beta-1|+|\beta^{-1}-1|=|\beta-\beta^{-1}|$.
\end{Lemma}
\begin{Proof}\rm
Let $t>0$ and $f \in \mathscr{D}$. Then we see that
\begin{eqnarray}
|\mathfrak{h}_{\beta}[f]-\mathfrak{h}_1[f]|&=&
|((\beta P^*-I)f,(\beta^{-1}P^*-I)f)-((P-I)f,(P^*-I)f)|
\nonumber
\\
&\leq&
|\beta-1||(f,Pf)|
+|\beta^{-1}-1||(f,P^*f)|
\nonumber
\\
&\leq&
|\beta-1|\{
(\Delta^*f,f)+(f,f)
\}
+|\beta^{-1}-1|
\{
(f,\Delta^*f)+(f,f)
\}
\nonumber
\\
&\leq&
C(\beta)\|f\|\|\Delta^*f\|+C(\beta)\|f\|^2
\nonumber
\\
&\leq&
C(\beta)
\left(
\cfrac{t}{2}\|\Delta^*f\|^2+\cfrac{1}{2t}\|f\|^2+\|f\|^2
\right).
\nonumber
\end{eqnarray}
\qed
\end{Proof}
\begin{Lemma}\label{6.2}
Let $z\in \mathbb{C}^+\coloneqq \{z\in \mathbb{C}\mid \text{\rm Im}z>0\}$. Let $0<\gamma<1$, and let $\eta_z$, $m_z$, and $\alpha_z$ be
\begin{eqnarray}
\eta_z&=& {\rm dist}(z, \sigma(H)),
\nonumber
\\
m_z&=& \cfrac{\eta_z}{\sqrt{\eta_z+|z|}+1},
\nonumber
\\
\alpha_z(\gamma)&=& \cfrac{1}{4}(\gamma m_z+\sqrt{(\gamma m_z)^2+16}).
\nonumber
\end{eqnarray}
Then for any $i,j \in \mathbb{N}$,
\begin{equation}
|(\delta_i, (H-z)^{-1}\delta_j)|\leq
{\alpha_z(\gamma)}^{-|i-j|}
\cfrac{1}{\eta_z}
\left(
\cfrac{1+\gamma}{1-\gamma}
\right)^2.
\nonumber
\end{equation}
\end{Lemma}
\begin{Proof}\rm
It follows from $\|(H-z)^{-1}\|=\eta_z^{-1}$ that for any $t>0$
\begin{eqnarray}
2
\left\|
C(\beta)
\left(
\cfrac{t}{2}H+1+\frac{1}{2t}
\right)
(H-z)^{-1}
\right\|
\leq
C(\beta)
\left\{
\left(
1+\cfrac{|z|}{\eta_z}
\right)t+
\left(
2+\cfrac{1}{t}
\right)
\cfrac{1}{\eta_z}
\right\}
.
\label{Y}
\end{eqnarray}
Let $\gamma \in (0,1)$ and
\begin{eqnarray}
t_z
&=&\cfrac{1}{\sqrt{\eta_z+|z|}},
\nonumber
\\
\beta_z&=&
\frac{1}{4}(\gamma m_z+\sqrt{(\gamma m_z)^2+16})>1.
\nonumber
\end{eqnarray}
By the inequality of arithmetic and geometric means, we see that for any $z\in \mathbb{C^+},$
\begin{eqnarray
C(\beta_z)
\left\{
\left(
1+\cfrac{|z|}{\eta_z}
\right)t_z+
\left(
2+\cfrac{1}{t_z}
\right)
\cfrac{1}{\eta_z}
\right\}
&=&
C(\beta_z)
\left\{
\left(
1+\cfrac{|z|}{\eta_z}
\right)t_z+
\cfrac{1}{\eta_z}\cfrac{1}{t_z}
+
\cfrac{2}{\eta_z}
\right\}
\nonumber
\\
&=&
2\cfrac{\sqrt{\eta_z+|z|}+1}{\eta_z}\:C(\beta_z)
\nonumber
\\
&=&
\cfrac{2}{m_z}
\left(
\beta_z-\frac{1}{\beta_z}
\right)
\nonumber
\\
&=& \gamma.
\label{QAZ}
\end{eqnarray}
$(\ref{Y})$ and $(\ref{QAZ})$ imply that for any $z\in \mathbb{C}^+$,
\begin{eqnarray}
2
\left\|
C(\beta_z)
\left(
\cfrac{t_z}{2}H+1+\frac{1}{2t_z}
\right)
(H-z)^{-1}
\right\|
\leq \gamma
\label{ghj}
\end{eqnarray}
By $(\ref{5.2.2})$, $(\ref{ghj})$ and Lemma \ref{quadratic thm},
there exists the m-sectoral operator $H_{\beta_z}$ associated with $\mathfrak{h}_{\beta_z}$ and for any $z \in \mathbb{C}^{+}$,
\begin{eqnarray
\|(H_{\beta_z}-z)^{-1}-(H-z)^{-1}\|\leq
\cfrac{4\gamma}{(1-\gamma)^2}\|(H-z)^{-1}\|.
\nonumber
\end{eqnarray}
Therefore we see that
\begin{equation
\|(H_{\beta_z}-z)^{-1}\|\leq
\cfrac{1}{\eta_z}
\left(
\cfrac{1+\gamma}{1-\gamma}
\right)^2.
\nonumber
\end{equation}
Let $i,j \in \mathbb{N}$ with $i<j$. Then we see that
\begin{eqnarray}
|(\delta_i,(H-z)^{-1}\delta_j)|
&=&
|(M_{\beta_z}\delta_i,(H_{\beta_z}-z)^{-1}M_{\beta_z}^{-1}\delta_j)|
\nonumber
\\
&\leq&
{\beta_z}^{i-j}\|(H_{\beta}-z)^{-1}\|
\nonumber
\\
&\leq&
{\beta_z}^{-|i-j|}
\cfrac{1}{\eta_z}
\left(
\cfrac{1+\gamma}{1-\gamma}
\right)^2.
\label{kjh}
\end{eqnarray}
This implies our assertion in the case of $i<j$. In the case of $i \geq j$, let
\[
\beta_z=
\frac{1}{4}(-\gamma m_z+\sqrt{(\gamma m_z)^2+16}).
\]
Then we can prove $(\ref{kjh})$ similarly.
\qed
\end{Proof}
Let $f\in C^{n}(\mathbb{R})$, and the norm $\opnorm{\cdot}_n$ on $C^{n}(\mathbb{R})$ be defined by
\[
\opnorm{f}_n=
\sum_{r=0}^{n}
\int_{\mathbb{R}}|f^{(r)}(x)|\langle x \rangle^{r-1}dx.
\]
The next lemma is used to estimate the intermittency function.
\begin{Lemma}\label{6.2.4}
Suppose that $f \in C^{2k+3}(\mathbb{R})$ and $\opnorm{f}_{2k+3}<\infty$.
Then there exists $C_2=C_2(k)>0$ such that for any $i,j\in \mathbb{N}$,
\[
|(\delta_i, f(H)\delta_j)|\leq C_2\opnorm{f}_{2k+3}\langle i-j \rangle^{-k},
\]
where $\langle x \rangle = (1+|x|^2)^{\frac{1}{2}}$.
\end{Lemma}
\begin{Proof}\rm
Let $n \geq 0$ and
$\tau \in C^{\infty}_0(\mathbb{R})$ such that
$\begin{displaystyle}
\tau(x)=
\begin{cases}
1&(|x|\leq 1)\\
0&(|x|\geq 2)
\end{cases}
\end{displaystyle}$.
By Helffer-Sj\"{o}strand formula \cite[2.2 The Helffer-Sj\"{o}strand formula]{Hollfer}, we see that
\begin{eqnarray}
f(H)=\cfrac{1}{\pi}
\int_{\mathbb{C}}
\cfrac{\partial \tilde{f}}{\partial \overline{z}}(z)(H-z)^{-1}dxdy,
\nonumber
\\
\tilde{f}(z)=
\left\{
\sum_{r=0}^{n}
f^{(r)}(x)\cfrac{(iy)^r}{r!}
\right\}
\tau
\left(\frac{y}{\langle x \rangle}
\right).
\nonumber
\end{eqnarray}
We see that
\begin{eqnarray}
\left|
\cfrac{\partial \tilde{f}}{\partial \overline{z}}(z)
\right|
&\leq&\;
\cfrac{1}{2}
\left|
f^{(n+1)}(x)
\cfrac{(iy)^n}{n!}
\tau
\left(
\cfrac{y}{\langle x \rangle}
\right)
\right|
\nonumber
\\
&
+&
\cfrac{1}{2}
\left|
\sum_{r=0}^{n}
f^{(r)}(x)\cfrac{(iy)^r}{r! \langle x \rangle}
\right|
\left|
\left(1+xy\langle x \rangle^{-1}
\right)
\tau^{\prime}
\left(
\cfrac{y}{\langle x \rangle}
\right)
\right|.
\nonumber
\end{eqnarray}
Let $A$, $B \subset \mathbb{R}^2$ be
\begin{eqnarray}
A&=&
\left\{
(x,y) \in \mathbb{R}^2
\mid
\left|
\frac{y}{\langle x \rangle}
\right|
\leq 2
\right\}
,
B=
\left\{
(x,y) \in \mathbb{R}^2
\mid
1 \leq
\left|
\frac{y}{\langle x \rangle}
\right|
\leq 2
\right\}.
\nonumber
\end{eqnarray}
Then we see that there exists $C_2^{\prime}=C_2^{\prime}(\tau)>0$ such that
\begin{eqnarray}
\left|
\cfrac{\partial \tilde{f}}{\partial \overline{z}}(z)
\right|
\leq
\cfrac{1}{2}
\left|
f^{(n+1)}(x)
\cfrac{(iy)^n}{n!}
\right|
\mathbbm{1}_{A}(x,y)
+
C_2^{\prime}
\left|
\sum_{r=0}^{n}
f^{(r)}(x)\cfrac{(iy)^r}{r! \langle x \rangle}
\right|
\mathbbm{1}_{B}(x,y).
\nonumber
\end{eqnarray}
Let $i, j \in \mathbb{N}$. Then we have
\begin{eqnarray}
|(\delta_i,f(H)\delta_j)|
&
\leq
&
\cfrac{1}{2\pi}
\int_{\mathbb{C}}
\left|
f^{(n+1)}(x)
\cfrac{(iy)^n}{n!}
\right|
|
(\delta_i, (H-z)^{-1} \delta_j)
|
\mathbbm{1}_{A}(x,y)
dxdy
\nonumber
\\
&+&
\cfrac{C_2^{\prime}}{\pi}
\int_{\mathbb{C}}
\left|
\sum_{r=0}^{n}
f^{(r)}(x)\cfrac{(iy)^r}{r! \langle x \rangle}
\right|
|
(\delta_i, (H-z)^{-1} \delta_j)
|
\mathbbm{1}_{B}(x,y)
dxdy.
\label{ker17}
\end{eqnarray}
Let $\gamma_z= \cfrac{1}{\sqrt{\eta_z+|z|}+1}<1$ and
$\alpha_z=\alpha_z(\gamma_z)$.
By Lemma $\ref{6.2}$, we see that
\begin{eqnarray}
&&
\int_{\mathbb{C}}
\left|
\sum_{r=0}^{n}
f^{(r)}(x)\cfrac{(iy)^r}{r! \langle x \rangle}
\right|
|
(\delta_i, (H-z)^{-1} \delta_j)
|
\mathbbm{1}_{B}(x,y)
dxdy
\nonumber
\\
&&\leq
\sum_{r=0}^{n}
\int_{\mathbb{C}}
\left|
f^{(r)}(x)\cfrac{(iy)^r}{r! \langle x \rangle}
\right|
\alpha_z^{-|i-j|}
\cfrac{1}{\eta_z}
\left(
\cfrac{1+\gamma_z}{1-\gamma_z}
\right)^2
\mathbbm{1}_{B}(x,y)
dxdy.
\label{ker6}
\end{eqnarray}
We estimate the lower bound of $\alpha_z$.
Suppose that $(x,y)\in B$, then $1\leq |y|\leq \eta_z$ and $|z|\leq \sqrt{2}|y|$.
Therefore
\begin{eqnarray}
\gamma_z m_z
=\cfrac{\eta_z}{
\eta_z+|z|+1
+2\sqrt{\eta_z+|z|}
}
\geq
\cfrac{1}{2+\sqrt{2}+2\sqrt{1+\sqrt{2}}}
\label{lkj}
\end{eqnarray}
Let
$b=\cfrac{1}{2+\sqrt{2}+2\sqrt{1+\sqrt{2}}}$.
By the definition of $\alpha_z$,
$(\ref{lkj})$ implies that
\begin{eqnarray}
\alpha_z \geq
\cfrac{1}{4}
\left(
b+\sqrt{b^2+16}
\right)
>1.
\label{ker8}
\end{eqnarray}
Let $B=\cfrac{1}{4}
\left(
b+\sqrt{b^2+16}
\right).$
We see that
\begin{eqnarray}
\cfrac{1+\gamma_z}{1-\gamma_z}
=1+
\cfrac{2}{\sqrt{\eta_z+|z|}}
\leq1+\sqrt{2}.
\label{ker9}
\end{eqnarray}
By $(\ref{ker6})$, $(\ref{ker8})$, and $(\ref{ker9})$, we see that
\begin{eqnarray}
&&
\int_{\mathbb{C}}
\left|
\sum_{r=0}^{n}
f^{(r)}(x)\cfrac{(iy)^r}{r! \langle x \rangle}
\right|
|
(\delta_i, (H-z)^{-1} \delta_j)
|
\mathbbm{1}_{B}(x,y)
dxdy
\nonumber
\\
&&
\leq
\left(
1+\sqrt{2}
\right)^2
B^{-|i-j|}
\sum_{r=0}^{n}
\cfrac{1}{r!}
\int_{\mathbb{C}}
|f^{(r)}(x)|
\left|
\cfrac{y^{r-1}}{\langle x \rangle}
\right|
\mathbbm{1}_{B}(x,y)
dxdy
\nonumber
\\
&&
\leq
\left(
1+\sqrt{2}
\right)^2
B^{-|i-j|}
\sum_{r=0}^{n}
\cfrac{2^{r-1}}{r!}
\int_{\mathbb{C}}
|f^{(r)}(x)|
\langle x \rangle^{r-2}
\mathbbm{1}_{B}(x,y)
dxdy
\nonumber
\\
&&
\leq
\left(
1+\sqrt{2}
\right)^2
B^{-|i-j|}
\sum_{r=0}^{n}
\int_{\mathbb{R}}
|f^{(r)}(x)|
\langle x \rangle^{r-1}
dx.
\label{ker17.1}
\end{eqnarray}
By Lemma $\ref{6.2}$, we see that
\begin{eqnarray}
&&
\int_{\mathbb{C}}
\left|
f^{(n+1)}(x)
\cfrac{(iy)^n}{n!}
\right|
|
(\delta_i, (H-z)^{-1} \delta_j)
|
\mathbbm{1}_{A}(x,y)
dxdy
\nonumber
\\
&&
\leq
\int_{\mathbb{C}}
\left|
f^{(n+1)}(x)
\cfrac{(iy)^n}{n!}
\right|
\alpha_z^{-|i-j|}
\cfrac{1}{\eta_z}
\left(
\cfrac{1+\gamma_z}{1-\gamma_z}
\right)^2
\mathbbm{1}_{A}(x,y)
dxdy.
\label{kerq}
\end{eqnarray}
Note that for any $k \in \mathbb{Z}_{\geq 0}$ and $t>0$
\begin{equation}
e^{-t}\leq
\cfrac{e^{-k}k^k}{t^k}.
\nonumber
\end{equation}
This implies that, for $i,j \in \mathbb{N}$ with $i\neq j$,
\begin{eqnarray}
\alpha_z^{-|i-j|}
\leq
\left(
1+
\left(
\frac{\gamma_z m_z}{4}
\right)^2
\right)^{-\frac{|i-j|}{2}}
\leq
\cfrac{e^{-k}(2k)^k}{
|i-j|^k
\left(
{\rm log}
\left(
1+
\left(
\frac{\gamma_z m_z}{4}
\right)^2
\right)
\right)^k
}.
\label{ker11}
\end{eqnarray}
Suppose that $(x,y) \in A$, then
$|y|\leq 2\langle x\rangle$ and $|z|\leq \sqrt{5}\langle x \rangle$. We see that
\begin{eqnarray}
\gamma_z m_z
\geq
\cfrac{1}{2}\:
\cfrac{\eta_z}{\eta_z+|z|+1}
\geq
\cfrac{1}{2}\:
\cfrac{|y|}{|y|+|z|+1}
\geq
\cfrac{3-\sqrt{5}}{8}\:
\cfrac{|y|}{\langle x \rangle}.
\label{ker12}
\end{eqnarray}
$(\ref{ker11})$ and $(\ref{ker12})$ imply
\begin{equation}
\alpha_z^{-|i-j|}
\leq
\cfrac{e^{-k}(2k)^k}{
|i-j|^k
\left(
{\rm log}
\left(
1+
\left(
\frac{3-\sqrt{5}}{32}
\;
\frac{|y|}{\langle x \rangle}
\right)^2
\right)
\right)^k
}.
\label{ker13}
\end{equation}
We see that
\begin{equation}
\cfrac{1+\gamma_z}{1-\gamma_z}\leq
1+\sqrt{\frac{2}{|y|}}.
\label{ker14}
\end{equation}
By $(\ref{kerq})$, $(\ref{ker13})$, and $(\ref{ker14})$, we see that
\begin{eqnarray}
&&
\int_{\mathbb{C}}
\left|
f^{(n+1)}(x)
\cfrac{(iy)^n}{n!}
\right|
|
(\delta_i, (H-z)^{-1} \delta_j)
|
\mathbbm{1}_{A}(x,y)
dxdy
\nonumber
\\
&&\leq
\cfrac{2e^{-k}(2k)^k}{n! |i-j|^k}
\int_{\mathbb{C}}
\cfrac{|f^{(n+1)}(x)||y|^{n-1}}{
\left(
{\rm log}
\left(
1+
\left(
\frac{3-\sqrt{5}}{32}
\;
\frac{|y|}{\langle x \rangle}
\right)^2
\right)
\right)^k
}
\left(
1+\frac{2}{|y|}
\right)
\mathbbm{1}_{A}(x,y)
dxdy
\nonumber
\\
&&\leq
\cfrac{8e^{-k}(2k)^k}{n! |i-j|^k}
\int_{\mathbb{R}}
dx|f^{(n+1)}(x)|
\int_{0}^{2\langle x \rangle}dy
\cfrac{|y|^{n-1}+|y|^{n-2}}{
\left(
{\rm log}
\left(
1+
\left(
\frac{3-\sqrt{5}}{32}
\;
\frac{|y|}{\langle x \rangle}
\right)^2
\right)
\right)^k
}
\nonumber
\\
&&\leq
\cfrac{8e^{-k}(2k)^k}{n! |i-j|^k}
\int_{\mathbb{R}}
dx|f^{(n+1)}(x)|
\langle x \rangle^{n}
\int_{0}^{2}dt
\cfrac{t^{n-1}+t^{n-2}}{
\left(
{\rm log}
\left(
1+
\left(
\frac{3-\sqrt{5}}{32}
\;
t
\right)^2
\right)
\right)^k
}.
\label{ker15}
\end{eqnarray}
If $n>2k+1$,
\begin{eqnarray}
C_2^{\prime \prime}(n)
\coloneqq
\int_{0}^{2}dt
\cfrac{t^{n-1}+t^{n-2}}
{
\left(
{\rm log}
\left(
1+
\left(
\frac{3-\sqrt{5}}{32}
\;
t
\right)^2
\right)
\right)^k
}<\infty.
\label{ker16}
\end{eqnarray}
Let $n=2k+2$. Then $(\ref{ker17})$, $(\ref{ker17.1})$, $(\ref{ker15})$, and $(\ref{ker16})$ imply that there exists $C_2=C_2(k)>0$ such that for any $i,j \in \mathbb{N}$,
\begin{eqnarray}
&&|(\delta_i,f(H)\delta_j)|
\nonumber
\\
&&\leq
\cfrac{C_2^{\prime}}{\pi}
\left(
1+\sqrt{2}
\right)^2
B^{-|i-j|}
\opnorm{f}_{2k+2}
+
\cfrac{1}{2\pi}
C_2^{\prime\prime}(2k+2)
\cfrac{8e^{-k}(2k)^k}{(2k+2)! |i-j|^k}
\opnorm{f}_{2k+3}
\nonumber
\\
&&\leq
C_2\opnorm{f}_{2k+3}\langle i-j \rangle^{-k}.
\nonumber
\end{eqnarray}
This implies our assertion.
\qed
\end{Proof}
\section{Intermittency function and proof of the main result}
In this section, we mainly consider the distribution of $a_{\psi}^{(k)}(n,T)$ and estimate the lower and upper bounds the momentum $\langle |X|^p\rangle_{\psi}^{(k)}(T)$.
From this, we calculate the intermittency function exactly.
Finally, we prove Theorem $\ref{main}$ by using the intermittency function.
\subsection{Lower bound of intermittency function}
Let $k\in \mathbb{N}$, $\psi \in l^2(\mathbb{N})$, and $T > 0$. We define for $S\in2^{\mathbb{N}}$,
\begin{eqnarray}
P_{\psi}^{(k)}(S, T)= \sum_{n \in S}^{}a_{\psi}^{(k)}(n,T).
\nonumber
\end{eqnarray}
For $M \geq N \geq1$, let subsets $\{N \sim M \}$ and $\{M \sim\infty \}$ of $\mathbb{N}$ be
\begin{eqnarray}
\{N \sim M \}&=&\{n\in \mathbb{N} \mid N \leq n\leq M\},
\nonumber
\\
\{M \sim\infty \}&=&\{n\in \mathbb{N} \mid n\geq M\}.
\nonumber
\end{eqnarray}
\begin{Lemma}\label{6.1.1}
Let $T>0$ and $\epsilon>0$.
Suppose that $B\in \mathcal{B}^1$ and $A \coloneqq \mu_{\psi}^{(k)}(B)>0$.
Let
\begin{eqnarray}
M_T&=&
\cfrac{A^2}{16J^{(k)}_{\psi}(T^{-1},B)},
\nonumber
\\
J_{\psi}^{(k)}(\epsilon, B)&=& \int_{B}\mu_{\psi}^{(k)}(dx)
\int_{\mathbb{R}}\mu_{\psi}^{(k)}(dy)
\cfrac{\epsilon^2}{(x-y)^2+\epsilon^2}.
\nonumber
\end{eqnarray}
Then for any $T>0$
\[
P_{\psi}^{(k)}(\:\{M_T \sim \infty\}, T)\geq \cfrac{A}{2}>0.
\]
\end{Lemma}
\begin{Proof}\rm
We denote $P_{\psi}^{(k)}$, $\mu_{\psi}^{(k)}$, and $J_{\psi}^{(k)}$ by $P_{\psi}$, $ \mu_{\psi}$, and $J_{\psi}$, respectively for simplicity of notation.
Let $\rho = E^{(k)}(B)\psi$ and $\chi = \psi - \rho$.
Note that $\rho \neq 0$. We see that
\begin{eqnarray}
P_{\psi}(\{1 \sim M\}, T)&=&
\sum_{n=1}^{M}\cfrac{1}{T}\int_{\mathbb{R}}dt\:e^{-\frac{t}{T}}|\chi(n,t)+\rho(n,t)|^2
\nonumber
\\
&=&
P_{\chi}(\{1 \sim M\},T)+
P_{\rho}(\{1 \sim M\},T)
+2\sum_{n=1}^{M}\cfrac{1}{T}\int_{\mathbb{R}}dt\:e^{-\frac{t}{T}}
{\rm Re}(\chi(n,t) \overline{\rho(n,t)})
\nonumber
\\
&=&
P_{\chi}(\{1 \sim M\},T)-
P_{\rho}(\{1 \sim M\},T)
+2\sum_{n=1}^{M}\cfrac{1}{T}\int_{\mathbb{R}}dt\:e^{-\frac{t}{T}}
{\rm Re}(\psi(n,t)\overline{\rho(n,t)})
\nonumber
\\
&\leq&
P_{\chi}(\{1 \sim M\},T)
+2\sum_{n=1}^{M}\cfrac{1}{T}\int_{\mathbb{R}}dt\:e^{-\frac{t}{T}}
{\rm Re}
(\psi(n,t)\overline{\rho(n,t)}).
\label{OIU}
\end{eqnarray}
Since $P_{\psi}(\{M \sim \infty \},T)=\|\psi\|^2-P_{\psi}(\{1,M-1\},T)$ and $\|\psi\|^2=\|\rho\|^2+\|\chi\|^2$,
$( \ref{OIU})$ implies that
\begin{equation}
P_{\psi}(\{M \sim \infty \},T)\geq
\|\rho\|^2-2|D(M-1,T)|,
\label{7.2}
\end{equation}
where
\begin{equation}
D(M,T)=
\sum_{n=1}^{M}\cfrac{1}{T}\int_{\mathbb{R}}dt\:e^{-\frac{t}{T}}
\psi(n,t)\overline{\rho(n,t)}
=
\sum_{n=1}^{M}\cfrac{1}{T}\int_{\mathbb{R}}dt\:e^{-\frac{t}{T}}
(\delta_n,\psi(t))(\rho(t),\delta_n).
\nonumber
\end{equation}
Since $U_{\psi}:\mathcal{H}_{\psi}\rightarrow L^2(\mathbb{R}, d\mu_{\psi}^{(k)})$ is unitary, by Schwarz inequality we see that
\begin{eqnarray}
&&|D(M,T)|
\nonumber
\\
&&=
\left|
\sum_{n=1}^{M}
\cfrac{1}{T}\int_{\mathbb{R}}dt\; e^{-\frac{t}{T}}
\int_{\mathbb{R}}\mu_{\psi}(dx)
\int_{B}\mu_{\psi}(dy)
e^{-it(x-y)}
\overline{U_{\psi}\delta_n(x)}U_{\psi}\delta_n(y)
\right|
\nonumber
\\
&&=
\left|
\int_{\mathbb{R}}\mu_{\psi}(dx)
\int_{B}\mu_{\psi}(dy)\:
\cfrac{1-iT(x-y)}{1+T^2(x-y)^2}
\:
\sum_{n=1}^{M}
\overline{U_{\psi}\delta_n(x)}U_{\psi}\delta_n(y)
\right|
\nonumber
\\
&&\leq
\int_{\mathbb{R}}\mu_{\psi}(dx)
\int_{B}\mu_{\psi}(dy)\:
\cfrac{1}{\sqrt{1+T^2(x-y)^2}}
\left|
\sum_{n=1}^{M}
\overline{U_{\psi}\delta_n(x)}U_{\psi}\delta_n(y)
\right|
\nonumber
\\
&&\leq
\int_{B}\mu_{\psi}(dy)
\left(
\int_{\mathbb{R}}
\cfrac{\mu_{\psi}(dx)}{1+T^2(x-y)^2}
\right)^{\frac{1}{2}}
\left(
\int_{\mathbb{R}}
\mu_{\psi}(dx)
\left|
\sum_{n=1}^{M}
\overline{U_{\psi}\delta_n(x)}U_{\psi}\delta_n(y)
\right|^2
\right)^{\frac{1}{2}}.
\nonumber
\\
&&\leq
\left(
\int_{B}\mu_{\psi}(dy)
\int_{\mathbb{R}}
\cfrac{\mu_{\psi}(dx)}{1+T^2(x-y)^2}
\right)^{\frac{1}{2}}
\left(
\int_{B}\mu_{\psi}(dy)
\int_{\mathbb{R}}
\mu_{\psi}(dx)
\left|
\sum_{n=1}^{M}
\overline{U_{\psi}\delta_n(x)}U_{\psi}\delta_n(y)
\right|^2
\right)^{\frac{1}{2}}.
\label{7.3}
\end{eqnarray}
Since $U_{\psi}:\mathcal{H}_{\psi}\rightarrow L^2(\mathbb{R}, d\mu_{\psi}^{(k)})$ is unitary, we have
\begin{eqnarray}
\int_{\mathbb{R}}
\mu_{\psi}(dx)
\left|
\sum_{n=1}^{M}
U_{\psi}\delta_n(x)
\overline{
U_{\psi}\delta_n(y)}
\right|^2
&=&
\int_{\mathbb{R}}
\mu_{\psi}(dx)
\left|
U_{\psi}\left(
\sum_{n=1}^{M}
\overline{
U_{\psi}\delta_n(y)}
\delta_n
\right)
(x)
\right|^2
\nonumber
\\
&=&
\left\|
\sum_{n=1}^{M}
\overline{
U_{\psi}\delta_n(y)}
\delta_n
\right\|^2_{l^2(\mathbb{N})}
\nonumber
\\
&=&
\sum_{n=1}^{M}
\left|
U_{\psi}\delta_n(y)
\right|^2.
\label{7.4}
\end{eqnarray}
By $(\ref{7.3})$ and $(\ref{7.4})$, we see that
\begin{eqnarray}
|D(M-1,T)|^2&\leq&
\left(
\int_{B}\mu_{\psi}(dy)
\int_{\mathbb{R}}
\cfrac{\mu_{\psi}(dx)}{1+T^2(x-y)^2}
\right)
\left(
\sum_{n=1}^{M}
\int_{B}\mu_{\psi}(dy)
\left|
U_{\psi}\delta_n(y)
\right|^2
\right)
\nonumber
\\
&\leq&
J_{\psi}(T^{-1},B)\sum_{n=1}^{M}
\|U_{\psi}\delta_n\|_{l^2}^2
\nonumber
\\
&\leq&
MJ_{\psi}(T^{-1},B)
\nonumber
\end{eqnarray}
Let $M=M_T$. Then
\begin{equation}
|D(M_T-1,T)|\leq
\sqrt{ M_TJ_{\psi}(T^{-1},B)}
=
\cfrac{\|\rho\|^2}{4}.
\label{7.6}
\end{equation}
By $(\ref{7.2})$ and $(\ref{7.6})$, we obtain that
\begin{equation}
P_{\psi}(\{ M_T\sim \infty\}, T) \geq \frac{\|\rho \|^2}{2}=\frac{A}{2}.
\nonumber
\end{equation}
This implies our assertion.
\qed
\end{Proof}
For $\psi \in l^2(\mathbb{N})$,
let an analytic function $m_{\psi}^{(k)}: \mathbb{C}^+ \rightarrow \mathbb{C}^+$ be defined by
\begin{eqnarray}
m_{\psi}^{(k)}(z) = \int_{\mathbb{R}}\cfrac{\mu_{\psi}^{(k)}(d\lambda)}{\lambda-z}
=(\psi, (H^{(k)}-z)^{-1}\psi).
\nonumber
\end{eqnarray}
Let $\epsilon>0$ and $B\in \mathcal{B}^1$.
We define
\[
I_{\psi}^{(k)}(\epsilon,B)=\epsilon
\int_{B}dE |{\rm Im}\: m_{\psi}^{(k)}(E+i\epsilon)|^2.
\]
\begin{Lemma}\label{pro612}
Let $B=[a,b]\subset \mathbb{R}$. Then there exists $C_3=C_3(a,b)>0$ such that
for any $\epsilon \in(0,1)$
\begin{equation}
J_{\psi}^{(k)}(\epsilon, B)\leq C_3 I_{\psi}^{(k)}(\epsilon,B).
\label{RDX}
\end{equation}
\end{Lemma}
\begin{Proof}\rm
We denote $J_{\psi}^{(k)}$, $I_{\psi}^{(k)}$, and $\mu_{\psi}^{(k)}$ by
$J_{\psi}$, $I_{\psi}$, and $\mu_{\psi}$, respectively for simplicity of notation. We see that
\begin{eqnarray}
I_{\psi}(\epsilon,B)
&=&
\epsilon^3\int_{B}dE
\left(
\int_{\mathbb{R}}
\cfrac{\mu_{\psi}(dx)}{\epsilon^2+(E-x)^2}
\right)^2
\nonumber
\\
&=&
\epsilon^3\int_{B}dE
\int_{\mathbb{R}}
\cfrac{\mu_{\psi}(dx)}{\epsilon^2+(E-x)^2}
\int_{\mathbb{R}}
\cfrac{\mu_{\psi}(dy)}{\epsilon^2+(E-y)^2}
\nonumber
\\
&\geq&
\int_{B}\mu_{\psi}(dx)
\int_{\mathbb{R}}\mu_{\psi}(dy)\:
\epsilon^3
\int_B
\cfrac{dE}{(\epsilon^2+(E-x)^2)(\epsilon^2+(E-y)^2)}.
\label{7.7}
\end{eqnarray}
Let $s= \frac{x-y}{\epsilon}$. Since
$x \in B=[a,b]$ and $0<\epsilon<1$, we have
\begin{eqnarray}
\epsilon^3
\int_B
\cfrac{dE}{(\epsilon^2+(E-x)^2)(\epsilon^2+(E-y)^2)}
&=&
\int_{\frac{a-x}{\epsilon}}^{\frac{b-x}{\epsilon}}\cfrac{dt}{(1+t^2)(1+(t+s)^2)}
\nonumber
\\
&\geq&
\int_{a-x}^{b-x}\cfrac{dt}{(1+t^2)(1+(|t|+|s|)^2)}.
\nonumber
\end{eqnarray}
If $|s|\leq1$, then
\begin{eqnarray}
\int_{a-x}^{b-x}\cfrac{dt}{(1+t^2)(1+(|t|+|s|)^2)}
&\geq&
\int_{a-x}^{b-x}\cfrac{dt}{(1+t^2)(1+(|t|+1)^2)}.
\label{TY}
\end{eqnarray}
If $|s|\geq 1$, then there exists $C_3^{\prime}=C_3^{\prime}(a,b)>0$ such that for any $x\in B$,
\begin{eqnarray}
\int_{a-x}^{b-x}\cfrac{dt}{(1+t^2)(1+(|t|+|s|)^2)}
&=&
\int_{a-x}^{b-x}
dt\:
\left(
\cfrac{1}{1+t^2}-
\cfrac{1}{1+(|t|+|s|)^2}
\right)
((|t|+|s|)^2-t^2)^{-1}
\nonumber
\\
&\geq&
\int_{a-x}^{b-x}
dt\:
\left(
\cfrac{1}{1+t^2}-
\cfrac{1}{1+(|t|+1)^2}
\right)
\cfrac{C_3^{\prime}}{1+s^2}
\label{7.10}
\end{eqnarray}
By $(\ref{TY})$ and $(\ref{7.10})$, there exists $C_3=C_3(a,b)>0$ such that for any $x\in B$ and any $y\in \mathbb{R}$,
\begin{equation}
\epsilon^3
\int_B
\cfrac{dE}{(\epsilon^2+(E-x)^2)(\epsilon^2+(E-y)^2)}
\geq
\cfrac{C_3}{1+s^2},
\qquad s=\frac{x-y}{\epsilon} .
\label{7.11}
\end{equation}
$(\ref{7.7})$ and $(\ref{7.11})$ imply our assertion.
\qed
\end{Proof}
\begin{Definition}
Let $f:\mathbb{Z}_{\geq 0} \rightarrow \mathbb{C}$ and
$n\in \mathbb{N}$.
Let $\left(\tilde{H}^{(k)}f \right) (n)$ be defined by
\[
\left( \tilde{H}^{(k)}f \right)(n)
= -a_k(n)f({n+1})+d_k(n)f(n)-a_k({n-1})f({n-1}),
\]
where $a_k(0)=1$.
\end{Definition}
Let $z \in \mathbb{C}^+$, and $n,m \in \mathbb{N}$ such that $n\geq m$. We define
\begin{eqnarray}
T_z(n)
&=&
\begin{cases}
\left(
\begin{array}{cc}
0&1\\
-\sqrt{\frac{g_{n-1}}{g_n}}&\frac{g_n+1-z}{\sqrt{g_n}}\\
\end{array}
\right)
&
(n \geq 1),
\\
\left(
\begin{array}{cc}
0&1\\
-1&1-z\\
\end{array}
\right)
&(n=0),
\end{cases}
\nonumber
\\
S_z(n,m)&=& T_z(n)T_z(n-1)\cdots T_z(m),
\nonumber
\\
S_z(n)&=& S_z(n,0).
\nonumber
\end{eqnarray}
Let $f : \mathbb{Z}_{\geq 0}\rightarrow \mathbb{C}$ and $z\in \mathbb{C}^+$.
Suppose that $\left( \tilde{H}^{(k)}f \right) (n)=zf(n)$ for each $n \in \mathbb{N}$. Then
\begin{eqnarray}
\left(
\begin{array}{c}
f(n)\\
f(n+1)
\end{array}
\right)
&=&
T_z(n+N(k)-1)
\left(
\begin{array}{c}
f(n-1)\\
f(n)
\end{array}
\right)
\nonumber
\\
&=&
S_z(n+N(k)-1, N(k))
\left(
\begin{array}{c}
f(0)\\
f(1)
\end{array}
\right).
\nonumber
\end{eqnarray}
\begin{comment}
\begin{Proof}
\rm
We see that
\begin{eqnarray}
\left(
\begin{array}{c}
f(n)\\
f(n+1)
\end{array}
\right)=
\left(
\begin{array}{cc}
0&1\\
-\frac{a_k(n-1)}{a_k(n)}&\frac{d_k(n)-z}{a(n)}\\
\end{array}
\right)
\left(
\begin{array}{cc}
f(n-1)\\
f(n)
\end{array}
\right).
\nonumber
\end{eqnarray}
By Definition $\ref{Jacobi}$, this implies our assertion.
\qed
\end{Proof}
\end{comment}
\begin{Lemma}\label{7.1.4}
Let $K>0$ and $z=E+i \epsilon$ with $0<E<4$ and $\epsilon>0$.
Then there exists $C_4=C_4(E, K)>0$ such that
\begin{enumerate}[$(1)$]
\item
if $L_m +1\leq n < L_{m+1}$ and $n \epsilon<K$, then
\begin{equation}
\|S_z(n)^{-1}\|\leq C_4^{m+1}\prod_{j=1}^{m}L_j^{\frac{1-\Gamma}{2\Gamma}},
\nonumber
\end{equation}
\item
if $n \leq L_{m}$ and $n \epsilon<K$, then
\begin{equation}\label{7.15}
\|S_z(n)^{-1}\|\leq C_4^{m}\prod_{j=1}^{m-1}L_j^{\frac{1-\Gamma}{2\Gamma}}.
\nonumber
\end{equation}
\end{enumerate}
\end{Lemma}
\begin{Proof}\rm
By Definition $\ref{sparse}$, we see that
\begin{equation}\label{7.16}
T_z(n)=
\begin{cases}
\left(
\begin{array}{cc}
0&1\\
-[n^{\frac{1-\Gamma}{\Gamma}}]^{-\frac{1}{2}}&
\frac{[n^{\frac{1-\Gamma}{\Gamma}}]+1-z}{[n^{\frac{1-\Gamma}{\Gamma}}]^{\frac{1}{2}}}\\
\end{array}
\right)
&(n\in \{L_m \mid m\in \mathbb{N}\}),\\
\left(
\begin{array}{cc}
0&1\\
-[n^{\frac{1-\Gamma}{\Gamma}}]^{\frac{1}{2}}&2-z\\
\end{array}
\right)
&(n\in \{L_m+1 \mid m\in \mathbb{N}\}),\\
\left(
\begin{array}{cc}
0&1\\
-1&2-z\\
\end{array}
\right)
&({\rm otherwise}).
\end{cases}
\nonumber
\end{equation}
If $L_m+1 \leq n < L_{m+1}$, then
\begin{eqnarray}
S_z(n)
&=&
R^{n-L_m-1}
S(L_m+1)\nonumber
\\
&=&
R^{n-L_m-1}
S(L_m+1,L_m)
S(L_m-1)
\nonumber
\\
&=&
R^{n-L_m-1}
S(L_m+1,L_m)
R^{L_m-L_{m-1}-2}
S(L_{m-1}+1)
\nonumber
\\
&=&\cdots
\nonumber
\\
&=&
R^{n-L_m-1}
S(L_m+1,L_m)
R^{L_m-L_{m-1}-2}
\cdots
S(L_1+1,L_1)
R^{2}
\nonumber
\label{7.17}
\nonumber
\end{eqnarray}
where
$\begin{displaystyle}
R
=
\left(
\begin{array}{cc}
0&1\\
-1&2-z\\
\end{array}
\right).
\end{displaystyle}$
Let $L_0=-2$. Then we see that
\begin{equation}
\|S_z(n)^{-1}\|
\leq
\|R^{-n+L_m+1}\|
\|T_z(0)\|
\prod_{j=1}^{m}
\|S(L_j+1,L_j)^{-1}\|
\|R^{-L_j+L_{j-1}+2}\|.
\label{7.18}
\end{equation}
Note that
$\begin{displaystyle}
R^{-1}
=
\left(
\begin{array}{cc}
2-z&-1\\
1&0\\
\end{array}
\right)
\end{displaystyle}$
and $\|R^m\|=\|R^{-m}\|$ for any $m\in \mathbb{N}$.
\begin{eqnarray}
R=
\left(
\begin{array}{cc}
0&1\\
-1&2-E\\
\end{array}
\right)
-i
\left(
\begin{array}{cc}
0&0\\
0&\epsilon\\
\end{array}
\right).
\nonumber
\end{eqnarray}
Since $E\in(0,4)$, there exist invertible matrix $A_E$ and $\lambda_{\pm}\in \mathbb{C}$ with
$|\lambda_{\pm}|=1$
such that
\begin{equation}
A_E^{-1}
\left(
\begin{array}{cc}
0&1\\
-1&2-E\\
\end{array}
\right)
A_E
=
\left(
\begin{array}{cc}
\lambda_+&0\\
0&\lambda_-\\
\end{array}
\right).
\label{7.20}
\nonumber
\end{equation}
Therefore we see that
\begin{eqnarray}
\|A_E^{-1}R^nA_E\|
&\leq&
\left\|
\left(
\begin{array}{cc}
\lambda_+&0\\
0&\lambda_-\\
\end{array}
\right)
-
i\epsilon
A_E^{-1}
\left(
\begin{array}{cc}
0&0\\
0&1\\
\end{array}
\right)
A_E
\right\|^n
\nonumber
\\
&\leq&
\left(
1+\epsilon\|A_E^{-1}\|\|A_E\|
\right)^n.
\label{7.21}
\nonumber
\end{eqnarray}
If $\epsilon<\frac{K}{n}$, then
\begin{equation}
\|A_E^{-1}R^nA_E\|
\leq
2{\rm exp}(K\|A_E^{-1}\|\|A_E\|).
\nonumber
\end{equation}
Therefore we obtain that for $\epsilon<\frac{K}{n}$,
\begin{eqnarray}
\|R^n\|=\|A_EA_E^{-1}R^nA_EA_E^{-1}\|
\leq 2\|A_E\|\|A_E^{-1}\|{\rm exp}(K\|A_E^{-1}\|\|A_E\|).
\label{SD}
\end{eqnarray}
If $0<\epsilon<1$, then there exists $C_4^{\prime}>0$ such that
\begin{equation}
\|S_z(L_j+1,L_j)^{-1}\|
\leq
C_4^{\prime}[L_j^{\frac{1-\Gamma}{\Gamma}}]^{\frac{1}{2}}
\leq
C_4^{\prime}L_j^{\frac{1-\Gamma}{2\Gamma}}.
\label{7.23}
\end{equation}
By $(\ref{7.18})$, $(\ref{SD})$ and $(\ref{7.23})$, we see that for $L_m +1\leq n < L_{m+1}$,
\begin{equation}
\|S_z(n)^{-1}\|
\leq
(2C_4^{\prime}\|A_E\|\|A_E^{-1}\|)^{m+1}{\exp}((m+1)K\|A_E^{-1}\|\|A_E\|)
\prod_{j=1}^{m}
L_j^{\frac{1-\Gamma}{2\Gamma}}.
\nonumber
\end{equation}
This implies our first part of the assertion. The second part of the assertion can be proved similarly.
\qed
\end{Proof}
\begin{Lemma}\label{7.1.5}
Let $\psi \in l^2(\mathbb{N})$ and $n \in \mathbb{N}$.
Then for any $T>0$,
\begin{equation}
\cfrac{1}{T}\int_{0}^{\infty} e^{-\frac{t}{T}}|\psi_k(t,n)|^2 \:dt
=
\cfrac{\epsilon}{\pi}\int_{\mathbb{R}}|(H^{(k)}-(E+i\epsilon))^{-1}\psi(n)|^2\:dE,
\nonumber
\end{equation}
where $\epsilon=\frac{1}{2T}$.
\end{Lemma}
\begin{Proof}\rm
Abbreviate $H^{(k)}$ and $\psi_k$ to $H$ and $\psi$, respectively. We see that
\begin{eqnarray}
\cfrac{1}{T}\int_{0}^{\infty} e^{-\frac{t}{T}}|\psi(t,n)|^2 \:dt
&=&
\cfrac{1}{T}\int_{0}^{\infty} dt e^{-\frac{t}{T}}(\delta_n, e^{-itH}\psi)(e^{-itH}\psi,\delta_n)
\nonumber\\
&=&
\int_{\mathbb{R}}(\delta_n, E(dx)\psi)
\int_{\mathbb{R}}(E(dy)\psi,\delta_n)
\cfrac{1}{T}\int_{0}^{\infty} dt e^{-\frac{t}{T}-it(x-y)}
\nonumber\\
&=&
\int_{\mathbb{R}}(\delta_n, E(dx)\psi)
\int_{\mathbb{R}}(E(dy)\psi,\delta_n)
(1+iT(x-y))^{-1},
\nonumber
\end{eqnarray}
and
\begin{eqnarray}
&&\cfrac{\epsilon}{\pi}\int_{\mathbb{R}}|(H^{(k)}-(E+i\epsilon))^{-1}\psi(n)|^2\:dE
\nonumber
\\
&&=
\int_{\mathbb{R}}(\delta_n, E(dx)\psi)
\int_{\mathbb{R}}(E(dy)\psi,\delta_n)
\cfrac{\epsilon}{\pi}
\int_{\mathbb{R}}(E-x+i\epsilon)^{-1}(E-y-i\epsilon)^{-1}
\nonumber\\
&&=
\int_{\mathbb{R}}(\delta_n, E(dx)\psi)
\int_{\mathbb{R}}(E(dy)\psi,\delta_n)
(1+iT(x-y))^{-1}.
\nonumber
\end{eqnarray}
These imply our assertion.
\qed
\end{Proof}
\begin{Definition}
Let $f:\mathbb{R} \rightarrow \mathbb{C}$ be measurable and
$B_{\nu}=[\nu, 4-\nu]$ with $0<\nu<1$.
We say that $f$ is the first kind, if there exist
$\nu>0$ and $x_0 \in B_\nu$ such that $f \in C^{\infty}_0(B_\nu)$ and $f(x_0)\neq 0$,
and we say that
$f$ is the second kind, if
$f$ is bounded and there exist $E_0\in (0,4)$ and $\nu>0$ with
$[E_0-\nu,E_0+\nu]\subset B_\nu$ such that
$f \in C^{\infty}([E_0-\nu,E_0+\nu])$ and
$|f(x)|\geq c>0$ for $x\in [E_0-\nu,E_0+\nu]$.
\end{Definition}
\begin{Lemma}\label{7.1.6}
Let $f:\mathbb{R} \rightarrow \mathbb{C}$ be the second kind and $\psi=f(H^{(k)})\delta_1$.
Let $N$ be sufficiently large. Then there exists $C_5=C_5(\nu)>1$ such that
\begin{enumerate}[$(1)$]
\item
if $L_N\leq T \leq \cfrac{L_{N+1}}{4}$, then
\begin{eqnarray}
P_{\psi}^{(k)}(\{T \sim \infty\}, T) \geq C_5^{-(N+1)}T
\left(
\cfrac{1}{T}+I_{\delta_1}^{(k)}(T^{-1}, B_{\nu})
\right)
\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}},
\label{620}
\end{eqnarray}
\item
if $\cfrac{L_N}{4}\leq T \leq L_{N}$, then
\begin{eqnarray}
P_{\psi}^{(k)}(\{T \sim \infty\}, T) \geq C_5^{-(N+1)}L_N
\left(
\cfrac{1}{T}+I_{\delta_1}^{(k)}(T^{-1}, B_{\nu})
\right)
\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}},
\label{621}
\end{eqnarray}
\item
if $\cfrac{L_N}{4} \leq T$, then
\begin{eqnarray}
P_{\psi}^{(k)}(
\{\frac{L_N}{4} \sim \frac{L_N}{2}\}, T) \geq C_5^{-N}L_N
\left(
\cfrac{1}{T}+I_{\delta_1}^{(k)}(T^{-1}, B_\nu)
\right)\prod_{j=1}^{N-1}L_j^{\frac{1-\Gamma}{\Gamma}}.
\label{622}
\end{eqnarray}
\end{enumerate}
\end{Lemma}
\begin{Proof}\rm
Firstly, we prove the lemma in the case of $\psi=\delta_1$.
Let $z \in \mathbb{C}^+$ and $f_k: \mathbb{Z}_{\geq0} \rightarrow \mathbb{C}$ be
\begin{equation}
f_k(n)=
\begin{cases}
(H^{(k)}-z)^{-1}\delta_1(n)&(n \in \mathbb{N}),\\
1&(n=0).
\end{cases}
\nonumber
\end{equation}
Let $g=(H^{(k)}-z)^{-1}\delta_1 \in l^2(\mathbb{N})$. We see that $g(n)=f_k(n)$ for each $n \in \mathbb{N}$ and that
\begin{eqnarray}
&&
(H^{(k)}-z)g(n)=\delta_1(n)
\nonumber\\
&&
\Leftrightarrow
\begin{cases}
-a_k(n)g(n+1)+d_k(n)g(n)-a_k(n-1)g(n-1)-zg(n)=0&(n\geq2)\\
-a_k(1)g(2)+d_k(1)g(1)-zg(1)=1&(n=1)
\end{cases}
\nonumber\\
&&
\Leftrightarrow
\begin{cases}
-a_k(n)f_k(n+1)+d_k(n)f_k(n)-a_k(n-1)f_k(n-1)=zf_k(n)&(n\geq2)\\
-a_k(1)f_k(2)+d_k(1)f_k(1)-1=zf_k(1)&(n=1)
\end{cases}
\nonumber\\
&&
\Leftrightarrow
\begin{cases}
-a_k(n)f_k(n+1)+d_k(n)f_k(n)-a_k(n-1)f_k(n-1)=zf_k(n)&(n\geq2)\\
-a_k(1)f_k(2)+d_k(1)f_k(1)-a_k(0)f_k(0)=zf_k(1)&(n=1)
\end{cases}.
\nonumber
\end{eqnarray}
This implies that $f_k$ satisfies the equation $\left(\tilde{H}^{(k)}f_k\right)(n)=z f_k(n)$ for each $n\in \mathbb{N}$.
We see that
\begin{eqnarray}
\left(
\begin{array}{c}
f_k(n)\\
f_k(n+1)
\end{array}
\right)=
S_z(n+N(k)-1,N(k))
\left(
\begin{array}{c}
f_k(0)\\
f_k(1)
\end{array}
\right).
\label{7.29}
\end{eqnarray}
Note that $f_k(1)=m_{\delta_1}^{(k)}(z)$. Let $z=E+i\epsilon$.
By $(\ref{7.29})$, we obtain that
\begin{equation}
|f_k(n)|^2+|f_k(n+1)|^2
\geq
\cfrac{1+|m_{\delta_1}^{(k)}(E+i\epsilon)|^2}{\|S_z(n+N(k)-1,N(k))^{-1}\|^2}
.
\nonumber
\end{equation}
Suppose that $L_N+1\leq n+N(k) \leq L_{N+1}$ and $\epsilon<\cfrac{K}{n}$.
By Lemma $\ref{7.1.4}$, we see that
\begin{equation}
|f_k(n)|^2+|f_k(n+1)|^2
\geq
C_4^{-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}
\left(
1+|{\rm Im}\;m_{\delta_1}^{(k)}(E+i\epsilon)|^2
\right).
\label{SDF}
\end{equation}
By Lemma $\ref{7.1.5}$, we see that $f_k(n)=(H^{(k)}-z)^{-1}\delta_1(n)$ and that
\[
\cfrac{\epsilon}{\pi}\int_{\mathbb{R}}|(H^{(k)}-(E+i\epsilon))^{-1}\delta_1(n)|^2\:dE
=\cfrac{1}{T}\int_{0}^{\infty} e^{-\frac{t}{T}}|{\delta_1}_k(t,n)|^2 \:dt=
a_{\delta_1}^{(k)}(n,T),
\qquad \epsilon=(2T)^{-1}.
\]
Let $(2T)^{-1}<\cfrac{K}{n}$. By $(\ref{SDF})$, we see that there exist
$C_4$ and $C_4^{\prime}>0$ such that
\begin{eqnarray}
a_{\delta_1}^{(k)}(n,T)+
a_{\delta_1}^{(k)}(n+1,T)
&\geq&
C_4^{-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}
\cfrac{1}{2T}
\int_{B_{\nu}}
dE
\left(
1+|{\rm Im}\;m_{\delta_1}^{(k)}(E+i(2T)^{-1})|^2
\right)
\nonumber
\\
&\geq&
C_4^{-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}
\left(
\cfrac{1}{2T}
+
I_{\delta_1}^{(k)}((2T)^{-1},B_{\nu})
\right)
\nonumber
\\
&\geq&
C_4^{\prime-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right).
\nonumber
\end{eqnarray}
Let $L_N<T<\cfrac{L_{N+1}}{4}$, and $K$ sufficiently large. Then we have
\begin{equation}
\sum_{T\leq n \leq 2T}
a_{\psi}^{(k)}(n,T)+
a_{\psi}^{(k)}(n+1,T)
\geq
C_4^{\prime-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}T
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right),
\nonumber
\end{equation}
and hence
\begin{equation}
P_{\psi}^{(k)}(\{T\sim 2T\},T)
\geq
C_4^{\prime-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}T
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right).
\label{7.33}
\end{equation}
Let $\cfrac{L_N}{4}<T< L_N$. Then we see that
\begin{equation}
P_{\psi}^{(k)}(\{2L_N \sim 3L_N\},T)
\geq
C_4^{\prime-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}L_N
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right).
\nonumber
\end{equation}
Let $\cfrac{L_N}{4} \leq T$. Then we see that
\begin{equation}
P_{\psi}^{(k)}(\{\cfrac{L_N}{4}\sim \cfrac{L_N}{2}\},T)
\geq
C_4^{\prime-N}\prod_{j=1}^{N-1}L_j^{\frac{\Gamma-1}{\Gamma}}
L_N
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right).
\nonumber
\end{equation}
Therefore, we can prove the lemma in the case of $\psi=\delta_1$.
Next we take $g\in C^{\infty}_0([0,4])$ such that $g(x)=1$ on $B_{\frac{\nu}{2}}$.
We prove the lemma in the case of $\psi=g(H^{(k)})\delta_1$.
Let $\chi = \delta_1-\psi$ and $z \in \mathbb{C}^+$. Then
\begin{eqnarray}
|(H^{(k)}-z)^{-1}\psi(n)|^2
\geq
\cfrac{1}{2}|(H^{(k)}-z)^{-1}\delta_1(n)|^2
-|(H^{(k)}-z)^{-1}\chi(n)|^2.
\nonumber
\end{eqnarray}
Let $L_N<T<\cfrac{L_{N+1}}{4}$. Then
we see that, by $(\ref{7.33})$,
\begin{eqnarray}
P_{_\psi}^{(k)}(\{T\sim \infty \},T)
&\geq&
\cfrac{1}{2}\:
C_4^{\prime-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}T
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right)
\nonumber
\\
&-&
\cfrac{1}{T}
\int_{B_{\nu}}dE
\sum_{T\leq n \leq 2T}
|(H^{(k)}-(E+i\epsilon))^{-1}\chi(n)|^2.
\label{FGH}
\end{eqnarray}
Let $f_z(x)= \cfrac{1-g(x)}{x-z}$. Then $(H^{(k)}-z)^{-1}\chi(n)=f_z(H^{(k)})\delta_1(n)$ and
Lemma $\ref{6.2.4}$ implies that for $l>1$,
\begin{eqnarray}
|(H^{(k)}-z)^{-1}\chi(n)|=|f_z(H^{(k)})\delta_1(n)|&\leq& C_2\opnorm{f_z}_{2l+3} n^{-l},
\nonumber
\end{eqnarray}
and that
\begin{eqnarray}
\sum_{T\leq n \leq 2T}|(H^{(k)}-z)^{-1}\chi(n)|=\sum_{T\leq n \leq 2T}|f_z(H^{(k)})\delta_1(n)|&\leq& C_2\opnorm{f_z}_{2l+3} T^{-(l-1)}.
\label{FGH1}
\end{eqnarray}
Let $z=E+i\epsilon$. Note that there exists $C_5^{\prime}=C_5^{\prime}(g,\nu, l)>0$ such that
$\begin{displaystyle}
\sup_{E\in B_{\nu}, 0<\epsilon <1}\opnorm{f_z}_{2l+3} \leq C_5^{\prime}.
\end{displaystyle}$
By $(\ref{FGH})$ and $(\ref{FGH1})$, we obtain
\begin{eqnarray}
P_{_\psi}^{(k)}(n \geq T,T)
&\geq&
\cfrac{1}{2}\:
C_4^{\prime-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}T
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right)
-
\cfrac{4}{T}
\:
C_2C_5^{\prime} T^{-(l-1)}
\nonumber
\\
&\geq&
\left\{
\cfrac{1}{2}\:
C_4^{\prime-(N+1)}
-
4C_2 C_5^{\prime}T^{-l}
\prod_{j=1}^{N}L_j^{\frac{1-\Gamma}{\Gamma}}
\right\}
\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}T
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right).
\label{7.37}
\nonumber
\end{eqnarray}
Let $l$ sufficiently large, then we can prove $(\ref{620})$
in the case of $\psi=g(H^{(k)})$.
We can prove $(\ref{621})$ and $(\ref{622})$ in the case of $\psi=g(H^{(k)})$ similarly.
Finally, let $f$ be the second kind, and we prove the lemma in the case of $\psi=f(H^{(k)})$.
Let $\nu$ satisfy $f \in C^{\infty}([E_0-\nu,E_0+\nu])$ and
$|f(x)|\geq c>0$ for $x\in [E_0-\nu,E_0+\nu]$.
We take $g \in C^{\infty}_0([E_0-\nu,E_0+\nu])$ such that
$g(x)=1$ on $[E_0-\frac{3\nu}{4},E_0+\frac{3\nu}{4}]$.
Then there exists $h\in C^{\infty}_0([E_0-\nu,E_0+\nu])$ such that $g(x)=h(x)f(x)$.
Since $\sup_n \sum_{m=1}^{\infty}
\langle n-m \rangle^{-l}<\infty$, by Lemma $\ref{6.2.4}$, we see that
\begin{eqnarray}
&&|(H^{(k)}-z)^{-1}g(H^{(k)})\delta_1(n)|^2
\nonumber
\\
&&=
|(h(H^{(k)})\delta_n, (H^{(k)}-z)^{-1}f(H^{(k)})\delta_1 )|^2
\nonumber
\\
&&=
\left|
\sum_{m=1}^{\infty}
(h(H^{(k)})\delta_n,\delta_m)(\delta_m,(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1)
\right|^2
\nonumber
\\
&&
\leq
C_2
\left|
\sum_{m=1}^{\infty}
\langle n-m \rangle^{-l}|(\delta_m,(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1)|
\right|^2
\nonumber
\\
&&
\leq
C_2
\left(
\sum_{m=1}^{\infty}
\langle n-m \rangle^{-l}
\right)
\left(
\sum_{m=1}^{\infty}
\langle n-m \rangle^{-l}
|(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1(m)|^2
\right)
\nonumber
\\
&&
\leq
C_2^{\prime}
\sum_{m=1}^{\infty}
\langle n-m \rangle^{-l}
|(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1(m)|^2.
\nonumber
\end{eqnarray}
This implies the inequality
\begin{eqnarray}
A(2L,T)&\coloneqq&
\epsilon
\sum_{n\geq 2L}
\int_{B_\nu}dE\:
|(H^{(k)}-z)^{-1}g(H^{(k)})\delta_1(n)|^2
\nonumber
\\
&=&
C_2^{\prime}\epsilon
\sum_{m=1}^{\infty}\sum_{n\geq 2L}
\langle n-m \rangle^{-l}
\int_{B_\nu}dE\:
|(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&\leq&
\epsilon
\sum_{m=1}^{\infty}
h_l(m,L)
\int_{B_\nu}dE\:
|(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1(m)|^2,
\label{BGT}
\end{eqnarray}
where $z=E+i\epsilon$, $\epsilon=\cfrac{1}{2T}$ and
$\begin{displaystyle}
h_l(m, L)=
\sum_{n\geq 2L}\cfrac{C_2^{\prime}}{1+ |n-m|^l}.
\end{displaystyle}$
It follows for $\phi \in l^2(\mathbb{N})$ and $\epsilon>0$ that
\begin{eqnarray}
\epsilon\sum_{n=1}^{\infty}
\int_{\mathbb{R}}dE
|(H^{(k)}-z)^{-1}\phi(n)|^2
=
\pi
\|\phi\|^2,
\qquad z=E+i\epsilon.
\nonumber
\end{eqnarray}
There exists
$C_2^{\prime \prime}>\max\{C_2^\prime, \sup_{m\geq T}h_l(m,T) \}$.
By $(\ref{BGT})$, we obtain that
\begin{eqnarray}
A(2T,T)
&\leq&
\epsilon
\sum_{m<T}^{}
h_l(m,T)
\int_{B_\nu}dE\:
|(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&+&
\epsilon
\sum_{m\geq T}^{}
h_l(m,T)
\int_{B_\nu}dE\:
|(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&\leq& \pi \|f(H^{(k)})\delta_1\|^2C^{\prime \prime}_2T^{1-l}+
C_2^{\prime\prime}\epsilon\sum_{m \geq T}
\int_{B_\nu}dE\:
|(H^{(k)}-z)^{-1}f(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&=&
\pi \|f(H^{(k)})\delta_1\|^2C_2^{\prime\prime}T^{1-l}+
{C_2^{\prime \prime}}
P_{\psi}^{(k)}(\{T\sim \infty \} , T),
\label{7.38}
\end{eqnarray}
where $z=E+i\epsilon$ and $\epsilon=\cfrac{1}{2T}$.
Let $L_N<T<\cfrac{L_{N+1}}{4}$. Then the previous argument shows that
\begin{eqnarray}
\mbox{
$\begin{displaystyle}
A(2T,T)=\pi P_{g(H^{(k)})\delta_1}^{(k)}(\{2T \sim \infty \}, T)\geq
C_5^{-(N+1)}\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}T
\left(
\cfrac{1}{T}
+
I_{\delta_1}^{(k)}(T^{-1},B_{\nu})
\right).
\end{displaystyle}$
}
\label{7.39}
\end{eqnarray}
We take $l$ sufficiently large. Then $(\ref{7.38})$ and $(\ref{7.39})$ imply
$(\ref{620})$ in the case of $\psi=f(H^{(k)})$.
We can also prove $(\ref{621})$ and $(\ref{622})$ similarly.
\qed
\end{Proof}
\begin{Lemma}\label{618}
Let $f$ be the first kind with $\sup_{x}|f(x)|\leq 1$ and $\psi = f(H^{(k)})\delta_1$.
If $N$ is sufficient large, then
there exist $C_6>0$ and $q_N\in \mathbb{R}$ such that
$\begin{displaystyle}
\lim_{N \rightarrow \infty}q_N=0
\end{displaystyle}$ and it follows for $\cfrac{L_N}{4}<T<\cfrac{L_{N+1}}{4}$ that
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)
&\geq&
C_6I_{\delta_1}^{(k)}(T^{-1}, B_\nu)^{-p}
\nonumber\\
&+&
C_6
\left(
L_N^{p+1+q_N}
+T^{p+1}L_N^{\frac{\Gamma-1}{\Gamma}+q_N}
\right)
I_{\delta_1}^{(k)}(T^{-1},B_\nu).
\nonumber
\end{eqnarray}
\end{Lemma}
\begin{Proof}\rm
Let $M\in \mathbb{N}$. Then it follows that
$\langle |X|^p \rangle_{\psi}^{(k)}(T)
\geq
M^pP_{\psi}^{(k)}(\{M \sim \infty \},T)$.
Lemma $\ref{6.1.1}$ implies that
\[
\langle |X|^p \rangle_{\psi}^{(k)}(T)
\geq
M_T^pP_{\psi}^{(k)}(\{M_T \sim \infty \},T)
\geq C_6^{\prime}J_{\psi}^{(k)}(T^{-1},B_{\nu})^{-p}
\geq C_6^{\prime}J_{\delta_1}^{(k)}(T^{-1},B_{\nu})^{-p}.
\]
By $(\ref{RDX})$, we have
\begin{equation}
\langle |X|^p \rangle_{\psi}^{(k)}(T)\geq C_6^{\prime\prime}I_{\delta_1}^{(k)}(T^{-1},B_{\nu})^{-p}.
\label{WER}
\end{equation}
Note that $f$ is the second kind.
For $\frac{L_N}{4}\leq T \leq L_N$, by $(\ref{621})$ we have
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)&\geq&
T^p
P_{\psi}^{(k)}(\{T \sim \infty\}, T)
\nonumber\\
&\geq&
4C_5^{-(N+1)}T^{p+1}
I_{\delta_1}^{(k)}(T^{-1}, B_{\nu})
\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}.
\nonumber
\end{eqnarray}
Let $q_N>0 $ satisfy
\[
L_N^{q_N}=C_5^{-(N+1)}\prod_{j=1}^{N-1}L_j^{\frac{\Gamma-1}{\Gamma}}.
\]
Then $\lim_{N \rightarrow \infty}q_N=0$ and we see that for $\frac{L_N}{4}\leq T \leq L_N$,
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)\geq
4T^{p+1}
L_N^{\frac{\Gamma-1}{\Gamma}+q_N}
I_{\delta_1}^{(k)}(T^{-1}, B_{\nu}).
\label{REW}
\end{eqnarray}
For $L_N\leq T \leq \cfrac{L_{N+1}}{4}$, by $(\ref{620})$ we have
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)&\geq&
T^p
P_{\psi}^{(k)}(\{T \sim \infty\}, T)
\nonumber\\
&\geq&
C_5^{-(N+1)}T^{p+1}
I_{\delta_1}^{(k)}(T^{-1}, B_{\nu})
\prod_{j=1}^{N}L_j^{\frac{\Gamma-1}{\Gamma}}
\nonumber\\
&\geq&
T^{p+1}
L_N^{\frac{\Gamma-1}{\Gamma}+q_N}
I_{\delta_1}^{(k)}(T^{-1}, B_{\nu}).
\label{XXXX}
\end{eqnarray}
By $(\ref{REW})$ and $(\ref{XXXX})$, we see that for
$\cfrac{L_N}{4}\leq T \leq \cfrac{L_{N+1}}{4}$,
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)\geq
T^{p+1}
L_N^{\frac{\Gamma-1}{\Gamma}+q_N}
I_{\delta_1}^{(k)}(T^{-1}, B_{\nu}).
\label{HFS}
\end{eqnarray}
For $\cfrac{L_N}{4}\leq T$, by $(\ref{622})$ we see that
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)&\geq&
T^p
P_{\psi}^{(k)}(\{T \sim \infty\}, T)
\nonumber\\
&\geq&
C_5^{-N}L_N^{p+1}I_{\delta_1}^{(k)}(T^{-1}, B_{\nu})
\prod_{j=1}^{N-1}L_j^{\frac{\Gamma-1}{\Gamma}}
\nonumber\\
&\geq&
L_N^{p+1+q_N}I_{\delta_1}^{(k)}(T^{-1}, B_{\nu}).
\label{IJN}
\end{eqnarray}
$(\ref{WER})$, $(\ref{HFS})$, and $(\ref{IJN})$ imply our assertion.
\qed
\end{Proof}
\begin{Lemma}
Let $f$ be the first kind with $\sup_{x}|f(x)|\leq 1$ and $\psi = f(H^{(k)})\delta_1$. Then
\begin{eqnarray}
\beta_{\psi}^{(k)}(p)\geq \cfrac{p+1}{p+\frac{1}{\Gamma}}.
\label{WWWWW}
\end{eqnarray}
\end{Lemma}
\begin{Proof}\rm
By Lemma $\ref{618}$, for $x=I_{\delta_1}^{(k)}(T^{-1},B_\nu)$,
we obtain that
\[
\langle |X|^p \rangle_{\psi}^{(k)}(T)
\geq
C_6x^{-p}+
C_6
\left(
L_N^{p+1-q_N}
+T^{p+1}L_N^{\frac{\Gamma-1}{\Gamma}-q_N}
\right)x.
\]
Let $f(x)=x^{-p}+Kx$. Then
$\begin{displaystyle}
\inf_{x>0}f(x)=
c(p)K^{\frac{p}{p+1}}
\end{displaystyle}$, where $c(p)=p^{-\frac{p}{p+1}}+p^{\frac{1}{p+1}}$.
Let $\cfrac{L_N}{4} \leq T \leq \cfrac{L_{N+1}}{4}$. Then there exists $C_6^{\prime}=C_6^{\prime}(p)>0$ such that
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)
&\geq&
c(p)C_6
\left(
L_N^{p+1-q_N}
+T^{p+1}L_N^{\frac{\Gamma-1}{\Gamma}-q_N}
\right)^{\frac{p}{p+1}}
\nonumber
\\
&\geq&
C_6^{\prime}
L_N^{-\frac{p}{p+1}q_N}
\left(L_N^{p+1}+T^{p+1}L_N^{\frac{\Gamma-1}{\Gamma}}\right)^\frac{p}{p+1}.
\nonumber
\end{eqnarray}
For $\cfrac{L_N}{4} \leq T \leq L_N^A$ with $A=\cfrac{p+\frac{1}{\Gamma}}{p+1}$,
we have
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)
\geq
C_6^{\prime}L_N^{-\frac{p}{p+1}q_N}L_N^p
\geq
C_6^{\prime}L_N^{-\frac{p}{p+1}q_N}T^{\frac{p}{A}}.
\label{UHB}
\end{eqnarray}
For $L_N^A \leq T \leq \cfrac{L_{N+1}}{4}$, we have
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)
\geq
C_6^{\prime}L_N^{-\frac{p}{p+1}q_N}T^pL_N^{\frac{\Gamma-1}{\Gamma}\frac{p}{p+1}}
\geq
C_6^{\prime}L_N^{-\frac{p}{p+1}q_N}T^{\frac{p}{A}}.
\label{UHB1}
\end{eqnarray}
$(\ref{UHB})$ and $(\ref{UHB1})$ imply that for sufficiently large $T>0$ and any $\epsilon>0$,
\begin{eqnarray}
\langle |X|^p \rangle_{\psi}^{(k)}(T)\geq
C_6^{\prime}T^{\frac{p}{A}-\epsilon}.
\nonumber
\end{eqnarray}
Therefore we obtain that
\[
\beta_{\psi}^{(k)}(p)=
\cfrac{1}{p}\liminf_{T \rightarrow \infty}\cfrac{{\rm log}\langle |X|^p\rangle_{\psi}^{(k)}(T)}{{\rm log}T}\geq \cfrac{p+1}{p+\frac{1}{\Gamma}}-\cfrac{\epsilon}{p}.
\]
This implies our assertion.
\qed
\end{Proof}
\subsection{Upper bound of intermittency function}
\begin{Lemma}
Let $f$ be the first kind, $\psi = f(H^{(k)})\delta_1$, and $p>0$.
Then there exists $C_7=C_7(p)>0$ such that
for $L_N \leq T \leq L_N^{\frac{1}{\Gamma}}$ with $N$ sufficiently large,
\begin{eqnarray}
\sum_{n \geq 2L_N}n^pa_{\psi}^{(k)}(n,T)\leq C_7T^{p+1}L_N^{-\frac{1}{\Gamma}}.
\end{eqnarray}
\end{Lemma}
\begin{Proof}\rm
We have
\begin{eqnarray}
\sum_{n \geq 2L_m}n^pa_{\psi}^{(k)}(n,T)
=
\sum_{n=2L_m}^{T^3}n^pa_{\psi}^{(k)}(n,T)
+
\sum_{n> T^3}n^pa_{\psi}^{(k)}(n,T).
\nonumber
\end{eqnarray}
Let $G_t(x)= e^{-itx}$.
Lemma $\ref{6.2.4}$ shows that for any $l>1$, there exists
$C_7^{(4)}=C_7^{(4)}(l)>0$ such that
\begin{eqnarray}
\sum_{n> T^3}n^p a_{\psi}^{(k)}
&=&
\sum_{n> T^3}n^p\cfrac{1}{T}\int_{\mathbb{R}}dt\; e^{-\frac{t}{T}}
|(\delta_n, G_t(H^{(k)})f(H^{(k)})\delta_1)|^2
\nonumber
\\
&\leq&
\sum_{n> T^3}n^p\cfrac{C_2}{T}\int_{\mathbb{R}}dt\; e^{-\frac{t}{T}}
\opnorm{G_tf}_{2l+3} n^{-l}
\nonumber
\\
&\leq&
\sum_{n> T^3}n^p
\cfrac{C_7^{(4)}}{T}\int_{\mathbb{R}}dt\; e^{-\frac{t}{T}}
\; t^{2l+3}
n^{-l}
\nonumber
\\
&\leq&
C_7^{(4)}
T^{2l+3}
\sum_{n> T^3}n^{p-l}
\nonumber
\\
&\leq&
C_7^{(4)}T^{-l+3p+6}.
\nonumber
\end{eqnarray}
We take $l$ large enough, so it is sufficient to prove
$\begin{displaystyle}
\sum_{n=2L_N}^{T^3}n^pa_{\psi}^{(k)}(n,T)
\leq
C_7T^{p+1}L_N^{-\frac{1}{\Gamma}}.
\end{displaystyle}$
Since $f$ is the first kind, $f \in C^{\infty}_0(B_{\nu})$.
By Lemma $\ref{7.1.5}$, we have
\begin{eqnarray}
\sum_{n=2L_N}^{T^3}n^pa_{\psi}^{(k)}(n,T)
&=&
\sum_{n=2L_N}^{T^3}n^p
\cfrac{\epsilon}{\pi}
\int_{B_{\frac{\nu}{2}}}dE \; |(H^{(k)}-E-i\epsilon)^{-1}\psi(n)|^2
\nonumber
\\
&+&
\sum_{n=2L_N}^{T^3}n^p
\cfrac{\epsilon}{\pi}
\int_{\mathbb{R} \setminus B_{\frac{\nu}{2}}}dE \; |(H^{(k)}-E-i\epsilon)^{-1}\psi(n)|^2,
\nonumber
\qquad \epsilon=(2T)^{-1}.
\end{eqnarray}
Let $\chi_z(x) = (x-z)^{-1}$. Then Lemma $\ref{6.2.4}$ shows that for any $l>0$, there exists
$C_7^{(3)}=C_7^{(3)}(l)>0$ such that
\begin{eqnarray}
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{-\infty}^{\frac{\nu}{2}}dE \; |(H^{(k)}-E-i\epsilon)\psi(n)|^2
&\leq&
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{-\infty}^{\frac{\nu}{2}}dE \;
|(\delta_n, \chi_{E+i\epsilon}(H^{(k)})f(H^{(k)})\delta_1)|^2
\nonumber
\\
&\leq&
C_2
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{-\infty}^{\frac{\nu}{2}}dE \;
\opnorm{\chi_{E+i\epsilon}f}^{2}_{2l+3}n^{-2l}
\nonumber
\\
&\leq&
C_2
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^{p-2l}
\int_{-\infty}^{\frac{\nu}{2}}dE \;
\cfrac
{C_7^{(3)}}
{
(E-\nu)^2+\epsilon^2
}
\nonumber
\\
&\leq&
C_2C_7^{(3)}L_N^{-2l+p+1}.
\label{FDS}
\end{eqnarray}
Similarly, there exists $D_7^{(3)}=D_7^{(3)}(l)>0$ such that
\begin{eqnarray}
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{4-\frac{\nu}{2}}^{\infty}dE \; |(H^{(k)}-E-i\epsilon)^{-1}\psi(n)|^2
\leq
C_2D_7^{(3)}L_N^{-2l+p+1}.
\label{FDS1}
\end{eqnarray}
$(\ref{FDS})$ and $(\ref{FDS1})$ imply that
\[
\sum_{n=2L_N}^{T^3}n^p
\cfrac{\epsilon}{\pi}
\int_{\mathbb{R} \setminus B_{\frac{\nu}{2}}}dE \; |(H^{(k)}-E-i\epsilon)^{-1}\psi(n)|^2
\leq
\max\{C_2C_7^{(3)}, C_2D_7^{(3)}\}L_N^{-2l+p+1}.
\]
We take $l$ large enough, so it is sufficient to prove
\begin{eqnarray}
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{B_{\frac{\nu}{2}}}dE \; |(H^{(k)}-E-i\epsilon)^{-1}\psi(n)|^2
\leq
C_7T^{p+1}L_N^{-\frac{1}{\Gamma}},
\qquad \epsilon=(2T)^{-1}.
\nonumber
\end{eqnarray}
Lemma $\ref{6.2.4}$ implies that there exists $C=C(l,f)>0$ such that
\begin{eqnarray}
|(H^{(k)}-E-i\epsilon)^{-1}\psi(n)|^2
\leq
C \sum_{m=1}^{\infty}
(1+|n-m|^2)^{-l}
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2.
\nonumber
\end{eqnarray}
Therefore it is sufficient to prove
\[
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{B_{\frac{\nu}{2}}}dE \;
\sum_{m=1}^{\infty}
(1+|n-m|^2)^{-l}
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\leq
C_7T^{p+1}L_N^{-\frac{1}{\Gamma}},
\qquad \epsilon=(2T)^{-1}.
\nonumber
\]
We see that for any $l>1$, by Lemma $\ref{7.1.5}$, there exists
$C_7^{(2)}=C_7^{(2)}(l)>0$ such that
\begin{eqnarray}
&&
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{B_{\frac{\nu}{2}}}dE \;
\sum_{m=1}^{L_N}
(1+|n-m|^2)^{-l}
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&&\leq
\sum_{n=2L_N}^{T^3}n^p
L_N^{-2l}
\sum_{m=1}^{L_N}
\cfrac{\epsilon}{\pi}
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&&\leq
\sum_{n=2L_N}^{T^3}n^p
L_N^{-2l}
\nonumber
\\
&&\leq
C_7^{(2)}
T^{3(p+1)}
L_N^{-2l}, \qquad \epsilon=(2T)^{-1}.
\label{oiu}
\end{eqnarray}
Similarly, we have
\begin{eqnarray}
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{B_{\frac{\nu}{2}}}dE \;
\sum_{m=T^3+L_N}^{\infty}
(1+|n-m|^2)^{-l}
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\leq
D_{7}^{(2)}T^{3(p+1)}
L_N^{-2l}.
\label{oiu1}
\end{eqnarray}
By $(\ref{oiu})$ and $(\ref{oiu1})$, it is sufficient to prove
\begin{eqnarray}
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{B_{\frac{\nu}{2}}}dE \;
\sum_{m=L_N}^{T^3+L_N}
(1+|n-m|^2)^{-l}
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\leq
C_7T^{p+1}L_N^{-\frac{1}{\Gamma}},\qquad \epsilon=(2T)^{-1}.
\nonumber
\end{eqnarray}
Let $A^{(k)}_N$ and $D^{(k)}_N:l^2(\mathbb{N})\rightarrow l^2(\mathbb{N})$ be
\begin{displaymath}
A^{(k)}_N=
\renewcommand{\arraystretch}{1.8}
\left(
\begin{tabular}{Wc{10mm}Wc{10mm}Wc{10mm}Wc{10mm}Wc{10mm}|
Wc{10mm}Wc{10mm}Wc{10mm}cccccccccc}
0&$a_{k}({1})$&&&&&\\
$a_{k}({1}$)&0&$a_{k}({2})$&&&&&&\\
&$a_{k}({2})$&$\ddots$&$\ddots$&&&&\\
&&$\ddots$&0&$a_{k}(L_N)$&&&&\\
&&&$a_{k}(L_N)$&0&1\\
\hline
&&&&1&0&1\\
&&&&&1&0&1\\
&&&&&&1&$\ddots$&$\ddots$\\
&&&&&&&$\ddots$&\\
\end{tabular}\right)
\renewcommand{\arraystretch}{1}
\end{displaymath}
and
\begin{displaymath}
D^{(k)}_N=
\renewcommand{\arraystretch}{1.8}
\left(
\begin{tabular}{Wc{8mm}Wc{8mm}Wc{8mm}Wc{10mm}|Wc{8mm}cccccc}
$d_{k}({1})$&&&&\\
&$d_{k}({2})$&&\\
&&$\ddots$&\\
&&&$d_{k}({L_N})$&\\
\hline
&&&&2\\
&&&&&2\\
&&&&&&$\ddots$\\
\end{tabular}\right).
\renewcommand{\arraystretch}{1}
\end{displaymath}
Let $H_N^{(k)}= D^{(k)}_N-A^{(k)}_N$. We see that for any $l>1$ there exists
$C^{\prime}=C^{\prime}(l)>0$ such that
\begin{eqnarray}
&&
\cfrac{\epsilon}{\pi}
\sum_{n=2L_N}^{T^3}n^p
\int_{B_{\frac{\nu}{2}}}dE \;
\sum_{m=L_N}^{T^3+L_N}
(1+|n-m|^2)^{-l}
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&&=
\cfrac{\epsilon}{\pi}
\sum_{m=L_N}^{T^3+L_N}
\left(
\sum_{n=2L_N}^{T^3}n^p
(1+|n-m|^2)^{-l}
\right)
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&&\leq
C^{\prime}
\cfrac{\epsilon}{\pi}
\sum_{m=L_N}^{T^3+L_N}
m^p
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)|^2
\nonumber
\\
&&\leq
2C^{\prime}
\cfrac{\epsilon}{\pi}
\sum_{m=L_N}^{T^3+L_N}
m^p
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)-\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1(m)|^2
\nonumber
\\
&&+
2C^{\prime}
\cfrac{\epsilon}{\pi}
\sum_{m=L_N}^{T^3+L_N}
m^p
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1(m)|^2.
\label{6.28}
\end{eqnarray}
By the resolvent equation, we have
\begin{eqnarray}
&&
\|(\chi_{E+i\epsilon}(H^{(k)})-\chi_{E+i\epsilon}(H^{(k)}_N))\delta_1\|
\nonumber
\\
&&\leq
\cfrac{1}{\epsilon}
\|(H^{(k)}-H_N^{(k)})\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1\|
\nonumber
\\
&&\leq
\cfrac{1}{\epsilon}
\|(D^{(k)}-D_N^{(k)})\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1\|
+
\cfrac{1}{\epsilon}
\|(A^{(k)}-A^{(k)}_N)\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1\|.
\label{6.29}
\end{eqnarray}
Let $z=E+i\epsilon$ and
$\phi:\mathbb{Z}_{\geq 0}\rightarrow \mathbb{C}$ be
$\phi(0)=1$,
$\phi(n)= \chi_{z}(H^{(k)}_N)\delta_1(n)$ $(n\geq 1)$. Then
it follows for
$n>L_N$ that
\begin{eqnarray}
-\phi(n+1)+2\phi(n)-\phi(n-1)=z \phi(n).
\nonumber
\end{eqnarray}
This implies for $n>L_N$,
\begin{eqnarray}
\left(
\begin{array}{c}
\phi(n)\\
\phi(n+1)
\end{array}
\right)
=
\left(
\begin{array}{cc}
0&1\\
-1&2-z\\
\end{array}
\right)
\left(
\begin{array}{c}
\phi(n-1)\\
\phi(n)
\end{array}
\right).
\nonumber
\end{eqnarray}
Let $\lambda_\pm = \cfrac{2-z\pm \sqrt{(2-z)^2+4}}{2}$, then
there exists $C_\pm$ such that
\begin{eqnarray}
\phi(n)=C_+ \lambda_+^{n-L_N}+C_-\lambda_-^{n-L_N}.
\nonumber
\end{eqnarray}
Since $\epsilon>0$, we get $|\lambda_-|<1$ and $|\lambda_+|>1$.
Since $\|\phi \|_{l^2}<\infty$, $C_+=0$, $C_-=\phi(L_N)$ and there exists $c,c^{\prime}>0$ such that
it follows for $0<E<4$ and $0<\epsilon<1$ that
\begin{equation}
e^{-c^{\prime}\epsilon}\leq |\lambda_-|\leq e^{-c\epsilon}.
\nonumber
\end{equation}
If $\beta>1$, then it follows for sufficiently large $N$ that
\begin{eqnarray}
L_{N+1}-L_N^\beta>\cfrac{1}{2}L_{N+1}.
\nonumber
\end{eqnarray}
Therefore, we see that there exist $C^{\prime\prime}>0$ such that
\begin{eqnarray}
\|(D^{(k)}-D_N^{(k)})\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1\|^2
&=&
\sum_{j=N+1}^{\infty}|d_k(L_j)-2|^2|\phi(L_j)|^2
\nonumber
\\
&\leq&
2|\phi(L_N)|^2
\sum_{j=N+1}^{\infty}L_j^{\frac{2(1-\Gamma)}{\Gamma}}
{\rm exp}(-2c\epsilon(L_j-L_N))
\nonumber
\\
&\leq&
2\epsilon^{-2}
\sum_{j=N+1}^{\infty}L_j^{\frac{2(1-\Gamma)}{\Gamma}}
{\rm exp}(-2c\epsilon(L_j-L_N^{\beta}))
{\rm exp}(-2c\epsilon(L_N^{\beta}-L_N))
\nonumber
\\
&\leq&
4T^2
\sum_{j=N+1}^{\infty}L_j^{\frac{2(1-\Gamma)}{\Gamma}}
{\rm exp}(-c\epsilon L_j)
{\rm exp}(-c\epsilon L_N^{\beta})
\nonumber
\\
&\leq&
C^{\prime\prime}{\rm exp}(-c\epsilon L_N^{\beta}),
\label{6.30}
\\
\nonumber
\end{eqnarray}
and
\begin{eqnarray}
\|(A^{(k)}-A^{(k)}_N)\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1\|^2
&\leq&
\sum_{j=N+1}^{\infty}
|1-a_k(L_j)|^2
\{
|\phi(L_j)|^2+|\phi(L_j+1)|^2
\}
\nonumber
\\
&\leq&
C^{\prime\prime}T^2\sum_{j=N+1}^{\infty}
{\rm exp}(-2c\epsilon(L_j-L_N))
\nonumber
\\
&\leq&
C^{\prime\prime}T^2{\rm exp}(-c\epsilon L_N^{\beta}).
\label{6.31}
\end{eqnarray}
By $(\ref{6.29})$, $(\ref{6.30})$, and $(\ref{6.31})$, we see that there exists
$C_7^{(1)}>0$ such that
\begin{eqnarray}
&&
\epsilon
\sum_{m=L_N}^{T^3+L_N}
m^p
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)})\delta_1(m)-\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1(m)|^2
\nonumber
\\
&&\leq
\epsilon
\sum_{m=L_N}^{T^3+L_N}
m^p
\int_{B_{\frac{\nu}{2}}}dE \;
\|\chi_{E+i\epsilon}(H^{(k)})\delta_1-\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1\|^2
\nonumber
\\
&&\leq
C_7^{(1)}
T^{3(p+1)}{\rm exp}(-c\epsilon L_N^{\beta}),\qquad \epsilon=(2T)^{-1}.
\label{WAZ}
\end{eqnarray}
By $(\ref{6.28})$ and $(\ref{WAZ})$, it is sufficient to prove
\begin{eqnarray}
\epsilon
\sum_{m=L_N}^{T^3+L_N}
m^p
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1(m)|^2
\leq
C_7T^{p+1}L_N^{-\frac{1}{\Gamma}}.
\nonumber
\end{eqnarray}
Let $F_N(z)= (\delta_1, \chi_z(H_N^{(k)})\delta_1)=(\delta_1,(H_N^{(k)}-z)^{-1}\delta_1)$.
We see that there exists $C^{\prime\prime\prime}>0$ such that
\begin{eqnarray}
\cfrac{1}{\epsilon}{\rm Im}F_N(E+i\epsilon)
=\|\chi_{E+i\epsilon}(H_N^{(k)})\delta_1\|^2
\geq
\sum_{m>L_N}
|\phi(m)|^2
\geq
\sum_{m>L_N}e^{-c^{\prime}\epsilon(m-L_N)}|\phi(L_N)|^2
\geq
\cfrac{C^{\prime\prime\prime}}{\epsilon}|\phi(L_N)|^2.
\nonumber
\end{eqnarray}
This implies that $|\phi(L_N)|\leq C^{\prime\prime\prime} {\rm Im}F_N(E+i\epsilon)$.
Let $L_{N-1}<n<L_{N+1}$. Then
\[
\left\{
\begin{array}{ll}
-\phi(n+1)+2\phi(n)-\phi(n-1)=z\phi(n)&(n \neq L_N, L_N+1)
\\
-\sqrt{[L_N^{\frac{1-\Gamma}{\Gamma}}]}\phi(L_N+1)+([L_N^{\frac{1-\Gamma}{\Gamma}}]+2)\phi(L_N)-\phi(L_N-1)=z\phi(L_N)&(n= L_N)
\\
-\phi(L_N+2)+2\phi(L_N+1)-\sqrt{[L_N^{\frac{1-\Gamma}{\Gamma}}]}\phi(L_N)=z\phi(L_N+1)&(n =L_N+1).
\end{array}
\right.
\]
Let
$\begin{displaystyle}
R
=
\left(
\begin{array}{cc}
0&1\\
-1&2-z\\
\end{array}
\right)
\end{displaystyle}$. Then it follows for $L_N+1<n < L_{N+1}$ that
\begin{eqnarray}
\left(
\begin{array}{c}
\phi(n)\\
\phi(n+1)
\end{array}
\right)
=
R^{n-L_N}
\left(
\begin{array}{c}
\phi(L_N)\\
\phi(L_N+1)
\end{array}
\right).
\nonumber
\end{eqnarray}
Similary, for $L_{N-1} < n<L_N-1$, we have
\begin{eqnarray}
\nonumber
\left(
\begin{array}{c}
\phi(n)\\
\phi(n+1)
\end{array}
\right)
=
R^{n-L_N+1}
\left(
\begin{array}{c}
\phi(L_N-1)\\
\phi(L_N)
\end{array}
\right).
\end{eqnarray}
There exists $B=B(K)>0$ such that $\|R^n\|<B$ for $\epsilon<\frac{K}{|n|}$.
It follows for $L_N<n<2L_N$ that
\begin{eqnarray}
|\phi(n)|^2+|\phi(n+1)|^2 \geq B^{-1}(|\phi(L_N)|^2+|\phi(L_N+1)|^2).
\nonumber
\end{eqnarray}
Therefore, we have
\begin{eqnarray}
\cfrac{1}{\epsilon}\:{\rm Im}F_N(E+i\epsilon)
=
\|\phi\|^2 \geq B^{-1}L_N(|\phi(L_N)|^2+|\phi(L_N+1)|^2).
\label{++}
\end{eqnarray}
Similarly, it follows that for $\cfrac{L_N}{2}<n<L_N$,
\[
|\phi(n)|^2+|\phi(n+1)|^2 \geq B^{-1}(|\phi(L_N)|^2+|\phi(L_N+1)|^2),
\]
and that
\begin{eqnarray}
\cfrac{1}{\epsilon}{\rm Im}F_N(E+i\epsilon)
=
\|\phi\|^2 \geq B^{-1}L_N(|\phi(L_N-1)|^2+|\phi(L_N)|^2).
\label{+++}
\end{eqnarray}
$(\ref{++})$ and $(\ref{+++})$ imply that
\begin{eqnarray}
|\phi(L_N-1)|^2+|\phi(L_N+1)|^2
\leq
\cfrac{2B}{\epsilon L_N}{\rm Im}F_N(E+i\epsilon).
\label{IJM}
\end{eqnarray}
We see that
\begin{eqnarray}
([L_N^{\frac{1-\Gamma}{\Gamma}}]+2-z)\phi(L_N)
=\phi(L_N-1)+\sqrt{[L_N^{\frac{1-\Gamma}{\Gamma}}]}\phi(L_N+1).
\nonumber
\end{eqnarray}
This shows that
\begin{eqnarray}
|[L_N^{\frac{1-\Gamma}{\Gamma}}]+2-z|^2|\phi(L_N)|^2
\leq2[L_N^{\frac{1-\Gamma}{\Gamma}}](|\phi(L_N-1)|^2+|\phi(L_N+1)|^2).
\nonumber
\end{eqnarray}
Let $|z|<5$. Then there exists $B^{\prime}>0$ such that
\begin{eqnarray}
|\phi(L_N)|^2
\leq B^{\prime}L_N^{\frac{\Gamma-1}{\Gamma}}(|\phi(L_N-1)|^2+|\phi(L_N+1)|^2).
\label{IJM1}
\end{eqnarray}
$(\ref{IJM})$ and $(\ref{IJM1})$ imply that
\begin{eqnarray}
|\phi(L_N)|^2
\leq2BB^{\prime}
\cfrac{L_N^{-\frac{1}{\Gamma}}}{\epsilon}{\rm Im}F_N(E+i\epsilon).
\nonumber
\end{eqnarray}
Therefore, there exists $C_7=C_7(p)$ such that
\begin{eqnarray}
&&
\epsilon
\sum_{m=L_N}^{T^3+L_N}
m^p
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1(m)|^2
\nonumber
\\
&&\leq
C_7
\epsilon
\sum_{m=L_N}^{T^3+L_N}
m^p
{\rm exp}(-2c\epsilon(m-L_N))
\int_{B_{\frac{\nu}{2}}}dE \;
|\chi_{E+i\epsilon}(H^{(k)}_N)\delta_1(L_N)|^2
\nonumber
\\
&&\leq
C_7
\epsilon^{-p}
\int_{B_{\frac{\nu}{2}}}dE \;
|\phi(L_N)|^2
\nonumber
\\
&&\leq
2BB^{\prime}
C_7
\epsilon^{-p-1}
L_N^{-\frac{1}{\Gamma}}
\int_{B_{\frac{\nu}{2}}}dE \; {\rm Im}F_N(E+i\epsilon)
\nonumber
\\
&&\leq
2BB^{\prime}
C_7
T^{p+1}L_N^{-\frac{1}{\Gamma}},\qquad \epsilon=(2T)^{-1}.
\nonumber
\end{eqnarray}
\qed
\end{Proof}
\begin{_corollary}\label{PQ}
Let $p>0$, $f$ be the first kind, and $\psi = f(H^{(k)})\delta_1$.
Then there exists $C_8=C_8(p)>0$ such that for
$L_N \leq T \leq L_N^{\frac{1}{\Gamma}}$ with $N$ sufficiently large,
\begin{equation}
\langle|X|^p \rangle^{(k)}_{\psi}(T)
\leq
C_8L_N^p+C_8T^{p+1}L_N^{-\frac{1}{\Gamma}}.
\nonumber
\end{equation}
\end{_corollary}
\begin{Lemma}\label{HH}
Let $f$ be the first kind and $\psi = f(H^{(k)})\delta_1$. Then
\begin{eqnarray}
\beta_{\psi}^{(k)}(p)= \cfrac{p+1}{p+\frac{1}{\Gamma}}.
\nonumber
\end{eqnarray}
\end{Lemma}
\begin{Proof}\rm
Let $L_N \leq T=L_N^A \leq L_N^{\frac{1}{\Gamma}}$, where $A=\cfrac{p+\frac{1}{\Gamma}}{p+1}$. Then Corollary $\ref{PQ}$ shows that
\begin{eqnarray}
\langle|X|^p \rangle^{(k)}_{\psi}(L_N^A)
\leq
C_8L_N^p.
\nonumber
\end{eqnarray}
Therefore we have
\begin{eqnarray}
\beta_{\psi}^{(k)}(p)\leq
\cfrac{1}{p}\lim_{N \rightarrow \infty}
\cfrac{{\rm log}\langle |X|^p\rangle_{\psi}^{(k)}(L_N^A)}{{\rm log}L_N^A}
\leq
A^{-1}
=
\cfrac{p+1}{p+\frac{1}{\Gamma}}.
\nonumber
\end{eqnarray}
Since $f$ is the first kind, $(\ref{WWWWW})$ holds.
\qed
\end{Proof}
\subsection{Proof of the main result}
\begin{Lemma}\label{last lemma}
Let $A\in \mathcal{B}^1$. Then $E(A)=0$ if and only if
$\mu_{\delta_1}^{(k)}(A)=0$ for any $k\in\mathbb{N}$.
\end{Lemma}
\begin{Proof}\rm
Assume that $E(A)=0$. Then we see that $E^{(k)}(A)=0$ and
\[
\mu_{\delta_1}^{(k)}(A)=(\delta_1,E^{(k)}(A)\delta_1)=0.
\]
Conversely, assume that $\mu_{\delta_1}^{(k)}(A)=0$ for any $k \in \mathbb{N}$. It is sufficient to prove that $E^{(k)}(A)=0$ for any $k \in \mathbb{N}$. Let $p$ ba a polynomial, then we see that
\[
\mu_{p(H^{(k)})\delta_1}^{(k)}(A)=(p(H^{(k)})\delta_1,E^{(k)}(A)p(H^{(k)})\delta_1)=\int_{A}|p(\lambda)|^2 \mu_{\delta_1}^{(k)}(d\lambda)=0.
\]
This implies that $E^{(k)}(A)p(H^{(k)})\delta_1=0$.
Since $\delta_1$ is a cyclic vector for $H^{(k)}:l^2(\mathbb{N}) \rightarrow l^2(\mathbb{N})$,
$\{p(H^{(k)})\delta_1 \in l^2(\mathbb{N})\mid \text{ $p$ is a polynomial}\}$ is dense in $l^2(\mathbb{N})$. Therefore $E^{(k)}(A)=0$ follows.
\qed
\end{Proof}
\begin{Lemma}\label{last last lamma}
Let $A\in \mathcal{B}^1$ and $A \subset (0,4)$. Then $\tilde{E}(A)=0$ if and only if
$\mu_{\psi}^{(k)}(A)=0$ for any $k\in\mathbb{N}$ and any $\psi=f(H^{(k)})\delta_1$, with the first kind $f$.
Moreover, ${\rm dim}_{*}\tilde{E}={\rm dim}_{*}\mu_{\psi}^{(k)}$ and ${\rm dim}^{*}\tilde{E}={\rm dim}^{*}\mu_{\psi}^{(k)}$ follow for any $k\in \mathbb{N}$ and any $\psi=f(H^{(k)})\delta_1$, with the first kind $f$.
\end{Lemma}
\begin{Proof}\rm
Assume that $\tilde{E}(A)=0$. Then we see that $E^{(k)}(A)=0$ and
\[
\mu_{\psi}^{(k)}(A)=(\psi,E^{(k)}(A)\psi)=0.
\]
Assume that $\mu_{\psi}^{(k)}(A)=0$ for any $k\in\mathbb{N}$ and $\psi=f(H^{(k)})\delta_1$, where $f$ is the first kind.
Let $f_n \in C^{\infty}_0(\frac{1}{n}, 4-\frac{1}{n})$, $|f_n|\leq 1$, and $f_n=1$ on the interval $(\frac{2}{n}, 4-\frac{2}{n})$, $n=1,2,...$. Let $\psi_n=f_n(H^{(k)})\delta_1$.
Since $f_n$ is the first kind, $\mu_{\psi_n}^{(k)}(A)=0$ for any $k \in \mathbb{N}$.
It is sufficient to prove that $E^{(k)}(A)=0$ for any $k \in \mathbb{N}$.
We see that
\[
\mu_{\psi_n}^{(k)}(A)=(f_n(H^{(k)})\delta_1,E^{(k)}(A)f_n(H^{(k)})\delta_1)=\int_{A}|f_n(\lambda)|^2 \mu_{\delta_1}^{(k)}(d\lambda)=0.
\]
By the Lebeasgue's dominated convergence theorem,
\[
0=\lim_{n\rightarrow \infty}\mu_{\psi_n}^{(k)}(A)=
\lim_{n\rightarrow \infty}\int_{A}|f_n(\lambda)|^2 \mu_{\delta_1}^{(k)}(d\lambda)
=\mu_{\delta_1}^{(k)}(A).
\]
By Lemma $\ref{last lemma}$, we see that $E^{(k)}(A)=0$. Then we prove the first part of our assertion. The second part is straightforward to prove by the first part and the definition of the lower and upper Hausdorff dimensions.
\qed
\end{Proof}
\begin{proof_main}\rm
By Lemma $\ref{last last lamma}$, it is sufficient to prove that ${\rm dim}_{*}\mu_{\psi}^{(k)}={\rm dim}^{*}\mu_{\psi}^{(k)}=\Gamma$ for any $k\in \mathbb{N}$ and any $\psi=f(H^{(k)})\delta_1$, with the first kind $f$. By Lemma $\ref{former-result}$, we see that
\[
\Gamma \leq {\rm dim}_{*}\mu_{\psi}^{(k)}\leq{\rm dim}^{*}\mu_{\psi}^{(k)}.
\]
By Lemma $\ref{H}$ and Lemma $\ref{HH}$, for any $p>0$,
\[
{\rm dim}^*(\mu_{\psi}^{(k)}) \leq \beta_{\psi}^{(k)}(p)= \cfrac{p+1}{p+\frac{1}{\Gamma}}.
\]
This imlies that ${\rm dim}_{*}\mu_{\psi}^{(k)}={\rm dim}^{*}\mu_{\psi}^{(k)}=\Gamma$.
\qed
\end{proof_main}
\begin{appendices}
\section{}
\label{Decomposition of the graph Laplacian}
We discuss the decomposition of the graph Laplacian and represent the graph Laplacian as a Jacobi matrix. See $[1]$.
We assume that $G=(V,E)$ is a spherically homogeneous tree.
Let $\pi_n:l^2(S_n)\rightarrow l^2(S_{n+1})$, $n=0,1,... ,$ be defined by
\[
\pi_nf(u)= \sum_{v\in S_{n}:v \sim u} f(v), \:\: {\text \rm u \in S_{n+1}}.
\]
The adjoint $\pi_n^*:l^2(S_{n+1})\rightarrow l^2(S_n)$ is given by
\begin{displaymath}
\pi_n^*g(u)=\sum_{v\in S_{n+1}: v\sim u}g(v), \:\: {\text \rm u \in S_n}.
\end{displaymath}
\begin{Lemma}\label{3.1.1}
Let $f$, $g \in l^2(S_n)$. Then $(\pi_nf,\pi_ng)=g_n(f,g)$.
\end{Lemma}
\begin{Proof}\rm
Let $f$, $g \in l^2(S_n)$. Since $G$ is a spherically homogeneous,
\[
\langle \pi_n f, \pi_n g \rangle=\sum_{u\in S_{n+1}}\overline{\pi_n f(u)} \pi_n g(u)
=g_n\sum_{u\in S_{n}}\overline{ f(u)} g(u).
\]
\qed
\end{Proof}
We see that $V$ is a disjoint union $V=\cup_{n=0}^{\infty}S_n$,
and that $l^2(V)=\bigoplus_{n=0}^{\infty}l^2(S_n)$.
Let $\Pi:l^2(V)\rightarrow l^2(V)$ be defined by
$\Pi = \bigoplus_{n=0}^{\infty}\pi_n$.
\begin{Lemma}\label{RT}
Let $f \in \mathcal{D}$. Then $Af=(\Pi +\Pi^*)f$.
\end{Lemma}
\begin{Proof}\rm
Let $f \in \mathcal{D}$ and $u \in S_n$.
Since $G$ is a spherically homogeneous tree,
$u$ is adjacent with only vertices in $S_{n-1}$ and $S_{n+1}$. Therefore we see that
\[
(\Pi +\Pi^*)f(u)= \sum_{v\in S_{n-1};v\sim u}f(v)+ \sum_{v\in S_{n+1};v\sim u}f(v)
=\sum_{v \in V;v\sim u}f(v)
=Af(u).
\]
\qed
\end{Proof}
Let $\alpha_n= \#S_n={\rm dim}(l^2(S_n))$, $n=0,1,...$.
Suppose that $\{ e^{(n)}_k\}_{k=1}^{\alpha_n}$ is a CONS of $l^2(S_n)$.
Then we can construct a CONS $\{ e^{(n+1)}_k\}_{k=1}^{\alpha_{n+1}}$ of $l^2(S_{n+1})$ by the following procedure.
Let
$
e^{(n+1)}_k = \| \pi_n e^{(n)}_k\|^{-1}\pi_n e^{(n)}_k
$, $k=1,2,...,\alpha_n$.
By Lemma $\ref{3.1.1}$,
$\{e^{(n+1)}_k\}_{k=1}^{\alpha_{n}}$ is an ONS of $l^2(S_{n+1})$.
If $\alpha_n=\alpha_{n+1}$, then $\{e^{(n+1)}_k\}_{k=1}^{\alpha_{n}}$ is a CONS of $l^2(S_{n+1})$.
If $\alpha_n<\alpha_{n+1}$,
by the Gram-Schmidt process, we can obtain $e^{(n+1)}_k\in l^2(S_{n+1})$, $k=\alpha_n+1,...,\alpha_{n+1}$, such that
$\{e^{(n+1)}_k\}_{k=1}^{\alpha_{n}}\cup\{e^{(n+1)}_k\}_{k=\alpha_n+1}^{\alpha_{n+1}}$
is a CONS of $l^2(S_{n+1})$.
Suppose that a CONS of $l^2(S_0)$ is given.
Then we can costruct a CONS $\{e^{(n)}_k\}_{k=1}^{\alpha_n}$ of $l^2(S_n)$,
$n=0,1,... ,$ inductively.
Hence,
$
\bigcup_{n=0}^{\infty}\{e^{(n)}_k\}_{k=1}^{\alpha_n}
$ is a CONS of $l^2(V)$.
Assume that
$
\sup_{n=0,1,...} \alpha_n=\infty
$, and let $\alpha_{-1}=0$.
Since
$\{\alpha_n\}_{n=0}^{\infty}$ is non-decreasing,
there exists a unique $N(k)\in \mathbb{N}\cup \{0\}$ such that
$\alpha_{N(k)-1}< k \leq \alpha_{N(k)}$ for every $k \in \mathbb{N}$.
We see that
\begin{displaymath}
\bigcup_{n=0}^{\infty}\{e^{(n)}_k \mid k=1,2,...,\alpha_n \}=
\bigcup_{k=1}^{\infty}\{e^{(n)}_k \mid n=N(k),N(k)+1,... \}.
\end{displaymath}
\begin{Lemma}\label{invariance of M_k}
Let the closed subspace $M_k$ of $l^2(V)$, $k=1,2,...$, be defined by
\begin{displaymath}
M_k = \overline{\langle
\{e^{(n)}_k \mid n=N(k),N(k)+1,...
\}\rangle}.
\end{displaymath}
Then $M_k$ is invariant under $A$, $D$ and $H$.
\end{Lemma}
\begin{Proof}\rm
By the definition of $e^{(n)}_k$ and
Lemma \ref{3.1.1}, we see that
\begin{eqnarray}
\Pi e^{(n)}_k&=&\| \pi_n e^{(n)}_k\| e^{(n+1)}_k ,
\nonumber
\\
\Pi^* e^{(n)}_k&=&
\begin{cases}
\cfrac{g_{n-1}}{\| \pi_{n-1} e^{(n-1)}_k\|}\; e^{(n-1)}_k & (n \geq N(k)+1),
\\
\bm{o}&(n=N(k)).
\nonumber
\end{cases}
\end{eqnarray}
This implies that $M_k$ is invariant under $\Pi$ and $\Pi^*$, and hence,
by Lemma \ref{RT}, we see that
$M_k$ is invariant under $A$.
Since $G$ is a spherically homogeneous tree, we have
\begin{equation}
D e^{n}_k =
\begin{cases}
(g_n+1) e^{(n)}_k & (n\geq 1),\\
g_0 e^{(0)}_1 & (n=0).
\end{cases}
\nonumber
\label{3.3}
\end{equation}
Hence, $M_k$ is also inavariant under $D$.
Since $H=\overline{D-A}$, we see that $M_k$ is invariant under $H$.
\qed
\end{Proof}
By Lemma \ref{invariance of M_k},
let $H^{(k)},A^{(k)},D^{(k)}:M_k\rightarrow M_k$, $k=1,2,...$, be defined by the restriction of $H,A$ and $D$ to $M_k$, respectively.
We see that $H^{(k)}$ is self-adjoint and
$H=\bigoplus_{k=1}^{\infty}H^{(k)}$.
We consider the matrix representation of $H^{(k)}$
with respect to the CONS $\{e^{(n)}_k \mid n=N(k),N(k)+1,... \}$ of $M_k$ for $k=2,3,...$. Then it follows for $n,m \geq N(k)$ that
\begin{eqnarray}
(e_k^{(n)},H^{(k)}e_k^{(n)})&=&(e_k^{(n)},D^{(k)}e_k^{(n)})=g_n+1,
\nonumber
\\
(e_k^{(n)},H^{(k)}e_k^{(n+1)})&=&
-(e_k^{(n)},A^{(k)}e_k^{(n+1)})=-\sqrt{g_n},
\nonumber
\\
(e_k^{(n)},H^{(k)}e_k^{(m)})&=&0, \text{ if $|n-m|\geq2$ } .
\nonumber
\end{eqnarray}
We have the matrix representation of $H^{(1)}$.
It follows for $n,m \geq N(1)=0$ that
\begin{eqnarray}
(e_1^{(n)},H^{(1)}e_1^{(n)})&=&
(e_1^{(n)},D^{(1)}e_1^{(n)})=
\begin{cases}
g_0&(n=0),\\
g_{n}+1&(n \geq 1),
\end{cases}
\nonumber
\\
(e_1^{(n)},H^{(1)}e_1^{(n+1)})&=&
-(e_1^{(n)},A^{(1)}e_1^{(n+1)})=-\sqrt{g_n},
\nonumber
\\
(e_1^{(n)},H^{(1)}e_1^{(m)})&=&0, \text{ if $|n-m|\geq 2$}.
\nonumber
\end{eqnarray}
Let $k,n\in \mathbb{N}$ and let
$d_{k}=(d_k(n))_{n=1}^{\infty}$ and $a_{k}=(a_{k}(n))_{n=1}^{\infty}$ be defined by
\begin{eqnarray}
d_k(n)&=&
(e_{k}^{(n+N(k)-1)},D^{(k)}e_{k}^{(n+N(k)-1)}),\nonumber\\
a_{k}(n)&=&
(e_{k}^{(n+N(k))},A^{(k)}e_{k}^{(n+N(k)-1)}).\nonumber
\end{eqnarray}
We can identify $H^{(k)}:l^2(\mathbb{N})\rightarrow l^2(\mathbb{N})$, $k=1,2,...$, with the following Jacobi matrix :
\begin{displaymath}
H^{(k)}=
\renewcommand{\arraystretch}{1.6}
\left(
\begin{array}{ccccccc}
d_{k}({1})&-a_{k}({1})\\
-a_{k}({1})&d_{k}({2})&-a_{k}({2})\\
&-a_{k}({2})&d_{k}({3})&-a_{k}({3})\\
&&-a_{k}({3})&\ddots&\ddots\\
&&&\ddots\\
\end{array}\right).
\renewcommand{\arraystretch}{1}
\end{displaymath}
This implies our assertion of Lemma $\ref{jacobi identification}$.
Similary, we can identify $A^{(k)}, D^{(k)}:l^2(\mathbb{N}) \rightarrow l^2(\mathbb{N})$ with the following Jacobi matrices :
\begin{displaymath}
A^{(k)}=
\renewcommand{\arraystretch}{1.6}
\left(
\begin{array}{ccccccc}
0&a_{k}({1})\\
a_{k}({1})&0&a_{k}({2})\\
&a_{k}({2})&0&a_{k}({3})\\
&&a_{k}({3})&\ddots&\ddots\\
&&&\ddots\\
\end{array}\right)
\renewcommand{\arraystretch}{1},
D^{(k)}=
\renewcommand{\arraystretch}{1.6}
\left(
\begin{array}{ccccccc}
d_{k}({1})&\\
&d_{k}({2})&\\
&&d_{k}({3})&\\
&&&\ddots&\\
&&&\\
\end{array}\right).
\renewcommand{\arraystretch}{1}
\end{displaymath}
\section{}
In this section, we introduce a result about the Fourier analysis of the fractal measure.
\begin{comment}
We review Marcinkiewicz interpolation.
Let $p,q\in [1,\infty]$.
We say that $T:L^p(\mathbb{R},d\mu)\rightarrow L^q(\mathbb{R},d\mu)$
is quasi-linear, if
there exists a non-negative number $C>0$ such that for any $f,g \in L^p(\mathbb{R},d\mu)$,
\[
|T(f+g)(x)|\leq C(|Tf(x)|+|Tg(x)|), \qquad \text{ \rm a.e. } x
\nonumber
\]
with respect to $\mu$.
We say that a measurable function $f:\mathbb{R}\rightarrow \mathbb{C}$ is weak $L^p$, if there exists a positive number $A>0$ such that
for $s>0$,
\begin{equation}
\mu(\{ x \in \mathbb{R} \mid |f(x)|>s \})\leq \cfrac{A^p}{s^p}. \label{5w}
\nonumber
\end{equation}
If $p<\infty$, let
$\|f\|_{L^p_w}= \inf\{A>0\mid \text{\rm$A$ satisfies } (\ref{5w})\}$, and
if $p=\infty$, let $\|f\|_{L^{\infty}_w}= \|f\|_{L^{\infty}}$.
We say that a quasi-linear map $T:L^p(\mathbb{R},d\mu)\rightarrow L^q(\mathbb{R},d\mu)$
is strong $(p,q)$ type, if there exists a positive number
$M_{p,q}>0$ such that for $f \in L^p(\mathbb{R}, d\mu)$,
\[
\|Tf\|_{L^q}\leq M_{p,q}\|f\|_{L^p}.
\]
We say that a quasi-linear map
$T:L^p(\mathbb{R},d\mu)\rightarrow L^q(\mathbb{R},d\mu)$
is weak $(p,q)$ type, if
there exists a positive number $M_{p,q}>0$ such that for $f \in L^p(\mathbb{R}, d\mu)$,
\[
\|Tf\|_{L^q_w}\leq M_{p,q}\|f\|_{L^p}.
\]
\begin{Lemma}\label{Marcin}
Let $p,q \in [1,\infty]$ be $p<q$, and
let $T$ be a quasi-linear map on both $L^p(\mathbb{R}, d\mu)$ and
$L^q(\mathbb{R}, d\mu)$.
Suppose that
$T$ is weak $(p,p)$ type and weak $(q,q)$ type. Then
$T:L^r(\mathbb{R},d\mu)\rightarrow L^r(\mathbb{R}, d\mu)$ is quasi-linear and strong $(r,r)$ type for any $r \in (p,q)$.
\end{Lemma}
\begin{Proof}
\rm
\cite[VIII, 9 The Marcinkiewicz interpolation theorem, Theorem 9.1]{real analysis}
\end{Proof}
\end{comment}
Let $B_r(x)=[x-r,x+r]\subset \mathbb{R}$. Let
$\mathcal{L}$ be the Lebeage measure on $\mathbb{R}$, and
$\mu:\mathcal{B}^1 \rightarrow [0,\infty]$ be a locally finite measure.
Let $M_{\mu}f:\mathbb{R} \rightarrow \mathbb{R}$ be defined by
\[
M_{\mu}f(x)
=
\sup_{r>0}
\cfrac{1}{\mu(B_r(x))}
\int_{B_r(x)}|f|d\mu
\]
for $f \in L^1_{\rm loc}(\mathbb{R}, d\mu)$,
where we take $\frac{0}{0}=0$ if $\mu(B_r(x))=0$.
$M_{\mu}f$ is measurable and called the Maximal function.
\begin{comment}
\begin{Lemma}\label{Besicovitch}
There exists $C_0>0$ such that for any bounded subset $A\subset \mathbb{R}$, and any closed covering $\{B_{r_x}(x)\subset \mathbb{R} \mid x \in A\}$ of $A$ ,
there exists a countable subset $\{x_j\}_{j=1}^{\infty}$ of $A$ such that
$
A\subset \bigcup_{j=1}^{\infty}B_j
$
and for any $y \in \mathbb{R}$,
\[
\mathbbm{1}_A(y)
\leq
\sum_{j=1}^{\infty}
\mathbbm{1}_{B_j}(y)
\leq
C_0,
\]
where $B_j=B_{r_{x_j}}(x_j)$.
\end{Lemma}
\begin{Proof}
\rm
\cite[Section II, 18, Theorem18.1]{real analysis}
\qed
\end{Proof}
\end{comment}
\begin{Lemma}\label{max}
$M_{\mu}:L^p(\mathbb{R},d\mu)\rightarrow L^p(\mathbb{R},d\mu)$ is bounded for any $p \in (1, \infty)$.
\begin{comment}
It follows for any $f \in L^1(\mathbb{R}, d\mu)$ and $s>0$,
\[
\mu(\{
x \in \mathbb{R} \mid M_{\mu}f(x)>s
\})
\leq C_0s^{-1}\|f\|_{L^1}.
\]
Especially $M_{\mu}:L^1(\mathbb{R},d\mu)\rightarrow L^1(\mathbb{R},d\mu)$ is
weak $(1,1)$ type quasi-linear. Moreover,
\end{comment}
\end{Lemma}
\begin{Proof}\rm
Let $E_s^{n}=\{x \in \mathbb{R} \mid |x|\leq n, M_{\mu}f(x)>s\}$ and
$x \in E_s^n$. There exists $r_x>0$ such that
\[
\int_{B_{r_x}(x)}|f|d\mu\geq s\mu(B_{r_x}(x)).
\]
Note that $\{B_{r_x}(x)\mid x \in E_s^n \}$ is a Besicovitch covering of $E_s^n$.
By \cite[II, 18 The Besicovitch covering theorem, Theorem18.1]{real analysis}, we see that there exists a countable subcollections $\{B_j^n\}_{j=1}^{\infty}$ of $\{B_{r_x}(x)\mid x \in E_s^n \}$ such that $\{B_j^n\}_{j=1}^{\infty}$ is a closed covering of $E_s^n$
and there exist $C>0$ which is independent of $E_s^n$ such that for any $x \in \mathbb{R}$,
\[
\1_{E_s^n}(x) \leq \sum_{j=1}^{\infty} \1_{B_j^n}(x) \leq C.
\]
Hence we have
\[
s\;\mu(E_s^n)\leq s\sum_{j=1}^{\infty}\mu(B_j^n)
\leq
\sum_{j=1}^{\infty} \int_{B_j^n}|f|d\mu \leq C\int_{\mathbb{R}}|f|d\mu.
\]
Let $n \rightarrow \infty$. Then we see that for any $f \in L^1(\mathbb{R}, d\mu)$ and $s>0$,
\[
\mu(\{
x \in \mathbb{R} \mid M_{\mu}f(x)>s
\})
\leq Cs^{-1}\|f\|_{L^1}.
\]
This implies that $M_{\mu}:L^{1}(\mathbb{R},d\mu)\rightarrow L^{1}(\mathbb{R},d\mu)$ is weak $(1,1)$ type.
We also see that $M_{\mu}:L^{\infty}(\mathbb{R},d\mu)\rightarrow L^{\infty}(\mathbb{R},d\mu)$ is
weak $(\infty,\infty)$ type.
Thus we have our assertion by \cite[VIII, 9 The Marcinkiewicz interpolation theorem, Theorem 9.1]{real analysis}.
\qed
\end{Proof}
We consider the Fourier transformation of the fractal measure.
Let $f \in L^1(\mathbb{R},d\mu)$ and
let $\widehat{f\mu}(\xi)$, $\xi \in \mathbb{R}$, be defined by
\[
\widehat{f\mu}(\xi)=\int_{\mathbb{R}}f(x)e^{-i\xi x} \mu(dx).
\]
\begin{Lemma}
Suppose $\mu$ be a finite measure. Then
\[
\int_{\mathbb{R}}
|\widehat{f\mu}(\xi)|^2 e^{-t\xi^2}d\xi<\infty.
\]
for any $f\in L^2(\mathbb{R}, d\mu)$ and $t>0$.
\end{Lemma}
\begin{Proof}
\rm
Let $f\in L^1(\mathbb{R},d\mu)$. We see that
\begin{eqnarray}
\int_{\mathbb{R}}|\widehat{f\mu}(\xi)|e^{-t\xi^2}d\xi
&\leq&
\int_{\mathbb{R}}\int_{\mathbb{R}}|f(x)|\mu(dx)e^{-t\xi^2}d\xi
<\infty.
\nonumber
\end{eqnarray}
This implies that
$L^1(\mathbb{R}, d\mu)\ni f \rightarrow \widehat{f\mu}\in L^1(\mathbb{R}, e^{-t\xi^2}d\xi)$ is bounded.
Since $\mu$ is a finite measure,
$L^{\infty}(\mathbb{R}, d\mu)\ni f \rightarrow \widehat{f\mu}\in L^{\infty}(\mathbb{R}, e^{-t\xi^2}d\xi)$ is bounded. We have our assertion by the Riesz interpolation theorem.
\qed
\end{Proof}
\begin{Definition}Let $\alpha \in (0,1)$.
We say that a measure $\mu$ is uniformly $\alpha$-H\"{o}lder continuous,
if there exists
$\widetilde{C}_1>0$ such that $\mu(I)<\widetilde{C}_1\mathcal{L}(I)^{\alpha}$
for any interval $I\subset \mathbb{R}$ with $\mathcal{L}(I)<1$.
\end{Definition}
\begin{Lemma}\label{fourierdecay}
Let $\mu$ be a uniformly $\alpha$-H\"{o}lder continuous and finite measure.
Then there exists $\widetilde{C}_2=\widetilde{C}_2(\alpha,\mu)>0$ such that for any $f \in L^2(\mathbb{R}, d\mu)$,
\begin{equation}
\sup_{0<t \leq1}
t^{\frac{1-\alpha}{2}}
\int_{\mathbb{R}}
|\widehat{f\mu}(\xi)|^2 e^{-t\xi^2}d\xi<\widetilde{C}_2\|f\|_{L^2}^2.
\nonumber
\end{equation}
\end{Lemma}
\begin{Proof}\rm
Let $f \in L^2(\mathbb{R}, d\mu)$. We see that
\begin{eqnarray}
t^{\frac{1-\alpha}{2}}
\int_{\mathbb{R}}
|\widehat{f\mu}(\xi)|^2 e^{-t\xi^2}d\xi
&=&
t^{\frac{1-\alpha}{2}}
\int_{\mathbb{R}}\mu(dx)
\int_{\mathbb{R}}\mu(dy)
f(x)\overline{f(y)}
\int_{\mathbb{R}}e^{-t\xi^2-i\xi(x-y)}d\xi
\nonumber\\
&=&
\pi
t^{-\frac{\alpha}{2}}
\int_{\mathbb{R}}\mu(dx)
\int_{\mathbb{R}}\mu(dy)
f(x)\overline{f(y)}
e^{-\frac{(x-y)^2}{4}}
\nonumber\\
&=&
\pi
t^{-\frac{\alpha}{2}}
\int_{\mathbb{R}}\mu(dx)
\int_{\mathbb{R}}\mu(dy)
f(x)\overline{f(y)}
\int_{|x-y|}^{\infty}
\cfrac{r}{2t}\;e^{-\frac{r^2}{4t}}dr
\nonumber\\
&\leq&
\pi
t^{-\frac{\alpha}{2}}
\int_{\mathbb{R}}\mu(dx)\;
|f(x)|
\int_{0}^{\infty}
dr
\cfrac{r}{2t}\;
e^{-\frac{r^2}{4t}}
\int_{B_r(x)}\mu(dy)
|f(y)|
\nonumber\\
&\leq&
\pi
t^{-\frac{\alpha}{2}}
\int_{\mathbb{R}}\mu(dx)\;
|f(x)|
\int_{0}^{\infty}
dr
\cfrac{r}{2t}\;
e^{-\frac{r^2}{4t}}
\mu(B_r(x))
M_{\mu}f(x).
\nonumber
\end{eqnarray}
Since $\mu$ is uniformly $\alpha$-H\"{o}lder continuous and finite,
there exists $\widetilde{C}_2^{\prime}=\widetilde{C}_2^{\prime}(\alpha,\mu)>0$ such that for any $t \in (0,1]$,
\begin{eqnarray}
t^{-\frac{\alpha}{2}}
\int_{0}^{\infty}
dr
\cfrac{r}{t}\;
e^{-\frac{r^2}{4t}}
\mu(B_r(x))
&\leq&
\mu({\mathbb{R}})
t^{-\frac{\alpha}{2}}
\int_{1}^{\infty}
dr
\cfrac{r}{t}\;
e^{-\frac{r^2}{4t}}
+\widetilde{C}_1t^{-\frac{\alpha}{2}}
\int_{0}^{1}
dr
\cfrac{r^{1+\alpha}}{t}\;
e^{-\frac{r^2}{4t}}
\nonumber\\
&\leq&
2\mu({\mathbb{R}})
t^{-\frac{\alpha}{2}}
e^{-\frac{1}{4t}}
+2^{2+\alpha}\widetilde{C}_1
\int_{0}^{\infty}
s^{1+\alpha}
e^{-s^2}
ds
\nonumber
\\
&\leq&
\widetilde{C}_2^{\prime}.
\nonumber
\end{eqnarray}
Let $\widetilde{C}_2=\pi \widetilde{C}_2^{\prime}>0$. By Schwarz inequality, we see that for any $t\in (0,1)$,
\begin{eqnarray}
&&
t^{\frac{1-\alpha}{2}}
\int_{\mathbb{R}}
|\widehat{f\mu}(\xi)|^2 e^{-t\xi^2}d\xi
\leq
\widetilde{C}_2
\int_{\mathbb{R}}\mu(dx)\;
|f(x)|
M_{\mu}f(x)
\leq
D_2
\|f\|_{L^2}\|M_{\mu}f\|_{L^2}.
\nonumber
\end{eqnarray}
By Lemma \ref{max}, we have our assertion.
\qed
\end{Proof}
\begin{Lemma}\label{cor}
Let $\mu$ be a uniformly $\alpha$-H\"{o}lder continuous and finite measure.
Then there exists $\widetilde{C}_3=\widetilde{C}_3(\alpha,\mu)>0$ such tha for any $f \in L^2(\mathbb{R}, d\mu)$,
\[
\sup_{T\geq 1}
T^{\alpha-1}
\int_{0}^T
|\widehat{f\mu}(\xi)|^2d\xi
\leq \widetilde{C}_3\|f\|^2.
\]
\end{Lemma}
\begin{Proof}\rm
Let $t \in (0,1)$ and $T=t^{-\frac{1}{2}}$.
By Lemma \ref{fourierdecay}, we see that for any $T>1$,
\begin{eqnarray}
\widetilde{C}_2\|f\|^2
\geq
T^{\alpha-1}
\int_{\mathbb{R}}
|\widehat{f\mu}(\xi)|^2 e^{-(\frac{\xi}{T})^2}d\xi
\geq
e^{-1}
T^{\alpha-1}
\int_{0}^T
|\widehat{f\mu}(\xi)|^2 d\xi.
\nonumber
\end{eqnarray}
This implies our assertion.
\qed
\end{Proof}
\begin{Lemma}\label{cor1}
Let $\mu$ be a uniformly $\alpha$-H\"{o}lder continuous and finite measure.
Then there exists $\widetilde{C}_4=\widetilde{C}_4(\alpha,\mu)>0$ such that for any $f \in L^2(\mathbb{R}, d\mu)$,
\begin{equation}
\sup_{T\geq 1}
T^{\alpha-1}
\int_{0}^{\infty}
e^{-\frac{t}{T}}
|\widehat{f\mu}(t)|^2dt
\leq
\widetilde{C}_4\|f\|^2.
\nonumber
\end{equation}
\end{Lemma}
\begin{Proof}\rm
Let $f \in L^2(\mathbb{R}, d\mu)$.
Then, by Lemma \ref{cor}, we see that for any $T>1$,
\begin{eqnarray}
T^{\alpha-1}
\int_{0}^{\infty}e^{-\frac{t}{T}}|\widehat{f\mu}(t)|^2dt
&=&
\lim_{N\rightarrow \infty}
T^{\alpha-1}
\int_{0}^{T(N+1)}e^{-\frac{t}{T}}|\widehat{f\mu}(t)|^2dt
\nonumber
\\
&=&
\lim_{N\rightarrow \infty}
T^{\alpha-1}
\sum_{n=0}^{N}
\int_{Tn}^{T(n+1)}e^{-\frac{t}{T}}|\widehat{f\mu}(t)|^2dt
\nonumber
\\
&=&
\sum_{n=0}^{\infty}
(n+1)^{1-\alpha}e^{-n}
\{T(n+1)\}^{\alpha-1}
\int_{Tn}^{T(n+1)}|\widehat{f\mu}(t)|^2dt
\nonumber
\\
&\leq&
\widetilde{C}_3 \|f\|^2
\sum_{n=0}^{\infty}
(n+1)^{1-\alpha}e^{-n}.
\nonumber
\end{eqnarray}
This implies our assertion.\qed
\end{Proof}
\section{}
In this section we give some lemmas about quadratic form theory which is used in Seciotin 4.
Let $\mathcal{H}$ ba a complex Hilbert space. Let $\mathfrak{s}:\mathcal{H}\times \mathcal{H}\rightarrow \mathbb{C}$
be a closed sesquilinear form, and
$T:\mathcal{H}\rightarrow \mathcal{H}$ be a closed linear operator.
We say that $\mathfrak{s}$ is symmetric, if
$\mathfrak{s}(f,g)=\overline{\mathfrak{s}(g,f)}$
for $f,g \in \mathcal{D}(\mathfrak{s})$, and that
$\mathfrak{s}$ is sectorial, if
there exist $r\in \mathbb{R}$ and $\theta \in (-\frac{\pi}{2}, \frac{\pi}{2})$ such that
for $f \in \mathcal{D}(\mathfrak{s})$ with $\|f\|=1$,
\[
{\rm arg}(\mathfrak{s}[f]-r)\leq \theta,
\]
where $\mathfrak{s}[f]= \mathfrak{s}(f,f)$. We say that $T$ is sectorial, if
there exist $r\in \mathbb{R}$ and $\theta \in (-\frac{\pi}{2}, \frac{\pi}{2})$ such that
for $f \in \mathcal{D}(\mathfrak{s})$ with $\|f\|=1$,
\[
{\rm arg}((f,Tf)-r)\leq \theta,
\]
and that $T$ is m-accretive, if
${\rm Re}(f,Tf)\geq 0$ for $f \in \mathcal{D}(T)$ and
$(T+\lambda)^{-1}$ is bounded and $\|(T+\lambda)^{-1}\|\leq ({\rm Re}\lambda)^{-1}$
for $\lambda \in \mathbb{C}$ with ${\rm Re} \lambda > 0$.
In particular, $T$ is said to be quasi m-accretive, if
there exists $\gamma\in \mathbb{R}$ such that
$T+\gamma$ is m-accretive, and
$T$ is said to be m-sectorial, if
$T$ s quasi m-accretive and sectorial.
\begin{Lemma}
Let $\mathfrak{s}:\mathcal{H}\times \mathcal{H}\rightarrow \mathbb{C}$
be a densely defined, closed, and sectorial sesquilinear form. Then
there exist a unique m-sectorial operator $S:\mathcal{H}\rightarrow \mathcal{H}$
such that for $f \in \mathcal{D}(\mathfrak{s}), g\in \mathcal{D}(S)$,
\begin{equation}
\mathfrak{s}(f,g)=(f,Sg).
\nonumber
\end{equation}
\end{Lemma}
\begin{Proof}\rm
\cite[VI, \S2, Theorem 2.1]{Kato}
\end{Proof}
\begin{Lemma}\label{quadratic thm}
Let $\mathfrak{t}:\mathcal{H}\times \mathcal{H}\rightarrow \mathbb{C}$ be
a densely defined, closed, and symmetric form bounded from below, and
let $T$ be the self-adjoint operator associated with $\mathfrak{t}$.
Suppose that $\mathfrak{s}:\mathcal{H}\times \mathcal{H}\rightarrow \mathbb{C}$ is
a relatively bounded sesquilinear form with respect to $\mathfrak{t}$ such that
for any $f \in \mathcal{D}(\mathfrak{t}) \subset \mathcal{D}(\mathfrak{s})$,
\[
|\mathfrak{s}[f]|\leq a\mathfrak{t}[f]+b\|f\|^2, \qquad 0<a<1,\;b\geq0.
\]
Then $\mathfrak{t}^{\prime}= \mathfrak{s}+\mathfrak{t}$ is sectorial and closed.
Let $T^{\prime}$ be the m-sectorial operators associated with $\mathfrak{t}^{\prime}$.
If $0<\gamma<1$, $z\in \rho(T)$ and
\[
2\|(aT+b)(T-z)^{-1}\|\leq \gamma <1,
\]
then
$z\in \rho (T^{\prime})$ and
\[
\|(T^{\prime}-z)^{-1}-(T-z)^{-1}\|\leq \cfrac{4\gamma}{(1-\gamma)^2}\|(T-z)^{-1}\|.
\]
\end{Lemma}
\begin{Proof}\rm
\cite[VI, \S3, Theorem 3.9]{Kato}
\qed
\end{Proof}
\end{appendices}
\section*{Acknowledgement}
This work was supported by JST SPRING, Grant Number JPMJSP2136.
|
\section{Introduction}
Let $S_g$ be a closed orientable surface of genus $g$. A \emph{simple closed curve} $\alpha$ on $S_g$ is a continuous injective map $\alpha:S^1\rightarrow S_g$, where $S^1$ is the unit circle. For two simple closed curves $\alpha$ and $\beta$, the \emph{geometric intersection number} between them, denoted by $i(\alpha,\beta)$, is defined by $$i(\alpha,\beta) = \min\limits_{\alpha'\sim \alpha} |\alpha'\cap \beta|,$$ where $\sim$ stands for the free homotopy relation. Two simple closed curves $\alpha$ and $\beta$ are said to be in \emph{minimal position} if their geometric intersection number satisfies $i(\alpha,\beta)=|\alpha\cap\beta|$ (for more details, we refer the readers to~\cite{farb2011primer}, Section 1.2.3).
A set $\Omega=\{\alpha_1,\hdots,\alpha_n\}$ of simple closed curves on $S_g$ is called a \emph{filling set} if for $i\neq j$, the curves $\alpha_i$ and $\alpha_j$ are in minimal position and $S_g\setminus(\cup_{i=1}^{n}\alpha_i)$ is a disjoint union of topological disks. If the number of curves in a filling set is $n=2$, then it is called as \emph{filling pair} and if the number of discs in $S_g\setminus(\cup_{i=1}^{n}\alpha_i)$ is minimum or equivalently the total number of intersection points between the curves is minimum, then the filling system is called minimally intersecting.
The study of filling pairs has its importance for its various applications. For example, Thurston has used filling sets of size two to construct the most nontrivial mapping class group elements, so-called pseudo-Anosov homeomorphisms (for details see~\cite{farb2011primer}, Section 13.2.3). More recently, in~\cite{penner1988construction}, Penner has extended Thurston's construction to pairs of filling multi-curves as follows: given multi-curves $A= \{ a_1, \dots, a_m \}$ and $B=\{b_1,\dots,b_n \}$, such that $A\cup B$ fills the surface S. Then a homeomorphism $f$ which is the product of $T_{a_i}$ and $T_{b_j}^{-1}$, where each $a_j$ and $b_k$ occur at least once, is a pseudo-Anosov homeomorphism. Note that $T_{a_i}$ denotes the Dehn twist about the simple closed curve $a_i$. Another motivation for studying filling sets is in the problem of construction of spine for the moduli space of $S_g$. Thurston has proposed $\chi_g$, the set of all closed hyperbolic surfaces of genus $g$ having a filling set of systolic geodesics, as a candidate spine of moduli space of $S_g$. He has given a sketch of a proof but unfortunately, that appears difficult to understand.
Fillings of surfaces has been studied extensively by Aougab, Huang \cite{aougab2015minimally}, Sanki \cite{sanki2018filling}, Jeffreys \cite{jeffreys2019minimally}, Aougab, Menasco and Nieland \cite{aougab2022origamis} and many others \cite{anderson2011small}, \cite{MR3877282}. In \cite{aougab2015minimally}, the authors have shown the existence of minimally intersecting filling pair on $S_g$ for $g\geq3$ and estimated the bounds of the number of $\mathrm{Mod}(S_g)$-orbits of such filling. In \cite{sanki2018filling}, the author has generalised the result in \cite{aougab2015minimally}, where he has shown the existence of filling pair on $S_g$ for $g\geq2$ with given any number of complementary regions. Luke Jeffryes further extended the study of minimally intersecting filling pair on punctured surfaces. In \cite{jeffreys2019minimally}, he has shown the existence of minimally intersecting filling pair on punctured surfaces. Furthermore, in \cite{aougab2022origamis}, the authors have considered those minimally intersecting filling pair for which the geometric intersection number and algebraic intersection number are same and they find an upper bound for number of $\mathrm{Mod}(S_g)$-orbits of such pairs.
The simple closed curves on a surface are classified into two classes: separating and non-separating. A simple closed curve $\alpha$ on $S_g$ is called \emph{separating} if $S_g\setminus\alpha$, the complement of $\alpha$ in $S_g$, is disconnected. Otherwise, it is called \emph{non-separating}. As per our knowledge, all the works done so far on filling systems of closed surfaces are with non-separating closed curves. In this paper, we study the topology, geometry and combinatorics of filling pairs containing at least one separating curve. A filling pair $(\alpha,\beta)$, where at least one of $\alpha$ and $\beta$ is a separating simple closed curve, is called a \emph{separating filling pair}. We consider $\alpha$ is a separating curve in a separating filling pair $(\alpha,\beta)$.
Suppose $(\alpha,\beta)$ is a filling pair on $S_g$. We have the following natural question.
\begin{question}\label{q:1}
How to determine whether a filling pair $(\alpha,\beta)$ is a separating filling pair?
\end{question}
In our first result, we answer Question~\ref{q:1}. In particular, we prove a characterisation theorem which gives a necessary and sufficient condition for a curve in a filling pair to be separating. The key ingredient used in the proof is topological graph theory.
Given a filling pair $(\alpha,\beta)$, we associate a number $b$, the number of topological disks in the complement. A simple Euler's characteristic argument implies that
$i(\alpha,\beta)=2g-2+b.$
If $(\alpha,\beta)$ is a separating filling pair, then we have $b\geq2$ and hence $i(\alpha,\beta)\geq2g$. If $i(\alpha,\beta)=2g$, we call it as a \emph{minimally intersecting separating filling pair} or simply a \emph{minimal separating filling pair}. In~\cite{aougab2015minimally}, Aougab and Huang have shown that for $g\geq 3$, there exists a filling pair on $S_g$ with exactly one connected component in the complement. In the context of minimal separating filling pair, we study an analogous question. We prove the existence of minimal separating filling pair in the following theorem.
\begin{theorem}\label{theorem:2}
There exists a minimally intersecting separating filling pair on $S_g$ if and only if g is even and $g\geq4$.
\end{theorem}
The proof of Theorem~\ref{theorem:2} is constructive. We explicitly construct such a filling pair on $S_4$ and for the existence on $S_g, g\geq 6$, we use the method of induction on $g$. The converse part is proved using the standard Euler's characteristic formula.
The mapping class group $\mathrm{Mod}(S_g)$ of a surface $S_g$ of genus $g$ is the group of all orientation preserving self-homeomorphisms up to isotopy (\cite{farb2011primer}, Chapter 2). The group $\mathrm{Mod}(S_g)$ acts on the set of minimally intersecting separating filling pair as follows: for a minimally intersecting separating filling pairs $(\alpha,\beta)$ and $f\in \mathrm{Mod}(S_g)$, $f\cdot(\alpha,\beta)=(f(\alpha),f(\beta))$. We estimate the number of $\mathrm{Mod}(S_g)$-orbits of this action in the following theorem
\begin{theorem}\label{thm:1.3}
If $N(g)$ is the number of $\mathrm{Mod}(S_g)$-orbits of minimally intersecting separating filling pair, then
\begin{align*}
\frac{\prod\limits_{k=1}^{\frac{g-4}{2}}(3k+5)}{4\times (2g)^2\times (\frac{g-4}{2})!}\leq N(g) \leq 2(2g-2)(2g-2)!.
\end{align*}
\end{theorem}
The moduli space $\mathcal{M}_g$ of genus $g\geq 2$ is the collection of all hyperbolic metrics on $S_g$ up to isometry. As in \cite{aougab2015minimally}, we construct a topological Morse function on $\mathcal{M}_g$ using the length function of filling pair. Consider the set, $$ \mathcal{C}_g=\{(\alpha,\beta): (\alpha,\beta)\text{ is a minimally intersecting separating filling pair on } S_g\}.$$
For $X\in \mathcal{M}_g$, the length $l_X(\alpha,\beta)$ of $(\alpha,\beta)\in \mathcal{C}_g$ is defined as $l_X(\alpha,\beta)=l_X(\alpha)+l_X(\beta)$ where $l_X(\alpha)$ is the length of the geodesic in the free homotopy class of $\alpha$ with respect to the hyperbolic metric X. Now we define a function $\mathcal{F}_g :\mathcal{M}_g \longrightarrow \mathbb{R}$, by $$
\mathcal{F}_g(X)=\min \{ l_X(\alpha,\beta) | (\alpha,\beta) \in \Sigma_g\}.$$
We show that,
\begin{theorem}\label{theorem_lenth_function}
$\mathcal{F}_g$ is a proper and topological Morse function. For any $X\in \mathcal{M}_g$, we have
\begin{center}
$\mathcal{F}_g\geq m_g,$
\end{center}
where
\begin{align*}
m_g =4g\times \cosh ^{-1}\left( 2\left[\cos \left(\frac{2\pi}{4g}\right)+\frac{1}{2}\right]\right)
\end{align*}
denotes the perimeter of a regular right angled $4g$-gon.\\
Suppose $\mathcal{B}_g=\{X\in\mathcal{M}_g:\mathcal{F}_g(X)=m_g\}$. Then $\mathcal{B}_g$ is a finite set and $|\mathcal{B}_g|=N(g)$.
\end{theorem}
\section{Fat Graphs}
In this section, we recall some definitions and notations on fat graphs which are essential in the subsequent sections. We begin with recalling the definition of a graphs. The definition of graphs used here is not the standard one, but it is straightforward to see that it is equivalent to the standard one. Such a definition is convenient for defining fat graphs.
\begin{definition}
A graph $G$ is a triple $G=(E,\sim,\sigma_1)$, where
\begin{enumerate}
\item $E=\{\vec{e}_1,\cev{e}_1,\dots,\vec{e}_n,\cev{e}_n\}$ is a finite, non-empty set, called the set of directed edges.
\item $\sim$ is an equivalence relation on $E$.
\item $\sigma_1$ is a fixed point free involution.
\end{enumerate}
\end{definition}
The fixed-point free involution $\sigma_1$ maps a directed edge to its reverse directed edge, i.e.,
$\sigma_1(\vec{e})=\cev{e},$ for all $\vec{e}\in E$. The equivalence relation $\sim$ is defined by $\vec{e}_1\sim\vec{e}_2$ if $\vec{e}_1$ and $\vec{e}_2$ have same initial vertex. In an ordinary language, $V=E/\sim$ is the set of all vertices and $E/\sigma_1$ is the set of un-directed edges of the graph. The number of edges incident at a vertex is called the \textit{degree} of the vertex (for more details, we refer to~\cite{sanki2018filling}, section 2). Now, we define fat graphs in the following.
\begin{definition}
A \textit{fat graph} is a quadruple $\Gamma=(E,\sim,\sigma_1,\sigma_0)$, where
\begin{enumerate}
\item $G=(E,\sim,\sigma_1)$ is a graph.
\item $\sigma_0$ is a permutation on E so that each cycle corresponds to a cyclic order on the set of oriented edges going out from a vertex.
\end{enumerate}
\end{definition}
\begin{definition}
A fat graph is called \textit{decorated} if degree of each vertex is even and at least 4.
\end{definition}
\begin{definition}
A simple cycle in a decorated fat graph is called a \textit{standard cycle} if every two consecutive edges in the cycle are opposite to each other in the cyclic order on the set of edges incident at their common vertex. If a cycle is not standard,
we call it as non-standard.
\end{definition}
\textbf{Surface associated to a fat graph.}
Given a fat graph $\Gamma=(E, \sim, \sigma_1, \sigma_0)$, we construct an oriented topological surface $\Sigma(\Gamma)$ with boundary by following: Consider a closed disk corresponding to each vertex and a rectangle corresponding to each un-directed edge. Then identify the sides of the rectangles with the boundary of the discs according to the order of the edges incident to the vertex. See Figure~\ref{loc_pic_of_fat_graph} for a local picture. The boundary of a fat graph $\Gamma$ is defined as the boundary of the surface $\Sigma(\Gamma).$
\tikzset{->-/.style={decoration={
markings,
mark=at position #1 with {\arrow{>}}},postaction={decorate}}}
\tikzset{-<-/.style={decoration={
markings,
mark=at position #1 with {\arrow{<}}},postaction={decorate}}}
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw (-1.5,0)--(1.5,0);
\draw (0,1.5)--(0,-1.5);
\draw[->-=1] (0.6062,.35) arc (30:240:.7);
\draw (5,0) circle (.7cm);
\draw (3,.2)--(4.34,.2);
\draw (3,-.2)--(4.34,-.2);
\draw (5.66,.2)--(7,.2);
\draw (5.66,-.2)--(7,-.2);
\draw (4.8,.66)--(4.8,2);
\draw (5.2,.66)--(5.2,2);
\draw (4.8,-.66)--(4.8,-2);
\draw (5.2,-.66)--(5.2,-2);
\end{tikzpicture}
\end{center}
\caption{Local picture of the surface obtained from a fat graph.}
\label{loc_pic_of_fat_graph}
\end{figure}
\begin{example}
Let us consider the fat graph $\Gamma=(E,\sim,\sigma_1,\sigma_0)$, where
\begin{enumerate}[label=(\roman*)]
\item $E=\{\vec{e}_i,\cev{e}_j|i,j=1,2,3\}$, is the set of all directed edges and $E_1=\{e_1, e_2, e_3\}$ is the set of all un-directed edges, where $e_i=\{\vec{e}_1, \cev{e}_i\}, i=1,2,3.$
\item $V=\{v_1, v_2\}$, where $v_1=\{\vec{e}_1,\vec{e}_2,\vec{e}_2\}$ and $v_2=\{\cev{e}_1,\cev{e}_2,\cev{e}_3\}$, is the set of vertices.
\item $\sigma_1(\vec{e_i})=\cev{e_i}$, $i=1,2,3,$ and
\item $\sigma_0=(\vec{e}_1,\vec{e}_2,\vec{e}_2)(\cev{e}_1,\cev{e}_2,\cev{e}_3)$.
\end{enumerate}
The fat graph $\Gamma$ and the associated surface $\Sigma(\Gamma)$ is given in Figure~\ref{fat_graph_example_two_vertices}. A simple Euler characteristic argument and topological classification of surfaces imply that $\Sigma(\Gamma)$ is homeomorphic to a compact surface of genus one and with a single boundary component. We also observe that if the boundary of the surface is labelled with the convention as in Figure \ref{fat_graph_example_two_vertices}, the boundary corresponds to the cycles of $\sigma_0^{-1}\sigma_1$ which is $(\vec{e}_1,\cev{e}_3,\vec{e}_2,\cev{e}_1,\vec{e}_3,\cev{e}_2)$. We generalize this observation in Lemma \ref{lemma:2.6}.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1, yscale=1]
\draw[cyan][->-=.35] (0,0)..controls(1,-0.1)and(2.5,.2)..(3,-.25);
\draw[cyan][-{Stealth[color=cyan]}] (4,0)..controls(5,.1)and(5,-.1)..(4.5,-.3);
\draw[cyan] (3.15,-.4)..controls(3.4,-.6)and(4.3,-.4)..(4.5,-.3);
\draw[blue, ->-=.25, -<-=.74] (0,0)..controls(1.5,.7)and(1.5,.7)..(4,0);
\draw[lime, ->-=.25, -<-=.85] (0,0)..controls(1.5,-.7)and(2.5,-.7)..(4,0);
\draw[magenta] (0,0) node{$\bullet$} (4,0) node{$\bullet$};
\draw[blue, ->-=.5] (6.5,0)..controls(8,.7)and(9,.7)..(10.5,0);
\draw[blue, -<-=.5] (6.1,0)to[bend left=40](10.9,0);
\draw[lime, -<-=.5] (6.5,-.22) to [bend right=40] (9,-.3);
\draw[cyan](6.5,0)..controls(7,0)and(8,.2)..(8.8,-.45);
\draw[cyan](6.5,-.22)..controls(7,-.28)and(7.8,-.2)..(8.52,-.57);
\draw[cyan, -<-=.4](9,-.69)..controls(9.5,-1)and(10.5,-1.2)..(10.8,-1);
\draw[cyan](10.8,-1)to [bend right=50] (10.8,-.3);
\draw[cyan, ->-=.3](8.75,-.8)..controls(9.3,-1.2)and(10.5,-1.4)..(10.9,-1.2);
\draw[cyan](10.9,-1.2)..controls(11.2,-1)and(11.3,-.7)..(11.1,-.25);
\draw[lime, ->-=.5] (6.2,-.3) to [bend right=35] (9,-.65);
\draw[lime](9,-.65)..controls(9.5,-.35)and(10,-.25)..(10.5,-.25);
\draw[lime] (9,-.3)..controls(9.5,-0.05)and(10,0)..(10.5,0);
\draw[lime] (6.1,0) to[bend right=15](6.2,-.3);
\draw[lime] (10.8,-.3) to[bend right=10](10.5,-.25);
\draw[blue](10.9,0)to[bend left=10](11.1,-.25);
\draw (1.05,.65) node{\tiny$\vec{e}_1$};
\draw (1.2,.2) node{\tiny$\vec{e}_2$};
\draw (1.1,-.6) node{\tiny$\vec{e}_3$};
\draw (3,.55) node {\tiny{$\cev{e_1}$}} (3.3,0) node{\tiny$\cev{e_3}$} (4.3,-.2) node{\tiny$\cev{e_2}$};
\draw (8.45,1.15) node {\tiny{$\cev{e_1}$}} (8.45,.3) node{\tiny$\vec{e}_1$} (7.5,-.52) node{\tiny$\cev{e}_3$} (7.7,-1.15) node{\tiny$\vec{e_3}$} (9.9,-.8) node{\tiny$\cev{e_2}$} (9.4,-1.3) node{\tiny$\vec{e}_2$};
\end{tikzpicture}
\end{center}
\caption{Example of fat graph and associated surface.}
\label{fat_graph_example_two_vertices}
\end{figure}
\end{example}
\begin{lemma}\label{lemma:2.6}
The number of boundary components of a fat graph $\Gamma=(E, \sim, \sigma_1, \sigma_0)$ is same as the number of disjoint cycles in $\sigma_0^{-1}\sigma_1$.
\end{lemma}
\begin{proof}
For proof see Lemma 2.3 of \cite{sanki2018filling}.
%
\end{proof}
\section{Classification of curves on fat graphs}
Decorated fat graphs play an important role in our construction of fillings on surfaces. Such a graph can be written as an edge disjoint union of its standard cycles. The set of standard cycles corresponds to a set of simple closed curves that gives a filling. In this section, we classify the standard cycles of a decorated fat graph in Theorem~\ref{classif_of_std_cycle_in_fat_graph}. Before going to the theorem directly, let us consider the following motivating example.
\begin{example}\label{example}
Consider the filling pair $(\alpha, \beta)$ on $S_2$ as shown in Figure~\ref{sep_fill_on_S_2} and the corresponding fat graph $\Gamma(\alpha,\beta)=(E,\sim,\sigma_1,\sigma_0)$ as described below (see Figure~\ref{fatgraph_for_fill_on_S_2}):
\begin{enumerate}
\item $E=\{x_i,x_i^{-1},y_i,y_i^{-1}\mid i=1,\dots,6\}$.
\item The equivalence classes of $\sim$ are
\begin{align*}
v_1&=\{x_1,y_6^{-1},x_6^{-1},y_1\},\\
v_2&=\{x_2,y_2^{-1},x_1^{-1},y_3\},\\
v_3&=\{x_3,y_6,x_2^{-1},y_5^{-1}\},\\
v_4&=\{x_4,y_4,x_3^{-1},y_3^{-1}\},\\
v_5&=\{x_5,y_2,x_4^{-1},y_1^{-1}\} \text{ and}\\
v_6&=\{x_6,y_4^{-1},x_5^{-1},y_5^{-1}\}.
\end{align*}
\item $\sigma_1(x_i)=x_i^{-1}, \sigma_1(y_i)=y_i^{-1}, i=1,\dots,6.$
\item $\sigma_0=\prod\limits_{i=1}^6\sigma_{v_i},$ where
\begin{align*}
\sigma_{v_1}&=(x_1,y_6^{-1},x_6^{-1},y_1),\\
\sigma_{v_2}&=(x_2,y_2^{-1},x_1^{-1},y_3),\\
\sigma_{v_3}&=(x_3,y_6,x_2^{-1},y_5^{-1}),\\
\sigma_{ v_4}&=(x_4,y_4,x_3^{-1},y_3^{-1}),\\
\sigma_{ v_5}&=(x_5,y_2,x_4^{-1},y_1^{-1}) \text{ and}\\
\sigma_{v_6}&=(x_6,y_4^{-1},x_5^{-1},y_5^{-1}).
\end{align*}
\end{enumerate}
In this graph, $\alpha=(x_1, \dots,x_6)$ and $\beta=(y_1,\dots,y_6)$ are the standard cycles.
The curve $\alpha$ is separating.
We construct a $6\times4$ matrix $M(\alpha,\beta)$ whose rows are the cycles of the fat graph structure which are arranged such that edges $x_i$'s occur in the first position. For instance, the first row of $M(\alpha,\beta)$ is $(x_1,y_6^{-1},x_6^{-1},y_1)$. The matrix is shown in Figure~\ref{fatgraph_matrix}. We call this matrix $M(\alpha,\beta)$ as the \emph{normal matrix}.
We observe that for each $i$, $y_i$ and $y_i^{-1}$ are in the same column.
\end{example}
For any fat graph with two standard cycles, its normal matrix is similarly defined. Furthermore, it has all the properties as discussed in example~\ref{example}. Now, we are ready to state and prove the classification theorem.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1.5,yscale=1.5]
\draw [blue] (0,0) ellipse (3cm and 1.2cm);
\draw (-2,.05) to [bend right] (-1.2,.05);
\draw (-2,-0.05) to [bend left] (-1.2,-.05);
\draw (1.2,.05) to [bend right] (2,.05);
\draw (1.2,-.05) to [bend left] (2,-.05);
\draw [red, -<-=.4] (0,1.2) arc
[
start angle=90,
end angle=270,
x radius=.3cm,
y radius =1.2cm
] ;
\draw [dashed, red] (0,-1.2) arc
[
start angle=270,
end angle=450,
x radius=.3cm,
y radius =1.2cm
] ;
\draw[red] (-.5,0) node {$\alpha$};
\draw[green] (-1.3,-.8) node {$\beta$};
\draw (1.1,.8) node {$y_1$};
\draw (-1,-.32) node {$y_2$};
\draw (2.8,0) node {$y_3$};
\draw (-2,-.41) node {$y_4$};
\draw (1.3,.2) node {$y_5$};
\draw (-1,.88) node {$y_6$};
\draw (-.3,1.1) node {$x_1$};
\draw (.45,0) node {$x_2$};
\draw (-.37,-.9) node {$x_3$};
\draw (-.45,-.5) node {$x_4$};
\draw (-.15,0) node {$x_5$};
\draw (-.1,.67) node {$x_6$};
\draw [green,-<-=.35] (2.3,0) arc
[
start angle=0,
end angle=180,
x radius=2.5cm,
y radius =.8cm
] ;
\draw [green] (-2.7,0) arc
[
start angle=180,
end angle=258,
x radius=2cm,
y radius =1.12cm
] ;
\draw [dashed, green] (-1.2,-1.09) arc
[
start angle=265,
end angle=360,
x radius=2.2cm,
y radius =1.12cm
] ;
\draw [green] (1.2,0) arc
[
start angle=0,
end angle=260,
x radius=1.7cm,
y radius =.6cm
] ;
\draw [green] (-.8,-.59) arc
[
start angle=240,
end angle=300,
x radius=2.3cm,
y radius =.5cm
] ;
\draw [green] (1.5,-.59) arc
[
start angle=300,
end angle=360,
x radius=2.3cm,
y radius =.5cm
] ;
\draw [green] (2.65,-.18) arc
[
start angle=0,
end angle=60,
x radius=2.4cm,
y radius =1.41cm
] ;
\draw [dashed, green] (1.45,1.045) arc
[
start angle=70,
end angle=180,
x radius=1.98cm,
y radius =1.12cm
] ;
\draw [green] (-1.211,0) arc
[
start angle=180,
end angle=360,
x radius=1.752cm,
y radius =.4cm
] ;
\end{tikzpicture}
\end{center}
\caption{Filling pair $(\alpha, \beta)$ on $S_2$}
\label{sep_fill_on_S_2}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw[red] (0,0) to (0,7);
\draw [green] (0,1) to (2,1);
\draw [green] (0,2) to (4,2);
\draw [green] (0,3) to (3,3);
\draw [green] (0,4) to (2,4);
\draw [green] (0,5) to (3,5);
\draw [green] (0,6) to (4,6);
\draw [green] (4,6) to (4,2);
\draw [green] (3,3) to (3,5);
\draw [green] (2,4) to (2,3.1);
\draw [green] (2,2.9) to (2,2.1);
\draw [green] (2,1.9) to (2,1);
\draw[red] (0,0) to (5,0);
\draw[red] (5,0) to (5,7);
\draw[red] (5,7) to (0,7);
\draw [green] (0,2) to (-2,2);
\draw [green] (-2,2) to (-2,4);
\draw [green] (-2,4) to (0,4);
\draw [green] (0,1) to (-3,1);
\draw [green] (-3,1) to (-3,5);
\draw [green] (-3,5) to (0,5);
\draw [green] (0,6) to (-4,6);
\draw [green] (-4,6) to (-4,3);
\draw [green] (-4,3) to (-3.1,3);
\draw [green] (-2.9,3) to (-2.1,3);
\draw [green] (-1.9,3) to (0,3);
\draw [-{Stealth[scale=1.1]}] (0,0.5) -- (0,0.51);
\draw [-{Stealth[scale=1.1]}] (0,1.5) -- (0,1.51);
\draw [-{Stealth[scale=1.1]}] (0,2.5) -- (0,2.51);
\draw [-{Stealth[scale=1.1]}] (0,3.5) -- (0,3.51);
\draw [-{Stealth[scale=1.1]}] (0,4.5) -- (0,4.51);
\draw [-{Stealth[scale=1.1]}] (0,5.5) -- (0,5.51);
\draw [-{Stealth[scale=1.1]}] (0,6.5) -- (0,6.51);
\draw [-{Stealth[scale=1.1]}] (5,3.52) -- (5,3.51);
\draw [-{Stealth[color=black]}] (4,4.01)--(4,4);
\draw [-{Stealth[color=black]}] (3,4.01)--(3,4);
\draw [-{Stealth[color=black]}] (2,2.51)--(2,2.5);
\draw [-{Stealth[color=black]}] (-3,2.5)--(-3,2.51);
\draw [-{Stealth[color=black]}] (-4,4.5)--(-4,4.51);
\draw [-{Stealth[color=black]}] (-2,3.5)--(-2,3.51);
\draw (.25,6.42) node {$x_2$};
\draw (.25,5.42) node {$x_1$};
\draw (.25,4.42) node {$x_6$};
\draw (.25,3.42) node {$x_5$};
\draw (.25,2.42) node {$x_4$};
\draw (.25,1.42) node {$x_3$};
\draw (.25,0.42) node {$x_2$};
\draw (4.22,4.1) node{$y_3$};
\draw (3.22,4.1) node{$y_1$};
\draw (2.22,2.6) node{$y_5$};
\draw (-3.22,2.4) node{$y_6$};
\draw (-4.22,4.4) node{$y_2$};
\draw (-1.78,3.4) node{$y_4$};
\end{tikzpicture}
\end{center}
\caption{Fat graph corresponding to the filling pair $(\alpha, \beta)$.}
\label{fatgraph_for_fill_on_S_2}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1.5,yscale=1.5]
\draw (2,-1) node {$y_3$} (-1,-1) node {$x_2$} (0,-1) node {$y_2^{-1}$} (1,-1) node {$x_1^{-1}$};
\draw (2,-.5) node {$y_1$} (-1,-.5) node {$x_1$} (0,-.5) node {$y_6^{-1}$} (1,-.5) node {$x_6^{-1}$};
\draw (2,-3) node {$y_5$} (-1,-3) node {$x_6$} (0,-3) node {$y_4^{-1}$} (1,-3) node {$x_5^{-1}$};
\draw (2,-2.5) node {$y_1^{-1}$} (-1,-2.5) node {$x_5$} (0,-2.5) node {$y_2$} (1,-2.5) node {$x_4^{-1}$};
\draw (2,-2) node {$y_3^{-1}$} (-1,-2) node {$x_4$} (0,-2) node {$y_4$} (1,-2) node {$x_3^{-1}$};
\draw (2,-1.5) node {$y_5^{-1}$} (-1,-1.5) node {$x_3$} (0,-1.5) node {$y_6$} (1,-1.5) node {$x_2^{-1}$};
\end{tikzpicture}
\end{center}
\caption{The normal matrix $M(\alpha,\beta)$.}
\label{fatgraph_matrix}
\end{figure}
\begin{theorem}\label{classif_of_std_cycle_in_fat_graph}
Let $\Gamma=(E,\sim,\sigma_1,\sigma_0)$ be a fat graph with two standard cycles $\alpha=(x_1, \dots,x_n)$ and $\beta=(y_1,\dots,y_n)$. The cycle $\alpha$ is separating if and only if $y_i$ and $y_i^{-1}$ are in the same column of the normal matrix $M(\alpha,\beta)$, for every $i=1,\dots,n$.
\end{theorem}
\begin{proof}
Consider the surface $\Sigma(\Gamma)$, associated to the fat graph $\Gamma$. We construct a closed and oriented surface $S$ from $\Sigma(\Gamma)$ by attaching a disk along each boundary component. The simple closed curves on $S$ corresponding to the standard cycles $\alpha$ and $\beta$ are again denoted by $\alpha$ and $\beta$ respectively. If we cut the surface along $(\alpha \cup \beta)$, then we get a finite number of polygons with labelled boundary corresponding to the cycles of the permutation $\sigma_0^{-1}\sigma_1$. The curve $\alpha$ is separating if and only if whenever the polygons are identified along the edges $y_i$ and $y_i^{-1}$, then the resulting surface is still disconnected.
($\Leftarrow$) Assume that the normal matrix $M(\alpha,\beta)$ is as in Theorem~\ref{classif_of_std_cycle_in_fat_graph}. We show that $\alpha$ is separating. Consider the disjoint cycle representation of the permutation $\sigma_0^{-1}\sigma_1$. We claim that the $y$-edges contained in each cycle are either from second or forth column of $M(\alpha,\beta)$, but not from the both. The entries in each cycle are alternatively coming from $x$-edges and $y$-edges. Consider a generic cycle $(y_i,x_j,y_t,\dots)$ of the permutation $\sigma_0^{-1}\sigma_1$. To prove the claim, it suffices to show that $y_i$ and $y_t$ are in the same column. Suppose $y_i$ is in the second column, then $\sigma_1(y_i)=y_i^{-1}$ is also in the second column, by assumption on matrix $M(\alpha,\beta)$. Therefore, $\sigma_0^{-1} \left(y_i^{-1}\right) =x_j$ is in first column and $\sigma_1(x_j)=x_j^{-1}$ in third column, by the properties of $M(\alpha,\beta)$. Now, it follows that $\sigma_0^{-1}(x_j^{-1})$ is in the second column (see Figure~\ref{general_fat_graph_matrix} and Figure~\ref{multiplication_rule}). In case $y_i$ is in forth column, a similar argument shows that $y_t$ is also in forth column.
Suppose $D_2$ and $D_4$ be the collection of the polygons whose y-edges are in second and forth column respectively. Then, both $D_2$ and $D_4$ are nonempty disjoint sets. Therefore, after identifying the polygons along the y-edges, we get at least two disconnected sets. Hence, $\alpha$ is a separating curve.
($\Rightarrow$) The above process is reversible and this completes the proof.
\begin{figure}[htbp]\label{x}
\begin{center}
\begin{tikzpicture}[xscale=1.5,yscale=1.5]
\draw (-2,0) node {C1} (-1,0) node {C2} (0,0) node {C3} (1,0) node {C4};
\draw (-2,0) node {C1} (-1,0) node {C2} (0,0) node {C3} (1,0) node {C4};
\draw (-2,-1) node {$x_1$} (-1,-1) node {-} (0,-1) node {$x_{n}^{-1}$} (1,-1) node {-};
\draw (-2,-1.5) node {$x_2$} (-1,-1.5) node {-} (0,-1.5) node {$x_{1}^{-1}$} (1,-1.5) node {-};
\draw (-1.5,-1.8) node {$\vdots$} (0,-1.8) node {$\vdots$};
\draw (-2,-2.3) node {$x_r$} (-1,-2.3) node {$y_i$} (0,-2.3) node {$x_{r-1}^{-1}$} (1,-2.3) node {$y_s^{-1}$};
\draw (-1.5,-2.6) node {$\vdots$} (0,-2.6) node {$\vdots$};
\draw (-2,-3.1) node {$x_j$} (-1,-3.1) node {$y_i^{-1}$} (0,-3.1) node {$x_{j-1}^{-1}$} (1,-3.1) node {$y_k$};
\draw (-2,-3.6) node {$x_{j+1}$} (-1,-3.6) node {$y_t$} (0,-3.6) node {$x_{j}^{-1}$} (1,-3.6) node {$y_u$};
\draw (-1.5,-3.9) node {$\vdots$} (0,-3.9) node {$\vdots$};
\draw (-2,-4.4) node {$x_n$} (-1,-4.4) node {-} (0,-4.4) node {$x_{n-1}^{-1}$} (1,-4.4) node {-};
\end{tikzpicture}
\end{center}
\caption{Normal matrix}
\label{general_fat_graph_matrix}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw (-8,0) node {$y_i$};
\draw[dashed,red] (-8,0) circle (.3cm);
\draw [-{Stealth[color=black]}] (-7.6,0)--(-6,0);
\draw (-6.8,.2) node {$\sigma_1$};
\draw (-5.6,0) node {$y_i^{-1}$};
\draw [-{Stealth[color=black]}] (-5.2,0)--(-3.6,0);
\draw (-4.4,.3) node {$\sigma_0^{-1}$};
\draw (-3.2,0) node {$x_j$};
\draw[dashed,red] (-3.2,0) circle(.3cm);
\draw [-{Stealth[color=black]}] (-2.8,0)--(-1.2,0);
\draw (-2,.2) node {$\sigma_1$};
\draw (-.8,0) node {$x_j^{-1}$};
\draw [-{Stealth[color=black]}] (-.4,0)--(1.2,0);
\draw (.4,.3) node {$\sigma_0^{-1}$};
\draw (1.6,0) node {$y_t$} (2.3,0) node { $\dots$};
\draw[dashed,red] (1.6,0) circle(.3cm);
\draw [-{Stealth[color=black]}] (-8,-.3)--(-8,-1.3);
\draw (-8,-1.6) node {2nd column};
\draw [-{Stealth[color=black]}] (-5.6,-.3)--(-5.6,-1.3);
\draw (-5.6,-1.6) node {2nd column};
\draw [-{Stealth[color=black]}] (-3.2,-.3)--(-3.2,-1.3);
\draw (-3.2,-1.6) node {1st column};
\draw [-{Stealth[color=black]}] (-.8,-.3)--(-.8,-1.3);
\draw (-.8,-1.6) node {3rd column};
\draw [-{Stealth[color=black]}] (1.6,-.3)--(1.6,-1.3);
\draw (1.6,-1.6) node {2nd column};
\end{tikzpicture}
\end{center}
\caption{Encircled elements lie on same cycle of $\sigma_0^{-1}\sigma_1$.}
\label{multiplication_rule}
\end{figure}
\end{proof}
\section{Existence of minimal separating filling pair}
The goal of this section is to prove Theorem~\ref{theorem:2} which gives a necessary and sufficient condition for the existence of a minimally intersecting separating filling pair on $S_g$. We prove Proposition~\ref{prop:4.1} below which has a consequence giving the proof of only if part of Theorem~\ref{theorem:2}.
\begin{proposition}\label{prop:4.1}
Let $(\alpha, \beta)$ be a separating filling pair on $S_g$ and $S_g\setminus \alpha=S_{g_1,1}\sqcup S_{g_2,1}$, where $g_1$ and $g_2$ are the genera of $S_{g_1,1}$ and $S_{g_2,1}$ respectively. Then the number of connected components in $S_{g_1,1}\setminus \beta$ and $S_{g_2,1} \setminus \beta$ are equal if and only if $g_1=g_2$.
\end{proposition}
\begin{proof}
($\Rightarrow$) Assume that $S_{g_1,1}\setminus\beta$ and $S_{g_2,1}\setminus\beta$ both have same number of connected components, say n. The curve $\alpha$ is separating implies that the geometric intersection number $i(\alpha,\beta)$ is an even integer, say $2v$. Now, the union $(\alpha\cup\beta)$ gives a cell decomposition of $S_{g_1,1}$ and $S_{g_2,1}$. In each cell decomposition, the number of $0$-cells is $2v$, the number of $1$-cells is $3v$ and the number of 2-cells is $n$. Euler's characteristic formula implies that
\begin{align*}
&2v-3v+n=2-2g_1-1\hspace{.5cm} \\
&\implies g_1=(1-n+v)/2.
\end{align*}
A similar calculation gives $g_2=(1-n+v)/2.$
Thus, we conclude $g_1=g_2$.
($\Leftarrow$) Suppose $g_1=g_2$. Applying Euler's characteristic formula we have $g_1=(1-n_1+v)/2$ and $g_2=(1-n_2+v)/2$, where $n_i$'s are the number of components in $S_{g_i,1}\setminus \beta$, $i=1,2$ and $v=i(\alpha,\beta)/2$. Now, $g_1=g_2$ implies that $n_1=n_2$. This completes the proof.
\end{proof}
\begin{lemma}\label{lemma:4.2}
There exists no minimally intersecting separating filling pair on $S_2$.
\end{lemma}
\begin{proof}
The proof is by contradiction. Assume that there is a minimally intersecting separating filling pair $(\alpha,\beta)$ on $S_2$. Euler's characteristic formula implies that the geometric intersection number $i(\alpha,\beta)$ between the curves $\alpha$ and $\beta$ is $4$. The curves $\alpha$ and $\beta$ correspond to two standard cycles, each of length $4$, in the decorated fat graph $\Gamma(\alpha,\beta)$. We denote them by $\alpha=(x_1,x_2,x_3,x_4)$ and $\beta=(y_1,y_2,y_3,y_4)$ (see Figure~\ref{possible_normal_matrix}). Now Theorem~\ref{classif_of_std_cycle_in_fat_graph} implies that the edges $y_i$ and $y_i^{-1}$ are in the same column of the normal matrix $M(\alpha,\beta)$. There is a unique such normal matrix up to relabelling (see Figure~\ref{possible_normal_matrix}) and in the corresponding fat graph $\Gamma(\alpha,\beta)$, there is a boundary component of length two which implies that the curves $\alpha$ and $\beta$ form a bigon on the surface $S_2$. This contradicts that the curves $\alpha$ and $\beta$ are in minimal position.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw (2,0) node {$y_1$} (-1,0) node {$x_2$} (0,0) node {$y_4^{-1}$} (1,0) node {$x_1^{-1}$};
\draw (2,1) node {$y_1^{-1}$} (-1,1) node {$x_1$} (0,1) node {$y_2$} (1,1) node {$x_4^{-1}$};
\draw (2,-2) node {$y_3^{-1}$} (-1,-2) node {$x_4$} (0,-2) node {$y_4$} (1,-2) node {$x_3^{-1}$};
\draw (2,-1) node {$y_3$} (-1,-1) node {$x_3$} (0,-1) node {$y_2^{-1}$} (1,-1) node {$x_2^{-1}$};
\draw[blue] (6,-2)--(9,-2)--(9,3)--(6,3)--(6,-2);
\draw[cyan] (6,-1)--(5,-1)--(5,1)--(7,1)--(7,0)--(5.1,0);
\draw[cyan] (4.9,0)--(4,0)--(4,2)--(8,2)--(8,-1)--(6,-1);
\draw [-{Stealth[color=black]}] (6,-.5)--(6,-.49);
\draw [-{Stealth[color=black]}] (6,-1.5)--(6,-1.49);
\draw [-{Stealth[color=black]}] (6,.5)--(6,.51);
\draw [-{Stealth[color=black]}] (6,1.5)--(6,1.51);
\draw [-{Stealth[color=black]}] (6,2.5)--(6,2.51);
\draw [-{Stealth[color=black]}] (5.5,-1)--(5.49,-1);
\draw [-{Stealth[color=black]}] (5.5,0)--(5.49,0);
\draw [-{Stealth[color=black]}] (5.5,1)--(5.51,1);
\draw [-{Stealth[color=black]}] (5.5,2)--(5.51,2);
\draw [-{Stealth[color=black]}] (6.5,-1)--(6.49,-1);
\draw [-{Stealth[color=black]}] (6.5,0)--(6.49,0);
\draw [-{Stealth[color=black]}] (6.5,1)--(6.51,1);
\draw [-{Stealth[color=black]}] (6.5,2)--(6.51,2);
\draw (6.28,-.5) node{$x_1$};
\draw (6.28,.5) node{$x_2$};
\draw (5.7,1.5) node{$x_3$};
\draw (5.7,2.5) node{$x_4$};
\draw (5.4,-.75) node{$y_2$};
\draw (5.4,{1-.75}) node{$y_4$};
\draw (6.7,{3-.8}) node{$y_1$};
\draw (6.7,{2-.8}) node{$y_3$};
\end{tikzpicture}
\end{center}
\caption{Normal matrix $M(\alpha,\beta)$ and the corresponding fat graph.}
\label{possible_normal_matrix}
\end{figure}
\end{proof}
\begin{proof}[Proof of Theorem~\ref{theorem:2}]\label{proof_th_2}
($\Rightarrow$) Suppose $(\alpha,\beta)$ is a minimally intersecting separating filling pair on $S_g$. Consider $S_{g_1,1},S_{g_2,1},g_1$ and $g_2$ are as in Proposition~\ref{prop:4.1}. Then the number of components in each of $S_{g_1,1}\setminus \beta$ and $S_{g_2,1} \setminus \beta$ is one. Therefore, by Proposition~\ref{prop:4.1}, we have $g_1=g_2$ and this implies $g=g_1+g_2$, an even integer. Finally, by Lemma~\ref{lemma:4.2}, we have $g\geq 4$.
($\Leftarrow$) Suppose $g\geq 4$ is an even integer. We show that there is a minimal separating filling pair on $S_g$. The proof is by induction on $g$.
For $g=4$, the existence of minimally intersecting separating filing pair follows from Figure~\ref{min_fill_pair_on_S_4}.
Let $(\alpha_g,\beta_g)$ be a minimal separating filling pair on $S_g$ (obtained by induction) and we prove the result for $g+2$.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=.8,yscale=.6]
\draw [] (0,0) ellipse (8cm and 4cm);
\draw [dashed, thick, red] (0,-4) arc
[
start angle=270,
end angle=450,
x radius=.8cm,
y radius =4cm
] ;
\draw [thick, red,->-=.45] (0,4) arc
[
start angle=90,
end angle=270,
x radius=.8cm,
y radius =4cm
] ;
\draw (-2.8,.05) .. controls (-3.5,-.3) .. (-4.2,.05);
\draw (-2.9,-.05) .. controls (-3.5,.25) .. (-4.05,-.05);
\draw (-5.8,.05) .. controls (-6.5,-.3) .. (-7.2,.05);
\draw (-5.95,-.05) .. controls (-6.5,.25) .. (-7.05,-.05);
\draw (2.8,.05) .. controls (3.5,-.3) .. (4.2,.05);
\draw (2.9,-.05) .. controls (3.5,.25) .. (4.05,-.05);
\draw (5.8,.05) .. controls (6.5,-.3) .. (7.2,.05);
\draw (5.95,-.05) .. controls (6.5,.25) .. (7.05,-.05);
\draw[thick, cyan,->-=.5] (-1,3.9) .. controls (-2,4.1) and (-8,2) .. (-7.5,-.4);
\draw[thick, cyan] (-7.5,-.4) .. controls (-7.4,-1) and (-5,-1.3) .. (0,0);
\draw[thick,cyan] (0,0) .. controls (3.5,.9) and (7.5,2) .. (7.5,-.1);
\draw[thick,cyan] (7.5,-.1) .. controls (7.5,-.4) and (5,-3.5) .. (3.3,-3.62);
\draw[dashed,thick,cyan] (3.3,-3.62) .. controls (0,-4) and (-5,-1) .. (-5,0);
\draw [dashed, thick, cyan] (0,2) arc
[
start angle=90,
end angle=180,
x radius=5cm,
y radius =2cm
] ;
\draw [dashed, thick, cyan] (2.9,0) arc
[
start angle=0,
end angle=90,
x radius=2.9cm,
y radius =2cm
] ;
\draw[thick, cyan] (-1,-3.945) .. controls (.5,-4) and (3,-1) .. (2.9,0);
\draw [dashed, thick, cyan] (-5.8,0) arc
[
start angle=180,
end angle=260,
x radius=5.8cm,
y radius =4cm
] ;
\draw[thick, cyan] (-5.8,0) .. controls (-5.8,.3)and (-3.5,.4) .. (-3.4,.2);
\draw[dashed, thick, cyan] (-3.4,.2) .. controls (0,0.4) and (5,-3) .. (6.5,2.2);
\draw[thick, cyan] (6.5,2.2) .. controls (5,4) and (-3.2,3) .. (-2.9,.05);
\draw [dashed, thick, cyan] (-2.9,0) arc
[
start angle=180,
end angle=270,
x radius=2.9cm,
y radius =2.6cm
] ;
\draw [dashed, thick, cyan] (0,-2.6) arc
[
start angle=270,
end angle=360,
x radius=5.9cm,
y radius =2.6cm
] ;
\draw[thick, cyan](5.9,0) .. controls (5.8,.3) and (4.5,.51) .. (3.5,.2);
\draw[dashed, thick, cyan] (3.5,.2) .. controls (3,.3) and (2,3) .. (-1,3.9);
\draw[red] (-1.05,1) node {$\alpha$};
\draw[cyan] (3,2.7) node {$\beta$};
\draw (-4,2.7) node{$y_1$} (5,1.4) node{$y_2$} (-3.7,-1.8) node{$y_3$} (2.7,-1.5) node{$y_4$} (-5.5,-2) node{$y_5$} (4.3,-.9) node{$y_6$} (-3.,1) node{$y_7$} (2.7,-2.65) node{$y_8$};
\draw (-.86,3) node{$x_1$} (-.45,.9) node{$x_2$} (-1.1,-1.5) node{$x_3$} (.6,-3.75) node{$x_4$} (.2,-2.9) node{$x_5$} (1,-2) node{$x_6$} (1.1,.9) node{$x_7$} (1.,2.3) node{$x_8$};
\end{tikzpicture}
\end{center}
\caption{$(\alpha,\beta)$ is a minimal separating filling pair on $S_4$.}
\label{min_fill_pair_on_S_4}
\end{figure}
For this, let us take a filling pair $(\alpha',\beta')$ as in Figure~\ref{a_sep_fill_on_S_2} on $S_2$. Then $S_2\setminus(\alpha\cup\beta)$ is a disjoint union of 4 disks as shown in Figure~\ref{the_disjoint_union_of_4_disks}.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=2,yscale=1.7]
\draw [blue] (0,0) ellipse (3cm and 1.2cm);
\draw (-2,.05) to [bend right] (-1.2,.05);
\draw (-2,-0.05) to [bend left] (-1.2,-.05);
\draw (1.2,.05) to [bend right] (2,.05);
\draw (1.2,-.05) to [bend left] (2,-.05);
\draw [red] (0,1.2) arc
[
start angle=90,
end angle=270,
x radius=.3cm,
y radius =1.2cm
] ;
\draw [dashed, red] (0,-1.2) arc
[
start angle=270,
end angle=450,
x radius=.3cm,
y radius =1.2cm
] ;
\draw (-.1,.9) node {A};
\draw (-.15,.48) node {B};
\draw (-.16,-.25) node {$C$};
\draw (-.35,-.75) node {$D$};
\draw (.3,-1) node {$E$};
\draw (.3,.95) node {F};
\draw[red] (-.5,.1) node {$\alpha'$};
\draw[green] (-1,.9) node {$\beta'$};
\draw [green] (2.3,0) arc
[
start angle=0,
end angle=150,
x radius=2.5cm,
y radius =.8cm
] ;
\draw[green] (-2.36,.402).. controls (-3.2,0)and(-2.2,-.95)..(-1.2,-1.09);
\draw [dashed, green] (-1.2,-1.09) arc
[
start angle=265,
end angle=360,
x radius=2.2cm,
y radius =1.12cm
] ;
\draw [green] (1.2,0) arc
[
start angle=0,
end angle=260,
x radius=1.7cm,
y radius =.6cm
] ;
\draw [green] (-.8,-.59) arc
[
start angle=240,
end angle=300,
x radius=2.3cm,
y radius =.5cm
] ;
\draw[green] (1.5,-.59)..controls (3,-.5) and (3,.5)..(1.45,1.045);
] ;
\draw [dashed, green] (1.45,1.045) arc
[
start angle=70,
end angle=180,
x radius=1.98cm,
y radius =1.12cm
] ;
\draw [green] (-1.211,0) arc
[
start angle=180,
end angle=360,
x radius=1.752cm,
y radius =.4cm
] ;
\end{tikzpicture}
\end{center}
\caption{Filling pair on $S_2$.}
\label{a_sep_fill_on_S_2}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw (2,0) .. controls (1.5705,0.656) .. (1.4142,1.4142) .. controls (0.6778,1.5705) .. (0,2) .. controls (-0.6778,1.5705) .. (-1.4142,1.4142) .. controls (-1.5705,0.656) .. (-2,0) .. controls (-1.5705,-0.656) .. (-1.4142,-1.4142) .. controls (-0.6778,-1.5705) .. (0,-2) .. controls (0.6778,-1.5705) .. (1.4142,-1.4142) .. controls (1.5705,-0.656) .. (2,0);
\draw (9,0) .. controls (8.5705,0.656) .. (8.4142,1.4142) .. controls (7.6778,1.5705) .. (7,2) .. controls (6.3222,1.5705) .. (5.5858,1.4142) .. controls (5.4295,0.656) .. (5,0) .. controls (5.4295,-0.656) .. (5.5858,-1.4142) .. controls (6.3222,-1.5705) .. (7,-2) .. controls (7.6778,-1.5705) .. (8.4142,-1.4142) .. controls (8.5705,-0.656) .. (9,0);
\draw (2,-5) .. controls (.8,-4.2) .. (0,-3) .. controls (-.8,-4.2) .. (-2,-5) .. controls (-.8,-5.8) .. (0,-7) .. controls (.8,-5.8) .. (2,-5);
\draw (9,-5) .. controls (7.8,-4.2) .. (7,-3) .. controls (6.2,-4.2) .. (5,-5) .. controls (6.2,-5.8) .. (7,-7) .. controls (7.8,-5.8) .. (9,-5);
\draw (0,2.2) node {A} (1.57,1.57) node{F} (2.2,0) node{C}(1.57,-1.57) node{D} (-.1,-2.2) node{B} (-1.57,-1.57) node{C} (-2.2,0) node{F} (-1.57,1.57) node{E};
\draw (7,2.2) node {C} (8.57,1.57) node{A} (9.2,0) node{B}(8.57,-1.57) node{E} (6.9,-2.2) node{F} (5.43,-1.57) node{D} (4.8,0) node{E} (5.43,1.57) node{B};
\draw (.1,-2.8) node {E} (2.2,-5) node{D} (0,-7.2) node{B} (-2.2,-5) node {A};
\draw (7.1,-2.8) node {D} (9.2,-5) node{F} (7,-7.2) node{A} (4.8,-5) node {C};
\end{tikzpicture}
\end{center}
\caption{Topological disks after cutting $S_2$ along $\alpha'\cup\beta'$.}
\label{the_disjoint_union_of_4_disks}
\end{figure}
The point A lies on all of the four polygons. Let us extract an open ball around A which contains no other intersection points. We extract another open ball around an intersection point of $\alpha_g$ and $\beta_g$ in a similar way. We attach these two surfaces along the boundary in such a way that the end points of the arc $\alpha_g$ are identified with the end points of arc of $\alpha'$ and the same holds for the $\beta$ arcs. Then we get a filling pair $(\alpha_{g+2},\beta_{g+2})$ on $S_{g+2}$ with $\alpha_{g+2}$ separating and with number of disks in the complements is two. These completes the proof.
\end{proof}
\begin{section}{Counting of Mapping Class Group Orbits.}
The mapping class group $\mathrm{Mod}(S_g)$ acts on the set $\mathcal{C}_g$ of all minimally intersecting separating filling pairs of $S_g$ by following: $f\cdot (\alpha, \beta)=(f(\alpha), f(\beta))$, where $f\in \mathrm{Mod}(S_g)$ and $(\alpha, \beta)\in \mathcal{C}_g$. In this section, we find an upper bound of the number of orbits of this action, in particular, we prove Theorem~\ref{thm:1.3}.
\begin{subsection}{Oriented filling pair and filling permutation}\label{sec:5.1} Consider $(\alpha,\beta)\in \mathcal{C}_g$. By Euler characteristic formula, we have the geometric intersection number between $\alpha$ and $\beta$ is $2g$.
These intersection points decompose $\alpha$ into $2g$ sub-arcs. We choose an initial arc and an orientation on $\alpha$. Now, we label the sub-arcs by $\alpha_1,\dots,\alpha_{2g},$ in accordance with the orientation with $\alpha_1$ is the initial arc. Similarly, we label the sub-arcs of $\beta$ by $\beta_1, \dots,\beta_{2g}$. A filling pair $(\alpha,\beta)$ with a chosen orientation on each of $\alpha$ and $\beta$ is called an \textit{oriented filling pair} and an oriented filling pair with a labelling, we call as a \textit{labelled filling pair}.
We define a permutation $\phi^{(\alpha,\beta)}$ on $\mathrm{N}_{g}= \{1, 2, \dots, 8g\}$ corresponding to a labelled filling pair $(\alpha,\beta)$ as described below. Consider the ordered set
$$ A_g=\left\{\alpha_1,\beta_1,\hdots,\alpha_{2g},\beta_{2g},\alpha_1^{-1},\beta_1^{-1},\hdots,\alpha_{2g}^{-1},\beta_{2g}^{-1}\right\}.$$
If we cut the surface $S_g$ along $(\alpha\cup\beta)$, then we get two $4g$-gons whose sides are labelled by the members of $A_g$. Now, we define,
\begin{equation*}
\phi^{(\alpha,\beta)}(i)=j;
\end{equation*}
if the $i^{\text{th}}$ element of $A_g$ is succeeded by the $j^{\text{th}}$ element of $A_g$ along the clockwise direction of the $4g$-gons.
\begin{example}
Consider the labelled filling pair $(\alpha,\beta)$ as in Figure~\ref{min_fill_pair_on_S_4}. The labelled polygons obtained by cutting the surface along $\alpha\cup\beta$ are shown in Figure~\ref{One_of_two_poly}. Then,
\begin{align*}
\phi^{(\alpha,\beta)}&=c_1c_2, \text{ where }\\ c_1&=(1,28,13,8,7,20,5,24,15,32,11,12,3,4,9,16)\text{ and}\\
c_2&=(17,2,19,14,25,6,29,21,18,31,22,23,26,27,30).
\end{align*}
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1.5,yscale=1.5]
\draw[thick, red] ({cos(0)},{sin(0)}) to [bend left] ({cos(22.5)},{sin(22.5)});
\draw[thick, cyan] ({cos(22.5)},{sin(22.5)}) to [bend left] ({cos(45)},{sin(45)});
\draw[thick, red] ({cos(45)},{sin(45)}) to [bend left] ({cos(67.5)},{sin(67.5)});
\draw[thick, cyan] ({cos(67.5)},{sin(67.5)}) to [bend left] ({cos(90)},{sin(90)});
\draw[thick, red] ({cos(90)},{sin(90)}) to [bend left] ({cos(112.5)},{sin(112.5)});
\draw[thick, cyan] ({cos(112.5)},{sin(112.5)}) to [bend left] ({cos(135)},{sin(135)});
\draw[thick, red] ({cos(135)},{sin(135)}) to [bend left] ({cos(157.5)},{sin(157.5)});
\draw[thick, cyan] ({cos(157.5)},{sin(157.5)}) to [bend left] (-1,0);
\draw[thick, red] (-1,0) to [bend left] ({cos(202.5)},{sin(202.5)});
\draw[thick, cyan] ({cos(202.5)},{sin(202.5)}) to [bend left] ({cos(225)},{sin(225)});
\draw[thick, red] ({cos(225)},{sin(225)}) to [bend left] ({cos(247.5)},{sin(247.5)});
\draw[thick, cyan] ({cos(247.5)},{sin(247.5)}) to [bend left] ({cos(270)},{sin(270)});
\draw[thick, red] ({cos(270)},{sin(270)}) to [bend left] ({cos(292.5)},{sin(292.5)});
\draw[thick, cyan] ({cos(292.5)},{sin(292.5)}) to [bend left] ({cos(315)},{sin(315)});
\draw [thick, red]({cos(315)},{sin(315)}) to [bend left] ({cos(337.5)},{sin(337.5)});
\draw [thick, cyan]({cos(337.5)},{sin(337.5)}) to [bend left] ({cos(360)},{sin(360)});
\draw[thick, red] ({4+cos(0)},{sin(0)}) to [bend left] ({4+cos(22.5)},{sin(22.5)});
\draw[thick, cyan] ({4+cos(22.5)},{sin(22.5)}) to [bend left] ({4+cos(45)},{sin(45)});
\draw[thick, red] ({4+cos(45)},{sin(45)}) to [bend left] ({4+cos(67.5)},{sin(67.5)});
\draw[thick, cyan] ({4+cos(67.5)},{sin(67.5)}) to [bend left] ({4+cos(90)},{sin(90)});
\draw[thick, red] ({4+cos(90)},{sin(90)}) to [bend left] ({4+cos(112.5)},{sin(112.5)});
\draw[thick, cyan] ({4+cos(112.5)},{sin(112.5)}) to [bend left] ({4+cos(135)},{sin(135)});
\draw[thick, red] ({4+cos(135)},{sin(135)}) to [bend left] ({4+cos(157.5)},{sin(157.5)});
\draw[thick, cyan] ({4+cos(157.5)},{sin(157.5)}) to [bend left] (4-1,0);
\draw[thick, red] (4+-1,0) to [bend left] ({4+cos(202.5)},{sin(202.5)});
\draw[thick, cyan] ({4+cos(202.5)},{sin(202.5)}) to [bend left] ({4+cos(225)},{sin(225)});
\draw[thick, red] ({4+cos(225)},{sin(225)}) to [bend left] ({4+cos(247.5)},{sin(247.5)});
\draw[thick, cyan] ({4+cos(247.5)},{sin(247.5)}) to [bend left] ({4+cos(270)},{sin(270)});
\draw[thick, red] ({4+cos(270)},{sin(270)}) to [bend left] ({4+cos(292.5)},{sin(292.5)});
\draw[thick, cyan] ({4+cos(292.5)},{sin(292.5)}) to [bend left] ({4+cos(315)},{sin(315)});
\draw [thick, red]({4+cos(315)},{sin(315)}) to [bend left] ({4+cos(337.5)},{sin(337.5)});
\draw [thick, cyan]({4+cos(337.5)},{sin(337.5)}) to [bend left] ({4+cos(360)},{sin(360)});
\draw ({1.1*cos(11.25)},{1.1*sin(11.25)}) node{\tiny$x_7$} ({1.18*cos(33.75)},{1.18*sin(33.75)}) node {\tiny$y_6^{-1}$} ({1.1*cos(56.25)},{1.1*sin(56.25)}) node{\tiny$x_1$} ({1.1*cos(78.75)},{1.1*sin(78.75)}) node {\tiny$y_8$} ({1.1*cos(101.25)},{1.1*sin(101.25)}) node{\tiny$x_5$} ({1.1*cos(123.75)},{1.1*sin(123.75)}) node {\tiny$y_2$} ({1.1*cos(146.25)},{1.1*sin(146.25)}) node{\tiny$x_2$} ({1.1*cos(168.75)},{1.1*sin(168.75)}) node {\tiny$y_6$} ({1.1*cos(191.25)},{1.1*sin(191.25)}) node{\tiny$x_6$} ({1.2*cos(213.75)},{1.2*sin(213.75)}) node {\tiny$y_8^{-1}$} ({1.1*cos(236.25)},{1.1*sin(236.25)}) node{\tiny$x_8$} ({1.13*cos(258.75)},{1.13*sin(258.75)}) node {\tiny$y_4^{-1}$} ({1.1*cos(281.25)},{1.1*sin(281.25)}) node{\tiny$x_3$} ({1.1*cos(303.75)},{1.1*sin(303.75)}) node {\tiny$y_2^{-1}$} ({1.1*cos(326.25)},{1.1*sin(326.25)}) node{\tiny$x_4$} ({1.1*cos(348.75)},{1.1*sin(348.75)}) node {\tiny$y_4$};
\draw ({4+1.18*cos(11.25)},{1.188*sin(11.25)}) node{\tiny$x_1^{-1}$} ({4+1.18*cos(33.75)},{1.18*sin(33.75)}) node {\tiny$y_7^{-1}$} ({4+1.18*cos(56.25)},{1.18*sin(56.25)}) node{\tiny$x_6^{-1}$} ({4+1.1*cos(78.75)},{1.1*sin(78.75)}) node {\tiny$y_5$} ({4+1.12*cos(101.25)},{1.12*sin(101.25)}) node{\tiny$x_4^{-1}$} ({4+1.1*cos(123.75)},{1.1*sin(123.75)}) node {\tiny$y_3^{-1}$} ({4+1.18*cos(146.25)},{1.18*sin(146.25)}) node{\tiny$x_8^{-1}$} ({4+1.15*cos(168.75)},{1.15*sin(168.75)}) node {\tiny$y_1^{-1}$} ({4+1.18*cos(191.25)},{1.18*sin(191.25)}) node{\tiny$x_3^{-1}$} ({4+1.2*cos(213.75)},{1.2*sin(213.75)}) node {\tiny$y_5^{-1}$} ({4+1.18*cos(236.25)},{1.18*sin(236.25)}) node{\tiny$x_7^{-1}$} ({4+1.13*cos(258.75)},{1.13*sin(258.75)}) node {\tiny$y_3$} ({4+1.16*cos(281.25)},{1.16*sin(281.25)}) node{\tiny$x_5^{-1}$} ({4+1.1*cos(303.75)},{1.1*sin(303.75)}) node {\tiny$y_7$} ({4+1.18*cos(326.25)},{1.18*sin(326.25)}) node{\tiny$x_2^{-1}$} ({4+1.1*cos(348.75)},{1.1*sin(348.75)}) node {\tiny\tiny $y_1$};
\draw[red] (0,0) node{$\curvearrowright$};
\draw[red] (4,0) node{$\curvearrowright$};
\end{tikzpicture}
\end{center}
\caption{The labelled polygons}
\label{One_of_two_poly}
\end{figure}
\end{example}
Now, we study the properties of the permutation $\phi^{(\alpha,\beta)}$ in the following proposition.
\begin{proposition}\label{prop_of_phi_alpha_beta}
The permutation $\phi^{(\alpha,\beta)}$ has the following properties.
\begin{enumerate}
\item $\phi^{(\alpha,\beta)}$ is pairity respecting and sends even entries to odds and vice-versa.
\item $\phi^{(\alpha,\beta)}$ is a product of two disjoint $4g$-cycles.
\item One cycle of $\phi^{(\alpha,\beta)}$ contains each element of $\{1,3,\hdots,4g-1\}$ and the other contains each element of $\{4g+1,4g+3,\hdots,8g-1\}$.
\item If a cycle contains some even integer $2i$ then it contains $(2i+4g) (\mathrm{mod}\ 8g).$
\item If a cycle contains an even integer of the form $(4k+2)$, then all the even integers in this cycle have this form. A similar statement is true for the integers of the form $4k$.
\item $\phi^{(\alpha,\beta)}$ satisfies the equation $\phi^{(\alpha,\beta)}Q^{4g}\phi^{(\alpha,\beta)}=\tau$,
where
\begin{align*}
Q &=(1,2,\hdots,8g) \text{ and}\\ \tau &=\tau_1\tau_2\tau_3\tau_4, \text{ where }\\
\tau_1&= (1,3,\hdots,4g-1)\\
\tau_2&=(2,4,\hdots,4g)\\
\tau_3&=(8g-1,8g-3,\hdots,4g+1) \text{ and }\\
\tau_4&=(8g,8g-2,\dots,4g+2).
\end{align*}
\end{enumerate}
\end{proposition}
\begin{proof}
Let $P_1$ and $P_2$ are the $4g$-gons obtained by cutting the surface $S_g$ along $\alpha\cup \beta.$
\begin{enumerate}
\item Among every two consecutive sides of the $4g$-gons $P_1$ and $P_2$, one comes from the $\alpha$-arcs and the other from the $\beta$-arcs, which implies that $\phi^{(\alpha,\beta)}$ is pairity respecting and sends even entries to odds and vice-versa.
\item
Each polygon corresponds to a cycle of $\phi^{(\alpha,\beta)}$ and the converse is also true. Now, the statement follows from the minimality of the separating filling pair $(\alpha,\beta)$.
\item
By consideration, the curve $\alpha$ is separating, which implies that
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw (0,0) ellipse (4cm and 1.2cm);
\draw [red] (0,1.2) arc
[
start angle=90,
end angle=270,
x radius=.3cm,
y radius =1.2cm
] ;
\draw [dashed, red] (0,-1.2) arc
[
start angle=270,
end angle=450,
x radius=.3cm,
y radius =1.2cm
] ;
\draw (-3.5,.1) .. controls (-3,-.2) .. (-2.5,.1);
\draw (-3.35,0) .. controls (-3,.1) .. (-2.65,0);
\draw (-1.5,.1) .. controls (-1,-.2) .. (-.5,.1);
\draw (-1.35,0) .. controls (-1,.1) .. (-.65,0);
\draw (3.5,.1) .. controls (3,-.2) .. (2.5,.1);
\draw (3.35,0) .. controls (3,.1) .. (2.65,0);
\draw (1.5,.1) .. controls (1,-.2) .. (.5,.1);
\draw (1.35,0) .. controls (1,.1) .. (.65,0);
\draw [red] (-.1,-.2) node {$\alpha$};
\draw [thick, cyan] (-1.5,.5) .. controls (0,.7) .. (1.5,.5);
\draw [cyan] (-1,.8) node {$\beta_1$} (1,.8) node {$\beta_2$};
\end{tikzpicture}
\end{center}
\caption{Separating Curve $\alpha$ on $S_g$.}
\label{sep_curve_alpha_on_S_g}
\end{figure}
$\alpha_1,\hdots,\alpha_{2g}$ are in one polygon and $\alpha_1^{-1},\hdots,\alpha_{2g}^{-1}$ are in the other polygon.
\item The proof follows from the fact that for each $i$, the sides labelled by $\beta_i$ and $\beta_i^{-1}$ are in the same polygon.
\item The edges labelled by $\beta_i$ and $\beta_i^{-1}$, for $i\in\{1,\dots,2g\}$ odd integers, are sides of one polygon and for even $i$'s, they are the sides of the other polygon (see Figure~\ref{One_of_two_poly}). Therefore the statement follows.
\item Let $k\in \{1,2,\hdots,8g\}$. We consider the following four cases to prove the statement.
\textbf{Case 1.} Consider $k\in \{1,2,\hdots,4g\}$ is odd. Then the $k$-th element of $A_g$ is $\alpha_i$, where $i=\frac{k+1}{2}$. We refer to Figure~\ref{fig11.3}(a) for a local picture near the intersection point between the arcs $\alpha_i$ and $\alpha_{i+1}$. By Proposition~\ref{prop_of_phi_alpha_beta}, $(4)$, the sides $\beta_j$ and $\beta_j^{-1}$ lie on the same polygon. We have $\phi^{(\alpha,\beta)}(\alpha_i)=\beta_j$, $Q^{4g}(\beta_j)=\beta_j^{-1}$ and $\phi^{(\alpha,\beta)}(\beta_j^{-1})=\alpha_{i+1}$. Therefore, $\phi^{(\alpha,\beta)}Q^{4g}\phi^{(\alpha,\beta)}(\alpha_i)=\alpha_{i+1}$ (see Figure~\ref{fig11.3}(b)). This shows that $\phi^{(\alpha,\beta)}Q^{4g}\phi^{(\alpha,\beta)}(k)=\tau(k)$.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw (-8,0) to (-4,0);
\draw (-6,2) to (-6,-2);
\draw (2,0) .. controls (1.7386,0.4658) .. (1.732,1) .. controls (1.2727,1.2727) .. (1,1.732);
\draw[dashed] (2,0) .. controls (1.7386,-0.4658).. (1.5588,-1);
\draw (-2,0) .. controls (-1.7386,-0.4658) .. (-1.732,-1) .. controls (-1.2727,-1.2727) .. (-1,-1.732);
\draw[dashed] (-2,0) .. controls (-1.7386,0.4658).. (-1.5588,1);
\draw[dashed] (1,1.732) .. controls (0.83857,1.7386) .. (0,2);
\draw[dashed] (-1,-1.732) .. controls (-0.83857,-1.7386) .. (0,-2);
\draw (1.45,1.45) node {$\alpha_i$};
\draw (2.08,.5) node {$\beta_j$};
\draw (-1.7,-1.55) node {$\beta_j^{-1}$};
\draw (-2.2,-.5) node {$\alpha_{i+1}$};
\draw (0,0) node {\LARGE{$\curvearrowright$}};
\draw [-{Stealth[color=blue]}] (-5,0)--(-4.99,0);
\draw [-{Stealth[color=blue]}] (-7,0)--(-6.99,0);
\draw [-{Stealth[color=blue]}] (-6,1)--(-6,.99);
\draw [-{Stealth[color=blue]}] (-6,-1)--(-6,-1.01);
\draw (-5,-.4) node {$\alpha_{i+1}$};
\draw (-7,-.4) node {$\alpha_{i}$};
\draw (-5.5,1) node {$\beta_{j-1}$};
\draw (-5.6,-1) node {$\beta_{j}$};
\draw (-6,-3) node {(a)} (0,-3) node {(b)};
\end{tikzpicture}
\end{center}
\caption{(a) Local picture of $\alpha_i$ on the surface. (b) Polygon containing $\alpha_i$.}
\label{fig11.3}
\end{figure}
\textbf{Case 2.} In this case, we consider $k\in\{1,2,\hdots,4g\}$ is even. Then the $k$-th element of $A_g$ is $\beta_j$, where $j=\frac{k}{2}$. We refer to Figure~\ref{fig11.4}(a) for a local picture near the intersection point between the arcs $\beta_j$ and $\beta_{j+1}$. We have $\phi^{(\alpha,\beta)}(\beta_j)=\alpha_i^{-1}$, $Q^{4g}(\alpha_i^{-1})=\alpha_i$ and $\phi^{(\alpha,\beta)}(\alpha_i)=\beta_{j+1}$. Therefore, $\phi^{(\alpha,\beta)}Q^{4g}\phi^{(\alpha,\beta)}(k)=\tau(k)$.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=.8,yscale=.8]
\draw (-8,0) to (-4,0);
\draw (-6,2) to (-6,-2);
\draw (2,0) .. controls (1.7386,0.4658) .. (1.732,1) .. controls (1.2727,1.2727) .. (1,1.732);
\draw[dashed] (2,0) .. controls (1.7386,-0.4658).. (1.5588,-1);
\draw (-2,0) .. controls (-1.7386,-0.4658) .. (-1.732,-1) .. controls (-1.2727,-1.2727) .. (-1,-1.732);
\draw[dashed] (-2,0) .. controls (-1.7386,0.4658).. (-1.5588,1);
\draw[dashed] (1,1.732) .. controls (0.83857,1.7386) .. (0,2);
\draw[dashed] (-1,-1.732) .. controls (-0.83857,-1.7386) .. (0,-2);
\draw (1.55,1.45) node {$\beta_j$};
\draw (2.2,.5) node {$\alpha_i^{-1}$};
\draw (0,0) node {\LARGE{$\curvearrowright$}};
\draw [-{Stealth[color=blue]}] (-5,0)--(-4.99,0);
\draw [-{Stealth[color=blue]}] (-7,0)--(-6.99,0);
\draw [-{Stealth[color=blue]}] (-6,1)--(-6,1.01);
\draw [-{Stealth[color=blue]}] (-6,-1)--(-6,-.99);
\draw (-5,-.4) node {$\beta_{j+1}$};
\draw (-7,-.4) node {$\beta_{j}$};
\draw (-5.5,1) node {$\alpha_{i+1}$};
\draw (-5.6,-1) node {$\alpha_{i}$};
\draw (8,0) .. controls (7.7386,0.4658) .. (7.732,1) .. controls (7.2727,1.2727) .. (7,1.732);
\draw[dashed] (8,0) .. controls (7.7386,-0.4658).. (7.5588,-1);
\draw (4,0) .. controls (4.2614,-0.4658) .. (4.268,-1) .. controls (4.7273,-1.2727) .. (5,-1.732);
\draw[dashed] (4,0) .. controls (4.2614,0.4658).. (4.4412,1);
\draw[dashed] (7,1.732) .. controls (6.83857,1.7386) .. (6,2);
\draw[dashed] (5,-1.732) .. controls (5.16143,-1.7386) .. (6,-2);
\draw (7.45,1.45) node {$\alpha_i$};
\draw (8.25,.5) node {$\beta_{j+1}$};
\draw (6,0) node {\LARGE{$\curvearrowright$}};
\draw (-6,-3) node {(a)} (0,-3) node {(b)} (6,-3) node {(c)};
\end{tikzpicture}
\end{center}
\caption{(a) Local picture of $\alpha_i$ on the surface. (b) Polygon containing $\alpha_i$.}
\label{fig11.4}
\end{figure}
\textbf{Case 3.} Consider $k\in \{4g+1,4g+2,\dots,8g\}$ is odd. Then the $k$-th element of $A_g$ is $\alpha_i^{-1}$, where $i=\frac{k-4g+1}{2}$. A similar argument as in Case 1 shows that $\phi^{(\alpha,\beta)}Q^{4g}\phi^{(\alpha,\beta)}(k)=\tau(k)$ (see Figure~\ref{fig11.5}).
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=.8,yscale=.8]
\draw (-8,0) to (-4,0);
\draw (-6,2) to (-6,-2);
\draw (2,0) .. controls (1.7386,0.4658) .. (1.732,1) .. controls (1.2727,1.2727) .. (1,1.732);
\draw[dashed] (2,0) .. controls (1.7386,-0.4658).. (1.5588,-1);
\draw (-2,0) .. controls (-1.7386,-0.4658) .. (-1.732,-1) .. controls (-1.2727,-1.2727) .. (-1,-1.732);
\draw[dashed] (-2,0) .. controls (-1.7386,0.4658).. (-1.5588,1);
\draw[dashed] (1,1.732) .. controls (0.83857,1.7386) .. (0,2);
\draw[dashed] (-1,-1.732) .. controls (-0.83857,-1.7386) .. (0,-2);
\draw (1.6,1.55) node {$\alpha_i^{-1}$};
\draw (2.2,.5) node {$\beta_j^{-1}$};
\draw (-1.7,-1.55) node {$\beta_j$};
\draw (-2.2,-.5) node {$\alpha_{i-1}^{-1}$};
\draw (0,0) node {\LARGE{$\curvearrowright$}};
\draw [-{Stealth[color=blue]}] (-5,0)--(-4.99,0);
\draw [-{Stealth[color=blue]}] (-7,0)--(-6.99,0);
\draw [-{Stealth[color=blue]}] (-6,1)--(-6,.99);
\draw [-{Stealth[color=blue]}] (-6,-1)--(-6,-1.01);
\draw (-5,-.4) node {$\alpha_{i}$};
\draw (-7,-.4) node {$\alpha_{i-1}$};
\draw (-5.5,1) node {$\beta_{j-1}$};
\draw (-5.6,-1) node {$\beta_{j}$};
\draw (-6,-3) node {(a)} (0,-3) node {(b)};
\end{tikzpicture}
\end{center}
\caption{(a) Local picture of $\alpha_i$ on the surface. (b) Polygon containing $\alpha_i$.}
\label{fig11.5}
\end{figure}
\textbf{Case 4.} Consider $k\in\{4g+1,4g+2,\hdots,8g\}$ is even.
Then the $k$-th element of $A_g$ is $\beta_j^{-1}$ where $j=\frac{k-4g}{2}$. A similar argument as in Case 2 shows that $\phi^{(\alpha,\beta)}Q^{4g}\phi^{(\alpha,\beta)}(k)=\tau(k)$ (see Figure~\ref{fig11.6}).
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=.8,yscale=.8]
\draw (-8,0) to (-4,0);
\draw (-6,2) to (-6,-2);
\draw (2,0) .. controls (1.7386,0.4658) .. (1.732,1) .. controls (1.2727,1.2727) .. (1,1.732);
\draw[dashed] (2,0) .. controls (1.7386,-0.4658).. (1.5588,-1);
\draw (-2,0) .. controls (-1.7386,-0.4658) .. (-1.732,-1) .. controls (-1.2727,-1.2727) .. (-1,-1.732);
\draw[dashed] (-2,0) .. controls (-1.7386,0.4658).. (-1.5588,1);
\draw[dashed] (1,1.732) .. controls (0.83857,1.7386) .. (0,2);
\draw[dashed] (-1,-1.732) .. controls (-0.83857,-1.7386) .. (0,-2);
\draw (1.7,1.45) node {$\beta_j^{-1}$};
\draw (2.2,.5) node {$\alpha_i^{-1}$};
\draw (0,0) node {\LARGE{$\curvearrowright$}};
\draw [-{Stealth[color=blue]}] (-5,0)--(-4.99,0);
\draw [-{Stealth[color=blue]}] (-7,0)--(-6.99,0);
\draw [-{Stealth[color=blue]}] (-6,1)--(-6,.99);
\draw [-{Stealth[color=blue]}] (-6,-1)--(-6,-1.01);
\draw (-5,-.4) node {$\beta_{j}$};
\draw (-7,-.4) node {$\beta_{j-1}$};
\draw (-5.65,1) node {$\alpha_{i}$};
\draw (-5.5,-1) node {$\alpha_{i+1}$};
\draw (8,0) .. controls (7.7386,0.4658) .. (7.732,1) .. controls (7.2727,1.2727) .. (7,1.732);
\draw[dashed] (8,0) .. controls (7.7386,-0.4658).. (7.5588,-1);
\draw (4,0) .. controls (4.2614,-0.4658) .. (4.268,-1) .. controls (4.7273,-1.2727) .. (5,-1.732);
\draw[dashed] (4,0) .. controls (4.2614,0.4658).. (4.4412,1);
\draw[dashed] (7,1.732) .. controls (6.83857,1.7386) .. (6,2);
\draw[dashed] (5,-1.732) .. controls (5.16143,-1.7386) .. (6,-2);
\draw (7.5,1.4) node {$\alpha_i$};
\draw (8.3,.5) node {$\beta_{j-1}^{-1}$};
\draw (6,0) node {\LARGE{$\curvearrowright$}};
\draw (-6,-3) node {(a)} (0,-3) node {(b)} (6,-3) node {(c)};
\end{tikzpicture}
\end{center}
\caption{(a) Local picture of $\alpha_i$ on the surface. (b) Polygon containing $\alpha_i$.}
\label{fig11.6}
\end{figure}
\end{enumerate}
\end{proof}
\textbf{Note:} If a permutation $\phi\in\Sigma_{8g}$ satisfies the properties $(1)-(6)$ of Proposition \ref{prop_of_phi_alpha_beta}, then we call it a \textit{ filling permutation}.\\
Given a minimally intersecting separating filling pair $(\alpha,\beta)$ on $S_g$, we have $\phi=\phi^{(\alpha,\beta)}$, a filling permutation. In Theorem \ref{thm:5.3}, we study the converse part, in particular, we prove that every filling permutation is realized by a minimally intersecting separating filling pair.
\begin{theorem} \label{thm:5.3}
Let $\phi\in\sum_{8g}$ be a filling permutation. There exists a minimally intersecting separating filling pair $(\alpha,\beta)$ such that $\phi=\phi^{(\alpha,\beta)}$.
\end{theorem}
\begin{proof}
Consider two disjoint $4g$-gons $P_1$ and $P_2$ and an ordered set $A_g$ of $8g$ symbols given by
$$A_g=\left\{\alpha_1,\beta_1,\alpha_2,\beta_2,\hdots,\alpha_{2g},\beta_{2g},\alpha_1^{-1},\beta_1^{-1},\hdots,\alpha_{2g}^{-1},\beta_{2g}^{-1}\right\}.$$
Suppose $\phi=\phi_1\phi_2$, where $\phi_1,\phi_2$ are two disjoint cycles of length $4g$. Now, we label the edges of $P_1$ and $P_2$ by elements of $A_g$ as follows. Let $\phi_1=(a_1,\dots,a_{4g})$. Choose an initial edge of $P_1$ and choose clockwise orientation on $P_1$. We label $i^{\text{th}}$ edge of $P_1$ by $a_i$$^{th}$ element of $A_g$. Similarly, we label $P_2$. Note that, the labelling of $P_1$ and $P_2$ are well defined up to cyclic order.
The above labelling on the polygons gives a side paring on $P_1\cup P_2$ and after identifying the sides pairwise, we obtain a closed, connected and orientable surface $S$.
Now, we compute the genus of the surface $S$ using Euler's characteristic formula. By the construction of the surface $S$, we have a cell decomposition, where the number of $2$-cells is $2$ and the number of $1$-cells is $4g$. Now, we calculate the number of $0$-cells. Let us begin with a vertex $v_1$ of $P_1\cup P_2$. Suppose $\alpha_i$ and $\beta_j$ are the incident edges at the vertex $v_1$ [see Figure~\ref{fig11.8}]. By Proposition \ref{prop_of_phi_alpha_beta}, (4), the edge labelled by $\beta_j^{-1}$ is in the same polygon. By Proposition \ref{prop_of_phi_alpha_beta}, (6), the edge immediately succeeding $\beta_j^{-1}$ is $\alpha_{i+1}$. Therefore the vertex $v_1$ is identified with the vertex $v_2$ between $\beta_j^{-1}$ and $\alpha_{i+1}$. \\
By a similar reasoning, we see that the vertex $v_2$ is identified with the vertex $v_3$ between the edges $\alpha_{i+1}^{-1}$ and $\beta_{j-1}^{-1}$ and the latter one is identified with the vertex $v_4$ between the edges $\beta_{j-1}$ and $\alpha_i^{-1}$ and $v_4$ is identified with $v_1$ which completes the cycle. This shows that four vertices are identified to a single point. Therefore, the total number of $0$-cells is $2g$. Therefore, the Euler characteristic of the surface is given by
$$ \chi(S)=2g-4g+2=2-2g.$$ So, by classification theorem of surfaces, we have the genus of the surface to be $g$.
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}[xscale=1,yscale=1]
\draw (2,0) .. controls (1.7386,0.4658) .. (1.732,1) .. controls (1.2727,1.2727) .. (1,1.732);
\draw[dashed] (2,0) .. controls (1.7386,-0.4658).. (1.5588,-1);
\draw (-2,0) .. controls (-1.7386,-0.4658) .. (-1.732,-1) .. controls (-1.2727,-1.2727) .. (-1,-1.732);
\draw[dashed] (-2,0) .. controls (-1.7386,0.4658).. (-1.5588,1);
\draw[dashed] (1,1.732) .. controls (0.83857,1.7386) .. (0,2);
\draw[dashed] (-1,-1.732) .. controls (-0.83857,-1.7386) .. (0,-2);
\draw (1.6,1.45) node {$\alpha_i$};
\draw (2.1,.5) node {$\beta_j$};
\draw (-1.7,-1.55) node {$\beta_j^{-1}$};
\draw (-2.3,-.5) node {$\alpha_{i+1}$};
\draw (0,0) node {\LARGE{$\curvearrowright$}};
\draw (8,0) .. controls (7.7386,0.4658) .. (7.732,1) .. controls (7.2727,1.2727) .. (7,1.732);
\draw[dashed] (8,0) .. controls (7.7386,-0.4658).. (7.5588,-1);
\draw (4,0) .. controls (4.2614,-0.4658) .. (4.268,-1) .. controls (4.7273,-1.2727) .. (5,-1.732);
\draw[dashed] (4,0) .. controls (4.2614,0.4658).. (4.4412,1);
\draw[dashed] (7,1.732) .. controls (6.83857,1.7386) .. (6,2);
\draw[dashed] (5,-1.732) .. controls (5.16143,-1.7386) .. (6,-2);
\draw (7.77,1.4) node {$\alpha_{i+1}^{-1}$};
\draw (8.3,.5) node {$\beta_{j-1}^{-1}$};
\draw (4.3,-1.55) node {$\beta_{j-1}$};
\draw (3.7,-.5) node {$\alpha_{i}^{-1}$};
\draw (6,0) node {\LARGE{$\curvearrowright$}};
\draw (0,-3) node {(a)} (6,-3) node {(b)};
\end{tikzpicture}
\end{center}
\caption{Vertices that are identified to a single point.}
\label{fig11.8}
\end{figure}
From the construction of the surface $S$, it is straightforward to see that the edges $\alpha_1,\alpha_2,\hdots,$ $\alpha_{2g}$ and $\beta_1,\beta_2,\hdots,\beta_{2g}$ project to two simple closed curves on the surface $S$, denoted by $\alpha$ and $\beta$ respectively. Therefore, we have $(\alpha,\beta)$ is a minimally intersecting separating filling pair of $S$ with $\alpha$ separating and $\phi^{(\alpha, \beta)}=\phi$.
\end{proof}
Next, we study a necessary and sufficient condition for two minimally intersecting separating filling pairs are in the same mapping class group orbit.
\begin{lemma}\label{lemma11.3}
Suppose $\Gamma=(\alpha,\beta)$ and $\tilde{\Gamma}=(\tilde{\alpha},\tilde{\beta})$ are two minimally intersecting filling pairs on $S_g$ in the same $\mathrm{Mod}(S_g)$-orbit. Then $\phi^{(\alpha,\beta)}=\phi^{(\tilde{\alpha},\tilde{\beta})}$ modulo conjugation by permutations of the form
\begin{center}
$\mu_g^l\kappa_g^k\delta_g^j\eta_g^i,\hspace{.5 cm} l,i\in\{0,1\};\hspace{.4 cm} j,k\in\{0,1,\hdots,2g-1\},$
\end{center}
where,
\begin{align*}
\mu_g&=(2,4g+2)(4,4g+4)\hdots(4g,8g),\\
\kappa_g&=(1,3,\hdots,4g-1)(4g+1,4g+1,\hdots,8g-1),\\
\delta_g&=(2,4,\hdots,4g)(4g+2,4g+4,\hdots,8g) \text{ and}\\
\eta_g&=(1,4g+1)(3,4g+3)\hdots(4g-1,8g-1).
\end{align*}
\end{lemma}
\begin{proof}
We consider labelling on $(\alpha,\beta)$ and $(\tilde{\alpha},\tilde{\beta})$ so that $\alpha=(\alpha_1,\hdots,\alpha_{2g})$, $\beta=(\beta_1,\hdots,\beta_{2g})$, $\tilde{\alpha}=(\tilde{\alpha}_1,\hdots,\tilde{\alpha}_{2g})$ and $\tilde{\beta}=(\tilde{\beta}_1,\hdots,\tilde{\beta}_{2g})$ (see Section \ref{sec:5.1}). The given filling pairs are in the same mapping class group orbit implies that there exists $f\in \mathrm{Mod}(S_g)$ such that $f(\alpha,\beta)=(\tilde{\alpha},\tilde{\beta})$. Now, there are following four cases to consider.
\noindent \textbf{Case 1.} Consider $f(\alpha_n)=\tilde{\alpha}_{n+k_0}$, for some $0\leq k_0\leq2g-1$ and $f(\beta_m)=\tilde{\beta}_m$, for all $m,n=1,\dots,2g$. In this case, $\phi^{(\alpha,\beta)}$ and $\phi^{(\tilde{\alpha},\tilde{\beta})}$ are conjugate by $\kappa_g^{k_0}$.
\noindent \textbf{Case 2.} Consider $f(\beta_m)=\tilde{\beta}_{m+j_0}$, for some $0\leq j_0\leq2g-1$ and $f(\alpha_n)=\tilde{\alpha}_n$ for all $m,n=1,\dots,2g$. Then $\phi^{(\alpha,\beta)}$ and
$\phi^{(\tilde{\alpha},\tilde{\beta})}$ are conjugate by $\delta_g^{j_0}$.
\noindent \textbf{Case 3.} Let $f$ preserves the orientation of $\alpha$ but reverses that of $\beta$. Then $\phi^{(\alpha,\beta)}$ and
$\phi^{(\tilde{\alpha},\tilde{\beta})}$ are conjugate by $\mu_g$.
\noindent \textbf{Case 4.} Let $f$ preserves the orientation of $\beta$ but reverses that of $\alpha$. Then $\phi^{(\alpha,\beta)}$ and
$\phi^{(\tilde{\alpha},\tilde{\beta})}$ are conjugate by $\eta_g$.
The general case comprises by the above cases and hence we conclude that $\phi^{(\alpha,\beta)}$ and
$\phi^{(\tilde{\alpha},\tilde{\beta})}$ are conjugate by $\mu_g^l\kappa_g^k\delta_g^j\eta_g^i \text{ where } l,i\in\{0,1\}$ and $j,k\in\{0,1,\hdots,2g-1\}.$
\end{proof}
\end{subsection}
\begin{subsection}{A lower bound of \texorpdfstring{$N(g)$}{TEXT}.}
In this section, we prove the following inequality
$$N(g)>\frac{\prod\limits_{k=0}^{\frac{g-4}{2}-1}(8+3k)}{4\times (2g)^2\times (\frac{g-4}{2})!}.$$
Recall that in the proof of Theorem 1.2 (see Section~\ref{proof_th_2}), to construct a minimal filling pair on the surface $S_{g+2}$ from that of $S_g$, we have used a copy of a $S_{2,1}$.
We label the sub-arcs corresponding to $\alpha'$ and $\beta'$ by $x_1,\hdots,x_6$ and $y_1,\hdots,y_6$, respectively. We call this as $Z$-piece on which the points below are distinct:
\begin{enumerate}
\item initial point of $y_1$,
\item initial point of $x_1$,
\item end point of $y_6$ and
\item end point of $x_6$.
\end{enumerate}
For a Z-piece $Z_1$, we define $Z_1^{(\alpha)}$ to be the interior arcs of the separating arc of the Z-piece. Here $Z_1^{(\alpha)}=\{x_2,\hdots,x_5\}$. Similarly we define $Z_1^{(\beta)}=\{y_2,\hdots,y_5\}$. For two Z-pieces $Z_1 $ and $Z_2$, we define $(Z_1,Z_2)^\cap=(Z_1^{(\alpha)}\cap Z_2^{(\alpha)})\cup(Z_1^{(\beta)}\cap Z_2^{(\beta)})$.
By $x_k(Z_1)$, we mean the end point of the arc $x_k$ in the Z-piece $Z_1$.
\begin{lemma}\label{lemma_equality_of_Z_piece}
If $Z_1$ and $Z_2$ are two Z-pieces on $(\alpha,\beta)$ with $(Z_1^{(\alpha)}\cap Z_2^{(\alpha)})\neq\emptyset$, then $Z_1=Z_2$.
\end{lemma}
\begin{proof}
Assume $Z_1$ starts before $Z_2$.
\noindent \textbf{Case 1:} Consider $x_6(Z_1)=x_3(Z_2)$. Then $x_4*x_5*y_4^{-1}$ is a loop implies $x_1*x_2*y_5^{-1}$ must be a loop which is not.
\noindent\textbf{Case 2:} Consider $x_6(Z_1)=x_2(Z_2)$. Then $x_4*x_5*y_4^{-1}$ is a loop implies the concatenation of the $x$-arc before $x_1,x_1$ and an $y$-arc incident at the endpoint of $x_1$ must be a loop, which is impossible.
\noindent\textbf{Case 3:} $x_6(Z_1)=x_4(Z_2)$. Then $x_2*x_3*x_4*y_2$ is a loop which implies $x_4*x_5*x_6*\tilde{y}$ is a loop, where $\tilde{y}$ is an $y$-arc incident at the endpoint of $x_6$. But this is not possible.
\noindent\textbf{Case 4:} $x_6(Z_1)=x_5(Z_2)$. Then $x_2*x_3*y_3^{-1}$ is a loop which implies $x_1*x_2*y'$ is a loop, where $y'$ is an $y$-edge incident at the end point of $x_2$. But this is not true.
Therefore, $x_6(Z_1)=x_6(Z_2)$ and this implies $Z_1=Z_2$.
\end{proof}
Now, we are ready to find a lower bound for $N(g)$. For $g\geq 6$, we define
\begin{align*}
\mathscr{A}_g&=\left\{\left(s_1,\dots,s_\frac{g-4}{2}\right)| s_i\in \mathbb{N}, s_i\leq4(i+1), i=1,\dots,\frac{g-4}{2}\text{ and } s_1<\dots<s_\frac{g-4}{2} \right\}\text{ and}\\
\mathscr{O}_g&=\left\{(\alpha,\beta)| (\alpha,\beta)\text{ is an oriented minimal filling pair on } S_g \right\}.
\end{align*}
\begin{lemma}\label{lemma5.6}
$|\mathscr{A}_g|\leq|\mathscr{O}_g|$.
\end{lemma}
\begin{proof}
To prove the lemma, we construct an injective function from $\mathscr{A}_g$ to $\mathscr{O}_g$. Let $\left(s_1,\dots,s_\frac{g-4}{2}\right)\in\mathscr{A}_g$. Consider an oriented minimal filling pair $(\alpha(4),\beta(4))$ on $S_4$ and label the vertices along $\alpha(4)$ by $v_1,v_2\dots,v_8$. At the vertex $v_{s_1}$, we extract an open disk and attach a $Z$-piece to obtain an oriented minimal filling pair $(\alpha(6),\beta(6))$ on $S_{6}$ whose vertices are labelled as follows: the labelling of the vertices $v_i,i<s_1$ are kept same, the vertices on the $Z$-piece are labeled as $v_{s_1},\dots,v_{s_1+4}$ and the subsequent vertices are labeled as $v_{s_1+5},\dots,v_{12}$.
Next, we choose the vertex $v_{s_2}$ and perform similar operation as above to obtain an oriented minimal filling pair $(\alpha(8),\beta(8))$ on $S_8$. We repeat this process to get an oriented minimal filling pair $(\alpha(g),\beta(g))$ on $S_g$.
Now we define $f:\mathscr{A}_g\longrightarrow\mathscr{O}_g$ by
$$f(s_1,\dots,s_\frac{g-4}{2})=(\alpha(g),\beta(g)).$$
We prove that $f$ is injective. Let $f(s_1,\dots,s_\frac{g-4}{2})=f(s_1',\dots,s_\frac{g-4}{2}')=(\alpha(g),\beta(g)).$ First, we show that $s_\frac{g-4}{2}=s_\frac{g-4}{2}'$. If not, suppose, without loss of generality, that $s_\frac{g-4}{2}< s_\frac{g-4}{2}'$. By the construction of the surface $S_g$, there are two distinct $Z$-pieces at $s_\frac{g-4}{2}$$^{th}$ and $ s_\frac{g-4}{2}'$$^{th}$ vertices and by Lemma \ref{lemma_equality_of_Z_piece}, they are disjoint. Now, we remove the $Z$-piece at $s_\frac{g-4}{2}$$^{th}$ vertex and attach a disk to obtain an oriented minimal filling pair $(\alpha(g-2),\beta(g-2))$ on $S_{g-2}$. We do this for the vertices corresponding to $s_{\frac{g-4}{2}-1},\dots,s_1$ and finally we get a minimal filling pair $(\alpha(4),\beta(4))$ on $S_4$ and it contains a Z-piece, which is not possible. So $s_\frac{g-4}{2}=s_\frac{g-4}{2}'$. A similar argument shows that $s_i=s_i^{'} \text{ for all, } i=1,\dots, \frac{g-4}{2}-1$. Hence, $f$ is injective.
\end{proof}
\begin{corollary}
$$N(g)\geq\frac{\prod\limits_{k=1}^{\frac{g-4}{2}}(3k+5)}{4\times (2g)^2\times (\frac{g-4}{2})!}.$$
\end{corollary}
\begin{proof}
Consider pairwise distinct natural numbers $t_1,\dots,t_\frac{g-4}{2}$ satisfying $t_k\leq4(k+1),\text{ for } k=1,\dots,\frac{g-4}{2}$. Then there exists a permutation $\sigma$ such that $\left(t_{\sigma(1)},\dots,t_{\sigma(\frac{g-4}{2})}\right)\in\mathscr{A}_g$. Furthermore, by consideration we have $3k+5$ many choices for each $t_k$ which implies
$$|\mathcal{A}_g|\geq\frac{\prod\limits_{k=1}^{\frac{g-4}{2}}(3k+5)}{(\frac{g-4}{2})!}.$$
Therefore, the lower bound is obtained by dividing by maximum number of twisting parameters,
$$N(g)\geq\frac{|\mathcal{O}_g|}{4 \times (2g)^2}\geq\frac{|\mathcal{A}_g|}{4 \times (2g)^2}\geq\frac{\prod\limits_{k=1}^{\frac{g-4}{2}}(3k+5)}{4\times (2g)^2\times (\frac{g-4}{2})!}.$$
\end{proof}
\end{subsection}
\end{section}
\begin{section}{Upper bound of \texorpdfstring{$N(g)$}{Lg}}
In this section, we prove the inequality $N(g)<2(2g-2)(2g-2)!$ in the following proposition which concludes the proof of Theorem~\ref{thm:1.3}.
\begin{proposition}\label{prop:6.1}
For $g\geq4$, we have $N(g)<2(2g-2)(2g-2)!$.
\end{proposition}
Before going into the proof of Proposition \ref{prop:6.1}, we prove a technical result in Lemma \ref{lemma:6.1} which is essential for the proof of the proposition.
\begin{lemma}\label{lemma:6.1}
Let $\phi\in\Sigma_{8g}$ be a filling permutation. Then $\phi$ is of the form $Q^{4g}C$, where C is a square root of $Q^{4g}\tau$. Conversely, if $C$ is a square root of $Q^{4g}\tau$ with $\phi=Q^{4g}C$ satisfying conditions $(1)-(5)$ of Proposition~\ref{prop_of_phi_alpha_beta}, then $\phi$ is a filling permutation.
\end{lemma}
\begin{proof}
Let $\phi$ be a filling permutation. Then,
$$\left(Q^{4g}\phi\right)^2=Q^{4g}\left(\phi Q^{4g}\phi\right)= Q^{4g}\tau.$$
Conversely,
$$\phi Q^{4g}\phi=Q^{4g}C Q^{4g}Q^{4g}C=Q^{4g}C^2=\tau.$$
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:6.1}]
In lite of Lemma \ref{lemma:6.1}, to prove the proposition, we find an upper bound of the number of square roots of the permutation $Q^{4g}\tau$. We have $$Q^{4g}\tau=(1,4g+3)(3,4g+5)\cdots(4g-1,4g+1)(2,4g+4)\dots(4g-2,8g).$$
A square root $C$ of $Q^{4g}\tau$ is obtained by taking the transpositions of $Q^{4g}\tau$ in pairs and interleaving them. For instance, for the pair of transpositions $(1,4g+3)$ and $(2,4g+4)$, the possible arrangements are $(1,2,4g+3,4g+4)$ and $(1,4g+4,2,4g+3)$. In each pair of transpositions, one comes from $\{(1,4g+3),(3,4g+5),\dots,(4g-1,4g+1)\}$ and the other from $\{(2,4g+4),(4,4g+6),\dots,(4g-2,8g)\}$, as $\phi=Q^{4g}C$ is parity respecting and sends even numbers to odds.
Suppose, for a suitable choice of $C$, the permutation $\phi$ satisfies conditions $(1)-(5)$ of Proposition \ref{prop_of_phi_alpha_beta} and $\phi=\phi_1\phi_2$, where $\phi_1$ and $\phi_2$ are disjoint $4g$-cycles. Now, we have the following cases to consider.
\noindent\textbf{Case 1:} The odd entries of $\phi_1$ come from $\{1,3,\hdots,4g-1\}$ and the even entries of $\phi_1$ are of the form $4k$ (following properties $(3)$ and $(5)$ of proposition \ref{prop_of_phi_alpha_beta}).
In this case, we have a unique choice for the $4$-cycle. More precisely, for the transpositions $(i,i+4g+2)$ and $(j,j+4g+2)$, the only possible $4$-cycle is $(i,j,i+4g+2,j+4g+2)$, where $i\leq4g$ is odd and $j\leq4g$ of the form $4k$. Such pairs can be chosen in $(2g)!$ ways implies the number of such square roots $C$ is bounded above by $(2g)!$.
\noindent\textbf{Case 2:} The odd entries of $\phi_1$ comes from $\{1,3,\dots,4g-1\}$ and the even entries of $\phi_1$ are of the form $4k+2$. A similar argument shows that in this case also there are at most $(2g)!$ choices.
Therefore, the square root $C$ and hence $\phi$ has at most $2(2g)!$ choices. Again using the conditions of Proposition \ref{prop_of_phi_alpha_beta}, we further rule out some more choices of $\phi$. Here, we find those $\phi$'s which satisfy $\phi^2(1)=1$ and subtract this from the whole. First, take the transpositions $(1,4g+3)$ and $(i,i+4g+2)$, where $i\leq4g$ is an even integer. Then $(1,i,4g+3,i+4g+2)$ is a choice for a cycle of $C$. Next, pair the transpositions $(i-2,4g+i)$ and $(4g-1,4g+1)$ to get another $4$-cycle $(4g-1,4g+i,4g+1,i-2)$ of $C$ and choose the other $4$-cycles arbitrarily.
As $\phi^2(1)=1,$ the number of filling permutations $\phi$ is bounded above by $2(2g)!-2g\times2(2g-2)!=4g(2g-2)(2g-2)!$.
Two minimally intersecting filling pairs are in the same $\mathrm{Mod}(S_g)$ orbit if and only if the corresponding filling permutations are conjugate by twisting permutations $\mu_g^l\kappa_g^k\delta_g^j\eta_g^i$, where $l,i\in\{0,1\},j,k\in\{0,1,\dots,2g-1\}$
and $\mu_g,\kappa_g,\delta_g,\eta_g$ are defined as in Lemma \ref{lemma11.3}. As there are at least $2g$ many twisting permutations, the number of $\mathrm{Mod}(S_g)$ orbits of minimally intersecting filling pair is bounded by $\frac{4g(2g-2)(2g-2)!}{2g}=2(2g-2)(2g-2)!.$
\end{proof}
\end{section}
\begin{section}{The length function}
In this section, we prove Theorem \ref{theorem_lenth_function}. First we prove a technical lemma which deals with injectivity radius (see Section $4.1$ in \cite{buser2010geometry}) and length of fillings. The injectivity radius of a hyperbolic surface $X$ is denoted by $\mathrm{inj}(X)$. We note that the lemma is essential in proving the length function $\mathcal{F}_g$ is proper.
\begin{lemma}\label{lemma_injectivity_radius}
Given $M\in \mathbb{R}_{>0}$, there exists an $\epsilon>0$ such that $\mathcal{F}_g(X)\geq M$ if $\mathrm{inj}(X)\leq \epsilon$.
\end{lemma}
\begin{proof}
Consider the width function $w:\mathbb{R}^{+}\longrightarrow\mathbb{R}$, defined by
$$w(x)=\mathrm{arcsinh}(1/\sinh(x/2)),$$
as in Collar Lemma, satisfying following: for $M\in \mathbb{R}_{>0}$, there exists an $\epsilon>0$ such that $w(x)\geq M$, for all $x<\epsilon$ (for details, see Theorem $4.1.1$ in \cite{buser2010geometry}).
Let $X\in \mathcal{M}_g$ with $\mathrm{inj}(X)\leq\frac{\epsilon}{2}$. Then there exists an essential simple closed geodesic $\gamma$ with $l_X(\gamma)\leq\epsilon$, as $\mathrm{inj}(X)=\frac{1}{2}\mathrm{sys}(X)$ (see Lemma 4.1.2 in \cite{buser2010geometry}). By Collar's lemma, $\gamma$ has a tubular neighbourhood of width at least $M$. Now it follows that $\mathcal{F}_g(X)\geq M$.
\end{proof}
\begin{lemma}\label{lemma7.2}
$\mathcal{F}_g$ is proper.
\end{lemma}
\begin{proof}
Let $C$ be a compact set in $\mathbb{R}$. Then there exists $M\in \mathbb{R}_{>0}$ such that $C\subset[-M,M]$. By Lemma~\ref{lemma_injectivity_radius}, there exists an $\epsilon>0$ such that $\mathcal{F}_g(X)>M$, whenever $\mathrm{inj}(X)<\epsilon$. This shows that $\mathcal{F}_g^{-1}(C)\subset \mathcal{M}_g^{\epsilon}$, where $\mathcal{M}_g^{\epsilon}$ is the $\epsilon$-thick part of $\mathcal{M}_g$ (see Section $12.4$ \cite{farb2011primer}). As $\mathcal{F}_g$ is continuous and $\mathcal{M}_g^{\epsilon}$ is compact, it follows that $\mathcal{F}_g$ is a proper function.
\end{proof}
\begin{lemma}\label{lemma7.3}
The function $\mathcal{F}_g$ is a topological Morse function.
\end{lemma}
\begin{proof}
The proof follows from a similar argument as in the proof of Theorem $1.3$ in \cite{aougab2015minimally}.
\end{proof}
\begin{lemma}\label{lemma7.4}
$\mathcal{F}_g\geq m_g$, where $m_g$ is the perimeter of the regular right angled hyperbolic $4g$-gon.
\end{lemma}
\begin{proof}
Let the minimum value of $\mathcal{F}_g$ occurs at $X\in \mathcal{M}_g$ and $(\alpha,\beta)$ be a shortest length filling pair on $X$, where both of $\alpha$ and $\beta$ are geodesics. We cut the surface along $\alpha\cup\beta$ and we obtain two hyperbolic $4g$-gons $P_1$ and $P_2$, each of area $\pi(2g-2)$. It follows from the fact that the hyperbolic $n$-gon with least perimeter is regular (see Bezdek \cite{bezdek1984elementarer}), the polygons $P_1$ and $P_2$ are regular. The Gauss-Bonnet theorem implies that $P_1$ and $P_2$ are regular right angled $4g$-gons. Hence, the proof follows.
\end{proof}
Next, we find the growth of the set $\mathcal{B}_g$ in the following lemma.
\begin{lemma} \label{lemma7.5}
For $g\geq4,$ $ \mathcal{B}_g$ is a finite set and $|\mathcal{B}_g|=N(g)$.
\end{lemma}
\begin{proof}
By Lemma~\ref{lemma7.2}, $\mathcal{F}_g$ is proper and hence $\mathcal{B}_g$ is compact. All the points of $\mathcal{B}_g$ are isolated as they are critical points of the Morse function $\mathcal{F}_g$. Therefore, $\mathcal{B}_g$ is a finite set.
Now we find an one to one correspondence between the set $\mathcal{B}_g$ and the collection of $\mathrm{Mod}(S_g)$-orbits of minimal filling pairs. For each of $\mathrm{Mod}(S_g)$-orbits, we associate an element $X$ of $\mathcal{B}_g$ by simply identifying the edges of two disjoint regular right angled $4g$-gons in accordance with the chosen orbit. It is straightforward to see that this association is surjective. The injectivity follows from Proposition~\ref{lemma:7.4}. This completes the proof.
\end{proof}
\begin{proposition} \label{lemma:7.4}
Let $(\alpha,\beta)$ and $(\tilde\alpha,\tilde\beta)$ be two minimal filling pairs on a hyperbolic surface $X\in \mathcal{B}_g$ each of length $m_g$. Then there exists a homeomorphism $\phi:X\longrightarrow X$ such that $\phi(\alpha,\beta)=(\tilde\alpha,\tilde\beta)$.
\end{proposition}
Before going to the proof, we recall the notion of equivariant tiling, chamber system and Delaney-Dress symbols and their equivalence which are essential for the proof of Proposition~\ref{lemma:7.4}. An \emph{equivariant tiling} is a pair $(\mathcal{T},\Gamma)$, where $\mathcal{T}$ is a tiling of $\mathbb{H}$ and $\Gamma$ is a discrete subgroup of $\mathrm{PSL}(2, \mathbb{R})$ satisfying $\gamma\mathcal{T}=\{\gamma A|A\in \mathcal{T}\}=\mathcal{T}$ for all $\gamma\in \Gamma$. Two equivariant tilings $(\mathcal{T},\Gamma)$ and $(\mathcal{T}',\Gamma')$ are said to be equivariantly equivalent if there exists a homeomorphism $\phi:\mathbb{H}\to \mathbb{H}$ such that $\phi \mathcal{T}=\mathcal{T}'$ and $\Gamma'= \phi \Gamma \phi^{-1}$.
To every vertex, edge and tile of $\mathcal{T}$ choose an interior point, called a $0$-, $1$- and $2$-center, respectively. For every tile $A\in\mathcal{T}$, join its $2$-center with $0$- and $1$-centers by geodesics (see Figure~\ref{Fig:Chamber}) which decompose the tiling $A$ into triangles. We call each triangle a chamber and the set of all triangles, denoted by $\mathcal{C}_\mathcal{T}$, is called \emph{chamber system}. Consider $\mathscr{D}=\mathcal{C}_\mathcal{T}/\Gamma$. The edge opposite the $i$-center of a chamber is called an $i$-edge. For $T\in \mathcal{C}_\mathcal{T}$, the neighbour triangle that share the $i$-edge of $T$, is denoted by $\sigma_i(T)$. Thus we have functions $\sigma_i:\mathcal{C}_\mathcal{T}\to \mathcal{C}_\mathcal{T}$ satisfying $\sigma_i(\gamma T)=\gamma \sigma_i(T)$, for $i=0,1,2$, $T\in \mathcal{C}_\mathcal{T}$ and $\gamma \in \Gamma$. This induces functions $\sigma_i^*:\mathscr{D}\longrightarrow\mathscr{D}$.
For $0\leq i<j \leq2$, we define
$m_{ij}:\mathscr{D}\longrightarrow\mathbb{N}$ with
$$m_{ij}(D)=\min \left\{m\in \mathbb{N}|C(\sigma_i\sigma_j)^m=C \text{ for all } C\in D\right\}.$$
Then the system $(\mathscr{D},m):=((\mathscr{D},\mathscr{E});m_{0,1},m_{0,2},m_{1,2})$ is called a Delaney-Dress symbol if the followings hold:
\begin{description}
\item[DS1] $m_{ij}(D)=m_{ij}(D\sigma_i)=m_{ij}(D\sigma_j)$
\item[DS2] $D(\sigma_i\sigma_j)^{m_{ij}(D)}=D(\sigma_j\sigma_i)^{m_{ij}(D)}=D$
\item[DS3] $m_{0,2}(D)=2$
\item[DS4] $m_{0,1}\geq2$
\item[DS5] $m_{1,2}(D)\geq3.$
\end{description}
Two Delaney-Dress symbols $(\mathscr{D},m)$ and $(\mathscr{D'},m')$ are called isomorphic if
and only if there exists a bijection $\pi:\mathscr{D}
\longrightarrow\mathscr{D}'$ with $(\pi D)\sigma_k=\pi(D\sigma_k)$ and $m_{ij}'(\pi D)=m_{ij}(D)$ for all $D\in \mathscr{D}, 0\leq k\leq2$ and $0\leq i<j \leq 2$. For more details, we refer the reader to Section 1 of \cite{huson1993generation}.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\draw ({2*cos(45)},{2*sin(45)}) to[bend left=20] ({2*cos(90)},{2*sin(90)}) to[bend left=20] ({2*cos(135)},{2*sin(135)}) to[bend left=20] ({2*cos(180)},{2*sin(180)}) to[bend left=20] ({2*cos(225)},{2*sin(225)})to[bend left=20] ({2*cos(270)},{2*sin(270)}) to[bend left=20] ({2*cos(315)},{2*sin(315)}) to[bend left=20] ({2*cos(360)},{2*sin(360)}) to[bend left=20] ({2*cos(45)},{2*sin(45)});
\draw ({5+2*cos(45)},{2*sin(45)}) to[bend left=20] ({5+2*cos(90)},{2*sin(90)}) to[bend left=20] ({5+2*cos(135)},{2*sin(135)}) to[bend left=20] ({5+2*cos(180)},{2*sin(180)}) to[bend left=20] ({5+2*cos(225)},{2*sin(225)})to[bend left=20] ({5+2*cos(270)},{2*sin(270)}) to[bend left=20] ({5+2*cos(315)},{2*sin(315)}) to[bend left=20] ({5+2*cos(360)},{2*sin(360)}) to[bend left=20] ({5+2*cos(45)},{2*sin(45)});
\draw [thin,cyan] (5,0)--({5+2*cos(45)},{2*sin(45)});
\draw [thin,cyan] (5,0)--({5+2*cos(90)},{2*sin(90)});
\draw [thin,cyan] (5,0)--({5+2*cos(135)},{2*sin(135)});
\draw [thin,cyan] (5,0)--({5+2*cos(180)},{2*sin(180)});
\draw [thin,cyan] (5,0)--({5+2*cos(225)},{2*sin(225)});
\draw [thin,cyan] (5,0)--({5+2*cos(270)},{2*sin(270)});
\draw [thin,cyan] (5,0)--({5+2*cos(315)},{2*sin(315)});
\draw [thin,cyan] (5,0)--({5+2*cos(360)},{2*sin(360)});
\draw [thin,blue] (5,0)--({5+1.7*cos(22.5)},{1.7*sin(22.5)});
\draw [thin,blue] (5,0)--({5+1.7*cos(90-22.5)},{1.7*sin(90-22.5)});
\draw [thin,blue] (5,0)--({5+1.7*cos(135-22.5)},{1.7*sin(135-22.5)});
\draw [thin,blue] (5,0)--({5+1.7*cos(180-22.5)},{1.7*sin(180-22.5)});
\draw [thin,blue] (5,0)--({5+1.7*cos(225-22.5)},{1.7*sin(225-22.5)});
\draw [thin,blue] (5,0)--({5+1.7*cos(270-22.5)},{1.7*sin(270-22.5)});
\draw [thin,blue] (5,0)--({5+1.7*cos(315-22.5)},{1.7*sin(315-22.5)});
\draw [thin,blue] (5,0)--({5+1.7*cos(360-22.5)},{1.7*sin(360-22.5)});
\draw[thick, green] ({5+1.7*cos(22.5)},{1.7*sin(22.5)}) node{\tiny$\bullet$} ({5+1.7*cos(90-22.5)},{1.7*sin(90-22.5)}) node{\tiny$\bullet$}
({5+1.7*cos(135-22.5)},{1.7*sin(135-22.5)}) node{\tiny$\bullet$}
({5+1.7*cos(180-22.5)},{1.7*sin(180-22.5)}) node{\tiny$\bullet$}
({5+1.7*cos(225-22.5)},{1.7*sin(225-22.5)}) node{\tiny$\bullet$}
({5+1.7*cos(270-22.5)},{1.7*sin(270-22.5)}) node{\tiny$\bullet$}
({5+1.7*cos(315-22.5)},{1.7*sin(315-22.5)}) node{\tiny$\bullet$}
({5+1.7*cos(360-22.5)},{1.7*sin(360-22.5)}) node{\tiny$\bullet$};
\draw[red] ({5+2*cos(45)},{2*sin(45)}) node{\tiny$\bullet$} ({5+2*cos(90)},{2*sin(90)}) node{\tiny$\bullet$}
({5+2*cos(135)},{2*sin(135)}) node{\tiny$\bullet$}
({5+2*cos(180)},{2*sin(180)}) node{\tiny$\bullet$}
({5+2*cos(225)},{2*sin(225)}) node{\tiny$\bullet$}
({5+2*cos(270)},{2*sin(270)}) node{\tiny$\bullet$}
({5+2*cos(315)},{2*sin(315)}) node{\tiny$\bullet$}
({5+2*cos(360)},{2*sin(360)}) node{\tiny$\bullet$};
\draw (5,0) node{$\bullet$};
\draw (2.5,-3) node{\textcolor{red}{$\bullet$} : $0$-center} (2.5,-3.5) node{\textcolor{green}{$\bullet$} : $1$-center} (2.5,-4) node{$\bullet$ : $2$-center};
\end{tikzpicture}
\end{center}
\caption{A tile and its chambers}
\label{Fig:Chamber}
\end{figure}
Next we will state an useful theorem (for more details, see Lemma 1.1 of \cite{huson1993generation}).
\begin{lemma}\label{lemma0.2}
Two equivariant tilings $(\mathcal{T},\Gamma)$ and $(\mathcal{T}',\Gamma')$ are (equivariantly)
equivalent if and only if the corresponding Delaney-Dress symbols $(\mathscr{D},m)$ and $(\mathscr{D'},m')$ are isomorphic.
\end{lemma}
\begin{proof}[Proof of Proposition \ref{lemma:7.4}]
Consider the filling pairs $(\alpha,\beta)$ and $(\tilde\alpha,\tilde\beta)$ as in the proposition. Let $\Gamma$ be the discrete subgroup of $\mathrm{PSL}_2(\mathbb{R})$ such that $X\cong \mathbb{H}/\Gamma$. The lifts $\mathscr{T}$ and $\mathscr{T}'$ of $(\alpha,\beta)$ and $(\tilde\alpha,\tilde\beta)$, respectively are two tilings of $\mathbb{H}$ by regular right-angled $4g$-gons. Thus we have two equivariant tilings $(\mathcal{T},\Gamma)$ and $(\mathcal{T}',\Gamma)$.
Consider the chamber systems $\mathcal{C}_\mathcal{T}$ and $\mathcal{C}_{\mathcal{T}'}$ . Let $(\mathscr{D},m)$ and $(\mathscr{D'},m')$ be the corresponding Delaney-Dress symbols. Then,
\begin{align*}
& m_{0,2}(D)=2= m_{0,2}'(D'),\\
&m_{1,2}(D)=4=m_{1,2}'(D') \text{ and}\\
&m_{0,1}(D)=4g=m_{0,1}'(D'), \text{ for all } D\in \mathscr{D}, D'\in \mathscr{D'}.
\end{align*}
As $|\mathscr{D}|=|\mathscr{D'}|=16g$ and value of $m$ and $m'$ are identical, the Delaney-Dress symbols $(\mathscr{D},m)$ and $(\mathscr{D'},m')$ are isomorphic and hence by Lemma 1.1 of \cite{huson1993generation}, $(\mathcal{T},\Gamma)$ and $(\mathcal{T}',\Gamma)$ are equivariantly equivalent. Therefore, there exists a homeomorphism $\phi:\mathbb{H}\longrightarrow\mathbb{H}$ such that $\phi\mathscr{T}=\mathscr{T'}$. This homeomorphism projects to a homeomorphism $\tilde\phi$ of the surface such that $\tilde\phi(\alpha,\beta)=(\tilde\alpha,\tilde\beta)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{theorem_lenth_function}]
By Lemma \ref{lemma7.2}, $\mathcal{F}_g$ is a proper function and that $\mathcal{F}_g$ is a Morse function follows from Lemma \ref{lemma7.3}. The inequality $\mathcal{F}_g\geq m_g$ follows from Lemma \ref{lemma7.4}. Finally, the fact that $g\geq4,$ $ \mathcal{B}_g$ is a finite set and $|\mathcal{B}_g|=N(g)$ follows from Lemma~\ref{lemma7.5}.
\end{proof}
\end{section}
|
\section{Introduction}
The task of formation control is central to many problems in multiagent coordination and cooperative control, as it is usually the first problem one typically has to solve to achieve some collective objective with multiple agents. In the standard formation control problem, it is usually of interest to control a group of agents --- so that they converge to unique terminal states and with the goal of attaining a desired geometric pattern --- to facilitate a specific task. Such a control objective finds direct application in several areas such as reconnaissance, aerial coverage and monitoring, mobile target tracking, and in mobile communication network maintenance, to name a handful. In problems involving formation control, the prevailing assumptions are usually that all the agents have either the same forward velocity \cite{maCooperativeTargetTracking2015a, marshallFormationsVehiclesCyclic2004} or angular velocity \cite{brinonarranzContractionControlFleet2010}, and that information about the state of each agent is available to its neighbors. The agents obtain this state information from either a central station broadcasting to all agents or from a more complex distributed network topology that can be fixed, stochastic, or even have intrinsic dynamics \cite{caoOverviewRecentProgress2013}.
Conventionally, the formation control problem takes one of two broad forms: group reference formation control and non-group reference formation control \cite{caoOverviewRecentProgress2013}. Group-reference formation control, also known as formation tracking, is the case where the agents move in formation while tracking a reference trajectory or \textit{group reference}. In non-group reference formation control on the other hand, the agents are tasked with maintaining a specific geometric shape without following any trajectory setpoint.
Unsurprisingly, much of the research on formation control is centered around the more challenging problem of formation tracking, and several methods have been proposed (see \cite{ohSurveyMultiagentFormation2015} for a detailed survey on the topic). There is the well-researched leader-follower paradigm where one agent is taken as the leader and the other agents, as followers, that must track the leader's motion while maintaining some pre-specified distance from themselves and the leader. Defining rules that govern the evolution of these inter-agent distances thus leads to the desired formation, and by varying the rules, a new formation results. Simultaneous tracking under this formation is then achieved by specifying the desired trajectory as the leader's path setpoint.
To effectively track the leader, follower agents require sufficiently accurate estimates of the leader's pose in the inertial frame, which can be affected by noisy sensor measurements, exogenous disturbances from the environment, such as wind or downwash from nearby agents --- in the case where the agents are aerial vehicles --- or even uncertainty in the communication network from delays and packet drops. Thus, it is often the case that the multiagent system (MAS) will fail to track the reference trajectory while keeping formation, or deviate from the desired formation altogether, causing unintended and even unsafe effects \cite{liuFurtherResultsDistance2021}. Furthermore, the disturbances themselves may be difficult, computationally expensive, or impossible to estimate, making formation tracking under uncertainty both a safety-critical requirement and a nontrivial problem.
Several studies have approached the formation tracking problem from an optimal control viewpoint. One of the earliest efforts at formulating the tracking with formation maintenance task as an optimal control problem was presented in \cite{wangIntegratedOptimalFormation2013}. Here, using an approach derived from the Riccati equation, the authors designed a distributed optimal formation control law --- for multiple UAVs with linear models --- by minimizing a non-quadratic cost function. The optimal control formulation was given here, however, without any consideration to the pairwise distances between agents. Following standard thinking based on Pontryagin's Minimum Principle (PMP), the authors in \cite{liuFinitetimeFormationControl2015} presented an optimal formation control approach by minimizing the control energy of the system, with the agents evolving under perfect-state dynamic models. More recently, an identifier-critic-actor reinforcement learning based method was employed in \cite{wenOptimizedFormationControl2020} to select the optimal control policy for an MAS comprising agents with unknown and adaptively-identified nonlinear dynamics.
In our work, we study the problem of formation tracking under localization errors where the leader in the MAS is required to track a sinusoidal reference. Simultaneously, the followers are required to keep their assigned planar positions with respect to the leader and themselves as defined by a triangular formation rule. In contrast to the aforementioned research articles, our work focuses on designing optimal formation tracking laws for a specific case where the agents are modeled as quadrotors in the plane under uncertainty (from sensor noise). We also formulate the formation tracking task as a dynamic optimization problem with a constrained-augmented Lagrangian and solve it using optimization software tools, as opposed to traditional analytical optimal control methods like PMP or the Riccati equation.
\subsection{Contributions}
Our contributions are as follows:
\begin{enumerate}[i.]
\item Application of optimal control theory to a uniquely-formulated multiagent formation tracking problem.
\item Simulative validation of the effectiveness of the optimal control law in both the nominal setting and the case with Gaussian noise in state measurements.
\end{enumerate}
In what follows, we introduce the notation used in this work and discuss the setting under which we study the formation tracking problem (see Section \ref{sec:probform}). Next, in Section \ref{sec:2dquad}, we provide details about the planar quadrotor model under consideration. Section \ref{sec:optconc} puts forward the optimal control component of our work. Following that, in Section \ref{ssec:formgen}, we discuss motivations for electing a triangular formation as the reference formation in our work, along with a brief description of the properties of this desired formation. The simulation setup is provided in Section \ref{sec:simstudies}, with key simulation results following in Section \ref{sec:res}. Finally, we conclude the paper with recommendations for future research in Section \ref{sec:conc}.
\section{Preliminaries}
\label{sec:probform}
We pose the formation tracking problem, as considered in this article, under the following assumptions:
\begin{enumerate}[i.]
\item All agents are homogeneous, i.e., they are identical, hence (\ref{it:limod}) follows.
\item \label{it:leadfoll} All but one (randomly-chosen) agent (the leader) belong to the follower group; information about the leader's state is available to all the follower agents through a common communication network shared by all agents.
\item \label{it:limod} The model of each agent can be approximated by a linear time-invariant continuous-time model. See Section \ref{sec:2dquad} for a description of the model.
\item Each agent's roll angle -- and thus, rate -- is approximately zero, i.e., the agent moves to its position in the formation by maintaining a near-hover state.
\item Each follower agent is driven independently to execute the local task of keeping its pre-assigned position in the inertial frame and also to simultaneously achieve the collective task of maintaining a desired group formation with the other agents. This assumption implies that there are no adversarial agents within the group.
\item The uncertainty in the MAS is only due to localization errors from the state estimation module (see Figure \ref{fig:optcontarch}), hence the agents' states are perturbed by sensor noise, and are thus taken to be imperfect. Effects from external disturbances such as wind gust and downwash are neglected.
\end{enumerate}
We denote the $i^{\text{th}}$ agent as $a_i \in \mathcal{A}$, where $\mathcal{A}$ is the set of all agents. $a_\mc{L} \in \mc{A}$ denotes the leader agent, while the follower agents are in the set $\mathcal{A} \setminus \{a_\mathcal{L}\}$. Additionally, while it is possible to segment $\mathcal{A}$ into a finite number of leader-follower subsets (e.g., in the multi-leader case \cite{sorensenUnifiedFormationControl2007}), we have assumed that there is only one leader (see assumption (\ref{it:leadfoll})) and that all other agents are followers within any optimization horizon. A few other assumptions will be introduced in later sections as we specify the notation required for their definition. However, with the above setting, we can now present the formation tracking problem as follows: Given $N$ agents in total, $N-1$ followers must keep their positions in the formation while the randomly-selected leader tracks a particular trajectory in space, with possibly inaccurate state information from sensor measurements. Essentially, we require that the group formation be preserved, with the least possible formation error, while the leader agent tracks a specified trajectory.
\section{Planar Quadrotor Model}
\label{sec:2dquad}
We model the agents as quadrotors in the plane (see Figure \ref{fig:quadmod}), governed by the dynamics of a planar quadrotor linearized at the equilibrium (hover) state:
\begin{subequations}
\label{eq:masmodelss}
\begin{align}
&\ddot{y}_i = -g\phi_i\\
&\ddot{z}_i = -g + \frac{u_{1_i}}{m}\\
&\ddot{\phi}_i = \frac{u_{2_i}}{I_{xx}},
\end{align}
\end{subequations}
where $m$ is the mass of each agent, $g$ is the gravitation constant, and $I_{xx}$ is the $x$ component of the (diagonal) inertia matrix. To simulate sensor noise, we introduce White Gaussian Noise (WGN) terms to the formulation in (\ref{eq:masmodelss}) to get:
\begin{subequations}
\label{eq:masmodelssawgn}
\begin{align}
&\ddot{y}_i = -g\phi_i + {w}_1\\
&\ddot{z}_i = -g + \frac{u_{1_i}}{m} + {w}_2\\
&\ddot{\phi}_i = \frac{u_{2_i}}{I_{xx}} + {w}_3,
\end{align}
\end{subequations}
We can now write the planar quadrotor model in state-space form as:
\begin{equation}
\label{eq:statewithw}
\dot{{\mbf x}}_i = f_i({{\mbf x}_i}, {\mbf u}_i, \mbf{w}) = {\mbf A}{\mbf x}_i + {\mbf B}{\mbf u}_i + {\mbf G}_{c}g + {\mbf K}{\mbf w}.
\end{equation}
where ${\mbf A}$ and ${\mbf B}$ are the plant and input matrices of the state-space model with appropriate dimensions, respectively, with $\text{det}({\mbf A}) \neq 0$. ${\mbf G}_{c}$ is the vector $\begin{bsmallmatrix}0 & 0 & 0 & 0 & -1 & 0 & 0\end{bsmallmatrix}^T$, which accounts for gravity compensation in the $z_\mc{O}$-direction. ${\mbf K} \in \mb{R}^{6\times 6}$ is the noise gain matrix while ${\mbf w} \in \mb{R}^{6}$ is the vector $\begin{bsmallmatrix}0 & w_1 & 0 & w_2 & 0 & w_3\end{bsmallmatrix}^T$, where each $w_j\ (j = [1,2,3]$) follows a Gaussian distribution with mean $\mu \in \mb{R}$ and standard deviation, $\sigma \in \mb{R}$. Table \ref{tab:simquadparam}, partly adapted from \cite{forsterSystemIdentificationCrazyflie2015a}, lists the discussed parameters for the simulated planar quadrotor.
\begin{table}[htb]
\caption{Planar Quadrotor Model Parameters}
\label{tab:simquadparam}
\centering
\begin{tabular}{cc}
\hline
Parameter & Value\\
\hline
\hline
$m$ [kg] & 0.028\\
$I_{xx}$ [kg$m^2$] & $6.4893\cdot10^{-6}$\\
$[\mu \ \sigma ]$ & $[0.0\ 0.2]$ \\
\hline
\end{tabular}
\end{table}
\section{Optimal Quadrotor Formation Control}
\label{sec:optconc}
To achieve the formation tracking task as set forth in Section \ref{sec:probform}, we solve the following initial-value, finite-horizon optimal control problem (FHOCP) for the $i^\text{th}$ agent's control input, ${\mbf u}_i; \ i = [1, 2, \dots, N]$, on the interval $\tau = [0, T]$:
\begin{subequations}
\label{eq:optprob}
\begin{align}
\min_{{\mbf u}_i} \quad & J_i
\nonumber\\%\int_{\tau = t_0}^{T}L_i({{\mbf{x}}_i}[\tau],{\mbf u}_i[\tau])d\tau
\textrm{subject to:} \quad & \dot{{\mbf x}}_i(\tau) = f_i({{\mbf x}_i}(\tau),{\mbf u}_i(\tau),{\mbf w})\\
&{{{\mbf{x}}}_i}(0) = {\mbf x}_i^0,\\ \
\label{eq:iadist}
\quad & \big\lvert\big\lvert {\mbf \Gamma}_i(\tau) -{\mbf \Gamma}_j(\tau)\big\rvert\big\rvert_2 = d_{ij}^r;\ i \neq j \\
\label{eq:ubound}
\quad & \lvert {u}_{1_i}\rvert \leq u_{1_{\text{max}}}; \quad \lvert {u}_{2_i}\rvert \leq u_{2_{\text{max}}}.
\end{align}
\end{subequations}
Here, $J_i$ is the objective for the $i^{\text{th}}$ agent equal to the total expectation of the trajectory error, control, and Mayer costs defined as:
\begin{equation}
\label{eq:objexp}
\mb{E}\Bigg[\int_{\tau = 0}^{T}L_i({{\mbf{\Gamma}}_i}[\tau],{\mbf u}_i[\tau])d\tau + h(\mbf{x}_i(T))\Bigg],
\end{equation}
where $L_i({{\mbf{x}}_i}[\tau],{\mbf u}_i[\tau]:\mb{R}^{2}\times\mb{R}^2 \mapsto \mb{R}$ is the Lagrangian defined as follows (the $\tau$ argument has been omitted for brevity):
\begin{equation}
\label{eq:lagrgn}
{\mbf u}_i^T{\mbf R}_i{\mbf u}_i + ({\mbf \Gamma}_i-{{\mbf \Gamma}_i}^r)^T{\mbf Q}_i({\mbf \Gamma}_i-{{\mbf \Gamma}_i}^r),
\end{equation}
and $h:\mb{R}^{6} \mapsto \mb{R}$ is the terminal (Mayer) cost for the $i^{\text{th}}$ optimal control problem, given as:
\begin{equation}
h({{\mbf x}_i}(T)) = {{\mbf x}_i}^T(T){\mbf P}_i{{\mbf x}_i}(T).
\end{equation}
In (\ref{eq:objexp}), $\mb{E}$ is the expectation operator --- defined in terms of the instantaneous probabilities $p(s = s(*(\tau))$ --- as:
\begin{equation}
\label{eq:expop}
\mb{E}[s(*)] = \int_{\tau = 0}^T{s(*(\tau))p(s)}d\tau,
\end{equation}
where $*$ here represents a generic time-dependent argument of $s$, a generic function. In the preceding equations, we denote the set of admissible control laws or policies as $\mc{U} \subseteq \mathbb{R}^2$. $T \in \mb{R}$ is the time horizon for the optimal control problem, and ${{\mbf x}_i} = \begin{bsmallmatrix}
y_i & \dot{y_i} & z_i & \dot{z_i} & \phi_i & \dot{\phi_i}
\end{bsmallmatrix}^T \in \mb{R}^{6}$ is the state of the $i^{\text{th}}$ agent. $y_i$ and $z_i$ are the respective positions of the $i^{\text{th}}$ agent along the $Y$ and $Z$ inertial axes, while $\phi_i$ is its roll angle. The ($\dot{\quad}$) variables in $\mbf{x}_i$ represent the corresponding linear ($y$ and $z$) and roll rates. ${{\mbf x}_i}^r\in \mb{R}^{6}$ is the $i^{\text{th}}$ agent's reference state, ${\mbf u}_i = \begin{bsmallmatrix} u_{1_i} & u_{2_i} \end{bsmallmatrix}\in \mc{U}$ is the $i^{\text{th}}$ control, with $u_{1_i}$ and $u_{2_i}$ respectively equal to the effective gravity-opposing force produced by the propellers of the $i^{\text{th}}$ agent and the torque about the suppressed inertial $X$ axis. $\mbf{\Gamma}_i$ is the $i^{\text{th}}$ agent's trajectory equal to the vector $ \begin{bsmallmatrix} y_i & z_i \end{bsmallmatrix}^T \in \mb{R}^{2}$ while $d_{ij}^r \in \mb{R}_+$ is the prescribed inter-agent distance, which can be thought of as representing the limited communication range between agents or as a simple inter-agent proximity constraint for collision avoidance. $\mathbf{x}_i^0$ is the initial state of the $i^{\text{th}}$ agent. ${\mbf P}_i \in \mb{R}^{6\times 6}$ is the weight matrix of the $i^{\text{th}}$ terminal cost, ${\mbf R}_i \in \mb{R}^{2\times 2}$ is the weight matrix corresponding to the control cost, and ${\mbf Q}_i \in \mb{R}^{2\times 2}$ is the weight matrix for the cost corresponding to the $i^{\text{th}}$ trajectory error $({\mbf \Gamma}_i - {{\mbf \Gamma}_i}^r)$, with the $i^{\text{th}}$ time-varying trajectory ${{\mbf \Gamma}_i}^r(\tau)$ as reference. ${\mbf P}_i$ and ${\mbf R}_i$ are taken to be positive definite ($p.d.$) matrices, while ${\mbf Q}_i$ is chosen as follows:
\begin{equation*}
{\mbf Q}_i \ is \
\begin{cases}
p.d., & \text{if } a_i = a_\mc{L} \\
0 \in \mb{R}^{2\times 2}, & \text{otherwise. }
\end{cases}
\end{equation*}
\begin{figure}[!t]
\centering
\includegraphics[width=0.7\linewidth]{quad_model_alt.eps}
\caption{Planar Quadrotor Model and Coordinate Frames: $\mc O$ is the origin of the inertial frame. A moving body frame, subscripted by $\mc B$ and pictured in \textit{blue}, is attached to the agent's center of mass. $u_1$ and $u_2$ retain their former definitions.}
\label{fig:quadmod}
\end{figure}
Since the desired position $\begin{bsmallmatrix} y_{i} & z_{i} \end{bsmallmatrix}$ in the $YZ$-plane encodes the trajectory of the $i^{\text{th}}$ agent, the choice of Lagrangian in (\ref{eq:lagrgn}) -- with ${\mbf Q}_i$ as defined -- ensures that the leader tracks a specific trajectory determined by a high-level trajectory planner (see Figure \ref{fig:optcontarch}), while the other agents maintain their position in the formation. We define this desired formation by specifying rules that guide the inter-agent distances between the leader and follower agents and between the follower agents themselves (see Section \ref{ssec:formgen}). We also set an upper bound on the magnitude of the control signals for each agent ($u_{1_{\text{max}}}$ and $u_{2_{\text{max}}}$), which is standard in practice. $f_i: \mb{R}^{6}\times\mb{R}^2 \mapsto \mb{R}^{6}$ is the continuous linear time-variant state-space model describing the $i^{\text{th}}$ agent, presented in Section \ref{sec:2dquad}.
With the model in (\ref{eq:statewithw}), we rewrite the optimal control problem (\ref{eq:optprob}) in a more compact fashion as:
\begin{subequations}
\label{eq:optprobalt}
\begin{align}
\label{eq:newobjfunc}
\begin{split}
\min_{{\mbf u}_i} \quad & J_i + \ \lambda_i \cdot (d_{ij}^r - \big\lvert\big\lvert {\mbf \Gamma}_i(t) - {\mbf \Gamma}_j(t)\big\rvert\big\rvert_2) \nonumber
\end{split}\\
\textrm{subject to:} \quad & {\dot{\mbf x}_i}(\tau) = f_i({{\mbf x}_i}(\tau),{\mbf u}_i(\tau),{\mbf w})\\
&{{\mbf x}_i}(0) = {\mbf x}_i^0 \\%\ {\mbf x}_i(T) = {\mbf x}_i^T\\
\quad & \lvert {u}_{1_i}\rvert \leq u_{1_{\text{max}}}; \quad \lvert {u}_{2_i}\rvert \leq u_{2_{\text{max}}},
\end{align}
\end{subequations}
where we have introduced the inter-agent distance constraint as a penalty term in $J_i$. The objective function in (\ref{eq:optprobalt}) is the $i^{\text{th}}$ augmented Lagrangian. $\lambda_i$ is a non-negative real term that specifies whether the inter-agent distance constraint is taken into account in the $i^\text{th}$ optimal control problem, and to what degree if so. Thus, we set the value for $\lambda_i$ as follows:
\begin{equation*}
\lambda_i =
\begin{cases}
0, & \text{if } a_i = a_\mc{L} \\
\beta > 0, & \text{otherwise.}
\end{cases}
\end{equation*}
With this problem formulation and choice of $\lambda_i$, only the leader tracks the desired trajectory, while the follower agents simply keep their respective positions in the formation as determined by the triangular formation rule and corresponding inter-agent distances.
\section{Formation Specification}
\label{ssec:formgen}
In addition to tight trajectory tracking, we require the MAS to maintain a triangular formation. We elect this formation because it is geometrically well suited to the leader-follower concept and also ensures that the followers are uniformly distributed spatially on a line segment behind the leader. This pattern has been shown to be locally asymptotically stable under the assumption that the formation is infinitesimally rigid \cite{ohFormationControlMobile2011}.
To this end, we require that the formation be rigid, and translation and rotation invariant, i.e., that:
\begin{equation}
\big\lvert\big\lvert {\mbf \Gamma}_i - {\mbf \Gamma}_j\big\rvert\big\rvert_2 = d_{ij}^r \ \forall \ i, j \in \{1,2, \dots, N\};\ i\neq j,
\end{equation}
and that there exists a $\bm{\xi}_i \in \mb{R}^{2}$ for the $i^{\text{th}}$ agent such that:
\begin{equation}
\begin{bmatrix}
\mbf{R}_{\theta} & \bm{\tau}_d\\
0 & 1
\end{bmatrix} \cdot
\begin{bmatrix}
\bm{\xi}_i \\
1
\end{bmatrix}
= \begin{bmatrix}
{\mbf \Gamma}_i \\
1
\end{bmatrix},
\end{equation}
respectively, for some homogeneous transformation in $SE(3)$ comprising a fixed rotation about the $X$ axis by $\theta \in \mb{R}$ -- encoded by the rotation matrix $\mbf{R}_{\theta} \in SO(2)$ -- and a fixed translation by $\bm{\tau}_d \in \mb{R}^2$. Figure \ref{fig:triform} depicts three agents in triangular leader-follower formation with the desired inter-agent distances labeled.
\begin{figure}[H]
\centering
\includegraphics[width=0.45\textwidth]{triformfig.ps}
\caption{Three agents in the plane assuming a triangular leader-follower formation. $d^r_{\mc{L}f}$ and $d^r_{ff}$ are the desired leader-follower and follower-follower distances, respectively.}
\label{fig:triform}
\end{figure}
\section{Simulation Studies}
\label{sec:simstudies}
For simulation, we solve (\ref{eq:optprob}) using the Ipopt Python optimization package \cite{wachterImplementationInteriorpointFilter2006}, with the problem setup parameters outlined in Table \ref{tab:params}. $\mathbbm{1}_{n}$ is the $n\times n$ identity matrix. In our problem setup, we also assume that the dynamics of each agent is propagated forward in time.
\begin{table}[htb]
\caption{FHOCP Setup Parameters}
\label{tab:params}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ccccccccccc}
\hline
\multirow{2}{*}{Parameter} & $u_{1_{\text{max}}}$ & $u_{2_{\text{max}}}$ & \multicolumn{2}{c}{${\mbf P}_i$, ${\mbf K}_i$} & \multicolumn{2}{c}{${\mbf Q}_i$, ${\mbf R}_i$} & $d^r_{\mc{L}f}$ & $d^r_{ff}$ & $\beta$ & $T$\\
& [N] & [N-m] & & & & &[m] & [m]\\
\hline \hline
Value & $1.2mg$ & $\frac{I_{xx}\pi}{10}$ & \multicolumn{3}{c}{$\mathbbm{1}_{6}$} & $\mathbbm{1}_{2}$ & 0.5 & 0.5 & 1 & 10\\
\hline
\end{tabular}%
}
\end{table}
\section{Results \& Analysis}
\label{sec:res}
\subsection{Trajectory Tracking}
Figures \ref{fig:allplots} and \ref{fig:ctsine} show the optimal state variables and control inputs, respectively, for the leader agent. As expected, the optimal $\phi$ and $u_2$ values over the time horizon are approximately zero implying near hover state. Concerning tracking performance, we can see from Figure \ref{fig:yvsyd} that the optimal trajectory closely tracks the desired sinusoidal reference trajectory. As expected, in the case with no measurement noise, a much better tracking performance is recorded -- near zero trajectory error and hence, tight trajectory tracking (Figure \ref{fig:leadtrajerr}). For a numerical comparison, Table \ref{tab:rmse} presents root-mean-square error (RMSE) values for the leader agent's trajectory error, along with those for the error between the desired and actual inter-agent distances for both follower agents (abbreviated as $f_1$ and $f_2$).
\begin{figure}[htb]
\centering
\includegraphics[trim=0pt 0pt 0pt 0pt, clip, width=0.8\columnwidth]{st_n_ct_lead_sine_wd.eps}
\caption{Time evolution of the optimal configuration for the leader agent (model with $\mbf{w}$).}
\label{fig:allplots}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[trim=5pt 0pt 0pt 0pt, clip, width=0.8\columnwidth]{ct_lead_both.eps}
\caption{Optimal control inputs for the leader agent.}
\label{fig:ctsine}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[trim=5pt 0pt 0pt 0pt, clip, width=0.8\columnwidth]{lead_traj_err.eps}
\caption{Leader trajectory error.}
\label{fig:leadtrajerr}
\end{figure}
\begin{figure*}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[trim=0pt 0pt 0pt 0pt, clip, width=0.9\columnwidth]{trvstrd_lead_sine_both.eps}
\caption{}
\label{fig:yvsyd}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[width=0.9\columnwidth]{track_plus_form.eps}
\caption{}
\label{fig:trackplusform}
\end{subfigure}
\caption{(a) Optimal leader agent trajectory with a sinusoidal reference (for both model cases) and (b) A snapshot demonstrating the planar formation tracking performance for the model with $\mbf w$ (over $T$).}
\end{figure*}
\subsection{Triangular Formation Maintenance}
Results that portray the inter-agent distance maintenance results are presented in Figure \ref{fig:trackplusform} and Table \ref{tab:rmse}, from where we observe that, over the time horizon, the triangular formation is maintained as the leader closely tracks the desired trajectory. Upon the introduction of the WGN term, a comparable performance is recorded, with only slight deviations from the trajectory reference and inter-agent distance constraints --- evidenced also by the low RMSE values obtained for all three agents. A sample simulation video (showing the formation tracking performance) is available at the following link: \tee{\url{https://youtu.be/aJJXmN3UJoQ}}.
\begin{table}[htb]
\caption{RMSE values for the leader and follower agents}
\label{tab:rmse}
\centering
\begin{tabular}{ccc}
\hline
\multirow{2}{*}{Agent} & \multicolumn{2}{c}{RMSE}\\
\cline{2-3}
& Nominal Model & Model with ${\mbf w}$\\
\hline
\hline
\multirow{1}{*}{Leader} & 1.18$\times10^{-7}$ & 0.0836\\
\multirow{1}{*}{$f_1$} & 1.34$\times 10^{-12}$ & 1.4$\times10^{-9}$\\
\multirow{1}{*}{$f_2$} & 2.6$\times 10^{-9}$ & 1.3$\times10^{-9}$\\
\hline
\end{tabular}%
\end{table}
\section{Conclusion \& Future Work}
\label{sec:conc}
This paper presented results from formation tracking experiments for a special class of multiagent systems: quadrotors in the plane, restricted to states near the equilibrium. We showed that for the studied model, an overall acceptable tracking performance was recorded for the MAS, even with simulated white Gaussian noise in the model. In future work, we will apply similar ideas to the case where the agents' states evolve in space under the standard nonlinear quadrotor model. Another promising direction would be to study how the performance of the optimal control scheme is affected by scaling the number of agents.
\section*{Acknowledgements}
The authors gratefully acknowledge Microsoft Corporation and the Maryland Robotics Center for their kind financial support.
\bibliographystyle{ieeetr}
|
\section{Introduction}
Artificial Intelligence (AI) systems are increasingly adopted in various decision-making scenarios \cite{dastin2018amazon, dilsizian2014artificial, khandani2010consumer, wang2014improving, yang2018insurance}. However, AI is still far from 100\% accurate in many real-world applications \cite{bansal2020optimizing, buccinca2021trust}. Besides, due to legal and ethical concerns, it remains risky for AI to make a decision autonomously, especially in high-stake domains such as medicine, criminal justice, etc. \cite{cai2019hello, lee2021human, binns2018s}. Hence, a paradigm named AI-assisted decision-making \cite{buccinca2021trust, zhang2020effect, wang2021explanations, bansal2021does} is proposed and widely studied in HCI and AI communities. In this paradigm, AI performs an assistive role by providing a recommendation, while the human decision-maker can choose to accept or reject AI's suggestion in the final decision.
One key challenge in AI-assisted decision-making is whether the human-AI team can achieve complementary performance, i.e., the collaborative decision outcome outperforming human or AI alone \cite{bansal2021does, lai2019human, zhang2020effect}. A critical step toward complementary performance is that human decision-makers could properly determine when to take the AI's suggestion into consideration and when to be skeptical about it \cite{zhang2020effect, buccinca2021trust, rastogi2020deciding}.
Since well-calibrated AI confidence scores can represent the model's actual correctness likelihood (CL) \cite{guo2017calibration, bansal2021does, bansal2019updates}, several recent studies propose different designs to help humans allocate appropriate trust to AI based on this information \cite{zhang2020effect, bansal2021does, rastogi2020deciding}.
For example, Zhang et al. \cite{zhang2020effect} directly display AI's confidence score to human decision-makers. Bansal et al. \cite{bansal2021does} show AI's explanations for the alternative predictions if the AI's confidence is below a threshold to make humans doubt the AI. Rastogi et al. \cite{rastogi2020deciding} propose leaving more time for humans to make a decision when the AI's confidence is lower than a threshold to reduce anchoring bias.
Nevertheless, the empirical results from these studies are mixed at best \cite{zhang2020effect, bansal2021does, rastogi2020deciding, lai2022human}. There are two potential reasons. First, these works assume humans have an appropriate perception of their capability (CL) in a task instance to make reasonable decisions after knowing AI's CL. However, people usually have poorly-calibrated self-confidence that cannot reliably reflect their CL \cite{moore2020perfectly, miller2015meta, weber2004confidence, meyer2013physicians, kahneman2011thinking}. Second, these methods try to steer how much humans value AI's suggestions solely based on AI's correctness likelihood (illustrated in Figure \ref{fig:paradigm} (a)) while largely overlooking humans' correctness likelihood in each case.
This poses a question: \emph{When AI's correctness likelihood is low (high) but that of humans is even lower (higher), should we still encourage humans to doubt (trust) the AI?
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{Illustration.pdf}
\caption{The difference between prior work and this work. (A) In prior work, AI's calibrated confidence is usually used to represent the AI's correctness likelihood (CL), a value ranging from 0.5 to 1.0. They calibrate humans' trust based on an empirically set threshold, i.e., when AI's confidence exceeds this threshold they will calibrate humans to trust the AI, and when AI's confidence falls below this threshold they will calibrate humans to distrust the AI (trust themselves). (B) In this work, besides considering AI's CL, we also estimate humans' CL in each task instance. We propose calibrating humans' trust based on the relative CL of both parties, rather than solely relying on whether AI's confidence is above a preset threshold. For example, if the AI's CL is higher than the human's, we will calibrate humans to trust AI; otherwise, we will calibrate humans to trust themselves.}
\label{fig:paradigm}
\end{figure*}
To explore the answer to this question, in this paper, we propose a framework that aims to promote appropriate human trust in AI and complementary team performance according to the predicted human-AI correctness likelihood (CL) at a task instance level.
In this framework, as shown in Figure \ref{fig:paradigm} (b), we no longer have to calibrate human trust based solely on whether the AI's confidence exceeds a preset threshold. Instead, the CL of both humans and AI on a given task instance will be taken into consideration.
To verify the feasibility and efficacy of the proposed framework for promoting appropriate trust and complementary performance,
our investigation is divided into two phases: 1) How to model humans' capability (CL) on a given task? And 2) How to leverage human-AI capabilities (CL) to promote appropriate trust in AI-assisted decision-making?
In the \emph{first} phase, based on the theories from cognitive science that humans usually adopt similar solutions to deal with similar problems \cite{cacciabue1992cosimo, goldstein2014cognitive, moyer1976mental, kahneman2011thinking}, we propose to estimate people's CL on a new task according to their performance in similar tasks. For example, if a person performs well on similar tasks, her CL on the current task is also likely to be high.
However, it is often difficult to obtain enough decision data to compute human performance on similar tasks. To solve this problem, we propose a method to first approximate a human's decision-making model (a mapping from task input to human decision), then apply this model to predict the human's possible decisions in similar task instances.
In the process, a question arises, \textbf{RQ1: How to effectively approximate a human's decision-making model?}
To explore the answer, we propose to combine data-driven initialization and interactive modification to derive the possible decision rules employed by each individual. And we design an interface called \emph{interactive rule set} for users to revise the initial model to better align with their inner decision-making process. We verified the appropriateness of the designed interface compared with another interface \emph{interactive decision tree} through a preliminary study (N=20).
We take the system-initiated \& human-revised decision rule set as an approximated human decision-making model. For each new task case, we retrieve the closest cases from the existing task dataset (used for training the AI), then apply the derived models of individual decision-makers to get their likely predictions for those cases. Afterward, based on the estimated predictions and ground truth, we can calculate the probable performance of an individual, and further use this information to estimate the correct likelihood (CL) of that person on the current new task case. Combining the human CL and AI CL together, we can identify who has a higher capability in each task instance. Through a crowdsourcing study (N=30), we validated the effectiveness of our method in identifying complementary task instances (only one in the team can do it right) compared to the traditional AI confidence-based method.
In the \emph{second} phase of our work, after obtaining the estimated human-AI CL on an input task case, we further explore how to exploit this information to foster appropriate human trust and ultimately reach complementary performance in AI-assisted decision-making. In particular, we attempt to reduce human trust in AI when humans have a higher CL than AI, and increase human trust otherwise. Based on the relevant literature on people's cognitive processes \cite{bansal2019updates, nourani2021anchoring, buccinca2021trust, fogliato2022goes}, we propose three CL exploitation strategies to \emph{communicate} the CL of both sides to the responsible human decision-maker explicitly or implicitly, namely \emph{Direct Display}, \emph{Adaptive Workflow}, and \emph{Adaptive Recommendation}. Two related research questions emerge concerning these three CL exploitation strategies: \textbf{RQ2: How do different strategies affect human trust appropriateness and team performance?} And \textbf{RQ3: How do different strategies affect humans' perceptions and experiences in the decision-making process?}
Through a between-subjects crowdsourcing experiment with 293 participants, we found that our proposed three CL exploitation strategies resulted in more appropriate user trust in AI compared to baseline conditions, especially when the AI gave wrong recommendations. The three proposed CL exploitation strategies also led to improved team performance. However, different conditions did not lead to significantly different human perceptions or experiences in most subjective measures.
Our work provides a new perspective on promoting appropriate human trust in AI-assisted decision-making. In summary, our key contributions include:
\begin{itemize}
\item We propose a framework to promote humans' appropriate trust in AI-assisted decision-making at a task instance level based on the capabilities of both sides.
\item Accordingly, we design a method for estimating humans' CL on a new task instance with a data-driven initialization and interactive modification method to derive decision rules to estimate users' decision-making models.
\item We conduct two preliminary studies to verify the appropriateness of the interactive decision rule creation interface, and to verify the effectiveness of the human CL modeling method.
\item Based on the human-AI CL and related theories of humans' cognitive processes, we propose three CL exploitation strategies to foster humans' appropriate trust in AI explicitly or implicitly.
\item We conduct a user study to analyze the impact of different CL exploitation strategies on user trust appropriateness, team performance, and user experience. Based on our key findings, we provide design implications for more effective human-AI collaborative decision-making.
\end{itemize}
\section{Related work}
\subsection{Trust Calibration in AI-Assisted Decision-Making}
Trust calibration refers to the correspondence between people's trust in the AI and the AI's actual capabilities \cite{lee1994trust}. When trust exceeds the AI's capabilities, over-trust leads to misuse, which refers to when people trust AI while they shouldn't \cite{lee2004trust, parasuraman1997humans}. Under-trust, when trust is less than the AI's capabilities, leads to disuse, which refers to people failing to use it when they should \cite{lee2004trust}. These flawed human-AI partnerships can result in costly and even catastrophic outcomes. Successful decision-making requires humans to calibrate their trust in AI on a case-by-case basis \cite{zhang2020effect, bansal2019beyond, bansal2019updates, bansal2021does, turnercalibrating}.
A pivotal approach to calibrating human trust is to convey AI's capability (also called reliability or trustworthiness) to humans \cite{bansal2021does, wang2021explanations, turnercalibrating, poursabzi2021manipulating}. There are several cues that can reflect AI's capability, such as the AI's accuracy (including stated accuracy \cite{yin2019understanding, rechkemmer2022confidence} and observed accuracy \cite{yin2019understanding, rechkemmer2022confidence, ma2022glancee}), explanation \cite{poursabzi2021manipulating, lai2020chicago, lai2019human}, the actual behavior/output \cite{gero2020mental, bansal2020optimizing, bansal2019beyond}, and confidence \cite{zhang2020effect, rastogi2020deciding, bansal2021does}, etc. For example, some works help people build a mental model of AI's error boundaries by observing AI's outputs \cite{bansal2019beyond}. Also, several studies expected that if humans were shown explanations for AI decisions \cite{bansal2021does, wang2021explanations, turnercalibrating, poursabzi2021manipulating}, they would be able to identify the trustworthiness behind the prediction.
One of the most commonly used capability indicators is AI's \emph{calibrated confidence score}, as well-calibrated confidence can accurately reflect the actual correctness likelihood (CL) of the AI in a specific task instance \cite{bansal2021does, bansal2019updates}. Therefore, many recent works calibrate human trust based on AI confidence. One line of work directly displays the calibrated confidence score to people. For example, Zhang et al. \cite{zhang2020effect} compared the effects of showing and not showing AI's confidence on people's trust calibration and task performance. Another line of work integrates AI confidence into the interface design. For example, Rastogi et al. \cite{rastogi2020deciding} discovered that if given longer thinking time, people would have more cognitive resources to invest in analytical thinking and reduce being anchored by AI. Therefore, they assigned different lengths of decision-making time to humans based on AI's confidence. In addition, Bansal et al. \cite{bansal2021does} developed an adaptive explanation strategy that explains the alternative predicted classes when the AI confidence is below a threshold, otherwise only explaining the top prediction.
There are two flaws in these works. On the one hand, they assume people have an appropriate perception of their capability (CL) in a task instance to make reasonable decisions after knowing AI's CL. However, people's subjective self-confidence usually cannot accurately represent their actual CL \cite{moore2020perfectly, miller2015meta, meyer2013physicians, kahneman2011thinking}. On the other hand, these approaches calibrate humans' trust only based on AI's CL and ignore human CL. For example, existing methods make people doubt AI when AI's confidence (CL) is low. But what if the human's CL is even lower? Note that the confidence of AI just represents a ``likelihood''; thus, a prediction with low confidence can still be correct, and a high-confidence prediction may also err. To solve these problems, our work proposes a novel method for calibrating human' trust based on human and AI capability (CL).
\subsection{Mental Model in Human-AI Collaboration}
Mental models are presentations of external reality that people use to interact with the world around them \cite{johnson1983mental, norman2014some}. In human-AI collaboration, some studies investigate building humans' mental model of the AI partner \cite{bansal2019beyond, gero2020mental, nourani2021anchoring, ma2022modeling}, so that humans know whether and when to assign a task to the AI. For example, Gero et al. \cite{gero2020mental} find those who win more often have better estimates of the AI agent's abilities in a cooperative game setting. Bansal et al. \cite{bansal2021does} help humans build a mental model for the AI system's error boundary, and they found that a good mental model can help humans achieve better performance. Besides building a mental model of how AI works, a faithful mental model of how human works is also essential. For example, in human-robot interaction, some works approximate human decision policy by modeling how people will behave in different environments \cite{ng2000algorithms, deng2018prediction}. However, little attention has been paid to leveraging the model of how humans make decisions in AI-assisted decision-making. In this paper, we approximate humans' decision-making (mental) models at the instance level (i.e., given a task instance, what prediction will people make), then based on the model, we can estimate humans' CL on a new task instance.
One approach to building humans' mental models is through data-driven methods. For example, in a loan approval task, Wang et al. \cite{wang2022will} construct a general human prediction model via a neural network with crowdsourcing data. Another approach is through rule-based methods. For instance, Bansal et al. \cite{bansal2019beyond} use simple rules to build humans' mental model of AI's error boundary, such as ``(age = old \& bloodePressure = high)''. Mozannar et al. \cite{mozannar2022teaching} ask humans to formalize their mental model of AI's error regions by writing a rule describing the region after solving a set of selected examples. Especially, rule-based methods have the advantage of interpretability \cite{kulesza2013too, lim2009and, lakkaraju2016interpretable, liao2021human, arrieta2020explainable}.
In this work, we propose to combine data-driven initialization and interactive rule modification to derive the possible decision-making mental model employed by individuals. This method has two advantages. First, it saves people's time by training an initial model via a small amount of user decision data, so that the model does not need to be built from scratch. Second, the model can also be presented to the user for manual interactive refinement.
\subsection{Cognitive Bias and Human Reliance in AI}
In human-AI interaction, as people are generally inclined to engage in System 1 thinking \cite{kahneman2011thinking}, there are often various cognitive biases, including common anchoring bias \cite{nourani2021anchoring}, confirmation bias \cite{nickerson1998confirmation}, automation bias/aversion \cite{cummings2004automation}, availability bias \cite{wang2019designing}, illusion of validity \cite{simkute2020experts}, etc. These cognitive biases can (negatively) affect people's trust in AI. For example, after observing model behaviors early on, people often have an anchoring bias towards AI's suggestions \cite{nourani2021anchoring}, leading to over-rely on AI's suggestions. People are also often brought by the illusion of validity of the information displayed by AI \cite{lai2019human, kaur2020interpreting, eiband2019impact}. For example, Kaur et al. \cite{kaur2020interpreting} find that the existence of explanations could mistakenly lead to data scientists' over-confidence that the model is ready for deployment. Eiband et al. \cite{eiband2019impact} find that even placebic explanations, which do not convey useful information, invoke a similar level of trust as real explanations do.
In order to reduce the adverse effects of cognitive biases on human-AI cooperation, existing works have proposed some mitigation methods. One way is to provide interventions to nudge people to engage deeper in System 2 thinking \cite{kahneman2011thinking}. For example, research on ``cognitive forcing'' has explored methods for pushing human decision makers to spend more time deliberating about problems \cite{buccinca2021trust, park2019slow, rastogi2020deciding}, such as asking humans to make independent predictions before seeing AI's suggestions \cite{buccinca2021trust} or employing a ``slow algorithm'' \cite{park2019slow}. These cognitive forcing functions are found to be able to \textbf{decrease} humans' AI reliance. Other mitigation methods include enabling people to actively explore the data \cite{wang2019designing, simkute2020experts}, explaining clearly and training users on how to use explanations/AI \cite{lai2020chicago}, giving arguments for non-predicted outcomes \cite{bussone2015role}, monitoring user's anchored status \cite{echterhoff2022ai}, showing prior probabilities of outcome \cite{wang2019designing}, etc.
In this work, we ``\emph{leverage}'' humans' cognitive biases to help us calibrate humans' trust by incorporating cognitive biases into adaptive interaction design. Specifically, we do not blindly increase or decrease people's trust in AI. Instead, we regulate the distribution of people's trust according to the CL of both parties. When AI's CL is higher, we utilize anchoring bias to promote people's trust in AI, and when human's CL is higher, we deploy cognitive forcing to promote human-independent analytical thinking.
\section{\mxj{Phase I: Modeling Humans' Capability on A Given Task}}
\section{Phase I: Modeling Humans' Correctness Likelihood on a Given Task Instance}
\subsection{Overall process of human correctness likelihood modeling}
We investigate our human correctness likelihood (CL) modeling method in a typical AI-assisted decision-making scenario, where ground truth data is available for a task dataset but not for the current case. To promote more appropriate human trust in the AI-assisted decision-making process, the first phase of this work is to estimate humans' capabilities at a task instance level. Inspired by research in cognitive science which suggests that humans make decisions by weighing similar past experiences \cite{bornstein2017reminders, cacciabue1992cosimo, goldstein2014cognitive, moyer1976mental, kahneman2011thinking}, we propose to estimate humans' CL at a given task instance \cite{moore2020perfectly} based on their past performance in similar task instances. However, it is often difficult to obtain enough decision data to compute human performance in similar task instances, especially for a new task.
To solve this problem, we design a method first to approximate a human's decision-making model (i.e., to get a mapping from task input to human decision). Then we apply this model to predict the human's possible decisions in similar task instances and calculate their potential accuracy in these instances compared to the ground truth. To answer the question, \textbf{RQ1: How to effectively approximate a human's decision-making model?}, we propose to combine data-driven initialization and interactive modification to derive the possible decision rules employed by each individual.
Our proposed human capability modeling method goes through four steps, as illustrated in Figure \ref{fig:modeling}.
\begin{itemize}
\item Step 1: Collect user predictions: we gather the decision data of people on a small number of sampled task instances.
\item Step 2: Generate initialized decision models: we fit a classic decision tree model to infer humans' decision-making models and generate initial decision rules \cite{song2015decision, gennatas2020expert}.
\item Step 3: Interactively modify decision models: due to the limited amount of training data, the initial model generated may not reflect people's actual decision-making model well. Therefore, we design an interactive interface for users to revise the initial model to make it better align with their inner decision-making process.
\item Step 4: Apply decision models to estimate the correctness likelihood of new cases: We apply humans' approximated decision-making model to the neighbor cases of the current task case and compute humans' possible performance. Then, based on a distance-weighted method (see Eq. \ref{equation1}), we can estimate humans' CL in the current case.
\end{itemize}
In the following subsections, we present our task setup and introduce the details of the four steps through two small-scale studies (with Study I.1 focusing on Steps 1-3 and Study I.2 focusing on Step 4).
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{Pipeline.pdf}
\caption{The human capability modeling process. The whole process goes through four steps. Step 1: Collect user predictions. Step 2: Generate initialized decision models. Step 3: Interactively modify decision models. Step 4: Apply decision models to estimate correctness likelihood.}
\label{fig:modeling}
\end{figure*}
\subsection{Task setup}
\label{task-setup}
\subsubsection{Task selection}
We chose \emph{income prediction} as our testbed which has been used in several previous studies on AI-assisted decision-making \cite{zhang2020effect, hase2020evaluating, ribeiro2018anchors, ghai2021explainable}. In this task, a participant was asked to predict whether a given person's annual income would exceed \$50K or not based on some demographic and job information. The data used for the task came from the Adult Income dataset \cite{datasetucl} in UCI Machine Learning
Repository. The entire dataset has 48,842 instances of surveyed individuals, each described by 14 attributes such as age, occupation, etc. These people's actual annual income was recorded and binarized (greater/less than 50K) as the ground truth for assessing participants' prediction accuracy. This task is suitable for our study since it requires little domain expertise and imposes relatively limited risks, and thus is amenable for non-expert participants \cite{ghai2021explainable}.
To ensure the task has a reasonable complexity for lay people to establish a decision-making model, following \cite{zhang2020effect, ghai2021explainable}, we selected the five most important features out of the 14 attributes as the final attributes presented to participants, determined by the feature importance values based on the feature permutation method \cite{altmann2010permutation}. These attributes include age, year of education, occupation, marital status, and work hours per week. This number of features is suggested to be appropriate for non-expert users to form a decision-making model by experiencing several task samples (e.g., Bansal et al. \cite{bansal2019beyond, bansal2019updates} established users' mental models of AI's error boundaries using a three-feature task). Future work can be extended to simulate humans' decision-making models in more complex tasks.
\subsubsection{AI model}
Same as \cite{ghai2021explainable}, we chose a logistic regression model (using a default setting from \emph{sklearn}) as our AI model to assist humans in making decisions in the selected income prediction task. As the logistic regression model directly optimizes Log loss, it can return well-calibrated confidence scores
\cite{platt1999probabilistic}.
Calibrated confidence of a model can provide an accurate probability of correctness for the model's predictions. For example, if a model makes a prediction on a sample with 0.6 confidence (calibrated), there will be a 60\% chance that the prediction is correct, or equivalently, if a model makes predictions on M samples with 0.6 confidence, there will be around 0.6 * M samples that are actually correctly predicted.
Note that some ML classifiers (such as SVM and neural networks) cannot directly generate calibrated confidence scores \cite{guo2017calibration, bansal2021does, zhang2020effect}, so post-hoc calibration is required (such as Platt Scaling or Isotonic
Regression \cite{platt1999probabilistic, guo2017calibration}).
Our model was trained based on a 70\% random split of the original dataset, while the prediction trials given to the participants in the experiment were drawn from the remaining 30\%. For any new task cases in the testing set, our human capability estimation method will retrieve similar cases from the training set to predict humans' CL.
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{two_interface.pdf}
\caption{Two decision-making model creation interfaces. (A) The interactive decision tree. (B) The interactive rule set.}
\label{fig:rule_interfaces}
\end{figure*}
\subsubsection{Task cases selection}
The selected task cases for the user studies satisfy several criteria. First, to make the human-AI teaming setting more suitable for pursuing complementary performance, humans' independent accuracy on these samples should be comparable to that of AI \cite{bansal2021does, zhang2020effect}. Second, these cases should follow the data distributions in the test set \cite{wang2021explanations}. Third, AI's confidence scores in these samples should be well-calibrated to reflect its actual CL \cite{wang2021explanations, zhang2020effect}.
To keep the user studies at a proper length without causing fatigue in participants, we selected 40 task cases, which are split into two batches. The first 20 samples are used to get humans' decision data and build their decision-making models computationally. The remaining 20 samples are used in the main AI-assisted decision-making task. While the two batches of samples are fixed for all participants, the presentation order of samples inside each batch was randomized.
To make AI performance comparable to humans' independent accuracy, following \cite{bansal2021does}, we first conducted an additional pilot study to determine the average prediction accuracy of unassisted humans over 20 randomly picked task instances, which was around 70\% according to the results. We then selected 40 task samples over which the AI model had a 70\% accuracy with equal positive and negative labels, as well as equal false positive and false positive rates (similar setting as \cite{bansal2021does}).
To guarantee the representativeness of the selected samples, we made sure that most of the common values of each feature were included in these 40 samples. We also carefully controlled the AI's confidence in these instances to make it align with AI's actual CL.
Specifically, out of the 20 samples in each batch, half of them had a confidence score lower than 0.7 (representing low-confidence samples, with an average value of 0.6), of which 6 samples were correctly predicted by the AI (the CL was $6/10 = 60\%$). Another half of them had a confidence score higher than 0.7 (representing high-confidence samples, with an average of 0.8), of which 8 samples were correctly predicted by the AI (the CL was $8/10 = 80\%$).
Once the AI models and task cases are ready, we conducted a lab study to explore the suitable interface for non-expert users to interactively revise their decision-making models. Note that one may ask, \emph{why do we need to model human correctness likelihood (CL)? Just as AI's confidence can indicate its CL, can people's self-confidence represent their CL?} We carried out a small-scale user study and found that participants had poorly-calibrated subjective confidence. That is, the correlation between their actual accuracy and self-reported confidence is statistically unrelated, suggesting that self-reported confidence is not a reliable human CL indicator. The details can be found in the supplementary material.
\subsection{Study I.1: Comparison of interfaces for users to specify their decision-making models}
\label{decision_rule}
According to existing research, rules are considered to be an appropriate mechanism for approximating human decision-making processes \cite{gennatas2020expert, furnkranz2012foundations, furnkranz2020cognitive}. Humans, on the other hand, often make judgments based on decision tree-like structures \cite{damez2005fuzzy}. Hence, to compare the efficacy of these two representations, we design two interfaces for displaying and interactively updating humans' decision-making models (i.e. \emph{interactive decision tree} versus \emph{interactive rule set}). Both interfaces share the same initialization method where we fit a decision tree model (default setting from \emph{sklearn}) on the human decision data from the first 20 task cases. We chose the decision tree model instead of a black-box model because it can be easily understood even by people without machine learning knowledge \cite{damez2005fuzzy}.
These two interfaces are shown in Figure \ref{fig:rule_interfaces}. The \emph{interactive decision tree} interface (Figure \ref{fig:rule_interfaces} (a)) directly displays the decision tree model generated in the backend. On the interface, humans are first shown a tutorial about how to interpret and modify the decision tree (not included in the figure). Then, they can browse their past decision data on the first 20 instances to recall their decision-making rationale (Figure \ref{fig:rule_interfaces} (1)). Finally, they can add, delete or modify any tree node to reflect their actual decision process.
The \emph{interactive rule set} interface (Figure \ref{fig:rule_interfaces} (b)) presents the set of \emph{if-then} rules converted from the decision-tree \cite{lim2009and}.
With this interface, similarly, humans first view a tutorial, next revisit their historical decision data, and finally they can add, delete, or modify any rules or conditions in a rule (Figure \ref{fig:rule_interfaces} (2)).
\subsubsection{Study procedure and participants.} We conducted a between-subjects study, recruiting 20 participants (8 Female, 12 Male, average age: 27) from a local research university to build their decision models using the assigned interface (10 for each condition).
After giving their consent, they followed a tutorial to familiarize themselves with the income prediction task. Then, they proceeded to finish 20 prediction tasks (the first batch) without the help of AI. Upon completion, they were asked to use the assigned interface to create their decision model. We mainly focused on their qualitative perceptions of the interface, so we carried out an exit interview with them at the end of the study.
\subsubsection{Results} According to the interview results, seven out of the 10 participants using the \emph{interactive decision tree} interface reported that their actual decision process could not be well represented by a decision tree. For example, P3 (Male, 30, little knowledge in AI) noted, ``\emph{My actual decision process was not a single (decision) tree. Sometimes, I use `age' as the first criterion, but sometimes, I use the `year of education' as the main factor. }'' Furthermore, three out of the 10 participants found the decision tree to be visually complex. For instance, P9 (Male, 26, little knowledge in AI) mentioned that ``\emph{The tree is hard for me to read in a short time.}'' In comparison, the \emph{interactive rule set} interface was considered to be more visually interpretable and more in line with participants' decision-making process by users in this condition. Therefore, in the final version, we employ the \emph{interactive rule set} interface for participants to revise their decision-making models interactively.
Based on participants' feedback, we also improved the \emph{interactive rule set} interface. For each rule, we provide a ``check'' button, clicking on which allows users to check how many of their historical decisions conflict with this rule and whether this rule conflicts with other created rules (Figure \ref{fig:rule_interfaces} (3)).
\subsubsection{Discussion}
While a decision rule set is better suited for simulating human decision-making models, it also has some limitations. First, there are sometimes edge cases that are difficult to cover by limited decision rules \cite{furnkranz2020cognitive}. For these cases, we now use the system-initialized model to cover. Second, some users make decisions based on intuition, which can not be formulated as an explicit set of rules. Third, it may be difficult for non-expert users to form accurate decision rules by experiencing only a small set of task samples. We will discuss these in more detail in Sec. \ref{limitation}.
\subsection{Study I.2: Performance testing of our human correctness likelihood estimation method}
Based on the user-revised decision-making model, we can get their possible predictions for $N$ similar task instances retrieved from the training set. And comparing their possible predictions and ground-truth labels (already known), we can compute humans' potential performance on these task instances to obtain an estimated CL for the current task instance. We empirically set the number $N$ to 10 in this work to achieve a trade-off between sufficient similarity and coverage. If the number is set too large, a lot of dissimilar samples will be calculated and if the number is set too small, the sample size is insufficient to obtain a stable accuracy value. Note that the number can be different in other tasks with different properties. We calculate human correctness likelihood $CL_c$ on the current task instance $I_c$ based on the following equation.
\begin{small}
\begin{equation}
CL_c = \frac{\sum_{i=1}^{N} w^i \cdot IF(\hat{y}^{i} = y^i, 1, 0) + (1-w^i) \cdot 50\%}{N}, \quad where\ w^i = \frac{\alpha}{\alpha + d(\mathbf{x}_c,\mathbf{x}_n^i)}.
\label{equation1}
\end{equation}
\end{small}
where $\hat{y}^{i}$ is the human possible prediction in the $i$-th neighbor instance $I_n^i$, and $y^i$ is the ground-truth label of that instance. $IF(\hat{y}^{i} = y^i, 1, 0)$ means if $\hat{y}^{i} = y^i$, returns 1, otherwise, returns 0. And $w^i$ is the weight of each neighbor instance, $d(\mathbf{x}_c,\mathbf{x}_n^i)$ is the Euclidean distance between the current task instance $I_c$'s feature vector $\mathbf{x}_c$ and its neighbor instance $I_n^i$'s feature vector $\mathbf{x}_n^i$. We can see that the weight is negatively correlated with the distance. More similar neighbor instances will have a greater impact on performance computations. For example, if a human can make a correct prediction for a very close neighbor instance ($d(\mathbf{x}_c,\mathbf{x}_n^i) \rightarrow 0$), it will contribute $1/N$ to CL. If a human makes a correct prediction for an (extremely) distant task instance ($d(\mathbf{x}_c,\mathbf{x}_n^i) \rightarrow \infty$), the distance factor will discount its contribution and move $w^i$ closer to 0 (i.e., it only contributes $0.5/N$ to CL, which is equal to random guessing). We set the parameter $\alpha$ to 2 based on the median Euclidean distance between any two instances in the training set. While other values may be more appropriate, we leave this to future work.
Combining the human CL and AI CL (calibrated AI confidence), for a new task instance, we can estimate which member in the human-AI team has a higher correctness likelihood. Next, we verify the effectiveness of our method with two objectives. First, the estimated human CL should be significantly correlated with the actual human accuracy. Second, recall that a key purpose of our approach to modeling human capabilities is to better distinguish when to trust the AI and when to trust themselves. So we focus on the \emph{complementary region}, where for each case, only one member of the human-AI team can make a correct prediction. If the human is estimated to have a higher CL on a case, this case will be labeled ``human better''; otherwise, ``AI better''. In comparison, in the AI confidence-based method, same as previous works \cite{zhang2020effect}, when the AI's confidence exceeds the set threshold (0.7), we regard this case as ``AI better'' and otherwise ``human better''. We quantify the effectiveness as the $recall$ of complementary cases, i.e., the ratio of complementary cases that are correctly predicted by our method out of the whole \emph{complementary region}. We didn't focus on the $precision$ because in the case where both humans and AI can make correct or incorrect predictions, whoever has a higher likelihood won't lead to significantly different consequences.
\subsubsection{Study procedure and participants.} In the same setting as Study I.1 (Sec. \ref{decision_rule}), we conducted a crowdsourcing study to compare the effectiveness of our method and the AI confidence-based method. We recruited 30 participants from Prolific\footnote{www.prolific.co\label{prolific}} (18 Female, 11 Male, 1 non-binary, aged from 21 to 61, avg 35, all reside in the US). The study procedure was the same, except that we also asked participants to complete the remaining 20 tasks after creating their decision rules (again, without the assistance of AI, so that we could measure humans' independent correctness).
\subsubsection{Results.} We found that based on the auto-generated human decision-making model, the prediction accuracy of participants' decisions on the last 20 task instances was 77.5\%. In comparison, based on the human-revised decision-making model, the accuracy was 80.7\%. This shows a slight but not significant improvement. We speculate that this is because the default decision tree model is already close to the human decision-making process, so participants can only make minor adjustments to the initialized rules.
Following \cite{depaulo1997accuracy, rechkemmer2022confidence}, we calculated the Pearson correlation between our estimated humans' average CL and their actual accuracy on the last 20 tasks. The result showed a significantly positive correlation ($r$=0.482, $p$<.01).
Furthermore, we found that our human-AI CL method could recall 76.4\% of the \emph{complementary region} on average, while the AI confidence-based method could recall 66.7\% of the \emph{complementary region} on average. Paired T-tests showed significant differences ($p$<.05). The results validated that our method was more effective than the traditional AI confidence-only methods at guessing the human-AI CL on complementary cases.
\subsubsection{Discussion}
We note that our method highly relies on the accuracy of the approximated human decision-making models. Besides, although our method is better than the AI confidence-based method in identifying complementary cases, the improvement is not very large. We speculate it might be due to the limited complementarity of humans and AI, which affects the superiority of our method. We will discuss this issue in the final Discussion.
\section{Phase 2: Communicating Human-AI Correctness Likelihood to Promote Appropriate Trust}
\section{Phase II: Communicating Human's and AI's Correctness Likelihood to Promote Appropriate Trust}
The second phase of this work is to explore how to integrate the modeled human-AI correctness likelihood (CL) to empower the AI-assisted decision-making process. Specifically, we propose three different strategies to exploit CL, i.e., \emph{Direct Display}, \emph{Adaptive Workflow}, and \emph{Adaptive Recommendation}. Then, through a between-subjects experiment, we aim to investigate two research questions: \textbf{RQ2: How do different CL exploitation strategies affect human trust appropriateness and team performance?} and \textbf{RQ3: How do different CL exploitation strategies affect humans' perceptions and user experiences in the decision process?}
\subsection{Experimental Conditions and Interface Design}
To help people realize when to refer to the AI's suggestion and when to rely on themselves, one intuitive mechanism is to \emph{explicitly} display the human and AI CL information to human decision-makers.
\begin{itemize}
\item \textbf{Direct Display}: We directly present the estimated human and AI CL and the AI's recommendations to the human (Figure \ref{fig:condition-interfaces} C) in this condition. To be more specific, on the experimental website, alongside the profile area (the five attributes of the person to predict, Figure \ref{fig:condition-interfaces} A1), the system illustrates the estimated CL of the human and AI side by side (Figure \ref{fig:condition-interfaces} C3). At the top of this area is a summary sentence, "According to the system's estimation, in this task case, the AI (you) might have a higher probability of making a correct decision than you (the AI)". Below are two gauge graphs showing the CL values of humans and AI, respectively, followed by the recommendation (i.e., the predicted income) from AI. Finally, people need to input their final decision (Figure \ref{fig:condition-interfaces} B4).
\end{itemize}
In this condition, it is up to human users to decide how to interpret the CL information and whether to trust the AI.
However, we acknowledge that the estimates of humans' and AI's capabilities are far from perfect, and relying on this information to assess AI's suggestions may have serious consequences, especially in high-risk areas. For example, in a clinical decision-making scenario, physicians may develop false self-confidence in their diagnosis if our model overestimates their abilities.
To mitigate this issue, we propose two other \emph{implicit} CL exploitation strategies based on theories in cognitive science. On the one hand, according to the \emph{anchoring bias} theory in decision-making \cite{epley2006anchoring, furnham2011literature, rastogi2020deciding, buccinca2021trust, epley2001putting}, if human decision-makers have access to anchors (such as AI's opinions), they are likely to diminish further exploration of alternative hypotheses and \textbf{increase} humans' reliance on AI. On the other hand, research on ``cognitive forcing'' has explored methods for pushing human decision-makers to spend more time with deliberating about problems \cite{buccinca2021trust, park2019slow, rastogi2020deciding}.
These cognitive forcing functions are found to be able to \textbf{decrease} humans' reliance on AI. Based on these theoretical supports, we propose the following condition.
\begin{itemize}
\item \textbf{Adaptive Workflow}: In this condition (Figure \ref{fig:condition-interfaces} D), we adaptively change the order of human and AI decisions based on the estimated human and AI CL. If the predicted human CL is higher than that of the AI, our interface will first ask human users to input their initial decision and then reveal the AI's recommendation (Figure \ref{fig:condition-interfaces} D5). Likewise, if the AI's CL is estimated to be higher than the human's, our interface will directly present the AI's suggestion to the human (Figure \ref{fig:condition-interfaces} D5 will not be displayed). Both cases require the human to make the final decision after reviewing the AI's recommendations.
\end{itemize}
Another implicit way to leverage \emph{cognitive forcing} to promote appropriate trust is not to show AI suggestions to people if they have higher CL than the AI. However, this will prevent people from taking advantage of AI's assistance in such cases. A trade-off solution is to provide AI's explanations but not the prediction result to users when humans have a higher CL than AI so that they have to make their own decisions. Garhos et al. \cite{gajos2022people} found that people engaged more in analytical thinking based on their own knowledge if AI's explanations were shown without concrete recommendations. Therefore, we propose the following condition.
\begin{itemize}
\item \textbf{Adaptive Recommendation}: In this condition, the AI shows the explanation of its prediction (generated by LIME \cite{ribeiro2016should}, a widely used XAI method) by default. We control the display of AI's recommendation based on the comparison between the estimated human and AI CL. If the AI's CL is higher, our interface will display AI's recommendation (Figure \ref{fig:condition-interfaces} E6) along with its explanation (Figure \ref{fig:condition-interfaces} E7). If the human CL is higher, our interface will \emph{not} disclose the AI's recommendation to users (Figure \ref{fig:condition-interfaces} E6 hidden).
\end{itemize}
Besides the above-mentioned three conditions, we also include two baseline conditions following \cite{bansal2021does, rastogi2020deciding}.
\begin{itemize}
\item \textbf{Human Only}: In this condition, humans must make their own decisions without any AI assistance (Figure \ref{fig:condition-interfaces} A).
\item \textbf{AI Confidence}: In this condition, humans are presented with AI's recommendation and its calibrated confidence but \emph{without} human CL information (Figure \ref{fig:condition-interfaces} B2). We incorporate this baseline because it is a broadly acknowledged design to calibrate humans' trust in AI-assisted decision-making \cite{zhang2020effect, bansal2021does}.
\end{itemize}
The interfaces were tested through a pilot study to ensure that the workflow was clear for participants to follow.
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{study_interface.pdf}
\caption{Interface of different conditions. (A) \textbf{Human Only}. (B) \textbf{AI Confidence}. (C) \textbf{Direct Display}. (D) \textbf{Adaptive Workflow}. (E) \textbf{Adaptive Recommendation}. The interfaces of all conditions share a similar layout: the left side is a person's profile area and the right side is a decision-making area. To save space, we do not draw the person profile area repeatedly in Figure \ref{fig:condition-interfaces} (C), (D), (E).}
\label{fig:condition-interfaces}
\end{figure*}
\subsection{Study Design}
We employ a between-subjects study design with the five conditions. The study is approved by the University IRB.
\subsubsection{Task and Procedure}
We adopted the same task as in Phase I (Sec. \ref{task-setup}), i.e., predicting whether a person's annual income exceeds \$50K. Participants went through five stages during the study (shown in Figure \ref{fig: procedure}): (1) Introduction: After obtaining the consent of participants, we provided a tutorial walk-through to familiarize them with the task where we detailed the meaning and value range of each attribute in the profile table to the participants. For each attribute, we presented a graph showing the distribution of the corresponding income in the entire dataset, giving participants a basic understanding of the salary situation. We inserted two attention-check questions at the end of the tutorial to help filter out participants who did not read the introduction carefully. After the tutorial, we provided participants with two training examples with ground truth. (2) First batch of 20 tasks: Next, participants proceeded to complete the first 20 task cases independently (no AI advice or ground truth information was displayed). (3) Interactive decision rule creation: Participants entered the decision rule creation page (see Sec. \ref{decision_rule} for details). (4) Second batch of 20 tasks: Once done customizing their own decision rules, participants moved on to the last 20 task cases, this time, with AI's assistance (except for \emph{Human Only} condition). Depending on the assigned conditions, different interfaces were presented to the participants (as shown in Figure \ref{fig:condition-interfaces}). (5) Exit survey: Finally, participants were asked to fill out an exit survey in which we collected basic demographic information as well as subjective measures and open-ended feedback about their perceptions in the decision-making process, which are described later in Sec. \ref{questions}.
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{procedure.pdf}
\caption{Procedure of the experiment. Participants go through five stages in the whole study.}
\label{fig: procedure}
\end{figure*}
\subsubsection{Participants}
We recruited 300 participants (60 for each condition) from Prolific\textsuperscript{\ref{prolific}}. To ensure high-quality responses, all participants met the following criteria: (1) residing in the United States (as the task was to predict income for adults in the United States); (2) at least 99\% approval rate for previous submissions; (3) using English as the first language; (4) owning a bachelor's degree or above; and (5) using a desktop computer for the experiment. The study followed a between-subjects design, so we did not allow any repeated participation. In total, we got 293 complete submissions. After filtering based on the attention-check questions, we obtain 289 valid responses (\emph{Human Only}: 59, \emph{AI Confidence}: 59, \emph{Direct Display}: 59, \emph{Adaptive Workflow}: 56, \emph{Adaptive Recommendation}: 56). Among the final participants, there were 174 self-reported male, 110 female, and 5 non-binary. A total of 77 participants were aged 18-29, 116 aged 30-39, 45 aged 40-49, 28 aged 50-59, and 23 aged over 59. Participants also rated their knowledge of artificial intelligence: 40 had no knowledge, 205 knew basic concepts in AI, 43 had used AI algorithms, and one was an expert in AI. To motivate high-quality work, in addition to the base payment, we gave participants a \$0.50 bonus if their overall accuracy exceeded 80\%. The entire study lasted about 20 minutes. The average wage for participants was about \$9.34 per hour.
\subsection{Evaluation Metrics}
\label{questions}
\subsubsection{Measures for RQ2.} We investigate the effects of different conditions on humans' trust appropriateness and human-AI team performance through two main measurements. (1) \emph{Human-AI Agreement} \cite{zhang2020effect, bansal2021does}: the fraction of tasks where the participant's final decision agreed with the AI's recommendation, whether it is right or wrong. (2) \emph{Team Performance} \cite{bansal2021does, zhang2020effect, rastogi2020deciding, wang2021explanations}: the final decision accuracy. We also collected participants' \emph{Perceived correctness likelihood (CL)}, where in each task instance, we asked participants which one (human, AI, or both) they thought had a higher CL.
\subsubsection{Measures for RQ3.} Here, we focus on participants' experiences and perceptions in different conditions. Specifically, referring to and adapted from related works, we investigate the following subjective measures as 7-point Likert scale questions in the exit survey (1: Strongly Disagree, 7: Strongly Agree): (1) \emph{Trust in AI} \cite{ghai2021explainable, buccinca2021trust}; (2) \emph{Confidence in the decision-making process} \cite{lai2022human, rechkemmer2022confidence}; (3) \emph{Perceived complexity of the system} \cite{buccinca2021trust}; (4) \emph{Mental demand} \cite{hart2006nasa, lai2022human, ghai2021explainable, buccinca2021trust}; (5) \emph{Perceived autonomy} \cite{hong2019racism}; (6) \emph{Satisfaction} \cite{ghai2021explainable}; (7) \emph{Future use} \cite{brooke1996sus}; (8) \emph{Trust in the estimation of human-AI CL}; (9) \emph{Perceived usefulness of estimation of human-AI CL} \cite{laugwitz2008construction}; (10) \emph{Perceived helpfulness to decide when to trust the AI} \cite{laugwitz2008construction}); and (11) \emph{Acceptance of estimation of their CL}. Besides these questions, we also asked participants open-ended questions about how they used and perceived the communicated human-AI CL, and how their decision-making processes were affected by different interface designs. Detailed questions can be found in the supplementary material.
\subsubsection{Analysis Methods}
\label{analysismethod}
We conducted mixed-methods analyses on the aforementioned metrics. For quantitative analysis of the objective data for RQ2 and participants' subjective data for RQ3, since most of the data did not follow a normal distribution, we carried on non-parameter tests. Specifically, for pair-wise comparison, we ran Mann-Whitney U Test or Wilcoxon Signed Ranks Test based on whether the sample was from the same group of participants. And for analysis among more than two groups of participants, we ran Kruskal-Wallis Test and post-hoc analysis with Bonferroni correction. For qualitative analysis, two authors coded the open-ended feedback via inductive thematic analysis \cite{hsieh2005three}. The final themes were discussed and harmonized over several iterations, and specific examples were identified from the source texts for demonstration in this paper.
\section{Results}
\subsection{Effects of CL Exploitation Strategies on Team Performance and Human Trust in AI}
We organize the results into two parts. In the first part, we analyze the participants' overall team performance and trust in AI. In the second part, we dig deeper into the data and analyze the results according to different situations (e.g., different human-AI CL). All results are organized into ``Findings'' for easy reading.
\textbf{Part 1}
\textbf{Finding 1: The trend of human trust in AI was consistent with the trend of estimated human-AI CL in the proposed three conditions.}
Since the basic intention of our three designs is to make people rely more on the member with higher CL in the human-AI team, we want to see if our approaches made people trust AI \textbf{\emph{more}} when the AI's CL was higher and trust AI \textbf{\emph{less}} otherwise.
Figure \ref{fig:agreement} shows the human-AI agreement in different human-AI CL under three conditions. Results showed that all three conditions made people's agreement with AI significantly higher when AI's CL was higher than that when the human's CL was higher ($p$<.001 in all three conditions).
\begin{figure*}[t]
\centering
\includegraphics[width=0.5\linewidth]{humanAIagreement_different_CL_NEW.pdf}
\caption{Human-AI agreement in different human and AI correctness likelihood (with mean and 95\% confidence interval). We can see that when the AI's CL is higher than the human's, participants tend to agree with AI more in the final decision.}
\label{fig:agreement}
\end{figure*}
\textbf{Finding 2: The proposed three CL exploitation conditions achieved complementary performance while the \emph{AI Confidence} condition did not.} Figure \ref{fig:team_performance} (a) shows the overall team performance (i.e., the accuracy of humans' final decisions) in all conditions. The team performance in \emph{Direct Display} ($M$=0.758, $SD$=0.078), \emph{Adaptive Workflow} ($M$=0.767, $SD$=0.085), \emph{Adaptive Recommendation} ($M$=0.754, $SD$=0.072) surpassed both AI alone (0.7) and \emph{Human Only} ($M$=0.721, $SD$=0.113). However, the team performance of \emph{AI Confidence} ($M$=0.720, $SD$=0.099) did not outperform \emph{Human Only}. From Figure \ref{fig:team_performance} (b), we can see that although \emph{AI Confidence} made humans agree with AI more when AI's recommendation was correct, it also made humans agree with AI's wrong recommendation more. This finding is consistent with previous work revealing that showing AI confidence does not necessarily improve team performance \cite{zhang2020effect, rastogi2020deciding}. Compared with \emph{AI Confidence}, the proposed CL exploitation methods achieved marginally to significantly higher team performance (\emph{Direct Display}: $p$=.078; \emph{Adaptive Workflow}: $p$=.027; \emph{Adaptive Recommendation}: $p$=.078).
\textbf{Finding 3: Humans in the proposed three CL exploitation conditions trusted the AI more appropriately when the AI's recommendation was wrong.} As shown in Figure \ref{fig:team_performance} (b) (the red bars), the human-AI agreement in \emph{AI Confidence} was significantly higher than \emph{Adaptive Workflow} ($p$=.017), and \emph{Adaptive Recommendation} ($p$<.001), and it was marginally higher than \emph{Direct Display} ($p$=.07). Note that when AI is wrong, a lower agreement with AI is better. These results suggest humans' \textbf{less over-trust} in AI in our proposed CL exploitation conditions. Through qualitative feedback from participants, we found that our three designs prompted people to rely more on their own thinking when AI's advice was wrong (AI's CL was often lower than human's). Specifically, in \emph{Direct Display}, displaying a higher human CL made them more confident in their own judgment. For example, P26 (49, female, knew basic knowledge of AI) said, ``\emph{Sometimes the AI's opinion differed from mine. When I saw that my (CL) value was higher than the AI's, it confirmed my opinion.}'' While in \emph{Adaptive Workflow}, in most cases, when participants had to make their own judgment first, then did not change their decision later, even if the AI's recommendation was the opposite. For example, P16 (31, male, knew or used AI algorithms) mentioned ``\emph{I had made careful thinking before the AI's suggestion, and I would stick to my own opinion.}'' Similar phenomenon can be found in \emph{Adaptive Recommendation}. P4 (40, female, no knowledge of AI) said, ``\emph{AI didn't tell me any answers, I could only make decisions according to my own thoughts.}''
However, when the AI's recommendation was correct (the red bars in Figure \ref{fig:team_performance} (b)), we did not observe significant differences in human-AI agreement between the three proposed conditions and \emph{AI Confidence} baseline. We infer this might be because the task instances where humans and AI could make correct predictions were highly overlapped. It can be seen from \emph{Human Only} that in the case where AI gave correct advice, even if people did not get any assistance from the AI, their performance also reached ~80\% (agree with AI on 80\% cases), which indicates the complementarity of human and AI in such situations was relatively weak, and the room for improvement was thus limited.
\begin{figure*}[h]
\centering
\subfigure[]{
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth]{overall_performance.pdf}
\end{minipage}%
}%
\subfigure[]{
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth]{agreement_wrong_correct.pdf}
\end{minipage}%
}%
\centering
\caption{Overall team performance and trust appropriateness (with average accuracy and 95\% confidence interval) in different conditions. (A) The overall team performance in five conditions. The proposed three communication strategies achieve complementary performance compared with AI accuracy (0.7) and \emph{Human Only} (0.72). (B) Humans' trust appropriateness which is indicated by human-AI agreement when AI gives correct recommendations and when AI gives wrong recommendations.}
\label{fig:team_performance}
\end{figure*}
\textbf{Part 2}
\textbf{Finding 4: Team performance was better when humans' CL was higher.}
In the proposed three CL exploitation conditions, we show participants different information or change the decision-making workflow based on human-AI CL. Therefore, we want to analyze how the team performance differs in different CL situations and different AI recommendation correctness. Overall, as shown in Figure \ref{fig:team_performance_CL} (a), when the human CL was higher, the team performance was significantly better than when the AI's CL was higher ($p$<.001 in all conditions).
Specifically, as shown in Figure \ref{fig:team_performance_CL} (b), (1) when AI's CL was higher \& AI's recommendation was correct, there was no significant difference in team performance between the three conditions. (2) When AI's CL was higher \& AI's recommendation was wrong, there was no significant difference in team performance between the three conditions. (3) When human's CL was higher \& AI's recommendation was correct, compared with \emph{Adaptive Recommendation}, the team performance in \emph{Direct Display} and \emph{Adaptive Workflow} was significantly higher ($p$<.05, $p$<.01 respectively). This might be because in \emph{Adaptive Recommendation} condition, when the human's CL was higher, the AI's suggestions were not shown, thus participants could not get the help of the AI's correct suggestions. (4) When human's CL was higher \& AI's recommendation was wrong, the team performance in \emph{Direct Display} was marginally significantly lower than \emph{Adaptive Recommendation} ($p$<.1) and significantly lower than \emph{Adaptive Recommendation} ($p$<.05).
Furthermore, as expected, we found that when the AI's recommendation was wrong, the team performance when the human's CL was higher was significantly better than when the AI's CL was higher ($p$<.001 in all conditions).
But to our surprise, when the AI's recommendation was correct, the team performance when the human's CL was higher was significantly better than when the AI's CL was higher, \emph{Direct Display} ($p$<.05); \emph{Adaptive Workflow} ($p$<.01); \emph{Adaptive Recommendation} ($p$=.058, marginally).
The possible reason was that when AI's CL was higher, since human's capability was not as good as AI's, people sometimes listened to their own wrong judgments and thus \emph{under-trust}ed AI.
\begin{figure*}[h]
\centering
\subfigure[]{
\begin{minipage}[t]{0.46\linewidth}
\centering
\includegraphics[width=\linewidth]{performan_human_AI_CL_2_NEW.pdf}
\end{minipage}%
}%
\subfigure[]{
\begin{minipage}[t]{0.54\linewidth}
\centering
\includegraphics[width=\linewidth]{performan_CL_correctness_NEW.pdf}
\end{minipage}%
}%
\centering
\caption{Team performance (with average accuracy and 95\% confidence interval) in different CL situations. (A) Team performance in different human-AI CL situations under three conditions. (B) We combine the correctness of AI suggestions with human-AI CL to analyze the team performance of the three conditions in different situations in detail.}
\label{fig:team_performance_CL}
\end{figure*}
\textbf{Finding 5: The \emph{Direct Display} and \emph{Adaptive Workflow} conditions worked better when the AI's confidence level ``contradicted'' the correctness of the AI's recommendation.}
There is an insufficiency of only utilizing AI confidence to calibrate humans' trust. Specifically, even if the AI's confidence is higher than a threshold (we used 0.7 \cite{zhang2020effect, wang2021explanations}), AI may still output wrong predictions (denoted as \emph{High \& Wrong region}). Sometimes even if AI's confidence is low, AI can give correct recommendations (denoted as \emph{Low \& Correct region}). We name these two situations as \emph{Conflict region}. We argue that just showing AI's confidence is insufficient for people to recognize these situations.
In general, as shown in Figure \ref{fig:four_situation} (a), in \emph{Conflict region}, the team performance in \emph{AI Confidence} was significantly lower than \emph{Direct Display} ($p$=.002), \emph{Adaptive Workflow} ($p$=.006). No significant difference was found between \emph{AI Confidence} and \emph{Adaptive Recommendation}.
Specifically, as shown in Figure \ref{fig:four_situation} (b), in the \emph{Low \& Correct region}, there was no significant difference between \emph{Direct Display}, \emph{Adaptive Workflow} and \emph{AI Confidence}. The possible reason may be that the room for improvement is limited (already exceeds 90\%). We noted that \emph{Adaptive Recommendation} was significantly lower than \emph{AI Confidence} ($p$=.032) probably because, in \emph{Adaptive Recommendation} condition, people developed a mode of independent thinking without relying on AI advice because they often could not see AI advice, which might lead to \emph{under-trust}. In the \emph{High \& Wrong region}, team performance in \emph{AI Confidence} was significantly lower than \emph{Direct Display} ($p$=.005), \emph{Adaptive Workflow} ($p$=.022) and marginally lower than \emph{Adaptive Recommendation} ($p$=.056).
\begin{figure*}[h]
\centering
\subfigure[]{
\begin{minipage}[t]{0.29\linewidth}
\centering
\includegraphics[width=\linewidth]{performan_strong_weak.pdf}
\end{minipage}%
}%
\subfigure[]{
\begin{minipage}[t]{0.71\linewidth}
\centering
\includegraphics[width=\linewidth]{performan_4_situation.pdf}
\end{minipage}%
}%
\centering
\caption{(A) Team performance (with the mean accuracy and 95\% confidence interval) when the AI's confidence level is in conflict (denoted as \emph{Conflict region}) and consistent (denoted as \emph{Consistent region}) with the correctness of the recommendation given by the AI. (B) Specifically, we divide the \emph{Conflict region} into (1) \emph{Low \& Correct} (AI's confidence is below the threshold but the recommendation is correct) and (2) \emph{High \& Wrong} (AI's confidence is above the threshold but the recommendation is wrong), and we divide the \emph{Consistent region} into (1) \emph{Low \& Wrong} (AI's confidence is below the threshold and the recommendation is wrong) and (2) \emph{High \& Correct} (AI's confidence is above the threshold and the recommendation is correct).}
\label{fig:four_situation}
\end{figure*}
\textbf{Finding 6: The \emph{Adaptive Workflow} and \emph{Adaptive Recommendation} conditions worked better when the AI's confidence level was ``consistent'' with the correctness of the AI's recommendation.} Another category of task instance is called \emph{Consistent region}, which includes (1) \emph{Low \& Wrong} (when the AI's confidence is below the threshold and the recommendation given is wrong), and (2) \emph{High \& Correct} (when the AI's confidence is above the threshold and the recommendation given is correct).
In general, as shown in Figure \ref{fig:four_situation} (a), the team performance in \emph{AI Confidence} was marginally significantly lower than \emph{Adaptive Workflow} ($p$=.074) and significantly lower than \emph{Adaptive Recommendation} ($p$<.001). But no significant difference can be found between \emph{AI Confidence} and \emph{Direct Display}. We then dig deeper into the two sub-regions. In the \emph{Low \& Wrong region}, the team performance in \emph{AI Confidence} was marginally significantly lower than \emph{Adaptive Workflow} ($p$=.088) and significantly lower than \emph{Adaptive Recommendation} ($p$<.001). But no significant differences were observed between \emph{AI Confidence} and \emph{Direct Display}. In the \emph{High \& Correct region}, there was no significant difference between the \emph{AI Confidence} baseline and any of the proposed three CL exploitation conditions.
In addition to the above results, we also found that (1) the proposed three conditions effectively conveyed the estimated CL information to humans, and (2) participants performed better on \emph{Consistent CL} task instances (where their perceived human-AI CL was consistent with the system's communicated human-AI CL). Detailed results can be found in the supplementary material.
\textbf{Summary.} Overall, our proposed three CL exploitation methods promoted humans' appropriate trust in AI (especially reducing humans' over-trust and without causing under-trust) and thus led to better team performance. Also, the proposed three CL exploitation strategies could effectively communicate the system's estimated human-AI CL to humans. In addition, our methods outperformed the \emph{AI Confidence} baseline when the AI's confidence contradicted its correctness. We also notice some pitfalls in our designs and will discuss them in the later section.
\subsection{Effects of CL Exploitation Strategies on Human Perceptions and Experiences.}
To answer RQ3, we analyze participants' subjective perceptions in different conditions in the exit survey, with a 7-point Likert scale (1: Strongly disagree, 7: Strongly agree). Figure \ref{fig:subjective} shows the results.
\textbf{Perceived complexity of the system.} Overall, participants' perceived system complexity is relatively low to neutral in the four conditions. Kruskal-Wallis test reveals significant differences among different conditions ($\chi^2$=19.223, $p$<.001). Post-hoc analysis shows that compared with \emph{AI Confidence}, participants found the system significantly more complex in \emph{Direct Display} ($p$<.001) and \emph{Adaptive Workflow} ($p$<.01), perhaps because the two conditions display more information and require more complex workflow.
\textbf{Mental demand.} Overall, participants were neutral about whether the decision-making process was mentally demanding in the four conditions. There are no statistically significant differences among different conditions. However, we observe a trend that \emph{Adaptive Workflow} leads to high mental demand for participants.
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{subjective.pdf}
\caption{Participants' subjective ratings in the exit survey (with the mean values and 95\% confidence interval).}
\label{fig:subjective}
\end{figure*}
\renewcommand{\arraystretch}{1.2}
\begin{table*}[tp]
\centering
\fontsize{8}{8}\selectfont
\caption{Participants' qualitative feedback in the open-ended questions. (Note that we excluded answers that just gave positive feedback but not specific, such as ``helpful'', ``like it'', so the sum of the participants may not reach 100\%.)}\label{table:openend}
\begin{tabular}{p{2.5cm}p{10cm}p{1.5cm}}
\toprule
Themes&Definitions and Examples&\#Participants\\
\midrule
\multicolumn{3}{l}{\textbf{How did participants perceive the estimated human-AI CL (in \emph{Direct Display})?}}\\
\emph{Doubt the CL}&Some participants did not believe their abilities could be easily and accurately estimated.&14 (23\%)\\
&``\emph{It was just a guess. The AI did not actually know about me so it did not seem reliable.}'' (P41)&\\
\cline{2-3}
\emph{Feeling of confirmation}&If the human and AI had the same views and the human's CL was high, it made participants more confident.&8 (13.5\%)\\
&``\emph{It made me feel very confident on those that we all agreed on}'' (P3)&\\
\cline{2-3}
\emph{Realize AI's flaws}&When people saw AI's lower CL, they realized that AI was not always trustworthy.&6 (10\%)\\
&``\emph{It made me recognize the AI could also be flawed on this task. I was better than the AI!}''(P5)&\\
\cline{2-3}
\emph{Decrease confidence}&Sometimes people became less confident when the displayed CL of them was low.&5 (8\%)\\
&``\emph{I began to question my capability somewhat when I viewed my estimated capability that was displayed low in some cases.}'' (P14)&\\
\midrule
\multicolumn{3}{l}{\textbf{How did participants use the displayed human-AI CL to make a decision (in \emph{Direct Display})?}}\\
\emph{Rely on the higher one}&They first looked at the two CL values. If AI's was higher, they followed AI's recommendation. Otherwise, they would make their own decisions.&19 (32\%)\\
&``\emph{The two charts (displaying human-AI CL) showed me when I should go with my own gut instincts and when I should rely on the AI instead.}''(P12)&\\
\cline{2-3}
\emph{Just ignore the CL}&Some participants only believe in themselves and completely ignore the CL. &15 (25\%)\\
&``\emph{I trusted my capability more, and did not put much stock in the displayed values}''(P23)&\\
\cline{2-3}
\emph{Reflect upon it}&Some participants reflected their decisions after seeing the estimated CL.&10 (17\%)\\
&``\emph{It helped me reflect on my decision once I saw my score was not as high as I thought.}''(P12)&\\
\cline{2-3}
\emph{Refer to it in inconsistent cases}&Some participants first made their own decisions. If AI agreed with them, they would ignore the CL. Otherwise, they mainly listened to the party with the higher CL.&9 (15\%)\\
&``\emph{I directly made the decision if I saw I and AI were the same. If the AI disagreed with me, I compared our abilities and chose the higher one to follow.}''(P49)&\\
\cline{2-3}
\emph{Refer to it on uncertain cases}&In cases where people were not confident, they would refer to the CL. And when they are confident, they would ignore it.&6 (10\%)\\
&``\emph{It did help me make some decisions where I was unsure. However, I would not consider it when I felt I was totally correct.}''(P39)&\\
\midrule
\multicolumn{3}{l}{\textbf{How were participants' decision processes influenced by the adaptive workflow?}}\\
\emph{Devote more cognitive resources}&When people were asked to make decisions first, they devoted more cognitive resources to the task itself, avoiding being influenced by AI's judgment.&30 (53\%)\\
&``\emph{I paid closer attention to what I thought was correct if the AI didn't make a recommendation first. If it did, then I more or less yielded to the AI's judgment.}''(P42)&\\
\cline{2-3}
\emph{Little influence}&Some participants regarded AI as a double-check or second opinion.&26 (46\%)\\
&``\emph{It did not affect me much. I always treated the AI as a second opinion, no matter whether the AI allowed me to decide first or not.}''(P54)&\\
\midrule
\multicolumn{3}{l}{\textbf{How were participants' decision processes influenced by the adaptive AI recommendation?}}\\
\emph{Independent thinking}&Not showing AI's concrete recommendation required people to think independently.&35 (62\%)\\
&``\emph{When I looked at the AI's recommendation, it was harder to trust myself and instead I found myself defaulting to the AI.}''(P33)&\\
\cline{2-3}
\emph{Little influence}&It did not affect their decision because they often relied on themselves.&19 (33\%)\\
&``\emph{It didn't affect my decision-making. I was still thinking about each information card that is presented.}''(P44)&\\
\bottomrule
\end{tabular}
\end{table*}
\textbf{Perceived helpfulness to decide when to trust the AI.} Overall, participants thought the estimated human-AI CL was helpful to make a trust choice. Kruskal-Wallis test reveals significant difference among different conditions ($\chi^2$=7.039, $p$<.05). Post-hoc analysis shows that participants in \emph{Adaptive Workflow} found the estimated human-AI CL marginally more helpful than \emph{Direct Display} ($p$=.056) and \emph{Adaptive Recommendation} ($p$=.073).
\textbf{Acceptance of estimation of their CL.} Overall, participants thought the estimation of their CL is acceptable. Kruskal-Wallis test reveals significant difference among different conditions ($\chi^2$=9.162, $p$<.01). Post-hoc analysis shows that compared with \emph{Direct Display} condition, participants in \emph{Adaptive Workflow} thought the estimation of their CL significantly more acceptable ($p$<.01).
However, in terms of \textbf{Trust in AI}, \textbf{Confidence in the decision-making process}, \textbf{Perceived autonomy}, \textbf{Satisfaction}, \textbf{Future use}, \textbf{Trust in the estimation of human-AI CL}, and \textbf{Perceived usefulness of estimation of human-AI CL}, there is no significant difference among different conditions.
In summary, except for the perceived system complexity, our proposed three conditions did not cause significantly different perceptions of participants on other aspects. Within our proposed three CL exploitation methods, \emph{Adaptive Workflow} seems to earn people's higher perceptions regarding its helpfulness and acceptance. This also echoes the result that \emph{Adaptive Workflow} achieved the highest team performance among all conditions. However, we also note that the benefits come with increased complexity and mental demand. Therefore, future work is suggested to explore a trade-off between effectiveness and user experience with human-involved empirical studies.
\subsection{Qualitative Analysis on How Participants Perceived, Used, and Were Affected by the Human-AI CL.}
To better understand the effects of different CL exploitation conditions, in the exit survey, we leave open-ended questions asking how participants perceived and utilized the human-AI CL (in \emph{Direct Display}) and were influenced by the adaptive decision-making process (in \emph{Adaptive Workflow} and \emph{Adaptive Recommendation}). Following an inductive thematic analysis process \cite{hsieh2005three} (Sec. \ref{analysismethod}), two authors created a codebook (Table \ref{table:openend}). We highlight the following findings which explain the quantitative results afore-mentioned.
\textbf{Some participants doubted the displayed CL information and ignored it in the decision-making process.}
In \emph{Direct Display}, 23\% of participants doubted the displayed CL, and 25\% ignored the CL in the decision-making process.
This also echoes the results that even though the AI's CL was higher, the human-AI agreement still did not reach 100\% (Figure \ref{fig:agreement}), and it might explain why when AI' CL was higher and AI's recommendation was correct, the team performance still did not reach 100\% (Figure \ref{fig:team_performance_CL} (b)).
\textbf{Most participants referred to or were influenced by the displayed CL or CL-based adaptation.}
In \emph{Direct Display} 74\% of the participants referred to CL, and in \emph{Adaptive Workflow} and \emph{Adaptive Recommendation}, 53\% and 62\% of participants were affected by the adaptive process respectively. This is consistent with the results that the three proposed methods could effectively affect people's agreement with AI (Figure \ref{fig:agreement}), and promote humans' appropriate trust (Figure \ref{fig:team_performance}).
\textbf{Participants in \emph{Adaptive Workflow} and \emph{Adaptive Recommendation} were forced to think independently.}
In \emph{Adaptive Workflow} and \emph{Adaptive Recommendation}, most participants were influenced by the adaptive process to think independently when human CL is higher. This supports the reason why our methods helped reduce over-trust when AI was wrong (Figure \ref{fig:team_performance} (b)). In particular, 62\% of participants in \emph{Adaptive Recommendation} would think independently when they could not see an AI recommendation, and 33\% said they always thought on their own. This reflects that participants formed a pattern of not relying on the AI, so the human-AI agreement is lowest when the AI's CL is higher (Figure \ref{fig:agreement}). This also explains why team performance is the lowest in \emph{Adaptive Recommendation} in the \emph{Low \& Correct} region (Figure \ref{fig:four_situation}). For other findings please refer to our codebook (Table \ref{table:openend}).
\section{Discussion}
Through investigating the three research questions, our study shows the promise of modeling and communicating humans' CL for promoting appropriate human trust in AI-assisted decision-making. Based on our main findings, we discuss several key issues for improving decision-making with human-AI teams and the limitations of our work.
\subsection{Human Perceptions of Self-confidence and Understanding of AI's confidence}
\textbf{Maintaining proper self-confidence is critical for humans to establish appropriate trust in AI.} Evidence shows that people's confidence in themselves significantly affects whether they will take AI's advice \cite{vodrahalli2022humans, chong2022human}. However, individuals' confidence in their own capabilities may mismatch with their actual capabilities \cite{moore2020perfectly, miller2015meta, weber2004confidence, meyer2013physicians, kahneman2011thinking} for both experts and lay people, leading to overconfidence (or underconfidence) \cite{turnercalibrating, meyer2013physicians, weber2004confidence, miller2015meta}.
From our results, in \emph{AI Confidence} condition, when AI offered correct recommendations (also with high confidence), some participants, however, still followed their wrong judgments. This is because it is difficult for humans to maintain a ``\emph{calibrated}'' self-confidence \cite{moore2020perfectly}, thus overlooking AI's suggestions. We believe that if people could accurately perceive their abilities (e.g., correctness likelihood) and calibrate self-confidence accordingly, the collaboration between humans and AI will be more successful.
The human CL modeling and communication method proposed in this paper is an initial step toward this goal.
We hope our work can inspire researchers to explore effective ways to adjust humans' confidence in AI and in themselves with human-centered computational modeling and interface design.
\textbf{Humans' understanding of probability affects the effectiveness of trust calibration.} Both the AI's confidence and the human's CL are numerical probabilities. However, previous works suggest that it is difficult for humans to act on numbers (e.g., confidence, accuracy) to coordinate the efforts of AI, especially with limited cognitive resources in some time-critical scenarios \cite{buccinca2020proxy, berwick1981doctors, lai2019human, slovic2006risk}. Furthermore, people, especially who are not good at applying mathematical thinking, lack the ability to easily interpret what a probability value actually means \cite{cosmides1996humans, peters2006numeracy, reyna2008numeracy}. This is possibly one reason why displaying AI's confidence score to humans is insufficient for calibrating their trust, which is revealed by both our work and existing studies \cite{zhang2020effect, rastogi2020deciding}.
Thus, it can be challenging to rely solely on people to make rational reliance choices and coordination (e.g., \emph{who makes a prediction first}). Our work proposes leaving the computational probability estimation/comparison task to the system and calibrating human trust by automatically adapting the decision-making process/interface. This can counter possible human cognitive biases \cite{bertrand2022cognitive, wang2019designing} and avoid making people directly deal with probabilities. Future work could explore two other directions. One is to design more effective algorithm-in-the-loop task coordination methods (e.g., \emph{learning to defer} \cite{madras2018predict}) while retaining a proper level of human autonomy. The other is to design interfaces to improve people's comprehension of probabilities, such as adding a simple tutorial about probability and frequency \cite{moore2017confidence}, presenting probabilities in more understandable manners \cite{lai2019human}, etc.
\subsection{Achievement of Complementary Performance beyond Trust Calibration}
\textbf{Exploiting knowledge complementarity is beneficial for team performance.} Our study found that although the team performance in the proposed three conditions exceeded \emph{AI Only} and \emph{Human Only}, the improvement was not ``remarkable'' (about 3-4\%). One of the key reasons is that the \textbf{complementary region/zone} between humans' knowledge space and that of AI's is relatively small.
It is reflected by the performance analysis in \emph{Human Only} that there are only a few instances that only one member of the human-AI team can handle correctly, making it hard to achieve substantial complementary performance just by calibrating human trust.
In comparison, in Bansal et al.'s work \cite{bansal2021does} where complementary performance is achieved, humans' independent accuracy is even higher when AI \emph{cannot} gives a correct recommendation than when AI \emph{can}. Therefore, echoing \cite{bansal2021does, zhang2020effect}, to ensure complementary performance, besides calibrating people's trust in AI, it is necessary to harness the complementarity of human and AI intelligence to achieve optimal outcome \cite{bansal2019updates, wilder2020learning}, perhaps by training an AI that can complement humans' knowledge and error regions.
\textbf{The modeling of human capability can empower more elaborate designs.}
In addition to approximating people's CL, our proposed modeling method is able to estimate people's predictions. We think this information can be valuable because it can help us project in advance whether humans will make consistent judgments with AI. We can combine this information with human-AI CL to enable more sophisticated strategies for assisting humans in making better decisions. For example, when the judgments of humans and AI are predicted to be consistent and neither of their CL is high, both of them are likely to make a wrong prediction. In such a case, AI can focus on encouraging people to think analytically rather than affirming people's decisions. For example, as suggested in \cite{bansal2021does}, AI sometimes can play the role of devil's advocate and question humans' judgment. Future work could explore more advanced approaches to leveraging humans' decision-making models and human-AI capabilities proposed in this paper.
\subsection{Design of Appropriate CL Communication Methods}
\textbf{Appropriately communicating the CL information is as important as correctly modeling it.} Because humans are the ultimate decision-makers, how people receive, perceive, and use this information in their decision process is essential to the outcome. Although the three CL exploitation mechanisms proposed in this paper improved people's appropriate trust in AI, we found that there were still some participants who held onto their misjudgments.
From the open-ended feedback, we found that some participants did not think that their ability could be easily and reliably estimated by the system, which undoubtedly hindered the potential of our method. Although we had told them the necessary information,
the underlying process was still a kind of \emph{black box} to some participants. Thus, we suggest providing a more detailed and easy-to-understand guide to introduce and explain the rationale behind the CL modeling to humans, increasing their understanding and acceptance.
\textbf{Other potential effective CL communication designs.}
Besides the proposed adaptive design, other types of information may also be adapted to facilitate the calibration of human trust. Previous works show that the availability of AI's explanations, regardless of their correctness, is likely to increase people's trust in AI \cite{wang2021explanations, bansal2021does, poursabzi2021manipulating, lai2019human}.
Hence, we may design an \emph{Adaptive Explanation} strategy to provide AI's explanation only when AI's CL is higher than humans'.
In addition, some studies found that the framing of confidence may affect people's perception of risk \cite{christopoulos2009neural}. We thus can apply a positive tone to describe the AI's CL when it is high, e.g., ``AI has a 75\% chance to make a correct prediction'', and use an uncertain tone otherwise, e.g., ``AI has a 25\% chance to make a wrong prediction'' (although equivalent to the former).
Besides, recent work highlights the dual-process of cognition when people process information in decision-making \cite{buccinca2020proxy, kahneman2011thinking, cacioppo1984elaboration, wason1974dual}, where System 1 processes stimuli in a fast and automatic manner which could lead to cognitive biases if applied inappropriately \cite{kahneman2011thinking}, whereas System 2 engages in deliberative and analytical thinking.
One general way to leverage this theory is when people need to rely more on their own judgment such as when their CL exceeds the AI's, the interface should stimulate people's System 2 thinking. Through these theoretical lenses, we can design more effective usage of CL information.
\subsection{Pitfalls of Current CL Modeling and Exploitation Methods}
\textbf{Potential side effects of the interface.}
Despite the effectiveness of our proposed designs in promoting humans' appropriate trust, we still suggest designers be cautious of their potential pitfalls. For example, in \emph{Adaptive Recommendation}, we found that the participants seemed to form a pattern of \textbf{skepticism} of AI because they often could not see AI suggestions, which might hinder their utilization of AI's assistance when AI's correct advice is shown. In addition, \emph{Adaptive Workflow} may lead to humans' confirmation bias \cite{nickerson1998confirmation}. For example, after people made an initial judgment and then found AI's ``confirmation'', they would be very sure that this was the correct answer. But in fact, sometimes people and AI make wrong judgments simultaneously. Therefore, we recommend that, in addition to grounding a design in existing cognitive theories, it is necessary to verify the potential impact and adverse effects of the design empirically.
\textbf{The drawback of human-AI CL and its ethical and accountability issues.}
There are two issues surrounding human-AI CL. First, even if the AI's confidence is calibrated and the human's CL is accurately modeled, inconsistent cases still exist: In the human-AI team, for a specific task instance, the member with higher CL makes a wrong prediction while the member with lower CL makes a correct prediction. It may lead to humans' inappropriate trust in AI. Since CL is just a \emph{probability} of being correct, such \emph{inconsistency of CL and correctness} is inevitable. Second, our estimation of human CL can be imperfect, and an AI model's confidence can sometimes be poorly-calibrated. So, using human-AI CL inappropriately may lead to negative results. For example, if we mistakenly estimate a human's CL to be lower, our method may lead the human to accept the wrong advice from AI when she/he could have made a correct decision independently. It can induce severe consequences and even become dangerous in high-risk scenarios. Therefore, for human-AI CL to play a positive role, confirming the reliability of human-AI correctness likelihood is essential before deployment. Besides, it may be beneficial to communicate the uncertainty behind the CL wherever appropriate to warn human decision-makers of the risk of such information.
Another possible way to mitigate the negative impact of the estimated CL is to avoid conveying a sense of confirmed, precise information, such as using specific percentages or judgmental words \cite{reyes1980judgmental}. Instead, researchers could communicate this information implicitly, embedding it in the decision-making process through designs similar to our proposed adaptive methods.
\subsection{On the Generalizability of Our Method and Results}
Proper caution should be used when generalizing our method and results to different task domains and subject populations. First, we choose a rule-based approach to help users understand and modify the auto-generated decision model. However, this approach may not be suitable for more complex decision tasks such as those involving text or image data. Thus, we need to design proper knowledge representation and modeling algorithms based on the specific characteristics of the task and data. For example, in a textual sentiment analysis task, users can specify keywords or example sentences to represent their decision model \cite{lai2022human}.
Second, our study was conducted on non-expert users in low-stake decision-making tasks. While it is a suitable testbed for exploring humans' trust appropriateness in AI-assisted decision-making \cite{zhang2020effect, ghai2021explainable}, we caution readers to generalize our results to other populations or other tasks. For example, it is unclear whether our results will still hold when our designed interfaces are adopted in high-stake tasks (where the decision-maker might have different cognitive routes \cite{suresh2021beyond}). And whether domain experts' capabilities can be well modeled by our method is also unknown. Nevertheless, we believe our proposed framework to calibrate humans' trust based on both sides' capabilities can be generalized to different AI-assisted decision-making scenarios where collaboration is needed. Future work can adapt our human CL modeling and communication method to other decision-making tasks with different stakeholders \cite{liao2021human}.
\subsection{Limitations and Future Work}
\label{limitation}
There are several limitations in our proposed methods and experimental setting. First, we used decision rules to approximate humans' decision-making models. However, rules only provide a general model and cannot cover all edge cases. Future solutions can consider integrating the ``\emph{behavioral testing}'' method \cite{beizer1995black} where the system can use test cases to ``check'' users' ability, just like testing a software or NLP model \cite{ribeiro2020beyond}. Second, we did not update humans' decision-making models in the last 20 tasks because we focused on studying the impact of our method on humans in the scope of this paper. We assume that in the absence of correctness feedback (e.g., no access to ground truth), the user's decision model is relatively fixed in the short term, which is reasonable for our experiments. However, in real-world decision-making, users' decision-making models can change as users interact with AI services and encounter more task instances \cite{ma2019smarteye}, so a static model is not enough. In the future, we plan to explore how to maintain a real-time updated user decision model in long-term AI-assisted decision-making. Third, we measured human trust in AI by human-AI agreement. Although it is widely used \cite{zhang2020effect, wang2021explanations,bansal2021does}, an obvious shortcoming is that, when people's final judgment is consistent with the AI's, we cannot distinguish whether it is because they listened to the AI's advice or because their own decisions are consistent with the AI's. Future studies may explore more suitable measurements.
\section{Conclusion}
Humans' appropriate trust in AI is a fundamental challenge in AI-assisted decision-making, and our work makes a contribution toward calibrating humans' trust based on the capabilities of both humans and AI. Our investigation consists of two consecutive phases. In the first phase, we explore how to model humans' capability (correctness likelihood) on a given task instance. We propose a human decision-making model approximation method with an interactive decision rule modification interface. In the second phase, we explore how to leverage human-AI capabilities to promote appropriate trust in AI-assisted decision-making. Based on theories of people's cognitive processes, we propose three CL exploitation methods and investigate their effects on humans' trust appropriateness, task performance, and user experience. Our results highlight the effectiveness of the proposed human CL modeling and exploitation method in promoting more appropriate human trust in AI compared with the traditional AI confidence-based method. With the derived practical implications based on our main findings, we hope this work to be a step towards promoting appropriate human-AI decision-making by considering the mutual capability information of both sides.
\section*{Appendices}
\end{document}
|
\section{Introduction}
\label{sec:introduction}
Light-harvesting dendrimers have drawn much interest in many fields of applications for the past few decades, such as in the design of new photovoltaic technologies or photosynthetic processes.\cite{balzani_harvesting_1995,gilat_light_1999,adronov_light-harvesting_2000,balzani_light-harvesting_2003}
Phenylacetylene dendrimers (covalent tree-like macromolecules) are among the most promising $\pi$-conjugated organic systems for harvesting and amplifying the energy of sunlight, based on simple polyphenylene ethynylenes (PPEs) building blocks connected via meta-substitution at phenylene nodes.
Such building blocks are also of interest on their own, as some of their spectroscopic properties are still challenging to understand and interpret.
One of the best known phenylacetylene dendrimer is the nano-star\cite{xu_design_1994}, for which the building blocks are PPEs with different lengths and substitutions schemes (para- or meta-substitution).
The connectivity of the building blocks within the nano-star is such that the structure is two-dimensional, with a sequential threefold branched graph (two donors on each acceptor).
Experimental and theoretical studies so far have suggested two main results.
On the one hand, the absorption spectra of the building blocks of PPE dendrimers are similar and additive, with mainly local electronic excitations (LE).\cite{kopelman_spectroscopic_1997,swallen_dendrimer_1999,swallen_correlated_2000,rana_steady-state_2001,wu_exciton_2006,palma_electronic_2010}
On the other hand, the building blocks are arranged such that a local excitation at the periphery of the nano-star is transferred to its center thanks to an intramolecular electronic excitation energy gradient from the external chromophores to the energy-releasing core.\cite{devadoss_energy_1996,shortreed_directed_1997,kleiman_ultrafast_2001,melinger_optical_2002,fernandez-alberti_nonadiabatic_2009,fernandez-alberti_shishiodoshi_2012,fernandez-alberti_non-adiabatic_2016,nelson_electronic_2017}
Among these building blocks, the 1,3-bis(phenylethynyl)benzene (\cref{fig:p2_m22}, bottom, called m22 in the following) shows interesting steady-state spectroscopic properties.
Alike other building blocks based on diphenylacetylene (\cref{fig:p2_m22}, top, called p2 in the following; also known as tolane), it has an absorption spectrum similar in its vibronic structure to the absorption spectrum of diphenylacetylene, only with a different intensity.
\begin{figure}[htp]
\centering
\includegraphics[width=0.70\linewidth]{p2_and_m22_81_87_88_axes.pdf}
\caption{Diphenylacetylene p2 and 1,3-bis(phenylethynyl)benzene m22 molecules. A scheme of the displacement associated to normal modes of vibration 81, 87 and 88 is given for m22.
The Mulliken axis convention used here is such that in the C\textsubscript{2v} symmetry point group, the normal mode of vibration 87 is of in-plane $y$-type symmetry and belongs to B\textsubscript{2}.}
\label{fig:p2_m22}
\end{figure}
This peculiarity was assigned to the additivity of local single excitations on p2 branches in the building blocks, and further rationalized with a pseudo-fragmentation scheme for m22.\cite{thompson_variable_2004,ho_diabatic_2019}
In addition, a minimum energy conical intersection (MECI) was characterized between the first two singlet electronic excited states.\cite{ho_diabatic_2019}
The nature of the coupled electronic singlet excited states was further studied, with two quasi-diabatic representations: one with a locally excited state on each of the p2 branch (right or left), the other with delocalized excited states with a well-defined symmetry: B\textsubscript{2} (right minus left) or A\textsubscript{1} (right plus left).
However, whereas the absorption spectra are almost identical for m22 and p2, the emission spectrum of m22 is completely different from the emission spectrum of p2 experimentally.
An unusual Stokes shift $\Delta \bar{\nu}\simeq \SI{2000}{\per\centi\meter}$ has been reported for the vibronic spectra of m22 separately by different authors.\cite{thompson_variable_2004,gaab_meta-conjugation_2003,chu_vibronic_2004}
Let us stress out here that we are not considering the typical textbook Stokes shift between the maxima of bell-shaped absorption and emission spectra, essentially due to the influence of a significant displacement between ground- and excited-state equilibrium geometries on vertical absorption and emission transition energies (hence on maximal Franck-Condon factors $0' \rightarrow v''>0$ \emph{vs.} $0'' \rightarrow v'>0$), but rather focusing on the relative spectral shift of the two band origins ($0' \leftrightarrow 0''$), which occur to coincide indeed and be the most intense for p2 but are strangely shifted in the triangular-shaped emission of m22 and not in its triangular-shaped absorption spectrum at low temperature.
A first hypothesis was suggested by some authors\cite{chu_vibronic_2004} of a missing band origin associated to the $0-0$ transition in both the absorption and emission spectra.
Such a proposition does not seem satisfactory because the $0-0$ transition is not missing in the absorption spectrum of m22, which is similar to the one of p2 and for well-justified reasons.
It thus seems more likely to investigate some particular effect influencing the emission spectrum of m22 compared to that of p2.
This is precisely the objective of the present work, where we propose some potential theoretical explanations based on symmetry, yet cannot provide any definitive answer until further time-resolved experiments are carried out.
Gaab, Thompson and co-workers produced an extensive experimental study for m22-like molecules, and proposed a variable excitonic coupling model for the reproduction and interpretation of the Stokes shift in H-terminated meta-substituted phenylacetylene instead of phenyl-terminated ones\cite{thompson_variable_2004,gaab_meta-conjugation_2003}.
Their model qualitatively agrees with \emph{ab initio} calculations for the ordering of electronic states of interest in the absorption and emission phenomena for meta-substituted PPEs.
In their work, the theoretical Stokes shift is reproduced by only looking at vertical transition energies for absorption and emission.
Once again, such a picture implies some type of usual Stokes shift due to some geometrical displacement between the absorbing and emissive species and/or incomplete vibrational relaxation.
Further along this line, we could invoke some additional effect due to the relaxed or unrelaxed state of the solvent.
Yet, there seems to some more fundamental effect due to the nature of the absorbing and emissive species while considering steady-state spectra, expected to exhibit the \emph{a priori} common $0' \leftrightarrow 0''$ band origin connecting via light-matter resonance the long-lived (stationary) eigenstates of the system.
As such, we propose hereby a minimal (three-dimensional; 3D) vibronic coupling model, in the fashion of Cederbaum, Köppel, Cattarius, and co-workers\cite{cederbaum_strong_1977,koppel_multimode_1984,cattarius_all_2001} for the construction of a model vibronic Hamiltonian, able to reproduce the vibrational structure of the electronic absorption and emission spectra for m22 via nonadiabatic quantum dynamics techniques beyond the Born-Oppenheimer approximation. This has never been tried on the present system to the extent of our knowledge; we shall use it to discriminate Stokes and non-Stokes shift contributions to the emission spectrum according to light polarization. Let us also stress that such a strategy based on the spectral response of quantum dynamics has been largely tested on photoabsorption or photoelectron/photoionization spectra, but rarely on photoemission spectra, which somewhat remain uncharted territory in terms of spontaneous \emph{vs.} stimulated emission contributions.
With the vibronic coupling models used in this work, the coupling explicitly depends on the geometry of the molecule, through variations of normal mode coordinates, following the primary idea of a variable excitonic coupling proposed by Gaab and Thompson.
If one may say, note that our interpretation is more the one of a chemist point of view (vibronic coupling) than of a physicist point of view (excitonic coupling).
Indeed, our vibronic coupling models allow for the interpretation of the absorption and emission spectra in terms of vibronic transitions and vibrational overlaps.
Because the electronic excited states of interest are strongly coupled, the simulation of absorption and emission spectra is a non-trivial procedure for which quantum dynamics was used to overcome the breakdown of the Born-Oppenheimer approximation.
The low-dimensionality of the model allows us to use the MCTDH (multiconfiguration time-dependent Hartree) method for the propagation of vibrational wave packets on the coupled electronic excited states, thus paving the way to further high-dimensional simulations and/or other types of calculations including the role of the solvent and/or of soft modes in a more effective way, which are works in progress.
As usual in this context, the vibronic spectra are then obtained using the Fourier transform of the autocorrelation functions obtained via the dynamics.
For the absorption spectra, the results are in very good agreement in spite of the low-dimensionality of the models.
However, the emission spectra show at this stage two types of contributions: one that complies very well with the observed Stokes shift of the band origin, $\Delta \bar{\nu}\simeq \SI{2000}{\per\centi\meter}$, and another that does not and merely provides coincidence of both band origins.
The Stokes-shifted band is the one that corresponds with a coherent flip of light polarization consistent with the odd character of the interstate coupling and could be attributed to an emissive Herzberg-Teller contribution or even perhaps a geometric phase; to the extent of our knowledge, this is not an effect that has been reported in large molecules.
However, this may reflect some symmetry-driven selection rule concerning the conservation of orbital momentum that is rarely to be mentioned in this context.
In the first part, we recall our main insights into the electronic structure calculations on the most important points of the energy landscape of m22, and we give relevant computational details associated to electronic structure and quantum dynamics calculations for this work.
We also describe the vibronic coupling Hamiltonian models that were used in the quantum dynamics calculations hereby.
Next, we discuss the results of these models fitted on accessible \textit{ab initio} data and describe the vibronic eigenstates produced by the models and their molecular-symmetry/light-polarization peculiarities.
Finally, computed absorption and emission spectra are provided and discussed, and further compared to their experimental counterparts, before conclusions and outlooks are given so as to open to further investigations.
\section{Theoretical background and methods}
\subsection{Conical Intersections and branching-space vectors}
The lowest two excited states of the energy landscape of m22 have been studied previously. \cite{ho_diabatic_2019}
Their potential energy surfaces (PESs) share a conical intersection (CoIn) seam, which lies locally within an $(N-8)$-dimensional subspace (where $N=108$ is the number of internal coordinates for the molecule).
The local complementary subspace to the seam is denoted the branching space, where the energy difference increases linearly from zero along the two branching-space vectors.
The minimum-energy conical intersection (MECI) of the seam was characterized\cite{ho_diabatic_2019} and will be used as a central point in this work.
For high-symmetry points, the branching-space vectors can be associated to simply defined quantities: the (halved) gradient difference vector (\textbf{GD}) and the derivative coupling vector (\textbf{DC}).
In our case, the point group of the m22 molecule is C\textsubscript{2v}, and the \textbf{GD} and \textbf{DC} expand along A\textsubscript{1} normal modes (totally symmetric) and B\textsubscript{2} normal modes (in-plane, non-totally symmetric), respectively.
The usual Born-Oppenheimer approximation fully breaks down over regions such as CoIn seams, which requires considering both electronic excited states simultaneously together with their non-adiabatic couplings.
The adiabatic (eigenstate) representation of the corresponding electronic Hamiltonian is not practical because of the singularity with respect to the energy difference in the non-adiabatic couplings.
Thus, a quasi-diabatic representation should be preferred, with smoothly varying functions (with respect to the internal coordinates) for the diabatic potentials and the inter-state couplings.
In this work, such quantities are, indeed, smoothly varying functions of the internal coordinates associated to acetylenic and quinoidal normal modes of vibration of m22 (\cref{fig:p2_m22}).
\subsection{Computational details}
\subsubsection*{Quantum chemistry}
\label{sec:details_qm}
All quantum chemistry calculations were performed with the Gaussian16 package (revision A.03)\cite{g16} using DFT (for the ground state) and TD-DFT (for the excited states) at the CAM-B3LYP/6-31+G* level of theory.
Such computational choices were already validated for para-conjugated PPE\cite{adamo_calculations_2013,ho_vibronic_2017} and this same meta-substituted molecule for the study of its absorption spectrum, with good agreement to the experiment.\cite{ho_diabatic_2019}
The defect of TD-DFT for the evaluation of charge transfer (CT) contributions to the excited states is not troublesome in the excited state of interest in this work for m22 (the first two excited states are essentially locally excited (LE) states).
The MECI of interest here was taken from the work of E. K. L. Ho \emph{et al.}\cite{ho_diabatic_2019}, and has been optimized in the C\textsubscript{2v} subspace of the molecular geometries.
In the current TD-DFT implementation of the Gaussian package, analytic non-adiabatic couplings (NAC) between excited states and thus the derivative-coupling vector (DC) are not routinely available.
For this reason, the branching-space vectors are evaluated at the MECI using the energy-based and wave-function free method of Gonon \emph{et al.}, along with their lengths, through the diagonalization of the Hessian of the squared energy difference.\cite{gonon_applicability_2017}
The main contribution in \textbf{GD} comes from the quinoidal mode 81 (43\% of the GD), the main contribution in \textbf{DC} comes from the acetylenic mode 87 (73\% of the DC).
Because mode 87 mainly is the antisymmetric elongation of left and right acetylene bonds (B\textsubscript{2}), its symmetric counterpart, the acetylenic mode 88 (A\textsubscript{1}), is added to the description; it also participates in the \textbf{GD} (10\% of the GD).
Both A\textsubscript{1} modes 81 and 88 are important for the description of the region between the Franck-Condon point (FC) and the MECI. Indeed, both modes have non-negligible contributions to the gradients at the Franck-Condon (FC) point and are associated to important shifts between the FC point and the MECI ($d_{81} = -7.36 \sqrt{m_e}a_0$ and $d_{88}=6.78 \sqrt{m_e}a_0$).
Rigid scans of the PESs of S\textsubscript{0}, S\textsubscript{1}, and S\textsubscript{2} were obtained starting from the MECI and following the normal modes directions computed at the minimum of the electronic ground state.
The grid points are described in \cref{tab:grid}, and the displacement associated to the chosen normal modes of vibration are schematized in \cref{fig:p2_m22}.
\begin{table*}[!hbt]
\caption{Frequency calculation results, symmetry, and generated points along modes 81, 87, and 88 originated from the MECI.}
\label{tab:grid}
\centering
\begin{tabular}{rllll}
\toprule
Mode S\textsubscript{0} (S\textsubscript{1}) & Frequency (cm\textsuperscript{-1}) & Reduced mass (AMU) & Symmetry & Grid ($\sqrt{m_e}a_0$)\\
\midrule
81 & 1656 & 5.465 & A\textsubscript{1} & 15 points [-18,18]\\
87 & 2367 (2184) & 11.998 (11.979) & B\textsubscript{2} & 15 points [-18,18]\\
88 & 2367 (2353) & 11.998 (11.997) & A\textsubscript{1} & 15 points [-18,18]\\
\bottomrule
\end{tabular}
\end{table*}
B\textsubscript{2} displacements break C\textsubscript{2v} geometries to C\textsubscript{s} geometries (exploring the minima of the S\textsubscript{1} PES) whereas A\textsubscript{1} displacements conserve the C\textsubscript{2v} symmetry of the molecule allowing for an assignation of the diabatic PES using unambiguous symmetry arguments.
\subsubsection*{Quantum dynamics}
\label{sec:details_qd}
The Quantics package was used for quantum dynamics calculations\cite{beck_multiconfiguration_2000,worth_quantics_2015}, mainly for the multiconfiguration time-dependent Hartree (MCTDH) wave packet propagation method.
Details concerning the single-particule functions (SPFs) and the primitive basis set are given in \cref{tab:quantics_parameters}.
The multi-set formulation of the MCTDH ansatz was used, with different sets of SPFs (8 per state).
The primitive basis set consists in harmonic oscillator wave functions for the three degrees of freedom.
Two types of calculations were done: relaxations and propagations.
Relaxations are used for computing vibronic excited eigenstates within our three-mode two-state model.
As a result, vibronic states expand in a basis of two electronic wave functions, which in most of this work are delocalized diabatic electronic wave functions.
Propagations are used for computing autocorrelation functions and associated spectra (for absorption and emission).
The autocorrelation function is computed as:
\begin{equation}
C(t)=\Braket{\Psi(0)|\Psi(t)}=\Braket{\Psi(t/2)^*|\Psi(t/2)}
\end{equation}
and the intensity of the spectra are:
\begin{equation}
I(\omega)\propto \int_{-\infty}^{+\infty}C(t)\exp{(i\omega t)}\mathrm{d}t.
\end{equation}
Because the propagation time is finite, the autocorrelation function was multiplied as usual with a decaying function
\begin{equation}
g(t)=\cos^n(\pi t/2T),
\end{equation}
where $T$ is the propagation time and $n=1$ in our calculations.
Additionally, the autocorrelation function was multiplied by a Gaussian function characterized by the damping time $\tau = (2\sqrt{2\log{2}})/(\Delta \omega)$ to simulate the experimental full width at half-maximum (FWHF), $\Delta \omega$.
\begin{table*}[!hbt]
\caption{Single-particle functions and primitive basis parameters for the relaxation and propagation of nuclear wave packets. Reduced masses for the HO primitive basis are 1 because the coordinates are mass-weighted.}
\label{tab:quantics_parameters}
\centering
\begin{tabular}{llllll}
\toprule
Coordinates & SPFs per state & & & & \\
\midrule
Q\textsubscript{87} & multi-set 8,8,8 & & & & \\
Q\textsubscript{81} & multi-set 8,8,8 & & & & \\
Q\textsubscript{88} & multi-set 8,8,8 & & & & \\
\midrule
Coordinates & Basis Type & Size of the Basis & Eq. Position & Frequency & Reduced Mass\\
\midrule
Q\textsubscript{87} & HO & 15 & 0.0 & 2365 cm\textsuperscript{-1} & 1.0\\
Q\textsubscript{81} & HO & 15 & 0.0 & 1655 cm\textsuperscript{-1} & 1.0\\
Q\textsubscript{88} & HO & 15 & 0.0 & 2365 cm\textsuperscript{-1} & 1.0\\
\bottomrule
\end{tabular}
\end{table*}
For absorption spectra calculations (ground state to excited states) and for vibronic excited states relaxation, the initial wave packet simply consists in the zero-th order harmonic oscillator wave function with corresponding frequencies for the three coordinates 81, 87, and 88.
For emission spectra calculations (excited states to ground state), the initial wave functions are the vibronic excited states obtained from the relaxations in the electronic excited states.
The different contributions to the emission spectra are discussed in the results section.
\subsection{The Quadratic and Linear vibronic coupling Hamiltonian models (QVC, LVC)}
As stated above, three modes are chosen for modelling the potential energy surfaces.
They have a specific symmetry, which makes more natural the definition of a diabatic vibronic coupling Hamiltonian model.\cite{cederbaum_strong_1977,koppel_multimode_1984,cattarius_all_2001}
Two levels of description are considered here: a quadratic vibronic coupling (QVC) model, relying on a second-order expansion of the diabatic potential energy surfaces, or a linear vibronic coupling (LVC) model, tuning and coupling the two diabatic potential energy surfaces through linear terms only.
In the following, electronic state 1 (notation superscript (1)) corresponds to the singlet excited A\textsubscript{1} state and electronic state 2 (notation superscript (2)) corresponds to the singlet excited B\textsubscript{2} state.
The diabatic representation of the Hamiltonian in the QVC model for the three modes of the m22 molecule reads:
\begin{widetext}
\begin{equation}
\label{eq:general}
\begin{aligned}
\hat{H}^{\text{dia}}&=
\left(
\hat{T}_{\text{nu}}(\boldsymbol{Q})
\right)\mathbb{1}_2
+
\begin{bmatrix}
E^{(1)}(\boldsymbol{Q}=0) & 0 \\
0 & E^{(2)}(\boldsymbol{Q}=0)
\end{bmatrix}\\
&+
\frac{1}{2}
\begin{bmatrix}
k_{81}^{(1)}Q_{81}^2 + k_{88}^{(1)}Q_{88}^2 + k_{87}^{(1)}Q_{87}^2 & 0 \\
0 & k_{81}^{(2)}Q_{81}^2 + k_{88}^{(2)}Q_{88}^2 + k_{87}^{(2)}Q_{87}^2 &
\end{bmatrix}\\
&+
\begin{bmatrix}
\kappa_{81}^{(1)}Q_{81} + \kappa_{88}^{(1)}Q_{88} & 0 \\
0 & \kappa_{81}^{(2)}Q_{81} + \kappa_{88}^{(2)}Q_{88}
\end{bmatrix}\\
&+
\begin{bmatrix}
0 & \lambda_{87} Q_{87} \\
\lambda_{87} Q_{87} & 0
\end{bmatrix}
+
\begin{bmatrix}
\gamma_{81,88}^{(1)}Q_{81}Q_{88}& 0 \\
0 & \gamma_{81,88}^{(2)}Q_{81}Q_{88}
\end{bmatrix}\\
& +
\begin{bmatrix}
0 & \mu_{87,81}Q_{87}Q_{81} + \mu_{87,88}Q_{87}Q_{88} \\
\mu_{87,81}Q_{87}Q_{81} + \mu_{87,88}Q_{87}Q_{88} & 0 \\
\end{bmatrix}.
\end{aligned}
\end{equation}
\end{widetext}
In this expression:
\begin{enumerate}
\item the first matrix represents the Kinetic Energy Operator (KEO), identical for each electronic state;
\item the second matrix represents the adiabatic energies (vertical excitation energies) at the Franck-Condon point (which is the reference point for the expansion);
\item the third matrix represents the distortion matrix in the framework of the Franck-Condon principle and of the harmonic approximation for vibronic transitions;
\item the fourth matrix represents the first-order description of the MECI with respect to the tuning modes;
\item the fifth matrix represents the first-order description of the MECI with respect to the coupling mode;
\item the sixth matrix represents the second-order description of the MECI with respect to the tuning modes, corresponding to the \emph{primary Duschinsky} transformation or mode mixing between modes of same symmetry;
\item the seventh matrix represents the second-order description of the MECI with respect to the coupling mode, corresponding to the \emph{secondary Duschinsky} transformation or mode mixing between modes of different symmetry.
\end{enumerate}
In this context, we call \emph{primary Duschinsky} parameters the second-order coupling parameters between normal modes of identical symmetry (81 with 88) and \emph{secondary Duschinsky} parameters the second order coupling parameters between normal modes of different symmetry (87 with 81 or 88).
The \textit{ab initio} Duschinsky coefficients obtained for the correlation between the first electronic excited state and the electronic ground state are given in the \cref{sec:duschinsky}, along with the shift vector.
Such a full-QVC 3D model should grasp most of the physics of the problem, with a total of 17 parameters.
However, as will be pointed out in the results section, further simplifications can be made without significant loss of information.
Hence, in the parent LVC model, all bilinear terms are neglected, such that the Hamiltonian reads:
\begin{widetext}
\begin{equation}
\label{eq:general_2}
\begin{aligned}
\hat{H}^{\text{dia}}&=
\left(
\hat{T}_{\text{nu}}(\boldsymbol{Q})
\right)\mathbb{1}_2
+
\begin{bmatrix}
E^{(1)}(\boldsymbol{Q}=0) & 0 \\
0 & E^{(2)}(\boldsymbol{Q}=0)
\end{bmatrix}\\
&+
\frac{1}{2}
\begin{bmatrix}
k_{81}^{(1)}Q_{81}^2 + k_{88}^{(1)}Q_{88}^2 + k_{87}^{(1)}Q_{87}^2 & 0 \\
0 & k_{81}^{(2)}Q_{81}^2 + k_{88}^{(2)}Q_{88}^2 + k_{87}^{(2)}Q_{87}^2 &
\end{bmatrix}\\
&+
\begin{bmatrix}
\kappa_{81}^{(1)}Q_{81} + \kappa_{88}^{(1)}Q_{88} & 0 \\
0 & \kappa_{81}^{(2)}Q_{81} + \kappa_{88}^{(2)}Q_{88}
\end{bmatrix}\\
&+
\begin{bmatrix}
0 & \lambda_{87} Q_{87} \\
\lambda_{87} Q_{87} & 0
\end{bmatrix}
\end{aligned}
\end{equation}
\end{widetext}
In both expansions of the model, diagonal terms are the diabatic potential energies associated to delocalized electronic states, with specified symmetry.
Thus, the electronic states \textsuperscript{1}A\textsubscript{1} and \textsuperscript{1}B\textsubscript{2} are coupled through the $\text{B\textsubscript{2}}=\text{A\textsubscript{1}}\bigotimes \text{B\textsubscript{2}}$ normal mode of vibration, 87.
The eigenvalues of the model Hamiltonian are to be compared with the adiabatic energies of the first two electronic excited states computed with \emph{ab initio} quantum chemistry calculations.
\subsection{Evaluating the parameters of the LVC and QVC models}
In the present L/QVC models, the diabatic states 1 and 2 have distinct symmetries when $Q_{87}=0$, that is in the C\textsubscript{2v} subspace of the molecular geometries.
This subspace is partly described by the profiles of 81 and 88, and energies from quantum chemistry calculation can be assigned to either the A\textsubscript{1} or the B\textsubscript{2} diabatic states (which fully identify to either adiabatic states S\textsubscript{1} or S\textsubscript{2} depending on the side of the MECI the calculation was made).
Thus, all parameters except for the one related to the mode 87 can be optimized by fitting directly the diabatic potential energies onto the symmetry-assigned potential energy surfaces from \emph{ab initio} calculations, such that $\{H_{11}(Q_{81},Q_{88}),H_{22}(Q_{81},Q_{88})\}$ are simultaneously fitted to the data $\{E_{\text{A\textsubscript{1}}},E_{\text{B\textsubscript{2}}}\}$ when $Q_{87}=0$. The fitting procedure uses the minimization of the squared-residual functions defined as:
\begin{equation}
\label{eq:first_fit}
\begin{aligned}
&L(E^{(1)},E^{(2)},\boldsymbol{\kappa},\boldsymbol{\gamma},\boldsymbol{k}_{81},\boldsymbol{k}_{88})=\\
&\sum_{n, Q_{87}(n)=0}\left(H_{11}[E^{(1)},\boldsymbol{\kappa}^{(1)},\boldsymbol{\gamma}^{(1)}](\boldsymbol{Q}(n))-E_{\text{A\textsubscript{1}}}(n)\right)^2\\
+&\sum_{n, Q_{87}(n)=0}\left(H_{22}[E^{(2)},\boldsymbol{\kappa}^{(2)},\boldsymbol{\gamma}^{(2)}](\boldsymbol{Q}(n))-E_{\text{B\textsubscript{2}}}(n)\right)^2.
\end{aligned}
\end{equation}
Then, taking the resulting parameters as granted, the eigenvalues of the full LVC and QVC models are fitted to the data from maps (87,81) and (87,88) so as to optimize the parameters related to mode 87.
Again, the parameters are obtained through a least-square fitting procedure minimizing the function:
\begin{equation}
\label{eq:second_fit}
\begin{aligned}
&L(\lambda_{87},\boldsymbol{\mu},\boldsymbol{k}_{87})=\\
&\sum_{n}\left(V_{1}[\lambda_{87},\boldsymbol{\mu},\boldsymbol{k}_{87}](\boldsymbol{Q}(n))-E_{\text{S\textsubscript{1}}}(n)\right)^2\\
+ &\sum_{n}\left(V_{2}[\lambda_{87},\boldsymbol{\mu},\boldsymbol{k}_{87}](\boldsymbol{Q}(n))-E_{\text{S\textsubscript{2}}}(n)\right)^2,
\end{aligned}
\end{equation}
where $V_1(\boldsymbol{Q})$ and $V_2(\boldsymbol{Q})$ are the eigenvalues of the model Hamiltonians.
For the LVC model, only separate profiles along 81, 87, and 88 are required, whereas the QVC model requires 2D-surfaces (87,81), (87,88), and (81,88) to optimize the second-order parameters.
Both yield parameters that allow the reconstruction of the PESs in the 3D-space but the LVC model does not account for full correlation between the coordinates, which will be discussed in the following section.
The fitting procedure with the determination of the L/QVC parameters was done using the functions of the \texttt{SciPy} library in \texttt{Python3.4}.
\section{Results and discussion}
\subsection{\emph{Ab initio} calculations and LVC parameters}
In this section we focus on the use of the LVC model only, for which the parameters are given in \cref{tab:parameters_LVC}.
Note that although the fitting procedure was carried out from data in atomic units and corresponding to mass-weighted coordinates, the parameters are given in understandable equivalent quantities, such as frequencies (actually, wavenumbers) and geometrical shifts instead of curvatures and gradients, respectively.
\begin{table}[!hbt]
\caption{LVC parameters obtained upon fitting \textit{ab initio} calculations. Parameters associated to the first-order expansion ($\lambda_{i}$ or $\kappa_{i}$) are given by the associated characteristic shift $d_{i}=\pm\frac{\lambda_{i}}{k_{i}}\text{ or }-\frac{\kappa_{i}}{k_{i}}^{(k)}$ in the mass-weighted framework for displacement along the normal modes directions.}
\label{tab:parameters_LVC}
\centering
\begin{tabular}{ccrl}
\toprule
Parameter & Equivalent & Value & Unit\\
\midrule
$E^{(1)}$ & -- & 4.405 & \si{\electronvolt}\\
$E^{(2)}$ & -- & 4.380 & \si{\electronvolt}\\
\midrule
$\kappa_{81}^{(1)}$ & ${d_{81}^{(1)}}$ & 7.900 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{81}^{(2)}$ & ${d_{81}^{(2)}}$ & 3.094 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{88}^{(1)}$ & ${d_{88}^{(1)}}$ & -7.205 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{88}^{(2)}$ & ${d_{88}^{(2)}}$ & -8.274 & $a_{0}\sqrt{m_{e}}$ \\
$\lambda_{87}$ & $d_{87}$ & 7.417 & $a_{0}\sqrt{m_{e}}$ \\
\midrule
$k_{81}^{(1)}$ & $\omega_{81}^{(1)}$& 1515 & \si{\per\centi\meter}\\
$k_{81}^{(2)}$ & $\omega_{81}^{(2)}$& 1552 & \si{\per\centi\meter}\\
$k_{88}^{(1)}$ & $\omega_{88}^{(1)}$& 2274 & \si{\per\centi\meter}\\
$k_{88}^{(2)}$ & $\omega_{88}^{(2)}$& 2281 & \si{\per\centi\meter}\\
$k_{87}^{(1)}$ & $\omega_{87}^{(1)}$& 2200 & \si{\per\centi\meter}\\
$k_{87}^{(2)}$ & $\omega_{87}^{(2)}$& 2201 & \si{\per\centi\meter}\\
\bottomrule
\end{tabular}
\end{table}
\Cref{fig:profiles_all} shows the adiabatic energies of states S\textsubscript{1} and S\textsubscript{2} from \textit{ab initio} calculations and the potential energy profiles computed with the latter optimized parameters.
The profiles for modes 81 and 88 are analogous.
Displacements along these A\textsubscript{1} normal modes lead to neither even nor odd potential energy surfaces with respect to the associated A\textsubscript{1} coordinates.
These displacements preserve the C\textsubscript{2v} geometry and thus the delocalized quasi-diabatic electronic states (based on symmetry labelling) maintain their nature.
Thus, we can assign \textit{ab initio} adiabatic energies ($E(\text{S\textsubscript{1,2}})$ symbol \textbf{+} in \cref{fig:profiles_all}, a and b) directly to the diabatic potentials ($E^{(1)}=E(\text{A\textsubscript{1}})$ and $E^{(2)}=E(\text{B\textsubscript{2}})$) of the LVC model.
\begin{figure}[!htb]
\centering
\includegraphics[height=0.7\textheight]{m22_LVC_fit_dia_abi_atCoIn_all.pdf}
\caption{Adiabatic energies from quantum chemistry calculations (symbol $\boldsymbol{+}$) and diabatic potential energies from the delocalized model (blue and red plain lines) along the 81\textsuperscript{th} (a), 88\textsuperscript{th} (b) and 87\textsuperscript{th} (c) normal modes of vibration.
Diabatic potential energies from the localized model are also given for 87 (c, purple dashed lines).
For unspecified coordinates, the values are those at the MECI.
All coordinates are mass-weighted and given in atomic units.}
\label{fig:profiles_all}
\end{figure}
Note that displacements along the normal mode 81 lift more efficiently the degeneracy than along 88, which is consistent with the fact that the gradient difference vector \textbf{GD} overlaps more with the direction associated to the normal mode 81 than 88.
This assignment is not so simple for the profile along the mode 87, where there are two relevant quasi-diabatic representations.
For mode 87, the eigenvalues of the LVC model Hamiltonian are directly fitted to the adiabatic energies from \textit{ab initio} calculations.
Thus, the symmetry-labelled quasi-diabatic representation is not necessarily the most relevant for comparison with the \textit{ab initio} calculations.
A more relevant representation is the localized representation fitting left and right wells to the S\textsubscript{1} potential energy surface, which is obtained through half-sum and difference of the symmetry-labelled quasi-diabatic states.\cite{ho_diabatic_2019}
Finally, these parameters give a three-dimensional picture in the vicinity of the MECI, with exact description of the MECI (starting point of our calculations) and accurate description of the minima in the first electronic excited state.
In \cref{fig:contours_all}, we give the LVC model PESs of the first two electronic excited states in the planes (87,81), (87,88), and (81,88).
The minima of S\textsubscript{1} are correctly reproduced, with $E=\SI{4.21}{\electronvolt}$ which is in a good agreement with the \textit{ab initio} calculations ($E=\SI{4.12}{\electronvolt}$).
The deviation from the \textit{ab initio} optimized minima of S\textsubscript{1} is reasonable considering the following: (i) the starting point for the rigid scans is a MECI and not one of the optimized minima, and (ii) the explored space is only a restricted three-dimensional subspace within all possible molecular geometries.
The positions of the minima are consistent with the shift vector, given in \cref{sec:duschinsky}, with two equivalent minima with respect to the non-totally symmetric coordinate 87.
The positions are also in good agreement for the totally symmetric coordinates 81 and 88, with the minima being in the positive region of 81 (quinoidal elongation) and negative region of 88 (acetylenic elongation).
The comparison of the critical points energy in the \textit{ab initio} PES and in the parameterized PES is summarized in \cref{tab:critical_points_pes}.
The most important deviation is found for the vertical transition energies and for the minimum of the S\textsubscript{1} state.
\begin{figure}[!hbt]
\centering
\includegraphics[width=1.0\linewidth]{m22_fit_LVC_all_contour_hspace0_wspace0.pdf}
\caption{LVC potential energy surfaces (in \si{eV}) of electronic states S\textsubscript{1} (a) and S\textsubscript{2} (b) in the planes (87,81), (87,88), and (81,88) from left to right.
For unspecified coordinates, the values are those at the MECI.
All coordinates are mass-weighted and given in atomic units, $\sqrt{m_e}a_0$.}
\label{fig:contours_all}
\end{figure}
\begin{table}[!htb]
\caption{Energies of the critical points in the \textit{ab initio} PESs (CAM-B3LYP/6-31+G\textsubscript{*} level of theory from Ho \emph{et al.}\cite{ho_diabatic_2019}) and in the parameterized LVC model PESs.
}
\label{tab:critical_points_pes}
\centering
\begin{tabular}{l|rrrrrr}
\toprule
Critical Point & MinS1 & TSB\textsubscript{2} & TSA\textsubscript{1} & $E_{S\textsubscript{1}}(\boldsymbol{0})$ & $E_{S\textsubscript{2}}(\boldsymbol{0})$ & MECI \\
\midrule
\small{CAM-B3LYP/6-31+G\textsubscript{*}} & 4.12 & 4.25 & 4.29 & 4.43 & 4.47 & 4.29 \\
\small{LVC Model} & 4.209 & 4.288 & 4.272 & 4.380 & 4.405 & 4.29 \\
\bottomrule
\end{tabular}
\end{table}
As mentioned earlier, the LVC model is limited for reproducing the correlation between the coordinates, in particular for taking into account the Duschinsky mixing of the modes between the ground state and the excited states.
The importance of reproducing this correlation with the QVC model is discussed in the next section.
\subsection{Comparison of linear and quadratic VC models}
The LVC and QVC models show almost identical values for the parameters that they share and no significant effects of the extra second-order parameters of the QVC model.
Similar figures and tables as the one provided for the LVC model earlier are given in the Supplementary Information (SI) for the QVC model, see tab. SI2 and fig. SI1.
The main difference lies in the ellipsoidal shapes of the isoenergy slices through the potential energy surfaces with respect to the direction of the A\textsubscript{1} normal mode 88.
Indeed, with the linear description only (LVC), the ellipsoids are aligned with the directions of the ground state normal modes of vibration (chosen for the generation of the grids), which is in contradiction with the Duschinsky transformation of the normal modes from the electronic ground state to excited states.
This transformation is taken into account in the quadratic description (through bilinear terms), for which the ellipsoids have different main axes, as shown in \cref{fig:axes}.
\begin{figure}[!hbt]
\hspace{-0.025\textwidth}
\centering
\includegraphics[width=1\linewidth,angle=0]{m22_fit_LVC_QVC_87_88_contour_V1_axes.pdf}
\caption{Potential surfaces of the first excited state in the LVC (left) and QVC (right) models with respect to the acetylenic coordinates 87 and 88. Central arrows correspond to the directions of the main axes of the ellipsoids.
All coordinates are mass-weighted and given in atomic units, $\sqrt{m_e}a_0$.}
\label{fig:axes}
\end{figure}
However, there is no significant effect of this transformation on the physics of the (dimensionally-reduced) system and in particular for the absorption and emission spectra.
This can be explained by the almost equivalent frequencies for modes 87 and 88, which make the rotation of the potential wells chemically motivated (pseudofragmentation, consistent with left or right local modes on each tolane moiety) but physically (or, say, numerically) not significant.
\subsection{Vibronic states in the excited-state landscape}
To calculate and interpret the details of vibronic transitions within UV-visible absorption and emission spectra, we must access the vibronic eigenstates of the molecule and their eigenenergies.
In the electronic ground state, the vibronic states are simple because the Born-Oppenheimer approximation is essentially valid there.
We describe those states as simple products of the electronic ground state wave function and the 3D harmonic oscillator wave functions, for which the parameters are the frequencies and reduced masses of the normal modes of vibration 81, 87, and 88 in the electronic ground state.
For the excited states, the description of the vibronic eigenstates is more involved.
We are in fact facing a situation whereby the Born-Oppenheimer approximation is completely off.
Due to the quasi-degeneracy of the two states of interest, already in the Franck-Condon region, and the presence of an accessible conical intersection in its close vicinity and inducing a strong diabatic mixing, the first two excited adiabatic electronic states are strongly coupled nondiabatically and cannot at all be treated separately for the description of the associated vibronic states.
In other words, and using the language of spectroscopy, we are in the unusual situation where the first-order Herzberg-Teller effect along mode 87 is clearly not a perturbation.
Within a description beyond Born-Oppenheimer the vibronic states expand as linear combinations of products of electronic and nuclear wave functions.
We choose the delocalized, symmetry-labelled, diabatic electronic basis set, as described earlier for the LVC Hamiltonian model.
Viewed as a good approximation of a crude adiabatic basis for the Herzberg-Teller effect, it presents the great advantage that transition dipoles will not vary much with displacements from C\textsubscript{2v} molecular geometries (such as the Franck-Condon ground-state equilibrium geometry); in particular, A\textsubscript{1} is polarized along $z$ and B\textsubscript{2} along $y$.
Within this representation, we thus get:
\begin{equation}
\begin{aligned}
\Psi^{\text{(exc)}}_k (\boldsymbol{q}, \boldsymbol{Q})
&=
\chi_{A_1,k} (\boldsymbol{q}) \phi_{A_1}(\boldsymbol{q};\boldsymbol{Q})\\
&+
\chi_{B_2,k} (\boldsymbol{q}) \phi_{B_2}(\boldsymbol{q};\boldsymbol{Q})
\end{aligned}
\end{equation}
where $\boldsymbol{Q}=(Q_{81},Q_{87},Q_{88})$ are the internal coordinates and $\boldsymbol{q}$ are the electronic coordinates.
The electronic states $\phi_{A_1}(\boldsymbol{q};\boldsymbol{Q})$ and $\phi_{B_2}(\boldsymbol{q};\boldsymbol{Q})$ are quasi-diabatic states, which vary moderately and smoothly (ideally not at all to first order) with the internal coordinates $\boldsymbol{Q}$.
The associated diabatic state populations are simply the integral of the vibrational contribution over the internal coordinates $\boldsymbol{Q}$.
In this work, such vibronic eigenstates are obtained through quantum dynamics calculations on the “coupled” PESs from the LVC Hamiltonian model.
The calculation consists in the propagation in imaginary time of an initial nuclear wave packet until variational convergence.
The so-called relaxation method is analogous to variational energy minimisation (parametric optimisation).
We consider two choices of initial states (they will be the same as those used later on for the absorption spectra simulations).
They are the vibrational ground state for the electronic ground singlet state (fundamental of the 3D harmonic oscillator for 81, 87, and 88) multiplied by either the diabatic state A\textsubscript{1} or the diabatic state B\textsubscript{2} (coupled electronic excited singlet states).
After an imaginary relaxation time of about \SI{20}{\femto\second}, which is short and indicates little internal vibrational redistribution, two quasi-degenerate ($\Delta E = \SI{0.0076}{\electronvolt}$) but distinct vibronic states are obtained (they could be viewed as a tunnelling pair, symmetric and antisymmetric, in an adiabatic context).
Their characteristics are collected in \cref{tab:vibronic_states}.
We notice that both vibronic states are complementary.
Vibrational contributions of the vibronic states in the planes (87,81), (87,88), and (81,88) are given (see \cref{fig:vibronic_states} for the lowest, and fig. SI3 for the second vibronic state).
The first vibronic state is mainly B\textsubscript{2}-populated (80\%) while the second one is mainly A\textsubscript{1}-populated (78\%) which is consistent with the ab initio energy and gradient information regarding the two delocalized diabatic states.
Indeed the minimum of B\textsubscript{2} is lower in energy than the minimum of A\textsubscript{1} in our 3D model, which is in good agreement with the energy ordering of the optimized transition states A\textsubscript{1} and B\textsubscript{2} (see \cref{tab:critical_points_pes}).
These results also indicate that with this model, in each vibronic state, about 80\% of the initial wave packet is preserved in the stationary vibronic states, which represents about 80\% of totally symmetrical vibrational contributions with respect to the three coordinates (see \cref{fig:vibronic_states}, (a)).
As a result, we notice about 20\% of population transfer into the second diabatic state, not-initially targeted, allowed by the coupling coordinate 87.
This 20\% transferred population corresponds to an odd shape of the vibrational contributions with respect to the coupling coordinate 87 (see \cref{fig:vibronic_states}, (b)).
We stress here that the part of the nuclear wave packet that transfers from one diabatic state to another must undergo a change in its parity or symmetry according to the odd behavior of the diabatic interstate coupling function with respect to mode 87.
This would be termed a geometric (Berry or Longuet-Higgins) phase effect if the adiabatic representation were used instead.
This change is easily identified in the delocalized and symmetry-labelled diabatic states basis set: the nuclear wave packet turns odd when leaving its initial diabatic state, due to the off-diagonal coupling being odd.
Let us notice that the first two vibronic states are distinct, but similar with only inversion of the roles of the diabatic states A\textsubscript{1} and B\textsubscript{2}.
\begin{table}[!hbt]
\caption{Vibronic states energy and diabatic populations.
The transition energy is computed by setting the energy reference to the zero-point energy of the 3D model, $E=\SI{0.396}{\electronvolt}$}
\label{tab:vibronic_states}
\centering
\begin{tabular}{lrrrr}
\toprule
Vibronic state & Total Energy (eV) & Transition Energy (eV) & $\mathcal{P}_{A_1}$ & $\mathcal{P}_{B_2}$\\
\midrule
$\Psi^{\text{(exc)}}_{0}$ & 4.576 & 4.180 & 0.20 & 0.80\\% 5759
$\Psi^{\text{(exc)}}_{1}$ & 4.584 & 4.188 & 0.78 & 0.22\\% 5835
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[!htb]
\centering
\includegraphics[width=1\linewidth]{m22_LVC_rlx_2_t50_all_wf_hspace0_wspace0.pdf}
\caption{Nuclear wave packet contributions in diabatic electronic states B\textsubscript{2} (a) and A\textsubscript{1} (b) in the first vibronic state, in planes (87,81), (87,88), and (81,88) from left to right.
For unspecified coordinates, the values are those at the MECI.
All coordinates are mass-weighted and given in atomic units, $\sqrt{m_e}a_0$.
The blank figure corresponds to the wave functions having the B\textsubscript{2} coordinate 87 in the nodal plane.
The second vibronic state is described in fig. SI3
}
\label{fig:vibronic_states}
\end{figure}
We propose to interpret again the vibronic states looking mostly at their character with respect to the non-totally symmetrical coordinate 87.
\Cref{fig:interaction_diagram} gives a schematic representation of the vibronic eigenstates in the ground and excited electronic states.
The pair of vibronic eigenstates in the electronic excited states manifold can be seen as vibronic eigenstates in diabatic states B\textsubscript{2} and A\textsubscript{1} interacting through the non-totally symmetrical coordinate 87.
This representation is analogous to an interaction diagram for electronic orbitals, only with coupled vibrational orbitals here.
The first two pairs of vibronic eigenstates are the results of four vibrational orbitals interacting and mixing their character.
As a result, the lower vibronic eigenstate has two vibrational contributions of different parity, odd and even with respect to the coordinate 87.
\begin{figure}[!hbt]
\centering
\includegraphics[width=1\linewidth]{interaction_diagram_transitions.pdf}
\caption{Representation of the vibronic eigenstates as interacting \textit{diabatic vibronic orbitals}.
The vibrational number $v$ refers to the excitation number of non-totally symmetrical normal mode of vibration 87.
The nuclear wave packets are given in the plane (87,88).}
\label{fig:interaction_diagram}
\end{figure}
For absorption, the initial wave packet in the ground electronic state overlaps with the even contribution of each of the vibronic eigenstates.
For emission, two cases arise.
The initial wave packets have even and odd vibrational contributions, that overlap with the first and second vibronic ground states, respectively.
Two contributions to the emission spectrum are thus obtained.
One is sharing the $0-0$ band origin with absorption and one shifted by the vibrational energy of the second vibrational state of the electronic ground state.
Analogous transitions occur for the second vibronic eigenstate. The transition energies to the vibronic eigenstates are ordered as the transition energies of the non-interacting diabatic vibronic eigenstates.
The latter ordering is consistent with the \textit{ab initio} potential energy minima of the electronic states in the C\textsubscript{2v} sub-space which are transition states of the first excited state in the C\textsubscript{S} space of molecular geometries.
\subsection{Absorption and emission spectra}
In this section, we describe the procedure for producing absorption and emission spectra (vibronic spectra) using quantum dynamics for our model of the m22 molecule.
The different contributions to the spectra will be discussed and compared to the low-temperature experimental spectra from Chu and Pang experiments.
Vibronic spectra are computed through Fourier transformation (FT) of the autocorrelation function associated to the propagation of the nuclear wave packet in the electronic excited states (for absorption) or in the electronic ground state (for emission).
The fact that only the autocorrelation function of the nuclear wave packet is required as a good approximation is helped by the use of the delocalized and symmetry-labelled diabatic basis set for the electronic states.
Indeed, in this basis set, the transition dipole moments have very little dependence on the nuclear coordinates (Franck-Condon-type approximation):
\begin{equation}
\begin{aligned}
\boldsymbol{\mu}_{\text{A\textsubscript{1}}}(\boldsymbol{Q}) &\simeq \boldsymbol{\mu}_{\text{A\textsubscript{1}}}(\boldsymbol{0}) =\num{-1.8303}\hat{\boldsymbol{z}} (a_0\sqrt{m_e})\\
\boldsymbol{\mu}_{\text{B\textsubscript{2}}}(\boldsymbol{Q}) &\simeq \boldsymbol{\mu}_{\text{B\textsubscript{2}}}(\boldsymbol{0}) =\num{3.9656}\hat{\boldsymbol{y}} (a_0\sqrt{m_e}).
\end{aligned}
\end{equation}
In other words, we can safely neglect Herzberg-Teller terms in the symmetry-adapted delocalized diabatic representation, which, in contrast, would be unusually large with respect to variations in the coordinate 87 in the strongly mixed adiabatic representation, as already pointed out.
As a consequence, the intensity spectra are here simply proportional to the FT of the autocorrelation function of the propagated wave function.
Here, the wave functions must be propagated in real-time to access the vibronic spectra, as opposed to the previous section where propagation was done in imaginary time to access relaxed vibronic states.
\subsubsection*{Modelling absorption}
\label{sec:modelling_absorption}
For absorption and the initial wave packet, we choose the vibrational ground state (fundamental of the 3D harmonic oscillator for 81, 87, and 88) multiplied by either the diabatic state A\textsubscript{1} or the diabatic state B\textsubscript{2} as for the vibronic states relaxations.
The choices of A\textsubscript{1} or B\textsubscript{2} has little or no influence on the shape of the absorption spectrum obtained in the end, even though they correspond to distinct polarization processes.
Before discussing the absorption spectrum, let us comment the early dynamics (first \SI{25}{\femto\second} of the \SI{200}{\femto\second} of simulation time) of the real-time propagation of the initial wave packets in the electronic excited states.
We discuss first the propagation of the A\textsubscript{1}-populated initial state.
The trajectories of the centers (coordinate expectation values) of the wave packets in the (81,88) plane for the two diabatic states are given in \cref{fig:early_dynamics} (a), where the PESs of the A\textsubscript{1} and B\textsubscript{2} states are shown, along with the positions of the conical intersection seam and the MECI.
We also report diabatic and adiabatic populations (computed using the Quantics package), \cref{fig:early_dynamics} (b).
First, let us notice that the initial state has a population of about 0.6 and 0.4 for the adiabatic states S\textsubscript{1} and S\textsubscript{2}.
This almost 50/50 population of the first two adiabatic states is not troublesome because, in the FC region, their energy difference is small ($\Delta E = \SI{0.04}{\electronvolt}$ at our level of theory) and both are optically bright (oscillator strengths $f_1= 1.71$, $f_2 = 0.37$ in atomic units).
Besides, in our 3D model, the ordering of the diabatic states B\textsubscript{2} and A\textsubscript{1} corresponding to adiabatic states S\textsubscript{1} and S\textsubscript{2} respectively is consistent with the \textit{ab initio} data.
It indicates the importance of taking into account both A\textsubscript{1} modes 81 and 88 for modelling m22 as the conical intersection seam and the FC gradients are oblique within the (81,88) plane.
Using a minimal model on the acetylenic vibrations with only modes 87 and 88 would initiate the propagation on the wrong side of the conical intersection seam, with an ordering of the diabatic states that is wrong compared to the \textit{ab initio} data.
The multiple crossings of the conical intersection seam in the early dynamics stress the importance of having a consistent ordering of the diabatic states at the FC point and thus the importance of taking into account the coordinate 81.
On a final note, the numerical kink of the trajectory in the B\textsubscript{2} state is unconsequential because of the almost-zero diabatic population of the state B\textsubscript{2} at this time of the simulation, and seems to be of no relevance for resuming a satisfactory trajectory for times with non-zero diabatic population.
Analogous interpretations of the early dynamics can be done for the other choice of initial diabatic state, with a B\textsubscript{2}-initially populated state, for which figures are given in fig. SI2.
\begin{figure}[!htb]
\centering
\includegraphics[width=1\linewidth]{early_dynamics_m22_LVC_abs_1_DRAFT.pdf}
\caption{Early dynamics: trajectories of the center of the wave packets in the (81,88)-plan.
Diabatic potential energy surfaces A\textsubscript{1} and B\textsubscript{2} in blue and red respectively, adiabatic PESs S\textsubscript{1} and S\textsubscript{2} in dashed lines and plain lines respectively (a).
Diabatic (plain lines) and adiabatic (dashed lines) populations associated to these trajectories (b).
Propagation for an initial state on A\textsubscript{1} is given here, propagation for an initial state on B\textsubscript{2} is given in fig. SI2}
\label{fig:early_dynamics}
\end{figure}
The two propagations performed here yield two absorption spectra according to the populated excited diabatic state and corresponding light polarization.
Let us notice that two representations are always given for each of our theoretical spectra.
A representation obtained with FT on the autocorrelation damped with $\tau = \SI{100}{\femto\second}$ and another one one with FT on an autocorrelation function damped with $\tau = \SI{19}{\femto\second}$.
The first representation resembles to a stick-spectrum while the second one shows realistically broadened bands.
The damping time $\tau$ was chosen so as to fulfil the experimental spectra, as described in the last sub-section.
With our model, we access two distinct but almost identical absorption spectra (\cref{fig:absorption_spectra}).
This is consistent with the results obtained for the vibronic states relaxation.
If analogous vibronic states are found from the two different initial states, then absorption spectra must also be analogous because the same eigenstates are accessible for the system.
Here we discard for now the differences in oscillator strength between the two diabatic states, but as both are bright, both simulations and both absorption spectra are relevant.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.95\linewidth]{m22_LVC-wide-19_abs_1_2-crop.pdf}
\caption{Simulated absorption spectra from electronic ground state to electronic excited state B\textsubscript{2} (plain lines) and A\textsubscript{1} (dashed lines).}
\label{fig:absorption_spectra}
\end{figure}
Notice that within the 3D LVC model, the $0-0$ band corresponds to a transition at \SI{296}{\nano\meter}.
The second most intense band is at \SI{278}{\nano\meter}, which corresponds to a difference of $\Delta \bar{\nu} = \SI{2190}{\per\centi\meter}$.
The latter energy difference is expected for the acetylenic mode 87, optically active in the absorption, for which the frequency is about \SI{2200}{\per\centi\meter} in the 3D LVC model of the electronic excited states.
The other transition at \SI{284}{\nano\meter}, corresponding to $\Delta \bar{\nu} = \SI{1430}{\per\centi\meter}$ with respect to the $0-0$ band, is the expected transition involving the quinoidal mode 81.
The absorption spectra are also reported in \cref{fig:abs_emi_spectra} for completeness of the theoretical results and in \cref{fig:exp_and_theo_spectra} for comparison with the experiments.
\subsubsection*{Modelling the emission spectrum}
The emission spectrum is simulated by propagating the initial nuclear wave packet in the electronic ground state, for which the PES is described as a 3D harmonic potential with frequency parameters for mode 81, 87, and 88 extracted from ab initio calculations at the minimum of the electronic ground state (see \cref{tab:grid}).
For the initial state, the choice is more complicated than for simulating absorption.
Indeed, as mentioned earlier, the first two vibronic states in the twofold excited electronic structure are quasi-degenerate and share similar characteristics in vibrational shape and populations.
The assumption that the emission would occur only from the lowest vibronic state should not be completely valid, but we focus first on this case as the second one is analogous.
Emitting from the lowest excited vibronic state (initially B\textsubscript{2}-populated), we identify two simple initial states for the quantum dynamics which will result in the end to two contributions to the emission spectrum.
The two initial states are: $\chi_{A_1,0}\phi_0$ and $\chi_{B_2,0}\phi_0$.
As described earlier the first vibrational part $\chi_{A_1,0}\phi_0$ is odd with respect to coordinate 87 while the second one $\chi_{B_2,0}\phi_0$ is even with respect to coordinate 87 (see \cref{fig:vibronic_states}).
For the first vibronic eigenstate, two contributions to the emission spectrum are obtained (see \cref{fig:emission_2a_2b}).
The $0-0$ band for emission is recovered from the contribution of the vibrational part on B\textsubscript{2}, at \SI{297}{\nano\meter}.
A vibronic progression (\SI{319}{\nano\meter} and \SI{345}{\nano\meter}) analogous to absorption is obtained as expected for frequencies in the acetylenic regions, with $\Delta \bar{\nu} = \SIlist{2320;2360}{\per\centi\meter}$.
Such a contribution to the emission spectrum will be called a non-Stokes contribution in the following.
The contribution of the vibrational part on A\textsubscript{1}, on the contrary, is shifted in the red wavelengths with an origin at \SI{319}{\nano\meter}, corresponding to a Stokes shift $\Delta \bar{\nu} = \SI{2320}{\per\centi\meter}$, and maintains the expected vibronic progression from this origin band to larger wavelengths. Such a contribution to the emission spectrum will be called a Stokes contribution in the following.
The positions of the first three transitions and the associated shifts with respect to the most intense band are summarized in \cref{tab:spectroscopic_data}.
\setlength{\tabcolsep}{18pt}
\begin{table*}[!hbt]
\centering
\begin{tabular}{l|l|l}
\toprule
Spectrum contribution & $\lambda_\text{max}$ (\si{\nano\meter}) & First vibronic progression (\si{\per\centi\meter}) \\
\midrule
Absorption $\rightarrow$ & \textbf{296}, (284), 278 & --, (1430), 2190\\
Emission\rotatebox[origin=c]{270}{$\Rsh$} & & \\
$\Psi_0$, non-Stokes & \textbf{297}, (312), 319 & --, (1620), 2320\\
$\Psi_0$, Stokes & \textbf{319}, (337), 345 & --, (1670), 2360\\
$\Psi_1$, non-Stokes & \textbf{296}, (311), 318 & --, (1630), 2340\\
$\Psi_1$, Stokes & \textbf{319}, (336), 344 & --, (1590), 2280\\
\bottomrule
\end{tabular}
\caption{Spectroscopic data for theoretical absorption and emission spectra, before any comparison with the experiment.
One row is given for absorption as both computed spectra are identical for reading the transition energies.
The four contributions to the emission spectrum are given next.
The most intense transition for each contribution is given in bold.
The first next transition with low intensity is given in parenthesis.
Associated vibronic progression are given with respect to the most intense transition of each contribution.
}
\label{tab:spectroscopic_data}
\end{table*}
\setlength{\tabcolsep}{6pt}
Let us notice that all these results are expected given the form of the vibrational contributions in the vibronic states.
Indeed, with a time-independent reasoning, the vibrational part $\chi_{B_2,0}$ of the vibronic state is expected to overlap with the vibrational ground state $\chi_0$ resulting in an band origin in the same energy region than for the band origin of the absorption.
However, the vibrational part $\chi_{A_1,0}$ of the vibronic state is expected to overlap with the first vibrational excited state $\chi_1$ resulting in the shifted origin band at \SI{319}{\nano\meter}.
The same interpretation holds for the second vibronic state, only interchanging the roles of A\textsubscript{1} and B\textsubscript{2} diabatic states.
\begin{figure}[!htb]
\centering
\includegraphics[width=1\linewidth]{m22_LVC-wide-19_emi_2a_2b-crop.pdf}
\caption{Simulated emission spectra from vibrational contributions of B\textsubscript{2} (plain lines) and A\textsubscript{1} (dashed lines) of the first vibronic eigenstate to the electronic ground state.
Simulated emission spectra from the second vibronic eigenstate are given in fig. SI4.
}
\label{fig:emission_2a_2b}
\end{figure}
All the transitions observed between the ground vibronic states and the excited vibronic states are schematized in the interaction diagram \cref{fig:interaction_diagram}, all the resulting contributions to the absorption and emission spectra (two for each vibronic state) are collected altogether in \cref{fig:abs_emi_spectra}.
\begin{figure}[!htb]
\centering
\includegraphics[width=1\linewidth]{m22_LVC-wide-19_abs_emi_2-crop.pdf}
\caption{Absorption spectra from the electronic ground state to both diabatic states (blue lines) and emission spectra from the different contributions of the first vibronic eigenstate (red lines).
Simulated emission spectra from the second vibronic eigenstate are given in fig. SI5.
Realistically broadened bands are obtained using a damping time $\tau = \SI{19}{\femto\second}$.}
\label{fig:abs_emi_spectra}
\end{figure}
\subsubsection*{Discussion and comparison to the experiment}
We chose to compare our theoretical results to the experiments of Chu and Pang because of the exact same substitution of the molecule in both studies and because of the temperature study\cite{chu_vibronic_2004}.
Indeed, the measurement of the spectra was done at \SIlist{25;-108;-198}{\celsius} and we choose the latter measurement for comparison.
The experimental spectra at \SI{-198}{\celsius} are reproduced in \cref{fig:exp_and_theo_spectra}, with courtesy of Chu and Pang, 2004\cite{chu_vibronic_2004}.
The theoretical spectra (considering the absorption to and emission from the first vibronic state) are superimposed to the experiment, shifting the $x$-axis of the theoretical spectra so that the band origins from theoretical absorption (\SI{297}{\nano\meter}) and experimental absorption (\SI{304}{\nano\meter}) match which represents a shift of \SI{7.3}{\nano\meter} at the $0-0$ band of absorption.
The shift applied to our theoretical spectra for the $0-0$ band of absorption to match with the experiment is thus $\Delta E = \SI{0.10}{\electronvolt}$ and $\Delta \bar{\nu} = \SI{818}{\per\centi\meter}$.
This difference between the experiment and our model can be understood as an underestimated zero-point energy difference for the minimal 3D model, because other vibrations should be included to the complete description of the system.
Besides, our model is built on data at a TD-DFT level of theory, which may imperfectly reproduce vertical transition energies.
We stress that we shift the $x$-axis of all our theoretical spectra of the same values of \SI{818}{\per\centi\meter} and thus do not introduce new or artificial Stokes shifts, between emission and absorption.
\begin{figure}[!htb]
\centering
\includegraphics[width=1\linewidth]{exp_and_theo_spectra_2a-2b_pl_LVC-wide-19_shifted2abs_energy-crop.pdf}
\caption{Experimental absorption and emission spectra (dashed lines), courtesy of Chu and Pang, 2004\cite{chu_vibronic_2004}, reproduced in (a) and (b).
Theoretical absorption spectrum to the first vibronic eigenstate reproduced in (a) and (b), theoretical emission spectrum from the first vibronic eigenstate, contributions of B\textsubscript{2} (a) and of A\textsubscript{1} (b).
Realistically broadened bands are obtained using a damping time $\tau = \SI{19}{\femto\second}$.
Simulated absorption and emission spectra from the second vibronic eigenstate are given in fig. SI6.
}
\label{fig:exp_and_theo_spectra}
\end{figure}
The broadening of the theoretical spectra was chosen so that the FWHM $\Delta \omega$ of the origin band for the absorption is the same as in the experiment.
To ensure this, we damp the autocorrelation function with the damping time $\tau = \SI{19}{\femto\second}$.
For absorption, the vibronic progression is well reproduced with our model for transitions involving acetylenic vibrations.
The missing transitions in-between (\SIlist{294;277}{\nano\meter}) should correspond to transitions involving other vibrations, not taken into account in our minimal 3D model, such as quinoidal and triangular vibrations, expected and identified in the absorption spectra of p2 and m22 in previous work\cite{ho_vibronic_2017}.
In the emission, the Stokes shift contribution is always identified as the contribution from the vibrational part of the vibronic states that underwent a transfer from one diabatic state to another.
The complementary vibrational part shows no contribution to a spectrum with a significant Stokes shift.
In the case of the Stokes contribution, the theoretical spectrum also reproduce the expected vibronic progression, with the missing bands expected to result from transitions involving other vibrations.
Strictly speaking, the spectrum to be compared to the experiment should be a combination of the four theoretical contributions described, involving the transition dipole moments for states A\textsubscript{1} and B\textsubscript{2} along with factors to take into account the population of these two states in the vibronic states.
However, searching such a combination within this model and reasoning would never result in a predominant Stokes contribution to the emission spectrum.
Indeed, the vibrational part of the vibronic states that overlaps with the $\chi_1$ ground vibrational state never correspond to more than 20\% of the diabatic population.
In addition, the oscillator strength for the B\textsubscript{2} state, in the direction of $y$, is much greater than the oscillator strength of the other state (along $z$).
As it stands, our model is still limited by several approximations on the system and preconceptions of how to simulate the phenomenon under study.
For the reasons mentioned above, it cannot explain why the Stokes contribution dominates over the non-Stokes one, but it shows that there is a rationale for the Stokes contribution that could be based on symmetry considerations and thus may provide some tentative avenues for further explorations.
\section{Conclusions and Outlooks}
In this work, we have devised diabatic linear and quadratic vibronic coupling models of Hamiltonians for the study of unusually strongly nondiabatically coupled excited states in a symmetric meta-substituted PPE (m22) and used them for spectral simulations based on quantum dynamics.
We showed that molecular symmetry may play a major role in the unusual Stokes shift observed for this system regarding its band origin; other effects are still under study, such as the role of soft modes, solvent relaxation dynamics, and perhaps the effect of possible competing photo-isomerizations.
On practical terms, our study is expected to serve two main purposes: laying the basic foundations for a more general model of nonadiabatically coupled PESs for symmetrical or asymmetrical meta- or tri-substituted PPEs, and using it for providing a possible geometric explanation to the unusual spectroscopic properties of m22 and its derivative species.
As pointed out, our modelling of m22 suggests some plausible rationale (based on molecular vibronic symmetry and light polarization transfer or conservation) for distinct different Stokes and non-Stokes contributions to the emission spectrum, with no fully conclusive elucidation yet on why the Stokes contributions should be predominant, as indeed observed in the experiments.
It must be understood that modelling emission (as opposed to absorption) spectra is not a routine task and potentially raises other fundamental questions about how we should consider the very role and nature of the stationary states involved in the experimental acquisition of absorption and emission spectra against the duration of the excitation light source and the time for the environment (external: solvent; internal: soft modes) to fully relax compared to the time it takes for an excited molecule to emit light.
Within our description, for the Stokes contribution to overcome the non-Stokes contribution, the effect of the coupling would have to be stronger, hence fully forcing the electronic population to be affected quantitatively by the non-totally symmetrical coordinate 87 and its odd character (according to the linear behavior of the coupling with respect to it), along the way of a geometric (Berry or Longuet-Higgins) phase effect such that the nuclear density goes to the other side of the conical intersection with a nodal line.
This is a possibility, but it obviously is very sensitive to the level of theory used for producing \emph{ab initio} data. However, in the extreme case of a completely transferred population to a non-totally symmetrical wave packet, the $0-0$ band in the absorption spectrum should also disappear, which thus seems to raise some contradiction.
Such hypotheses are thus to be evaluated carefully in future works, considering for example additional but relevant degrees of freedom for the vibronic transitions observed experimentally in the steady-state spectra, or dissipation effects (for example in the context of the hierarchical equations of motion that we recently explored together with our colleagues on the same system but within another context)\cite{jaouadi_laser-controlled_2022}.
More extensive simulations are currently under study and will be addressed soon in future work.
Finally, it must be noted as a last hypothesis that the non-Stokes contributions to the emission identified in this work seem to be associated to some unconventionally ultrafast coherently polarized resonant Raman process, while usual fluorescence spectra are typically viewed as incoherent, slow, and spontaneous, and they should exhibit both Stokes and non-Stokes contributions.
Such a mystery certainly now awaits for further detailed comparisons with time-resolved spectroscopic experimental studies.
\section*{Acknowledgments}
The authors warmly thank Y. Pang and Q. Chu for allowing the reproduction of the experimental absorption and emission spectra of 1,3-bis(phenylethynyl)benzene (m22).
The authors would also like to personally thank M. Desouter-Lecomte and E. Mangaud for very fruitful discussions on the system and the emission phenomenon, and their incentive for plausible extensions to further studies with complementary viewpoints.
J. G. acknowledges the French MESR (Ministère de l’Enseignement supérieur, de la Recherche) and the ENS (Ecole Normale Supérieure) of Lyon for funding his PhD grant, hosted at the University of Montpellier.
\section{Appendices}
\label{sec:appendices}
\subsection{Duschinsky matrix and shift vectors}
\label{sec:duschinsky}
Normal modes of vibration are computed at the minima of the ground electronic state and the first excited electronic state. The transformation from the normal modes of vibration at the minimum of the first electronic excited state, $\boldsymbol{Q}'$ to the normal modes of vibration at the minimum of the electronic ground state $\boldsymbol{Q}''$ is as follows:
\begin{equation}
\boldsymbol{Q}'=\boldsymbol{J}\boldsymbol{Q}'' + \boldsymbol{K}
\end{equation}
where $\boldsymbol{J}$ and $\boldsymbol{K}$ are the so-called Duschinsky matrix and shift vector, respectively.
For modes 81, 87, and 88, the Duschinsky matrix is:
\begin{equation}
\begin{bmatrix}
-0.118694 & -0.000010 & 0.000187\\
-0.050807 & 0.733369 & 0.662382\\
-0.005732 & -0.671423 & 0.740842
\end{bmatrix}
\end{equation}
and the shift vector is
\begin{equation}
\begin{bmatrix}
-4.01414\\
7.25456\\
7.59322
\end{bmatrix}.
\end{equation}
Note that the correlation between modes 81 and 87 or 88 is negligible (less than 0.3\% in any case). These small correlations suggest the validity for neglecting the primary Duschinsky paramaters, $\gamma_{81,88}^{(k)}$ in the QVC model.
The correlation between 87 and 88 correspond to a \SI{45}{\degree} angle of the ground state normal modes to obtain the excited state normal modes, which is consistent with the results found for the orientation of the ellipsoids in the plane (87,88) of the QVC model PESs.
\section*{References}
\bibliographystyle{article_title}
\section{Linear and Quadratic vibronic coupling Hamiltonian models}
\subsection{Parameters}
\begin{table}[H]
\caption{LVC model parameters obtained upon fitting \textit{ab initio} calculations.
Parameters associated to the first-order expansion ($\lambda_{i}$ or $\kappa_{i}$) are given by the associated characteristic shift $d_{i}=\pm\frac{\lambda_{i}}{k_{i}}\text{ or }-\frac{\kappa_{i}}{k_{i}}^{(k)}$ in the mass-weighted framework for displacement along the normal modes directions. a. u. stands for atomic units.}
\label{tab:parameters_LVC}
\centering
\begin{tabular}{cr|crl}
\toprule
Parameter & Value (a.u.) & Equivalent & Value & Unit\\
\midrule
$E^{(1)}$ & -- & -- & 4.405 & \si{\electronvolt}\\
$E^{(2)}$ & -- & -- & 4.380 & \si{\electronvolt}\\
\midrule
$\kappa_{81}^{(1)}$ & -0.000376319797672 & ${d_{81}^{(1)}}$ & 7.900 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{81}^{(2)}$ & -0.000154827457621 & ${d_{81}^{(2)}}$ & 3.094 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{88}^{(1)}$ & 0.000773557607664 & ${d_{88}^{(1)}}$ & -7.205 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{88}^{(2)}$ & 0.000894008947134 & ${d_{88}^{(2)}}$ & -8.274 & $a_{0}\sqrt{m_{e}}$ \\
$\lambda_{87}$ & 0.000745662804364 & $d_{87}$ & 7.417 & $a_{0}\sqrt{m_{e}}$ \\
\midrule
$k_{81}^{(1)}$ & 0.0000476374261878 & $\omega_{81}^{(1)}$& 1515 & \si{\per\centi\meter}\\
$k_{81}^{(2)}$ & 0.0000500377681987 & $\omega_{81}^{(2)}$& 1552 & \si{\per\centi\meter}\\
$k_{88}^{(1)}$ & 0.000107366759493 & $\omega_{88}^{(1)}$& 2274 & \si{\per\centi\meter}\\
$k_{88}^{(2)}$ & 0.000108052889481 & $\omega_{88}^{(2)}$& 2281 & \si{\per\centi\meter}\\
$k_{87}^{(1)}$ & 0.000100509641682 & $\omega_{87}^{(1)}$& 2200 & \si{\per\centi\meter}\\
$k_{87}^{(2)}$ & 0.000100546482724 & $\omega_{87}^{(2)}$& 2201 & \si{\per\centi\meter}\\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[H]
\caption{QVC model parameters obtained upon fitting \textit{ab initio} calculations.
Parameters associated to the first-order expansion ($\lambda_{i}$ or $\kappa_{i}$) are given by the associated characteristic shift $d_{i}=\pm\frac{\lambda_{i}}{k_{i}}\text{ or }-\frac{\kappa_{i}}{k_{i}}^{(k)}$ in the mass-weighted framework for displacement along the normal modes directions. a. u. stands for atomic units.
Parameters associated to the second-order expansion ($\gamma_{i,j}$ or $\mu_{i,j}$) are only given in atomic units.}
\label{tab:parameters_QVC}
\centering
\begin{tabular}{cr|crl}
\toprule
Parameter & Value (a.u.) & Equivalent & Value & Unit\\
\midrule
$E^{(1)}$ & -- & -- & 4.407 & \si{\electronvolt}\\
$E^{(2)}$ & -- & -- & 4.379 & \si{\electronvolt}\\
\midrule
$\kappa_{81}^{(1)}$ & -0.000389737234515 & ${d_{81}^{(1)}}$ & 8.203 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{81}^{(2)}$ & -0.00015231303099 & ${d_{81}^{(2)}}$ & 3.059 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{88}^{(1)}$ & 0.000790813573696 & ${d_{88}^{(1)}}$ & -7.365 & $a_{0}\sqrt{m_{e}}$ \\
$\kappa_{88}^{(2)}$ & 0.000894785951449 & ${d_{88}^{(2)}}$ & -8.289 & $a_{0}\sqrt{m_{e}}$ \\
$\lambda_{87}$ & 0.00065515654139 & $d_{87}$ & 6.569 & $a_{0}\sqrt{m_{e}}$ \\
\midrule
$\gamma_{81,88}^{(1)}$ & -0.00000259362606264 & -- & \num{-2.594e-6} & $E_h/(a_0^2m_e)$ \\
$\gamma_{81,88}^{(2)}$ & -0.000000511909024135 & -- & \num{-0.512e-6} & $E_h/(a_0^2m_e)$ \\
$\mu_{87,81}$ & -0.00000108420351464 & -- & \num{-1.084e-6} & $E_h/(a_0^2m_e)$ \\
$\mu_{87,88}$ & -0.0000118672374859 & -- & \num{-11.867e-6} & $E_h/(a_0^2m_e)$ \\
\midrule
$k_{81}^{(1)}$ & 0.0000475127642257 & $\omega_{81}^{(1)}$& 1513 & \si{\per\centi\meter}\\
$k_{81}^{(2)}$ & 0.0000498104486865 & $\omega_{81}^{(2)}$& 1549 & \si{\per\centi\meter}\\
$k_{88}^{(1)}$ & 0.000107374006285 & $\omega_{88}^{(1)}$& 2274 & \si{\per\centi\meter}\\
$k_{88}^{(2)}$ & 0.00010794607905 & $\omega_{88}^{(2)}$& 2280 & \si{\per\centi\meter}\\
$k_{87}^{(1)}$ & 0.000101563963056 & $\omega_{87}^{(1)}$& 2212 & \si{\per\centi\meter}\\
$k_{87}^{(2)}$ & 0.0000979165892662 & $\omega_{87}^{(2)}$& 2171 & \si{\per\centi\meter}\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Potential energy surfaces}
\begin{figure}[H]
\centering
\includegraphics[width=1.0\linewidth]{m22_fit_QVC_all_contour_hspace0_wspace0.pdf}
\caption{QVC potential energy surfaces (in \si{eV}) of electronic states S\textsubscript{1} (a) and S\textsubscript{2} (b) in the planes (87,81), (87,88), and (81,88) from left to right.
For unspecified coordinates, the values are those at the MECI.
All coordinates are mass-weighted and given in atomic units, $\sqrt{m_e}a_0$.}
\label{fig:contours_all_QVC}
\end{figure}
\section{Results of quantum dynamics}
\subsection{Early dynamics with different initial conditions}
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{early_dynamics_m22_LVC_abs_2_DRAFT.pdf}
\caption{Early dynamics: trajectories of the center of the wave packets in the (81,88)-plane.
The initial state for the propagation is entirely B\textsubscript{2}-populated.
Diabatic potential energy surfaces A\textsubscript{1} and B\textsubscript{2} in blue and red respectively, adiabatic PESs S\textsubscript{1} and S\textsubscript{2} in dashed lines and plain lines respectively (a).
Diabatic (plain lines) and adiabatic (dashed lines) populations associated to these trajectories (b).
}
\label{fig:early_dynamics_B2}
\end{figure}
\subsection{Relaxation and emission from the second vibronic eigenstate}
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{m22_LVC_rlx_1_t50_all_wf_hspace0_wspace0.pdf}
\caption{Nuclear wave packet contributions in diabatic electronic states B\textsubscript{2} (a) and A\textsubscript{1} (b) in the second vibronic state, in planes (87,81), (87,88), and (81,88) from left to right.
For unspecified coordinates, the values are those at the MECI.
All coordinates are mass-weighted and given in atomic units, $\sqrt{m_e}a_0$.
The blank figure corresponds to the wave functions having the B\textsubscript{2} coordinate 87 in the nodal plane.}
\label{fig:vibronic_states_2}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{m22_LVC-wide-19_emi_1a_1b-crop.pdf}
\caption{Simulated emission spectra from vibrational contributions of B\textsubscript{2} (plain lines) and A\textsubscript{1} (dashed lines) of the second vibronic eigenstate to the electronic ground state.
}
\label{fig:emission_1a_1b}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{m22_LVC-wide-19_abs_emi_1-crop.pdf}
\caption{Absorption spectra from the electronic ground state to both diabatic states (blue lines) and emission spectra from the different contributions of the second vibronic eigenstate (red lines).
Realistically broadened bands are obtained using a damping time $\tau = \SI{19}{\femto\second}$.}
\label{fig:abs_emi_spectra_1}
\end{figure}
For the comparison of the results corresponding to the second vibronic eigenstate and the experiment, the shift applied to our theoretical spectra for the 0-0 band of absorption to match with the experiment is $\Delta E = \SI{0.11}{\electronvolt}$ and $\Delta \bar{\nu} = \SI{883}{\per\centi\meter}$ (\SI{7.9}{\nano\meter} at the 0-0 band).
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{exp_and_theo_spectra_1a-1b_pl_LVC-wide-19_shifted1abs_energy-crop.pdf}
\caption{Experimental absorption and emission spectra (dashed lines), courtesy of Chu and Pang, 2004\cite{chu_vibronic_2004}, reproduced in (a) and (b).
Theoretical absorption spectrum to the second vibronic eigenstate reproduced in (a) and (b), theoretical emission spectrum from the first vibronic eigenstate, contributions of B\textsubscript{2} (a) and of A\textsubscript{1} (b).
Realistically broadened bands are obtained using a damping time $\tau = \SI{19}{\femto\second}$.
}
\label{fig:exp_and_theo_spectra}
\end{figure}
\section{Vibronic coupling operator files}
The operator files to be used with Quantics are given in the following for both LVC and QVC Hamiltonian models.
Note that the parameters are given for a mass-weighted system of coordinates.
As a consequence, the reduced masses of the three coordinate are 1, numerically.
Inputs files for propagation and relaxation calculations are available upon request.
\subsection{Quantics Operator file for the LVC Hamiltonian model}
\noindent
\includegraphics[page=1]{operator_file_LVC-crop.pdf}
\newpage
\noindent
\includegraphics[page=2]{operator_file_LVC-crop.pdf}
\newpage
\noindent
\includegraphics[page=3]{operator_file_LVC-crop.pdf}
\subsection{Quantics Operator file for the QVC Hamiltonian model}
\noindent
\includegraphics[page=1]{operator_file_QVC-crop.pdf}
\newpage
\noindent
\includegraphics[page=2]{operator_file_QVC-crop.pdf}
\newpage
\noindent
\includegraphics[page=3]{operator_file_QVC-crop.pdf}
\bibliographystyle{article_title}
|
\section{Introduction}\label{intro}
The waveguide (WG) couplers with tolerance to variations in fabrication and input wavelength are highly desirable with many potential and significant applications in area of optical circuits
and communications \cite{Ramadan:98,Paloczi2004,Sun:09,Chen:16}. Such devices have been widely used as switches \cite{Sekiguchi:12}, beam splitters \cite{Chenxi2018}, polarization rotators \cite{Chen:14} and other building blocks of optical waveguides \cite{PhysRevLett.101.200502,Yu2008}. They can be fabricated in the silicon-on-insulator (SOI) platform due to their simple configurations and the ease \cite{Lu:15}. To our knowledge, the most commonly used is adiabatic couplers because they do not require strict fabrication
control and are broadband \cite{Ho:15,PhysRevA.93.033802,PhysRevA.87.013806,PhysRevE.73.026607} in comparison to the conventional couplers which are known to be highly sensitive to structural parameters.
Building on the strong similarity between quantum mechanics and waveguide optics, coupled optical waveguides have proven to provide a very rich laboratory tool to investigate with optical waves the classical analogues of a wide variety of coherent quantum effects \cite{PhysRevA.71.065801} encountered in atomic, molecular or condensed-matter physics \cite{Longhi2009}. On the other hand, many coherent protocols based on quantum techniques have been exploited to manipulate light propagation in optical waveguide couplers, such as stimulated Raman adiabatic passage (STIRAP) \cite{Longhi2009,PhysRevA.70.063409,RevModPhys.89.015006}, shortcut to adiabaticity (STA) \cite{Ho:15,Tseng:14,Dou2022,Zhang_2021}, for instance, counter-diabatic driving (or quantum transitionless driving) \cite{Chenxi2018,Demirplak2003,Demirplak2008,PhysRevA.90.060301}, dressed-state technique \cite{PhysRevLett.116.230503,Dou2020}, adiabatic elimination \cite{PhysRevA.97.023811,PhysRevA.29.1438}, etc. These methods are committed to the exploitment of devices with high coupling or conversion efficiency at specific device lengths which feature good robustness against wavelength and fabrication variations \cite{Ho:15}. However, realistic propagation of light in the waveguide which is associated with nonlinear optical process \cite{PhysRevLett.101.193901}. When the nonlinearity come into play in waveguides, using these tools is a challenging question. At the same time, the general instability of the dynamics prevents the direct application of simple strategies \cite{PhysRevA.89.012123,PhysRevA.98.022102,PhysRevA.93.043419,PhysRevA.78.053410,PhysRevA.81.052112}.
The study of waveguide couplers that contain nonlinear materials have triggered great interest in the integrated optics \cite{Dang2017,Julius2022,Hanapi2021,Deka2018,Jia:10,Schiek2012,PhysRevB.88.045443,Goto2016} and are critically essential for the implementation in future quantum computers \cite{Goto2016} due to their novel all-optical switching applications \cite{Schiek2012,PhysRevB.88.045443} since being introduced by Jensen in 1982 \cite{Jensen1982}. It has been demonstrated that light propagation (light transfer, light split, light return, etc.) in waveguide systems exhibits intriguing features in the presence of optical nonlinearity \cite{Jia:10,Diebel2014,Friberg1987,deLima2018,Venugopal2012,Petr2016}. The nonlinear couplers can exhibit self-trapping, self-modulation and self-switching of the energy of the coupled modes \cite{Leoski2004}. For the analysis of problems of light pulse propagation in nonlinear optical couplers, the coupled-mode theory (CMT) \cite{Huang1994,Yuan1994} is also adopted which provides an easy and intuitive way for us to obtain physical insight. The important and natural phenomenon of optical fibers is the Kerr (third-order) nonlinearities due to the change of the refractive index of partial coupler structure as a result of high propagating power \cite{Zhang2021,PhysRevA.91.013840,Dacles-Mariani2007, Qi2014,Wang2004,Fleischer:05,PhysRevA.93.023848,PhysRevE.82.056605}. Of course, the question naturally arises on how nonlinear effects may influence such adiabatic transfer processes \cite{PhysRevLett.101.193901,PhysRevA.73.013617}. They are known to induce dynamically unstable regions in the parameter spaces \cite{PhysRevLett.93.250403,PhysRevA.73.013617,PhysRevLett.119.243902,PhysRevA.72.013608,PhysRevA.93.023848,PhysRevE.82.056605} and impair the efficiency of STIRAP drill on theory and experiment \cite{PhysRevLett.101.193901}. This failure was explained by the destruction of the dark state formed in the STIRAP configuration \cite{PhysRevLett.93.250403,PhysRevA.72.013608,PhysRevLett.101.193901,PhysRevLett.119.243902,PhysRevA.73.013617}. To overcome this detriment of nonlinearity, recent work established an exact resonance-locked inverse engineering technique allows one to induce a controlled population inversion which surpasses the usual nonlinear STIRAP efficiency in the case of quantum $\Lambda$ systems featuring second- and third-order nonlinearities \cite{PhysRevLett.119.243902}.
In this paper, we theoretically propose a design of the nonlinear waveguide couplers by the resonance-locked inverse engineering based on the STIRAP. The resonance-locked inverse protocol shows how the efficiency of the nonlinear STIRAP can be improved by the longitudinal varying detunings of the propagation constants depending on the desired evolution of the light intensity, which aims to nullify the nonlinear effect. We also demonstrate that our devices could achieve complete energy transfer, beam splitting and light return with robustness against coupling parameters fluctuations.
This paper is organized as follows. In Sec. \ref{sec:2}, we present the Hamiltonian of the nonlinear waveguide optical model and briefly explain the basic idea of the resonance-locked inverse engineering. The numerical results for the spatial evolution of the three light distributions in the single-mode waveguides for different coupling constants under study are reported in Sec. \ref{sec:3}. Finally, the conclusions are given in Sec. \ref{sec:4}.
\section{Model and approach}\label{sec:2}
For an array of $N+1$ optical waveguides with Kerr nonlinearity, in the framework of the CMT, the evolution of the wave amplitude with nearest-neighbor evanescent coupling is accurately described by a set of $N+1$ coupled differential equations \cite{Longhi2009}
\begin{equation}
\label{eq1}
i\frac{d}{dz}\mathbf{a}(z)=H(z)\mathbf{a}(z),
\end{equation}
where $\mathbf{a}(z)=[a_{1}(z),a_{2}(z),\ldots, a_{N+1}(z)]^T$ is the electric field amplitude of the individual waveguides. Replacing the spatial variation $z$ with the temporal variation $t$, Eq. (\ref{eq1}) is equivalent to the time-dependent Schr\"{o}dinger equation ($\hbar=1$). The $H(z)$ describes the interaction between the waveguide modes could be
explicitly read as (in matrix form) \cite{Christodoulides1988,Sukhorukov2003}
\begin{widetext}
\begin{equation}
\setlength{\arraycolsep}{2pt}
H(z)=\left[\begin{array}{ccccccc}
\beta_{1}+\Gamma|a_{1}(z)|^2 & C_{1} &0 &\cdots&0\\
C_{1}&\beta_{2}+\Gamma|a_{2}(z)|^2&C_{2}&\cdots&0\\
0 & C_{2} & \ddots &\ddots&\vdots\\
0&0 & \ddots & \ddots & C_{N}\\
0&0 & \cdots & C_{N} & \beta_{N+1}+\Gamma|a_{N+1}(z)|^2
\end{array}\right],\label{eq2}
\end{equation}
\end{widetext}
where the $C_{k}$ $(k=1,2,3,\cdots,N)$ is the $z$-dependent coupling coefficients which can be achieved with changing the distance between the waveguides. The dynamics preserves the normalization, which is fixed as $\sum_{i=1}^{N+1}|a_{i}(z)|^2=1$. $\beta_{i}$ denotes the linear propagation constant. It is assumed that the refractive index $n$ of the material is intensity-dependent, i.e., $n = n_{0} + n_{2}|E|^2$ \cite{PhysRevA.91.013840,Dacles-Mariani2007}. Such a dependence of $n$ results in linear term $\beta_{i}$ and nonlinear term $\Gamma|a_{i}(z)|^2$ in Eq. (\ref{eq1}). $\Gamma$ characterizes the Kerr nonlinearity which may lead to dynamical instability \cite{PhysRevLett.101.193901,PhysRevA.73.013617,PhysRevLett.93.250403,PhysRevLett.119.243902,PhysRevA.72.013608,PhysRevA.102.052203,PhysRevLett.98.050406,PhysRevLett.121.250405,Tsoy2018}. In the waveguide system, nonlinearity is introduced by increasing the power of the input beam \cite{PhysRevLett.81.3383,PhysRevLett.95.073902,PhysRevLett.74.2941}. At low light power levels, $\Gamma$ can be neglected ($\Gamma=0$) and the corresponding system is perfectly analogous to the linear case. With the emergence of the nonlinearity, the differential equation Eq. (\ref{eq1}) is no
longer analytically solvable, we therefore exploit a fourth-fifith order Runge-Kutta algorithm to trace the light evolution numerically. Without loss of generality, we take the coupling and nonlinear coefficients to be real and positive and the variation of the coupling coefficients due to nonlinearity at high input intensity are neglected in our works. Moreover, the study of beam dynamics in nonlinear waveguide couplers \cite{PhysRevA.93.023848,Wang2004,Fleischer:05} is of great significance for light routing in photonic circuits, for beam steering and switching \cite{Tsoy2018,RevModPhys.83.247}.
When the intensity of the optical pulses crosses a certain threshold value then the optical fiber behaves nonlinearly \cite{PhysRevLett.74.2941}. The existence of the nonlinearities make it not justified to apply the adiabatic condition of linear STIRAP to nonlinear case \cite{PhysRevLett.90.170404,PhysRevA.73.013617,MENG2009,PhysRevLett.99.223903} because of the absence of the superposition principle. More interestingly, the adiabatic
geometric phase or Berry phase is found also to be modified by the nonlinearity \cite{PhysRevA.81.052112}. There may exist more eigenstates than the dimension of the Hilbert space \cite{PhysRevLett.90.170404,MENG2009} and the non-adiabatic coupling between these eigenstates will drastically decrease the transfer efficiency \cite{PhysRevLett.99.223903,PhysRevA.73.013617}. Therefore, it is very difficult to analyze the adiabaticity of the nonlinear systems and follow their adiabatic conditions \cite{PhysRevLett.98.050406,PhysRevA.78.053410}. The key here for the success of nonlinear STIRAP is how to avoid these unstable regimes when designing the route of adiabatic passage.
We wish to nullify the nonlinear effect of nonlinear coupled-waveguide devices using resonance-locked engineering scheme \cite{PhysRevLett.119.243902} based on the STIRAP. This method is performed by introducing a general parametrization of the state vector and further derive the necessary conditions to obtain an efficient transfer. Now we employ resonance-locked inverse engineering technique to the system. We have to consider the phase $\gamma$ on parametrization
\begin{eqnarray}
a_{n}\mapsto a_{n}e^{-i\gamma},
\end{eqnarray}
this phase $\gamma$ is $z$-dependent and will be used to impose the resonance locking [see Eq. (\ref{eq6})]. Then the Eq. (\ref{eq1}) can be written as
\begin{equation}
i\frac{d}{dz}\mathbf{a}(z)=(H(z)-\dot{\gamma})\mathbf{a}(z)=H'(z)\mathbf{a}(z),
\end{equation}
and the resulting Hamiltonian $H'(z)$ becomes
\begin{widetext}
\begin{equation}
\setlength{\arraycolsep}{0.8pt}
H'(z)=\left[\begin{array}{ccccccc}
\beta_{1}+\Gamma|a_{1}(z)|^2-\dot{\gamma} & C_{1} &0 &\cdots&0\\
C_{1}&\beta_{2}+\Gamma|a_{2}(z)|^2-\dot{\gamma}&C_{2}&\cdots&0\\
0 & C_{2} & \ddots &\ddots&\vdots\\
0&0 & \ddots & \ddots & C_{N}\\
0&0 & \cdots & C_{N} & \beta_{N+1}+\Gamma|a_{N+1}(z)|^2-\dot{\gamma}
\end{array}\right].
\end{equation}
\end{widetext}
In order to completely eliminate the nonlinear effects dynamically, where the dynamics is exactly the same as the linear case, we shilling one of the elements on the main diagonal equal to zero, i.e., $\beta_{i}+\Gamma|a_{i}(z)|^2-\dot{\gamma}=0$, and then substitute it into the other remaining main diagonal elements, which leads to the resonance-locked conditions \cite{PhysRevLett.119.243902} as follows
\begin{equation}
\label{eq6}
\beta_{j(j\neq i)}=\beta_{i}+\Gamma|a_{i}(z)|^2-\Gamma|a_{j}(z)|^2,
\end{equation}
these propagation constants $\beta_{j}$ $(j=1,2,3\cdots,N+1)$, fully compensate the Kerr terms which based on population dynamics, thus we can just recover the preceding equations without Kerr terms.
\section{Adiabatic light propagation in nonlinear waveguide couplers}\label{sec:3}
For the sake of simplicity, we consider a waveguide structure with three waveguides in a nonlinear coupling configuration. In the framework of CMT, the evolution of the light amplitudes $a_{1}(z)$, $a_{2}(z)$ and $a_{3}(z)$, in the presence of the optical Kerr effect, can be described by the set of coupled discrete nonlinear Schr\"{o}dinger equation (DNLSE) with the Hamiltonian given by,
\begin{equation}
\label{eq3}
\setlength{\arraycolsep}{1pt}
H_{0}(z)=\left[\begin{array}{ccc}
\beta_{1}+\Gamma|a_{1}(z)|^2 & C_{1} &0\\
C_{1}&\beta_{2}+\Gamma|a_{2}(z)|^2&C_{2}\\
0&C_{2}& \beta_{3}+\Gamma|a_{3}(z)|^2
\end{array}\right].
\end{equation}
It is easy to show, as in the linear counterpart, the nonlinear systems also support the existence of the coherent population trapping (CPT) state or dark state with zero eigenvalue, with the corresponding state vector given by $|\Psi_{0}\rangle=(a_{1}^0,a_{2}^0,a_{3}^0)$, where $a_{1}^0=\frac{C_{2}}{\sqrt{C_{1}^2+C_{2}^2}}$, $a_{2}^0=0$ and $a_{3}^0=\frac{C_{1}}{\sqrt{C_{1}^2+C_{2}^2}}$. The existence of the CPT state cannot guarantee that it can always be followed adiabatically. The nonorthogonality between the CPT state and any of the other eigenstates is quite obvious. In order to keep the mathematics as simple as possible without sacrificing the central physics, we have set here $\dot{\gamma}=\beta_{2}+\Gamma|a_{2}|^{2}$ by the same procedures above, and the resonance-locked conditions can be obtained as
\begin{eqnarray}
\label{eq4}
\begin{aligned}
\beta_{1}=&\Gamma(|a_{2}|^{2}-|a_{1}|^{2})+\beta_{2},\\
\beta_{3}=&\Gamma(|a_{2}|^{2}-|a_{3}|^{2})+\beta_{2},
\end{aligned}
\end{eqnarray}
then the nonlinear Hamiltonian Eq. (\ref{eq3}) can be rewritten as
\begin{equation}
\setlength{\arraycolsep}{2.5pt}
H_{0}'(z)=\left[\begin{array}{ccc}
0 & C_{1} &0\\
C_{1}&0&C_{2}\\
0&C_{2}& 0
\end{array}\right].
\end{equation}
Subsequently, we can calculate the the propagation matrix $H_{0}'(z)$ corresponding eigenvalues $E_{0}=0$, $E_{\pm}=\pm\sqrt{C_{1}^2+C_{2}^2}$ and the eigenstates given by
\begin{eqnarray}
\begin{aligned}
|\Phi_{0}(z)\rangle=&\cos\theta|1\rangle-\sin\theta|3\rangle,\\
|\Phi_{\pm}(z)\rangle=&\frac{1}{\sqrt{2}}\sin\theta|1\rangle \pm\frac{1}{\sqrt{2}}|2\rangle \frac{1}{\sqrt{2}}\cos\theta|3\rangle,
\end{aligned}
\end{eqnarray}
where the mixing angle $\theta$ is defined as
\begin{equation}
\tan\theta=\frac{C_{1}}{C_{2}},
\end{equation}
and the adiabatic Hamiltonian
\begin{equation}
\setlength{\arraycolsep}{2.5pt}
H_{a}(z)=\left[\begin{array}{ccc}
\sqrt{C_{1}^2+C_{2}^2} & \frac{1}{\sqrt{2}}i\dot{\theta} &0\\
-\frac{1}{\sqrt{2}}i\dot{\theta}&0& -\frac{1}{\sqrt{2}}i\dot{\theta}\\
0& \frac{1}{\sqrt{2}}i\dot{\theta}& -\sqrt{C_{1}^2+C_{2}^2}
\end{array}\right].
\end{equation}
The off-diagonal terms of $H_{a}(z)$ are regarded as a non-adiabatic correction. When the adiabatic criterion, which can be written as $\sqrt{C_{1}^2+C_{2}^2}\gg \dot{\theta}$, is satisfied, there is no transition between the adiabatic states during the evolution. Mathematically, it means that for the entire dynamical evolution, the system remains at one of the systems eigenmodes \cite{Suchowski2014}.
When the evolution of the system is adiabatic, we can continue to predict the dynamics of light propagation in waveguides which obeys the following dark state $|\Phi_{0}(z)\rangle$, light intensity contained in WG1 and WG3 can be described as $|a_{1}|^{2}=\cos^2\theta$, $|a_{3}|^{2}=\sin^2\theta$ with $|a_{2}|^{2}=0$, then propagation constants mismatch under the resonance-locked inverse method can be simplified as
\begin{eqnarray}
\begin{aligned}
\Delta\beta_{12}=-&\Gamma\cos^{2}\theta,\\
\Delta\beta_{32}=-&\Gamma\sin^{2}\theta,
\end{aligned}
\end{eqnarray}
where $\Delta\beta_{12}=\beta_{1}- \beta_{2}$ and $\Delta\beta_{32}=\beta_{3}- \beta_{2}$. In addition, the nonlinear constant $\Gamma$ is evaluated to be $\Gamma=5m^{-1}W^{-1}$ \cite{PhysRevLett.101.193901,PhysRevLett.81.3383,PhysRevLett.95.073902} in our calculations. Based on the above, we can see the originally introduced phase $\gamma$ can be simplified to $\beta_{2}z$ which is commonly used in the waveguide systems \cite{PhysRevA.97.023811,Paspalakis2006,PhysRevE.73.026607,PhysRevA.103.053705}. The validity of the method is further substantiated by the fact that, when applied to the nonlinear system, the same result is obtained as in linear case. There is still much freedom to design these propagation constants detunings at will depending on the mixing angle $\theta$ and Kerr nonlinear constant $\Gamma$. It is noted that the corresponding CPT state via STIRAP are identical in form in both linear and nonlinear system, which are indeed governed by different adiabatic conditions \cite{PhysRevLett.98.050406}. When the above two conditions are satisfied at the same time, it can completely eliminate the nonlinear effect. Indeed, the adiabatic condition for the nonlinear system is more difficult to fulfill compared with its linear counterpart which limits the efficiency of the resonance-locked inverse method. Thus the result of the resonance-locked inverse engineering is not always strictly reachable desired final states in nonlinear system.
Since there are many possibilities of the two coupling coefficients, various light evolutions in the waveguide structure can be expected. Thus we are going to investigate three possible situations in the following subsections.
\subsection{Light transfer}
We first consider the situation that the two variable coupling coefficients, $C_{1}(z)$ and $C_{2}(z)$, are applied in a counterintuitive sequence, such that the coupling coefficient $C_{2}(z)$ precedes the coupling coefficient $C_{1}(z)$ and have the form as follows \cite{Paspalakis2006}
\begin{eqnarray}
\begin{aligned}
C_{1}(z)&=&\kappa \exp[-(z-z_{1})^2/\xi^2],\\
C_{2}(z)&=&\kappa \exp[-(z-z_{2})^2/\xi^2].
\end{aligned}
\end{eqnarray}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\textwidth]{figure1.eps}
\caption{Light transfer in the three-waveguide optical system ($P=200W$). The normalized light intensity against the propagation distance $z$ for (a) the original waveguide coupler and (b) the resonance-locked waveguide coupler; (c) The corresponding propagation constants detunings in resonance-locked structure; (d) The output light intensity in WG3 against the input power in original waveguide (solid black line) and resonance-locked coupler (dashed red line). Coupling parameters: $\kappa=15$$cm^{-1}$, $\xi=0.25$$cm^{-1}$, $z_{1}=1.6$cm, $z_{2}=1.4$cm, and $L=3$cm. The propagation constants $\beta_{1}$, $\beta_{2}$ and $\beta_{3}$ are set to a constant i.e., $\beta_{1}=\beta_{2}=\beta_{3}=20$$cm^{-1}$ in original case for convenience.}
\label{figure1}
\end{figure}
Here, $\kappa$ and $\xi$ are the maximum values and the width of the coupling coefficients, respectively. For light transfer, we assume the input as $\mathbf{a}(z_{i})=\lbrack 1,0,0\rbrack^T$ and expect the output final light field distributions among waveguides: $\mathbf{a}(z_{f})=\lbrack 0,0,1\rbrack^T$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figure2.eps}
\caption{The normalized intensity in WG3 at the output port $(z = L)$ as function of the coupling strength parameters $\kappa$ and $\xi$ with different input light power ($P =50W, 200W, 400W$) in original waveguide couplers (left) and resonance-locked couplers (right). The black star spots corresponding to the above coupling parameters in Fig. \ref{figure1}.}
\label{figure2}
\end{figure}
In Fig. \ref{figure1} (a) and (b), we show the normalized intensities in original waveguide coupler and resonance-locked coupler when the input light power $P=200W$. As seen, the original waveguide structure cannot realize the high-fidelity light transfer and normalized light intensities along beam propagation direction are varied irregularly owing to the nonlinear effect. However in the resonance locking case, the power adiabatically leaves WG1 and populates WG3, with very little energy remaining in the intermediate WG2 during the power exchange, in perfect analogy to the linear STIRAP process, which can be understood here the third-order Kerr nonlinearities here are compensated. We also give the shape of the detunings which depend essentially on the population dynamics in Fig. \ref{figure1} (c). And beyond that, the intensity at the end of the device at WG3 varies with the input optical power is depicted in Fig. \ref{figure1} (d), we note that the fidelity of the light transfer for the resonance-locked coupler is robust against variations of input optical power, while the output intensity in WG3 continue to reduce in original case.
To compare the performance of the proposed resonance-locked coupler to the original one, we depict the contour plots of the light intensity at the end of the device of WG3, as a function of coupling parameters $\kappa$ and $\xi$. We can find that at low input light power ($P=50W$), the output intensity of WG3 is stable in some regions of $\kappa$ and $\xi$, we do not notice any
difference with or without compensation of the Kerr terms. At this time, the influence of nonlinearity on adiabatic evolution can be almost ignored. When the input light power increases ($P=200W,400W$) where Kerr nonlinearities cannot be disregarded, the original waveguide structures are failed to realize adiabatical light transfer, in contrast, resonance locking structure can still achieve high-fidelity light transfer. The small light intensity discrepancies in Fig. \ref{figure2} can be attributed to the fact that the evolution is never perfectly adiabatic, and some non-adiabatic coupling are always present, which limits the efficiency of the STIRAP. In addition, the usual adiabatic condition for linear STIRAP does not hold in the nonlinear case \cite{PhysRevA.103.023307,PhysRevA.78.053410}, thus in some cases, our method cannot eliminate the nonlinearity completely.
\subsection{Light split}
Again we consider light initially is injected in the outer WG1 with $\mathbf{a}(z_{i})=\lbrack 1,0,0\rbrack^T$, the formulation of $C_{1}(z)$ and $C_{2}(z)$ as follows \cite{PhysRevA.80.013417},
\begin{eqnarray}
\label{15}
\begin{aligned}
C_{1}(z)&=\kappa \sin \varphi,\\
C_{2}(z)&=\kappa \cos \varphi,
\end{aligned}
\end{eqnarray}
where $\varphi(z)=\frac{\pi}{4}/(1+\exp(-(z-L/2)/\omega))$. $\kappa$ is the maximum values of the coupling coefficients and $\omega$ can influence the slope of the coupling constants.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\textwidth]{figure3.eps}
\caption{Light split in the three-waveguide optical system with the coupling coefficients given by Eq. (\ref{15}). The normalized light intensity against the propagation distance $z$ for (a) the original waveguide coupler and (b) the resonance-locked waveguide coupler; (c) The corresponding propagation constants detunings in resonance-locked structure. In (d) the output light intensity in WG1 for original waveguide (solid black line) and resonance-locked coupler (dashed red line) is shown. The parameters for this figure are $\kappa=25$$cm^{-1}$, $\omega=0.25$$cm^{-1}$, $P=200W$, and $L=3$cm.}
\label{figure3}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figure4.eps}
\caption{As the coupling strength parameters $\kappa$ and $\omega$ vary, the normalized intensity in WG1 at the output port $(z = L)$ with different input optical power in original waveguide couplers (left) and resonance-locked couplers (right) is shown.}
\label{figure4}
\end{figure}
Rather than light transfer, the goal of the light split is to divide equally light intensity in WG1 and WG3, the WG2 almost not excited during the optical evolution. In original waveguide coupler, as the input light power $P=200W$, one cannot obtain a complete split in the presence of three-order nonlinearities which we can see from Fig. \ref{figure3} (a). The nonlinear waveguide system with three-order nonlinearities demonstrate noncontrollable exactly in the sense that such nonlinearities prevent reaching the target state exactly \cite{PhysRevA.88.063622,PhysRevA.102.052203}. However, in resonance-locked scheme, light split can be realized perfectly. The issue of robustness of the output intensity of WG1 with respect to the input light power is further numerically analyzed in Fig. \ref{figure3} (d). As mentioned above, the output light intensity in WG1 for the resonance-locked case is independent of input light power. We also test the robustness of our resonance-locked couplers with respect to coupling parameters $\kappa$ and $\omega$ in Fig. \ref{figure4}, which leads to the same conclusion in light transfer case. As the nonlinear interaction grows ($P=200W, 400W$), non-adiabatic oscillations are significantly strengthened and the light intensity is reduced dramatically, while it is relatively stable in resonance-locked case.
\subsection{Light return}
Once more we consider the input as $\mathbf{a}(z_{i})=\lbrack 1,0,0\rbrack^T$ and choose the coupling coefficients to be Gaussian type of the form \cite{Paspalakis2006}
\begin{eqnarray}
\begin{aligned}
C_{1}(z)&=&\kappa \exp[-(z-L/2)^2/\xi_{1}^2],\\
C_{2}(z)&=&\kappa \exp[-(z-L/2)^2/\xi_{2}^2],
\end{aligned}
\end{eqnarray}
where the $\kappa$ denotes the maximum values of the coupling coefficients and the parameters $\xi_{1}$ and $\xi_{2}$ represent the the widths of the coupling coefficients. It is well known that the system will regularly undergo the complete population return (CPR) with $\mathbf{a}(z_{f})=\lbrack 1,0,0\rbrack^T$ when excited by a resonant coupling \cite{Vitanov1995}, i.e., the process is reciprocal. The coupling coefficients must only be sufficiently wide and overlap significantly in distance so that adiabatic evolution is ensured \cite{PhysRevA.103.053705}. Finally results of complete light return to the initially excited waveguide WG1 is shown in Fig. \ref{figure5}. During the propagation, the other two waveguides WG1 and WG3 take part in the coupling in Fig. \ref{figure5} (a). The special feature is the appearance of pronounced oscillations due to the non-adiabatic transition between the different eigenstates \cite{PhysRevA.73.013617}. We, however, notice that our resonance-locked waveguide coupler could obtain the almost complete population return even in the presence of Kerr terms in Fig. \ref{figure5} (b).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\textwidth]{figure5.eps}
\caption{Light return in the three-waveguide optical system. The normalized light intensity against the propagation distance $z$ for (a) the original waveguide coupler and (b) the resonance-locked waveguide coupler; (c) The corresponding propagation constants detunings in resonance-locked structure. In (d) the output light intensity in WG1 for original waveguide (solid black line) and resonance-locked coupler (dashed red line) is shown. In this example, we establish the parameters as $\kappa=40$$cm^{-1}$, $\xi_{1}=0.3$$cm^{-1}$, $\xi_{2}=0.6$$cm^{-1}$, $P=200W$, and $L=3$cm.}
\label{figure5}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figure6.eps}
\caption{The output light intensity in WG1 for original waveguide (left) and resonance-locked coupler (right) as a function of the widths of the coupling coefficients $\xi_{1}$ and $\xi_{2}$ with the different input optical power is depicted.}
\label{figure6}
\end{figure}
We continue to show the superiority of the resonance-locked waveguide couplers as compared to the original waveguide couplers by examining the light intensity return to WG1, as a function of the widths of the coupling coefficients $\xi_{1}$ and $\xi_{2}$ and the maximum value of the coupling coefficients is set to a constant $\kappa=40$$cm^{-1}$ for simplicity. Apart from that, it is shown that the output light intensity in WG1 is quite sensitive to the input light power in original case, indicating the nonadiabaticity of the light evolution in Fig. \ref{figure5} (d). We note that our resonance-locked waveguide coupler has a higher tolerance to the fluctuations in various parameters such as the width of Gaussian pulses and the input light power than original case in Fig. \ref{figure6}. The most intriguing property presented here comes from the three-order nonlinearity, as the nonlinearity gets stronger (corresponding to $P =200W,400W$), the adiabatic evolution is greatly disrupted and we observe a rapid fall of the robustness if one does not compensate the Kerr terms with the detunings.
We note that the coupling coefficients here need not be of a specific shape nor their parameters need to obey specific conditions, in contrast to the conventional nonlinear coupling approach. The choice of coefficients in the above three examples were made only for convenience.
\section{Conclusions}\label{sec:4}
In conclusion, we have theoretically employed the resonance-locked inverse method based on STIRAP for light evolution in nonlinear waveguide couplers. The presence of third-order Kerr nonlinearities which critically depends on the light intensity and the excitation power, impairs the efficiency of the STIRAP. It was shown that, upon an appropriate choice of the longitudinal evolution of the propagation constants detunings, the effect introduced by the third-order nonlinearities have been dynamically compensated, thus, the above nonlinear system exhibits like the linear regime subjected to the adiabatic passage process. Both the theoretical analysis and numerical calculations show that high-fidelity and robust light transfer, light split, and light return can be observed in the our resonance-locked waveguide structures in contrast to the case of the conventional nonlinear couplers. In addition, the proposed scheme is not constrained by the need for having exact system parameters and is not limited by Kerr nonlinearities parameters which also promises myriad possibilities in designing robust nonlinear waveguide structures such as beam splitters, switches, and directional couplers. Therefore, the light propagation in nonlinear waveguide couplers via the resonance-locked inverse based on STIRAP protocol is intriguing and vigorous topic in circuits and communications, magneto-optic data storage and many other practical fields.
The key point of implementation of resonance-locked inverse method based on STIRAP is the space-dependent propagation constants detunings in the waveguide coupler fabrication, which can be approximately realized by modifying the refractive index of waveguides along the propagation direction. Experimentally, the combination of the local illumination of the control beam and the electric field applied to the Sr$_{x}$Ba$_{1-x}$Nb$_{2}$O$_{6}$ (SBN) crystal \cite{PhysRevA.103.053705,PhysRevA.95.023811,Gorram2009} with $x = 0.61$ could lead to the above expectation, which also reflect the technical feasibility \cite{math8071128,Zhang2022}. It should
be noted here that, different from the linear waveguide systems works \cite{PhysRevA.103.053705,PhysRevA.95.023811,Gorram2009}, we consider the appropriate choices of the longitudinally varying detunings could fully compensate the nonlinear effect to realize the high-fidelity light propagation even in the nonlinear regime. Another distinguishing feature is that others discuss the spatial longitudinally varying detuning between the propagation constants of the waveguides could provide with more freedom than $z$-independent detuning to manipulate light evolution \cite{PhysRevA.103.053705,PhysRevA.95.023811}. Moreover, we believe this study may open new possibilities of exploiting the resonance-locked inverse scheme based on STIRAP for various applications in integrated optics.
\section*{Acknowledgments}
The work is supported by the National Natural Science Foundation of China (Grant No. 12075193).
|
\section{Introduction}
\label{sec:1}
\IEEEPARstart{O}{bject} detection in remote sensing images, which aims to determine the locations and categories of the interested targets, has become one of the most common and potential image interpretation steps for numerous applications\cite{deng2017toward,wang2019enhancing,shi2020orientation,yu2020orientation}, such as maritime rescuing\cite{varga2022seadronessee}, urban planning\cite{ma2018mobile} and traffic management\cite{asha2018vehicle}.
\begin{figure}[!tbp]
\centering
\includegraphics[width=1.0\linewidth]{imbalanced.png}
\caption{As shown in gray blocks, generic label assignment strategy selects samples on certain level of feature maps according to the scales of targets. This strategy will lead to the insufficient sampling for targets with extreme scales and aspect ratios. Best viewed in colour.}
\label{fig:Imbalance}
\vspace{-0.105cm}
\end{figure}
In the past few years, substantial methods have made significant progress for generic object detection based on CNN\cite{ren2015faster,redmon2016you,retinanet,tian2019fcos,duan2019centernet}, which have greatly promoted the development of object detection in remote sensing images\cite{wang2020learning,dai2022ace,cheng2022anchor} and attract wide attention from researchers.
However, object detection in remote sensing images is still facing challenges from the characteristics of targets, such as arbitrary orientations, large aspect ratios and large scale variations. To address above challenges, current detectors for remote sensing images have been designed to enhance orientation prediction based on generic detectors, such as adding extra prediction branch \cite{lin2019ienet,qin2021mrdet,yu2022object} and alleviating feature misalignment \cite{yu2022object,han2021align,yang2019r3det}.
Moreover, feature pyramid network (FPN)\cite{lin2017feature,hou2022refined,wang2020frpnet} is often adopted to represent multi-scale features, where the upper feature maps have semantic information and are often employed to detect large targets, whereas lower feature maps have more fine-gained details and are utilized to detect small targets \cite{zhu2019fsaf}. However, it usually faces the crucial issue of selecting training samples at reasonable pixel-wise locations (spatial assignment) on different feature levels (scale assignment), namely label assignment \cite{zhu2020autoassign}.
Many label assignment methods have been proposed for object detection due to their important role in determining positive or negative samples, which directly and significantly influences the performance \cite{zhang2020bridging}, such as maximum intersection over union (IoU) based strategies \cite{qin2021mrdet,yang2019r3det,yang2019scrdet} and center-based strategies \cite{cheng2021anchor,shi2021canet,feng2020toso}. However, they are barely designed specifically for remote sensing images and neglect the essential characteristics, e.g., the bird-eye-view and various scales of targets, which brings tough problems in object detection.
Concretely, when combined with FPN, current scale assignment strategies always heuristically assign samples in upper feature maps to large targets, and samples in lower feature maps to small targets\cite{zhu2019fsaf}. However, this heuristic often causes imbalance sampling, especially in remote sensing images. On the one hand, few positive samples will be assigned to targets with extreme scales and large aspect ratios. As shown in Fig.~\ref{fig:Imbalance}, samples will be treated as negatives due to the
heuristic scale constrains, even if these samples fall within the sampling range in multi-level feature maps. On the other hand, samples for small targets tend to be dominated in number, as illustrated in Fig.~\ref{fig:Imbalance}, where vehicles are far more numerous than planes. As a result, using these strategies will lead to a flood of positive samples assigned to the lower level of feature maps, thus overwhelm the sufficient attention to targets with other scales and introduce the scale-level bias. Therefore, a more flexible method is required for sample selection on multi-level feature maps.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.7\linewidth]{background.png}
\caption{Illustration of using different ranges as positive spatial area for sampling. (Left column) Oriented bounding box. (Right column) Central area.}
\label{fig:backgroundissue}
\end{figure}
In addition, current detectors usually use rectangle bounding box as the sampling range in spatial assignment, and tend to introduce background noise as shown in Fig.~\ref{fig:backgroundissue}~(Left column). While center-based strategies, such as center sampling \cite{zhang2020bridging}, can improve the quality of positive samples by only sampling central area of targets. However, it still rarely takes into account the large aspect ratios and arbitrary orientation of targets in remote sensing images, which leads to many foreground samples being misclassified as negatives and reduces the detection accuracy for rotated targets with large aspect ratios, as shown in Fig.~\ref{fig:backgroundissue}~(Right column). Hence, a more reliable spatial assignment strategy is necessary for detectors to capture positive samples with higher quality.
To address aforementioned limitations in both scale assignment and spatial assignment, in this paper, we propose an Elliptical Distribution aided Adaptive Rotation Label Assignment (EARL) strategy to select positive samples with higher quality on multi-level feature maps dynamically, thus improve the detection performance. Specifically, we designed a novel and intuitively simple strategy, namely adaptive scale sampling (ADS) strategy, to avoid insufficient sampling and the imbalanced distribution of positive samples on multi-level feature maps.
In addition, a dynamic elliptical distribution aided sampling (DED) strategy, together with a spatial distance weighting (SDW) module, is proposed to consistently exploit orientation and shape properties of targets, and mitigate the effect of low-quality samples.
Besides, the proposed EARL is deployed on a simple anchor-free detector, i.e., FCOS, with minimal modification to succinctly represent the effectiveness. Extensive experiments on popular remote sensing datasets, such as DOTA and HRSC2016, demonstrate the effectiveness and the superiority of our proposed EARL.
The contributions of this work are summarized as follows:
\begin{enumerate}
\item An adaptive scale sampling (ADS) strategy is proposed to dynamically select positive samples among all feature levels, which avoids insufficient sampling and alleviates the scale-level bias.
\item A dynamic elliptical distribution aided sampling (DED) strategy is proposed to select samples by taking full advantage of orientation and shape properties of targets.
\item A spatial distance weighting (SDW) module is employed to integrate the quality enhancement with loss function, and further mitigate the effect of low-quality positive samples.
\item The proposed label assignment strategy is evaluated by being integrated with simple detection architecture, without bells and whistles, EARL outperforms the state-of-the-art anchor-free methods, and achieves comparable performance with anchor-based algorithms, while keeping the efficiency of anchor-free detectors.
\end{enumerate}
The remainder of the paper is organized as follows: The related work is reviewed in Section \ref{sec:2}. Then, the details of the proposed method are introduced in Section \uppercase\expandafter{\romannumeral3}. Experiments and analyses are described and discussed in Section \uppercase\expandafter{\romannumeral4}. Finally, conclusions are summarized in Section \uppercase\expandafter{\romannumeral5}.
\section{Related Work}
\label{sec:2}
\subsection{Oriented Object Detection in Remote Sensing Images}
Due to its wide range of application scenarios, oriented object detection in remote sensing images has developed rapidly. Similar to generic detectors, recent methods can be divided into two categories, i.e., anchor-based detectors and anchor-free detectors. To detect targets with arbitrary orientations, some anchor-based detectors\cite{liu2017rotated,zhang2018toward,yang2018automatic}, which are built on the two-stage framework, i.e., Faster R-CNN\cite{ren2015faster}, densely preset multiple anchors with different scales, aspect ratios and angles for better regression while introducing heavy anchor-related computations. After that, RoI-Transformer\cite{ding2019learning} proposed an efficient region of interests (RoI) learner to transform horizontal proposals to rotated ones. While Gliding Vertex\cite{xu2020gliding} described oriented targets by the horizontal bounding boxes (HBB), which further avoided numerous anchors with multiple angles. To improve accuracy and efficiency, R$^3$Det\cite{yang2019r3det} and S$^2$A-Net\cite{han2021align} followed the one-stage schema and designed network modules to alleviate the feature misalignment issue. However, these methods still take a long time during training and inference\cite{yu2022object}.
To maintain high efficiency, anchor-free detectors are designed to avoid the usage of anchors. For instance, TOSO\cite{feng2020toso} directly regressed the surrounding HBB and transformation parameters to present rotated targets. IENet\cite{lin2019ienet} was built upon the classic one-stage anchor-free framework, i.e., FCOS, adding an additional branch with the interactive embranchment module to regress angles. Later, BBAVector\cite{yi2021oriented}, O$^2$-DNet\cite{wei2020oriented} and ACE\cite{dai2022ace} employed a simplified architecture and designed new oriented bounding box (OBB) representation methods.
However, these anchor-free detectors still not perform state-of-the-art results compared with anchor-based detectors, and most of them are more likely to improve the architecture while not considering the training strategy, especially the label assignment strategy, which has been proven to be the essential difference between these two kinds of detectors\cite{zhang2020bridging}.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.9\linewidth]{Architecture.png}
\caption{Overview of the proposed EARL. The upper gray block shows the baseline architecture of R-FCOS. The lower green block presents our proposed EARL strategy. The DED strategy is employed based on the shape of the target and generates foreground masks on each level of feature maps outlining area to select candidate samples. The candidate samples will be sorted from the highest to lowest level of feature maps where the samples in the same level will sorted by the distance from the center point of ground truth bounding box. Only the first top-$k$ samples are selected as positive samples (coloured points). Best viewed in colour.}
\label{fig:Architecture}
\end{figure*}
\subsection{Label Assignment Strategies}
Label assignment is a crucial component for CNN-based detectors to learn distribution of targets by determining positive or negative samples. Meanwhile, as the current detectors commonly adopt FPN\cite{lin2017feature} to alleviate the scale variation, label assignment needs to simultaneously assign labels to samples at different spatial locations (spatial assignment) on different feature levels (scale assignment)\cite{zhu2020autoassign}. Most anchor-based detectors\cite{ren2015faster,retinanet,qin2021mrdet} assign the labels to anchors among multi-level feature maps by comparing the preset thresholds and IoU between anchors and ground truth bounding boxes. However, this strategy always involves many hyperparameters to adjust the scales and aspect ratios of anchors based on the datasets\cite{zhang2021learning}. While anchor-free methods\cite{tian2019fcos,kong2020foveabox,xiao2020axis} always select samples inside the ground truth bounding box or the central area of targets for spatial assignment, and choose samples on certain level of feature maps according to the predefined scale constraint for scale assignment, therefore these can be regarded as fixed-scale sampling strategy. However, above strategies, which depend on the heuristic rules, may not be optimal enough. Thus, recent works provide dynamic label assignment strategies, which allow models to learn to select samples by itself. For instance, FSAF\cite{zhu2019fsaf} dynamically assigned targets to the suitable feature levels based on computed loss. ATSS\cite{zhang2020bridging} proposed an adaptive training sample selection method by adjusting the IoU threshold according to the statistics of targets, where ATSS also proposed the center sampling strategy to improve the quality of positive samples. Autoassign\cite{zhu2020autoassign} used two weighting modules to adjust the category-specific prior distribution according to the appearances of targets.
As for oriented object detection, GGHL\cite{huang2022general} used a 2-dimensional Gaussian heatmaps to define positive samples according to the size and direction properties of targets. For stable optimization, FSDet\cite{yu2022object} used a soft assignment mechanism to weight the training samples. Inspired by the excellent works for dynamical label assignment mechanism in object detection, the proposed method focuses on designing an adaptive label assignment strategy for orientation anchor-free detectors in remote sensing images, where targets have arbitrary orientations and large variations in scales and aspect ratios.
\section{The Proposed EARL Method}
\label{sec:3}
In this section, the architecture of our baseline anchor-free detector is firstly presented. Then the three components of the proposed EARL, i.e., ADS strategy, DED strategy and SDW module, are introduced in detail.
\subsection{Architecture of the Baseline Anchor-free Detector}
\label{sec:3-1}
To demonstrate the effectiveness of the proposed label assignment strategy, in this paper, a simple but classical anchor-free architecture, namely, R-FCOS, is employed as our baseline detector.
Compared with FCOS\cite{tian2019fcos}, our R-FCOS removes centerness branch to be more compact and achieves more simplicity and efficiency, which only contains classification branch and regression branch for oriented object detection.
As shown in Fig.~\ref{fig:Architecture}, the upper gray block shows the components of our baseline detector, which consists of the backbone network, FPN and prediction heads. Let $\mathbf{C}_\ell$ and $\mathbf{P}_{\ell}\in \mathbb{R}^{H_\ell \times W_\ell \times \mathcal{C}_\ell}$ be the feature maps from backbone network and FPN, respectively, where $\ell$ is the layer index, $H_\ell \times W_\ell$ represents the size of feature maps and $\mathcal{C}_\ell$ denotes the number of feature channels.
In this work, following FCOS, five levels of multi-scale feature maps $\{\mathbf{P}_3,\mathbf{P}_4,\mathbf{P}_5,\mathbf{P}_6,\mathbf{P}_7\}$ are used, where $\mathbf{P}_3,\mathbf{P}_4$ and $\mathbf{P}_5$ are generated from $\mathbf{C}_3,\mathbf{C}_4$ and $\mathbf{C}_5$, respectively. While $\mathbf{P}_6$ and $\mathbf{P}_7$ are generated via up-sampling $\mathbf{P}_5$ and $\mathbf{P}_6$, respectively.
After obtaining the feature maps from FPN, the prediction heads with two fully convolutional subnets are employed to predict the categories and regressions for each location on feature maps.
Specifically, given a set of ground truths $\mathcal{G}=\{g_{i}\}_{i=1}^n$ in the input image, where $n$ denotes the total number of ground truths and $i$ is the index, each $g_i$ in $\mathcal{G}$ is represented by $g_i=(x_i,y_i,w_i,h_i,\theta_i,c_i)$, as shown in Fig.
~\ref{fig:groundtruth}, where $(x_i, y_i)$ is the coordinate of center point of $g_i$ and $w_i$, $h_i$ represent the width and height of $g_i$, $\theta_i\in[0^\circ,180^\circ)$ represents the counterclockwise angle between the y-axis and long side, and $c_i$ is the class label that the ground truth belongs to.
To generate prediction, we adopt $\mathcal{A}_\ell=\{\alpha^{\ell}_{j}\}_{j=1}^{H_\ell \times W_\ell}$ to denote the set of anchor points, which are the pixel-wise locations on $\mathbf{P}_{\ell}$. For each anchor point $\alpha^{\ell}_{j}$, it can be mapped back to the input image via:
\begin{equation}
\begin{split}
x_j = \lfloor \frac{s_\ell}{2} \rfloor+\tilde{x}^{\ell}_j{s_\ell}\\
y_j = \lfloor \frac{s_\ell}{2} \rfloor+\tilde{y}^{\ell}_j{s_\ell}
\end{split}
\end{equation}
where $(\tilde{x}^{\ell}_j,\tilde{y}^{\ell}_j)$ and $(x_j,y_j)$ represent the locations of $\alpha_j^\ell$ on $\mathbf{P}_\ell$ and the input image, respectively. $s_\ell$ denotes the stride on $\mathbf{P}_\ell$ and $\lfloor\cdot\rfloor$ denotes the round-down operator. Here, we directly regress the target bounding box at each $\alpha_j^{\ell}$. Note that, the label assignment strategies are often employed to decide whether each $\alpha_j^{\ell}$ is a positive or negative sample, and allow the network to select higher-quality samples for training.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.6\linewidth]{groundTruth.png}
\caption{Illustration of ground truth. $(x_i,y_i,w_i,h_i,\theta_i)$ are the center, width, height and angle of $g_i$, and $(x_j^*,y_j^*)$ denotes the offset between $\alpha_j^\ell$ and the center of $g_i$. Best viewed in colour.}
\label{fig:groundtruth}
\end{figure}
Then, if $\alpha_j^{\ell}$ is a positive sample and assigned to the ground truth, e.g., $g_i$, its class label $c^*_j$ equals to $c_i$ and it has a 5-dimensional vector $t_j^*=(x^*_j,y^*_j,w^*_j,h^*_j,\theta^*_j)$ being the regression targets for the location, where $x^*_j$ and $y^*_j$ are the coordinate offsets between the center point of $g_i$ and $\alpha_j^{\ell}$, as shown in Fig.~\ref{fig:groundtruth}, which can be formulated as:
\begin{equation}
\begin{split}
x^*_j = x_i - x_j\\
y^*_j = y_i - y_j
\end{split}
\end{equation}
$w^*_j$, $h^*_j$ and $\theta^*_j$ equal to the width, height and angle of $g_i$, respectively.
During inference, the prediction heads will directly generate classification and regression for each anchor point and output the classification scores and locations after post-processing, hence our proposed strategy is inference cost-free.
The above represents the pipeline of our simple and effective baseline detector. In the next subsections, the components of our proposed EARL, which are given in the lower green block of Fig. \ref{fig:Architecture}, will be introduced in details.
\subsection{Adaptive Scale Sampling Strategy}
\label{sec:3-2}
Different from conventional label assignment strategies, which select positive samples on certain feature levels according to the predefined scale constraint on each $\mathbf{P}_{\ell}$ for regression and classification, we propose a simple but novel sampling strategy to adaptively select the positive samples according to the scales of targets. Its implementation can be summarized as sequentially sampling from high-level feature maps, i.e., $\mathbf{P}_7$, to low-level feature maps, i.e., $\mathbf{P}_3$.
Specifically, for each $g_i$ in the input image, a determined function $\tau(\cdot)$ is used to indicate whether each $\alpha^{\ell}_j$ is inside a certain region of $g_i$ when mapped back to the input image, which can be simply understood as selecting samples inside the mask region on each $\mathbf{P}_\ell$, as shown in Fig.~\ref{fig:Architecture}.
Then, if $\alpha^{\ell}_j$ is judged to be inside the mask region by $\tau(\cdot)$, it will be selected into the candidate samples set $\mathcal{R}_\ell$, where $\tau(\cdot)$ will be discussed in more detail in Section \ref{sec:3-3}.
\begin{algorithm}
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\caption{Algorithm of ADS strategy}\label{alg:cap}
\Input{The set of ground truths $\mathcal{G}$ in the input image, the set of all anchor points $\mathcal{A} = \{\mathcal{A}_\ell\}_{\ell=3}^{\ell=7}$ among all feature levels and the sample number $k$}
\Output{The set of positive samples $\mathcal{P}$ and negative samples $\mathcal{N}$}
\BlankLine
\For{each ground truth $g_i \in \mathcal{G}$}{
\emph{Build an empty set for positive samples $\mathcal{P}\gets\varnothing$}\;
\For{layer index $\ell\gets7$ to $3$}{
\If{$k > 0 $}{
\emph{Build an empty set for candidate samples $\mathcal{R}_{\ell}\gets\varnothing$}\;
\For{each anchor point $\alpha_j^\ell \in \mathcal{A}_{\ell}$}{
\If{$\tau(\alpha_j^\ell, g_i)$ is True}{
\emph{$\mathcal{R}_{\ell} = \mathcal{R}_{\ell} \cup \alpha_j^\ell$}
}
}
\emph{$\tilde{n}_{\mathcal{S}}^\ell=min(k,\tilde{n}_{\mathcal{R}}^\ell)$}\;
\emph{$\mathcal{S}_{\ell}\gets$ select top-$\tilde{n}_{\mathcal{S}}^\ell$ anchor points from $\mathcal{R}_{\ell}$}\;
\If{$\tilde{n}_{\mathcal{S}}^\ell > 0$}{
\emph{$k = k - \tilde{n}_{\mathcal{S}}^\ell$}\;
\emph{$\mathcal{P}=\mathcal{P} \cup \mathcal{S}_{\ell}$}\;
}
}
}
}
\emph{$\mathcal{N}=\mathcal{A}-\mathcal{P}$}\;
\emph{\Return {$\mathcal{P}$, $\mathcal{N}$}}\;
\end{algorithm}
When the candidate sample selection is applied from $\mathbf{P}_7$ to $\mathbf{P}_3$, we only keep the $k$ samples that are closest to the center point of $g_i$ as positive samples based on L2 distance, which is denoted as top-$k$ in this paper. Here $k$ is a robust hyperparameter that ensures a balanced number of positive samples for each target as possible, which is suggested in \cite{retinanet} that unbalanced positive samples will decrease the performance, and L2 distance indicates the Euclidean Distance.
In details, let $\mathcal{S}_\ell$ denote selection set on $\mathbf{P}_\ell$ and $\tilde{n}_{\mathcal{S}}^\ell$ denote the number of selected samples on $\mathbf{P}_{\ell}$, which is formulated as:
\begin{equation}
\begin{split}
\tilde{n}_{\mathcal{S}}^\ell = min(k,\tilde{n}_{\mathcal{R}}^\ell)
\end{split}
\end{equation}
where $\tilde{n}_{\mathcal{R}}^\ell$ denotes the number of candidate samples on $\mathbf{P}_{\ell}$. Finally, we subtract $\tilde{n}_{\mathcal{S}}^\ell$ from $k$ and add the selected candidate samples to the positive samples set $\mathcal{P}$, after which we repeat this operation if $k$ is greater than 0. Algorithm~\ref{alg:cap} specifies the proposed method for positive sample selection.
As a result, this strategy can adaptively adjust the level of feature maps that the positive samples belong to according to the scales of targets, which allows detector to have a smooth and adaptive positive sample selection on all feature levels, as shown in Fig.~\ref{fig:ADS} (a). In addition, if an anchor point is assigned to multiple ground truths, we force it to be assigned to the one with the longest side. Besides, as shown in Fig.~\ref{fig:ADS} (b), if a ground truth escapes from any samples, we assign it with the closest unassigned anchor point, which ensures the information from each target can be used reasonably and partially alleviates the insufficient sampling problem.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{ADS.png}
\caption{Illustration of fixed-scale strategy (left column) and ADS strategy (right column) for sampling positive samples of targets with different scales. Best viewed in colour.}
\label{fig:ADS}
\end{figure}
In this way, the sequentially adaptive sampling operation of ADS can obtain a better distribution of positive samples when compared with the fixed-scale sampling strategy and the top-$k$ sampling strategy (i.e., ADS without the sequentially adaptive sampling operation). As shown in Fig.~\ref{fig:AdaptiveSampleAssignment}, fixed-scale sampling strategy constrains the level distribution of positive samples. Whereas top-$k$ sampling strategy selects samples clustered at the bottom feature levels and has a discontinuous level distribution. Therefore, above strategies will lead to a scale-level bias, especially when the target scale distribution is unbalanced, which is a common phenomenon in remote sensing images. While our method can alleviate the problem mentioned above.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{NewSamplingScale.png}
\caption{Illustration of the comparison about the number of positive samples by using different sampling strategies on multi-level feature maps for targets with different scales. The size of the red range on the multi-level feature maps represents the number of selected positive samples. (a) ADS strategy. (b) Fixed-scale sampling strategy. (c) Top-$k$ sampling strategy. Best viewed in colour.}
\label{fig:AdaptiveSampleAssignment}
\end{figure}
\subsection{Dynamic Elliptical Distribution aided Sampling Strategy}
\label{sec:3-3}
In the Section \ref{sec:3-2}, the Algorithm~\ref{alg:cap} has a determined function $\tau(\cdot)$ that takes the ground truth $g_i$ and the anchor point $\alpha_j^\ell$ as input, and returns the true or false value depending on whether $\alpha_j^\ell$ is inside the certain region of $g_i$ or not. Here, the determined function achieves spatial assignment by constraining the certain region. Concretely, most existing anchor-free detectors use the center-based sampling strategies to select training samples, and they often adopt the bounding box or the central area as sampling range, as shown in Fig.~\ref{fig:ellipseSampling}. However, these methods do not properly consider the characteristics of targets in remote sensing images, such as the large aspect ratios and arbitrary orientations, which either introduce more background noise or result in insufficient sampling.
To this end, inspired by the dynamic label assignment strategies\cite{zhu2020autoassign,huang2022general,yu2022object}, we propose dynamic elliptical distribution aided sampling (DED) strategy that not only considers the aspect ratios and angles of the rotate targets, but also mitigates the background noise during sample selection.
Therefore, we select the candidate samples by obeying a dynamic elliptical distribution based on the shapes of targets.
Suppose that given a ground truth $g_i$ and an anchor point $\alpha_j^\ell$, which is mapped back to the input image, the function $\tau(\cdot)$ can be formulated as:
\begin{equation}
\tau(\cdot) = \begin{cases}
\text{true} ,& \frac{a^2}{(\frac{1}{2}w_i)^2} + \frac{b^2}{(\frac{1}{2}h_i)^2} < \xi \\
\text{false} ,& \text{otherwise}
\end{cases}
\end{equation}
where $a$ and $b$ are calculated by the angle of $g_i$ and the offsets of coordinates between $g_i$ and $\alpha_j^\ell$, which are formulated respectively as:
\begin{equation}
\begin{split}
a = x^*_j\cos{\theta_i}+y^*_j\sin{\theta_i}\\
b = x^*_j\sin{\theta_i}-y^*_j\cos{\theta_i}
\end{split}
\end{equation}
the ratio factor $\xi\in[0.5,1)$ is an adaptive threshold determined by the target shape that controls the range of elliptical distribution to select higher-quality samples with less background noise. Here, $\xi$ is calculated as follows
\begin{equation}
\xi = 1 - \frac{min(h_i,w_i)}{2\times max(h_i,w_i)}
\end{equation}
In this way, the sampling distribution can be adjusted dynamically according to the shapes of targets. As shown in Fig.~\ref{fig:ellipseSampling}~(a), the sampling range tends to be a compacted circular distribution when the shape of ground truth is close to a square. When the aspect ratio of ground truth is extremely large, the sampling range will be close to the inner tangent ellipse, which are more suitable for this shape of targets.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{NewEllipseSampling.png}
\caption{Schematic diagram of the comparison of different sampling range in spatial assignment. Best viewed in colour.}
\label{fig:ellipseSampling}
\end{figure}
\subsection{Spatial Distance Weighting Module}
\label{sec:3-4}
To mitigate the effect of low-quality predictions generated from the anchor points far away from the center of ground truth, instead of employing extra prediction branch such as centerness branch\cite{tian2019fcos} or IoU prediction branch\cite{yu2022object}, we introduce an SDW module to weight the regression loss and classification probability during training, hence integrate the quality enhancement branch with loss function directly, and mitigate the increase of computation cost and training time.
For each positive sample $\alpha_j^\ell\in\mathcal{P}$, SDW module calculates weight $\mathcal{W}_j$ according to the scale of $g_i$ that it belongs to, and the L2 distance between $\alpha_j^\ell$ and center point of $g_i$, which is represented as $d_{\alpha_j^\ell,g_i}$. For classification, $\mathcal{W}_j$ is used to weight the probability for the corresponding category. For box regression, we first establish the regression loss and then multiply by $\mathcal{W}_j$. Here, $\mathcal{W}_j$ is formulated as:
\begin{equation}
\label{eq:7}
\mathcal{W}_j = \frac{1-\frac{d_{\alpha_j^\ell,g_i}}{m}}{max(1-\frac{d_{\mathcal{P},g_i}}{m})}
\end{equation}
where $m=max(\frac{h_i}{2},\frac{w_i}{2})$, and $d_{\mathcal{P},g_i}$ is the set of all the distance of positive samples for $g_i$. The weight is normalized as in Equation (\ref{eq:7}) to make sure that the highest weight is given to the closest positive sample, which is regarded as the sample with the highest quality.
Finally, $\mathcal{W}_j$ is applied to weight the classification probability and box regression loss. The total loss $\mathcal{L}_{total}$ is constructed with two types of loss, i.e., the classification loss $\mathcal{L}_{cls}$ and box regression loss $\mathcal{L}_{reg}$, denoted as
\begin{equation}
\begin{split}
\mathcal{L}_{total} = \frac{1}{N_{pos}}\sum_\ell\sum_j\mathcal{L}_{cls}(\hat{c_j}, c^*_j, \mathcal{W}_j) \\
+ \frac{\lambda}{N_{pos}}\sum_\ell\sum_j\mathds{1}_{\{\alpha_j^\ell\in\mathcal{P}\}}\mathcal{L}_{reg}(\hat{t_j}, t^*_j, \mathcal{W}_j)
\end{split}
\end{equation}
where $\lambda$ is the penalty parameter to balance these two types of loss, $N_{pos}$ denotes the number of positive samples, which is used for normalization. $\mathds{1}_{\{condition\}}$ is the indicator function, which equals to 1 if the condition is satisfied, otherwise 0. The classification loss is formulated as:
\begin{equation}
\label{eq:classification}
\mathcal{L}_{cls}(\hat{c_j}, c^*_j, \mathcal{W}_j) =
\begin{cases}
- \delta(\mathcal{W}_j-\hat{c_j})^{\gamma}log(\hat{c_j}) ,& {c^*_j=1} \\
- \delta(\hat{c_j})^{\gamma}log(1-\hat{c_j}) ,& \mbox{otherwise}
\end{cases}
\end{equation}
where $c^*_j\in\{0,1\}$ is the ground truth for classification and $\hat{c_j}\in[0,1]$ is the classification score from the network, $\delta$ and $\gamma$ are hyperparameters of the focal loss\cite{retinanet}. The box regression loss is formulated as:
\begin{equation}
\label{eq:13}
\begin{aligned}
\mathcal{L}_{reg}(\hat{t_j}, t^*_j, \mathcal{W}_j)&= \mathcal{W}_j \times Smooth_{L1}(||\hat{t_j} - t^*_j||)
\end{aligned}
\end{equation}
where $Smooth_{L1}(\cdot)$ is smooth L1 loss function as defined in \cite{girshick2015fast}.
\section{Experiments}
\label{sec:4}
In this section, we present the results of experiments on two publicly available and challenging datasets, DOTA\cite{xia2018dota} and HRSC2016\cite{liu2017high}, to verify the effectiveness of our proposed EARL. Details of the datasets, implementation details, evaluation metrics, experimental results, together with the comparison with state-of-the-art methods are presented in the following subsections.
\subsection{Datasets}
\textbf{DOTA} is a large-scale dataset for object detection in aerial images with OBB annotations for oriented targets, which contains 2,806 aerial images collected from different sensors and platforms, where the size of each image ranges from around $800\times800$ to $4000\times4000$ pixels. The fully annotated images contain 188,282 instances with a wide variety of scales, orientations and shapes, involving 15 common target categories: plane (PL), baseball diamond (BD), bridge (BR), ground track field (GTF), small vehicle (SV), large vehicle (LV), ship (SH), tennis court (TC), basketball court (BC), storage tank (ST), soccer ball field (SBF), roundabout (RA), harbor (HA), swimming pool (SP) and helicopter (HC).
In our experiments, DOTA is used following \cite{lin2019ienet}, where both training and validation sets are used for training, and the test set is employed for evaluating by submitting results to the evaluation server from the dataset provider. Due to the large size of images, following \cite{yang2019r3det}, all images from the datasets are cropped to \(600 \times 600\) pixels with a stride of 450 for memory efficiency, and resize to \(800 \times 800\) pixels during both training and inference.
\textbf{HRSC2016} is a challenging high resolution ship detection dataset which contains 1,061 aerial images from two scenarios including inshore and offshore ships. The training, validation and test sets include 436 images, 181 images and 444 images, respectively. The size of the images ranges from $300\times300$ to $1500\times900$ pixels. In our experiments, the images of HRSC2016 are resized to $800\times800$ pixels during training and inference.
\subsection{Experimental Details}
\subsubsection{Implementation Details}
In our experiments, for simplicity and efficiency, ResNet-50\cite{he2016deep} pretrained on ImageNet\cite{deng2009imagenet} is used as the backbone network, and FPN\cite{lin2017feature} is employed as the neck network unless specified. Then, the convolution weight that composes the prediction head is initialized with normal distribution, where the standard deviation is set to be 0.01. Also, the convolution bias is enabled and initialized with 0. We use the group normalization\cite{wu2018group} in the prediction head, which performs better than the batch normalization\cite{ioffe2015batch}. $\delta$ and $\gamma$ are set to be 0.25 and 2.0 respectively in focal loss\cite{retinanet}, and $\beta$ in the smooth L1 loss is set as 0.01.
During training, we use stochastic gradient descent (SGD)\cite{bottou2010large} as the optimization approach, and the hyperparameters of SGD, i.e., weight decay, momentum and gamma, are set to be $1\times10^{-4}$, 0.9 and 0.1, respectively. In addition, to stabilize the training process and yield better performance, we adopt a learning rate scheduler with the combination of warm-up training and multi-step learning rate. Here, a linear warm-up method is applied, where the warm-up factor is set as $1\times10^{-3}$ with 1000 iterations at the beginning of training, and only random flipping is employed for data augmentation. For experiments on DOTA dataset, our network is trained with four NVIDIA A100 GPUs with 64 images per batch. Furthermore, the detectors are trained for 40k iterations in total and the learning rate is initialized with $4.5\times10^{-2}$ but reduced by a factor of 10 at the iterations of 25k and 34k. While for experiments on HRSC2016, we train the models for 90k iterations, and the learning rate decays at 60k and 80k steps with a total of 16 images per batch. In addition, the running time is measured on a single GeForce RTX 3090 GPU with 24GB memory.
\subsubsection{Evaluation Metrics} For evaluating the accuracy of detectors, the mean Average Precision (mAP) with IoU threshold 0.5 is adopted, which is the widely used metric in object detection tasks. The Average Precision (AP) of each category is employed to validate the detection accuracy of different categories. The frames per second (fps) is used to evaluate the inference speed.
\subsection{Ablation Study}
In this section, a series of ablative experiments are conducted with DOTA dataset to illustrate the advantages of each proposed component in EARL. Here, the components of proposed EARL are indicated in abbreviated form, i.e., `-A' indicates ADS strategy, `-E' refers to DED strategy, and `-W' means SDW module. In addition, `w/o' is applied to indicate without specified modules. The overall results of the ablative experiments are presented in Table \ref{tab:1}. Specifically, the first row represents the result of our baseline detector, i.e., R-FCOS, followed by the ablative result of 71.79 in mAP, obtained by replacing the scale assignment with our ADS strategy. When employing our DED strategy, the mAP performance is improved slightly by 0.54, as shown in the third row. In addition, by adopting ADS with DED strategy, we achieve an improvement of 1.50 in mAP, as compared with baseline.
When combining all the components of EARL together, we can achieve the mAP performance of 72.87, as shown in the last row of Table \ref{tab:1}.
\begin{table}[t]
\centering
\caption{Ablative experiments and evaluations of the proposed EARL method on the DOTA dataset. The best result is highlighted in bold.}
\begin{tabular}{c|ccc|c}
\hline
Method & ADS & DED & SDW & mAP (\%) \bigstrut\\
\hline
R-FCOS (Baseline) & & & & 71.07 \bigstrut[t]\\
EARL (w/o -E,-W) & $\checkmark$ & & & 71.79 (+0.72) \\
EARL (w/o -A,-W) & & $\checkmark$ & & 71.61 (+0.54) \\
EARL (w/o -W) & $\checkmark$ & $\checkmark$ & & 72.57 (+1.50)\\
EARL & $\checkmark$ & $\checkmark$ & $\checkmark$ & \textbf{72.87 (+1.80)} \bigstrut[b]\\
\hline
\end{tabular}%
\label{tab:1}%
\end{table}
Moreover, in order to introduce the improvement for different categories, a more detailed experimental results for each category are provided in Table \ref{tab:1.1}, and the contribution of improvement from each component is discussed in detail as follows.
\begin{table*}[t]
\centering
\caption{More detailed mAP results of ablation experiments on the DOTA Dataset. The best result is highlighted in bold.}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|ccccccccccccccc|c}
\hline
Methods & PL & BD & BR & GTF & SV & LV & SH & TC & BC & ST & SBF & RA & HA & SP & HC & mAP (\%) \bigstrut\\
\hline
R-FCOS (Baseline) & 88.60 & 81.08 & 45.74 & 64.58 & 79.83 & 77.89 & 86.95 & 90.52 & 80.24 & 84.09 & 40.55 & 58.31 & 66.24 & 71.78 & 49.70 & 71.07 \bigstrut[t]\\
EARL (w/o -E,-W) & 89.87 & 80.57 & 46.22 & \textbf{65.61} & 80.51 & 78.35 & 87.39 & \textbf{90.76} & 78.41 & \textbf{86.28} & 45.10 & 61.42 & 65.08 & 69.56 & 51.70 & 71.79 \\
EARL (w/o -A,-W) & 89.12 & 77.67 & 46.98 & 65.06 & 80.27 & 78.03 & \textbf{87.68} & 90.54 & 82.65 & 85.23 & 47.67 & 60.58 & \textbf{67.08} & 70.98 & 44.64 & 71.61 \\
EARL (w/o -W) & \textbf{90.05} & \textbf{81.16} & 47.01 & 65.17 & 80.39 & 79.85 & 87.10 & 90.37 & \textbf{83.11} & 86.17 & 44.32 & 65.41 & 66.82 & 71.10 & 50.56 & 72.57 \\
EARL & 89.76 & 78.79 & \textbf{47.01} & 65.20 & \textbf{80.98} & \textbf{79.99} & 87.33 & 90.74 & 79.17 & 86.23 & \textbf{49.09} & \textbf{65.87} & 65.75 & \textbf{71.86} & \textbf{55.21} & \textbf{72.87} \bigstrut[b]\\
\hline
\end{tabular}}
\label{tab:1.1}%
\end{table*}%
\subsubsection{Effect of Adaptive Scale Sampling Strategy}
The core of our proposed method is the ADS strategy, which resolves the insufficient sampling problem for targets with extreme scales.
For example, when PL and SV appear in one image, which have a huge difference in size, as illustrated in Fig.~\ref{fig:effectADS}, samples are restricted to be assigned on specific level of feature maps, and yields feature insufficient for various size of target. By employing ADS strategy, samples are selected more balancely for both PL and SV on each level of FPN, hence features with more various scales can be learned compared with the baseline method.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{EffectADS.png}
\caption{Comparison of sample numbers of PL and SV on each level of feature maps, where (a) and (b) are the visualization of sampling results, (c) and (d) are the statistics of the percent of samples on each level for PL and SV, respectively. Best viewed in colour.}
\label{fig:effectADS}
\end{figure}
Quantitatively, besides the improvement of 0.72 in mAP, the proposed ADS strategy boosts the AP performance of the targets with extreme scales, such as SV, GTF and SBF by 0.68, 1.03 and 4.55 compared with the baseline method, respectively, as shown in rows 1 and 2 of Table \ref{tab:1.1}, which indicates the effectiveness of ADS. However, although ADS strategy is able to obtain a better scale-level distribution, it introduces noise in higher feature maps for specific categories, e.g. BD, BC, HA, SP, and brings down the accuracy slightly. In order to mitigate the noise, DED with SDW are applied and discussed in the following subsection.
\subsubsection{Effect of Dynamic Elliptical Distribution aided Sampling Strategy}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{NewissueWithFcos.png}
\caption{Illustration of the positive samples selected by obeying different sampling range, such as (a) Bounding Box, (b) Central Area and (c) DED strategy on DOTA dataset. Best viewed in colour.}
\label{fig:exp_ed}
\end{figure}
DED strategy is proposed to alleviate noise from background in rectangle bounding box, and to obtain orientation and aspect ratio properties of targets. As shown in Fig.~\ref{fig:exp_ed} (a), when samples are assigned according to the rectangle bounding box (BB), as introduced in our baseline method, a part of samples are selected in area of background instead of within the scope of target. Although this problem is mitigated by assigning samples in the central area of rectangle bounding box in central area (CA) sampling strategy, samples are only gathered in the central part of targets, hence the features at the ends of the target are dropped, and yields the poor performance in orientation and the aspect ratio regression, as given in Fig.~\ref{fig:exp_ed} (b). When employing the proposed DED method, the distribution of samples can be controlled by fitting the shape of ellipse dynamically, hence is more suitable for oriented object detection tasks, and achieves the improvement of 0.54 and 1.00 in mAP when compared with BB and CA, respectively, in Table~\ref{tab:samplingrange}.
\begin{table}[t]
\centering
\caption{Comparison with different sampling range on the DOTA dataset. The best result is highlighted in bold.}
\begin{tabular}{cccccc}
\hline
Sampling Range & BB & CA & DED \bigstrut\\ \hline
mAP (\%) & 71.07 & 70.61 & \textbf{71.61} \bigstrut\\ \hline
\end{tabular}
\label{tab:samplingrange}
\end{table}
In addition, from rows 1 and 3 of Table~\ref{tab:1.1}, the AP of targets, such as BR and HA, is improved by 1.24 and 0.84, respectively, which further proves the effectiveness of our DED strategy for fitting targets with large aspect ratios. Moreover, when combined with DED strategy, the proposed ADS strategy achieved a further improvement of 0.78 in mAP, as shown in rows 2 and 4 of Table~\ref{tab:1.1}, due to its ability of mitigating the noise from selected samples for targets with various size, which further unlocking the potential of ADS strategy.
\subsubsection{Effect of Spatial Distance Weighting Module}
Based on DED, SDW module is proposed to further mitigate the influence of low-quality samples on detection performance. By weighting anchor points based on their distance from the center of ground truth, the samples with higher quality are given higher weights, hence can achieve samples with more feature information and avoid the influence from anchor points at the edge of the elliptical sampling area. In this way, as shown in rows 4 and 5 of Table \ref{tab:1.1}, EARL achieves mAP of 72.87 by employing SDW module, and it is obvious that detector with SDW module can significantly improve AP for difficult categories, such as SBF and HC, by 4.77 and 4.65. Moreover, to further comprehend the effect of SDW module, the qualitative comparison of detection results on DOTA dataset is displayed in Fig.~\ref{fig:effectSDW}, where the detector trained with SDW module can improve the detection performance by reducing the false alarms and increasing the recall rate in specific scenes. This indicates that the proposed SDW can focus on higher-quality samples thus obtain the accurate and appropriate features adaptively.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{effectSDW.png}
\caption{Effect of the SDW module on detection results on DOTA dataset. (Top row) Results of the method without SDW module. (Bottom row) Results of the method with SDW module. Best viewed in colour.}
\label{fig:effectSDW}
\end{figure}
\subsubsection{Visual Explanation}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{feature.png}
\caption{Gradient-based class-discriminative saliency maps visualization comparison of the detectors trained with different label assignment strategies. Best viewed in colour.}
\label{fig:feature}
\end{figure}
To further understand the effectiveness of each proposed component of EARL, we display the gradient-based class-discriminative saliency maps\cite{chattopadhay2018grad} to visualize the feature heatmaps of detectors with different label assignment strategies, as illustrated in Fig.~\ref{fig:feature}. As seen in row 1 of (a) and (c), EARL (w/o -A,-W) constricts the sampling range based on R-FCOS, so that less attention is paid to the background. From rows 2 and 3 of (b) and (c), ATSS focuses too much on the central region resulting in its failure to learn sufficient features and can hardly reflect the orientation characteristics, especially, for targets with large aspect ratios such as harbour and bridge. Row 3 of (a) and (d) show that EARL (w/o -E,-W) can extract richer information of targets with extreme scales. However, due to its ability to obtain training samples from multi-level feature maps, it also introduces a certain amount of background noise.
Whereas, it can be seen that from (d) and (e), using DED strategy can mitigate the impact of noise and release the potential of ADS strategy. In addition, as shown in (e) and (f), SDW module further reduces the influence of low-quality samples, and allows the detector to learn better feature representation of the targets, which can accurately focus on the target regions on the images.
\subsection{Evaluation of Hyperparameters}
\subsubsection{Sample Number $k$}
\begin{table}[t]
\centering
\caption{Evaluation of the various sample number $k$ in ADS strategy. The best result is highlighted in bold.}
\begin{tabular}{cccccccc}
\hline
$k$ & 12 & 13 & 14 & 15 & 16 & 17 & 18 \bigstrut\\
\hline
mAP (\%) & 71.52 & 72.36 & 72.42 & \textbf{72.87} & 72.05 & 71.87 & 71.93 \bigstrut\\
\hline
\end{tabular}%
\label{tab:2}%
\end{table}
Due to the importance of sample number $k$, which is introduced in the ADS strategy, we test from 12 samples to 18 samples per target on the DOTA dataset. The evaluation results are shown in Table \ref{tab:2} to illustrate its impact on performance. When $k$=12, although the quality of each selected sample is guaranteed, the number of positive samples is inadequate for the model to learn feature information, so the mAP has only reached 71.52. The performance of the model gradually improves with the increases of $k$, and the best performance is achieved when $k$=15. However, as $k$ continues to increase, more background noise are included since samples with further distance from the center of ground-truth are selected, so the mAP score drops from 72.87 to 71.93 when $k$=18. The performance is robust since the overall improvement keeps continuous with $k$ increases from 12 to 18. This result is crucial when use our ADS strategy to other datasets, which yields relatively stable detection performance and can be further improved by simply changing the value of $k$ to suit different categories.
\subsubsection{Ratio Factor $\xi$}
The ratio factor $\xi$ in DED strategy is introduced to ensure the quality of the selected samples by controlling the distribution of samples.
\begin{table}[t]
\centering
\caption{Evaluation of using various ratios factor $\xi$ in DED strategy. The best result is highlighted in bold.}
\begin{tabular}{cccccc}
\hline
$\xi$ & 0.4 & 0.6 & 0.8 & 1.0 & adaptive \bigstrut\\ \hline
mAP (\%) & 71.99 & 72.49 & 72.16 & 71.76 & \textbf{72.87} \bigstrut\\ \hline
\end{tabular}
\label{tab:4}
\end{table}
As given in Table \ref{tab:4}, different values for $\xi$ are tested to seek the optimum value, and display the influence from the diverse ratio factor. We can see that when $\xi$=0.4, the sampling range is small enough to avoid noise, and the mAP achieves 71.99. With the increase of $\xi$, the detector is able to learn more information from the selected high-quality samples, hence the detection performance increases and reaches a peak of 72.49 when $\xi$=0.6. However, the performance starts to drop, when $\xi$ further increases because it is difficult to guarantee the quality of selected samples, especially those far from the center of ground truth bounding box. To solve this problem, we set the value of $\xi$ to be adaptive according to the shapes of targets. It can be seen in Table \ref{tab:4} that the mAP of the detector with adaptive $\xi$ increases by 0.38 compared with the peak value when $\xi$=0.6, which demonstrates the effectiveness of our proposed adaptive method, and illustrates that the key to improve detection performance by spatial assignment is to obtain more positive samples while reducing the effect of noise.
\begin{table*}[tbp]
\centering
\caption{Performance comparison with different state-of-the-art methods on DOTA dataset. HG104 means Hourglass 104. $\ast$ indicates multi-scale training and testing. ATSS indicates the lite version. $\dagger$ means our re-implementation on our baseline detector. The best result is highlighted in black and red bold in anchor-based and anchor-free method, respectively.}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|ccccccccccccccc|cc}
\hline
Methods & Backbone & PL & BD & BR & GTF & SV & LV & SH & TC & BC & ST & SBF & RA & HA & SP & HC & mAP (\%) & fps\\
\hline
\textit{\textbf{Anchor-based} Methods:} & & & & & & & & & & & & & & & & & & \\
RoI-Transformer \cite{ding2019learning}$^{\ast}$ & Res101 & 88.53 & 77.91 & 37.63 & 74.08 & 66.53 & 62.97 & 66.57 & 90.50 & 79.46 & 76.75 & 59.04 & 56.73 & 62.54 & 61.29 & 55.56 & 67.74 & 7.8 \\
ICN\cite{azimi2018towards}$^{\ast}$ & Res101 & 81.36 & 74.30 & 47.70 & 70.32 & 64.89 & 67.82 & 69.98 & 90.76 & 79.06 & 78.20 & 53.64 & 62.90 & 67.02 & 64.17 & 50.23 & 68.16 & - \\
CenterMap-Net\cite{wang2020learning} & Res50 & 89.02 & 80.56 & 49.41 & 61.98 & 77.99 & 74.19 & 83.74 & 89.44 & 78.01 & 83.52 & 47.64 & 65.93 & 63.68 & 67.07 & 61.59 & 71.59 & - \\
SCRDet\cite{yang2019scrdet}$^{\ast}$ & Res101 & \textbf{89.98} & 80.65 & 52.09 & 68.36 & 68.36 & 60.32 & 72.41 & \textbf{90.85} & \textbf{87.94} & 86.86 & 65.02 & 66.68 & 66.25 & 68.24 & 65.21 & 72.61 & 9.5 \\
Gliding-Ver.\cite{xu2020gliding}$^{\ast}$ & Res101 & 89.64 & \textbf{85.00} & 52.26 & 77.34 & 73.01 & 73.14 & 86.82 & 90.74 & 79.02 & 86.81 & 59.55 & \textbf{70.91} & 72.94 & 70.86 & 57.32 & 75.02 & 13.1 \\
R$^3$Det\cite{yang2019r3det} & Res50 & 89.02 & 75.47 & 43.86 & 65.84 & 75.83 & 73.44 & 86.03 & 90.57 & 81.11 & 82.84 & 55.57 & 59.10 & 56.57 & 70.31 & 50.45 & 70.40 & - \\
S$^2$A-Net\cite{han2021align} & Res50 & 89.11 & 82.84 & 48.37 & 71.11 & 78.11 & 78.39 & 87.25 & 90.83 & 84.90 & 85.64 & 60.36 & 62.60 & 65.26 & 69.13 & 57.94 & 74.12 & \textbf{17.6} \\
S$^2$A-Net\cite{han2021align}$^{\ast}$ & Res50 & 88.89 & 83.60 & \textbf{57.74} & \textbf{81.95} & \textbf{79.94} & \textbf{83.19} & \textbf{89.11} & 90.78 & 84.87 & \textbf{87.81} & \textbf{70.30} & 68.25 & \textbf{78.30} & \textbf{77.01} & \textbf{69.58} & \textbf{79.42} & \textbf{17.6} \\
Hou \cite{hou2022refined} & Res101 & 89.32 & 76.05 & 50.33 & 70.25 & 76.44 & 79.45 & 86.02 & 90.84 & 82.80 & 82.50 & 58.17 & 62.46 & 67.38 & 71.93 & 45.52 & 72.63 & - \\
Hou \cite{hou2022refined}$^{\ast}$ & Res101 & 88.69 & 79.41 & 52.26 & 65.51 & 74.72 & 80.83 & 87.42 & 90.77 & 84.31 & 83.36 & 62.64 & 58.14 & 66.95 & 72.32 & 69.34 & 74.44 & - \\
\hline
\textit{\textbf{Anchor-free} Methods}: & & & & & & & & & & & & & & & & & & \\
TOSO\cite{feng2020toso} & Res101 & 80.17 & 65.59 & 39.82 & 39.95 & 49.71 & 65.01 & 53.58 & 81.45 & 44.66 & 78.51 & 48.85 & 56.73 & 64.40 & 64.24 & 36.75 & 57.92 & 16.9 \\
PIoU\cite{chen2020piou} & DLA34\cite{yu2018deep} & 80.90 & 69.70 & 24.10 & 60.20 & 38.30 & 64.40 & 64.80 & \textcolor{red}{\textbf{90.90}} & 77.20 & 70.40 & 46.50 & 37.10 & 57.10 & 61.90 & 64.00 & 60.50 & - \\
IENet\cite{lin2019ienet} & Res101 & 88.15 & 71.38 & 34.26 & 51.78 & 63.78 & 65.63 & 71.61 & 90.11 & 71.07 & 73.63 & 37.62 & 41.52 & 48.07 & 60.53 & 49.53 & 61.24 & 16.9 \\
Axis-Learning\cite{xiao2020axis} & Res101 & 79.53 & 77.15 & 38.59 & 61.15 & 67.53 & 70.49 & 76.30 & 89.66 & 79.07 & 83.53 & 47.27 & 61.01 & 56.28 & 66.06 & 36.05 & 65.98 & 14.1 \\
ATSS\cite{zhang2020bridging}$^{\dagger}$ & Res50 & 88.47 & 80.05 & 47.27 & 60.65 & 79.85 & 78.80 & 87.41 & 90.75 & 77.37 & 85.24 & 43.22 & 60.80 & 66.52 & 70.95 & 41.82 & 70.61 & - \\
DRN\cite{pan2020dynamic} & HG104\cite{newell2016stacked} & 88.91 & 80.22 & 43.52 & 63.35 & 73.48 & 70.69 & 84.94 & 90.14 & \textcolor{red}{\textbf{83.85}} & 84.11 & 50.12 & 58.41 & 67.62 & 68.60 & 52.50 & 70.70 & - \\
O$^2$-DNet\cite{wei2020oriented} & HG104\cite{newell2016stacked} & 89.31 & 82.14 & 47.33 & 61.21 & 71.32 & 74.03 & 78.62 & 90.76 & 82.23 & 81.36 & \textcolor{red}{\textbf{60.93}} & 60.17 & 58.21 & 66.98 & \textcolor{red}{\textbf{64.03}} & 71.04 & - \\
BBAVectors\cite{yi2021oriented} & Res101 & 88.35 & 79.96 & \textcolor{red}{\textbf{50.69}} & 62.18 & 78.43 & 78.98 & 87.94 & 90.85 & 83.58 & 84.35 & 54.13 & 60.24 & 65.22 & 64.28 & 55.70 & 72.32 & 18.4 \\
\textbf{EARL (Ours)} & Res50 & 89.76 & 78.79 & 47.01 & 65.20 & 80.98 & 79.99 & 87.33 & 90.74 & 79.17 & 86.23 & 49.09 & 65.87 & 65.75 & 71.86 & 55.21 & 72.87 & \textcolor{red}{\textbf{30.8}} \\
\textbf{EARL (Ours)}$^{\ast}$ & Res50 & \textcolor{red}{\textbf{90.13}} & \textcolor{red}{\textbf{83.90}} & 47.19 & \textcolor{red}{\textbf{72.17}} & \textcolor{red}{\textbf{81.54}} & \textcolor{red}{\textbf{84.26}} & \textcolor{red}{\textbf{88.24}} & 90.69 & 79.10 & \textcolor{red}{\textbf{86.71}} & 60.47 & \textcolor{red}{\textbf{72.21}} & \textcolor{red}{\textbf{71.26}} & \textcolor{red}{\textbf{73.81}} & 58.80 & \textcolor{red}{\textbf{76.03}} & \textcolor{red}{\textbf{30.8}} \\
\hline
\end{tabular}}
\label{tab:5}
\end{table*}
\begin{figure*}[tbp]
\centering
\includegraphics[width=0.9\linewidth]{DOTA-Label.png}
\caption{Visualization of detection results on the DOTA dataset with our method. Best viewed in colour.}
\label{fig:visualDOTA}
\end{figure*}
\begin{figure*}[tbp]
\centering
\includegraphics[width=1.0\linewidth]{exp_ads_percent.png}
\caption{Histogram of the percentage of the number of positive samples selected at different feature levels using different label assignment strategies on the DOTA dataset. Best viewed in colour.}
\label{fig:exp_ads}
\end{figure*}
\subsection{Comparison with the State-of-the-Art}
\subsubsection{Results on DOTA}
To evaluate the performance of the proposed method, EARL is compared with both anchor-based and anchor-free state-of-the-art detectors on DOTA dataset, and the experimental results are shown in Table \ref{tab:5}. It can be observed that anchor-based detectors can usually achieve higher performance on DOTA dataset. However, anchor-based detectors usually have limitations in inference speed due to complex structures and IoU calculations, which results in lower fps. EARL with ResNet50 as the backbone and FPN as the neck achieves mAP of 72.87 and inference speed of 30.8 fps, which is comparable in accuracy and much faster than anchor-based detectors, even without any tricks.
When compared with R$^3$Det with the same backbone, the mAP of our EARL increases 2.47, though R$^3$Det employs anchors. Moreover, EARL outperforms S$^2$A-Net in the aspect of speed by 13.2 fps, which is one of the state-of-the-art one-stage anchor-based detectors. In addition, EARL surpasses the most recent one-stage anchor-based oriented object detector (i.e., Hou \cite{hou2022refined}) by 0.24 in mAP, even though our method is based on anchor-free framework and only uses ResNet50 as backbone.
When compared with the anchor-free detectors, being integrated with the simple backbone network (i.e., ResNet50), our EARL method can provide better detection performance, even though they use the more complex backbone networks, i.e., Hourglass104, ResNet101. Specifically, when applying multi-scale augmentation in experiments, our method achieves the mAP of 76.03 and obtains the best AP results for small targets, such as SV and LV of 81.54 and 84.26, respectively, which further demonstrates the effectiveness of our method for targets with extreme scales. The visualization of detection results with our method is displayed in Fig.~\ref{fig:visualDOTA}.
In addition, to further illustrate that our method can obtain a better scale-level distribution, as shown in Fig.~\ref{fig:exp_ads}, the proposed EARL is compared with the most relevant assignment strategy, ATSS, together with
the baseline method R-FCOS on DOTA dataset, by comparing the distribution of the positive samples among different feature levels statistically. It can be seen that R-FCOS and ATSS achieve scale assignment according to the predefined scale constraints, which causes uneven sample distribution leading to the scale-level bias, especially for targets with extreme scales, such as GTF and SV. While our method with ADS strategy ensures that targets with different scales can adaptively assign training samples from all feature levels, which alleviates the scale-level bias thus improves detection performance. As a result, EARL achieves the mAP of 72.87 and has significant improvements by 1.80 and 2.26 on DOTA dataset over R-FCOS and ATSS, respectively, as shown in Table \ref{tab:1.1} and Table \ref{tab:5}.
Furthermore, the AP performance in Table \ref{tab:5} also confirm our method has the improvements on targets with both extreme large and small sizes, such as GTF, SBF, SV and LV. The above analysis proves that our EARL is advantageous for object detection in remote sensing images.
\begin{table}[tb]
\scriptsize
\centering
\caption{Performance comparison with different state-of-the-art methods on the HRSC2016 dataset. The best result is highlighted in bold.}
\begin{tabular}{cccc}
\hline
Methods & Backbone & mAP$_{07}$ (\%) & mAP$_{12}$ (\%)\\
\hline
\textit{\textbf{Anchor-based} Methods}\\
R$^2$CNN\cite{jiang2017r2cnn} & Res101 & 73.1 & 79.7 \\
RRPN\cite{ma2018arbitrary} & Res101 & 79.1 & 85.6\\
RetinaNet-H\cite{yang2019r3det} & Res101 & 82.9 & 89.2\\
R$^2$PN\cite{zhang2018toward} & VGG16 & 79.6 & - \\
RoI-Transformer\cite{ding2019learning} & Res101 & 86.2 & - \\
Gliding-Ver.\cite{xu2020gliding} & Res101 & 88.2 & - \\
DAL\cite{ming2020dynamic} & Res50 & 88.6 & - \\
CenterMap-Net\cite{wang2020learning} & Res50 & - & 92.8 \\
\hline
\textit{\textbf{Anchor-free} Methods}\\
RC2\cite{liu2017rotated} & VGG16 & 75.7 & - \\
Axis-Learning\cite{xiao2020axis} & Res101 & 78.2 & - \\
TOSO\cite{feng2020toso} & Res101 & 79.3 & - \\
BBAVectors\cite{yi2021oriented} & Res101 & 88.6 & - \\
DRN\cite{pan2020dynamic} & H104 & - & 92.7 \\
\textbf{EARL (Ours)} & Res50 & \textbf{89.0} & \textbf{93.0}\\
\hline
\end{tabular}
\label{tab:hrsc2016}
\end{table}
\begin{figure}[tb]
\centering
\includegraphics[width=1.0\linewidth]{hrsc2016.png}
\caption{Visualization of detection results on the HRSC2016 dataset with our method. Best viewed in colour.}
\label{fig:visualHRSC}
\end{figure}
\subsubsection{Results on HRSC2016} To further evaluate the robustness of the proposed method, we setup experiments on the HRSC2016 dataset, and the results are listed in Table \ref{tab:hrsc2016}. It is obvious that compared with both anchor-based and anchor-free methods, our EARL achieves the best performance of 89.0 and 93.0 in mAP under the Pascal VOC2007 and VOC2012 metrics\cite{everingham2010pascal}, respectively. The detection results on HRSC2016 dataset are illustrated in Fig.~\ref{fig:visualHRSC}, which also confirm our method is robust for detecting targets with large aspect ratios and further demonstrate the advancement of our method.
\section{Conclusion}
In this paper, we presented a novel and efficient strategy for label assignment, namely EARL. An adaptive scale sampling strategy was proposed to solve the problem of targets with large variation of scales, especially in remote sensing images, by selecting samples from suitable and continuous multi-level feature maps. Considering the large aspect ratio of the target and to obtain a more accurate and reasonable sampling range while reducing the influence of background noise, a dynamic elliptical distribution aided sampling method was designed, which was adaptively controlled by the target size. To learn high quality information from selected training samples, a spatial distance weighting module was developed. The experimental results demonstrated the effectiveness of our proposed method when being integrated with simple structure without any tricks. Our EARL is so simple that it has the potential to be easily combined with different orientation detectors to achieve better performance.
\bibliographystyle{IEEEtran}
\section{Introduction}
\label{sec:1}
\IEEEPARstart{O}{bject} detection in remote sensing images, which aims to determine the locations and categories of the interested targets, has become one of the most common and potential image interpretation steps for numerous applications\cite{deng2017toward,wang2019enhancing,shi2020orientation,yu2020orientation}, such as maritime rescuing\cite{varga2022seadronessee}, urban planning\cite{ma2018mobile} and traffic management\cite{asha2018vehicle}.
\begin{figure}[!tbp]
\centering
\includegraphics[width=1.0\linewidth]{imbalanced.png}
\caption{As shown in gray blocks, generic label assignment strategy selects samples on certain level of feature maps according to the scales of targets. This strategy will lead to the insufficient sampling for targets with extreme scales and aspect ratios. Best viewed in colour.}
\label{fig:Imbalance}
\vspace{-0.105cm}
\end{figure}
In the past few years, substantial methods have made significant progress for generic object detection based on CNN\cite{ren2015faster,redmon2016you,retinanet,tian2019fcos,duan2019centernet}, which have greatly promoted the development of object detection in remote sensing images\cite{wang2020learning,dai2022ace,cheng2022anchor} and attract wide attention from researchers.
However, object detection in remote sensing images is still facing challenges from the characteristics of targets, such as arbitrary orientations, large aspect ratios and large scale variations. To address above challenges, current detectors for remote sensing images have been designed to enhance orientation prediction based on generic detectors, such as adding extra prediction branch \cite{lin2019ienet,qin2021mrdet,yu2022object} and alleviating feature misalignment \cite{yu2022object,han2021align,yang2019r3det}.
Moreover, feature pyramid network (FPN)\cite{lin2017feature,hou2022refined,wang2020frpnet} is often adopted to represent multi-scale features, where the upper feature maps have semantic information and are often employed to detect large targets, whereas lower feature maps have more fine-gained details and are utilized to detect small targets \cite{zhu2019fsaf}. However, it usually faces the crucial issue of selecting training samples at reasonable pixel-wise locations (spatial assignment) on different feature levels (scale assignment), namely label assignment \cite{zhu2020autoassign}.
Many label assignment methods have been proposed for object detection due to their important role in determining positive or negative samples, which directly and significantly influences the performance \cite{zhang2020bridging}, such as maximum intersection over union (IoU) based strategies \cite{qin2021mrdet,yang2019r3det,yang2019scrdet} and center-based strategies \cite{cheng2021anchor,shi2021canet,feng2020toso}. However, they are barely designed specifically for remote sensing images and neglect the essential characteristics, e.g., the bird-eye-view and various scales of targets, which brings tough problems in object detection.
Concretely, when combined with FPN, current scale assignment strategies always heuristically assign samples in upper feature maps to large targets, and samples in lower feature maps to small targets\cite{zhu2019fsaf}. However, this heuristic often causes imbalance sampling, especially in remote sensing images. On the one hand, few positive samples will be assigned to targets with extreme scales and large aspect ratios. As shown in Fig.~\ref{fig:Imbalance}, samples will be treated as negatives due to the
heuristic scale constrains, even if these samples fall within the sampling range in multi-level feature maps. On the other hand, samples for small targets tend to be dominated in number, as illustrated in Fig.~\ref{fig:Imbalance}, where vehicles are far more numerous than planes. As a result, using these strategies will lead to a flood of positive samples assigned to the lower level of feature maps, thus overwhelm the sufficient attention to targets with other scales and introduce the scale-level bias. Therefore, a more flexible method is required for sample selection on multi-level feature maps.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.7\linewidth]{background.png}
\caption{Illustration of using different ranges as positive spatial area for sampling. (Left column) Oriented bounding box. (Right column) Central area.}
\label{fig:backgroundissue}
\end{figure}
In addition, current detectors usually use rectangle bounding box as the sampling range in spatial assignment, and tend to introduce background noise as shown in Fig.~\ref{fig:backgroundissue}~(Left column). While center-based strategies, such as center sampling \cite{zhang2020bridging}, can improve the quality of positive samples by only sampling central area of targets. However, it still rarely takes into account the large aspect ratios and arbitrary orientation of targets in remote sensing images, which leads to many foreground samples being misclassified as negatives and reduces the detection accuracy for rotated targets with large aspect ratios, as shown in Fig.~\ref{fig:backgroundissue}~(Right column). Hence, a more reliable spatial assignment strategy is necessary for detectors to capture positive samples with higher quality.
To address aforementioned limitations in both scale assignment and spatial assignment, in this paper, we propose an Elliptical Distribution aided Adaptive Rotation Label Assignment (EARL) strategy to select positive samples with higher quality on multi-level feature maps dynamically, thus improve the detection performance. Specifically, we designed a novel and intuitively simple strategy, namely adaptive scale sampling (ADS) strategy, to avoid insufficient sampling and the imbalanced distribution of positive samples on multi-level feature maps.
In addition, a dynamic elliptical distribution aided sampling (DED) strategy, together with a spatial distance weighting (SDW) module, is proposed to consistently exploit orientation and shape properties of targets, and mitigate the effect of low-quality samples.
Besides, the proposed EARL is deployed on a simple anchor-free detector, i.e., FCOS, with minimal modification to succinctly represent the effectiveness. Extensive experiments on popular remote sensing datasets, such as DOTA and HRSC2016, demonstrate the effectiveness and the superiority of our proposed EARL.
The contributions of this work are summarized as follows:
\begin{enumerate}
\item An adaptive scale sampling (ADS) strategy is proposed to dynamically select positive samples among all feature levels, which avoids insufficient sampling and alleviates the scale-level bias.
\item A dynamic elliptical distribution aided sampling (DED) strategy is proposed to select samples by taking full advantage of orientation and shape properties of targets.
\item A spatial distance weighting (SDW) module is employed to integrate the quality enhancement with loss function, and further mitigate the effect of low-quality positive samples.
\item The proposed label assignment strategy is evaluated by being integrated with simple detection architecture, without bells and whistles, EARL outperforms the state-of-the-art anchor-free methods, and achieves comparable performance with anchor-based algorithms, while keeping the efficiency of anchor-free detectors.
\end{enumerate}
The remainder of the paper is organized as follows: The related work is reviewed in Section \ref{sec:2}. Then, the details of the proposed method are introduced in Section \uppercase\expandafter{\romannumeral3}. Experiments and analyses are described and discussed in Section \uppercase\expandafter{\romannumeral4}. Finally, conclusions are summarized in Section \uppercase\expandafter{\romannumeral5}.
\section{Related Work}
\label{sec:2}
\subsection{Oriented Object Detection in Remote Sensing Images}
Due to its wide range of application scenarios, oriented object detection in remote sensing images has developed rapidly. Similar to generic detectors, recent methods can be divided into two categories, i.e., anchor-based detectors and anchor-free detectors. To detect targets with arbitrary orientations, some anchor-based detectors\cite{liu2017rotated,zhang2018toward,yang2018automatic}, which are built on the two-stage framework, i.e., Faster R-CNN\cite{ren2015faster}, densely preset multiple anchors with different scales, aspect ratios and angles for better regression while introducing heavy anchor-related computations. After that, RoI-Transformer\cite{ding2019learning} proposed an efficient region of interests (RoI) learner to transform horizontal proposals to rotated ones. While Gliding Vertex\cite{xu2020gliding} described oriented targets by the horizontal bounding boxes (HBB), which further avoided numerous anchors with multiple angles. To improve accuracy and efficiency, R$^3$Det\cite{yang2019r3det} and S$^2$A-Net\cite{han2021align} followed the one-stage schema and designed network modules to alleviate the feature misalignment issue. However, these methods still take a long time during training and inference\cite{yu2022object}.
To maintain high efficiency, anchor-free detectors are designed to avoid the usage of anchors. For instance, TOSO\cite{feng2020toso} directly regressed the surrounding HBB and transformation parameters to present rotated targets. IENet\cite{lin2019ienet} was built upon the classic one-stage anchor-free framework, i.e., FCOS, adding an additional branch with the interactive embranchment module to regress angles. Later, BBAVector\cite{yi2021oriented}, O$^2$-DNet\cite{wei2020oriented} and ACE\cite{dai2022ace} employed a simplified architecture and designed new oriented bounding box (OBB) representation methods.
However, these anchor-free detectors still not perform state-of-the-art results compared with anchor-based detectors, and most of them are more likely to improve the architecture while not considering the training strategy, especially the label assignment strategy, which has been proven to be the essential difference between these two kinds of detectors\cite{zhang2020bridging}.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.9\linewidth]{Architecture.png}
\caption{Overview of the proposed EARL. The upper gray block shows the baseline architecture of R-FCOS. The lower green block presents our proposed EARL strategy. The DED strategy is employed based on the shape of the target and generates foreground masks on each level of feature maps outlining area to select candidate samples. The candidate samples will be sorted from the highest to lowest level of feature maps where the samples in the same level will sorted by the distance from the center point of ground truth bounding box. Only the first top-$k$ samples are selected as positive samples (coloured points). Best viewed in colour.}
\label{fig:Architecture}
\end{figure*}
\subsection{Label Assignment Strategies}
Label assignment is a crucial component for CNN-based detectors to learn distribution of targets by determining positive or negative samples. Meanwhile, as the current detectors commonly adopt FPN\cite{lin2017feature} to alleviate the scale variation, label assignment needs to simultaneously assign labels to samples at different spatial locations (spatial assignment) on different feature levels (scale assignment)\cite{zhu2020autoassign}. Most anchor-based detectors\cite{ren2015faster,retinanet,qin2021mrdet} assign the labels to anchors among multi-level feature maps by comparing the preset thresholds and IoU between anchors and ground truth bounding boxes. However, this strategy always involves many hyperparameters to adjust the scales and aspect ratios of anchors based on the datasets\cite{zhang2021learning}. While anchor-free methods\cite{tian2019fcos,kong2020foveabox,xiao2020axis} always select samples inside the ground truth bounding box or the central area of targets for spatial assignment, and choose samples on certain level of feature maps according to the predefined scale constraint for scale assignment, therefore these can be regarded as fixed-scale sampling strategy. However, above strategies, which depend on the heuristic rules, may not be optimal enough. Thus, recent works provide dynamic label assignment strategies, which allow models to learn to select samples by itself. For instance, FSAF\cite{zhu2019fsaf} dynamically assigned targets to the suitable feature levels based on computed loss. ATSS\cite{zhang2020bridging} proposed an adaptive training sample selection method by adjusting the IoU threshold according to the statistics of targets, where ATSS also proposed the center sampling strategy to improve the quality of positive samples. Autoassign\cite{zhu2020autoassign} used two weighting modules to adjust the category-specific prior distribution according to the appearances of targets.
As for oriented object detection, GGHL\cite{huang2022general} used a 2-dimensional Gaussian heatmaps to define positive samples according to the size and direction properties of targets. For stable optimization, FSDet\cite{yu2022object} used a soft assignment mechanism to weight the training samples. Inspired by the excellent works for dynamical label assignment mechanism in object detection, the proposed method focuses on designing an adaptive label assignment strategy for orientation anchor-free detectors in remote sensing images, where targets have arbitrary orientations and large variations in scales and aspect ratios.
\section{The Proposed EARL Method}
\label{sec:3}
In this section, the architecture of our baseline anchor-free detector is firstly presented. Then the three components of the proposed EARL, i.e., ADS strategy, DED strategy and SDW module, are introduced in detail.
\subsection{Architecture of the Baseline Anchor-free Detector}
\label{sec:3-1}
To demonstrate the effectiveness of the proposed label assignment strategy, in this paper, a simple but classical anchor-free architecture, namely, R-FCOS, is employed as our baseline detector.
Compared with FCOS\cite{tian2019fcos}, our R-FCOS removes centerness branch to be more compact and achieves more simplicity and efficiency, which only contains classification branch and regression branch for oriented object detection.
As shown in Fig.~\ref{fig:Architecture}, the upper gray block shows the components of our baseline detector, which consists of the backbone network, FPN and prediction heads. Let $\mathbf{C}_\ell$ and $\mathbf{P}_{\ell}\in \mathbb{R}^{H_\ell \times W_\ell \times \mathcal{C}_\ell}$ be the feature maps from backbone network and FPN, respectively, where $\ell$ is the layer index, $H_\ell \times W_\ell$ represents the size of feature maps and $\mathcal{C}_\ell$ denotes the number of feature channels.
In this work, following FCOS, five levels of multi-scale feature maps $\{\mathbf{P}_3,\mathbf{P}_4,\mathbf{P}_5,\mathbf{P}_6,\mathbf{P}_7\}$ are used, where $\mathbf{P}_3,\mathbf{P}_4$ and $\mathbf{P}_5$ are generated from $\mathbf{C}_3,\mathbf{C}_4$ and $\mathbf{C}_5$, respectively. While $\mathbf{P}_6$ and $\mathbf{P}_7$ are generated via up-sampling $\mathbf{P}_5$ and $\mathbf{P}_6$, respectively.
After obtaining the feature maps from FPN, the prediction heads with two fully convolutional subnets are employed to predict the categories and regressions for each location on feature maps.
Specifically, given a set of ground truths $\mathcal{G}=\{g_{i}\}_{i=1}^n$ in the input image, where $n$ denotes the total number of ground truths and $i$ is the index, each $g_i$ in $\mathcal{G}$ is represented by $g_i=(x_i,y_i,w_i,h_i,\theta_i,c_i)$, as shown in Fig.
~\ref{fig:groundtruth}, where $(x_i, y_i)$ is the coordinate of center point of $g_i$ and $w_i$, $h_i$ represent the width and height of $g_i$, $\theta_i\in[0^\circ,180^\circ)$ represents the counterclockwise angle between the y-axis and long side, and $c_i$ is the class label that the ground truth belongs to.
To generate prediction, we adopt $\mathcal{A}_\ell=\{\alpha^{\ell}_{j}\}_{j=1}^{H_\ell \times W_\ell}$ to denote the set of anchor points, which are the pixel-wise locations on $\mathbf{P}_{\ell}$. For each anchor point $\alpha^{\ell}_{j}$, it can be mapped back to the input image via:
\begin{equation}
\begin{split}
x_j = \lfloor \frac{s_\ell}{2} \rfloor+\tilde{x}^{\ell}_j{s_\ell}\\
y_j = \lfloor \frac{s_\ell}{2} \rfloor+\tilde{y}^{\ell}_j{s_\ell}
\end{split}
\end{equation}
where $(\tilde{x}^{\ell}_j,\tilde{y}^{\ell}_j)$ and $(x_j,y_j)$ represent the locations of $\alpha_j^\ell$ on $\mathbf{P}_\ell$ and the input image, respectively. $s_\ell$ denotes the stride on $\mathbf{P}_\ell$ and $\lfloor\cdot\rfloor$ denotes the round-down operator. Here, we directly regress the target bounding box at each $\alpha_j^{\ell}$. Note that, the label assignment strategies are often employed to decide whether each $\alpha_j^{\ell}$ is a positive or negative sample, and allow the network to select higher-quality samples for training.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.6\linewidth]{groundTruth.png}
\caption{Illustration of ground truth. $(x_i,y_i,w_i,h_i,\theta_i)$ are the center, width, height and angle of $g_i$, and $(x_j^*,y_j^*)$ denotes the offset between $\alpha_j^\ell$ and the center of $g_i$. Best viewed in colour.}
\label{fig:groundtruth}
\end{figure}
Then, if $\alpha_j^{\ell}$ is a positive sample and assigned to the ground truth, e.g., $g_i$, its class label $c^*_j$ equals to $c_i$ and it has a 5-dimensional vector $t_j^*=(x^*_j,y^*_j,w^*_j,h^*_j,\theta^*_j)$ being the regression targets for the location, where $x^*_j$ and $y^*_j$ are the coordinate offsets between the center point of $g_i$ and $\alpha_j^{\ell}$, as shown in Fig.~\ref{fig:groundtruth}, which can be formulated as:
\begin{equation}
\begin{split}
x^*_j = x_i - x_j\\
y^*_j = y_i - y_j
\end{split}
\end{equation}
$w^*_j$, $h^*_j$ and $\theta^*_j$ equal to the width, height and angle of $g_i$, respectively.
During inference, the prediction heads will directly generate classification and regression for each anchor point and output the classification scores and locations after post-processing, hence our proposed strategy is inference cost-free.
The above represents the pipeline of our simple and effective baseline detector. In the next subsections, the components of our proposed EARL, which are given in the lower green block of Fig. \ref{fig:Architecture}, will be introduced in details.
\subsection{Adaptive Scale Sampling Strategy}
\label{sec:3-2}
Different from conventional label assignment strategies, which select positive samples on certain feature levels according to the predefined scale constraint on each $\mathbf{P}_{\ell}$ for regression and classification, we propose a simple but novel sampling strategy to adaptively select the positive samples according to the scales of targets. Its implementation can be summarized as sequentially sampling from high-level feature maps, i.e., $\mathbf{P}_7$, to low-level feature maps, i.e., $\mathbf{P}_3$.
Specifically, for each $g_i$ in the input image, a determined function $\tau(\cdot)$ is used to indicate whether each $\alpha^{\ell}_j$ is inside a certain region of $g_i$ when mapped back to the input image, which can be simply understood as selecting samples inside the mask region on each $\mathbf{P}_\ell$, as shown in Fig.~\ref{fig:Architecture}.
Then, if $\alpha^{\ell}_j$ is judged to be inside the mask region by $\tau(\cdot)$, it will be selected into the candidate samples set $\mathcal{R}_\ell$, where $\tau(\cdot)$ will be discussed in more detail in Section \ref{sec:3-3}.
\begin{algorithm}
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\caption{Algorithm of ADS strategy}\label{alg:cap}
\Input{The set of ground truths $\mathcal{G}$ in the input image, the set of all anchor points $\mathcal{A} = \{\mathcal{A}_\ell\}_{\ell=3}^{\ell=7}$ among all feature levels and the sample number $k$}
\Output{The set of positive samples $\mathcal{P}$ and negative samples $\mathcal{N}$}
\BlankLine
\For{each ground truth $g_i \in \mathcal{G}$}{
\emph{Build an empty set for positive samples $\mathcal{P}\gets\varnothing$}\;
\For{layer index $\ell\gets7$ to $3$}{
\If{$k > 0 $}{
\emph{Build an empty set for candidate samples $\mathcal{R}_{\ell}\gets\varnothing$}\;
\For{each anchor point $\alpha_j^\ell \in \mathcal{A}_{\ell}$}{
\If{$\tau(\alpha_j^\ell, g_i)$ is True}{
\emph{$\mathcal{R}_{\ell} = \mathcal{R}_{\ell} \cup \alpha_j^\ell$}
}
}
\emph{$\tilde{n}_{\mathcal{S}}^\ell=min(k,\tilde{n}_{\mathcal{R}}^\ell)$}\;
\emph{$\mathcal{S}_{\ell}\gets$ select top-$\tilde{n}_{\mathcal{S}}^\ell$ anchor points from $\mathcal{R}_{\ell}$}\;
\If{$\tilde{n}_{\mathcal{S}}^\ell > 0$}{
\emph{$k = k - \tilde{n}_{\mathcal{S}}^\ell$}\;
\emph{$\mathcal{P}=\mathcal{P} \cup \mathcal{S}_{\ell}$}\;
}
}
}
}
\emph{$\mathcal{N}=\mathcal{A}-\mathcal{P}$}\;
\emph{\Return {$\mathcal{P}$, $\mathcal{N}$}}\;
\end{algorithm}
When the candidate sample selection is applied from $\mathbf{P}_7$ to $\mathbf{P}_3$, we only keep the $k$ samples that are closest to the center point of $g_i$ as positive samples based on L2 distance, which is denoted as top-$k$ in this paper. Here $k$ is a robust hyperparameter that ensures a balanced number of positive samples for each target as possible, which is suggested in \cite{retinanet} that unbalanced positive samples will decrease the performance, and L2 distance indicates the Euclidean Distance.
In details, let $\mathcal{S}_\ell$ denote selection set on $\mathbf{P}_\ell$ and $\tilde{n}_{\mathcal{S}}^\ell$ denote the number of selected samples on $\mathbf{P}_{\ell}$, which is formulated as:
\begin{equation}
\begin{split}
\tilde{n}_{\mathcal{S}}^\ell = min(k,\tilde{n}_{\mathcal{R}}^\ell)
\end{split}
\end{equation}
where $\tilde{n}_{\mathcal{R}}^\ell$ denotes the number of candidate samples on $\mathbf{P}_{\ell}$. Finally, we subtract $\tilde{n}_{\mathcal{S}}^\ell$ from $k$ and add the selected candidate samples to the positive samples set $\mathcal{P}$, after which we repeat this operation if $k$ is greater than 0. Algorithm~\ref{alg:cap} specifies the proposed method for positive sample selection.
As a result, this strategy can adaptively adjust the level of feature maps that the positive samples belong to according to the scales of targets, which allows detector to have a smooth and adaptive positive sample selection on all feature levels, as shown in Fig.~\ref{fig:ADS} (a). In addition, if an anchor point is assigned to multiple ground truths, we force it to be assigned to the one with the longest side. Besides, as shown in Fig.~\ref{fig:ADS} (b), if a ground truth escapes from any samples, we assign it with the closest unassigned anchor point, which ensures the information from each target can be used reasonably and partially alleviates the insufficient sampling problem.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{ADS.png}
\caption{Illustration of fixed-scale strategy (left column) and ADS strategy (right column) for sampling positive samples of targets with different scales. Best viewed in colour.}
\label{fig:ADS}
\end{figure}
In this way, the sequentially adaptive sampling operation of ADS can obtain a better distribution of positive samples when compared with the fixed-scale sampling strategy and the top-$k$ sampling strategy (i.e., ADS without the sequentially adaptive sampling operation). As shown in Fig.~\ref{fig:AdaptiveSampleAssignment}, fixed-scale sampling strategy constrains the level distribution of positive samples. Whereas top-$k$ sampling strategy selects samples clustered at the bottom feature levels and has a discontinuous level distribution. Therefore, above strategies will lead to a scale-level bias, especially when the target scale distribution is unbalanced, which is a common phenomenon in remote sensing images. While our method can alleviate the problem mentioned above.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{NewSamplingScale.png}
\caption{Illustration of the comparison about the number of positive samples by using different sampling strategies on multi-level feature maps for targets with different scales. The size of the red range on the multi-level feature maps represents the number of selected positive samples. (a) ADS strategy. (b) Fixed-scale sampling strategy. (c) Top-$k$ sampling strategy. Best viewed in colour.}
\label{fig:AdaptiveSampleAssignment}
\end{figure}
\subsection{Dynamic Elliptical Distribution aided Sampling Strategy}
\label{sec:3-3}
In the Section \ref{sec:3-2}, the Algorithm~\ref{alg:cap} has a determined function $\tau(\cdot)$ that takes the ground truth $g_i$ and the anchor point $\alpha_j^\ell$ as input, and returns the true or false value depending on whether $\alpha_j^\ell$ is inside the certain region of $g_i$ or not. Here, the determined function achieves spatial assignment by constraining the certain region. Concretely, most existing anchor-free detectors use the center-based sampling strategies to select training samples, and they often adopt the bounding box or the central area as sampling range, as shown in Fig.~\ref{fig:ellipseSampling}. However, these methods do not properly consider the characteristics of targets in remote sensing images, such as the large aspect ratios and arbitrary orientations, which either introduce more background noise or result in insufficient sampling.
To this end, inspired by the dynamic label assignment strategies\cite{zhu2020autoassign,huang2022general,yu2022object}, we propose dynamic elliptical distribution aided sampling (DED) strategy that not only considers the aspect ratios and angles of the rotate targets, but also mitigates the background noise during sample selection.
Therefore, we select the candidate samples by obeying a dynamic elliptical distribution based on the shapes of targets.
Suppose that given a ground truth $g_i$ and an anchor point $\alpha_j^\ell$, which is mapped back to the input image, the function $\tau(\cdot)$ can be formulated as:
\begin{equation}
\tau(\cdot) = \begin{cases}
\text{true} ,& \frac{a^2}{(\frac{1}{2}w_i)^2} + \frac{b^2}{(\frac{1}{2}h_i)^2} < \xi \\
\text{false} ,& \text{otherwise}
\end{cases}
\end{equation}
where $a$ and $b$ are calculated by the angle of $g_i$ and the offsets of coordinates between $g_i$ and $\alpha_j^\ell$, which are formulated respectively as:
\begin{equation}
\begin{split}
a = x^*_j\cos{\theta_i}+y^*_j\sin{\theta_i}\\
b = x^*_j\sin{\theta_i}-y^*_j\cos{\theta_i}
\end{split}
\end{equation}
the ratio factor $\xi\in[0.5,1)$ is an adaptive threshold determined by the target shape that controls the range of elliptical distribution to select higher-quality samples with less background noise. Here, $\xi$ is calculated as follows
\begin{equation}
\xi = 1 - \frac{min(h_i,w_i)}{2\times max(h_i,w_i)}
\end{equation}
In this way, the sampling distribution can be adjusted dynamically according to the shapes of targets. As shown in Fig.~\ref{fig:ellipseSampling}~(a), the sampling range tends to be a compacted circular distribution when the shape of ground truth is close to a square. When the aspect ratio of ground truth is extremely large, the sampling range will be close to the inner tangent ellipse, which are more suitable for this shape of targets.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{NewEllipseSampling.png}
\caption{Schematic diagram of the comparison of different sampling range in spatial assignment. Best viewed in colour.}
\label{fig:ellipseSampling}
\end{figure}
\subsection{Spatial Distance Weighting Module}
\label{sec:3-4}
To mitigate the effect of low-quality predictions generated from the anchor points far away from the center of ground truth, instead of employing extra prediction branch such as centerness branch\cite{tian2019fcos} or IoU prediction branch\cite{yu2022object}, we introduce an SDW module to weight the regression loss and classification probability during training, hence integrate the quality enhancement branch with loss function directly, and mitigate the increase of computation cost and training time.
For each positive sample $\alpha_j^\ell\in\mathcal{P}$, SDW module calculates weight $\mathcal{W}_j$ according to the scale of $g_i$ that it belongs to, and the L2 distance between $\alpha_j^\ell$ and center point of $g_i$, which is represented as $d_{\alpha_j^\ell,g_i}$. For classification, $\mathcal{W}_j$ is used to weight the probability for the corresponding category. For box regression, we first establish the regression loss and then multiply by $\mathcal{W}_j$. Here, $\mathcal{W}_j$ is formulated as:
\begin{equation}
\label{eq:7}
\mathcal{W}_j = \frac{1-\frac{d_{\alpha_j^\ell,g_i}}{m}}{max(1-\frac{d_{\mathcal{P},g_i}}{m})}
\end{equation}
where $m=max(\frac{h_i}{2},\frac{w_i}{2})$, and $d_{\mathcal{P},g_i}$ is the set of all the distance of positive samples for $g_i$. The weight is normalized as in Equation (\ref{eq:7}) to make sure that the highest weight is given to the closest positive sample, which is regarded as the sample with the highest quality.
Finally, $\mathcal{W}_j$ is applied to weight the classification probability and box regression loss. The total loss $\mathcal{L}_{total}$ is constructed with two types of loss, i.e., the classification loss $\mathcal{L}_{cls}$ and box regression loss $\mathcal{L}_{reg}$, denoted as
\begin{equation}
\begin{split}
\mathcal{L}_{total} = \frac{1}{N_{pos}}\sum_\ell\sum_j\mathcal{L}_{cls}(\hat{c_j}, c^*_j, \mathcal{W}_j) \\
+ \frac{\lambda}{N_{pos}}\sum_\ell\sum_j\mathds{1}_{\{\alpha_j^\ell\in\mathcal{P}\}}\mathcal{L}_{reg}(\hat{t_j}, t^*_j, \mathcal{W}_j)
\end{split}
\end{equation}
where $\lambda$ is the penalty parameter to balance these two types of loss, $N_{pos}$ denotes the number of positive samples, which is used for normalization. $\mathds{1}_{\{condition\}}$ is the indicator function, which equals to 1 if the condition is satisfied, otherwise 0. The classification loss is formulated as:
\begin{equation}
\label{eq:classification}
\mathcal{L}_{cls}(\hat{c_j}, c^*_j, \mathcal{W}_j) =
\begin{cases}
- \delta(\mathcal{W}_j-\hat{c_j})^{\gamma}log(\hat{c_j}) ,& {c^*_j=1} \\
- \delta(\hat{c_j})^{\gamma}log(1-\hat{c_j}) ,& \mbox{otherwise}
\end{cases}
\end{equation}
where $c^*_j\in\{0,1\}$ is the ground truth for classification and $\hat{c_j}\in[0,1]$ is the classification score from the network, $\delta$ and $\gamma$ are hyperparameters of the focal loss\cite{retinanet}. The box regression loss is formulated as:
\begin{equation}
\label{eq:13}
\begin{aligned}
\mathcal{L}_{reg}(\hat{t_j}, t^*_j, \mathcal{W}_j)&= \mathcal{W}_j \times Smooth_{L1}(||\hat{t_j} - t^*_j||)
\end{aligned}
\end{equation}
where $Smooth_{L1}(\cdot)$ is smooth L1 loss function as defined in \cite{girshick2015fast}.
\section{Experiments}
\label{sec:4}
In this section, we present the results of experiments on two publicly available and challenging datasets, DOTA\cite{xia2018dota} and HRSC2016\cite{liu2017high}, to verify the effectiveness of our proposed EARL. Details of the datasets, implementation details, evaluation metrics, experimental results, together with the comparison with state-of-the-art methods are presented in the following subsections.
\subsection{Datasets}
\textbf{DOTA} is a large-scale dataset for object detection in aerial images with OBB annotations for oriented targets, which contains 2,806 aerial images collected from different sensors and platforms, where the size of each image ranges from around $800\times800$ to $4000\times4000$ pixels. The fully annotated images contain 188,282 instances with a wide variety of scales, orientations and shapes, involving 15 common target categories: plane (PL), baseball diamond (BD), bridge (BR), ground track field (GTF), small vehicle (SV), large vehicle (LV), ship (SH), tennis court (TC), basketball court (BC), storage tank (ST), soccer ball field (SBF), roundabout (RA), harbor (HA), swimming pool (SP) and helicopter (HC).
In our experiments, DOTA is used following \cite{lin2019ienet}, where both training and validation sets are used for training, and the test set is employed for evaluating by submitting results to the evaluation server from the dataset provider. Due to the large size of images, following \cite{yang2019r3det}, all images from the datasets are cropped to \(600 \times 600\) pixels with a stride of 450 for memory efficiency, and resize to \(800 \times 800\) pixels during both training and inference.
\textbf{HRSC2016} is a challenging high resolution ship detection dataset which contains 1,061 aerial images from two scenarios including inshore and offshore ships. The training, validation and test sets include 436 images, 181 images and 444 images, respectively. The size of the images ranges from $300\times300$ to $1500\times900$ pixels. In our experiments, the images of HRSC2016 are resized to $800\times800$ pixels during training and inference.
\subsection{Experimental Details}
\subsubsection{Implementation Details}
In our experiments, for simplicity and efficiency, ResNet-50\cite{he2016deep} pretrained on ImageNet\cite{deng2009imagenet} is used as the backbone network, and FPN\cite{lin2017feature} is employed as the neck network unless specified. Then, the convolution weight that composes the prediction head is initialized with normal distribution, where the standard deviation is set to be 0.01. Also, the convolution bias is enabled and initialized with 0. We use the group normalization\cite{wu2018group} in the prediction head, which performs better than the batch normalization\cite{ioffe2015batch}. $\delta$ and $\gamma$ are set to be 0.25 and 2.0 respectively in focal loss\cite{retinanet}, and $\beta$ in the smooth L1 loss is set as 0.01.
During training, we use stochastic gradient descent (SGD)\cite{bottou2010large} as the optimization approach, and the hyperparameters of SGD, i.e., weight decay, momentum and gamma, are set to be $1\times10^{-4}$, 0.9 and 0.1, respectively. In addition, to stabilize the training process and yield better performance, we adopt a learning rate scheduler with the combination of warm-up training and multi-step learning rate. Here, a linear warm-up method is applied, where the warm-up factor is set as $1\times10^{-3}$ with 1000 iterations at the beginning of training, and only random flipping is employed for data augmentation. For experiments on DOTA dataset, our network is trained with four NVIDIA A100 GPUs with 64 images per batch. Furthermore, the detectors are trained for 40k iterations in total and the learning rate is initialized with $4.5\times10^{-2}$ but reduced by a factor of 10 at the iterations of 25k and 34k. While for experiments on HRSC2016, we train the models for 90k iterations, and the learning rate decays at 60k and 80k steps with a total of 16 images per batch. In addition, the running time is measured on a single GeForce RTX 3090 GPU with 24GB memory.
\subsubsection{Evaluation Metrics} For evaluating the accuracy of detectors, the mean Average Precision (mAP) with IoU threshold 0.5 is adopted, which is the widely used metric in object detection tasks. The Average Precision (AP) of each category is employed to validate the detection accuracy of different categories. The frames per second (fps) is used to evaluate the inference speed.
\subsection{Ablation Study}
In this section, a series of ablative experiments are conducted with DOTA dataset to illustrate the advantages of each proposed component in EARL. Here, the components of proposed EARL are indicated in abbreviated form, i.e., `-A' indicates ADS strategy, `-E' refers to DED strategy, and `-W' means SDW module. In addition, `w/o' is applied to indicate without specified modules. The overall results of the ablative experiments are presented in Table \ref{tab:1}. Specifically, the first row represents the result of our baseline detector, i.e., R-FCOS, followed by the ablative result of 71.79 in mAP, obtained by replacing the scale assignment with our ADS strategy. When employing our DED strategy, the mAP performance is improved slightly by 0.54, as shown in the third row. In addition, by adopting ADS with DED strategy, we achieve an improvement of 1.50 in mAP, as compared with baseline.
When combining all the components of EARL together, we can achieve the mAP performance of 72.87, as shown in the last row of Table \ref{tab:1}.
\begin{table}[t]
\centering
\caption{Ablative experiments and evaluations of the proposed EARL method on the DOTA dataset. The best result is highlighted in bold.}
\begin{tabular}{c|ccc|c}
\hline
Method & ADS & DED & SDW & mAP (\%) \bigstrut\\
\hline
R-FCOS (Baseline) & & & & 71.07 \bigstrut[t]\\
EARL (w/o -E,-W) & $\checkmark$ & & & 71.79 (+0.72) \\
EARL (w/o -A,-W) & & $\checkmark$ & & 71.61 (+0.54) \\
EARL (w/o -W) & $\checkmark$ & $\checkmark$ & & 72.57 (+1.50)\\
EARL & $\checkmark$ & $\checkmark$ & $\checkmark$ & \textbf{72.87 (+1.80)} \bigstrut[b]\\
\hline
\end{tabular}%
\label{tab:1}%
\end{table}
Moreover, in order to introduce the improvement for different categories, a more detailed experimental results for each category are provided in Table \ref{tab:1.1}, and the contribution of improvement from each component is discussed in detail as follows.
\begin{table*}[t]
\centering
\caption{More detailed mAP results of ablation experiments on the DOTA Dataset. The best result is highlighted in bold.}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|ccccccccccccccc|c}
\hline
Methods & PL & BD & BR & GTF & SV & LV & SH & TC & BC & ST & SBF & RA & HA & SP & HC & mAP (\%) \bigstrut\\
\hline
R-FCOS (Baseline) & 88.60 & 81.08 & 45.74 & 64.58 & 79.83 & 77.89 & 86.95 & 90.52 & 80.24 & 84.09 & 40.55 & 58.31 & 66.24 & 71.78 & 49.70 & 71.07 \bigstrut[t]\\
EARL (w/o -E,-W) & 89.87 & 80.57 & 46.22 & \textbf{65.61} & 80.51 & 78.35 & 87.39 & \textbf{90.76} & 78.41 & \textbf{86.28} & 45.10 & 61.42 & 65.08 & 69.56 & 51.70 & 71.79 \\
EARL (w/o -A,-W) & 89.12 & 77.67 & 46.98 & 65.06 & 80.27 & 78.03 & \textbf{87.68} & 90.54 & 82.65 & 85.23 & 47.67 & 60.58 & \textbf{67.08} & 70.98 & 44.64 & 71.61 \\
EARL (w/o -W) & \textbf{90.05} & \textbf{81.16} & 47.01 & 65.17 & 80.39 & 79.85 & 87.10 & 90.37 & \textbf{83.11} & 86.17 & 44.32 & 65.41 & 66.82 & 71.10 & 50.56 & 72.57 \\
EARL & 89.76 & 78.79 & \textbf{47.01} & 65.20 & \textbf{80.98} & \textbf{79.99} & 87.33 & 90.74 & 79.17 & 86.23 & \textbf{49.09} & \textbf{65.87} & 65.75 & \textbf{71.86} & \textbf{55.21} & \textbf{72.87} \bigstrut[b]\\
\hline
\end{tabular}}
\label{tab:1.1}%
\end{table*}%
\subsubsection{Effect of Adaptive Scale Sampling Strategy}
The core of our proposed method is the ADS strategy, which resolves the insufficient sampling problem for targets with extreme scales.
For example, when PL and SV appear in one image, which have a huge difference in size, as illustrated in Fig.~\ref{fig:effectADS}, samples are restricted to be assigned on specific level of feature maps, and yields feature insufficient for various size of target. By employing ADS strategy, samples are selected more balancely for both PL and SV on each level of FPN, hence features with more various scales can be learned compared with the baseline method.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{EffectADS.png}
\caption{Comparison of sample numbers of PL and SV on each level of feature maps, where (a) and (b) are the visualization of sampling results, (c) and (d) are the statistics of the percent of samples on each level for PL and SV, respectively. Best viewed in colour.}
\label{fig:effectADS}
\end{figure}
Quantitatively, besides the improvement of 0.72 in mAP, the proposed ADS strategy boosts the AP performance of the targets with extreme scales, such as SV, GTF and SBF by 0.68, 1.03 and 4.55 compared with the baseline method, respectively, as shown in rows 1 and 2 of Table \ref{tab:1.1}, which indicates the effectiveness of ADS. However, although ADS strategy is able to obtain a better scale-level distribution, it introduces noise in higher feature maps for specific categories, e.g. BD, BC, HA, SP, and brings down the accuracy slightly. In order to mitigate the noise, DED with SDW are applied and discussed in the following subsection.
\subsubsection{Effect of Dynamic Elliptical Distribution aided Sampling Strategy}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{NewissueWithFcos.png}
\caption{Illustration of the positive samples selected by obeying different sampling range, such as (a) Bounding Box, (b) Central Area and (c) DED strategy on DOTA dataset. Best viewed in colour.}
\label{fig:exp_ed}
\end{figure}
DED strategy is proposed to alleviate noise from background in rectangle bounding box, and to obtain orientation and aspect ratio properties of targets. As shown in Fig.~\ref{fig:exp_ed} (a), when samples are assigned according to the rectangle bounding box (BB), as introduced in our baseline method, a part of samples are selected in area of background instead of within the scope of target. Although this problem is mitigated by assigning samples in the central area of rectangle bounding box in central area (CA) sampling strategy, samples are only gathered in the central part of targets, hence the features at the ends of the target are dropped, and yields the poor performance in orientation and the aspect ratio regression, as given in Fig.~\ref{fig:exp_ed} (b). When employing the proposed DED method, the distribution of samples can be controlled by fitting the shape of ellipse dynamically, hence is more suitable for oriented object detection tasks, and achieves the improvement of 0.54 and 1.00 in mAP when compared with BB and CA, respectively, in Table~\ref{tab:samplingrange}.
\begin{table}[t]
\centering
\caption{Comparison with different sampling range on the DOTA dataset. The best result is highlighted in bold.}
\begin{tabular}{cccccc}
\hline
Sampling Range & BB & CA & DED \bigstrut\\ \hline
mAP (\%) & 71.07 & 70.61 & \textbf{71.61} \bigstrut\\ \hline
\end{tabular}
\label{tab:samplingrange}
\end{table}
In addition, from rows 1 and 3 of Table~\ref{tab:1.1}, the AP of targets, such as BR and HA, is improved by 1.24 and 0.84, respectively, which further proves the effectiveness of our DED strategy for fitting targets with large aspect ratios. Moreover, when combined with DED strategy, the proposed ADS strategy achieved a further improvement of 0.78 in mAP, as shown in rows 2 and 4 of Table~\ref{tab:1.1}, due to its ability of mitigating the noise from selected samples for targets with various size, which further unlocking the potential of ADS strategy.
\subsubsection{Effect of Spatial Distance Weighting Module}
Based on DED, SDW module is proposed to further mitigate the influence of low-quality samples on detection performance. By weighting anchor points based on their distance from the center of ground truth, the samples with higher quality are given higher weights, hence can achieve samples with more feature information and avoid the influence from anchor points at the edge of the elliptical sampling area. In this way, as shown in rows 4 and 5 of Table \ref{tab:1.1}, EARL achieves mAP of 72.87 by employing SDW module, and it is obvious that detector with SDW module can significantly improve AP for difficult categories, such as SBF and HC, by 4.77 and 4.65. Moreover, to further comprehend the effect of SDW module, the qualitative comparison of detection results on DOTA dataset is displayed in Fig.~\ref{fig:effectSDW}, where the detector trained with SDW module can improve the detection performance by reducing the false alarms and increasing the recall rate in specific scenes. This indicates that the proposed SDW can focus on higher-quality samples thus obtain the accurate and appropriate features adaptively.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{effectSDW.png}
\caption{Effect of the SDW module on detection results on DOTA dataset. (Top row) Results of the method without SDW module. (Bottom row) Results of the method with SDW module. Best viewed in colour.}
\label{fig:effectSDW}
\end{figure}
\subsubsection{Visual Explanation}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{feature.png}
\caption{Gradient-based class-discriminative saliency maps visualization comparison of the detectors trained with different label assignment strategies. Best viewed in colour.}
\label{fig:feature}
\end{figure}
To further understand the effectiveness of each proposed component of EARL, we display the gradient-based class-discriminative saliency maps\cite{chattopadhay2018grad} to visualize the feature heatmaps of detectors with different label assignment strategies, as illustrated in Fig.~\ref{fig:feature}. As seen in row 1 of (a) and (c), EARL (w/o -A,-W) constricts the sampling range based on R-FCOS, so that less attention is paid to the background. From rows 2 and 3 of (b) and (c), ATSS focuses too much on the central region resulting in its failure to learn sufficient features and can hardly reflect the orientation characteristics, especially, for targets with large aspect ratios such as harbour and bridge. Row 3 of (a) and (d) show that EARL (w/o -E,-W) can extract richer information of targets with extreme scales. However, due to its ability to obtain training samples from multi-level feature maps, it also introduces a certain amount of background noise.
Whereas, it can be seen that from (d) and (e), using DED strategy can mitigate the impact of noise and release the potential of ADS strategy. In addition, as shown in (e) and (f), SDW module further reduces the influence of low-quality samples, and allows the detector to learn better feature representation of the targets, which can accurately focus on the target regions on the images.
\subsection{Evaluation of Hyperparameters}
\subsubsection{Sample Number $k$}
\begin{table}[t]
\centering
\caption{Evaluation of the various sample number $k$ in ADS strategy. The best result is highlighted in bold.}
\begin{tabular}{cccccccc}
\hline
$k$ & 12 & 13 & 14 & 15 & 16 & 17 & 18 \bigstrut\\
\hline
mAP (\%) & 71.52 & 72.36 & 72.42 & \textbf{72.87} & 72.05 & 71.87 & 71.93 \bigstrut\\
\hline
\end{tabular}%
\label{tab:2}%
\end{table}
Due to the importance of sample number $k$, which is introduced in the ADS strategy, we test from 12 samples to 18 samples per target on the DOTA dataset. The evaluation results are shown in Table \ref{tab:2} to illustrate its impact on performance. When $k$=12, although the quality of each selected sample is guaranteed, the number of positive samples is inadequate for the model to learn feature information, so the mAP has only reached 71.52. The performance of the model gradually improves with the increases of $k$, and the best performance is achieved when $k$=15. However, as $k$ continues to increase, more background noise are included since samples with further distance from the center of ground-truth are selected, so the mAP score drops from 72.87 to 71.93 when $k$=18. The performance is robust since the overall improvement keeps continuous with $k$ increases from 12 to 18. This result is crucial when use our ADS strategy to other datasets, which yields relatively stable detection performance and can be further improved by simply changing the value of $k$ to suit different categories.
\subsubsection{Ratio Factor $\xi$}
The ratio factor $\xi$ in DED strategy is introduced to ensure the quality of the selected samples by controlling the distribution of samples.
\begin{table}[t]
\centering
\caption{Evaluation of using various ratios factor $\xi$ in DED strategy. The best result is highlighted in bold.}
\begin{tabular}{cccccc}
\hline
$\xi$ & 0.4 & 0.6 & 0.8 & 1.0 & adaptive \bigstrut\\ \hline
mAP (\%) & 71.99 & 72.49 & 72.16 & 71.76 & \textbf{72.87} \bigstrut\\ \hline
\end{tabular}
\label{tab:4}
\end{table}
As given in Table \ref{tab:4}, different values for $\xi$ are tested to seek the optimum value, and display the influence from the diverse ratio factor. We can see that when $\xi$=0.4, the sampling range is small enough to avoid noise, and the mAP achieves 71.99. With the increase of $\xi$, the detector is able to learn more information from the selected high-quality samples, hence the detection performance increases and reaches a peak of 72.49 when $\xi$=0.6. However, the performance starts to drop, when $\xi$ further increases because it is difficult to guarantee the quality of selected samples, especially those far from the center of ground truth bounding box. To solve this problem, we set the value of $\xi$ to be adaptive according to the shapes of targets. It can be seen in Table \ref{tab:4} that the mAP of the detector with adaptive $\xi$ increases by 0.38 compared with the peak value when $\xi$=0.6, which demonstrates the effectiveness of our proposed adaptive method, and illustrates that the key to improve detection performance by spatial assignment is to obtain more positive samples while reducing the effect of noise.
\begin{table*}[tbp]
\centering
\caption{Performance comparison with different state-of-the-art methods on DOTA dataset. HG104 means Hourglass 104. $\ast$ indicates multi-scale training and testing. ATSS indicates the lite version. $\dagger$ means our re-implementation on our baseline detector. The best result is highlighted in black and red bold in anchor-based and anchor-free method, respectively.}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|ccccccccccccccc|cc}
\hline
Methods & Backbone & PL & BD & BR & GTF & SV & LV & SH & TC & BC & ST & SBF & RA & HA & SP & HC & mAP (\%) & fps\\
\hline
\textit{\textbf{Anchor-based} Methods:} & & & & & & & & & & & & & & & & & & \\
RoI-Transformer \cite{ding2019learning}$^{\ast}$ & Res101 & 88.53 & 77.91 & 37.63 & 74.08 & 66.53 & 62.97 & 66.57 & 90.50 & 79.46 & 76.75 & 59.04 & 56.73 & 62.54 & 61.29 & 55.56 & 67.74 & 7.8 \\
ICN\cite{azimi2018towards}$^{\ast}$ & Res101 & 81.36 & 74.30 & 47.70 & 70.32 & 64.89 & 67.82 & 69.98 & 90.76 & 79.06 & 78.20 & 53.64 & 62.90 & 67.02 & 64.17 & 50.23 & 68.16 & - \\
CenterMap-Net\cite{wang2020learning} & Res50 & 89.02 & 80.56 & 49.41 & 61.98 & 77.99 & 74.19 & 83.74 & 89.44 & 78.01 & 83.52 & 47.64 & 65.93 & 63.68 & 67.07 & 61.59 & 71.59 & - \\
SCRDet\cite{yang2019scrdet}$^{\ast}$ & Res101 & \textbf{89.98} & 80.65 & 52.09 & 68.36 & 68.36 & 60.32 & 72.41 & \textbf{90.85} & \textbf{87.94} & 86.86 & 65.02 & 66.68 & 66.25 & 68.24 & 65.21 & 72.61 & 9.5 \\
Gliding-Ver.\cite{xu2020gliding}$^{\ast}$ & Res101 & 89.64 & \textbf{85.00} & 52.26 & 77.34 & 73.01 & 73.14 & 86.82 & 90.74 & 79.02 & 86.81 & 59.55 & \textbf{70.91} & 72.94 & 70.86 & 57.32 & 75.02 & 13.1 \\
R$^3$Det\cite{yang2019r3det} & Res50 & 89.02 & 75.47 & 43.86 & 65.84 & 75.83 & 73.44 & 86.03 & 90.57 & 81.11 & 82.84 & 55.57 & 59.10 & 56.57 & 70.31 & 50.45 & 70.40 & - \\
S$^2$A-Net\cite{han2021align} & Res50 & 89.11 & 82.84 & 48.37 & 71.11 & 78.11 & 78.39 & 87.25 & 90.83 & 84.90 & 85.64 & 60.36 & 62.60 & 65.26 & 69.13 & 57.94 & 74.12 & \textbf{17.6} \\
S$^2$A-Net\cite{han2021align}$^{\ast}$ & Res50 & 88.89 & 83.60 & \textbf{57.74} & \textbf{81.95} & \textbf{79.94} & \textbf{83.19} & \textbf{89.11} & 90.78 & 84.87 & \textbf{87.81} & \textbf{70.30} & 68.25 & \textbf{78.30} & \textbf{77.01} & \textbf{69.58} & \textbf{79.42} & \textbf{17.6} \\
Hou \cite{hou2022refined} & Res101 & 89.32 & 76.05 & 50.33 & 70.25 & 76.44 & 79.45 & 86.02 & 90.84 & 82.80 & 82.50 & 58.17 & 62.46 & 67.38 & 71.93 & 45.52 & 72.63 & - \\
Hou \cite{hou2022refined}$^{\ast}$ & Res101 & 88.69 & 79.41 & 52.26 & 65.51 & 74.72 & 80.83 & 87.42 & 90.77 & 84.31 & 83.36 & 62.64 & 58.14 & 66.95 & 72.32 & 69.34 & 74.44 & - \\
\hline
\textit{\textbf{Anchor-free} Methods}: & & & & & & & & & & & & & & & & & & \\
TOSO\cite{feng2020toso} & Res101 & 80.17 & 65.59 & 39.82 & 39.95 & 49.71 & 65.01 & 53.58 & 81.45 & 44.66 & 78.51 & 48.85 & 56.73 & 64.40 & 64.24 & 36.75 & 57.92 & 16.9 \\
PIoU\cite{chen2020piou} & DLA34\cite{yu2018deep} & 80.90 & 69.70 & 24.10 & 60.20 & 38.30 & 64.40 & 64.80 & \textcolor{red}{\textbf{90.90}} & 77.20 & 70.40 & 46.50 & 37.10 & 57.10 & 61.90 & 64.00 & 60.50 & - \\
IENet\cite{lin2019ienet} & Res101 & 88.15 & 71.38 & 34.26 & 51.78 & 63.78 & 65.63 & 71.61 & 90.11 & 71.07 & 73.63 & 37.62 & 41.52 & 48.07 & 60.53 & 49.53 & 61.24 & 16.9 \\
Axis-Learning\cite{xiao2020axis} & Res101 & 79.53 & 77.15 & 38.59 & 61.15 & 67.53 & 70.49 & 76.30 & 89.66 & 79.07 & 83.53 & 47.27 & 61.01 & 56.28 & 66.06 & 36.05 & 65.98 & 14.1 \\
ATSS\cite{zhang2020bridging}$^{\dagger}$ & Res50 & 88.47 & 80.05 & 47.27 & 60.65 & 79.85 & 78.80 & 87.41 & 90.75 & 77.37 & 85.24 & 43.22 & 60.80 & 66.52 & 70.95 & 41.82 & 70.61 & - \\
DRN\cite{pan2020dynamic} & HG104\cite{newell2016stacked} & 88.91 & 80.22 & 43.52 & 63.35 & 73.48 & 70.69 & 84.94 & 90.14 & \textcolor{red}{\textbf{83.85}} & 84.11 & 50.12 & 58.41 & 67.62 & 68.60 & 52.50 & 70.70 & - \\
O$^2$-DNet\cite{wei2020oriented} & HG104\cite{newell2016stacked} & 89.31 & 82.14 & 47.33 & 61.21 & 71.32 & 74.03 & 78.62 & 90.76 & 82.23 & 81.36 & \textcolor{red}{\textbf{60.93}} & 60.17 & 58.21 & 66.98 & \textcolor{red}{\textbf{64.03}} & 71.04 & - \\
BBAVectors\cite{yi2021oriented} & Res101 & 88.35 & 79.96 & \textcolor{red}{\textbf{50.69}} & 62.18 & 78.43 & 78.98 & 87.94 & 90.85 & 83.58 & 84.35 & 54.13 & 60.24 & 65.22 & 64.28 & 55.70 & 72.32 & 18.4 \\
\textbf{EARL (Ours)} & Res50 & 89.76 & 78.79 & 47.01 & 65.20 & 80.98 & 79.99 & 87.33 & 90.74 & 79.17 & 86.23 & 49.09 & 65.87 & 65.75 & 71.86 & 55.21 & 72.87 & \textcolor{red}{\textbf{30.8}} \\
\textbf{EARL (Ours)}$^{\ast}$ & Res50 & \textcolor{red}{\textbf{90.13}} & \textcolor{red}{\textbf{83.90}} & 47.19 & \textcolor{red}{\textbf{72.17}} & \textcolor{red}{\textbf{81.54}} & \textcolor{red}{\textbf{84.26}} & \textcolor{red}{\textbf{88.24}} & 90.69 & 79.10 & \textcolor{red}{\textbf{86.71}} & 60.47 & \textcolor{red}{\textbf{72.21}} & \textcolor{red}{\textbf{71.26}} & \textcolor{red}{\textbf{73.81}} & 58.80 & \textcolor{red}{\textbf{76.03}} & \textcolor{red}{\textbf{30.8}} \\
\hline
\end{tabular}}
\label{tab:5}
\end{table*}
\begin{figure*}[tbp]
\centering
\includegraphics[width=0.9\linewidth]{DOTA-Label.png}
\caption{Visualization of detection results on the DOTA dataset with our method. Best viewed in colour.}
\label{fig:visualDOTA}
\end{figure*}
\begin{figure*}[tbp]
\centering
\includegraphics[width=1.0\linewidth]{exp_ads_percent.png}
\caption{Histogram of the percentage of the number of positive samples selected at different feature levels using different label assignment strategies on the DOTA dataset. Best viewed in colour.}
\label{fig:exp_ads}
\end{figure*}
\subsection{Comparison with the State-of-the-Art}
\subsubsection{Results on DOTA}
To evaluate the performance of the proposed method, EARL is compared with both anchor-based and anchor-free state-of-the-art detectors on DOTA dataset, and the experimental results are shown in Table \ref{tab:5}. It can be observed that anchor-based detectors can usually achieve higher performance on DOTA dataset. However, anchor-based detectors usually have limitations in inference speed due to complex structures and IoU calculations, which results in lower fps. EARL with ResNet50 as the backbone and FPN as the neck achieves mAP of 72.87 and inference speed of 30.8 fps, which is comparable in accuracy and much faster than anchor-based detectors, even without any tricks.
When compared with R$^3$Det with the same backbone, the mAP of our EARL increases 2.47, though R$^3$Det employs anchors. Moreover, EARL outperforms S$^2$A-Net in the aspect of speed by 13.2 fps, which is one of the state-of-the-art one-stage anchor-based detectors. In addition, EARL surpasses the most recent one-stage anchor-based oriented object detector (i.e., Hou \cite{hou2022refined}) by 0.24 in mAP, even though our method is based on anchor-free framework and only uses ResNet50 as backbone.
When compared with the anchor-free detectors, being integrated with the simple backbone network (i.e., ResNet50), our EARL method can provide better detection performance, even though they use the more complex backbone networks, i.e., Hourglass104, ResNet101. Specifically, when applying multi-scale augmentation in experiments, our method achieves the mAP of 76.03 and obtains the best AP results for small targets, such as SV and LV of 81.54 and 84.26, respectively, which further demonstrates the effectiveness of our method for targets with extreme scales. The visualization of detection results with our method is displayed in Fig.~\ref{fig:visualDOTA}.
In addition, to further illustrate that our method can obtain a better scale-level distribution, as shown in Fig.~\ref{fig:exp_ads}, the proposed EARL is compared with the most relevant assignment strategy, ATSS, together with
the baseline method R-FCOS on DOTA dataset, by comparing the distribution of the positive samples among different feature levels statistically. It can be seen that R-FCOS and ATSS achieve scale assignment according to the predefined scale constraints, which causes uneven sample distribution leading to the scale-level bias, especially for targets with extreme scales, such as GTF and SV. While our method with ADS strategy ensures that targets with different scales can adaptively assign training samples from all feature levels, which alleviates the scale-level bias thus improves detection performance. As a result, EARL achieves the mAP of 72.87 and has significant improvements by 1.80 and 2.26 on DOTA dataset over R-FCOS and ATSS, respectively, as shown in Table \ref{tab:1.1} and Table \ref{tab:5}.
Furthermore, the AP performance in Table \ref{tab:5} also confirm our method has the improvements on targets with both extreme large and small sizes, such as GTF, SBF, SV and LV. The above analysis proves that our EARL is advantageous for object detection in remote sensing images.
\begin{table}[tb]
\scriptsize
\centering
\caption{Performance comparison with different state-of-the-art methods on the HRSC2016 dataset. The best result is highlighted in bold.}
\begin{tabular}{cccc}
\hline
Methods & Backbone & mAP$_{07}$ (\%) & mAP$_{12}$ (\%)\\
\hline
\textit{\textbf{Anchor-based} Methods}\\
R$^2$CNN\cite{jiang2017r2cnn} & Res101 & 73.1 & 79.7 \\
RRPN\cite{ma2018arbitrary} & Res101 & 79.1 & 85.6\\
RetinaNet-H\cite{yang2019r3det} & Res101 & 82.9 & 89.2\\
R$^2$PN\cite{zhang2018toward} & VGG16 & 79.6 & - \\
RoI-Transformer\cite{ding2019learning} & Res101 & 86.2 & - \\
Gliding-Ver.\cite{xu2020gliding} & Res101 & 88.2 & - \\
DAL\cite{ming2020dynamic} & Res50 & 88.6 & - \\
CenterMap-Net\cite{wang2020learning} & Res50 & - & 92.8 \\
\hline
\textit{\textbf{Anchor-free} Methods}\\
RC2\cite{liu2017rotated} & VGG16 & 75.7 & - \\
Axis-Learning\cite{xiao2020axis} & Res101 & 78.2 & - \\
TOSO\cite{feng2020toso} & Res101 & 79.3 & - \\
BBAVectors\cite{yi2021oriented} & Res101 & 88.6 & - \\
DRN\cite{pan2020dynamic} & H104 & - & 92.7 \\
\textbf{EARL (Ours)} & Res50 & \textbf{89.0} & \textbf{93.0}\\
\hline
\end{tabular}
\label{tab:hrsc2016}
\end{table}
\begin{figure}[tb]
\centering
\includegraphics[width=1.0\linewidth]{hrsc2016.png}
\caption{Visualization of detection results on the HRSC2016 dataset with our method. Best viewed in colour.}
\label{fig:visualHRSC}
\end{figure}
\subsubsection{Results on HRSC2016} To further evaluate the robustness of the proposed method, we setup experiments on the HRSC2016 dataset, and the results are listed in Table \ref{tab:hrsc2016}. It is obvious that compared with both anchor-based and anchor-free methods, our EARL achieves the best performance of 89.0 and 93.0 in mAP under the Pascal VOC2007 and VOC2012 metrics\cite{everingham2010pascal}, respectively. The detection results on HRSC2016 dataset are illustrated in Fig.~\ref{fig:visualHRSC}, which also confirm our method is robust for detecting targets with large aspect ratios and further demonstrate the advancement of our method.
\section{Conclusion}
In this paper, we presented a novel and efficient strategy for label assignment, namely EARL. An adaptive scale sampling strategy was proposed to solve the problem of targets with large variation of scales, especially in remote sensing images, by selecting samples from suitable and continuous multi-level feature maps. Considering the large aspect ratio of the target and to obtain a more accurate and reasonable sampling range while reducing the influence of background noise, a dynamic elliptical distribution aided sampling method was designed, which was adaptively controlled by the target size. To learn high quality information from selected training samples, a spatial distance weighting module was developed. The experimental results demonstrated the effectiveness of our proposed method when being integrated with simple structure without any tricks. Our EARL is so simple that it has the potential to be easily combined with different orientation detectors to achieve better performance.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Predictive models that use images as inputs are constrained to any image alteration that can degrade the optimal performance of these models. Sometimes the degree of modification on the images can be regulated. A good example is when images are compressed before being sent to the algorithm for prediction. In case of earth observation satellites, the high cost of downloading the images can be significantly reduced by compressing the images first. \cite{Lofqvist2021}. One approach is to make images smaller to reduce the costs of downloading to earth \cite{rs13030447}. In this context, decision-makers need tools to study the optimal modification so that the performance of the predictive models is adequate despite the compression. \textsc{iquaflow}\footnote{\href{https://github.com/satellogic/iquaflow}{https://github.com/satellogic/iquaflow}} is a software tool that has been designed precisely to study image quality as well as the performance of models trained on top of provided datasets that are modified with any user-defined alteration. \cite{Lofqvist2021} studies object detection inference with compression algorithms based on decimation and scaling with interpolation in the context of earth observation from satellite applications. In the present work, the study is brought further with custom training for each level of compression, new kinds of compression, and new models of object detection that are suitable for oriented annotations as explained below.
\subsection{Compression}
Compression algorithms can be lossless or lossy \cite{Hussain2018ImageCT}. The first kind performs an operation on the image that allows the recovery of the original image before it was compressed. The second kind, on the other hand, does an irreversible operation. Using a lossy compression algorithm, we can achieve a greater reduction in file sizes than with a lossless one. A simple straightforward technique for lossy compression can be the interpolation of an image to fewer pixels. Then a smaller image will have lost information and it will also be smaller in file size. In this study the JPEG compression is used as explained in section \ref{compresssect}.
\subsection{Object detection}
A good example of predictive models on images is object detection (such as vehicles from aerial images). Most detectors such as Faster R-CNN \cite{renNIPS15fasterrcnn}, SSD \cite{LiuAESRFB16} and YOLOv2, v3 \cite{redmon2018yolov3} rely on a set of pre-defined anchors that consist in a small set of bounding boxes summarizing the most relevant geometric shapes covering relevant scale, aspect ratios and elongation directions. The idea is that any object can be associated with a specific anchor box without having to have a perfect fit.
However, the definition of this set of anchor boxes is a hyper-parameter that must be defined and has an effect on the detection performance. The models are, of course, sensitive to the sizes, aspect ratios, and a number of anchors defined in the set (see \cite{renNIPS15fasterrcnn} and \cite{lin2017focal}).
Another aspect to consider is the number of stages. Detectors can be composed of multiple stages and each of them has a trained model that solves a specific task in the workflow. A typical case in an object detection problem is the Region Proposal Network which is responsible for the task of generating bounding box proposals. Examples of that are \cite{DBLP2018}, \cite{Xu_2021} and \cite{SCRDet2018}. One advantage of the multistage approach is that each step in the workflow can be easily defined and understood by human logic. In single-stage detectors, the logic can be difficult to interpret inside an end-to-end network solution.
Depending on the annotations one can use a model that predicts with horizontal bounding boxes (HBB) or oriented bounding boxes (OBB). One problem with HBB is distinguishing between overlapping instances of different objects. This is usually approached with the logic of Non-Maximum Suppression (NMS) that involves the measure of Intersection Over Union between different instances to asses the overlapping and whether or not candidate boxes belong to the same sample. This logic struggles when there are elongated objects that are diagonal and parallel to each other. In aerial images, these can be ships in a harbor or trucks in parking. One solution for this is to consider more complex geometries that have a better fit with the object. The simplest complexity, in this case, is to orient the bounding box.
The models used in this study are explained in section \ref{objectdectsect}.
\subsection{Iquaflow}
Image quality can be often evaluated by the human eye. However, it is very challenging to define a numerical measurement for image quality. One of the reasons is that there are many aspects to consider such as the blur, the noise, the quality distribution along frequencies, etc. Moreover, image quality should be measured according to the particular application of the images being measured. Supervised super-resolution image prediction models are algorithms that translate an input image to a higher-resolution image that contains more pixels. These models are trained with a database containing pair samples of images with their respective higher resolution (also known as ground-truth or target images). In this context, the evaluation of quality will perform better by comparing the predicted image against the target image. These metrics are also known as similarity metrics and they include \cite{uSSIM2020}, \cite{DeepPerc2018}, \cite{Reisenhofer_2018} and \cite{Ding_2020}. Another context is when images are used as inputs for other predictive models with the aim to collect information from them. It is the case of an image classifier or object detection. For this case, a suitable image quality evaluation method can be the performance of this model on the images. This is assuming that changes in the input image quality are affecting the performance of the prediction model. Again, this is a way to measure image quality that is adapted to the actual application of the image.
\textsc{iquaflow}\cite{iquaflow} is a python package tool that measures image quality by using different approaches. Deterministic metrics include blind metrics which are measured directly on the image without comparing against a reference image or similarity metrics when they are measuring affinity against an ideal case. There are two metrics that have been designed for \textsc{iquaflow} which are implicit measurements of blur and noise levels. The first relies on edges found within the images and it measures the slope in the Relative Edge Response (RER) \cite{RER}. Then the second is based on homogeneous areas where the noise can be estimated (Signal to Noise ratio - SNR) \cite{SNR}. The Quality Metric Regression Network (QMRNet)\cite{qmrnet} has been designed, trained, and integrated into \textsc{iquaflow}. This is a classifier with small intervals that can predict quality parameters on images such as blur (measured as equivalent sigma from a gaussian bell), sharpness, pixel size and noise. Quality can also be measured by checking how predictive models trained on the image dataset are performing. A good example is the present study where object detection is trained on different quality datasets with different outcomes.
Apart from measuring image quality, \textsc{iquaflow} has a whole ecosystem that facilitates the design of new studies and experiment sets made of several training runs with variations. \textsc{iquaflow} wraps another open source tool named \href{https://mlflow.org/}{Mlflow} that is used for machine learning experiment tracking. It will record the executions in a standard format so that they are later easily visualized and compared from \href{https://mlflow.org/}{Mlflow} user interface tool in the browser. In \textsc{iquaflow} the user can add custom metrics and dataset modifiers that are easily integrated into a use case study.
\section{Materials and Methods}
The aim of the study is to measure the variation of the object detection algorithm's performance on a given image dataset that is modified with various compression ratios. Our goal is to evaluate what is the maximum compression level that still allows for acceptable model performance. In this section, the compression algorithm is described, and the object detection model(s) that we considered, as well as the tool used for managing our experiments.
\subsection{Data}
Two different datasets are used to carry out two experiments. The first analysis is based on the airplanes dataset\footnote{Contact <EMAIL> to request access to the dataset} which consists of 998 images of 1024 × 1024 pixels from airport areas with a total of almost $17000$ annotated planes. These captures were made using NewSat Satellogic constellation ($1~\mathrm{m}$ GSD) and the annotations were made using Happyrobot\footnote{\href{https://happyrobot.ai}{https://happyrobot.ai}} platform. The training partition contained $13731$ annotations and the remaining were used for evaluation.
The second experiment was based on the public dataset DOTA \cite{Xia_2018_CVPR}. It is a dataset for object detection in aerial images. The images are collected from different sensors the image sizes are ranging from $800~\times~800$ to $20000~\times~20000$ pixels and the pixel size varies from $0.3~\mathrm{m}$ to $2~\mathrm{m}$ resolution. DOTA has several versions and DOTA-v1.0 has been used in the present study which contains 15 common categories \footnote{plane, ship, storage tank, baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, large vehicle, small vehicle, helicopter, roundabout, soccer ball field and swimming pool.}, $2806$ images and more than $188k$ object instances. The annotations are oriented bounding boxes which allows us to train both oriented (OBB) and horizontal bounding boxes (HBB) models. The proportions of the training set, validation set, and testing set in DOTA-v1.0 are $1/2$, $1/6$, and $1/3$ \cite{Xia_2018_CVPR}. A disadvantage of this dataset is that the test set is not openly available, rather it is in a form of a remote service to query the predictions. This does not allow to alter the test on the same way the other partitions are modified in the present study. Because of that, 2 partitions are made from the validation set: half of is is used as actual validation and the other half for testing. Then the images are cropped to $1024~\times~1024$ with padding when necessary. After this operation the number of crops for the partitions train, validation and testing are respectively $9734$, $2670$ and $2627$.
\subsection{Compression} \label{compresssect}
In this study, JPEG compression \cite{Wallace91} is used. It is a lossy form of compression based on the discrete cosine transform (DCT) that converts images into the frequency domain and discards high-frequency information by a quantization process. The degree of compression in JPEG can be adjusted: the greater the quality the bigger the file size. In the present study, the compression is set at different levels with the aim to find an optimal value with respect to the performance of predictive models trained on them. We used the JPEG compression from OpenCV \cite{opencv_library} that can be regulated with the parameter CV\_IMWRITE\_JPEG\_QUALITY which can vary from $0$ to $100$ (the higher is the better) with a default value of 95.
Figure~\ref{figCompression} shows an example of the effect when compressing one of the images with JPEG method.
\begin{figure}
\centering
\includegraphics[width=10.5 cm]{Definitions/JPEG.png}
\caption{JPEG compression effects (original, JPG10,and JPG5 from left to right). This image is from the airplane dataset from Satellogic.
\label{figCompression}}
\end{figure}
\subsection{Object detection} \label{objectdectsect}
The first experiment has HBB annotated objects and the model YOLOv5 \cite{glenn_jocher_2021_4679653} was used because of its fast training and implementation.
For the second experiment, two OBB models were used. The first was Oriented R-CNN which is a two-stage oriented detector that uses Region Proposal Network (oriented RPN) in the first stage in order to generate high-quality oriented proposals in a nearly cost-free manner \cite{Xie_2021_ICCV}.
Then the other model used was FCOS \cite{tian2021fcos} which is originally designed for horizontal bounding boxes but it can be adapted with an added convolution layer channel on the top of the regression features that define the direction of the bounding box. Intersection Over Union is often used as a loss function in object detection. However, the IoU calculation between oriented boxes is complex and often not differentiable. There are rotated IoU that implements differentiable IoU calculation for oriented bounding boxes. In this case, the PolyIoULoss \cite{PIoU2021} between the OBB predictions and ground truths is used as a bounding box loss.
The performance of the detector is measured by calculating the average recall (AR) as well as the Mean Average Precision (mAP). AR is a ratio of correctly detected instances over the actual amount of objects. On the other hand, AP is defined with the same correctly detected instances over all the amount of detected cases (including wrong detection). The predicted bounding boxes do not have to have a perfect match with the ground truth. Because of that, the Intersection over Union (IoU) for each prediction and ground truth match candidate is measured to evaluate if they match. In which case it is considered a correct detection \cite{powers2011evaluation}. In this study, mAP is calculated by taking the mean AP over all classes and over a range of IoU thresholds.
\subsection{Experiment management}
The present study involves a workflow with multiple versions of the original dataset with the corresponding partitions for each altered version (train, validation and test) as well as many training experiment executions and tracking of results that must be organized correctly. All this can be managed easily with a typical \textsc{iquaflow} workflow as follows:
\begin{enumerate}
\item Optionally the user can start with a repository template of \textsc{iquaflow} use cases. This repository uses cookiecutter which is a python package tool for repository templates. By using this you can initialize a repository with the typical required files for a study in \textsc{iquaflow}.
\item The first step will be to set the modifications of the original dataset with different compression levels. This can be done with a list of Modifiers in \textsc{iquaflow}. There are some modifiers already available in \textsc{iquaflow} with performing specific alterations. However, one can set up a custom modifier by inheriting the DSModifier class of \textsc{iquaflow}. The list of modifiers will then be passed as an argument to the experiment setup.
\item Next step is to adapt the user training script to the \textsc{iquaflow} conventions. This is just to accept some input arguments such as the output path where the results are written. Optionally one can monitor in streaming the training by inputting additional arguments as explained in \textsc{iquaflow}'s guide.
\item All previous definitions are introduced in the experimental setup that can be executed afterward. The whole experiment will contain all runs which are the result of combining dataset modifications (the diverse compression levels) and the two different detectors that are used which will be defined as hyperparameter variations in the experiment setup.
\item The evaluation can be either done within the user's custom training script or by using a Metric in \textsc{iquaflow}. Similar to Modifiers there are some specific Metrics already defined in \textsc{iquaflow}. Alternatively, the user can make a custom metric by inheriting the class Metric from \textsc{iquaflow}. The results can be collected from \textsc{iquaflow} or directly by raising an \href{https://mlflow.org/}{mlflow} server which is a tool that is wrapped and used by \textsc{iquaflow}.
\end{enumerate}
As you can see using \textsc{iquaflow} we can automate the compression algorithm on the data, run the user custom training script and evaluate a model. All the results are logged using mlflow and can be handily compared and visualized. \textsc{iquaflow} is the ideal tool for this purpose.
\section{Results}
The airplanes dataset from Satellogic\footnote{\href{https://github.com/satellogic/iquaflow-airport-use-case}{https://github.com/satellogic/iquaflow-airport-use-case}} has the unique category of planes. The image format is tiff and the original average image size is $3.204$Megabytes. The average recall (AR) is measured and the Mean Average Precision (mAP) is calculated over different Intersection Over Union (IoU) thresholds varying from $0.5$ to $0.95$ with a step of $0.05$ and average again for the final score. Table~\ref{table:tab1} contains the resultant metrics and Figure~\ref{fig1} shows performances (mAP) along different levels of compression.
\begin{table}
\caption{Performance results at different compression levels using the airplanes dataset and two YOLOv5 model sizes with different architecture complexities. The scores for the different models are expressed as Mean Average Precision (mAP) and Average Recall (AR) as expressed in the methodology section. The last column shows the equivalent average image size from the dataset given the level of compression used.\label{tab1}}
\centering
\begin{tabular}{lllll}
\toprule
\multicolumn{2}{c}{\textbf{YOLOv5 NANO}} & \multicolumn{2}{c}{\textbf{YOLOv5 SMALL}} & size \\
\cmidrule(r){1-5}
\textbf{AR} & \textbf{mAP} & \textbf{AR} & \textbf{mAP} & \textbf{Mb} \\
\midrule
0.898 & \textbf{0.669} & \textbf{0.922} & \textbf{0.714} & 2.051 \\
\textbf{0.899} & 0.666 & 0.919 & 0.709 & 1.428 \\
0.892 & 0.663 & 0.917 & 0.708 & 1.256 \\
0.888 & 0.657 & 0.916 & 0.703 & 0.988 \\
0.872 & 0.636 & 0.891 & 0.675 & 0.874 \\
\bottomrule
\end{tabular}
\label{table:tab1}
\end{table}
\begin{figure}
\centering
\includegraphics[width=10.5 cm]{Definitions/ObjectDetection-YOLOv5-JPEGcompression.jpg}
\caption{Scatter plot that shows the performance of the models (mAP) evolution with different compression levels expressed as average image size of the files in the modified Satellogic's airplanes dataset. Red with "x" and blue with "+" correspond to model size nano and small of YOLOv5 model respectively.
\label{fig1}}
\end{figure}
The DOTAv1.0 dataset has 15 categories and different metrics are measured for each class. The categories of 'plane' and 'storage tank' are performing the best whereas the categories 'bridge' and 'soccer-ball-field' are performing the worst. Table~\ref{table:tab2} summarizes the averaged metrics for each run by aggregating with the mean of all the categories. Following the same logic, Figure~\ref{fig2} charts the evolution of performance (mAP) along different levels of compression. The original average image size of the $1024~\times~1024$ crops without compression was $1.13$Megabytes\footnote{\href{https://github.com/satellogic/iquaflow-dota-obb-use-case}{https://github.com/satellogic/iquaflow-dota-obb-use-case}}.
\begin{table}
\caption{Performance results at different compression levels using the DOTA1.0 dataset and two OBB models. The scores for the different models are expressed as Mean Average Precision (mAP) and Average Recall (AR) as expressed in the methodology section. The last column shows the equivalent average image size from the dataset given the level of compression used.\label{tab2}}
\centering
\begin{tabular}{lllll}
\toprule
\multicolumn{2}{c}{\textbf{FCOS}} & \multicolumn{2}{c}{\textbf{RCNN}} & size \\
\cmidrule(r){1-5}
\textbf{AR} & \textbf{mAP} & \textbf{AR} & \textbf{mAP} & \textbf{Mb} \\
\midrule
\textbf{0.869} & 0.688 & 0.806 & 0.662 & 0.332 \\
0.856 & 0.677 & 0.812 & 0.658 & 0.321 \\
0.865 & \textbf{0.692} & \textbf{0.813} & \textbf{0.668} & 0.311 \\
0.861 & 0.679 & 0.812 & 0.666 & 0.313 \\
0.861 & 0.679 & 0.810 & 0.663 & 0.305 \\
0.861 & 0.685 & 0.806 & 0.668 & 0.273 \\
0.849 & 0.677 & 0.811 & 0.669 & 0.245 \\
0.856 & 0.675 & 0.804 & 0.659 & 0.226 \\
0.847 & 0.673 & 0.800 & 0.660 & 0.209 \\
0.846 & 0.666 & 0.798 & 0.658 & 0.191 \\
0.835 & 0.651 & 0.785 & 0.649 & 0.171 \\
0.831 & 0.643 & 0.785 & 0.636 & 0.138 \\
0.799 & 0.598 & 0.741 & 0.588 & \textbf{0.097} \\
\bottomrule
\end{tabular}
\label{table:tab2}
\end{table}
The optimal compression ratio for the oriented-RCNN model seems to be around JPEG quality score of 70 which corresponds to an average image size of 0.245 Megabytes. This is because it corresponds to the minimum average file size that can be defined without lowering the performance. On the other hand, the adapted FCOS model seems to have an optimal around 80 for JPEG quality score which corresponds to an average image size of 0.273 Megabytes.
\begin{figure}
\centering
\includegraphics[width=10.5 cm]{Definitions/ObjectDetection-Oriented-JPEGcompression.jpg}
\caption{Scatter plot that shows the performance of the models (mAP) evolution with different compression levels expressed as average image size of the files using the DOTA1.0 dataset and two OBB models. Red dots correspond to the adapted FCOS model whereas blue dots are from the oriented RCNN model.
\label{fig2}}
\end{figure}
\unskip
\begin{figure}[ht]
\centering
\includegraphics[width=10.5 cm]{Definitions/tp10ships.png}
\caption{An example of prediction on an image with boats compressed with $CV\_JPEG\_QUALITY$ of $10$ which is equivalent to an average dataset image size of $0.097Mb$. The model used is adapted FCOS. The image belongs to the testing partition.
\label{figships}}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=10.5 cm]{Definitions/planes.png}
\caption{An example of prediction on an image with planes compressed with $CV\_JPEG\_QUALITY$ of $10$ which is equivalent to an average dataset image size of $0.097Mb$. The model used is adapted FCOS. The image belongs to the testing partition.
\label{figplanes}}
\end{figure}
\section{Conclusions}
In the experiment with Satellogic's airplanes dataset, the decrease in performance with compression is consistent for both models. The variations of mAP is small between the ranges of $0.15$ and $0.25$ average image size. The additional complexity of the Small model has a constant positive shift of $0.5$ in mAP with respect to the Nano model along all the analyzed compression rates.
In the context of the second experiment the adapted FCOS model seems to perform better than oriented RCNN because the AR and mAP are greater for all levels of compression. On the other hand, oriented-RCNN seems more resilient because the optimal compression ratio is higher than the optimal case for the other model. However, the degraded performance of FCOS model given the same compression setting as the optimal value for oriented-RCNN still offers higher performance. FCOS is also easier to implement because it is a single-stage detector that does not require setting anchors as hyperparameters. So far, given the data and context of the study, FCOS seems the best option.
Another interesting observation is the high resilience of the model for some specific applications. The figures \ref{figships} and \ref{figplanes} show a prediction with the FCOS model on an image with boats and airplanes respectively. Both of the images were set with a compression rate of $10$ for $CV\_JPEG\_QUALITY$ which is equivalent to an average dataset image size of $0.097Mb$. In the first image, $146$ ships were correctly detected (True positives), $9$ were wrongly detected (False positive) and $11$ ships were missed (False negative). In the other example all the planes (total amount: $39$) are correctly detected see \ref{figplanes} with no false positives or false negatives. This highlights the greater capacity of compressing images for usage such as the detection of airplanes over smaller or more difficult objects.
This study highlights the potential of \textsc{iquaflow} for decision-makers as well as researchers that want to study performance variation in an agile and ordered way. The key effort has been the development of the tool so that it facilitates further studies with the aim to scale it. The tool also allows for mitigating the uncertainty of image quality by using several strategies to measure that. This is helping also in studies that are exploring suitable solutions for satellite image Super Resolution.
\section*{Acknowledgments}
Conceptualization, P.G. and J.M.; methodology, P.G. and J.M.; software, P.G. and K.T.; validation, K.T. and J.M.; formal analysis, P.G.; investigation, P.G.; resources, J.M.; data curation, P.G.; writing---original draft preparation, P.G.; writing---review and editing, K.T. and J.M.; visualization, P.G.; supervision, J.M.; project administration, J.M.; funding acquisition, J.M. All authors have read and agreed to the published version of the manuscript.
This research was funded by the Ministry of Science and Innovation and by the European Union within the framework of Retos-Collaboration of the State Program of Research, Development and Innovation Oriented to the Challenges of Society, within the State Research Plan Scientific and Technical and Innovation 2017-2020, with the main objective of promoting technological development, innovation, and quality research. grant number: RTC2019-007434-7.
The authors declare no conflict of interest.
\bibliographystyle{unsrt}
|
\section{Introduction}
Heusler alloys are known to exhibit exotic phenomena as well as novel potential applications, which have stimulated a tremendous interest in physics and materials technology. Many systems from this family are reported to be promising spintronic materials such as half-metals (HM) \cite{PhysRevLett.50.2024} spin gapless semiconductors (SGS),\cite{bainsla2015spin} bipolar magnetic semiconductors (BMS),\cite{PhysRevB.104.134406} spin-valve \cite{PhysRevB.105.144409} etc. Most of the reported Heusler materials are superior to other materials from the application point of view because of their stable structure and high spin-polarization. The co-existence of different and interesting properties in these systems gives rise to new avenues for multifunctional materials suitable for technological applications such as spintronics. Recently, tuning the electronic structure by defects/impurities has become a major focus by various researchers to achieve the desired properties suitable for applications.\cite{PhysRevB.103.085202} As Heusler alloys are prone to anti-site disorder, complex magnetic/electronic structures can be realized in these materials, with a wide tuning capability. One of the main motives of this work is to understand the role of anti-site disorder in band engineering and hence in the tuning of the magneto-electronic properties.
In this article, we report the addition of a new member to the recently identified magnetic quantum material class namely spin semi-metals (SSM), with several complementary properties. This is a combined theoretical and experimental study where SSM nature is confirmed in a new quaternary Heusler alloy (QHA) CoRuVSi. The objective of this work is two-fold: (1) better understanding the key features of this relatively new class both from physics and materials perspectives, (2) highlight the importance of this class of materials for potential spintronic and thermoelectric applications.
In HMs, one of the spin bands shows metallic nature, while the other shows semiconducting/insulating behavior. SSM, on the other hand, is an unconventional class of spintronic materials in which one of the spin bands possesses semimetallic nature, while the other possesses a small gap near the Fermi level (E$_F$). Thus, electronic states of such materials can be easily controlled by an external perturbation (magnetic field, temperature etc.) and hence are advantageous for spin-transport based applications. This advantage is missing in the conventional spintronic systems such as HM and SGS. A schematic representation of the density of states (DoS) and overlap of conduction and valence bands for HM and SSM are shown in Fig. \ref{fig:schematic-SGS-SM}.
CoRuVSi is found to crystallize in the perfect cubic structure (space group $F\bar{4}3m$) with a partial L2$_1$-type disorder. The magnetization data indicates a weak ferri-/antiferro-magnetic ordering at very low temperature, with a very small saturation magnetization $\sim$ 0.13 $\mu_B$/f.u. The magnetization data indicate quenching of moment, attributed to the atomic disorder, a prediction also supported by our ab-initio disorder calculations. Theoretical studies reveal a fully compensated ferrimagnetic nature for CoRuVSi. Transport results provide strong evidence of semimetallic behavior dominated by two-band conduction, while low-T magnetoresistance data indicates the non-saturating, linear, positive magnetoresistance (LPMR), with a quadratic behavior with T. Close analysis of MR data hints toward the small-gap electronic structure near the E$_F$ as the origin of quantum LPMR, which indirectly hints toward the SSM nature present in this system. Point contact Andreev reflection (PCAR) measurements reveal a reasonably high spin polarization of $\sim$ 50\%. This matches fairly well with the theoretical calculations, again facilitating an indirect evidence of SSM feature in this system. CoRuVSi also shows a reasonably high thermopower value of $0.7$ $m Watt/ m-K^{2}$ at room temperature and hence can be further explored for its potential as a promising thermoelectric material. Our ab-initio simulation confirms the spin semimetallic feature in this alloy with a high spin polarization. Overall, the present study introduces a new member namely CoRuVSi to the magnetic quantum phase, having the potential for multifunctional applications, and gives a comprehensive analysis of the interplay between the non-trivial electronic states with magnetism and anti-site disorder. Such a combined theoretical and experimental study gives a unique platform to explore new exotic states of quantum matter.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{Fig1.png}
\caption{Schematic representation of spin polarized (a) density of states (DoS) for a conventional
half metal (HM) (b) bands (left and right panels) and DoS (middle) for spin semi-metal (SSM)}.
\label{fig:schematic-SGS-SM}
\end{figure}
\section{Experimental Details}
Polycrystalline samples of CoRuVSi were prepared using an arc melting system in a high purity Argon atmosphere using stoichiometric constituent elements having a purity of 99.99\%. To accomplish perfect homogeneity, the samples were melted several times and a very small weight loss ($<$ 0.15 \%) was observed after the final melting. To study the crystal structure, at room temperature (RT) X-ray diffraction (XRD) patterns were taken using Cu-K$\alpha$ radiation with the help of Panalytical X-pert diffractometer. For the crystal structure analysis, FullProf Suite software \cite{rodriguez1993recent} was used.
Magnetization measurements at various temperatures were carried out using a vibrating sample magnetometer (VSM) attached to a physical property measurement system (PPMS) (Quantum Design) for fields up to 70 kOe.
Temperature and field-dependent resistivity along with the MR measurements were carried out employing a physical property measurement system (PPMS-DynaCool; Quantum Design) using the electrical transport option (ETO) in a traditional four-probe method, applying a 10 mA current at a 15 Hz frequency. Hall measurements were carried out using PPMS with the van der Pauw method by applying a 5 mA current at 21 Hz frequency. Specific heat (C$_p$) measurements were done in a 14T/2 K PPMS. A small piece of the sample (18 mg) was used to measure C$_p$, down to 2 K in zero field and in 5T applied magnetic field using a relaxation calorimetry technique.
Thermoelectric power (TEP) in zero magnetic fields was measured using the differential dc sandwich method in a homemade setup in the temperature range of 4–300 K.
Point contact Andreev reflection (PCAR) measurements were performed in PPMS using a superconductive Nb tip. The landing of the tip on the sample is carefully controlled by a fully automated vertical Attocube piezo-stepper. Two additional horizontal Attocube piezo steppers are used to move the sample in horizontal directions, in order to probe the pristine area of the sample. The differential conductance spectra were fitted using the modified Blonder-Tinkham-Klapwijk (m-BTK) model, as detailed elsewhere.\cite{stamenov2013point,PhysRevB.94.094415}
\section{Computational details}
To study the electronic/magnetic structure of CoRuVSi, $\textit ab initio$ calculations were performed using spin-resolved density functional theory (DFT) \cite{hohenberg1964inhomogeneous} implemented within Vienna ab initio simulation package (VASP) \cite{kresse1996efficient,kresse1996efficiency,kresse1993ab} with a projected augmented-wave (PAW) basis.\cite{kresse1999ultrasoft} We used the electronic exchange-correlation potential due to Perdew, Burke, and Ernzerhof (PBE) \cite{perdew1996generalized} within the generalized gradient approximation (GGA) scheme. For the Brillouin zone integration within the tetrahedron method, a $24\times24\times24$ k-mesh was used. A plane wave energy cut-off of 420 eV was used for all the calculations. All the structures were fully relaxed with total energies (forces) converged to values less than 10$^{-6}$ eV (0.01 eV/\AA).
The Wannier90 \cite{wannier90,PhysRevB.65.035109,RevModPhys.84.1419}simulation tool was used to compute the tight-binding Hamiltonian. A total of 62-bands were wannierized by taking projections on atomic sites as: Co (s, p, d), Ru (s, p, d), V (s, p, d), Si (s,p) etc. Further, Berry curvature, Fermi surface and anomalous Hall conductivity were calculated to investigate the semimetallic nature. The intrinsic anomalous Hall conductivity ($\sigma_{int}^{AHE}$) was estimated by integrating the Berry curvature (-$\Omega_{z} (\bf k)$) over the entire Brillouin zone considering a k-grid of $40\times40\times40$ with adoptive refinement k-mesh size of $5\times5\times5$.
To capture the effect of disorder in L2$_1$ structure, a 64-atom special quasi-random structure (SQS)\cite{zunger1990special} was generated. SQS is a carefully generated ordered structure, which mimics the random correlations up to a certain neighboring distance in disordered compounds. To generate the SQSs, Alloy Theoretic Automated Toolkit (ATAT)\cite{van2013efficient} was used. Our generated SQSs mimic the random pair correlation functions accurately up to third-nearest neighbors.
\section{Experimental results}
\subsection{Crystal Structure}
CoRuVSi crystallizes in LiMgPdSn prototype structure (space group $F\bar{4}3m$) with the measured lattice parameter of 5.80 {\AA} as found from the Rietveld refinement. The crystal structure can be viewed as four interpenetrating fcc sub-lattices with Wyckoff positions 4$a(0, 0, 0)$, 4$b(0.5, 0.5, 0.5)$, 4$c(0.25, 0.25, 0.25)$, and 4$d(0.75, 0.75, 0.75)$. In general, for a QHA XX$'$YZ, there exist three possible energetically non-degenerate structural configurations\cite{PhysRevB.105.144409} (keeping Z-atom at 4$a$-site) as follows:
\begin{itemize}
\item (I) X at 4$d$, X$'$ at 4$c$ and Y at 4$b$ site,
\item (II) X at 4$b$, X$'$ at 4d and Y at 4c site,
\item (III) X at 4$d$ , X$'$ at 4$b$ and Y at 4$c$ site.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width= 1.0\linewidth]{Fig2.png}
\caption{For CoRuVSi, room temperature powder XRD pattern including the Rietveld refined data for configuration-I with 50\% disorder between tetrahedral site Co/Ru atoms. Left inset shows a zoomed in view near super-lattice peaks (111) and (200) with L2$_1$ structure. Right insets show primitive unit cell structures corresponding to the Y-type order (top) and L2$_1$-type disorder (bottom).}
\label{fig:xrd-CFVG}
\end{figure}
For a detailed structural analysis, we consider the structure factor for configuration-I, which can be expressed as,
\begin{equation}
F_{hkl} = 4(f_Z + f{_Y}e^{{\pi}i(h+k+l)} + f{_X}e^{\frac{{\pi}i}{2}(h+k+l)} + f_{X'}e^{-\frac{{\pi}i}{2}(h+k+l)}).
\label{eq:sfactor}
\end{equation}
where $(h, k, l)$ are the miller indices. $f_X$, $f_{X'}$, $f_Y$, and $f_Z$ are the atomic scattering factors. The structure factor for super lattice reflections [111] and [200] can be written as:
\begin{eqnarray}
F_{111} &=& 4{[( f{_Z} - f_Y ) - i( f{_X} - f_{X'})]}\nonumber\\
F_{200} &=& 4[( f{_Z} + f_Y ) - ( f{_X} + f_{X'})]\nonumber
\label{eq:sfactor200}
\end{eqnarray}
Figure \ref{fig:xrd-CFVG} shows the room temperature XRD pattern of CoRuVSi along with the Rietveld refinement for configuration-I with 50\% disorder between tetrahedral site atoms i.e. Co/Ru (X/X$'$). This is the best fit we got after carrying out rigorous refinement considering all possible disorders in all the configurations. Clearly, the low intensity of the superlattice peak (111) indicates the possibility of disorder in the octahedral/tetrahedral sites. For L2$_1$-type refinement, we have also considered 50\% anti-site disorder between octahedral site atoms for configuration-I which did not fit well. The best fit with the lowest $\chi^2$ (1.80) was found in configuration-I with the L2$_1$ order (also see inset of Fig. \ref{fig:xrd-CFVG}) in comparison with other refinements considering all possible other disorders like $B2$ ($\chi^2$= 5.24), $A2$($\chi^2$ =10.5), D$O_3$ ($\chi^2$ =8.1) and perfectly ordered Y-type ($\chi^2$=4.35). As such, we conclude that CoRuVSi crystallizes in the L2$_1$ structure. The crystal structure corresponding to the Y-type order and the best fit with the L2$_1$ order are shown in the right insets of Fig. \ref{fig:xrd-CFVG}.
\subsection{Magnetic properties}
Figure \ref{fig:mt-CFVG}(a) shows magnetization (M) vs. temperature (T) for CoRuVSi measured at H $=500 $ Oe. The field cooled warming (FCW) curve taken at H$=500 $Oe shows a rapid increase in M below 25 K, which hints toward the possibility of magnetic ordering at very low T.
Inset of Fig. \ref{fig:mt-CFVG}(a) shows the Curie-Weiss (C-W) law fitting of the susceptibility data in high T range at H=500 Oe.
\begin{figure}[t]
\centering
\includegraphics[width= 9cm,height=6cm]{Fig3.png}
\caption{For CoRuVSi, (a) M vs. T in field cooled warming (FCW) mode in $H=500$ Oe. Inset shows $T$-variation of inverse susceptibility ($1/\chi$) along with Curie-Weiss fitting in high $T$-regime. (b) M-H curves at 3K and 300K.}
\label{fig:mt-CFVG}
\end{figure}
The magnetic moment ($m$) can be calculated considering the Slater-Pauling (S-P) rule using the total number of valence electrons ($n_v$) of the constituent elements.\citep{graf2011simple} The total moment ($m$) per formula unit can be expressed as:\cite{ozdougan2013slater,zheng2012band} $ m = (n_v - 24) \ \ \ \mu_B/f.u. $
For CoRuVSi, the S-P rule predicts $ m$ =2.0 $\mu_B$/f.u. in the fully ordered state, but interestingly, the M-H curve shows a very small saturation magnetization ($0.13$ $\mu_B$/f.u.) even at 3 K, which is a complete deviation from the S-P rule. The presence of L2$_1$ disorder can be a plausible reason for the quenching of moment in this system. To get an idea about the magnetic interactions present in this system, the inverse-susceptibility data (H=500 Oe) has been fitted (solid red line) above 150 K using the C-W law ($\chi^{-1}=\frac{1}{\chi_0 + C/(T-\theta_P)}$) and from the fitting, we obtained effective moment $ m$ =0.2 $\mu_B$/f.u., $\chi_0$=$0.40$ $emu/mole-$Oe and Weiss temperature, $\theta_P$= $-$93 K, that indicates the presence of antiferromagnetic interactions in the system. The sharp shoot below 25 K may arise because of the moments of ferri/antiferromagnetic clusters can easily prevail over the paramagnetic regime at low T.
Additionally, non-saturating behavior (up to 70 kOe field) of low-T M-H curve (Fig. \ref{fig:mt-CFVG}(b)), along with no hysteresis indicates superparamagnetic-like behavior in this system, attributable to the L2$_1$ disorder, which gives rise to the moment quenching.
Thus, magnetization data reveal the possibility of small magnetic clusters, formed by weakly interacting moments, with no spontaneous magnetization. This confirms the absence of coherent long-range ordering, mediated by the atomic disorder, giving rise to complex magnetic nature in CoRuVSi.\cite{PhysRevLett.59.586}
\subsection{Transport properties}
\subsubsection{PCAR}
The electronic spin polarization $P$ at the Fermi level ($E_F$) is defined as:
\begin{equation}
P = \frac{{n_{\uparrow}(E_F)} - {n_{\downarrow}(E_F)}}{{n_{\uparrow}(E_F)} + {n_{\downarrow}(E_F)}}
\end{equation}
where ${n_{\uparrow}(E_F)}$ and ${n_{\downarrow}(E_F)}$ are the spin-projected density of states at ($E_F$) for spin-up and -down channels respectively. Figure \ref{fig:ac} summarize the spin polarization data as obtained by the PCAR measurement. It shows a maximum spin polarization of $\sim$ 50\% at the Fermi level, which is reasonably high to serve as potential spintronic material.\cite{bainsla2015spin} \textcolor{black}{The reduction in spin polarization value as compared to the theoretical value corresponds to a narrowing of the spin gap in the density of states, which is possibly due to the presence of small density of states attributed to the disorder in the real system.}
\begin{figure}[t]
\centering
\includegraphics[width= 1.0\linewidth]{Fig4.png}
\caption{For CoRuVSi, point-contact Andreev reflection (PCAR) spectra, along with fit, and extracted parameters. The extracted m-BTK-model parameters are provided inside the box. The measured spin polarization was found to be $P$= 48(5)\%.}
\label{fig:ac}
\end{figure}
\subsubsection{Resistivity}
Figure \ref{fig:RT-CFVG}(a) shows the T-dependence of resistivity ($\rho_{xx}$) at different applied fields. It reflects a semi-metallic behavior (also revealed by the electronic structure calculations shown later). To gain further understanding, we have fitted the zero-field resistivity data considering various scattering mechanisms in various T-ranges. A dip-like feature below 5 K in the zero-field data indicates the possibility of weak localization arising from disorder in this system. Resistivity follows a power law behaviour ($\rho(T)=\rho_0 + AT^{n}$) in the T-range of 5 K$<$T$<$30 K, as shown in Fig. \ref{fig:RT-CFVG} (b). Above 30 K, resistivity data fit well with the two-carrier model which supports the semi-metallic behavior in this system (also supported by the carrier concentration from Hall data, shown later).
For further investigation, we have fitted the conductivity ($\sigma$) data (see Fig. \ref{fig:RT-CFVG}(c)) with a modified two-carrier model (Eq. \ref{eq:tbm-final}),\cite{kittel2007introduction,jamer2017compensated} in the T-range $30-310$ K.
A two-carrier model for $\sigma$ can be written as,
\begin{equation}
\sigma(T) = e (n_e \mu_e + n_h \mu_h)
\label{eq:tbm}
\end{equation}
where, $n_i=n_{i0}\ e^{-\Delta E_i/k_\mathrm{B}T}$($i=e,h$) are the electron/hole carrier concentrations with mobilities $\mu_i$ and pseudo-energy gaps $\Delta E_i$.
Eq.(\ref{eq:tbm}) can be further expressed as,
\begin{equation}
\sigma(T) = [A_e(T) \ e^{-\Delta E_e/k_\mathrm{B}T} + A_h(T) \ e^{-\Delta E_h/k_\mathrm{B}T}].
\label{eq:tbm-final}
\end{equation}
After fitting the conductivity data with the above equation, we obtained the pseudo-energy gaps for electrons and holes to be 0.11 meV and 15.9 meV, which are quite small and resemble those of a narrow band gap semiconductor. It appears that the atomic disorder plays a crucial role in significantly reducing the pseudo-gaps, especially for electrons which have an extremely small gap and are likely to become metallic with small perturbations (e.g. applied field, thermal fluctuation etc.).
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig5.png}
\caption{ For CoRuVSi, (a) Longitudinal resistivity ($\rho_{xx}$) vs. T in three different fields, 0, 50 and 90 kOe. (b) $\rho(T)=\rho_0 + AT^{n}$ fitting in the T-range 5-30 K. Inset shows a zoomed in view of the $\rho_{xx}$ data in the low-$T$ range. (c) longitudinal conductivity ($\sigma_{xx}$) vs. $T$ in zero field along with a two-carrier model fit between 30-310 K. (d) MR vs. $H$ at four different temperatures 5, 25, 50, and 100K. (e) linear and quadratic fitting of MR vs. H at 5 K and 25 K respectively.}
\label{fig:RT-CFVG}
\end{figure}
\subsubsection{Magnetoresistance}
Figure \ref{fig:RT-CFVG}(d) shows the field dependence of MR at different T, where MR is defined as MR(H)=$ \left[ \rho(H) - \rho(0)\right]/\rho(0)$ $\times 100\%$. At 5 K, non-saturating linear positive magnetoresistance (LPMR) is observed, which is also confirmed by the linear fitting of MR vs. H (see Fig. \ref{fig:RT-CFVG} (e)). But with increasing T, field-dependent MR(H) becomes almost quadratic in nature and at 25K an unsaturated quadratic MR is observed. The origin of LPMR at the lowest T (5 K) is possibly due to the zero/small-gap electronic structure near the E$_F$.\cite{PhysRevB.103.104427} MR magnitude decreases gradually with T and at 100 K it becomes almost zero. To characterize the type of carriers, we have further performed the Hall measurement (as described below).
\subsubsection{Hall Measurements} Figure \ref{fig:Hall-CFVG}(a) shows the field-dependence of Hall resistivity $\rho_{xy}$ at various T. Generally, Hall resistivity for a magnetic material has two contributions expressed as,
\begin{equation}
\rho_{xy}(T)=\rho_{xy}^{O} + \rho_{xy}^{A}=R_{0}H+R_{A}M,
\label{eq:Hall}
\end{equation}
where, $\rho_{xy}^{O}$ and $\rho_{xy}^{A}$ are ordinary and anomalous contribution to $\rho_{xy}$, $R_0$ and $R_A$ denote the ordinary and anomalous Hall coefficients respectively. At 5 K and 10 K, both the contributions are observed, but at 50 K, the amplitude of $ \rho_{xy}$ drops abruptly to zero (see Figs. \ref{fig:Hall-CFVG}(b-c)), as anomalous contribution die out with increasing T and only ordinary contribution remains. We have extracted $\rho_{xy}^{O}$ and $\rho_{xy}^{A}$ contributions at 5 K and 10 K, as shown in Fig. \ref{fig:Hall-CFVG}(b-c).
To scale the anomalous Hall effect (AHE) contribution, we have fitted a linear curve to $\rho_{xy}$ data at large H, and extracted $\rho_{xy}^{A}$ contribution. From the AHE data, a small AHE contribution ($\sim$ $0.15 \mu\Omega$-cm) is observed. Figure \ref{fig:Hall-CFVG}(d) shows the field-dependence of anomalous Hall conductivity ($|\sigma_{xy}^A|$ $\sim$ $\frac{\rho_{xy}^A}{\rho_{xx}^2}$) at 5 K and 10 K. $\sigma_{xy}^{A}$ reaches a maxima $\sigma_{xy0}^{A}=$45 S cm$^{-1}$ at 5 K, confirming a non-saturating and non-linear behaviour. The measured value of the carrier concentration$(n)$ at 5 K is 7.4$\times10^{18}$ cm$^{-3}$, which falls well within the range of carrier densities for semimetals/semiconductors, again indicating the semi-metallic nature of CoRuVSi.\cite{chen2021large} The positive slope of $R_0$ reveals holes as majority charge carriers. The origin of this behavior may be attributed to the change in electronic structure brought about the atomic disorder.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig6.png}
\caption{For CoRuVSi, (a) Hall resistivity ($\rho_{xy}$) vs. applied field (H) at 5, 10 and 50 K. (b-c) Ordinary and anomalous contributions of $\rho_{xy}$ at 5 and 10 K respectively. (d) Anomalous Hall conductivity ($\sigma_{xy}^{A}$) vs. H at 5 and 10K.}
\label{fig:Hall-CFVG}
\end{figure}
\subsubsection{Thermoelectric power }
Figure \ref{fig:TEP} shows the T-dependence of the Seebeck coefficient (S) (left y-scale) along with the power factor ($S^2\sigma$) (right y-scale). S shows a sub-linear variation with T, which is typically seen in semimetals \cite{pan2021thermoelectric,sk2022experimental} The negative slope of S with T corresponds to electron-driven thermopower, which again reveals two-carrier conduction in this system. The linear behavior of S suggests the dominance of diffusion thermopower. $\vert$S$\vert$ attains a value of 23 $\mu V/K$ at 300 K, which is comparable to that of other potential thermoelectric (TE) materials, at RT.\cite{yu2009high,hayashi2017structural,lue2002thermoelectric}
To further evaluate the potential of CoRuVSi for thermoelectric applications, we have calculated the power factor (PF=S${^2}\sigma$), a key parameter determining the efficiency of thermoelectric material. PF varies linearly with T and attains a maximum value of $0.7$ $m Watt/m-K^{2}$ at RT, which is reasonably high as compared to many Heusler-based TE materials,\cite{yan2011enhanced} and also comparable with other reported promising TE materials.\cite{hinterleitner2019thermoelectric,huang2015new,fu2013electron,PhysRevB.103.085202,PhysRevB.105.144409} To get an idea about the carrier density (n) and E$_F$, S-data is fitted with the equation S$_d$=S$_0$+s$T$ in the high-T regime, where S$_d$ is the diffusion thermopower, S$_0$ is a constant and s$=\frac{\pi^2k_{B}^{2}}{3e\text{E}_F}$. From this fitting, we obtained E$_F$ = $1.41$ eV and n=$7.2\times{10^{18}}$cm$^{-3}$, which is comparable with the Hall data and falls well within the range of carrier densities of promising TE materials.
Further investigation on high T measurements and thermal transport can help in determining the potential of CoRuVSi as a promising thermoelectric material at high T.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig7.png}
\caption{For CoRuVSi, $T$-dependence of thermoelectric power (S) and power factor ($S^2\sigma$) along with the fitting of diffusion thermopower (S$_d$=S$_0$+s$T$) between 100-300 K.}
\label{fig:TEP}
\end{figure}
\subsubsection{Specific heat}
Figure \ref{fig:hc} shows the T-dependence of specific heat for 0 and 50 kOe. The low-T C$_p$ data is fitted with the equation $C(T)=$ $\gamma$$T + $$\beta$T$^3$, where the first term is electronic contribution to C$_p$ while the second term is the low-T phonon contribution. The inset of Fig. \ref{fig:hc} shows {C$_p$/T} vs. T$^2$ plot along with the linear fit. From this fitting, we obtained $\gamma$=0.05 $J/mole-K^{2}$ (Sommerfeld coefficient), which in turn gives the density of states at E$_F$ i.e. $n(E_F)$=$3\gamma/(\pi^{2}k_{B}^{2})$ $\sim$ 4.5 states/eV f.u.\cite{venkateswara2019coexistence} This value matches quite well with the theoretical results (see next section) and is in good agreement with small DoS near E$_F$ for semimetals. This is another indication of expected semimetallic feature in CoRuVSi. From the fitting, we also extracted the Debye temperature, $\theta_D$= 383 K using the value of $\beta$=1.3795$\times$ $10^{-4}$ $J/mole-K^{4}$. Interestingly {C$_p$/T} vs. T$^2$ plot shows a shallow minimum, which may be related to the AFM-like interaction present in CoRuVSi.
\begin{figure}[b]
\centering
\includegraphics[width=1.0\linewidth]{Fig8.png}
\caption{For CoRuVSi, specific heat (C$_p$) vs. $T$ for 0 and 50 kOe fields. Inset shows {C$_p$/T} vs. $T^2$ along with the linear fit (solid red line) for zero field.}
\label{fig:hc}
\end{figure}
\section{Theoretical Results}
\begin{table}[b]
\centering
\caption{For ordered CoRuVSi, theoretically optimized lattice parameter ($a_0$), total and atom-projected magnetic moments ($ \mu_B$), and relative energy ($\Delta E$) for type I, II and III configurations with respect to the energy of type-I configuration.}
\begin{tabular}{l c c c c c c}
\hline \hline
Type& $a_0$ (\AA) $ \ $ & $m^{\mathrm{Co}}$ & $\ $ $m^{\mathrm{Ru}}$ $\ $ & $m^{\mathrm{V}}$ & $\ $ $m^{\mathrm{Total}}$ $ \ $ & $\Delta E$(eV/f.u.) \\ \hline
I & 5.82 & 1.68 & -0.13 & 0.53 & 2.07 & 0 \\
II & 5.85 & 1.32 & 0.73 & -0.21 & 1.84 & 0.32 \\
III & 5.83 & 1.35 & 0.28 & -0.5 & 1.16 & 0.22 \\
\hline \hline
\end{tabular}
\label{tab:theory-CFVG}
\end{table}
We have used {\it ab-initio} simulation to investigate various magnetic states including para-, ferro-, antiferro-, and ferri-magnetic configurations in the ordered and L2$_1$-disordered phases for CoRuVSi. Out of all the configurations, type-I configuration (see Sec. IV(A)) with ferrimagnetic ordering turned out to be energetically the most favorable one. Table \ref{tab:theory-CFVG} shows the optimized lattice parameters, total and atom-projected moments and relative energies of three different ordered structures (type-I, II and III) in their respective lowest energy magnetic ground state. Figure \ref{fig:CRVS-band} shows the spin-resolved density of states and band structure for the lowest energy type-I configuration, which indicates a nearly half-metallic ground state with a net magnetization of $\simeq$2\ $\mu_B/f.u. $.
\begin{figure}[t]
\centering
\includegraphics[width= 1.0\linewidth]{Fig9.png}
\caption{For ordered CoRuVSi (in type-I configuration), spin resolved band structure and density of states (DoS) at the optimized lattice parameter ($a_0$). A few electron pockets at/around the E$_F$ are observed for minority spin-channel.}
\label{fig:CRVS-band}
\end{figure}
In order to further explore the half metallic/semimetallic nature, we have simulated the band structure of the ordered CoRuVSi (type-I configuration) including the spin-orbit coupling (SOC), as shown in Fig. \ref{fig:w90-1}(a). The corresponding band crossing near the Fermi level are shown in Fig. \ref{fig:w90-1}(b). This clearly illustrates the semimetallic nature with the bands 25, 26 and 27 crossing E$_F$. The Fermi surfaces corresponding to these three individual bands as well as the net combined Fermi surface are shown in Fig. \ref{fig:w90-1}(c-f). This confirms the emergence of hole pockets from bands 25 and 26, while band 27 gives rise to electron pocket. Figure \ref{fig:w90-2}(a) shows the z-component of the Berry curvature ($\Omega_z(\mathbf{k})$) along the high symmetry $\vec{k}$-points. The corresponding 2D projection of $\Omega_z(\mathbf{k})$ in k$_x$-k$_y$ plane is shown in Fig. \ref{fig:w90-2}(b). Here, black solid lines show intersections of the Fermi surface with this plane.
The large spike in the Berry curvature near the vicinity of $L$ point is attributed to the two spin-semimetallic bands (25 and 26), one of which is unoccupied (band 25) and the other (band 26) is occupied in a small k-interval. Due to spin-orbit coupling, a small energy gap opens up, giving rise to a small energy denominator in the definition of Berry curvature (i.e. $\Omega_n$ $\sim$ $1/(\Delta {\varepsilon_n}^2)$ from the Kubo-formula)\cite{PhysRevLett.92.037204}. So, these topological spin-semimetallic bands induce an appreciable Berry curvature, which is purely intrinsic in nature. The intrinsic anomalous Hall conductivity is calculated by integrating $\Omega_z(\mathbf{k})$ over the entire Brillouin zone (BZ), using the following expression,\cite{PhysRevLett.92.037204}
\begin{equation}
\sigma_{int}^{AHE}= - (e^2)/(8\pi^3\hbar) \int_{(BZ)} d^3k\ \Omega_z(\mathbf{k}),
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Fig10.png}
\caption{For ordered CoRuVSi (in type-I configuration), (a) band structure including spin-orbit coupling, (b) widths of various bands, illustrating the band-crossing near $E_F$. Fermi surfaces attributed to (c) band 25, (d) band 26, (e) band 27, and (f) combined bands, illustrate the emerging electron and hole pockets at the high symmetry points confirming the semimetallic feature.}
\label{fig:w90-1}
\end{figure}
The simulated AHE for ordered CoRuVSi (type I configuration) is $\vert\sigma_{int}^{AHE}\vert$=102 $S/cm$, a reasonably high value. The calculated $\vert\sigma_{int}^{AHE}\vert$ is almost double as compared to the experimentally measured value (45 S/cm).
It is important to note that the simulated net magnetization ($\simeq$2 $\mu_B/f.u. $) of the completely ordered phase is quite different as compared to the measured value ($\simeq$ 0.13 $\mu_B/f.u.$) To unveil the possible reason for these discrepancies between theory and experiment, we have simulated the band structure including Berry curvature, anomalous Hall conductivity and Fermi surface of the L$2_1$-disordered phase (50\% disorder between tetrahedral site atoms Co and Ru, as confirmed by XRD-refinement) of CoRuVSi using a 64 atom SQS cell.
This disordered structure gives a reduced magnetization of 0.29 $\mu_B/f.u. $ with a nearly compensated ferrimagnetic structure (see Table \ref{tab:theory-CRVS} for the optimized lattice parameter and the atom projected and total moments). This value of net moment agrees fairly well with our experimental finding. Interestingly, in the L2$_1$ disordered structure, CoRuVSi shows a spin-semimetal behavior (see Fig. \ref{fig:soc}(a)). This is due to a small overlap between the conduction and valence bands (CB and VB) close to the E$_F$, which in turn is tunable by the influence of impurity/disorder or external field, and hence plays a crucial role in the overall electronic structure of the material. To crosscheck the effect of SOC on SSM behavior, we have also simulated the electronic structure including SOC effect. This is shown in Fig. \ref{fig:soc}(b,c).
A close inspection of this band structure in the disordered phase reveals the existence of a linear band crossing at/around -0.4 eV below E$_F$, supported by a band inversion near the $X$ point (see the inset of Fig. \ref{fig:soc}(c)). The simulated Berry curvature, Fermi surface (for band \#25, 26) and the band positions for the L2$_1$ disordered phase of CoRuVSi are shown in Fig. \ref{fig:13}.
Interestingly, we also found remarkable agreement between the calculated AHE ($\vert\sigma_{int}^{AHE}\vert$=52.2 $S/cm$) of the L2$_1$ partially disordered phase and the corresponding experimental value ($\vert\sigma_{int}^{AHE}\vert$=45 $S/cm$). The semimetallic bands, which give rise to the topological non-trivial features are found to be robust against the disorder as indicated by the Berry curvature and Fermi surface calculations.
\begin{figure}[b]
\centering
\includegraphics[width=0.85\linewidth]{Fig11.png}
\caption{For ordered CoRuVSi (in type-I configuration), simulated Berry curvature (-$\Omega_{z}(\bf k)$) (a) along the high symmetry paths and (b) in the k$_x$-k$_y$ plane at E$_F$. Black solid lines show intersections of the Fermi surface with this plane. }
\label{fig:w90-2}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width= 1.0\linewidth]{Fig12.png}
\caption{For L$2_1$ partially ordered CoRuVSi (SQS structure) (a) spin resolved density of states without spin orbit coupling (SOC) at the optimized lattice parameter($a_0$) (b-c) DoS and band structure including the effect of SOC. Red circle in Fig. (c) highlights the linear band crossing at/near $X$ point with the band inversion, again confirming semimetallic nature. Inset shows a zoomed in view of this band crossing.}
\label{fig:soc}
\end{figure}
\begin{table}[t]
\centering
\caption{For L$2_1$ partially ordered CoRuVSi (SQS structure), optimized lattice parameter ($a_0$), total and atom-projected average moments ($ \mu_B$).}
\begin{tabular}{l c c c c }
\hline \hline
$a_0$ (\AA)\ \ \ \ \ \ \ \ & $m^{\mathrm{Co}}$ \ \ \ \ \ \ \ \ & $m^{\mathrm{Ru}}$ \ \ \ \ \ \ \ \ & $m^{\mathrm{V}}$ \ \ \ \ \ \ \ \ & $m^{\mathrm{Total}}$ \\ \hline
5.85\ \ \ \ \ \ \ \ \ \ & 0.29\ \ \ \ \ \ \ \ \ \ & -0.10\ \ \ \ \ \ \ \ \ \ & 0.10\ \ \ \ \ \ \ \ \ \ & 0.29 \\
\hline \hline
\end{tabular}
\label{tab:theory-CRVS}
\end{table}
\begin{figure}[b]
\centering
\includegraphics[width=\linewidth]{Fig13.png}
\caption{For L$2_1$ partially ordered CoRuVSi (SQS structure), simulated Berry curvature (-$\Omega_{z}(\bf k)$) (a) along the high symmetry paths and (b) in the k$_x$-k$_y$ plane at E$_F$. Black solid lines show intersections of the Fermi surface with this plane. Fermi surfaces attributed to (c) band 25 and (d) band 26, illustrate the emerging electron and hole pockets again confirming the semimetallic feature. (e) widths of various bands, illustrating the band-crossing at/near $E_F$.}
\label{fig:13}
\end{figure}
\section{Summary and Conclusion}
In summary, we report the identification of a new member, namely CoRuVSi, to the quantum material class namely {\it spin semi-metals} which can be quite promising for future spintronic and thermoelectric applications. Using a combined theoretical and experimental study, we have investigated the structural, magnetic, transport, and electronic properties of CoRuVSi. It crystallizes in the cubic structure (space group $F\bar{4}3m$) with a partial L2$_1$-type disorder in the tetrahedral site atoms Co/Ru, as confirmed from our XRD measurement. The magnetization data indicate a weak ferrimagnetic ordering at low T, with a very small moment $\sim$ 0.12 $\mu_B$/f.u. caused by the anti-site disorder. Resistivity results provide a strong evidence of semimetallic nature dominated by two-band conduction, while low-T magnetoresistance data indicate the non-saturating, linear positive magnetoresistance. The latter hints toward the small-gap electronic structure near the Fermi level, indirectly supporting the prediction of semimetallic nature. Specific heat data confirm a low value of density of states at/near E$_F$, supporting our theoretical findings about the semimetallic nature. PCAR measurements reveal a high spin polarization of $\sim$ 50\%. CoRuVSi also shows a high thermopower value of $0.7$ $m Watt/ m-K^{2}$ at room temperature, rendering it as a promising thermoelectric material as well. {\it Ab-initio} simulation of CoRuVSi with L2$_1$ disorder reveals a spin semimetal feature with nearly compensated ferrimagnetic configuration having a small net magnetization, as observed experimentally. Interestingly, the band structure hosts a linear band crossing at $\sim$-0.4 eV below the Fermi level, along with a band inversion, confirming the topological non-trivial nature of CoRuVSi. \textcolor{black}{This was further assessed from the simulated Berry curvature, anomalous Hall conductivity and Fermi surface.} The simulated anomalous Hall conductivity for L2$_1$ partially ordered CoRuVSi is 52 S/cm which agrees fairly well with experimentally measured value of 45 S/cm. The coexistence of many promising features in a single material is rare and hence it opens up new opportunities to search for other novel materials with multifunctional properties.
{\it \bf Acknowledgments:} JN acknowledges the financial help provided by IIT Bombay. JN also thanks Mr. Vinay Kaushik UGC-DAE-CSR Indore for setting up HC measurements. The authors thank Dr. Durgesh Singh for setting up TEP measurements. KGS acknowledges the funding from the Indo-Russian Project-TPN- 64868. A.A. acknowledges DST-SERB (Grant No. CRG/2019/002050) for funding to
support this research.
\bibliographystyle{apsrev4-2}
|
\section{Introduction}
Drug combination therapy has been widely applied in both traditional and modern medicine due to its diverse merits. Compared with monotherapy, administering drug combinations leads to improvement of efficacy \cite{csermely2013structure}, and reduction of side effects \cite{zhao2013systems} and host toxicity \cite{o2016unbiased}, further, it even overcomes drug resistance \cite{hill2013genetic}. Considering the fact that a single drug usually cannot be effective, drug combinations are increasingly used to treat a variety of complex diseases, such as human immunodeficiency virus (HIV) \cite{clercq2007design}, virus infections \cite{zheng2018drug}, and cancer \cite{kim2021anticancer, al2012combinatorial}. For instance, the combination of two clinically used drugs, colloidal bismuth subcitrate (CBS) and N-acetyl cysteine (NAC), suppresses the replication cycle of SARS-CoV-2 virus and reduces viral loads in the lung \cite{wang2022orally}. The combination provides a potential treatment for combating SARS-CoV-2, which can hardly be treated by any single drug. However, drug combinations can also be harmful without precise medicine \cite{hecht2009randomized, azam2021trends}. Therefore, it is pretty important to accurately find synergistic drug pairs for a cell type in case we want to take advantage of drug combination therapy.
Traditional methods for drug combination discovery are mainly based on clinical trial and error, which is time- and cost-consuming and can result in harm to patients \cite{day2016approaches, pang2014combinatorial}. Besides, the limited resources only satisfy web-lab tests on a few drug combinations \cite{li2015large}.
With the development of experimental technology, researchers are able to carry out high-throughput drug screening (HTS) \cite{macarron2011impact, torres2013high, he2018methods}, which is a kind of sensitive and fast synchronous experiment and makes the exploration of large drug combination space become a reality.
Due to HTS, the drug combination synergy data had increased tremendously. Some public databases make contributions to drug research for specific tissues, like ASDCD \cite{chen2014asdcd} provides antifungal drug combinations data, and a large HTS synergy study \cite{o2016unbiased} performed more than 20000 drug synergy measurements, which covers 38 drugs and 39 cancer cell lines. Part of these databases offer high quality training data for the development of computational methods, and also help the evaluation of computational methods for predicting novel drug combinations.
However, the in vivo and in vitro experiments cannot be exactly consistent. Although the original tumor and the derived cancer cell line share a high degree of genomic correlation, in vitro experiments are not able to restore the mode of drug action in vivo \cite{ferreira2013importance}, which means there still exist obstacles impede the effectiveness of HTS.
In recent years, with the advance of computational technology, some machine learning models and neural networks are effective and promising in finding novel drug combination candidates in large synergistic space. For example, DeepSynergy \cite{preuer2018deepsynergy} combined the three different types of chemical features of drugs and genomic information of cancer cells to predict drug pairs with synergistic effects. TranSynergy \cite{liu2021transynergy} is a transformer-based method that integrates information from gene-gene interaction networks, gene dependencies, and drug-target associations to predict synergistic drug combinations and deconvolute the cellular mechanisms. DeepDDS \cite{wang2022deepdds} converts molecular drugs into graphs and proposes a graph neural network with an attention mechanism to identify the synergistic drug combinations. MR-GNN \cite{xu2019mr} extracts features from different neighborhoods of each node in the drug molecular graph and uses a dual graph-state LSTM-based network to process features.
Some other methods like DeepDDI \cite{ryu2018deep}, DeepDrug \cite{yin2022deepdrug}, and GCN-BMP \cite{chen2020gcn}, focus on resolving drug-related tasks like drug-drug interaction, drug-food interaction, and drug relations, which provides useful information for synergistic drug combination prediction task.
Nevertheless, their prediction target usually remains in a specific pathway, cell line, or tissue because of the limitation of their used dataset. These studies are usually based on single databases, and the prediction was also made within the database. If we want to develop an unbiased and generalizable drug synergy prediction model, one of the key challenges is the problem of domain-shift data: the invitro drug responses of different tissues can be various. Previous methods focus on studying common tissues, like breast, skin, and lungs \cite{tang2019network, chen2018predict, zhao2011prediction, huang2014drugcomboranker}, these methods use drug combinations data of certain cell lines for training and try to discover novel drug combinations from other cell lines within the same tissue. Therefore, some tissues remain understudied due to some difficulties in data or bio-experiments. For instance, bone cancer is hard to deal with because of the technical limitations in culturing bone tissue as cell lines. The lack of cell line models leads to the obstacle in high-throughput screening, which in turn makes these tissues more difficult to study. As a consequence, finding out a way to develop a generalized drug combination effect prediction model is essential for resolving understudied tissues problem. Fortunately, with the help of the databases that systematically integrate multiple drug synergy datasets \cite{zagidullin2019drugcomb, liu2020drugcombdb}, the roadway to developing unbiased drug synergy prediction models was not that elusive.
In this stud
, we address the problems mentioned above by proposing an end-to-end deep learning framework that accurately predicts synergistic effects. Our method takes advantage of multi-modal data, graph neural networks, and large-scale unsupervised training to integrate and learn useful information for synergistic prediction. Specifically, our model takes chemical structure graphs of drugs and the protein expression of cell lines as input and applies a pre-trained molecular graph transformer \cite{li2022kpgt} to convert drug graphs into embeddings. Meanwhile, the model generates embeddings for every protein in the expression by utilizing a protein language model \cite{rives2021biological}. To enrich more features, we also include disease information, particularly, we apply RotatE \cite{sun2019rotate} to get the embedding of disease from PrimeKG \cite{chandak2022building}. Next, we utilize graph neural networks and take our generated embeddings as node representations. In order to inference on unseen drugs, we include drug-drug similarity edge and drug-target module/ drug-drug interaction module to generate pseudo edges and formed a refined graph with richer information. Finally, a synergistic prediction head is built on top of our graph and acts as a Perceptron (MLP) to predict the synergistic effect. We also include a self-training strategy to make use of the large amount of data in combination space.
We compare our model with five task-related deep learning models and two traditional machine learning models on the benchmark datasets, DrugComb and AstraZeneca. Our framework outperforms all the previous methods on various evaluation criteria. Moreover, we conduct experiments on some unseen drugs and cell lines in the training set to justify our robustness. We believe that our method is an effective tool for discovering novel synergistic drug combinations for further wet-lab experiment validation.
\section{Methodology}
In this section, we will dive into details of the dataset construction and how we establish our training pipeline. We divided this part into several subsections: Preprocessing section describes datasets manipulation and feature pre-training; Heterogeneous graph section delivers graph construction, graph neural network, and synergistic prediction head information; Graph structure learning section introduces our Drug-Target predictive module, Drug-Drug interaction module, and graph structure learning details; Self-training and inference section summarizes our self-training strategy and the way to perform inference.
\begin{figure}[!t]
\centering
\includegraphics[width=1\textwidth]{Untitled-2.png}
\caption{
\textbf{Overview of our drug synergistic combinations prediction framework.} \textbf{a.} We obtain features of drug, protein, and disease from three pre-trained models respectively, and build a heterogeneous graph upon these node features. Next, we perform drug-drug and drug-target inference to propagate information and refine the graph. The finalized embeddings of drugs and the cell line are obtained through a graph neural network based on the refined graph. The embedding vectors are subsequently concatenated to feed into a Multilayer Perception (MLP) to predict the drug combination synergy effect. \textbf{b.}Details of our self-training strategy to make use of large unlabelled synergistic data. \textbf{c.} The architecture of our Drug-Target predictive module, and our Drug-Drug interaction predictive module is built in the same way.}
\label{Fig.overview}
\end{figure}
\subsection{Preprocessing}
We take advantage of various types of datasets to meet our requirements for diverse information, in this section, we introduce each of the used datasets in detail by the order of our framework pipeline.
\paragraph{Datasets}
The datasets we used include originally published datasets as well as some following databases which integrated them together. They covered multiple aspects and biological relations, here we listed them below,\begin{itemize}
\item \textbf{\textit{Zinc250k\cite{irwin2005zinc}/ChemBL}}: Large scale drug SMILES dataset used for unsupervised training.
\item \textbf{\textit{PrimeKG\cite{chandak2022building}}}: A knowledge graph dataset integrating 20 high-quality datasets, biorepositories, and ontologies.
\item \textbf{\textit{Therapeutics Data Commons\cite{huang2021therapeutics}}}: TDC is a resource platform to access curated AI-ready datasets, machine learning tasks, and benchmark datasets.
\item \textbf{\textit{DrugComb\cite{zagidullin2019drugcomb}}}: DrugComb is an open-access, community-driven data portal where the results of drug combination screening studies for a large variety of cancer cell lines.
\item \textbf{\textit{AstraZeneca}}: AstraZeneca is an independent dataset published by AstraZeneca company and involves several drug combination screening studies on multiple cancer cell lines.
\end{itemize}
After we obtained the above datasets, we need to combine equivalent entries and delete duplicates. \begin{itemize}
\item Drug data are unified by their DrugBankID or ChemBLID across all our datasets and their corresponding SMILES representations are unified to their canonical form.
\item Protein data in PrimeKG and TDC are aligned via their Gene ID, Uniprot ID, and Protein ID.
\item Our cell line data are originally expressed on 13418 proteins obtained via CCLE dataset. After examining existing protein data in PrimeKG and TDC, we delete 10 protein columns in our cell expression.
\end{itemize}
Finally, after pre-processing, we can obtain around 22,000 numbers of protein/genes, 2000 Drugs, and 15,000 Diseases. They can be treated as nodes in our graph neural network but we still lack an important ingredient - their representations. We gained their representations through large-scale pre-training methods explained below.
\paragraph{Feature Representation}
Unsupervised learning shows brilliant results in the field of Natural Language Processing and bioinformatics, many sequence-based biological tasks benefit from it. Inspired by the progress of unsupervised frameworks, we take advantage of three models with large-scale information: ESM-1b \cite{rives2021biological}, KPGT \cite{li2022kpgt}, and PrimeKG \cite{chandak2022building} to obtain the initial embedding of our protein, drug, and disease respectively.
\textbf{\textit{Protein: ESM-1b.}}
ESM-1b is trained on Uniprot \cite{uniprot2015uniprot} database, utilizing 250 million protein sequences, which is large enough to
support the high-capacity protein language model. And deep transformer is chosen as the architecture, for its
great performance in many tasks. During the training process, the amino acid sequences extracted from database are further divided into different fractions and given a special mask token as inputs of model, and the output of neural network is the missing token of corresponding sequences. With the pretrained ESM-1b, we input a protein sequence and get residue-level sequence embeddings, and then average across all residue positions of such embeddings, so that we finally get a $768$-dimension feature for the input sequence.
\textbf{\textit{Drug: KPGT.}}
KPGT is trained on two million molecular SMILES from ChEMBL29 dataset \cite{gaulton2017chembl} for the representation learning of molecular graphs. With the Line Graph Transformer (LiGhT) structure, the model focuses on chemical bonds and can capture the structural information of molecular graphs. And the knowledge-guided pre-training strategy helps to exploit the additional message like abundant structural and semantic information, which are potentially more important to the downstream tasks, inside the molecular graphs. Given a SMILES string, the KPGT can convert it into a molecular line graph and input it to LiGhT. At last, we can get the $2304$-dimension embedding of our input SMILES string.
\textbf{\textit{Disease: RotatE.}}
Precision Medicine Knowledge Graph (PrimeKG) presents a holistic view of biological factors including diseases. To generate disease embeddings, we apply RotatE \cite{sun2019rotate} on top of PrimeKG to learn and gather the information. In our task, RotatE defines the disease relation as a rotation from the source node to the target node in the complex vector space, and the method is effective in modeling three relation patterns: symmetric/antisymmetric, inversion, and composition, therefore, it is able to resolve all relations in PrimeKG. Finally, we obtain a $512$-dimension embedding for each disease node.
\textbf{\textit{Cell line: Depmap CCLE}}
DepMap provides a “cancer dependency map” CCLE dataset by systematically identifying genetic dependencies. Our cell line data retrieved from CCLE are expressed on around 13400 genes/proteins. After preprocessing, a cell line therefore can be represented by a 13400-dimension vector, with each entry referring to the value of the corresponding gene/protein. We can then take advantage of our pre-trained protein embeddings and replace our original cell representation with the summation of weighted protein embeddings as shown below:
\begin{eqnarray}
h_{j} = \sum_{i \in protein} CCLE_{j}[i] \cdot E_i, j\in Cell,
\end{eqnarray}
where $CCLE_{j}$ is the CCLE 13400 dimension gene/protein expression vector of a cell $j$, $E_{i}$ corresponds to the $768$ dimension ESM-1b embedding for protein $i$, and $h_j$ is our derived cell line embedding. In this way, we encode richer information to our new cell line representations considering both gene expression and gene structural/co-evolutionary information.
\subsection{Heterogeneous Graph}
As shown in Figure \ref{Fig.overview}, after we obtain the initial representation for each node, we then construct the heterogeneous graph based on relations extracted from PrimeKG and TDC. In total, there are nine edge types and three node types in our graph. We listed all edge types in Table \ref{graph} including their connecting nodes and descriptions. Note that all edge types except Drug-Drug Similarity can be inferred in our datasets directly, we include Drug-Drug Similarity to benefit information propagation in our graph, especially for unseen drugs. Drug-Drug Similarity here is computed by measuring the distance between our KPGT drug embeddings as well as fingerprints Tanimoto similarity, if the distance between two drugs is smaller than $90$ or their similarity greater than $0.62$, we connect these drugs with Drug-Drug Similarity edge. Finally, with these relation types and nodes, we can construct our heterogeneous graph $G = (A,X)$ where $A,X$ here stands for the initial adjacency matrix and node embeddings.
\textbf{\textit{Graph neural networks}} $\psi$ are built upon our constructed $G$ to conduct message passing. Here in our framework, a simple Multilayer perception module and a Graph attention architecture are applied. We noticed that this simple architecture is adequate to reach similar performance \cite{10.1145/3447548.3467350} compared to computation demanding heterogeneous GNN framework such as MAGNN, graph-transformer, etc. Each node embeddings $\{Drug: 2304,\: Protein: 768, \: Disease:512 \}$ propagate through different MLP modules to form unified vectors with length $512$. Take protein embeddings as an example, an MLP takes 768-dimension vectors as input and outputs 512-dimension vectors for network propagation. These updated vectors with the same length of each instance served as
their new embeddings, and then these embeddings proceed through a homogeneous graph attention network to propagate
information as shown in Figure~\ref{Fig.overview}. We inserted three GNN layers in our framework. One is directly after the initial graph construction and computes upon the adjacent matrix $A$. Another two are after our graph refinement and compute upon the refined adjacent matrix $A^{*}$.
\textbf{\textit{Synergistic Prediction Head}} $f_{syn}$ is located at the last part of our framework. With the final embedding vectors of our graph node $X^{*}$, we can extract our final drug representations as well as the cell line representations being the weighted sum of its expressed genes. Then two drug embedding vectors and one cell line vector are sent through MLP, they are concatenated as the input of multiple fully-connected layers. The output of this MLP module is two prediction scores for synergistic or antagonistic classes. The label of the synergistic effect was computed by the softmax function that follows the output of prediction scores.
\begin{table}[!t]
\centering
\caption{Nodes and edges in our heterogeneous graph}
\label{graph}
\scalebox{1}{
\begin{tabular}{cccccc}
\toprule
Relation & Node & Node & Description \\
\midrule
Drug-Drug Interaction P & Drug & Drug &Summarized Drug-Drug effect obtained from DrugBank \\
Drug-Drug Interaction N & Drug & Drug &Summarized Drug-Drug effect obtained from DrugBank \\
Drug-Drug Similarity & Drug & Drug &Drug similarity based on fingerprints and KPGT.\\
Drug-Target Interaction & Drug & Protein/gene &Drug target interaction on specific proteins.\\
Drug-Disease Indication & Drug & Disease &Drug Disease interaction based on effect. \\
Drug-Disease Contraindication & Drug &Disease &Drug Disease interaction based on effect.\\
Protein-Protein & Protein/Gene & Protein/Gene &Protein protein interaction obtained from TDC and PrimeKG\\
Protein-Disease & Protein/Gene & Disease &Where disease is related with specific protein/gene.\\
Disease-Disease & Disease & Disease &Disease Disease interaction.\\
\bottomrule
\end{tabular}}
\end{table}
\subsection{Graph Structure Learning}
Graph structure learning(GSL) has emerged as a new technique for learning adaptive graphs when only insufficient data about the graph is available. Here, we introduce GSL into our framework in order to learn new pseudo edges to conduct inductive inference for unseen drugs during testing. Moreover, experimental results indicate that even the prediction accuracy of current nodes will increase by adding pseudo edges since it benefits information propagation. While there are 9 types of edges in our graph, Drug-Target interaction and Drug-Drug interaction are the two most informative edge types for unseen drugs. The idea of our graph structure learning is thus built upon a predictive module which consists of a Drug-Target interaction module and a Drug-Drug interaction network to generate pseudo edges. We will illustrate how these two modules are constructed and trained below.
\textbf{\textit{Drug-Target Interaction Module.}}
The most informative relation in our graph lies in the Drug-Target edges since they update the drug and cell line embeddings directly. Our Drug-Target interaction (DTI) module is pre-trained separately and finetuned into our framework. The data we utilized for DTI training are collected via PrimeKG and TDC which originates from BindingDB \cite{liu2007bindingdb}. The data pairs are in $(Drug, Protein)$ and form existing interaction. Thus, we trained an MLP classifier that takes two embeddings as input and outputs the prediction score. The DTI module architecture is presented in Figure \ref{Fig.overview}, it is a simple stack of several attention blocks and multiple perceptron layers. KPGT drug embeddings and ESM-1b protein representations go through separate computation branches and concatenate together through a predictive head. We train our DTI module with all positive existing data pairs and run negative sampling three times the number of positive pairs. The negative sampling factor can be modified, and we set it to be larger than one since we tend to control the numbers of predicted pseudo edges not to explode.
\textbf{\textit{Drug-Drug Interaction Module.}} Another informative relation in our graph is Drug-Drug interaction (DDI) edges since they also directly update the drug embeddings. Our Drug-Drug interaction module is also pre-trained separately and fine-tuned into our framework. The data we utilized for DDI training are collected via PrimeKG and TDC which originates from DrugBank and Twosides \cite{tatonetti2012data}. The data pairs are in $(Drug, Drug)$ form with $P$ or $N$ labels summarizing the interaction effect. The architecture is presented in Figure~\ref{Fig.overview}. Likewise, we trained an MLP classifier that takes two embeddings as input and outputs the predicted category $(P,N,$no edge). The DDI module architecture and training strategy are similar to DTI.
\textbf{\textit{Graph Structure Learning.}} After our DTI and DDI predictive module is pre-trained. We can fit them into our framework to generate pseudo edges and tune our framework end to end. Suppose our original heterogeneous graph $G$ can be presented by its adjacency matrix $A$, then the refined graph $A^{*}$ can be gained via our predictive module, $A^{*} = g(f_{DTI}, f_{DDI}, A)$, $g$ here act like a $sgn$ function,
\begin{eqnarray}
A^{*}_{ij} = g(f_{DTI}, f_{DDI},A ) = \left\{
\begin{aligned}
f_{DTI}(X_{i},X_{j}) & , & e_{ij} = e_{DTI} \: and \: A_{ij}=0, \\
f_{DDI}(X_{i},X_{j}) & , & e_{ij} = e_{DDI} \: and \: A_{ij}=0, \\
A_{ij}& , & Otherwise
\end{aligned}
\right.
\end{eqnarray}
Finally, the refined graph passes through a graph neural network and enters our synergistic prediction head. We denote the initial embedding for each node as $X$, and the finalized embedding as $X^{*}$. Thus, $X^{*} = \psi(A^{*}, X)$. The whole framework is then tuned according to the loss below:
\begin{eqnarray}
L &= &BCE(Y,f_{syn}(X^*))= BCE(Y,f_{syn}(\psi(A^{*}, X))\\\nonumber
&= &BCE(Y,f_{syn}(\psi(g(f_{DTI}, f_{DDI}, A), X)) \\ \nonumber
min\: L &=& \min_{f_{syn},\psi,f_{DTI}, f_{DDI}} BCE(Y,f_{syn}(\psi(g(f_{DTI}, f_{DDI}, A), X))
\end{eqnarray}
$Y$ here refers to the ground truth synergistic labels and $\psi$ stands for graph neural network which propagates messages on $A$ and $A^{*}$. $BCE$ stands for the binary cross entropy function to calculate our corresponding classifcation loss. Thus, the submodules that can be optimized in our framework include our synergistic prediction module $f_{syn}$, graph neural network $\psi$ and pre-trained predictive module $f_{DTI}, f_{DDI}$.
\subsection{Self-training and inference}
Self-training has shown positive effects in limited-data supervised learning tasks. Here in our case, although Drug-Comb provides over 300,000 entries, the combinatorial search space actually consists of over 0.7 billion possible cases. Labeled data doesn't even occupy $0.1\%$ of the whole space. Thus, we believe by expanding training data through self-training, our performance can be certainly boosted forward. Our main idea lies in using predicted confidence scores to filter out new training data. Figure \ref{Fig.overview} visualizes this procedure in our framework. First, we train our model on the original dataset $S$. Then, we run inference on the 0.7 billion combinatorial search space and obtained those entries $U$ whose confidence scores are greater than 0.8, we controlled the number of $U$ to be smaller than our original dataset $S$ and merged them together: $S^{'} = S\cup U$, to make a new training set. We then retrained our model on $S^{'}$. This process converges until our re-trained model almost cannot gain improvement.
After our framework is trained, for an incoming triple $(Drug_A,Drug_B,Cell\_line)$, inference can be conducted easily on our method. \begin{itemize}
\item First, generate both drug embeddings for $Drug_A,Drug_B$.
\item Inspect whether $Drug_A,Drug_B$ are in our graph. If yes, remained unchanged else let $Drug_A$ to be an unseen drug. Involve $Drug_A$ into our graph $G$ and generate Drug-Drug similarity edge for $Drug_A$ and obtain graph $G'$.
\item Run GNN on $G$ or $G'$, conduct Drug-Target, Drug-Drug interaction inference and generate pseudo edges with refined graph $G^{*}$.
\item Run GNN on $G^{*}$, and gained the finalized embedding $X^{*}$
\item Conduct synergistic prediction based on $f_{syn}$ and $X^{*}$.
\end{itemize}
\section{Results}
\begin{table}[!t]
\centering
\caption{Performance comparison of 10-fold cross-validation on DrugComb dataset.}
\label{test_set}
\scalebox{1.1}{
\begin{threeparttable}
\begin{tabular}{p{2.6cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}}
\toprule
\textbf{Metric} & \textit{AU ROC} & \textit{AU PRC} & \textit{ACC} & \textit{BACC} &\textit{Precision}\tnote{a} &\textit{F1-Score} \\
\midrule
\textbf{Ours} & \textbf{0.961} & \textbf{0.954} &0.878 & \textbf{0.862} & 0.883 & \textbf{0.972} \\
DeepDDS\cite{wang2022deepdds} & 0.942 & 0.934 &0.865 &0.855 & 0.862 &0.957 \\
TranSynergy\cite{liu2021transynergy} & 0.912 & 0.918 &0.892 &0.814 & 0.845&0.924 \\
DeepSynergy\cite{preuer2018deepsynergy} & 0.894 & 0.882 & \textbf{0.894} &0.862 & 0.843 &0.894 \\
MR-GNN\cite{xu2019mr} & 0.935 & 0.917 &0.885 &0.891 & \textbf{0.916} &0.904 \\
MatchMaker\cite{kuru2021matchmaker} & 0.927 & 0.914 &0.853 &0.876 & 0.786 &0.893 \\
XGBoost\cite{chen2016xgboost} & 0.802 & 0.814 &0.749 &0.662 &0.782 &0.813 \\
Adaboost\cite{freund1997decision} & 0.773 & 0.825 &0.763 &0.772 & 0.694 &0.790\\
\bottomrule
\end{tabular}
\begin{tablenotes}
\scriptsize
\item[a] The Precision and F1-score are the macro averages.
\end{tablenotes}
\end{threeparttable}}
\end{table}
\begin{table}[!t]
\centering
\caption{Performance comparison of different methods on AstraZeneca dataset.}
\label{AZ_set}
\scalebox{1.}{
\begin{threeparttable}
\begin{tabular}{p{2cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}}
\toprule
\textbf{Metric} & \textit{AU ROC} & \textit{AU PRC} & \textit{ACC} & \textit{BACC} &\textit{Precision}\tnote{a} &\textit{F1-Score} \\
\midrule
\textbf{Ours} &\textbf{0.841} &\textbf{0.887} &0.824 &\textbf{0.858} & \textbf{0.874} & \textbf{0.868} \\
DeepDDS & 0.722 & 0.801 &0.654 &0.627 & 0.824 &0.742 \\
DeepSynergy & 0.681 & 0.726 &0.662 &0.673 & 0.741 &0.735 \\
MR-GNN & 0.713 & 0.768 &\textbf{0.836} &0.621 & 0.690 &0.704 \\
MatchMaker & 0.702 & 0.698 &0.735 &0.728 & 0.790 &0.745 \\
XGBoost & 0.542 & 0.589 &0.697 &0.596 & 0.623 &0.606 \\
Adaboost & 0.521 & 0.546 &0.655 &0.620 & 0.633 &0.594\\
\bottomrule
\end{tabular}
\begin{tablenotes}
\scriptsize
\item[a] The Precision and F1-score are the macro averages.
\end{tablenotes}
\end{threeparttable}}
\end{table}
\begin{table}[t]
\centering
\caption{Performance of different methods on unseen drug and cell line experiments}
\label{unseen}
\scalebox{1.}{
\begin{threeparttable}
\begin{tabular}{cccc|ccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{3}{c}{Independent Drugs(39)\tnote{a}} & \multicolumn{3}{c}{Independent Cell-Lines(10)\tnote{b}} \\
\cmidrule(lr){2-4}\cmidrule(lr){5-7}
& \textit{AU ROC} & \textit{AU PRC} & \textit{F1-Score} & \textit{AU ROC} & \textit{AU PRC} & \textit{F1-Score}\tnote{c} \\
\midrule
\textbf{Ours} &\textbf{0.834} &\textbf{0.823} &\textbf{0.854} &\textbf{0.948} & \textbf{0.918} & \textbf{0.963} \\
DeepDDS & 0.697 & 0.795 &0.644 &0.889 & 0.854 &0.826 \\
DeepSynergy & 0.653 & 0.713 &0.676 &0.858 & 0.824 &0.863 \\
MatchMaker & 0.673 & 0.689 &0.675 &0.863 & 0.894 &0.878 \\
XGBoost & 0.510 & 0.581 &0.654 &0.794 & 0.754 &0.802 \\
Adaboost & 0.508 & 0.516 &0.592 &0.746 & 0.703 &0.776\\
\bottomrule
\end{tabular}
\begin{tablenotes}
\scriptsize
\item[a] 39 drugs are not included in the training set
\item[b] 10 cell lines show only a few times (around 10 times) in the training set
\item[c] The F1-score is the macro average.
\end{tablenotes}
\end{threeparttable}
}
\end{table}
\begin{table}[t]
\centering
\caption{Ablation study results}
\label{abalation}
\begin{threeparttable}
\begin{tabular}{cccc|ccc|cccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{3}{c}{DrugComb} & \multicolumn{3}{c}{AstraZeneca} & \multirow{2}{*}{Embedding} & \multicolumn{3}{c}{DrugComb}\\
\cmidrule(lr){2-4}\cmidrule(lr){5-7}\cmidrule(lr){9-11}
& \textit{AU ROC} & \textit{AU PRC} & \textit{F1-Score} & \textit{AU ROC} & \textit{AU PRC} & \textit{F1-Score} & &\textit{AU ROC} & \textit{AU PRC} & \textit{F1-Score} \\
\midrule
\textbf{Ours} &\textbf{0.961} &\textbf{0.954} &\textbf{0.972}&\textbf{0.841} &\textbf{0.887} & \textbf{0.868} & -Drug &0.914 &0.920 &0.939\\
Ours-S\tnote{b} & 0.946 & 0.925 &0.955 &0.832 & 0.854 &0.887 & -Protein &0.928 &0.915 &0.932\\
Ours-P\tnote{c} & 0.955 & 0.937 &0.943 &0.820 & 0.844 &0.868 & -Disease &0.946 &0.937 &0.952\\
Ours-PS\tnote{d} & 0.943 & 0.936 &0.948 &0.811 & 0.851 &0.837 & -All &0.889 &0.882 &0.913\\
\bottomrule
\end{tabular}
\begin{tablenotes}
\scriptsize
\item[a] The F1-score is the macro average.
\item[b] Ours-S means removing self-training strategy from our original model.
\item[c] Ours-P means removing predictive module (DDI and DTI) from our original model.
\item[d] Ours-PS means removing both self-training strategy and predictive module from our original model.
\end{tablenotes}
\end{threeparttable}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[width=1\textwidth]{com.jpg}
\caption{
\textbf{Performance analysis on DrugComb and Astrazeneca.} \textbf{a.} A scatter plot of predicted scores of Our method and DeepDDS. We measured the confidence scores of our method and DeepDDS for subsampled testing data in DrugComb (including both synergy and antagonism data). We outperform DeepDDS for more data points are below the diagonal, and our overall predicted score is higher than DeepDDS \textbf{b.} A scatter plot of predicted scores of Our method and DeepSynergy. We measured the confidence scores of our method and DeepSynergy for subsampled testing data in DrugComb. Our model can obviously predict the drug combination effect more accurately than DeepSynergy. \textbf{c.} Violin plots of predicted scores for five models. We compare our method with DeepDDS, DeepSynergy, MR-GNN, and MatchMaker. From the figure, our model's median and quartiles are all higher than the other four methods, indicating that we are among the best.
\textbf{d.} A similar scatter plot of predicted scores of our method and DeepDDS displaying all data points on the independent test set AstraZeneca. We outperform DeepDDS for more data points are below the diagonal, and our overall predicted score is higher than DeepDDS \textbf{e.} A similar scatter plot of predicted scores of Our method and DeepSynergy on AstraZeneca. Our model can obviously predict the drug combination effect more accurately than DeepSynergy. \textbf{f.} Violin plots of our model, DeepDDS, DeepSynergy, MR-GNN, and MatchMaker on AstraZeneca.
}
\label{Fig.DCperformance}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=1\textwidth]{5.pdf}
\caption{
\textbf{Performance on independent drugs and cell lines.} \textbf{a.} We want to test if our model can accurately predict unseen drugs in the training set. Totally we curated a testing dataset consisting of 39 unseen drugs. Our model surpasses DeepDDS with a higher confidence score. Most data points lie below the diagonal and in the right half of the subgraph. \textbf{b.} A scatter plot of predicted scores of Our method and DeepSynergy. We show more significant results. \textbf{c.} We compare our prediction score against drug similarity. The x-axis expresses the similarity between the drug in test set as well as its most similar drug in the training set, and the y-axis shows the corresponding predicted score of its synergistic effect. We apply Tanimoto Similarity to compute the similarity between the drugs. Although the performance fluctuates slightly, our model behaves robustly. \textbf{d.}
Here we choose 10 cell lines that hardly ever appear in the training set, and compare the performance of our method with DeepDDS. \textbf{e.} A scatter plot of predicted scores of our method compared with DeepSynergy. \textbf{f.} We compare our prediction score against cell line frequency. The x-axis refers to the frequency of cell lines in the training set, and the y-axis corresponds to the predicted score for the cell line. Our model shows stable performance across all the appearing frequencies of cell lines}
\label{Fig.celldrugspecific}
\end{figure}
Our model is a fully differentiable end-to-end method to perform Drug synergistic combinations prediction from given drug sequences and cell line representation alone. The overview is illustrated in Figure \ref{Fig.overview}. We performed several techniques such as large-scale pre-training to learn the structural information from a massive amount of unlabeled data, adaptive self-training to enrich our data space, and prediction module to enable inference via pseudo edges on independent nodes. In this section, we present a comprehensive evaluation study and compare our model with other existing methods across multiple metrics: area under the receiver operator characteristics curve \textit{AU ROC}, area under the precision-recall curve \textit{AU PRC}, accuracy \textit{ACC}, balanced accuracy \textit{BACC}, positive predictive value \textit{precision} and the harmonic mean of the precision and recall \textit{F1-Score}.
\paragraph{Cross-validation on DrugComb dataset}
The first step towards analyzing our performance is to compare with state-of-the-art methods on a large benchmark dataset: DrugComb \cite{zagidullin2019drugcomb}. Seven methods are selected including five deep learning methods: DeepDDS, TranSynergy, DeepSynergy, MR-GNN, MatchMaker, and two classical machine learning(ml) methods: XGBoost, and Adaboost. We utilized ChemicalX \cite{rozemberczki2022chemicalx} reimplementation of these methods except for the official version of DeepDDS and MatchMaker. Their reimplementations are different from the original paper, thus we compared their original and reimplemented versions and reported the best performing results. While MR-GNN is initially created to predict Drug-Drug Interactions (DDI), we can slightly adapt it to meet our scenarios. For Adaboost or XGBoost, we employed molecular fingerprints as input features of drugs.
The dataset is divided into ten splits of equal size randomly, and we perform cross-validation by iteratively masking out one split for testing and the remaining for training. The value for each metric takes the average numbers across every fold. Detailed results are presented in Table \ref{test_set} and the top value in each column is highlighted. We achieved the best result across all metrics except for ACC and precision. Specifically, for the most important measurements AU ROC and F1-Score, we surpass the second-best method DeepDDs by around 2\% and classical ml methods by almost 20\%. Deep learning methods perform better but it is notable that the overall performances for all methods are relatively high on DrugComb, thus the gap between ours and others is not so significant.
Since this is a binary classification task, we unified the last block of all methods to be a softmax layer. Intuitively, the higher the predicted score in the corresponding correct category, the better the model performs. Thus, we take a step further to examine this prediction score. Figure \ref{Fig.DCperformance} delivers the head-to-head comparison and violin plot of the prediction score. Figure \ref{Fig.DCperformance}.a and b compare our method with a competitive method: DeepDDS and an Astrazeneca-developed model: DeepSynergy. The x-axis refers to our method while the y-axis for others, each point represents a drug-pair-cell-line data and we subsample them down to 1100 data points. The majority of points are scattered below the diagonal and in the right part of the two subfigures suggesting we gained not only correct labels but also high prediction scores (confidence). Our results are more significant compared to DeepSynergy. Noted that points above the diagonal does not necessarily indicate wrong predictions, while points gathering in the lower left corner indicate hard samples where both methods failed. The boxplot around shows the average prediction scores for the correct category. The violin plot is prediction score distributions for selected four deep learning methods. We achieved the highest prediction score and DeepDDs/MR-GNN achieved similar second-best values. The margin is also not so large on DrugComb dataset.
\paragraph{Accurate predictions on domain-shift dataset}
The sound performance on DrugComb is somehow expected since it contains sufficient training samples and testing is done within the same data domain. Therefore, we tend to evaluate on a domain-shift dataset published by AstraZeneca to examine the performance of cross-domain inference. The data samples within AstraZeneca include plenty of unseen drugs and cell lines thus the drop of performance is no surprise. This time, all methods are then trained on the full DrugComb dataset and conduct inference on AstraZeneca.
Table \ref{AZ_set} summarized our detailed performance on the AstraZeneca dataset, the top result is highlighted in each column, and overall degraded performance is observed. Particularly, our \textit{AU ROC} fall from 0.961 to 0.852 and F1-Score from 0.972 to 0.863 compared with DrugComb test. This may be due to the data domain-shift but noticed that the second best DeepDDS or MR-GNN drop remarkably from 0.94 to 0.72 regarding their \textit{AU ROC}. The margin of our model with DeepDDs enlarges from 0.02 to 0.14, this is a substantial improvement since real application scenarios always exist data domain-shifts and new drugs. Thus, from the result, we can reduce more than 14 wrong cases out of 100 testing samples than DeepDDS, which can save great human effort and time costs in drug discovery. The results in terms of F1-Score behave similarly as our model performs better on imbalanced data.
We conduct a similar analysis of the prediction score as it is in cross-validation, Figure \ref{Fig.DCperformance} again delivers the head-to-head comparison and violin plot of prediction scores. Figure \ref{Fig.DCperformance}.d and e compares our method with two typical models in drug combination effect prediction methods: DeepDDs and DeepSynergy. The x-axis refers to our method while the y-axis for others. There are around 1200 datapoints in AstraZeneca, Majority of them are more scattered below the diagonal and in the right part of the two subfigures compared to DrugComb, suggesting that we're achieving more significant results on AstraZeneca dataset, which are not only correct labels but also high prediction scores (confidence). The box-plot around shows the average prediction score for the correct category. The violin plot in Figure \ref{Fig.DCperformance}.f shows prediction score distributions for four deep learning methods. We achieved the highest prediction score and DeepDDs/MR-GNN again achieved a similar second-best value.
\paragraph{Inferencing on independent drugs and cell lines}
The results from the AstraZeneca dataset indicate that domain-shift data is hard to infer. In fact, the AstraZeneca dataset still contains overlapping drugs or cell lines with DrugComb. To further test the generalization ability of our model, we tend to study our performance on non-overlapping drugs and cell lines. First, we create two datasets consisting of 39 independent drugs and 10 independent cell lines with 946 and 59 entries, respectively. Noted that these independent data are appointed from Astrazeneca on top of a cross-domain setting, and we delete the selected entries from our training set and the remaining data are used for training purposes. The independent drugs and cell lines are hence unrecognized to our model. Though these tasks are tough for the prior methods, according to our inference step, these unseen drugs can be linked in the graph via generated drug similarity edges or pseudo DTI or DDI edges. In this way, we enable information propagation even on unseen drugs and could provide richer representations than the initial embeddings. Notice that when a drug is dissimilar to all other drugs in datasets and has no predicted DTI/DDI interactions, our model would somehow degrade to an MLP-like classification pipeline, which we will investigate further.
Table \ref{unseen} summarized our performance in these two independent scenes. In the unseen drug study, our method still maintains a relatively high \textit{AU ROC} and \textit{AU PRC} over 80\% while the performance of other methods like DeepDDs and DeepSynergy drop under 70\% in some criteria. We achieved over 85\% regarding \textit{F1-Score} which is 20\% better than DeepDDS. From Figure \ref{Fig.celldrugspecific}.a and b, it can be clearly observed that most data points are below the diagonal and at the right side, such result indicates that our model is more robust against predicting unseen drugs than DeepDDS and DeepSynergy. Figure \ref{Fig.celldrugspecific}.c shows the trend between predicted score and drug similarity. Here the similarity implies the Tanimoto Similarity between unseen drugs and their most similar drugs in training set. Although the performance drops at low similarity region, our model is able to precisely predict the data above $0.5$, which in the sense conducting correct classification. Also, the regression trend shows our robust performance on the whole.
In the unseen cell line experiment, all the methods obtained a relatively high performance, as shown in Table \ref{unseen} since they all applied CCLE cell line expression as embedding vectors. However, we're different in the sense that we express the cell line as a weighted sum of rich protein embeddings while they treat protein as a one-hot vector. Thus, our model is superior to others for at least 5\% across all criteria. Figure \ref{Fig.celldrugspecific}.d and e present an intuitive view of unseen cell line study results of our model, DeepDDS, and DeepSynergy. Take HCC1569\_BREAST as an example, for most data points of this cell line, our model predicts them with very high confidence, on the contrary, DeepDDS and DeepSynergy can hardly make the right decision. The box plots around also show that our model has a significant lead on overall predicted scores. Besides, we calculated the trend between the predicted score and the occurrence frequency of the cell line in the training set, which is demonstrated in Figure \ref{Fig.celldrugspecific}.f. Our model behaves stably against the changes in cell line frequency. To some extent, the two data perturbation experiments verify that our model maintains high capability on challenging tasks.
\paragraph{Ablation study}
With the aforementioned experiments validating the strengths of our model in predicting drug combination effect, we first conduct an ablation study to evaluate the effectiveness of our submodules: self-training strategy and pre-trained DDI and DTI respectively. To investigate the self-training strategy, we remove it from our model and then regard this pipeline as Ours-S. Furthermore, to analyze the two predictive modules, we banned them from our framework and similarly named this pipeline as Ours-P. At last, we skip both self-training strategy and predictive modules to obtain the ‘Ours-PS’ pipeline. We conduct ablation studies on both DrugComb and AstraZeneca. Table \ref{abalation} shows all four models delivered high \textit{AU ROC}, \textit{AU PRC}, and \textit{F1-Score}, our original model has been consistently ranked as the best performance across all the measures. Results indicate that our self-training strategy and predictive modules nicely mine more useful information about drugs and proteins, and it helps our model to behave better in drug combination effect prediction.
Meanwhile, we found that even with our basic model ‘Ours-PS,’ we achieved noteworthy results on DrugComb, which could be attributed to our innovative use of rich embeddings over all instances. Therefore, we investigate how performance varies when moving to simpler representations. We used molecular fingerprints as an alternative for drug representations (‘-Drug’), one-hot encoding for either protein (‘-Protein’) or Disease alternative representations (‘-Disease’). ‘-All’ denotes the model using all the above mentioned simple alternative representations. The right part of Table \ref{abalation} concludes that instance embeddings have a huge and explicit impact on our performance. While drug embeddings are closely related to synergistic combinations, the term ‘-Drug’ yields the poorest results. One-hot encoding of Protein and Disease would introduce more noise and less information into our pipeline and thus lower our performance. When all rich representations are deleted, ‘-All’ definitely yields the worst results.
\paragraph{Hyperparameters setting} We listed the model parameters we used in our pipeline here to offer a better understanding. Every layer of our graph neural network is based on graph attention nets with input and output of both 512 dimensions and increased heads 4,8,12. The MLP prediction for synergistic classification has hidden layers of 3072, 768, and 128 dimensions. For DTI and DDI prediction modules, we use 1 attention block with 8 heads to encode either protein or drug representations and 2 attention blocks with 12 heads to process their concatenated embeddings. An MLP with 2048 and 256 dimensions of hidden layers is followed to predict the outcomes. We trained with a learning rate of $10^{-4}$ and a dropout rate of 0.2.
\section{Conclusion}
We develop an end-to-end model to facilitate the detection of drug combinations, aggregating various types of drug-related information. Comprehensive experiments including cross-fold validation, independent test, ablation study, and unseen experiment, demonstrate the effectiveness and robustness of our proposed method, where our model consistently and significantly outperforms all counterparts. Most known drug combination prediction methods utilize one or two kinds of features and their discovery ability are limited to only a few cell lines or tissues, they are not able to handle novel drugs, leading to poor performance in our unseen experiment.
Moreover, since large-scale pre-trained models show brilliant results in many fields, extending our model to do multi-tasks is a promising research direction. In the future, we will try to develop a method that is able to conduct multiple drugs, cell lines, and disease-related tasks, not limited to drug combination effect prediction.
We believe that our model can serve as a powerful tool to facilitate precise combination medicine and novel combination discovery. We will try to incorporate other kinds of information, such as 3D molecular structure into our framework to further improve our method’s performance.
\section{Acknowledgement}
The research reported in this publication was partially supported by Research Grants Council of the Hong Kong Special Administrative Region, China (Research Impact Fund (RIF), R5034-18, CUHK 2410021).
\bibliographystyle{unsrt}
|
\section{Introduction}
Solutions to integrable partial differential equations (PDEs) in
terms of multi-dimensional theta functions on compact Riemann
surfaces appeared in the 1970s in the search for quasi-periodic
solutions, see for instance \cite{Dub,BBEIM} for a historic account.
These solutions were constructed via the \emph{Baker-Akhiezer}
function, a function with an essential singularity on the Riemann
surface first introduced by Clebsch and Gordan. Mumford and coworkers
introduced in \cite{Mum} a complementary approach based on Fay's
celebrated \textit{trisecant identity} for theta functions
\cite{fay},
\begin{equation*}
\Theta^{*}_{ad}\Theta^{*}_{cb}\Theta_{ac}\Theta_{bd}
+\Theta^{*}_{ca}\Theta^{*}_{db}\Theta_{bc}\Theta_{ad}
= \Theta^{*}_{cd}\Theta^{*}_{ab}\Theta\Theta_{a+b,c+d},
\end{equation*}
where we have introduced the notation
\begin{equation}
\Theta^{*}_{ab}=\Theta^{*}\left(\int_{a}^{b}\right),\quad
\Theta_{ab}=\Theta\left(\mathrm{z}+\int_{a}^{b}\right)
\label{notation};
\end{equation}
here $\Theta(\mathrm{z})$, $\mathrm{z}\in \mathbb{C}^{g}$,
is the $g$-dimensional Riemann theta function, $\Theta^{*}(\mathrm{z})$ is
a theta function with an odd non-singular characteristic, see the
definitions (\ref{theta}), (\ref{thchar}),
$\Theta=\Theta(\mathrm{z})$, $\Theta^{*}=\Theta^{*}(0)=0$, and
$a,b,c,d$ are points on a
Riemann surface $\mathcal{R}$ with genus $g$. The Abel map
$\int_{a}^{b}$ between two points $a$ and $b$ on $\mathcal{R}$ is
defined at the beginning of section 2. Note that the name trisecant
identity refers to secants on the so-called Kummer variety, see
\cite{Tai} for a comprehensive review.
Since Fay's identity (\ref{Fay}) holds for arbitrary points $a$, $b$,
$c$, $d$ on the Riemann surface $\mathcal{R}$, it is possible to
consider the identity in the limit that two or more points
coincide\footnote{Note that there are generalizations of Fay's
identity to more than 4 points and degenerations thereof, see for
instance \cite{Fay2,Dub2,Ber} and references therein.}.
This leads to identities between derivatives of theta functions making
it possible to identify solutions to certain PDEs from degenerated
identies. In \cite{Mum} this was done for the Sine-Gordon equation
and the Kadomtsev-Petviasvili (KP) equation. On special Riemann surfaces
(hyperelliptic, trigonal) the latter solutions lead to
algebro-geometric solutions for the Korteweg-de Vries (KdV) \cite{Mum} and
the Boussinesq equation \cite{BBEIM}. In \cite{KKS} previously known solutions to the
Ernst equation \cite{Kor} were reconstructed via Fay's identity, see
also \cite{ernstbook}, in \cite{KKch}
known
solutions to the Camassa-Holm equation \cite{GH} were obtained with Mumford's
approach. In \cite{KalIMRN} Kalla
presented a new degenerated identity allowing to identify known
solutions to the nonlinear Schr\"odinger \cite{Its,Pre} and Davey-Stewartson
equations \cite{Mal} and to construct solutions to vector nonlinear
Schr\"odinger equations in terms of theta functions. For a recent
review on completely integrable dispersive PDEs, we refer to
\cite{book}. In this paper we generalize Kalla's approach
to higher order in the local parameter near the point $a$. We obtain
with the above notation\\
\textbf{Main theorem Part I}\\
Let $a$, $b$ be points on a compact Riemann surface $\mathcal{R}$,
and let $U:=D_{b}\ln \Theta
\Theta^{*}_{ba}$. Then $U$ satisfies
\begin{equation}
\begin{split}
0&=2(D_{a}U)^2 D_{a}''D_{a}U-2D_{a}U D_{a}''UD_{a}^{2}U
-(D_{a}U)^2 D_{a}^{4}U+4D_{a}UD_{a}^{3}U D_{a}^{2}U\\
&-3(D_{a}^{2}U)^{3}+3(D_{a}'U)^{2}D_{a}^{2}U
-3(D_{a}U)^2(D_{a}')^{2}DU.
\end{split}
\label{pillet3}
\end{equation}
This identity has similarities to the classical identity (\ref{Fay3})
by Fay in the sense that it involves the derivatives $D_{a}''$,
$D_{a}'$ and $D_{a}$ of $\Theta(\mathrm{z})$, but appears to be new.
In contrast to the potential in (\ref{Fay3}), the function $U$ also
depends on a point $b$ on the Riemann surface $\mathcal{R}$ which is
distinct from $a$, but otherwise
arbitrary.
We also prove \\
\textbf{Main theorem Part II}\\
The function \( \phi(x,y,t) := \mathrm{D}_b \ln{\Theta^*_{ab} \Theta
(x\mathbf{v}_{0}(a)+y \mathbf{v}_{1}(a)+ t
\mathbf{v}_{2}(a)+\mathbf{d} }) \), $(x,y,t)\in \mathbb{R}^{3}$, solves the Schwarzian KP equation:
\begin{equation}
\Big( \frac{\phi_t}{\phi_x}-\frac{1}{2} \{ \phi ; x \} \Big)_x - \frac{3}{2} \Big( \frac{\phi_y}{\phi_x} \Big)_y-\frac{3}{4} \Big( \frac{\phi_y^2}{\phi_x^2} \Big)_x=0
\label{SKP}
\end{equation}
where \( \{ \phi ; x \} \) denotes the Schwarzian derivative along \(
x \): \( \{ \phi ; x \} := \frac{\phi_{xxx}}{\phi_x}-\frac{3}{2}
\Big( \frac{\phi_{xx}}{\phi_x}\Big)^2 \), where
the indices denote partial derivatives with respect to the respective
variable, and where $\mathbf{v}_{j}$, $j=0,1,2$ has the components
$v_{ij}$, $i=1,\ldots,g$ defined in (\ref{abelexp}).
The solution $\phi$ in terms of multi-dimensional theta functions for
the Schwarzian KP equation seems to be new.
The Schwarzian KP equation (\ref{SKP}) appeared first in the
Painlev\'e analysis of the KP equation in \cite{WeissI} as a singularity manifold equation. Its
integrability was established in \cite{BK}. As in the case of the KP
equation, a reduction to a Schwarzian KdV and Boussinesq equation is
possible.
The paper is organised as follows: in section 2 we collect some basic
definitions of quantities defined on a compact Riemann surface and
known facts on Fay's identities. In section 3 we rederive identities
(\ref{Fay2}) and (\ref{kalla1}) from identity (\ref{Fay1}) and prove
the first part of the main theorem. In section 4 this is applied to
integrable PDEs. We add some concluding remarks in section 5.
\section{Preliminaries}
In this section, we will collect some basic definitions and known
facts on Fay's identities and applications.
\subsection{Basic definitions}
In this paper we always consider a Riemann surface $\mathcal{R}$ of
genus $g\in \mathbb{N}$ equipped with a canonical basis of cycles
$a_{1},\ldots,a_{g}, b_{1},\ldots,b_{g}$
satisfying the intersection conditions
$$ a_{i}\circ b_{j}=\delta_{ij},\quad a_{i}\circ a_{j}=0,\quad b_{i}\circ
b_{j}=0, \quad i,j=1,\ldots,g.
$$ The $g$-dimensional vector of holomorphic
1-forms is denoted by $\mathrm{d}\omega$ and normalized by
$\int_{a_{i}}^{}\mathrm{d}\omega_{j}=\delta_{ij}$, $i,j=1,\ldots,g$.
The matrix of $b$-periods
$\mathbb{B}_{ij}=\int_{b_{i}}^{}\mathrm{d}\omega_{j}$,
$i,j=1,\ldots,g$, is a Riemann matrix, i.e., it is symmetric and has
a positive definite imaginary part.
The Abel map $\omega:P\mapsto \int_{P_{0}}^{P}\mathrm{d}\omega$
is a bijective map from the Riemann surface $\mathcal{R}$ into the \emph{Jacobian}
$Jac(\mathcal{R}):=\mathbb{C}^{g}/\Lambda$ where $\Lambda$ is the
lattice formed by the periods of the holomorphic 1-forms,
$$\Lambda=\left\{\mathrm{m}+\mathbb{B}\mathrm{n}:
m,n\in\mathbb{Z}^{g}\right\}.$$ The expansion of the Abel map at a point
$P\in\mathcal{R}$ near a point $a\in\mathcal{R}$ is
written in the form,
\begin{equation}
\omega_{i}(P)=\sum_{j=0}^{\infty}v_{ij}\frac{\tau^{j}}{j!},\quad i=1,\ldots,g
\label{abelexp},
\end{equation}
where $\tau$ is a local parameter in the vicinity of $a$
containing also $P$. We define the derivatives acting on a function
$f(z)$, $z\in \mathbb{C}^{g}$ as
\begin{equation}
\begin{split}
D_{a}&:= \sum_{i=1}^{g}v_{i0}\partial_{z_{i}},\quad
D_{a}':= \sum_{i=1}^{g}v_{i1}\partial_{z_{i}},\quad
D_{a}'':= \sum_{i=1}^{g}\frac{v_{i2}}{2}\partial_{z_{i}},\\
D_{a}^{(n)}&:=
\sum_{i=1}^{g}\frac{v_{in}}{n!}\partial_{z_{i}},\quad n \in
\mathbb{N}.
\end{split}
\label{Ddef}
\end{equation}
Multi-dimensional theta functions are the building blocks of meromorphic functions
on Riemann surfaces. The theta
function with characteristic $\left
[\mathrm{p},\mathrm{q}\right]$ is defined
as an infinite series,
\begin{equation}\label{theta}
\Theta_{\mathrm{p}\mathrm{q}}(\mathrm{z},\mathbb{B})=
\sum\limits_{\mathrm{N}\in\mathbb{Z}^g}\exp\left\{
\mathrm{i}\pi\left\langle\mathbb{B}\left(\mathrm{N}+\mathrm{p}\right),
\mathrm{N}+\mathrm{p}
\right\rangle+2\pi \mathrm{i}
\left\langle \mathrm{z}+\mathrm{q},\mathrm{N}+\mathrm{p}
\right\rangle\right\}
\;,
\end{equation}
with $\mathrm{z}\in\mathbb{C}^g$ and $\mathrm{p}$, $\mathrm{q}\in{
\mathbb{R}}^g$, where
$\left
\langle\cdot,\cdot\right\rangle$ denotes the Euclidean scalar product
$\left\langle \mathrm{N},\mathrm{z}\right\rangle=\sum_{i=1}^gN_iz_i$.
The properties of the Riemann matrix ensure that the series converges
absolutely and that the theta function is an entire function on
$\mathbb{C}^{g}$. A characteristic is called \emph{singular} if the
corresponding theta function vanishes identically.
Half-integer
characteristics with $2\mathrm{p},2\mathrm{q}\in \mathbb{Z}^{g}$ are
called \emph{even} if $4\langle
\mathrm{p},\mathrm{q}\rangle=0\mbox{ mod } 2$ and \emph{odd}
otherwise.
Theta functions with odd (even) characteristic are odd
(even) functions of the argument $\mathrm{z}$. The theta function with
characteristic is related to the Riemann theta function $\Theta$, the
theta function with zero characteristic $\Theta:= \Theta_{\mathrm{00}}$,
via
\begin{equation}
\Theta_{\mathrm{pq}}(\mathrm{z},\mathbb{B})=\Theta(\mathrm{z}
+\mathbb{B}\mathrm{p} + \mathrm{q})\exp\left\{\mathrm{i}\pi
\left\langle\mathbb{B}\mathrm{p},\mathrm{p}\right\rangle+
2\pi \mathrm{i}\left\langle\mathrm{p},\mathrm{z} + \mathrm{q}\right\rangle
\right\}\;.
\label{thchar}
\end{equation}
A theta function with a nonsingular half-integer
characteristic is denoted by $\Theta^{*}$.
\subsection{Fay's identities}
Theta functions on Jacobians satisfy Fay's celebrated trisecant identity
\cite{fay}. It can be seen as a generalization of the classical
relation between cross ration functions
for four arbitrary points $a$, $b$, $c$, $d$ in the euclidean
plane
\begin{equation}
\lambda_{abcd}=\frac{\Theta^{*}(\smallint_{a}^{b})\Theta^{*}(\smallint_{c}^{d})}{
\Theta^{*}(\smallint_{a}^{d})\Theta^{*}(\smallint_{c}^{b})}
\label{eq:cross3}\;,
\end{equation}
which is a function on $\mathcal{R}$ that vanishes for $a=b$
and $c=d$ and has poles for $a=d$ and $b=c$.
\begin{theo}[Fay \cite{fay}]
Let $a$, $b$, $c$, $d$ be four points on the Riemann
surface $\mathbb{R}$. Then with the above
definitions
the following identity holds
\begin{equation}
\begin{aligned}
\lambda_{cabd}\,\Theta(\mathrm{z}+\smallint_{b}^{c})\,
&\Theta(\mathrm{z}+\smallint_{a}^{d}) +\lambda_{cbad}\,
\Theta(\mathrm{z}+ \smallint_{a}^{c})\,\Theta(\mathrm{z}+
\smallint_{b}^{d})\\
& =\Theta(\mathrm{z})\;\Theta(\mathrm{z}+\smallint_{b}^{c}+
\smallint_{a}^{d})\;,
\end{aligned}
\label{Fay}
\end{equation}
$\forall\mathrm{z}\in \mathbb{C}^{g}$. The
integration paths in (\ref{Fay}) have to be chosen in a way not to
intersect the canonical cycles.
\end{theo}
Degenerated versions of Fay's
identity lead to identities for derivatives of
theta functions. In the limit $d\to b$, one finds for (\ref{Fay})
\begin{cor}[Fay \cite{fay}]
Let $a$, $b$, $c$ be points on the Riemann surface
$\mathcal{R}$. Then the following identity holds,
\begin{equation}
D_{b}\ln \frac{\Theta(\mathrm{z}+
\smallint_{a}^{c})}{\Theta(\mathrm{z})}=p_{1}(a,b,c)+p_{2}(a,b,c)
\frac{\Theta(\mathrm{z}+
\smallint_{b}^{c})\Theta(\mathrm{z}+
\smallint_{a}^{b})}{\Theta(\mathrm{z}+
\smallint_{a}^{c})\Theta(\mathrm{z})}
\label{Fay1},
\end{equation}
where
\begin{align}
p_{1}(a,b,c) &= D_{b}\ln \frac{\Theta^{*}(
\smallint_{a}^{b})}{\Theta^{*}(\smallint_{c}^{b})},
\nonumber\\
p_{2}(a,b,c) & =\frac{\Theta^{*}(
\smallint_{a}^{c})D_{b}\Theta^{*}}{\Theta^{*}(
\smallint_{b}^{c})\Theta^{*}(
\smallint_{b}^{a})}
\label{pi}.
\end{align}
\end{cor}
In the limit $c\to a$, equation (\ref{Fay1}) yields
\begin{cor}[Fay \cite{fay}]
Let $a$, $b$ be points on the Riemann surface
$\mathcal{R}$. Then the following identity holds,
\begin{equation}
D_{a}D_{b}\ln \Theta(\mathrm{z})=q_{1}(a,b)+q_{2}(a,b)
\frac{\Theta(\mathrm{z}+
\smallint_{b}^{a})\Theta(\mathrm{z}+
\smallint_{a}^{b})}{\Theta(\mathrm{z})^{2}}
\label{Fay2},
\end{equation}
where
\begin{align}
q_{1}(a,b) &= D_{a}D_{b}\ln \Theta^{*}(
\smallint_{a}^{b}),
\nonumber\\
q_{2}(a,b) & =\frac{D_{a}\Theta^{*}D_{b}\Theta^{*}}{\Theta^{*}(
\smallint_{a}^{b})^{2}}
\label{qi}.
\end{align}
\end{cor}
In the limit $b\to a$, equation (\ref{Fay2}) can be cast into
the form (we suppress the index $a$ at the derivatives)
\begin{cor}[Fay \cite{fay}]
The following identity holds on the Riemann surface $\mathcal{R}$,
\begin{equation}
D^{4}\ln\Theta(\mathrm{z})+6(D^{2}\ln
\Theta(\mathrm{z}))^{2}+3D'D'\ln \Theta(\mathrm{z})-2DD''\ln
\Theta(\mathrm{z})+c_{1}D^{2}\ln \Theta(\mathrm{z}) +c_{2}=0
\label{Fay3},
\end{equation}
where
\begin{equation}
c_{1}=2\frac{D''\Theta^{*}}{\Theta^{*}}-4\frac{D^{3}\Theta^{*}}{D\Theta^{*}}
-3\left(\frac{D'\Theta^{*}}{D\Theta^{*}}\right)^{2};
\label{c1}
\end{equation}
the constant $c_{2}$
can be obtained by expanding $q_{1}$ and $q_{2}$
(\ref{qi}) in the considered limit to fourth order in the local
parameter near $a$.
\end{cor}
Kalla generalized the identity (\ref{Fay2}) to
\begin{theo}[Kalla \cite{KalIMRN}]\label{Ktheorem}
Let $a$, $b$ be points on $\mathcal{R}$, then the following
identity holds,
\begin{equation}
\begin{split}
0=&D_{a}'\ln
\frac{\Theta(\mathrm{z}+\smallint_{a}^{b})}{\Theta(\mathrm{z})}
+D_{a}^{2}\ln
\frac{\Theta(\mathrm{z}+\smallint_{a}^{b})}{\Theta(\mathrm{z})}
+\left(D_{a}\ln
\frac{\Theta(\mathrm{z}+\smallint_{a}^{b})}{\Theta(\mathrm{z})}-K_{1}(a,b)\right)^{2}\\
&+2D_{a}^{2}\ln \Theta(\mathrm{z})+K_{2}(a,b)
\end{split}
\label{kalla1}
\end{equation}
where $K_{1}(a,b)$, $K_{2}(a,b)$ depend on the points $a$, $b$,
but not on $\mathrm{z}$.
\end{theo}
\section{Proof of Main theorem Part I}
In this section we will consider various identities following from
Fay's identity (\ref{Fay1}) in the limit $c\to a$. First we
identify the known relations (\ref{Fay2}) and (\ref{kalla1}), then we
prove the first part of the main theorem.
To this end we write (\ref{Fay1}) in the form
\begin{equation}
V^{2}D_{b}
\frac{V(c)}{V}=D_{b}\Theta^{*}\Theta^{*}_{ac}\Theta_{ab}\Theta_{bc},
\label{Fay1V}
\end{equation}
where we have put
\begin{equation}
V(c) = \Theta_{ac}\Theta^{*}_{cb},\quad V = V(a)=\Theta\Theta^{*}_{ab}.
\label{V}
\end{equation}
\subsection{Known identities}
For identity (\ref{Fay1V})) we consider a Taylor expansion in the limit $c\to a$ in the local
parameter $\tau$ in (\ref{abelexp}). In lowest
order we get (\ref{Fay2}) in the form
\begin{equation}
D_{a}D_{b}\ln V =
\frac{1}{V^{2}}D_{b}\Theta^{*}D_{a}\Theta^{*}\Theta_{ab}\Theta_{ba}
\label{Fay2V}.
\end{equation}
In order $\tau^{2}$, we get for (\ref{Fay1V})
\begin{equation*}
\frac{V^{2}}{2}D_{b}(D_{a}^{2}\ln V+D_{a}'\ln V+(D_{a}\ln V)^{2})=
D_{b}\Theta^{*}D_{a}\Theta^{*}\Theta_{ab}\Theta_{ba}\left(\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}
+D_{a}\ln \Theta_{ba}\right),
\end{equation*}
which can be written with (\ref{Fay2V}) in the form
\begin{equation}
\frac{1}{2}D_{b}(D_{a}^{2}\ln V+D_{a}'\ln V+(D_{a}\ln V)^{2})=
D_{a}D_{b}\ln V\left(\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}
+D_{a}\ln \Theta_{ba}\right)
\label{kalla2V}.
\end{equation}
Since this identity holds for all $\mathrm{z}\in\mathbb{C}$, it also holds
for $\mathrm{z}$ replaced by $\mathrm{z}+\int_{a}^{b}$. This means $\Theta\mapsto
\Theta_{ab}$, $\Theta_{ba}\mapsto \Theta$. As shown by Kalla
\cite{KalIMRN}, the difference between
identity (\ref{kalla2V}) and (\ref{kalla2V}) after this shift of $\mathrm{z}$
reads
\begin{equation}
\begin{split}
0=&\frac{1}{2}D_{b}\left(D_{a}'\ln \frac{\Theta_{ab}}{\Theta}
+\frac{D_{a}^{2}\Theta_{ab}}{\Theta_{ab}}-\frac{D_{a}^{2}\Theta}{\Theta}
+2D_{a}\ln
\Theta^{*}_{ba}D_{a}\ln \frac{\Theta_{ab}}{\Theta}\right)\\
&+D_{a}D_{b}\ln \Theta\Theta^{*}_{ab}D_{a}\ln
\Theta_{ba}
-D_{a}D_{b}\ln \Theta_{ab}\Theta^{*}_{ab}D_{a}\ln
\Theta -\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}D_{a}D_{b}\ln \frac{\Theta_{ab}}{\Theta}
\end{split}
\label{kalla3}.
\end{equation}
With (\ref{Fay2}), we get
\begin{equation}
\begin{split}
0=&\frac{1}{2}D_{b}\left(D_{a}'\ln \frac{\Theta_{ab}}{\Theta}
+D_{a}^{2}\ln\frac{\Theta_{ab}}{\Theta}+
\left(D_{a}\ln\frac{\Theta_{ab}}{\Theta}\right)^{2}
+2D_{a}^{2}\ln \Theta \Theta^{*}_{ba}\right)
\\
&+\left(D_{a}\ln
\Theta^{*}_{ba}-\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}\right)
D_{a}D_{b}\ln \frac{\Theta_{ab}}{\Theta}
\end{split}
\label{kalla4}.
\end{equation}
Introducing the derivative $\nabla_{b}:=
\sum_{i=1}^{g}v_{i0}(b)\partial_{\mathrm{z}_{i}}$ acting only on $\mathrm{z}$, we
can write (\ref{kalla4}) in the form
\begin{equation}
\begin{split}
0=&\nabla_{b}\left\{\frac{1}{2}\left(D_{a}'\ln \frac{\Theta_{ab}}{\Theta}
+D_{a}^{2}\ln\frac{\Theta_{ab}}{\Theta}\right)
+\left(D_{a}\ln
\Theta^{*}_{ba}-\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}\right)
D_{a}\ln \frac{\Theta_{ab}}{\Theta}\right.\\
&\left.+D_{a}^{2}\ln \Theta\Theta^{*}_{ab}
+\frac{1}{2}\left(D_{a}\ln
\frac{\Theta_{ab}}{\Theta}\right)^{2}
\right\}
\end{split}
\label{kalla5}.
\end{equation}
This implies that relation (\ref{kalla1}) holds with
\begin{equation}
\begin{split}
K(a,b)=&D_{a}'\ln \frac{\Theta_{ab}}{\Theta}
+D_{a}^{2}\ln\frac{\Theta_{ab}}{\Theta}
+2\left(D_{a}\ln
\Theta^{*}_{ba}-\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}\right)
D_{a}\ln \frac{\Theta_{ab}}{\Theta}\\
&+2D_{a}^{2}\ln \Theta\Theta^{*}_{ab}
+\left(D_{a}\ln
\frac{\Theta_{ab}}{\Theta}\right)^{2}
\end{split}
\label{kalla6},
\end{equation}
where $K(a,b)$ just depends on $a$, $b$, but not on $\mathrm{z}$. It
can be computed for instance by putting $\mathrm{z}=0$ on the right
hand side of (\ref{kalla6}). This reproduces the proof of Theorem
\ref{Ktheorem} from \cite{KalIMRN}.
\subsection{Third order in $\tau$}
In third order of the local parameter $\tau$ we get for (\ref{Fay1})
\begin{equation}
\begin{split}
0=&D_{b}\left(\frac{1}{6}D_{a}''\ln V+\frac{1}{2}D_{a}D_{a}'
\ln V + \frac{1}{2}D_{a}'\ln V D_{a}\ln
V+\frac{1}{6}D_{a}^{3}\ln V \right.\\
&\left.+\frac{1}{2}D_{a}^{2}\ln V
D_{a}\ln V+\frac{1}{6}(D_{a}\ln V)^{3}\right)\\
&-D_{a}D_{b}\ln V
\left(\frac{D_{a}''\Theta^{*}}{6D_{a}\Theta^{*}}
+\frac{D_{a}^{3}\Theta^{*}}{6D_{a}\Theta^{*}}
+\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}
D_{a}\ln \Theta_{ba}+\frac{1}{2}D_{a}'\ln
\Theta_{ba}+\frac{D_{a}^{2}\Theta_{ba}}{2\Theta_{ba}}\right),
\end{split}
\label{pillet1aV}
\end{equation}
where we have used (\ref{Fay2V}). With (\ref{kalla2V}), we can
replace $\Theta_{ba}$ in (\ref{pillet1aV}) to obtain a relation
only involving $V$. To this end we put
\begin{equation}
F:= \frac{1}{2}D_{b}\left(D_{a}'\ln
V+D_{a}^{2}\ln V+(D_{a}\ln
V)^{2}\right)
\label{F}
\end{equation}
as well as
\begin{equation}
C_{1}:=\frac{D_{a}'\Theta^{*}}{2D_{a}\Theta^{*}}
\label{C1}
\end{equation}
which leads to (\ref{kalla2V}) in the form
\begin{equation}
D_{a}\ln \Theta_{ba}=\frac{F}{D_{a}D_{b}\ln V}-C_{1}
\label{kalla2a}.
\end{equation}
In addition we put
\begin{equation}
\begin{split}
G:=&D_{b}\left(\frac{1}{6}D_{a}''\ln
V
+\frac{1}{2}D_{a}D_{a}'\ln V+\frac{1}{2}D_{a}'\ln VD_{a}\ln V
\right.\\
&\left. +\frac{1}{6}D_{a}^{3}\ln V
+\frac{1}{2}D_{a}^{2}\ln VD_{a}\ln V
+\frac{1}{6}(D_{a}\ln V)^{3}
\right)
\end{split}
\label{G}
\end{equation}
and
\begin{equation}
C_{2}=\frac{D_{a}''\Theta^{*}}{6D_{a}\Theta^{*}}
+\frac{D_{a}^{3}\Theta^{*}}{6D_{a}\Theta^{*}}.
\label{C2}
\end{equation}
with which (\ref{pillet1aV}) takes the form
\begin{equation}
G= D_{a}D_{b}\ln V\left(C_{2}+C_{1}D_{a}\ln \Theta_{ba}+\frac{1}{2}D_{a}'\ln
\Theta_{ba}+\frac{1}{2}D_{a}^{2}\ln
\Theta_{ba}+\frac{1}{2}(D_{a}\ln\Theta_{ba})^{2}\right)
\label{pillet1a2}.
\end{equation}
Eliminating $\Theta_{ba}$ with (\ref{kalla2a}) from (\ref{pillet1a2})
leads in a first step to
\begin{equation}
\frac{G}{D_{a}D_{b}\ln V}-\frac{1}{2}D_{a}\left(\frac{F}{D_{a}D_{b}\ln V}\right)
-\frac{F^{2}}{2(D_{a}D_{b}\ln V)^{2}}=C_{2}-\frac{C_{1}^{2}}{2}
+\frac{1}{2}D_{a}'\ln\Theta_{ba}
\label{pillet1a3}.
\end{equation}
Differentiating with respect to $D_{a}$ (note that the derivatives of
the odd theta functions in $c_{1}$, $c_{2}$ vanish), we obtain with (\ref{kalla2a})
\begin{equation}
D_{a}\left(\frac{G}{D_{a}D_{b}\ln V}-\frac{1}{2}D_{a}\left(\frac{F}{D_{a}D_{b}\ln V}\right)
-\frac{F^{2}}{2(D_{a}D_{b}\ln V)^{2}}\right)=\frac{1}{2}
D_{a}'\left(\frac{F}{D_{a}D_{b}\ln V}\right)
\label{pillet2}.
\end{equation}
We have with $U:=D_{b}\ln V$
\begin{equation}
\begin{split}
&\frac{G}{D_{a}D_{b}\ln V}-\frac{1}{2}D_{a}\left(\frac{F}{D_{a}D_{b}\ln V}\right)
-\frac{F^{2}}{2(D_{a}D_{b}\ln V)^{2}}\\
&=\frac{D_{a}''U}{6D_{a}U}+\frac{D_{a}'D_{a}U}{4D_{a}U}
+\frac{1}{2}D_{a}'\ln V\\
&-\frac{D_{a}^{3}U}{12D_{a}U}+\frac{(D_{a}^{2}U)^{2}}{8(D_{a}U)^{2}}-\frac{(D_{a}'U)^{2}}{8(D_{a}U)^{2}},
\end{split}
\label{pillet2a}.
\end{equation}
Thus we get for (\ref{pillet2})
\begin{equation}
\begin{split}
0&=\frac{D_{a}''D_{a}U}{6D_{a}U}-\frac{D_{a}''UD_{a}^{2}U}{(6D_{a}U)^{2}}
-\frac{D_{a}^{4}U}{12D_{a}U}+\frac{D_{a}^{3}UD_{a}^{2}U}{3(D_{a}U)^{2}}\\
&-\frac{(D_{a}^{2}U)^{3}}{4(D_{a}U)^{3}}+\frac{(D_{a}'U)^{2}D_{a}^{2}U}{4(D_{a}U)^{3}}
-\frac{(D_{a}')^{2}DU}{4D_{a}U}.
\end{split}
\label{pillet3U}
\end{equation}
identical to equation (\ref{pillet3}) which
concludes the proof. Note that there are terms in (\ref{pillet2})
without a derivative $D_{a}$, but remarkably these terms all cancel
leaving (\ref{pillet2a}) a relation for terms all involving this
derivative.
\section{Applications to integrable PDEs}
In this section we will apply relation (\ref{pillet3}) to integrable
equations, prove the second part of the main theorem and consider
various reductions on special Riemann surfaces as known for the KP case.
\subsection{Main theorem Part II}
To prove the second part of the main theorem, we
define the function \( \phi(x,y,t) := \mathrm{D}_b \ln{\Theta^*_{ab} \Theta
(x\mathbf{v}_{0}(a)+y \mathbf{v}_{1}(a)+ t
\mathbf{v}_{2}(a)+\mathbf{d} }) \)
and show that it solves the Schwarzian KP equation (\ref{SKP}).
With our previous notations $ \phi = D_{b}\ln V=U$, moreover we can identify: $D_{a}=\partial_{x}$, $D_{a}'=\partial_{y}$
and $ D_{a}''=\partial_{t}$. Inserting the function \( \phi \) into (\ref{SKP}) we get:
\begin{align*}
\frac{D_{a}''D_{a} U}{D_{a} U}-\frac{D_{a}^2 U}{(D_{a}
U)^2}D_{a}'' U-\frac{1}{2}\frac{D_{a}^4 U}{D_{a}
U}+\frac{2}{(D_{a} U)^2}D_{a}^2 U D_{a}^3 U \\
-\frac{3}{2(D_{a} U)^{3}}(D_{a} U)^{3}+\frac{3}{2}\frac{D_{a}^2
U}{(D_{a} U)^{3}}(D'_{a} U)^{2}-\frac{3}{2}\frac{(D'_{a})^2
U}{D_{a} U}=0
\end{align*}
which is equivalent to (\ref{pillet3}) thus proving this part of the
theorem.
\begin{remark}
It is well known, see for instance \cite{Mum,BBEIM}, that one way
to represent meromorphic functions on a Riemann surface is in
terms of second logarithmic derivatives of theta functions. The
function $D_{b}\ln \Theta^{*}_{ab}\Theta$ is a priori not
independent of the path between $a$ and $b$ in
$\int_{a}^{b}\mathrm{d}\omega$. Both points have to be in
the same fundamental polygon. A possibility to avoid this
condition would be to consider $U=\partial_{x}^{-1}u$, where $u:=D_{a}D_{b}\ln
V$ is path independent; the
anti-derivative is defined as
$\partial_{x}^{-1}=\int_{a}^{x}dx'$.
\end{remark}
The Schwarzian KP equation
can also be written in the form
\begin{equation}
\left(\frac{U_{t}}{U_{x}}-\frac{1}{2}\frac{U_{xxx}}{U_{x}}+\frac{3}{4}
\frac{U_{xx}^{2}}{U_{x}^{2}}\right)_{x}+\frac{3}{2}\frac{U_{xx}}{U_{x}^{3}}U_{y}^{2}
-\frac{3}{2}\frac{U_{yy}}{U_{x}}=0.
\label{KP2}
\end{equation}
Integrating with respect to $x$, we get
\begin{equation}
U_{t}-\frac{1}{2}U_{xxx}+\frac{3}{4}\frac{U_{xx}^{2}-U_{y}^{2}}{U_{x}} +\frac{3}{2}U_{x} \partial_{x}^{-1}\left(\frac{U_{xy}U_{y}}{U_{x}^{2}}
-\frac{U_{yy}}{U_{x}}\right)
\label{KP3},
\end{equation}
which can be written in the form
\begin{equation}
U_{t}-\frac{1}{2}U_{xxx}+\frac{3}{4}\frac{U_{xx}^{2}-U_{y}^{2}}{U_{x}} -\frac{3}{2}U_{x} W_{y}
\label{KP4},
\end{equation}
where
$W_{x}:=U_{y}/U_{x}$. This is equation (13) in \cite{BK} after the
change of time $t\mapsto 2t$.
\subsection{Reductions on special Riemann surfaces}
Let us restrict our attention to the special case of the Riemann
surface \( \mathcal{R} \) being hyperelliptic, i.e., given by the
zero locus of the polynomial \( P(\lambda, \mu) = \mu^2-\prod_{j=1}^N
(\lambda-\lambda_j) \) where \( \lambda_j \in \mathbb{C} \),
$j=1,\ldots,N$, and \( N=2g+1 \) or \( N=2g+2 \), and denote by \( \pi: \mathcal{R} \longrightarrow \mathbb{CP}^1 \) the projection onto the Riemann sphere. \\
If \( a \) is a branch point of \( \pi \) then the hyperelliptic
involution locally reads \( \sigma: k_a \mapsto -k_a \) and its
pullback on \( \omega_j \) is given by: \( \sigma^* \omega_j =
-\omega_j \), hence the Taylor expansion of \( \omega_j \) around \(a
\) must be even in \(k_a \), i.e \( \mathbf{v}_{1} =0 \) and thus \( \mathrm{D_a'}=0 \).
Eliminating $\Theta_{ba}$ from (\ref{pillet1a2}) via
(\ref{F}), we get
\begin{cor}
Identity (\ref{pillet3}) reduces on hyperelliptic surfaces with
$a$ being a branch point to
\begin{equation}
0=\frac{1}{6}D_{a}''D_{b}\ln V - \frac{1}{12}D_{a}^{3}D_{b}\ln
V- C_{2}D_{a}D_{b}\ln V +\frac{1}{8}\frac{(D_{a}^{2}D_{b}\ln
V)^{2}}{D_{a}D_{b}\ln V}
\label{red1}.
\end{equation}
\end{cor}
If we put again $U=D_{b}\ln V$, $D_{a}''=\partial_{t}$ and
$D_{a}=\partial_{x}$, we get for (\ref{red1})
\begin{equation}
\frac{1}{6}U_{t}-\frac{1}{12}U_{xxx}-C_{2}U_{x}+\frac{1}{8}
\frac{U_{xx}^{2}}{U_{x}}=0
\label{red2}.
\end{equation}
This implies
\begin{cor}
The function \( \phi(x,t) := \mathrm{D}_b \ln{\Theta^*_{ab} \Theta
(x\mathbf{v}_{0}(a)+ t \mathbf{v}_{2}(a)+\mathbf{d} }) \) solves the Schwarzian KdV equation:
\begin{equation}
\frac{1}{6}\frac{\phi_t}{\phi_x}-\frac{1}{12} \{ \phi ; x \} =
C_{2}
\label{SKdV}.
\end{equation}
\end{cor}
As in Chapter 3.4 of \cite{BBEIM} for KP, there is also a reduction
to a Schwarzian Boussinesq equation. If the surface $\mathcal{R}$ is
given by a trigonal curve, i.e., a curve on which a meromorphic
function with a third order pole at a point $a\in \mathcal{R}$ and no
other singularities exists. A simple example of such a curve is
\begin{equation}
\mu^{4}=\prod_{i=1}^{4}(\lambda-E_{i}).
\label{trig}
\end{equation}
In this case $D_{a}''=0$ which leads for (\ref{pillet3}) to
\begin{equation}
0=
-\frac{D_{a}^{4}U}{12D_{a}U}+\frac{D_{a}^{3}UD_{a}^{2}U}{3(D_{a}U)^{2}}\\
-\frac{(D_{a}^{2}U)^{3}}{4(D_{a}U)^{3}}+\frac{(D_{a}'U)^{2}D_{a}^{2}U}{4(D_{a}U)^{3}}
-\frac{(D_{a}')^{2}DU}{4D_{a}U}.
\label{pillet3b}
\end{equation}
The function
\( \phi(x,t) := \mathrm{D}_b \ln{\Theta^*_{ab} \Theta
(x\mathbf{v}_{0}(a)+t \mathbf{v}_{1}(a)+\mathbf{d} }) \) then gives a
solution to the Schwarzian Boussinesq equation \cite{WeissI}
\begin{equation}
\Big( -\frac{1}{2} \{ \phi ; x \} \Big)_x - \frac{3}{2} \Big(
\frac{\phi_t}{\phi_x} \Big)_t-\frac{3}{4} \Big( \frac{\phi_t^2}{\phi_x^2} \Big)_x=0
\label{SB}
\end{equation}
\section{Conclusion}
In this paper we have studied degenerations of Fay's identities in higher order of the
local parameter $\tau$ near one of the points. The starting point was Fay's
identity for 3 points on a Riemann surface in the form (\ref{Fay1V}).
The case of order $\tau^{3}$ was studied in detail as well as its
application to the Schwarzian KP equation. It appears straight
forward to generalize this approach to higher orders of the parameter
$\tau$. A standard Taylor expansion of the quantity $V(c)$ yields for
the left hand side of equation (\ref{Fay1V})
\begin{equation}
D_{b}\sum_{m=1}^{\infty}\frac{1}{m!V}\left(\tau
D_{a}+\frac{\tau^{2}}{2}D_{a}'+\frac{\tau^{3}}{6}D_{a}''+\ldots
\frac{\tau^{k}}{k!}D_{a}^{(k)}+\ldots\right)^{m}V
\label{Vex}.
\end{equation}
Thus one gets in order $\tau^{n}$
\begin{equation}
D_{b}\left(\frac{D_{a}^{n}V}{n!
V}+\frac{nD_{a}^{n-1}D_{a}'V}{2(n-1)!V}+\ldots+\frac{D_{a}^{(n)}V}{n!V}\right)
\label{Vn}.
\end{equation}
The same expansion can be obtained on the right hand side of
(\ref{Fay1V}) for $\Theta^{*}_{ac}$, where all even order derivatives
vanish for symmetry reasons, and for $\Theta_{bc}$ leading to
derivatives of $\Theta_{ba}$. As in section 3, the latter terms can be
replaced via (\ref{kalla2a}) by derivatives of $\Theta$. As in
(\ref{pillet2}), it will be necessary to differentiate with respect
to $D_{a}$ in general in order to eliminate all terms with
$\Theta_{ba}$. It is beyond the scope of the current paper to detail
the resulting relations and to establish a potential relation to
integrable PDEs and whether these are from a hierarchy of Schwarzian
KP equations. An interesting question is also whether a similar
approach can be applied to the degeneration of identity (\ref{Fay2})
in the limit $b\to a$
in higher orders of the local parameter near $a$, which would lead to
a generalisation of relation (\ref{Fay3}). This will be the subject
of future research.
Another interesting aspect would be to relate the present work to the
bilinear approach studied in \cite{BK2}. In this article the authors
describe the tau and Baker-Akhiezer functions associated to some generalized KP hierarchy (the Schwarzian KP hierarchy being one of them). Their derivation is based on a generalized Hirota's bilinear identity based on the equation:
\begin{equation*}
\int_{\partial G} \chi( \nu, \mu ; g_1) g_1(\nu) g_2^{-1}(\nu) \chi(\lambda, \nu ; g_2) d \nu =0
\end{equation*}
Where (following the notations of \cite{BK2}) \( G \) is the unit disk, \( (\lambda, \mu) \in \mathbb{C}^2 \) are spectral parameters, \( g_1(\nu) = g(\nu, \mathbf{x})=\exp{\big( \sum_{i=1}^{\infty} x_i \nu ^{-i} \big) } \), \( g_2(\nu)= g(\nu, \mathbf{x'}) \) and \( \chi(\lambda, \mu) \) is an unknown meromorphic function in both variables. \\
The associated tau function is given by:
\begin{equation*}
\chi(\lambda, \mu, \mathbf{x}) = \frac{1}{(\lambda-\mu)}\frac{\tau(\mathbf{x}-[\lambda]+[\mu])}{\tau(\mathbf{x})}
\end{equation*}
where \( \mathbf{x}+[\mu]=x_i+[\mu]_i \), with \( [\mu]_i=\frac{1}{i} \mu^i \), \( 0 \leq i < \infty \).
They proved that this tau function satisfies the following addition formula for \(a\), \(b\), \(c\) and \(d\) some arbitrary complex numbers :
\begin{equation*}
\begin{split}
& (a-c)(d-b) \tau(\mathbf{x}+[a]+[c])\tau(\mathbf{x}+[d]+[b])+ \\
& +(d-a)(b-c)\tau(\mathbf{x}+[d]+[a])\tau(\mathbf{x}+[b]+[c])+\\
& +(b-a)(d-c)\tau(\mathbf{x}+[b]+[a])\tau(\mathbf{x}+[d]+[c])=0
\end{split}
\end{equation*}
As established in the seminal paper \cite{Shio}, this addition formula is nothing more than Fay's trisecant identity when a tau function can be written in terms of Riemann theta functions. All these considerations lead to the following conjecture:
\begin{conj}
There exists a quadratic form \( \mathbf{Q}(\mathbf{t}):= \sum_{i,j=1}^4 Q_{ij}t_i t_j \), \( Q_{ij} \in \mathbb{C} \), such that the function
\begin{equation*}
\tau(\mathbf{t}):=\exp{ (\mathbf{Q}(\mathbf{t})) }\Theta(t_1\mathbf{v}_{0}(a)+t_2 \mathbf{v}_{1}(a)+ t_3
\mathbf{v}_{2}(a) + t_4 \mathbf{v_0}(b)+ \mathbf{d})
\end{equation*}
is a tau function for the Schwarzian KP equation \ref{SKP} in the sense of \cite{BK2}, the variables \( t_1 \), \( t_2 \), \( t_3 \) being respectively identified with \( x \), \(y \), \(t \) and \(t_4 \) is an auxiliary parameter.
\end{conj}
This conjecture relies on the fact that the Schwarzian KP hierarchy
is generated in a very similar fashion as the classical KP hierarchy \cite{BK2}. Hence a tau function for the full Schwarzian KP hierarchy could be of the form \( \Tilde{\tau} := \exp{ (\mathbf{Q}(\mathbf{t})) } \Theta( \mathbb{V} \mathbf{t} + \mathbf{d}) \), with \( \mathbb{V} \) a \( g \times \infty \) matrix whose columns are the vectors associated to the different "time" variables, and the higher order degenerations discussed above could have an explicit and compact form when expressed with Hirota's symbols.
|
\section{Introduction}
The Bochner technique is a highly acclaimed method of proof in classical differential geometry, which is attributed to \textsc{Bochner} \cite{Boch} and sometimes also \textsc{Yano} \cite{YB}. It discusses how Ricci curvature affects the types of vector fields that a manifold admits. A modern introduction to the topic can be found in \cite[Chapter 7]{Pete}, which also served as inspiration for the generalization in this article.
Let $ (M,g) $ be a connected closed oriented Riemannian manifold and let $ \Ric $ denote the Ricci curvature tensor of $ (M,g) $. The Bochner technique culminates in the following two theorems \cite[Theorems 36 \& 48]{Pete}:
\begin{theorem} \label{negBoch}
If $ \Ric $ is negative semi-definite everywhere, then every Killing vector field is parallel. If additionally $ \Ric $ is negative definite at one point, then no non-trivial Killing vector fields exist.
\end{theorem}
\begin{theorem} \label{posBoch}
If $ \Ric $ is positive semi-definite everywhere, then every harmonic vector field (i.e.~the $ g $-dual one-form is harmonic) is parallel. If additionally $ \Ric $ is positive definite at one point, then no non-trivial harmonic vector fields exist.
\end{theorem}
By combining these results with the famous Hodge theorem, which states that the space of harmonic one-forms is isomorphic to the first cohomology group, we obtain the following interesting consequences:
\begin{corollary} \label{van}
If $ \Ric $ is positive semi-definite everywhere and positive definite at one point, then the first Betti number $ b_1(M) = 0 $.
\end{corollary}
\begin{corollary} \label{flat}
If $ \Ric $ vanishes everywhere, then the dimension of the isometry group of $ (M,g) $ is equal to $ b_1(M) $.
\end{corollary}
Over the years \Cref{negBoch,posBoch} have been adapted to work with various additional structures on $ M $, in particular \emph{Riemannian foliations}. We will go through all of the necessary preliminaries about Riemannian foliations thoroughly in \Cref{Riem}. \textsc{Kamber} and \textsc{Tondeur} devised the following analogue of \autoref{negBoch} \cite[Theorem B]{KT2}:
\begin{theorem} \label{tnegBoch} Let $ M $ be a connected closed oriented manifold, endowed with a Riemannian foliation $ (\mc{F},g) $ and let $ \Ric^T $ denote the transverse Ricci curvature of $ (\mc{F},g) $. If $ \Ric^T $ is negative semi-definite everywhere, then every transverse Killing vector field is transverse parallel. If additionally $ \Ric^T $ is negative definite at one point, then no non-trivial transverse Killing vector fields exist.
\end{theorem}
In fact, \autoref{tnegBoch} is even a generalization of \autoref{negBoch}, where the latter corresponds to the special case that $ \mc{F} $ is the trivial foliation of $ M $ by singletons. In the direction of \autoref{van}, there is the following vanishing theorem for basic cohomology, which was discovered independently by \textsc{Min-Oo}, \textsc{Ruh} and \textsc{Tondeur} \cite[Theorem 8.16]{Tond} as well as \textsc{Hebda} \cite[Theorem 1]{Hebd}:
\begin{theorem} \label{bvan}
If $ (M,\mc{F},g) $ are as in \autoref{tnegBoch} and $ \Ric^T $ is positive definite everywhere, then the first basic Betti number $ b_1(\mc{F}) = 0 $.
\end{theorem}
The main goal of this article is to close the remaining gap by developing a generalization of \autoref{posBoch}:
\begin{theorem} \label{introBoch}
Let $ M $ be a connected closed oriented manifold, endowed with a transversely oriented harmonic Riemannian foliation $ (\mc{F}, g) $. If $ \Ric^T $ is positive semi-definite everywhere, then every basic harmonic vector field is transverse parallel. If additionally $ \Ric^T $ is positive definite at one point, then no non-trivial basic harmonic vector fields exist.
\end{theorem}
By combining \Cref{tnegBoch,introBoch} with a basic Hodge theorem, we obtain the following variations of \autoref{flat} and \autoref{bvan}:
\begin{corollary}
If $(M, \mc{F},g) $ are as in \autoref{introBoch} and $ \Ric^T $ is positive semi-definite everywhere, then $ b_1(\mc{F}) \leq \codim \mc{F} $. If additionally $ \Ric^T $ is positive definite at one point, then $ b_1(\mc{F}) = 0 $.
\end{corollary}
\begin{corollary} \label{tflat}
If $(M, \mc{F},g) $ are as in \autoref{introBoch} and $ \Ric^T $ vanishes everywhere, then the dimension of the vector space $ \mf{iso}(\mc{F}) $ of transverse Killing vector fields of $ (\mc{F},g) $ is equal to $ b_1(\mc{F}) $.
\end{corollary}
Finally, we apply \autoref{tflat} to two classes of spaces which naturally naturally satisfy all of the required conditions, namely \emph{degenerate 3-$(\alpha,\delta) $-Sasaki} and certain \emph{Sasaki-$ \eta $-Einstein manifolds}:
\begin{theorem} \label{3adS}
\vspace{-.2cm} Let $ M $ be a connected closed degenerate 3-$(\alpha, \delta) $-Sasaki manifold with characteristic foliation $ \mc{F} $. Then the dimension of the automorphism group $ \Aut(M) $ is at most $ b_1(\mc{F}) + 3 $. \\
In particular, if $ M $ arises via \cite[Theorem 3.1]{GRS} as a $ T^3 $-bundle over a compact hyperkähler manifold $ N $ with integral Kähler classes, then $ \dim \Aut(M) \leq b_1(N) + 3 $.
\end{theorem}
\begin{theorem} \label{SeE}
Let $ M $ be a connected closed Sasaki-$ \eta $-Einstein manifold with transverse Calabi-Yau structure and characteristic foliation $ \mc{F} $. Then the dimension of the automorphism group $ \Aut(M) $ is at most $ b_1(\mc{F}) + 1 $. \\
In particular, if $ M $ arises as the Boothby-Wang bundle over a compact Calabi-Yau mani- fold $ N $ with integral Kähler class, then $ \dim \Aut(M) \leq b_1(N) + 1 $.
\end{theorem}
About the structure of this article: We start with a self-contained explanation of the required fundamentals about Riemannian foliations (\Cref{Riem}) and basic Hodge theory (\Cref{sHodge}). In \Cref{sBochner} we complete the proof of the main \autoref{introBoch} as well as its consequences and in \Cref{Sas} we provide the promised applications. \\
\textbf{Acknowledgements:} The author was partially supported by the German Academic Scholarship Foundation. The author thanks Oliver Goertsches for various fruitful discussion about the subject and Leander Stecker for the suggestion to apply the technique to Sasaki-$ \eta $-Einstein manifolds.
\section{Riemannian Foliations} \label{Riem}
Let $ M $ be a smooth manifold and $ (\mc{F},g_T) $ a \emph{Riemannian foliation} on $ M $. This means that $ \mc{F} $ is a foliation on $ M $ defined by an integrable subbundle $ E \subset TM $ and $ g_T $ is a \emph{transverse metric}, i.e.~a symmetric positive semi-definite $ (0,2) $-tensor field such that $ \ker g_T = E $ and $ \mc{L}_X g_T = 0 $ for all $ X \in \Gamma_\ell(E) $, where $ \Gamma_\ell $ denotes the set of all local sections of a fiber bundle. In order to avoid having to deal with quotient bundles, we shall choose and fix a so-called \emph{bundle-like metric} $ g $ on $ M $, i.e.~a Riemannian metric such that $ g(X^\perp, Y^\perp) = g_T(X,Y) $ for all $ X,Y \in TM $, where $ Z^\perp $ denotes the $ g $-orthogonal projection of $ Z \in TM $ to $ E^\perp $.
\begin{definition} \label{vfs}
The Lie algebra of \emph{foliated vector fields} and the vector space of \emph{transverse vector fields} are given by $ \mf{fol}(\mc{F}) := N_{\Gamma(TM)}\big(\Gamma(E)\big) $, the normalizer of $ \Gamma(E) $ inside $ \Gamma(TM) $, as well as $ \mf{trans}(\mc{F}) := \mf{fol}(\mc{F}) \cap \Gamma(E^\perp) $, respectively. We call a function $ f:M \to \R $ \emph{basic} if $ X(f) = 0 $ for all $ X \in \Gamma_\ell(E) $.
\end{definition}
\begin{lemma} \label{grad}
\begin{enumerate}
\item[a)] If $ X \in \mf{fol}(\mc{F}) $, then $ f := \frac{1}{2} \, g_T(X,X) $ is basic.
\item[b)] If $ f $ is basic, then its gradient (with respect to $ g $) satisfies $ \nabla f \in \mf{trans}(\mc{F}) $.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item[a)] For all $ Y \in \Gamma_\ell(E) $: $ Y(f) = g_T([Y,X],X) = 0 $.
\item[b)] First, $ 0 = X(f) = g(\nabla f, X) $ for all $ X \in \Gamma_\ell(E) $, so $ \nabla f \in \Gamma(E^\perp) $. Furthermore, for all $ X \in \Gamma(E) $, $ Y \in \Gamma_\ell(E^\perp) $:
\[ g_T([\nabla f,X],Y) = g_T(\nabla f,[X,Y]) - X\big(g_T(\nabla f,Y)\big) = [X,Y](f) - X\big(Y(f)\big) = 0 \, . \]
\end{enumerate}
\end{proof}
\begin{definition} \label{bott}
Let $ \nabla $ denote the Levi-Civita connection of $ g $. The \emph{transverse Levi-Civita} or \emph{Bott connection} $ \nabla^T $ is the connection in the vector bundle $ E^\perp $ given by
\[ \nabla^T_X Y := \begin{cases} \left(\nabla_X Y\right)^\perp & , \; X \in \Gamma_\ell(E^\perp) \, , \\ [X,Y]^\perp &, \; X \in \Gamma_\ell(E) \, . \end{cases} \]
The condition $ \mc{L}_X g_T = 0 $ ensures that $ [X,Y]^\perp $ is indeed tensorial in $ X \in \Gamma_\ell(E) $. Note that if $ Y \in \mf{trans}(\mc{F}) $, then $ \nabla^T_X Y = [X,Y]^\perp = 0 $ for all $ X \in \Gamma_\ell(E) $. The connection $ \nabla^T $ is the unique metric and torsion-free connection in $ E^\perp $ \cite[Theorem 5.9]{Tond}, i.e. for all $ X \in \Gamma_\ell(TM) $ and $ Y,Z \in \Gamma_\ell(E^\perp) $:
\begin{align*}
X\big(g_T(Y,Z)\big) &= g_T(\nabla^T_X Y,Z) + g_T(Y, \nabla^T_X Z) \, , \\ \nabla^T_Y Z - \nabla^T_Z Y &= [Y,Z]^\perp \, .
\end{align*}
Furthermore, $ \nabla^T $ may be characterized via a Koszul formula \cite[Proposition 1.7]{KT1}, i.e.~for all $ X, Z \in \Gamma_\ell(TM) $, $ Y \in \Gamma(E^\perp) $:
\begin{align*} \label{koszul}
2g_T(\nabla^T_X Y, Z) &= X\big(g_T(Y,Z)\big) + Y\big(g_T(Z,X)\big) - Z\big(g_T(X,Y)\big) \\
&\quad + g_T([X,Y],Z) + g_T([Z,X],Y) - g_T([Y,Z],X) \, .
\end{align*}
\end{definition}
\begin{definition}
Let $ f $ be a basic function. The \emph{transverse Hessian} $ \Hess_T f $ is the symmetric $ (0,2) $-tensor field given by
\[ \Hess_T f (X,Y) := g_T(\nabla^T_X \nabla f,Y) \, , \quad X,Y \in \Gamma_\ell(TM) \, . \]
Clearly, $ \iota_X \Hess_T f = 0 $ for all $ X \in \Gamma_\ell(E) $. The \emph{transverse Laplacian} $ \Delta_T f $ is defined as
\[ \Delta_T f := \tr_g \Hess_T f = \sum_i \Hess_T f(E_i,E_i) \, , \]
where $ E_i $ is a local $ g $-orthonormal frame. The \emph{transverse Riemann curvature tensor} $ R^T $ is given by
\[ R^T(X,Y)Z := \nabla^T_X \nabla^T_Y Z - \nabla^T_Y \nabla^T_X Z - \nabla^T_{[X,Y]} Z \, , \quad X,Y \in \Gamma_\ell(TM) , \, Z \in \Gamma_\ell(E^\perp) \, . \]
Again, $ \iota_X R^T = 0 $ for all $ X \in \Gamma_\ell(E) $ \cite[Proposition 3.6]{Tond}. As usual:
\[ R^T(X,Y,Z,V) := g_T\big(R^T(X,Y)Z,V\big) \, , \quad V \in TM \, . \]
Finally, the \emph{transverse Ricci curvature} $ \Ric^T $ is defined as
\[ \Ric^T (X,X) := \tr \big(Y \mapsto R^T(Y,X)X \big) = \sum_i R^T(E_i,X,X,E_i) \, , \quad X \in \Gamma_\ell(E^\perp) \, . \]
\end{definition}
\begin{remark} \label{curv}
It is well-known that a Riemannian foliation can be characterized equi- valently via local Riemannian submersions $ \phi: U \to N $ onto a Riemannian model space $ (N,g_N) $. The transverse Riemann curvature tensor $ R^T $ then reflects the Riemann curvature tensor $ R^N $ of the local model $ N $, as made precise by the following equation \cite[Equation 5.40]{Tond}:
\[ \phi_\ast R^T(X,Y)Z = R^N(\phi_\ast X, \phi_\ast Y)\phi_\ast Z \, , \quad X,Y,Z \in \Gamma_\ell (E^\perp) \, . \]
Likewise, $ \Ric^T $ mirrors the Ricci curvature tensor $ \Ric^N $ of $ N $, viz.
\[ \Ric^T(X,X) = \Ric^N(\phi_\ast X,\phi_\ast X) \circ \phi \, , \quad X \in \Gamma_\ell(E^\perp) \, . \]
Hence, if the Riemannian foliation $ (\mc{F},g) $ was simply given by one (global) Riemannian submersion $ \phi: M \to N $ onto a Riemannian manifold $ N $, then we could just apply the ordinary Bochner technique to $ N $ instead of the more complicated approach presented here. However, the advantage of a Bochner technique for foliations is that it also works if the model space (globally) is not as well-behaved as a smooth manifold, which is often a non-trivial condition.
\end{remark}
From now on let $ n:= \dim M $, $ p:= \rk E $ and $ q := n-p $.
\begin{definition}
We call $ \mc{F} $ \emph{transversely orientable} if $ E^\perp $ is orientable. Suppose that $ M $ is orientable and $ \mc{F} $ is tranversely orientable. Then we shall orient $ M $ and $ E^\perp $ using their \emph{Riemannian volume forms} $ \mu $ and $ \mu_T $, respectively. This means we choose a local oriented orthonormal frame $ E_1, \ldots, E_n $ of $ TM $ such that $ E_{p+1}, \ldots, E_n $ is an oriented frame of $ E^\perp $ and require $ \mu(E_1,\ldots,E_n) = \mu_T(E_{p+1},\ldots,E_n) = 1 $. \\
If $ X \in \mf{fol}(\mc{F}) $ is a foliate vector field, then $ \iota_Y \mc{L}_X \mu_T = 0 $ for all $ Y \in \Gamma_\ell(E) $. Thus, $ \mc{L}_X \mu_T $ may be viewed as a section of the vector bundle $ \Lambda^q (E^\perp)^* $, which has rank one. Hence, we can define the \emph{transverse divergence} $ \Div_T X $ as the unique function which satisfies
\[ \mc{L}_X \mu_T = \Div_T X \cdot \mu_T \, . \]
\end{definition}
\begin{lemma} \label{dlem}
For any transverse vector field $ X \in \mf{trans}(\mc{F}) $:
\[ \Div_T X = \tr \nabla^T X \, . \]
In particular, for any basic function $ f $:
\[ \Div_T \nabla f = \Delta_T f \, . \]
\end{lemma}
\begin{proof}
Since $ X $ is transverse, we have $ \nabla^T_Y X = 0 $ for all $ Y \in \Gamma_\ell(E) $. If $ E_{p+1}, \ldots, E_n $ is an oriented local orthonormal frame of $ E^\perp $, then:
\begin{align*}
(\mc{L}_X\mu_T)(E_{p+1},\ldots,E_n) &= X\big(\mu_T(E_{p+1},\ldots,E_n)\big) - \sum_i \mu_T(E_{p+1},\ldots,[X,E_i],\ldots,E_n) \\
&= - \sum_i g_T([X,E_i],E_i) \, \mu_T(E_{p+1},\ldots,E_n) = - \sum_i g_T([X,E_i],E_i) \, .
\end{align*}
On the other hand:
\[ \tr \nabla^T X = \sum_i g_T(\nabla_{E_i}^T X, E_i) = \sum_i g_T(\nabla_X^T E_i, E_i) - g_T([X,E_i],E_i) = - \sum_i g_T([X,E_i],E_i) \, . \]
\end{proof}
\begin{definition}
A foliation $ \mc{F} $ is called \emph{taut} if there exists a Riemannian metric $ g $ on $ M $ such that the leaves of $ \mc{F} $ are minimal submanifolds of $ M $ with respect to $ g $. If a Riemannian foliation is taut, then $ g $ may be chosen to be bundle-like \cite[Proposition~7.6]{Tond}, \linebreak in which case $ (\mc{F},g) $ is called \emph{harmonic}.
\end{definition}
One key tool for us will be following transverse divergence \mbox{theorem \cite[Theorem 4.35]{Tond}:}
\begin{theorem} \label{div}
Let $ M $ be a closed oriented manifold, endowed with a transversely orien- ted harmonic Riemannian foliation $ (\mc{F}, g) $. Then for any foliate vector field \mbox{$ X \in \mf{fol}(M) $:}
\[ \int_M \Div_T X \cdot \mu = 0 \, . \]
\end{theorem}
\section{Basic Hodge Theory} \label{sHodge}
Let $ M $ be a smooth manifold, endowed with a foliation $ \mc{F} $ of codimension $ q $ defined by an integrable subbundle $ E \subset TM $.
\begin{definition}
A differential $ k $-form $ \omega \in \Omega^k(M) $ is called \emph{basic} if $ \iota_X \omega = 0 $ as well as $ \mc{L}_X \omega = \iota_X d\omega = 0 $ for all $ X \in \Gamma_\ell(E) $. Note that for $ f \in \Omega^0(M) $ this coincides with the \autoref{vfs} of a basic function. \\
If $ \omega $ is basic, then so is $ d\omega $, meaning the basic differential forms constitute a subcomplex $ \Omega_B(\mc{F}) $ of the de Rham complex $ \Omega(M) $. Clearly, $ \Omega_B^k(\mc{F}) = 0 $ for $ k > q $. We denote the restriction of $ d $ to $ \Omega_B(\mc{F}) $ by $ d_B $. The cohomology ring of the complex $ (\Omega_B(\mc{F}), d_B) $ is called the \emph{basic cohomology of $ \mc{F} $} and will be denoted by $ H_B(\mc{F}) $. The \emph{basic Betti numbers of $ \mc{F} $} are defined as $ b_k(\mc{F}) := \dim H_B^k(\mc{F}) $.
\end{definition}
The inclusion $ \Omega_B^1(\mc{F}) \to \Omega^1(M) $ induces an injective map $ H_B^1(\mc{F}) \to H^1(M) $ \cite[Proposition 4.1]{Tond}. Furthermore, if $ M $ is closed and $ (\mc{F},g_T) $ is a Riemannian foliation on $ M $, then $ b_k(\mc{F}) < \infty $ for $ k=0, \ldots, q $ \cite[Chapter 4]{Tond}. \\
From now on we assume that $ M $ is closed and oriented, $ (\mc{F},g_T) $ is a transversely oriented Riemannian foliation on $ M $ and $ g $ is a bundle-like metric compatible with $ g_T $. As usual, the metric $ g $ induces an inner product on $ \Lambda^k T_x^*M $ for every $ x \in M $. We let $ \mu \in \Omega^n(M) $ denote the Riemannian volume form of $ (M,g) $ and consider the inner product $ \langle \cdot, \cdot \rangle $ on $ \Omega^k(M) $ given by
\[ \langle \omega,\omega'\rangle := \int_M g(\omega, \omega') \cdot \mu \, , \quad \omega, \omega' \in \Omega^k(M) \, . \]
We write $ \langle \cdot, \cdot\rangle_B $ for the restriction of $ \langle \cdot, \cdot\rangle $ to the subspace $ \Omega_B^k(\mc{F}) \subset \Omega^k(M) $.
\begin{definition}
The \emph{basic codifferential} $ \delta_B: \Omega^k_B(\mc{F}) \to \Omega_B^{k-1}(\mc{F}) $ is the formal adjoint of $ d_B: \Omega_B^{k-1}(\mc{F}) \to \Omega_B^k(\mc{F}) $ with respect to $ \langle \cdot, \cdot\rangle_B $, viz.
\[ \langle d_B\omega, \eta\rangle_B = \langle \omega, \delta_B \eta\rangle_B \, , \quad \omega \in \Omega_B^{k-1}(\mc{F}), \, \eta \in \Omega^k_B(\mc{F}) \, . \]
The \emph{basic Laplacian} is given by
\[ \Delta_B := d_B \delta_B + \delta_B d_B: \Omega_B^k(\mc{F}) \to \Omega_B^k(\mc{F}) \, . \]
A basic form $ \omega \in \Omega_B^k(\mc{F}) $ is called \emph{basic harmonic} if $ \Delta_B \omega = 0 $. The vector space of all basic harmonic $ k $-forms will be denoted by $ \mc{H}_B^k(\mc{F}) $.
\end{definition}
Beware that $ \Delta_B $ is not the restriction of the ordinary Laplacian $ \Delta = d\delta+\delta d $ to $ \Omega_B^k(\mc{F}) $ \cite[Equation 7.28]{Tond}. On basic functions $ \Delta_B $ differs from the previously defined transverse Laplacian $ \Delta_T $ by a sign, see \autoref{laps}.
By definition of $ \delta_B $, every $ \omega \in \Omega_B^k(\mc{F}) $ satisfies
\[ \langle \Delta_B\omega,\omega\rangle_B = \langle d_B\delta_B\omega,\omega\rangle_B + \langle \delta_B d_B\omega,\omega\rangle_B = \langle d_B\omega,d_B\omega \rangle_B + \langle \delta_B \omega, \delta_B \omega \rangle_B \, . \]
Therefore $ \Delta_B \omega = 0 $ if and only if both $ d_B \omega = 0 $ and $ \delta_B \omega = 0 $. In particular, we have a natural map $ \mc{H}_B^k(\mc{F}) \to H_B^k(\mc{F}) $. In case the bundle-like metric is chosen appropriately, there is the following basic Hodge theorem \cite[Theorem 7.51]{Tond}:
\begin{theorem} \label{Hodge}
Let $ M $ be a closed oriented manifold, endowed with a transversely orien- ted harmonic Riemannian foliation $ (\mc{F}, g) $. Then the natural map $ \mc{H}_B^k(\mc{F}) \to H_B^k(\mc{F}) $ is an isomorphism.
\end{theorem}
In preparation for the Bochner technique in the next section, we now specialize to one-forms: Recall the usual one-to-one correspondence between vector fields $ X \in \Gamma(TM) $ and their $ g $-dual one-forms $ \omega_X := \iota_X g \in \Omega^1(M) $. One easily checks that $ X \in \mf{trans}(\mc{F}) $ if and only if $ \omega_X \in \Omega_B^1(\mc{F}) $.
\begin{lemma} \label{symm}
We have $ d_B\omega_X = 0 $ if and only if $ \nabla^T X $ is $ g_T $-symmetric, i.e.
\[ g_T(\nabla^T_Y X,Z) = g_T(Y,\nabla^T_Z X) \, , \quad Y, Z \in \Gamma_\ell(TM) \, . \]
\end{lemma}
\begin{proof}
The Koszul formula from \autoref{bott} can be rewritten as
\[ 2g_T(\nabla^T_Y X,Z) = (d_B\omega_X)(Y,Z) + (\mc{L}_X g_T)(Y,Z) \, , \quad Y,Z \in \Gamma_\ell(TM) \, . \]
Since $ d_B\omega_X $ is skew-symmetric and $ \mc{L}_X g_T $ is symmetric, this yields the claim.
\end{proof}
\begin{lemma} \label{codif}
If $ (\mc{F},g) $ is harmonic, then $ \delta_B \omega_X = - \Div_T X $.
\end{lemma}
\begin{proof}
By definition of $ \delta_B $, we need to show that for all basic functions $ f \in \Omega_B^0(\mc{F}) $:
\[ \int_M g(d_B f, \omega_X) \cdot \mu = - \int_M f \cdot \Div_T X \cdot \mu \, . \]
Using \autoref{dlem} we calculate:
\[ \Div_T (f \cdot X) = f \cdot \Div_T X + g(\nabla f,X) = f \cdot \Div_T X + g(d_Bf,\omega_X) \, . \]
If we integrate over $ M $, then the left-hand side vanishes by \autoref{div}, since $ f \cdot X $ is foliate and $ (\mc{F},g) $ is harmonic, and the claim follows.
\end{proof}
\begin{remark} \label{laps}
\cref{dlem,codif} imply that for all basic functions $ f $:
\[ \Delta_B f = \delta_B d_B f = \delta_B \omega_{\nabla f} = - \Div_T \nabla f = - \Delta_T f \, . \]
\end{remark}
\begin{cordef} \label{cordef}
If $ (\mc{F},g) $ is harmonic, then $ \omega_X $ is basic harmonic if and only if $ \nabla^T X $ is $ g_T $-symmetric and $ \Div_T X = 0 $. In this case we call $ X $ \emph{basic harmonic}.
\end{cordef}
\section{A Bochner Technique} \label{sBochner}
From now on, let $ M $ be a connected closed oriented manifold, endowed with a transversely oriented harmonic Riemannian foliation $ (\mc{F}, g) $ of codimension $ q $.
\begin{definition}
A transverse vector field $ X \in \mf{trans}(\mc{F}) $ is \emph{transverse parallel} if \mbox{$ \nabla^T X = 0 $.}
\end{definition}
Beware that a transverse vector field $ X \in \mf{trans}(\mc{F}) $ which is parallel in the usual sense that $ \nabla X = 0 $ is also transverse parallel, but the converse is not true. By virtue of \autoref{dlem} and \autoref{cordef}, every transverse parallel vector field is basic harmonic.
\begin{lemma} \label{const}
Transverse parallel vector fields have constant length. Hence, they are uniquely determined by their value at one point.
\end{lemma}
\begin{proof}
If $ X \in \mf{trans}(\mc{F}) $ is transverse parallel and $ f := \frac{1}{2} g(X,X) = \frac{1}{2} g_T(X,X) $, then for all $ Y \in \Gamma_\ell(TM) $:
\[ Y(f) = g_T(\nabla^T_Y X,X) = 0 . \]
Since $ M $ is connected, this implies that $ f $ is constant.
\end{proof}
For an endomorphism field $ A \in \Gamma(\End(TM)) $, we set
\[ |A|^2 := \tr (A \circ A^*) = \sum_i g\big(A(E_i), A(E_i)\big) \, , \]
where $ E_1, \ldots, E_n $ is a local orthonormal frame.
\begin{proposition} \label{harmeq}
Let $ X \in \mf{trans}(\mc{F}) $ be a basic harmonic vector field and consider the basic function $ f := \frac{1}{2} \, g_T(X,X) $. Then:
\begin{enumerate}
\item[a)] $ \nabla f = \nabla_X^T X $.
\item[b)] $ \Hess_T f (Y,Y) = g_T(\nabla_Y^T X, \nabla_Y^T X) + R^T(Y,X,X,Y) + g_T(\nabla^T_X \nabla^T_Y X,Y) - g_T(\nabla^T_{\nabla^T_X Y} X, Y) $ for all $ Y \in \Gamma_\ell(E^\perp) $.
\item[c)] $ \Delta_T f = |\nabla^T X|^2 + \Ric^T (X,X) $.
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{enumerate}
\item[a)] By virtue of \autoref{symm}:
\[ g(\nabla f, Y) = Y(f) = g_T(\nabla^T_Y X, X) = g_T(\nabla^T_X X, Y) = g(\nabla^T_X X, Y) \, , \quad Y \in \Gamma_\ell(TM) \, . \]
\item[b)] Part a) and \autoref{symm} imply that for all $ Y \in \Gamma_\ell(E^\perp) $:
\begin{align*}
\Hess_T f(Y,Y) &= g_T(\nabla^T_Y \nabla f, Y) = g_T(\nabla_Y^T \nabla^T_X X, Y) \\
&= R^T(Y,X,X,Y) + g_T(\nabla^T_X \nabla_Y^T X,Y) + g_T(\nabla^T_{[Y,X]} X,Y) \\
&= R^T(Y,X,X,Y) + g_T(\nabla^T_X \nabla_Y^T X,Y) + g_T(\nabla^T_{\nabla^T_Y X}X,Y) - g_T(\nabla^T_{\nabla_X^T Y}X,Y) \\
&= R^T(Y,X,X,Y) + g_T(\nabla^T_X \nabla_Y^T X,Y) + g_T(\nabla^T_Y X, \nabla^T_Y X) - g_T(\nabla^T_{\nabla_X^T Y}X,Y) \\
&= g_T(\nabla_Y^T X, \nabla_Y^T X) + R^T(Y,X,X,Y) + g_T(\nabla^T_X \nabla^T_Y X,Y) - g_T(\nabla^T_{\nabla^T_X Y} X, Y) \, .
\end{align*}
From the second to the third line we implicitly used that $ \nabla_{[Y,X]}^T X = \nabla_{[Y,X]^\perp}^T X $, since $ X \in \mf{trans}(\mc{F}) $.
\item[c)] If we sum b) over any local orthonormal frame, then the first two terms on the right-hand side yield $ |\nabla^T X|^2 $ and $ \Ric^T(X,X) $, as desired. \\
Fix a point $ x \in M $. As shown in \cite[Section 3]{KTT}, there exists a local orthonormal frame $ E_1, \ldots, E_n $ in a neighborhood of $ x $ such that $ E_1, \ldots, E_p \in \Gamma_\ell(E) $, $ E_{p+1}, \ldots, E_n \in \Gamma_\ell(E^\perp) $ and $ (\nabla^T E_i)_x = 0 $ for $ i = p+1, \ldots, n $. If we sum b) at $ x $ over such a frame, then the last term on the right-hand side vanishes and the third term reduces to
\[ \sum_i g_T(\nabla^T_X \nabla^T_{E_i} X, E_i) = \sum_i X\big(g_T(\nabla_{E_i}^T X, E_i)\big) = X( \Div_T X) = 0 \, . \]
\end{enumerate}
\end{proof}
We can now finally come to our main result:
\begin{theorem} \label{Bochner}
Let $ M $ be a connected closed oriented manifold, endowed with a transversely oriented harmonic Riemannian foliation $ (\mc{F}, g) $. If $ \Ric^T $ is positive semi-definite everywhere, then every basic harmonic vector field is transverse parallel. If additionally $ \Ric^T $ is positive definite at one point, then no non-trivial basic harmonic vector fields exist.
\end{theorem}
\begin{proof}
Let $ X \in \mf{trans}(\mc{F}) $ be a basic harmonic vector field and $ f := \frac{1}{2} \, g_T(X,X) $. By virtue of \autoref{dlem}, \autoref{div} and \autoref{harmeq}:
\[ 0 = \int_M \Delta_T f \cdot \mu = \int_M \Big(|\nabla^T X|^2 + \Ric^T (X,X) \Big) \cdot \mu \geq \int_M |\nabla^T X|^2 \cdot \mu \geq 0 \, . \]
Therefore $ | \nabla^T X|^2 $ vanishes everywhere, meaning $ X $ is transverse parallel. Furthermore, also $ \Ric^T(X,X) $ vanishes everywhere, so if additionally $ \Ric^T $ is positive definite at one point, then $ X $ vanishes at that point. But then $ X $ vanishes everywhere by virtue of \autoref{const}.
\end{proof}
\begin{remark}
Note that \autoref{Bochner} is indeed a generalization of \autoref{posBoch}: If $ \mc{F} $ is the trivial foliation of $ M $ by singletons (i.e.~the corresponding integrable distribution $ E = 0 $), then transverse orientability of $ \mc{F} $ coincides with ordinary orientability of $ M $, the Riemannian foliation $ (\mc{F},g) $ is trivially harmonic and $ \Ric^T = \Ric $. Furthermore, basic harmonic and transverse parallel vector fields are nothing else than ordinary harmonic and parallel vector fields in this case.
\end{remark}
\begin{corollary}
If $ (M,\mc{F}, g) $ are as in \autoref{Bochner} and $ \Ric^T $ is positive semi-definite everywhere, then $ b_1(\mc{F}) \leq q = \codim \mc{F} $. If additionally $ \Ric^T $ is positive definite at one point, then $ b_1(\mc{F}) = 0 $.
\end{corollary}
\begin{proof}
\autoref{Hodge} states that $ b_1(\mc{F}) = \dim \mc{H}_B^1(\mc{F}) $. Fix a point $ x \in M $ and consider the linear map $ \mc{H}_B^1(\mc{F}) \to T_xE^\perp , \, \omega_X \mapsto X_x $. By virtue of \autoref{const} and \autoref{Bochner}, this map is injective, meaning $ b_1(\mc{F}) \leq \dim T_x E^\perp = q $. If additionally $ \Ric^T $ is positive definite at one point, then \autoref{Bochner} even yields $ b_1(\mc{F}) = 0 $.
\end{proof}
We conclude this section by deriving \autoref{tflat}, for which we first need the following
\begin{definition}
\vspace{-.5cm} A transverse vector field $ X \in \mf{trans}(\mc{F}) $ is \emph{transverse Killing} if \mbox{$ \mc{L}_X g_T = 0 $.} We denote the vector space of all transverse Killing fields of $ (\mc{F},g) $ by $ \mf{iso}(\mc{F}) $.
\end{definition}
Again, a transverse vector field $ X \in \mf{trans}(\mc{F}) $ which is Killing in the usual sense that $ \mc{L}_X g = 0 $ is also transverse Killing, but the converse is not true. The same argument as in the proof of \autoref{symm} shows that $ X \in \mf{trans}(\mc{F}) $ is transverse Killing if and only if $ \nabla^T X $ is $ g_T $-skew-symmetric. This also demonstrates that every transverse parallel vector field is transverse Killing. Combining \Cref{tnegBoch,Bochner} yields the following
\begin{corollary} \label{tflat2}
If $ (M,\mc{F}, g) $ are as in \autoref{Bochner} and $ \Ric^T $ vanishes everywhere, then $ \dim \mf{iso}(\mc{F}) = b_1(\mc{F}) $.
\end{corollary}
\section{Applications} \label{Sas}
We conclude this article by applying \autoref{tflat2} to two classes of spaces which naturally satisfy all of the required conditions, namely \emph{degenerate 3-$(\alpha,\delta)$-Sasaki} and certain \emph{Sasaki-$ \eta $-Einstein manifolds}, which arise for example as Boothby-Wang bundles over hyperkähler and Calabi-Yau manifolds, respectively. We only give minimal expositions of these geometries here and refer the interested reader to the comprehensive monograph \cite{BG} as well as the introductory articles \cite{AD} and \cite{ADS} about 3-$ (\alpha,\delta) $-Sasaki manifolds and the recent publication \cite{GRS} which focusses specifically on the degenerate case.
\begin{definition}
Let $ (M^{2n+1}, g, \xi, \eta,\varphi) $ be a Riemannian manifold, endowed with a unit length vector field $ \xi $, its $ g $-dual one-form $ \eta $ and an almost Hermitian structure $ \varphi $ on $ \ker \eta $. Then $ M $ is an \emph{almost contact metric manifold} if
\[ \varphi \,\xi = 0 \, , \qquad \varphi^2 = - \id + \xi \otimes \eta \, , \qquad g \circ (\varphi \times \varphi) = g - \eta \otimes \eta \, . \]
The \emph{Reeb vector field} $ \xi $ spans an integrable distribution $ E $ which defines the so-called \emph{characteristic foliation} $ \mc{F} $. The \emph{fundamental 2-form} is given by $ \Phi(X,Y) := g(X,\varphi Y) $ and $ M $ is a \emph{Sasaki manifold} if $ [\varphi,\varphi] + d\eta \otimes \xi = 0 $ as well as $ d\eta = 2 \Phi $. A Sasaki manifold is called \emph{$ \eta $-Einstein} if its Ricci curvature tensor satisfies $ \Ric = a g + b \eta \otimes \eta $ for some constants $a,b \in \R $.
\end{definition}
We can endow any Sasaki manifold with an orientation and its characteristic foliation with a compatible transverse orientation by using the volume forms $ (d\eta)^n \wedge \eta $ and $ (d\eta)^n$, respectively. The characteristic foliation of any Sasaki manifold harmonic and admits a transverse Kähler structure \cite{BG}. \\
In order to apply \autoref{tflat2} we limit ourselves to the case where the Kähler structure is Ricci-flat, i.e.~Calabi-Yau. In this case the Sasaki manifold is not Einstein in the ordinary sense that $ \Ric = a g $ but instead $ \eta $-Einstein with $ b \neq 0 $ \cite[Theorem 11.1.3]{BG}. Examples of Sasaki-$ \eta $-Einstein manifolds can be constructed via the famous \emph{Boothby-Wang bundle} \cite{BW}:
\begin{theorem}
Let $ N $ be a Calabi-Yau manifold with integral Kähler class. Then a certain $ S^1 $-bundle over $ N $ admits a Sasaki-$ \eta $-Einstein structure.
\end{theorem}
3-$ (\alpha, \delta) $-Sasaki geometry was devised by \textsc{Agricola} and \textsc{Dileo} as a common gene-ralization to accomodate both 3-Sasaki manifolds and other interesting examples like the quaternionic Heisenberg groups \cite{AD}. This new class of manifolds retains favorable properties like hypernormality and canonicity \cite[Theorem 2.2.1 \& Corollary 2.3.3]{AD}.
\begin{definition}
Let $ (M^{4n+3}, g, \xi_i, \eta_i,\varphi_i)_{i=1,2,3} $ be a Riemannian manifold, endowed with three almost contact metric structures. Then $ M $ is an \emph{almost 3-contact metric manifold} if their interrelation is governed by
\begin{gather*}
\varphi_i \, \xi_j=\xi_k \, ,\qquad \eta_i\circ\varphi_j=\eta_k\, , \qquad \varphi_i \circ \varphi_j=\varphi_k+\xi_i\otimes\eta_j
\end{gather*}
for any even permutation $ (i,j,k) $ of $ (1,2,3) $. The three \emph{Reeb vector fields} $ \xi_i $ span an integrable distribution $ E $ which defines the so-called \emph{characteristic foliation $\mc{F} $}. The \emph{fundamental 2-forms} are given by $ \Phi_i(X,Y) := g(X, \varphi_iY) $ and $ M $ is a \emph{3-$(\alpha,\delta)$-Sasaki manifold} if there exist $ \alpha, \delta \in \R $, $ \alpha \neq 0 $, such that
\[ d\eta_i = 2\alpha \Phi_i + 2 (\alpha-\delta) \eta_j \wedge \eta_k \]
for any even permutation $ (i,j,k) $ of $ (1,2,3) $. Finally, $ M $ is called \emph{positive} if $ \alpha \delta > 0 $, \emph{negative} if $ \alpha \delta < 0 $ and \emph{degenerate} if $ \delta = 0 $.
\end{definition}
Again, we can endow any 3-$(\alpha,\delta) $-Sasaki manifold with an orientation and its characteristic foliation with a compatible transverse orientation using $ (d\eta_1)^{2n} \wedge \eta_1 \wedge \eta_2 \wedge \eta_3 $ and $ (d\eta_1)^{2n}$, respectively. Furthermore, the characteristic foliation is harmonic \cite[Corollary 2.3.1]{AD} and we can easily control the sign of the transverse Ricci curvature tensor, cf.~\autoref{curv} \cite[Theorem 2.2.1]{ADS}:
\begin{theorem}
The characteristic foliation of a 3-$(\alpha, \delta) $-Sasaki manifold induces local Riemannian submersions onto a quaternionic Kähler manifold whose Ricci curvature is positive/negative/zero if $ M $ is positive/negative/degenerate.
\end{theorem}
Interesting examples of \emph{degenerate} 3-$(\alpha,\delta)$-Sasaki manifolds can be constructed as \linebreak \emph{3-Boothby-Wang bundles} over hyperkähler manifolds \cite[Theorem 3.1]{GRS}:
\begin{theorem} \label{BW}
Let $ N $ be a hyperkähler manifold with integral Kähler classes. Then a certain $ T^3 $-bundle over $ N $ admits a degenerate 3-$(\alpha,\delta) $-Sasaki structure.
\end{theorem}
As outlined above, \autoref{tflat2} can be applied to both 3-$ (\alpha,\delta) $-Sasaki manifolds and Sasaki-$ \eta $-Einstein manifolds with transverse Calabi-Yau structure:
\begin{corollary} \label{corapp}
Let $ M $ be a connected closed degenerate 3-$(\alpha, \delta) $-Sasaki manifold or Sasaki-$ \eta $-Einstein manifold with transverse Calabi-Yau structure. Then the characteristic foliation $ \mc{F} $ satisfies $\dim \mf{iso}(\mc{F}) = b_1(\mc{F}) $.
\end{corollary}
In this context, one is usually not so much interested in transverse Killing fields per se, but rather in the following notion:
\begin{definition}
An \emph{automorphism of a Sasaki manifold} $ M $ is an isometry which satisfies one of the equivalent conditions $ \phi_\ast \xi = \xi $, $ \phi^* \eta = \eta $ or $ \phi_\ast \circ \varphi = \varphi \circ \phi_\ast $. The collection of all such automorphisms constitutes a Lie group, which we denote by $ \Aut(M) $. The Lie algebra $ \mf{aut}(M) $ of $ \Aut(M) $ is comprised of all complete Killing vector fields $ X $ which satisfy $ \mc{L}_X \xi = 0 $, $ \mc{L}_X \eta = 0 $ and $ \mc{L}_X \varphi = 0 $. \\
An \emph{automorphism of a 3-$(\alpha,\delta)$-Sasaki manifold} $ M $ is an isometry $ \phi:M \to M $ which satisfies one of the equivalent conditions $ \phi_\ast \xi_i = \xi_i $, $ \phi^* \eta_i = \eta_i $ or $ \phi_\ast \circ \varphi_i = \varphi_i \circ \phi_\ast $ for $ i=1,2,3 $. The collection of all such automorphisms constitutes a Lie group, which we denote by $ \Aut(M) $. The Lie algebra $ \mf{aut}(M) $ of $ \Aut(M) $ is comprised of all complete Killing vector fields $ X $ which satisfy $ \mc{L}_X \xi_i = 0 $, $ \mc{L}_X \eta_i = 0 $ and $ \mc{L}_X \varphi_i = 0 $ for $ i = 1,2,3 $.
\end{definition}
From now on, let $ M $ be a connected closed degenerate 3-$(\alpha, \delta) $-Sasaki manifold or Sasaki-$ \eta $-Einstein manifold with transverse Calabi-Yau structure, let $ E $ be the integrable distribution spanned by the Reeb vector field(s) and $ \mc{F} $ the characteristic foliation.
\begin{lemma}
The orthogonal projection of any infinitesimal automorphism to $ E^{\perp} $ is a transverse Killing field.
\end{lemma}
\begin{proof}
Let $ X \in \mf{aut}(M) $ be an infinitesimal automorphism and let $ X^\top $, $ X^\perp $ denote its orthogonal projections to $ E $, $ E^\perp $, respectively. Because $ X $ commutes with the Reeb vector field(s), it follows that $ X \in \mf{fol}(\mc{F}) $. Since $ X^\top $ is trivially foliate, we obtain that $ X^\perp = X - X^\top \in \mf{fol}(\mc{F}) $. This implies that $ (\mc{L}_{X^\perp}g_T)(Y,Z) = 0 $ if $ Y $ or $ Z $ lies in $ \Gamma_\ell(E) $. Furthermore $ \mc{L}_X g = 0 $, since $ X $ is Killing and $ \mc{L}_{X^\top} g_T = 0 $ because $ g_T $ is a transverse metric. Hence, if both $ Y,Z \in \Gamma_\ell(E^\perp) $:
\begin{align*}
(\mc{L}_{X^\perp} g_T) (Y,Z) &= (\mc{L}_X g_T)(Y,Z) \\
&= X\big(g_T(Y,Z)\big) - g_T([X,Y],Z) - g_T(Y,[X,Z]) \\
&= X\big(g(Y,Z)\big) - g([X,Y],Z) - g(Y,[X,Z]) \\
&= (\mc{L}_X g)(Y,Z) = 0 \, .
\end{align*}
\end{proof}
Therefore $ \pi: \mf{aut}(M) \to \mf{iso}(\mc{F}) , \, X \mapsto X^\perp $ is a well-defined linear map. We can determine the kernel of $ \pi $ using an argument from \cite[Lemma 4.2]{GRS}:
\begin{lemma}
The kernel of $ \pi $ is comprised of all the linear combinations of the Reeb vector field(s) with constant coefficients.
\end{lemma}
\begin{proof}
Clearly, all linear combinations of the Reeb vector field(s) with constant coefficients lie in the kernel of $ \pi $. Conversely, let $ X \in \ker \pi = \mf{aut}(M) \cap \Gamma(E) $. Then $ X = f \xi $ or $ X = \sum_{i=1}^3 f_i \xi_i $ with $ f,f_1,f_2,f_3 \in C^\infty(M) $, respectively. For all $ Y \in \Gamma_\ell(TM) $, we have
\[ 0 = \big(\mathcal{L}_X \eta\big)(Y) = f \underbrace{(\mathcal{L}_{\xi} \eta)}_{=0}(Y) + Y(f) \underbrace{\eta(\xi)}_{=1} = Y(f) \, , \]
respectively
\[ 0 = \big(\mathcal{L}_X \eta_j\big)(Y) = \sum_{i=1}^3 \Big(f_i \underbrace{(\mathcal{L}_{\xi_i} \eta_j)}_{=0}(Y) + Y(f_i) \underbrace{\eta_j(\xi_i)}_{=\delta_{ij}}\Big) = Y(f_j) \, . \]
Since $ M $ is connected, it follows that $ f,f_1,f_2,f_3 $ have to be constant.
\end{proof}
Hence, the rank-nullity theorem and \autoref{corapp} yield:
\[ \dim \mf{aut}(M) = \dim \im \pi + \dim \ker \pi \leq \dim \mf{iso}(\mc{F}) + \rk E = b_1(\mc{F}) + \rk E \, . \]
We have thus arrived at our final two theorems:
\begin{theorem}
Let $ M $ be a connected closed degenerate 3-$(\alpha, \delta) $-Sasaki manifold with characteristic foliation $ \mc{F} $. Then the dimension of the automorphism group $ \Aut(M) $ is at most $ b_1(\mc{F}) + 3 $. \\
In particular, if $ M $ arises via \cite[Theorem 3.1]{GRS} as a $ T^3 $-bundle over a compact hyperkähler manifold $ N $ with integral Kähler classes, then $ \dim \Aut(M) \leq b_1(N) + 3 $.
\end{theorem}
\begin{theorem}
Let $ M $ be a connected closed Sasaki-$ \eta $-Einstein manifold with transverse Calabi-Yau structure and characteristic foliation $ \mc{F} $. Then the dimension of the automorphism group $ \Aut(M) $ is at most $ b_1(\mc{F}) + 1 $. \\
In particular, if $ M $ arises as the Boothby-Wang bundle over a compact Calabi-Yau mani- fold $ N $ with integral Kähler class, then $ \dim \Aut(M) \leq b_1(N) + 1 $.
\end{theorem}
\begin{remark}
One might ask if there is even equality $ \dim \Aut(M) = b_1(\mc{F}) + \rk E $ in the above theorems. This is equivalent to the question if $ \pi : \mf{aut}(M) \to \mf{iso}(\mc{F}) $ is surjective or if every transverse Killing field can be extended to an infinitesimal automorphism. In the context of Sasaki manifolds, this problem was further rephrased in \cite[Theorem 8.1.8]{BG}, where they obtain that a transverse Killing field $ X \in \mf{iso}(\mc{F}) $ extends to an infinitesimal automorphism if and only if the basic cohomology class $ [\iota_X d\eta] \in H_B^1(\mc{F}) $ vanishes. The same holds for degenerate 3-$(\alpha,\delta) $-Sasaki manifolds if the classes $ [\iota_X d\eta_i] $ vanish for \mbox{$ i =1,2,3 $.} In the special case $ b_1(\mc{F}) = 0 $ this leads to alternative proofs of the above theorems.
\end{remark}
\printbibliography
\vspace{1cm}
\textsc{Leon Roschig, Philipps-Universität Marburg, Fachbereich Mathematik und Informatik, Hans-Meerwein-Straße 6, 35043 Marburg} \\
\textit{E-mail address}: \texttt{<EMAIL>}
\end{document} |
\section{Introduction}
The problem of constructing martingales which match given marginales has been the subject of many works in probability theory and applications thereof. The origin of the problem dates back to Strassen \cite{Strassen}. He showed that a discrete-time martingale $(M_n)_{n\in \mathbb{N}}$ can match a sequence of real probability measures $(\mu_n)_{n\in \mathbb{N}}$ if and only if the measures $(\mu_n)_{n\in \mathbb{N}}$ are convexly ordered, i.e., $M_n \sim \mu_n$ is possible for all $n \in \mathbb{N}$ if and only if $ \int f(x) \, \mathrm{d}\mu_n(x)$ is an increasing sequence in $n$ for any positive convex function $f$. This result also extends to the continuous-time setting, see Kellerer's theorem \cite{Kellerer}.
\bigskip
Strassen's and Kellerer's theorems do not provide a method for constructing the matching martingales. Many techniques are dedicated to Kellerer's setting, where one is looking to construct a martingale that matches a given peacock \cite{Hirsch}: Skorokhod's embedding problem, Brownian random time changes, local volatility models, scaling peacocks method, etc., see \cite{Madan}. Since the target is a dynamical measure, the matching is in a weak sense, and the martingale is in continuous time. On the other hand, the matching in Strassen's setting can be weak or almost sure, depending on the nature of the targets (measures or random variables), and the martingale is in discrete time. A widely known approach for this task is martingale optimal transport (MOT, \cite{Beiglbock}, \cite{Beiglbock2}). Inspired by Kantorovich's optimal transport theory (\cite{Kantorovich}, \cite{Villani1}, \cite{Bogachev}), it provides a framework for selecting a martingale coupling by minimizing a cost functional over the set of all martingale couplings.
\bigskip
Our interest lies at the intersection of Strassen's and Kellerer's settings: how to construct a continuous-time martingale that matches, in law or almost surely, a given finite set of convexly ordered random variables, at pre-specified given dates? For example, this is of interest in a financial setting, where knowledge of the prices of vanilla call and put options provides an implied distribution for the underlying asset price at future dates under a no arbitrage requirement. This is what we call the martingale interpolation problem, since we wish to interpolate with a continuous-time martingale between the components of a discrete-time martingale. Imposing the martingale property on the interpolating process is the challenging part of this problem. If one put the martingale condition aside, then bridging the gap between a coupling and an interpolating measure would usually be a manageable proposition. For instance, in Kantorovich's optimal transport theory, it is well known that there is a one-to-one correspondence between optimal couplings and optimal interpolating measures in Wasserstein spaces, called McCann's interpolations, under mild conditions. However, this property is not present in the martingale counterpart of optimal transport. Solutions to the martingale interpolation problem exist. For example, the martingale Benamou-Brenier problem (MBBP) \cite{Veraguas} aims to solve the optimization problem
\begin{equation*}
\sup\limits_{\mathcal M_t(\mu,\nu,B) } \mathbb{E} \left[\int_{0}^{1} \sigma_{t} \, \mathrm{d} t\right]
\end{equation*}
over the set
\begin{align}
\mathcal M_t(\mu,\nu,B) &= \left\{(\sigma_t)_{t \geqslant 0} \in L^1(B_t) \| M_t = M_0 + \int_0^t \sigma_s \, \mathrm{d} B_s \implies \right. \nonumber \\
&\hspace{5.1cm} M_0 \sim \mu, M_1 \sim \nu, (M_t) \text{ is a martingale} \Bigg\},
\end{align}
where $(B_t)_{t \geqslant 0}$ is a standard Brownian motion and $(\mu,\nu)$ are convexly ordered probability measures with finite second moments. If $(\mu,\nu)$ is irreducible in the sense of \cite{Beiglbock}, Appendix A.1, the solutions to the MBBP are called $(\mu,\nu)$-stretched Brownian motions, and have the form $\mathbb{E}[ F(B_1 + Y_0) \| \mathcal{F}_t]$, where $(\mathcal{F}_t)$ is the filtration generated by $(B_t+Y_0)$. The function $F:\mathbb{R} \rightarrow \mathbb{R} $ and the random variable $Y_0$ must be selected such that $F(B_1 + Y_0) \sim \nu$, $\mathbb{E}[ F(B_1 + Y_0) \| Y_0] \sim \mu$ and $Y_0 \perp\!\!\!\perp (B_t)$. In the case where $(\mu,\nu)$ have irreducible components, the solution to the MBBP is a stretched Brownian motion on each irreducible component, see Theorem 3.1 in \cite{Veraguas}. Stretched Brownian motions are interpolating martingales between the given measures $\mu$ and $\nu$. Another weak solution to the martingale interpolation problem is given by Schrödinger's volatility models (SVMs) \cite{Labordere}. Inspired by the entropic relaxation of optimal transport (\cite{Leonard1}, \cite{Peyre1}), also known as Schrödinger's problem in specific cases \cite{Schrodinger}, SVMs rely on a measure change. Consider two, possibly correlated, Brownian motions $(B_t)_{t \geqslant 0}$ and $(\tilde B_t)_{t \geqslant 0}$, respectively, under a measure $\mathbb{P}$, along with the system of SDEs
\begin{align}
&\, \mathrm{d} M_{t} = a_{t}\, M_{t}\, \mathrm{d}\widetilde B_t, \nonumber\\
&\, \mathrm{d} a_{t}=b(a_t)+ c\left(a_{t}\right)\, \mathrm{d} B_t,
\end{align}
where the real functions $b$ and $c$ are given, and are such that $a_t$ is an Itô process. The goal is to change the measure $\mathbb{P}$ by an equivalent one, such that the process $(M_t)_{0 \leqslant t \leqslant 1}$ matches the given measures $\mu$ and $\nu$ at times zero and one, respectively, while being a true martingale. If several equivalent measures that satisfy these constraints exist, one must select the one that is closest to $\mathbb{P}$ in terms of the Kullback-Leibler divergence. By Girsanov's theorem, changing $\mathbb{P}$ by an equivalent measure $\mathbb{Q}$ transforms the system of SDEs as follows:
\begin{align}
&\, \mathrm{d} M_{t}= a_{t} M_{t} (\, \mathrm{d}\widetilde B_t + \widetilde \lambda_t \, \mathrm{d} t), \nonumber\\
&\, \mathrm{d} a_{t}=b(a_t)+ c\left(a_{t}\right) (\, \mathrm{d} B_t + \lambda_t \, \mathrm{d} t).
\end{align}
One must then select the drifts $(\tilde \lambda_t)$ and $(\lambda_t)$, generated by the same measure change, that minimize the functional $\mathbb{E}_\mathbb{Q} [ \int_{T_0}^{T_1} \widetilde{\lambda}^2_s + \lambda^2_s \, \mathrm{d} s]$ while satisfying the constraints $M_0 \sim \mu$ and $M_1 \sim \nu$, and that $(M_t)$ be a martingale.
\bigskip
In this paper, inspired by the information-based approach, see \cite{BHM1} and \cite{BHM2}, we construct a class of martingales that match, almost surely, any set of convexly ordered random variables, at an arbitrary set of fixed times. These interpolating martingales, called filtered arcade martingales (FAMs), need two main ingredients: a discrete-time martingale (for example, a solution to an MOT problem) and an interpolating process that we call the randomized arcade process.
\bigskip
Section 2 is concerned with the development of {\it arcade processes} (AP). Defined as a functional of a given stochastic process called the driver, APs are sample-continuous stochastic processes that interpolate between zeros on the whole probability space, i.e., omega by omega. To execute the interpolation, they rely on deterministic functions called interpolating coefficients. APs may be viewed as multi-period anticipative stochastic bridges. We study their properties and focus on Gaussian APs, which play an important role in the definition of FAMs. Starting from a Gauss-Markov driver, we show that it is always possible to construct a Markovian AP by utilizing the covariance structure of the driver. The resulting AP from this procedure is called a standard AP. This process is not unique, since one can construct infinitely many Markovian APs that are driven by the same Gauss-Markov process.
\bigskip
Section 3 treats the additive randomization of APs, that is, an AP added to a stochastic process that interpolates deterministically between the given random variables. We term such processes {\it randomized arcade processes} (RAPs), which can be thought of as a sum of a signal function and a noise process. By construction, a RAP can match any random variables on the whole probability space at any given time, i.e, it is a stochastic interpolator between target random variables in the strong sense. The notion of the Markov property does not suit RAPs in general, since their natural filtration contains the $\sigma$-algebras generated by each previously matched target random variable. For this reason, we introduce a weaker notion, the nearly-Markov property, and show under which conditions a RAP is nearly-Markov.
\bigskip
In the fourth section, we introduce the {\it filtered arcade martingales} (FAMs). A FAM is defined as the conditional expectation of the final target random variable, given the information generated by a RAP and, hence, inherits the filtering framework of the information-based approach that appears in \cite{BHM1},\cite{BHM2}: the signal is the final target random variable and it can only be observed through a noisy version, the RAP. FAMs are tractable, thanks to the nearly-Markov property enjoyed by the underlying RAPs, and can be simulated using Bayes formula. Applying Itô's lemma, we derive the stochastic differential equation satisfied by a FAM, which reveals the structure of the underlying innovations process that is adapted to the filtration generated by the respective RAP. Finally, we introduce the {\it information-based martingale optimal transport} (IB-MOT) problem, a similar problem to martingale optimal transport, that incorporates noise in the optimization process, inspired by the entropic regularization of optimal transport and Schrödinger's problem. IB-MOT selects the martingale coupling that maximizes the expectation of the cumulative weighted squared error between a target random variable and its associated FAM, for a given RAP.
\bigskip
In what follows, we consider the collection of fixed dates $\{T_i \in \mathbb{R} \| n \in \mathbb{N}_0 \mbox{ and } i=0, 1, \ldots, n\}$ such that $0\leqslant T_0 <T_{1}<T_{2}<\ldots < T_n < \infty$, and introduce the ordered sets
\begin{eqnarray*}
\{T_0,T_n\}_*&=&\{T_0, T_1, \ldots,T_n\},\\
(T_0,T_n)_*&=&\bigcup\limits_{i=0}^{n-1}(T_i,T_{i+1}).
\end{eqnarray*}
Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space, $\left(D_{t}\right)_{t \in [T_0,T_n]}$ a sample-continuous stochastic process such that $\mathbb{P}[D_t=0]<1$ whenever $t\in (T_0,T_n)_*$, and $X$ an $\mathbb{R}^{n+1}$-valued random vector that is independent of $(D_t)$.
\section{Arcade processes}
We construct stochastic processes on $[T_0,T_n]$, as a functional of $(D_t)$, which match exactly $0$ (for all $\omega \in \Omega$) at the given times $\{T_0,T_n\}_*$. The first step is to introduce deterministic functions called interpolating coefficients.
\begin{defn} \label{interpolating}
The functions $f_0,f_1,\ldots, f_n$ are interpolating coefficients on $\{T_0,T_n\}_*$ if $f_0, $ $f_1,\ldots, f_n \in C^0\left([T_0,T_n],\mathbb{R}\right), f_i(T_i)=1,$ and $f_i(T_j)=0 \text{ for }i,j=0,\ldots,n,\, i \neq j$.
\end{defn}
We can now give the definition of what we call an \emph{arcade process}.
\begin{defn} \label{AP}
An arcade process (AP), denoted $(A_{t}^{(n)})_{t \in [T_0,T_n]}$, on the partition $\{T_0,T_n\}_*$ is a stochastic process of the form
\begin{equation}
A_{t}^{(n)}:=D_{t}-\sum_{i=0}^n f_i(t)D_{T_i},
\end{equation}
where $f_0,\ldots, f_n$ are interpolating coefficients on $\{T_0,T_n\}_*$. The process $(D_t)_{t \in [T_0,T_n]}$ is the driver of the AP. We denote by $(\mathcal{F}^A_t)_{t \in [T_0,T_n]}$ the filtration generated by $(A_{t}^{(n)})_{t \in [T_0,T_n]}$.
\end{defn}
We observe that $A_{T_0}^{(n)}=A_{T_1}^{(n)}=\ldots=A_{T_n}^{(n)}=0$ by construction, for all $\omega \in \Omega$.
\begin{ex} \label{ABB}
For $n=1$, ${f_0(t)= \frac{t-T_1}{T_0-T_1}} $, $ {f_1(t)= \frac{t-T_0}{T_1-T_0}} $, the AP driven by a standard Brownian motion $\left(B_{t}\right)_{t \geqslant 0}$ is the anticipative Brownian bridge on $[T_0,T_1]$,
\begin{equation}
A_{t}^{(1)}= B_t - \frac{T_1-t}{T_1-T_0} B_{T_0} - \frac{t-T_0}{T_1-T_0} B_{T_1}.
\end{equation}
\end{ex}
\begin{ex}
For $n>1$, we can generalize the anticipative Brownian bridge by taking
\begin{equation}
f_0(t)= \frac{T_{1}-t}{T_1-T_{0}} \mathds{1}_{[T_0, T_1]} (t), \quad f_n(t)= \frac{t-T_{n-1}}{T_n-T_{n-1}} \mathds{1}_{(T_{n-1}, T_n]} (t),
\end{equation}
and
\begin{equation}
f_i(t)=\frac{t-T_{i-1}}{T_i-T_{i-1}} \mathds{1}_{(T_{i-1}, T_{i}]} (t) + \frac{T_{i+1} - t}{T_{i+1}-T_i} \mathds{1}_{(T_{i}, T_{i+1}]} (t), \quad \text{for } i=1,\ldots, n-1.
\end{equation}
We call this AP the stitched Brownian AP for it can be written as
\begin{equation}
A_{t}^{(n)} = \left\{
\begin{array}{ll}
B_t - \frac{T_1-t}{T_1-T_0} B_{T_0} - \frac{t-T_0}{T_1-T_0} B_{T_1}, & \mbox{if } t \in [T_0,T_{1}), \\ \\
B_t - \frac{T_2-t}{T_2-T_1} B_{T_1} - \frac{t-T_1}{T_2-T_1} B_{T_2}, & \mbox{if } t \in [T_1,T_{2}), \\
\, \vdots \\
B_t - \frac{T_n-t}{T_n-T_{n-1}} B_{T_{n-1}} - \frac{t-T_{n-1}}{T_n-T_{n-1}} B_{T_n}, & \mbox{if } t \in [T_{n-1},T_{n}].
\end{array}
\right.
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=0.7\textwidth]{BBAP.png}
\caption{Paths simulation of a stitched Brownian AP with $n=5$, using the equidistant partition $\{T_i = 2i \| i=0,1,\ldots, 5\}$.}
\end{figure}
Another way of generalizing the anticipative Brownian bridge to obtain an AP driven by Brownian motion is by using Lagrange's polynomial interpolation, that is
\begin{equation}
f_i(t)= \prod\limits_{k=0,k\neq i}^n\frac{ T_k-t}{ T_k-T_i} \quad \text{for } i=0,\ldots, n.
\end{equation}
We may call the resulting AP the Lagrange-Brownian AP, which has the form
\begin{equation}
A_t^{(n)}=B_t - \sum_{i=0}^n \prod\limits_{k=0,k\neq i}^n\frac{ T_k-t}{ T_k-T_i} B_{T_i}.
\end{equation}
More generally, the Lagrange AP driven by a stochastic process $(D_t)$ has the form
\begin{equation}
A_t^{(n)}=D_t - \sum_{i=0}^n \prod\limits_{k=0,k\neq i}^n\frac{ T_k-t}{ T_k-T_i} D_{T_i}.
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=0.73\textwidth]{LagrangeAP.png}
\caption{Paths simulation of a Lagrange-Brownian AP with $n=5$, using the equidistant partition $\{T_i = 2i \| i=0,1,\ldots, 5\}$.}
\end{figure}
The Lagrange APs inherit Runge's phenomenon from their interpolation coefficients: when $n$ is big, the AP oscillates around the edges of the interval.
\begin{figure}[H]
\centering
\includegraphics[width=0.73\textwidth]{LagrangeAPbign.png}
\caption{Paths simulation of a Lagrange-Brownian AP with $n=10$, using the equidistant partition $\{T_i = i \| i=0,1,\ldots, 10\}$, illustrating Runge's phenomenon.}
\end{figure}
One can control this effect by applying a transformation to the interpolating coefficients. For instance, the map $x\mapsto |x|^{2(1-|x|)}$ applied to each interpolating coefficients $$ f_i(t)= \prod\limits_{k=0,k\neq i}^n\frac{ T_k-t}{ T_k-T_i}$$ for $i=0,\ldots, n$, yields another set of interpolating coefficients $\tilde f_0, \ldots, \tilde f_n$ which do no suffer from Runge's phenomenon.
\begin{figure}[H]
\centering
\includegraphics[width=0.75\textwidth]{RegLagrangeAPbign.png}
\caption{Paths simulation of a Brownian AP with $n=10$ and interpolating coefficients $\tilde f_0, \ldots, \tilde f_{10}$, using the equidistant partition $\{T_i = i \| i=0,1,\ldots, 10\}$.}
\end{figure}
\end{ex}
\begin{ex}
Elliptic APs have interpolating coefficients given by
\begin{equation}
f_0(t)= \sqrt{1 - \left( \frac{t-T_0}{T_1-T_0} \right)^2} \mathds{1}_{[T_0, T_1]} (t), \quad f_n(t)= \sqrt{1 - \left( \frac{t-T_{n}}{T_n-T_{n-1}} \right)^2} \mathds{1}_{(T_{n-1}, T_n]} (t),
\end{equation}
\begin{equation}
f_i(t)= \sqrt{1 - \left( \frac{t-T_{i}}{T_i-T_{i-1}} \right)^2} \mathds{1}_{(T_{i-1}, T_{i}]} (t) + \sqrt{1 - \left( \frac{t-T_{i}}{T_{i+1}-T_{i}} \right)^2} \mathds{1}_{(T_{i}, T_{i+1}]} (t) \quad \text{for } i=1,\ldots, n-1.
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=0.75\textwidth]{EllipticAP.png}
\caption{Paths simulation of an elliptic-Brownian AP with $n=5$, using the equidistant partition $\{T_i = 2i \| i=0,1,\ldots, 5\}$.}
\end{figure}
\end{ex}
The expectation and the covariance of an AP, when they exist, are fully characterised by the driver and the interpolating coefficients.
\begin{prop}
If the driver $(D_t)$ has a mean function $\mu_D$, a variance function $\sigma_D^2$, and an covariance function $K_D$, we have:
\begin{align}
\mu_A(t):=\mathbb{E} [A_{t}^{(n)}]&= \mu_D(t) - \sum\limits_{i=0}^n f_i(t) \mu_D(T_i), \label{expect} \\
\sigma_A^2(t):=\mathrm{Var}{[A_{t}^{(n)}]} &= \sigma_D^2(t) + \sum\limits_{i=0}^n f_i^2(t) \sigma_D^2(T_i)-2 f_i(t) K_D(t,T_i) \nonumber \\
&\hspace{1.55cm}+ 2 \sum\limits_{i=0}^n \sum\limits_{j=i+1}^{n} f_i(t) f_j(t) K_D(T_i,T_j), \label{variance} \\
K_A(s,t):=\mathrm{Cov}[{A_{s}^{(n)}, A_{t}^{(n)}}] &= K_D(s,t) - \sum\limits_{i=0}^n f_i(t)K_D(s,T_i) +f_i(s)K_D(t,T_i) \nonumber \\
&\hspace{1.55cm}+ \sum\limits_{i=0}^n\sum\limits_{j=0}^{n} f_i(s)f_j(t) K_D(T_i, T_j). \label{cov}
\end{align}
\end{prop}
\begin{proof}
Eq. (\ref{expect}) follows from the linearity of the expectation and Eq. (\ref{variance}) follows from Eq. (\ref{cov}). Hence, it is enough to show Eq. (\ref{cov}).
\begin{align}
\mathrm{Cov}[{A_{s}^{(n)}, A_{t}^{(n)}}] &= \mathrm{Cov}\left[ D_{s}-\sum_{i=0}^n f_i(s)D_{T_i}, D_{t}-\sum_{i=0}^n f_i(t)D_{T_i} \right] \\
&= K_D(s,t) - \mathrm{Cov}\left[ D_{s}, \sum_{i=0}^n f_i(t)D_{T_i} \right] - \mathrm{Cov}\left[ \sum_{i=0}^n f_i(s)D_{T_i}, D_{t} \right] \nonumber \\
&\hspace{1.55cm} + \mathrm{Cov}\left[\sum_{i=0}^n f_i(s)D_{T_i}, \sum_{i=0}^n f_i(t)D_{T_i} \right] \\
&= K_D(s,t) - \sum_{i=0}^n f_i(t) K_D(s,T_i) - \sum_{i=0}^n f_i(s)K_D(T_i,t) \nonumber \\
& \hspace{1.55cm}+ \sum_{i=0}^n\sum_{j=0}^n f_i(s)f_j(t) K_D(T_i,T_j) \\
&=K_D(s,t) - \sum\limits_{i=0}^n f_i(t)K_D(s,T_i) +f_i(s)K_D(t,T_i) \nonumber \\
&\hspace{1.55cm}+ \sum\limits_{i=0}^n\sum\limits_{j=0}^{n} f_i(s)f_j(t) K_D(T_i, T_j).
\end{align}
\end{proof}
Markov APs will play a crucial role in the construction of FAMs. An immediate property of these processes is the following:
\begin{prop} \label{ind}
Let $(A_t^{(n)})_{t \in [T_0,T_n]}$ be an AP on $\{T_0,T_n\}_*$ that is Markov with respect to $(\mathcal{F}^A_t)_{t \in [T_0,T_n]}$. If $t>T_i$ for $T_i \in \{T_0,T_n\}_*$, then $A_t^{(n)} \perp\!\!\!\perp \mathcal{F}^A_{T_{i}}$.
\end{prop}
In particular, if $(A_t^{(n)})$ is Markov and has a covariance function, then $\mathrm{Cov} (A_s^{(n)},A_t^{(n)}) = 0 $ whenever $s\in[T_i,T_{i+1}]$, $t\in[T_j,T_{j+1}]$ and $j \neq i$.
\bigskip
If we think about APs as noise processes, the most natural subclass to study is the Gaussian subclass. But instead of looking at all APs that are also Gaussian processes, we restrict to APs driven by Gaussian processes.
\begin{defn}
An AP $(A_t^{(n)})_{t \in [T_0,T_n]}$ is said to be a Gaussian AP if its driver $(D_t)$ is a Gaussian stochastic process.
\end{defn}
Not all APs that are Gaussian processes are Gaussian APs, though all Gaussian APs are Gaussian processes. For example, take $D_t=B_t + tY$, where $(B_t)$ is a Gaussian process and $Y$ a non-Gaussian random variable, and $A_t=D_t - (1-t) D_{0} - t D_{1}$. The driver is not Gaussian while the AP is a Gaussian process.
\bigskip
We give a first result about the Markov property of Gaussian APs.
\begin{thm} \label{thmmarkov}
Let $(A_t^{(n)})_{t \in [T_0,T_n]}$ be a Gaussian AP on $\{T_0,T_n\}_*$ with covariance function $K_A$. Then $(A_t^{(n)})$ is Markov with respect to its own filtration if and only if $ \forall \, (r,s,t) \in (T_0,T_n)_*^3$ such that $r \leqslant s < t$, there exists $a(s,t) \in \mathbb{R}_0$ such that
\begin{equation} \label{form1}
K_A(r,t)=\left\{
\begin{array}{cc}
0, &\mbox{ if } t\in [T_i,T_{i+1}], r\in [T_j,T_{j+1}], i\neq j, \\ \\
K_A(r,s)a(s,t), &\mbox{ otherwise}.
\end{array}
\right.
\end{equation}
\end{thm}
\begin{proof}
Suppose $K_A$ is of the form \ref{form1}, we shall show that $(A_t^{(n)})$ is Markov. Let $k>1$ and $(s_1 , s_2 , \ldots , s_k , t ) \in [T_0,T_n] ^{k+1}$ such that $s_1 < s_2 < \ldots < s_k < t$. Then, $(A_t^{(n)})$ is Markov if and only if
\begin{equation}
\mathbb{P}\left[A_{t}^{(n)} \in \cdot \mid A_{s_1}^{(n)}, \ldots, A_{s_{k}}^{(n)}\right] = \mathbb{P}\left[A_{t}^{(n)} \in \cdot \mid A_{s_k}^{(n)} \right].
\end{equation}
Since $K_A(s,t)=0$ unless $s$ and $t$ are in the same sub-interval, we can assume without loss of generality that $(s_1 , s_2 , \ldots , s_k , t ) \in (T_m,T_{m+1})^{k+1}$ for a fixed $m \in \{0,\ldots,n-1\}$. Define
\begin{equation}
\Delta_q=\sum_{i=1}^{k} c_{i,q} A_{s_i}^{(n)}, \quad q=1,\ldots,k-1,
\end{equation}
where the coefficients $(c_{i,q})$ are chosen such that
\begin{equation}
\sum\limits_{i=1}^{k} c_{i,q} K_A(s_i,t)=0 \quad \mbox{ and } \quad \mathbb \det
\begin{pmatrix}
c_{1,1} & \ldots & c_{1,k-1} \\
\vdots & & \vdots \\
c_{k-1,1} & \ldots & c_{k-1,k-1} \\
\end{pmatrix} \neq 0.
\end{equation}
We notice that
\begin{equation}
\sigma (A_{s_1}^{(n)}, \ldots, A_{s_{k}}^{(n)}) = \sigma (\Delta_1, \ldots, \Delta_{k-1}, A_{s_k}^{(n)}) \iff \mathbb \det
\begin{pmatrix}
c_{1,1} & \ldots & c_{1,k-1} \\
\vdots & & \vdots \\
c_{k-1,1} & \ldots & c_{k-1,k-1} \\
\end{pmatrix} \neq 0.
\end{equation}
It remains to be shown that
\begin{equation}
A_{t}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1}) \text{ and } A_{s_k}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1}).
\end{equation}
Equivalently, because we are treating the Gaussian case, we shall show that
\begin{equation}
\mathrm{Cov}[A_{t}^{(n)},\Delta_q]=0 \quad \mbox{ and } \quad \mathrm{Cov}[A_{s_k}^{(n)},\Delta_q]=0 \quad \forall q=1,\ldots,k-1.
\end{equation}
Expanding these covariances, we get
\begin{equation}
\mathrm{Cov}[A_{t}^{(n)},\Delta_q]=
\sum_{i=1}^{k} c_{i,q} K_A(s_i,t)=0,
\end{equation}
which is guaranteed by the choice of the coefficients $(c_{i,q})$, and
\begin{equation}
\mathrm{Cov}[A_{s_k}^{(n)},\Delta_q]=
\sum_{i=1}^{k} c_{i,q} K_A(s_k,s_i)=0.
\end{equation}
The first equation implies the second because $K_A(s_i,t) = 0 \implies K_A(s_k,s_i) =0$ by \ref{form1}, and, when $K_A(s_i,t) \neq 0$, we have $K_A(s_i,t) = a(s_k,t)K_A(s_k,s_i)$, which means
\begin{equation}
\sum\limits_{i=1}^{k} c_{i,q} K_A(s_i,t)=0 \implies \sum\limits_{i=1}^{k} c_{i,q} a(s_k,t)K_A(s_k,s_i)=0 \implies \sum\limits_{i=1}^{k} c_{i,q} K_A(s_k,s_i)=0.
\end{equation}
This concludes one implication.
\bigskip
For the converse, suppose without lose of generality that the driver has mean $0$. We observe that $K_A(x,y)=0$ if $x\in [T_i,T_{i+1}], y\in [T_j,T_{j+1}], i\neq j$. Let $(r,s,t) \in (T_i,T_{i+1})^3$ such that $r\leqslant s <t$. Since $(A_{t}^{(n)})$ is Gaussian, we have
\begin{equation}
\mathbb{E}[A_{t}^{(n)} \| A_{s}^{(n)} ] = \frac{K_A(s,t)}{K_A(s,s)}A_{s}^{(n)}.
\end{equation}
Using the Markov property of $(A_t^{(n)})$, we get
\begin{equation} \label{blabla}
K_A(r,t) = \mathbb{E} [ A_{r}^{(n)}A_{t}^{(n)} ] = \mathbb{E} [ \mathbb{E} [ A_{r}^{(n)}A_{t}^{(n)} \| A_{s}^{(n)} ] ] = \mathbb{E}[\mathbb{E} [A_{r}^{(n)} \| A_{s}^{(n)} ] \mathbb{E} [ A_{t}^{(n)} \| A_{s}^{(n)} ] ],
\end{equation}
which implies
\begin{equation}
K_A(r,t) = \mathbb{E} \left[\frac{K_A(s,t)}{K_A(s,s)}A_{s}^{(n)} \frac{K_A(r,s)}{K_A(s,s)}A_{s}^{(n)} \right] = \frac{K_A(s,t)K_A(r,s)}{K_A(s,s)}.
\end{equation}
Hence $\displaystyle{a(s,t) = \frac{K_A(s,t)}{K_A(s,s)}}$.
\end{proof}
We can simplify the statement of Theorem \ref{thmmarkov} in the following way.
\begin{coro} \label{coromarkov}
A Gaussian AP $(A_t^{(n)})_{t \in [T_0,T_n]}$ is Markov with respect to its own filtration if and only if there exist real functions $A_1 : [T_0,T_n] \rightarrow \mathbb{R} $, and $A_2 : [T_0,T_n] \rightarrow \mathbb{R} $, such that
\begin{equation}
K_A(s,t)=\sum_{i=0}^{n-1} A_1(\min(s,t))A_2(\max(s,t)) \mathds{1}_{(T_i,T_{i+1})}(s,t).
\end{equation}
\end{coro}
\begin{proof}
If
\begin{equation}
K_A(s,t)=\sum_{i=0}^{n-1} A_1(\min(s,t))A_2(\max(s,t)) \mathds{1}_{(T_i,T_{i+1})}(s,t),
\end{equation}
then $\forall \, (r,s,t) \in (T_0,T_n)_*^3$ such that $r \leqslant s < t$,
\begin{equation}
K_A(r,t)=\left\{
\begin{array}{cc}
0, &\mbox{ if } t\in [T_i,T_{i+1}], r\in [T_j,T_{j+1}], i\neq j, \\ \\
K_A(r,s) \frac{A_2(t)}{A_2(s)}, &\mbox{ otherwise}.
\end{array}
\right.
\end{equation}
Hence, $(A_t^{(n)})$ is Markov.
\bigskip
For the converse, suppose that $(A_t^{(n)})$ is Markov. Let $T_m \in \{T_0,T_{n-1}\}_*$. Then,
\begin{equation}
\frac{K_A(r,t)}{K_A\left(r,\frac{T_m + T_{m+1}}{2}\right)} = \frac{K_A\left(\frac{T_m + T_{m+1}}{2},t\right)}{K_A\left(\frac{T_m + T_{m+1}}{2},\frac{T_m + T_{m+1}}{2}\right)}
\end{equation}
for any $(r,t)\in (T_m,T_{m+1})^2$ such that $r<\frac{T_m + T_{m+1}}{2} < t$ by Theorem \ref{thmmarkov}. Hence, if
\begin{equation}
A_1(x) = \sum_{i=0}^{n-1} K_A\left(x, \frac{T_i + T_{i+1}}{2}\right) \mathds{1}_{(T_i,T_{i+1})} (x),
\end{equation}
and
\begin{equation}
A_2(x) = \sum_{i=0}^{n-1} \frac{K_A\left(\frac{T_i + T_{i+1}}{2}, x\right)} {K_A\left( \frac{T_i + T_{i+1}}{2}, \frac{T_i + T_{i+1}}{2}\right )} \mathds{1}_{(T_i,T_{i+1})} (x),
\end{equation}
we have
\begin{equation}
K_A(s,t)=\sum_{i=0}^{n-1} A_1(\min(s,t))A_2(\max(s,t)) \mathds{1}_{(T_i,T_{i+1})}(s,t).
\end{equation}
\end{proof}
If $T_i \in \{T_0,T_{n-1}\}_*$ and $(s,t) \in (T_i,T_{i+1})^2$ such that $s<t$, then $\lim\limits_{s\rightarrow T_i} A_1(s)=0$ or $\lim\limits_{t\rightarrow T_{i+1}} A_2(t)=0$ by continuity of $K_A$, and $\displaystyle \frac{A_1}{A_2}(t)$ is positive and non-decreasing on each interval $(T_i,T_{i+1})$, since $K_A$ is a covariance function.
\bigskip
Starting from a Gauss-Markov driver $(D_t)$, it is always possible to construct a Markovian $(A_t^{(n)})$ by applying the following procedure.
\begin{thm}
For any Gauss-Markov driver $(D_t)$, there exists an AP $(A_t^{(n)})_{t \in [T_0,T_n]}$, driven by $(D_t)$, that is Markovian.
\end{thm}
\begin{proof}
Let $T_m \in \{T_0,T_{n-1}\}_*$ and $(s,t)\in (T_m,T_{m+1})^2$ such that $s<t$. Choose the interpolating coefficients $f_0,\ldots, f_n$ according to
\begin{equation}
\begin{pmatrix}
K_D(T_0,T_0) & \ldots & K_D(T_0,T_n)\\
\vdots & \ddots & \vdots \\
K_D(T_n,T_0) & \ldots & K_D(T_n,T_n)
\end{pmatrix}
\begin{pmatrix}
f_0(\cdot)\\
\vdots \\
f_n(\cdot)
\end{pmatrix} =
\begin{pmatrix}
K_D(\cdot,T_0)\\
\vdots \\
K_D(\cdot,T_n)
\end{pmatrix}.
\end{equation}
Then,
\begin{equation} \label{rel}
\sum_{j=0}^n f_j(\cdot) K_D(T_i,T_j) - K_D(\cdot,T_i)=0, \quad \forall i=0,\ldots, n,
\end{equation}
which implies
\begin{equation}
K_A(x,y)= K_D(x,y) - \sum_{i=0}^n f_i(x) K_D(y,T_i) = K_D(x,y) - \sum_{i=0}^n f_i(y) K_D(x,T_i), \quad \forall (x,y) \in [T_0,T_n]^2.
\end{equation}
Let $T_m \in \{T_0,T_{n-1}\}_*$ and $(s,t)\in (T_m,T_{m+1})^2$ such that $s<t$. Recalling that $(D_t)$ is Gauss-Markov, there exists two functions, $H_1: (T_m,T_{m+1}) \rightarrow \mathbb{R} $ and $H_2 : (T_m,T_{m+1}) \rightarrow \mathbb{R}$, such that $K_D(s,t)=H_1(s)H_2(t)$. Using this fact, we can write
\begin{align}
K_A(s,t)&=H_1(s)H_2(t) - \sum_{i=0}^m f_i(s) H_1(T_i)H_2(t) - \sum_{i=m+1}^n f_i(s) H_1(t)H_2(T_i) \nonumber \\
&=H_2(t)\left(H_1(s)- \sum_{i=0}^m f_i(s) H_1(T_i)\right) - H_1(t) \sum_{i=m+1}^n f_i(s) H_2(T_i)\nonumber \\
&= \left(\sum_{i=m+1}^n f_i(s) H_2(T_i)\right) \left( \lambda H_2(t) - H_1(t) \right)
\end{align}
for some $\lambda \in \mathbb{R}$, where we used Eq. (\ref{rel}) with $i=m$. Hence,
\begin{align}
A_1(x)\mathds{1}_{(T_m,T_{m+1})}(x) &= \left(\sum_{i=m+1}^n f_i(x) H_2(T_i)\right) \mathds{1}_{(T_m,T_{m+1})}(x), \label{A1} \\
A_2(x) \mathds{1}_{(T_m,T_{m+1})}(x) &= \left( \lambda H_2(x) - H_1(x) \right) \mathds{1}_{(T_m,T_{m+1})}(x). \label{A2}
\end{align}
$(A_t^{(n)})$ is Markov by Corollary \ref{coromarkov}.
\end{proof}
Recall that we imposed on all drivers $(D_t)$ of APs the property that $\mathbb{P}[D_t=0]<1$ whenever $t\in (T_0,T_n)_*$. If we extend this property to $[T_0,T_n]$ instead, then the above construction of a Markovian AP becomes explicit.
\begin{coro} \label{coroexplicit}
If $(D_t)$ is a Gauss-Markov process such that $\mathbb{P}[D_t=0]<1$ whenever $t\in [T_0,T_n]$, with $K_D(s,t)=H_1(\min(s,t)) H_2 (\max(s,t))$ for all $(s,t) \in [T_0,T_n]^2$ and for some real functions $H_1$ and $H_2$, then the solution to Eq. (\ref{rel}) is given by
\begin{align}
f_0(x)&= \frac{H_1(T_{1})H_2(x)-H_1(x)H_2(T_1)}{H_1(T_{1})H_2(T_0)-H_1(T_0)H_2(T_1)} \mathds{1}_{[T_0, T_1]} (x),\\
f_i(x)&=\frac{H_1(x)H_2(T_{i-1})-H_1(T_{i-1})H_2(x)}{H_1(T_{i})H_2(T_{i-1})-H_1(T_{i-1})H_2(T_{i})} \mathds{1}_{[T_{i-1}, T_i]} (x) \nonumber \\
& \quad + \frac{H_1(T_{i+1})H_2(x)-H_1(x)H_2(T_{i+1})}{H_1(T_{i+1})H_2(T_{i})-H_1(T_{i})H_2(T_{i+1})} \mathds{1}_{(T_{i}, T_{i+1}]} (x) , \quad \text{for } i=1,\ldots, n-1, \\
f_n(x)&= \frac{H_1(x)H_2(T_{n-1})-H_1(T_{n-1})H_2(x)}{H_1(T_n)H_2(T_{n-1})-H_1(T_{n-1})H_2(T_n)} \mathds{1}_{(T_{n-1}, T_n]} (x).
\end{align}
\end{coro}
\begin{proof}
Let ($T_{m^-},T_m,T_{m^+}) \in \{T_0,T_{n-1}\}_*^3$ such that $T_{m^-}\leqslant T_m<T_{m^+}$, and $x \in (T_m,T_{m+1})$. Then,
\begin{align}
\sum_{j=0}^n f_j(x) K_D(T_{m^-},T_j) &= f_m(x) H_1(T_{m^-}) H_2 (T_m) + f_{m+1}(x) H_1(T_{m^-}) H_2 (T_m) \nonumber \\
&= \frac{H_1(T_{m+1})H_2(x)-H_1(x)H_2(T_{m+1})}{H_1(T_{m+1})H_2(T_{m})-H_1(T_{m})H_2(T_{m+1})} H_1(T_{m^-}) H_2 (T_m) \nonumber \\
&\hspace{1.1cm}+ \frac{H_1(x)H_2(T_{m})-H_1(T_{m})H_2(x)}{H_1(T_{m+1})H_2(T_{m})-H_1(T_{m})H_2(T_{m+1})} H_1(T_{m^-}) H_2 (T_{m+1}) \nonumber \\ \nonumber \\
&= \frac{H_1(x) (H_1(T_{m^-})H_2(T_m) H_2 (T_{m+1}) - H_1(T_{m^-}) H_2 (T_m) H_2(T_{m+1}))}{H_1(T_{m+1})H_2(T_{m})-H_1(T_{m})H_2(T_{m+1})} \nonumber \\
&\hspace{1.1cm}+\frac{H_2(x)(H_1(T_{m+1}) H_1(T_{m^-}) H_2 (T_m) - H_1(T_{m}) H_1(T_{m^-}) H_2 (T_{m+1}) ) }{H_1(T_{m+1})H_2(T_{m})-H_1(T_{m})H_2(T_{m+1})}\nonumber \\ \nonumber \\
&=H_1(T_{m^-}) H_2(x) = K_D(T_{m^-},x)
\end{align}
The same argument applies to show $\displaystyle \sum_{j=0}^n f_j(x) K_D(T_{m^+},T_j) = K_D(x,T_{m^+}) $. Hence,
\begin{equation}
\begin{pmatrix}
K_D(T_0,T_0) & \ldots & K_D(T_0,T_n)\\
\vdots & \ddots & \vdots \\
K_D(T_n,T_0) & \ldots & K_D(T_n,T_n)
\end{pmatrix}
\begin{pmatrix}
f_0(\cdot)\\
\vdots \\
f_n(\cdot)
\end{pmatrix} =
\begin{pmatrix}
K_D(\cdot,T_0)\\
\vdots \\
K_D(\cdot,T_n)
\end{pmatrix}.
\end{equation}
\end{proof}
\begin{rem}
Notice that if $(D_t)$ does not satisfy $\mathbb{P}[D_t=0]<1$ for some $t\in \{T_0,T_n\}_*$, it is still straightforward to construct the above AP by removing all the rows and columns of zeros in the matrix
\begin{equation}
\begin{pmatrix}
K_D(T_0,T_0) & \ldots & K_D(T_0,T_n)\\
\vdots & \ddots & \vdots \\
K_D(T_n,T_0) & \ldots & K_D(T_n,T_n)
\end{pmatrix}.
\end{equation}
Since the driver already matches $0$ at $t \in \{T_0,T_n\}_*$, we will not need to find the corresponding interpolating coefficients because they will not appear in the AP expression. This is illustrated by taking the Brownian motion as a driver with $T_0=0$. Then $f_0$ does not matter since it is multiplied by $0$ in the AP expression.
\end{rem}
\begin{rem}
For the choice of coefficients in Corollary \ref{coroexplicit}, we can simplify Eqs. \ref{A1} and \ref{A2}:
\begin{align}
A_1(x)\mathds{1}_{(T_m,T_{m+1})}(x) &= f_{m+1}(x) H_2(T_{m+1}) \mathds{1}_{(T_m,T_{m+1})}(x),\\
A_2(x) \mathds{1}_{(T_m,T_{m+1})}(x) &= \left( \frac{H_1(T_{m+1})}{H_2(T_{m+1})} H_2(x) - H_1(x) \right) \mathds{1}_{(T_m,T_{m+1})}(x).
\end{align}
\end{rem}
This method of producing Markovian APs is not unique, but certainly feels natural. The resulting APs are called standard.
\begin{defn} \label{standard}
A standard AP $(A_t^{(n)})_{t \in [T_0,T_n]}$ is an AP driven by of a Gauss-Markov process $(D_t)$, with $K_D(s,t)=H_1(\min(s,t)) H_2 (\max(s,t)) \, \forall s,t \in [T_0,T_n]$ for some real functions $H_1$ and $H_2$, of the form
\begin{equation}
A_{t}^{(n)} = \left\{
\begin{array}{ll}
D_t - \frac{H_1(T_{1})H_2(t)-H_1(t)H_2(T_1)}{H_1(T_{1})H_2(T_0)-H_1(T_0)H_2(T_1)} D_{T_0} - \frac{H_1(t)H_2(T_{0})-H_1(T_0)H_2(t)}{H_1(T_{1})H_2(T_{0})-H_1(T_0)H_2(T_{1})} D_{T_1} & \mbox{if } t \in [T_0,T_{1}), \\ \\
D_t - \frac{H_1(T_2)H_2(t)-H_1(t)H_2(T_2)}{H_1(T_2)H_2(T_1)-H_1(T_1)H_2(T_2)} D_{T_1} - \frac{H_1(t)H_2(T_{1})-H_1(T_{1})H_2(t)}{H_1(T_2)H_2(T_{1})-H_1(T_{1})H_2(T_{2})} D_{T_2} & \mbox{if } t \in [T_1,T_{2}), \\
\, \vdots \\
D_t - \frac{H_1(T_n)H_2(t)-H_1(t)H_2(T_n)}{H_1(T_n)H_2(T_{n-1})-H_1(T_{n-1})H_2(T_n)} D_{T_{n-1}} & \\ \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad -\frac{H_1(t)H_2(T_{n-1})-H_1(T_{n-1})H_2(t)}{H_1(T_n)H_2(T_{n-1})-H_1(T_{n-1})H_2(T_{n})} D_{T_n} & \mbox{if } t \in [T_{n-1},T_{n}].
\end{array}
\right.
\end{equation}
\end{defn}
\begin{ex}
If $(D_t)$ is an Ornstein-Uhlenbeck process with parameters $\theta>0, \sigma>0, \mu \in \mathbb{R}$ and starting value $d_0\in \mathbb{R}$, that is, the solution to
\begin{equation}
\, \mathrm{d} D_{t}=\theta\left(\mu-D_{t}\right) \, \mathrm{d} t+\sigma \, \mathrm{d} W_{t}, \quad D_0=d_0,
\end{equation}
then $\displaystyle K_D(s,t)=\frac{\sigma^2}{2 \theta} \mathrm{e}^{\theta \min(s,t)} \mathrm{e}^{-\theta \max(s,t)}$. The standard AP driven by $(D_t)$ is
\begin{equation}
A_{t}^{(n)} = \left\{
\begin{array}{ll}
D_t - \frac{\mathrm{e}^{\theta (T_1-t)}-\mathrm{e}^{-\theta (T_1-t)}}{\mathrm{e}^{\theta (T_1-T_0)}-\mathrm{e}^{-\theta (T_1-T_0)}} D_{T_0} - \frac{\mathrm{e}^{\theta (t-T_0)}-\mathrm{e}^{-\theta (t-T_0)}}{\mathrm{e}^{\theta (T_1-T_0)}-\mathrm{e}^{-\theta (T_1-T_0)}} D_{T_1} & \mbox{if } t \in [T_0,T_{1}), \\ \\
D_t -\frac{\mathrm{e}^{\theta (T_2-t)}-\mathrm{e}^{-\theta (T_2-t)}}{\mathrm{e}^{\theta (T_2-T_1)}-\mathrm{e}^{-\theta (T_2-T_1)}} D_{T_1} - \frac{\mathrm{e}^{\theta (t-T_1)}-\mathrm{e}^{-\theta (t-T_1)}}{\mathrm{e}^{\theta (T_2-T_1)}-\mathrm{e}^{-\theta (T_2-T_1)}} D_{T_2} & \mbox{if } t \in [T_1,T_{2}), \\
\, \vdots \\
D_t - \frac{\mathrm{e}^{\theta (T_n-t)}-\mathrm{e}^{-\theta (T_n-t)}}{\mathrm{e}^{\theta (T_n-T_{n-1})}-\mathrm{e}^{-\theta (T_n-T_{n-1})}} D_{T_{n-1}} -\frac{\mathrm{e}^{\theta (t-T_{n-1})}-\mathrm{e}^{-\theta (t-T_{n-1})}}{\mathrm{e}^{\theta (T_n-T_{n-1})}-\mathrm{e}^{-\theta (T_n-T_{n-1})}} D_{T_n} & \mbox{if } t \in [T_{n-1},T_{n}].
\end{array}
\right.
\end{equation}
\end{ex}
There are infinitely many Markovian APs driven by the same Gauss-Markov driver. In general, when $T_m \in \{T_0,T_{n-1}\}_*$ and $(s,t)\in (T_m,T_{m+1})^2$ with $s<t$, we have
\begin{align}
K_A(s,t)=&\left(H_1(s)-\sum_{i=0}^m f_i(s) H_1(T_i) \right)\left(H_2(t)-\sum_{i=m+1}^n f_i(t) H_2(T_i) \right) \nonumber \\
&- \left(\sum_{i=m+1}^{n} f_i(s) H_2(T_i) \right) \left(H_1(t)-\sum_{i=0}^{m+1} f_i(t) H_1(T_i) \right)\nonumber \\
&+ \left(\sum_{i=m+1}^{n} f_i(s) H_1(T_i) \right)\sum_{i=m+2}^{n} f_i(t) H_2(T_i) \nonumber \\
&- \left(H_2(s)-\sum_{i=0}^m f_i(s) H_2(T_i) \right)\sum_{i=0}^{m} f_i(t) H_1(T_i) , \label{KA}\
\end{align}
where we use the convention that an empty sum is equal to zero. There are as many Markovian APs driven by $(D_t)$ as there are ways to separate the variable of the above expression of $K_A$.
\begin{ex} \label{nonstarcade}
If $(D_t)$ is a standard Brownian motion, applying Eq. (\ref{rel}) to find appropriate interpolating coefficients yields the stitched Brownian AP. But there are other Markovian APs driven by standard Brownian motion. For instance, in the two-period case, we may choose
\begin{align}
f_0(t)&= \frac{T_1-t}{T_1-T_0} \mathds{1}_{\left[T_0,\frac{T_1+T_2}{2}\right]}(t) - \frac{T_2-t}{T_1-T_0} \mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(t), \label{f0} \\
f_1(t)&= \frac{t-T_0}{T_1-T_0} \mathds{1}_{[T_0,T_1]}(t) + \frac{T_2-t}{T_2-T_1} \mathds{1}_{(T_1,T_2]}(t), \\
f_2(t)&= \frac{t-T_1}{T_2-T_1} \mathds{1}_{\left[T_1,T_2\right]}(t).
\end{align}
It is straightforward to verify that these are interpolating coefficients. Let $(A_t^{(2)})_{t \in [T_0,T_2]}$ be the AP with these interpolating coefficients driven by a standard Brownian motion. Then,
\begin{align}
&K_A(s,t)\nonumber \\
&= \frac{(\min(s,t) -T_0)(T_1-\max(s,t))}{T_1-T_0}\mathds{1}_{(T_0,T_1)}(s,t) \nonumber \\ \nonumber \\
&+ (\min(s,t)-T_1)\left(\frac{T_0 (\max(s,t) + T_0) - 3 T_0 T_1 + T_1^2}{(T_1-T_0)^2} + \frac{T_1-\max(s,t)}{T_2-T_1} \right)\mathds{1}_{\left(T_1,\frac{T_1+T_2}{2}\right]}(s,t)\nonumber \\ \nonumber \\
&+ \frac{(\min(s,t) -T_1)(T_2-\max(s,t))}{T_2-T_1} \frac{T_0^2 + T_1^2 + T_0 (T_2 -3 T_1)}{(T_1-T_0)^2}\nonumber \\ &\hspace{8.5cm}\times\mathds{1}_{\left(T_1,\frac{T_1+T_2}{2}\right]}(\min(s,t)) \mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(\max(s,t))\nonumber \\ \nonumber \\
&+ \frac{\min(s,t) (T_0^2 + T_1^2 - T_0 (T_1 + T_2)) - (T_1 - T_0)^2 T_1 + T_0 T_2 (T_2 - T_1)}{(T_1-T_0)^2} \frac{T_2-\max(s,t)}{T_2-T_1} \nonumber \\
&\hspace{8.5cm}\times\mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(s,t).
\end{align}
Thus,
\begin{align}
A_1(x)&= (x -T_0) \mathds{1}_{\left[T_0,T_1\right]}(x) + (x -T_1) \mathds{1}_{\left(T_1,\frac{T_1+T_2}{2}\right]}(x) \\
&+ \frac{x (T_0^2 + T_1^2 - T_0 (T_1 + T_2)) - (T_1 - T_0)^2 T_1 + T_0 T_2 (T_2 - T_1)}{T_0^2 + T_1^2 + T_0 (T_2 -3 T_1)}\mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(x),\nonumber
\end{align}
\begin{equation}
A_2(x)= \frac{T_1-x}{T_1-T_0}\mathds{1}_{[T_0,T_1]}(x) + \left(\frac{T_0 (x + T_0) - 3 T_0 T_1 + T_1^2}{(T_1-T_0)^2} + \frac{T_1-x}{T_2-T_1} \right)\mathds{1}_{\left(T_1,\frac{T_1+T_2}{2}\right]}(x)
\end{equation}
$$+\frac{(T_2-x)}{T_2-T_1} \frac{T_0^2 + T_1^2 + T_0 (T_2 -3 T_1)}{(T_1-T_0)^2}\mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(x). $$
Hence, recalling Theorem \ref{thmsemimarkov}, this AP is Markovian. This AP is a slight modification of the stitched Brownian AP, where $f_0$ is not $0$ on $(T_1,T_2)$. Hence, $B_{T_0}$ still has an influence (a negative one since $f_0$ is negative on $(T_1,T_2)$) on the paths of the AP on $(T_1,T_2)$. Similarly, we can modify the stitched Brownian arcade by making $f_2$ not $0$ on $(T_0,T_1)$:
\begin{align}
f_0(t)&= \frac{T_1-t}{T_1-T_0} \mathds{1}_{[T_0,T_1]}(t),\\
f_1(t)&= \frac{t-T_0}{T_1-T_0} \mathds{1}_{[T_0,T_1]}(t) + \frac{T_2-t}{T_2-T_1} \mathds{1}_{(T_1,T_2]}(t),\\
f_2(t)&=\frac{T_0-t}{T_2-T_1}\mathds{1}_{\left[T_0,\frac{T_0+T_1}{2}\right]}(t) + \frac{t-T_1}{T_2-T_1} \mathds{1}_{\left(\frac{T_0+T_1}{2},T_2\right]}(t). \label{f2}
\end{align}
For this choice of interpolating coefficients, $B_{T_2}$ has an influence on the paths of the AP on $(T_0,T_1)$. Combining the interpolating coefficients $f_0$ from Eq. (\ref{f0}) and $f_2$ from Eq. (\ref{f2}), we can find an interpolating coefficient $f_1$, such that a Brownian AP with these interpolating coefficients is Markovian:
\begin{align}
f_0(t)&= \frac{T_1-t}{T_1-T_0} \mathds{1}_{\left[T_0,\frac{T_1+T_2}{2}\right]}(t) + \frac{t-T_2}{T_1-T_0} \mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(t),\\ \nonumber \\
f_1(t) &= \frac{(t-T_0)(T_2-T_0)}{(T_2-T_1)(T_1-T_0)}\mathds{1}_{\left[T_0,\frac{T_0+T_1}{2}\right]}(t) + \frac{T_1^2-T_0T_2 + t(T_0-2T_1+T_2)}{(T_1-T_0)(T_2-T_1)} \mathds{1}_{\left(\frac{T_0+T_1}{2},T_1\right]}(t) \nonumber \\ \nonumber \\
&+ \frac{T_1 (T_0 (T_1 - 2 T_2) + T_1 T_2) + t(T_0T_2-T_1^2)}{(T_2-T_1)(T_1-T_0)T_1} \mathds{1}_{\left(T_1,\frac{T_1+T_2}{2}\right]}(t)\nonumber \\ \nonumber \\
&+\frac{(T_2-t)(T_1^2+T_0(T_2-2T_1))}{(T_2-T_1)(T_1-T_0)T_1}\mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(t), \\ \nonumber \\
f_2(t)&=\frac{T_0-t}{T_2-T_1} \mathds{1}_{\left[T_0,\frac{T_0+T_1}{2}\right]}(t)+ \frac{t-T_1}{T_2-T_1} \mathds{1}_{\left(\frac{T_0+T_1}{2},T_2\right]}(t).
\end{align}
The key to building non-standard APs is to break each sub-interval into several pieces, and to define the interpolating coefficients by parts on these pieces while making sure that they remain continuous, and that the expression of $K_A$ in Eq. (\ref{KA}) has separable variables.
\end{ex}
\section{Randomized arcade processes}
We extend the idea of arcade processes to interpolate between the components of the random vector $X$ instead of interpolating between zeros. Two sets $\{f_0,\ldots,f_n\}$ and $\{g_0,\ldots,g_n\}$ of interpolating coefficients (see Definition \ref{interpolating}) are needed to ensure the matching of the target random variables.
We recall that the $\mathbb{R}^{n+1}$-valued random vector $X=(X_0,\ldots, X_n)$ is independent of the stochastic driver $(D_t)$, while the random variables $X_0,\ldots, X_n$ may be mutually dependent.
\begin{defn} \label{RAP}
An $X$-randomized arcade process (X-RAP) $(I_t^{(n)})_{t \in [T_0,T_n]}$ on the partition $\{T_0,T_n\}_*$ is a stochastic process of the form
\begin{equation}
I_t^{(n)}:=S_t^{(n)}+ A_t^{(n)} = D_{t}-\sum_{i=0}^n \left(f_i(t)D_{T_i}-g_i(t)X_{i}\right),
\end{equation}
where $f_0,\ldots, f_n$ and $g_0,\ldots, g_n$ are interpolating coefficients on $\{T_0,T_n\}_*$. We refer to
\begin{equation}
S_t^{(n)}=\sum\limits_{i=0}^n g_i(t)X_{i}
\end{equation}
as the signal function of $I_t^{(n)}$ and to
\begin{equation}
A_t^{(n)}=D_{t}-\sum\limits_{i=0}^n f_i(t)D_{T_i}
\end{equation}
as the noise process of $I_t^{(n)}$. We denote by $(\mathcal{F}^I_t)_{t \in [T_0,T_n]}$ the filtration generated by $(I_t^{(n)})$.
\end{defn}
We notice that $I_{T_0}^{(n)}= X_{0}, \ldots, I_{T_n}^{(n)}= X_{n}$, so $(I_t^{(n)})$ is a stochastic interpolator between the random variables $X_{0},\ldots, X_{n}$. We have that $(S_t^{(n)}) \perp\!\!\!\perp (A_t^{(n)})$ since $X \perp\!\!\!\perp (D_t)$.
\begin{rem}
A related but distinct class of processes, introduced in \cite{Menguturk} and called random $n$-bridges, shares some characteristics with RAPs. These processes, defined weakly, match given probability measures instead of random variables. They are constructed in a similar fashion as stochastic bridges, that is, by conditioning a stochastic process to match given distributions at given times. In special cases, the law of a RAP satisfies the conditions for the RAP to be a random $n$-bridge. For instance, the RAP obtained by randomizing the stitched Brownian bridge, using the same interpolating coefficients for the signal function as the ones used in the noise process, has a law that satisfies the conditions for the RAP to be a random $n$-bridge. However, any other RAP driven by Brownian motion is not a random $n$-bridge. Conversely, certain random $n$-bridges cannot have the same law as a RAP. In this paper, we chose not to consider the case in which the driver underlying a RAP jumps. But if we did, just for the sake of comparison, a random $n$-bridge built using a gamma process will not match the law of any RAP driven by a gamma process, since APs are sums, not a products.
\end{rem}
The paths of an $X$-RAP depend on the coupling of the random vector $X$, so not only on its marginal distributions. This property is illustrated in the following example.
\begin{ex}
Let $X=(X_0,\ldots,X_5)$ be a vector of independent uniform $\mathcal U\,(\{-1,1\})$ random variables, and $Y=(Y_0,\ldots,Y_5)$ be another vector of random variables such that $Y_0 \sim \mathcal U\,(\{-1,1\})$, $Y_i=-Y_{i-1}$ for $i=1,\ldots, 5$. Let $(A_t^{(5)})$ be an AP with elliptic interpolation coefficients driven by Brownian motion multiplied by 0.2, $g_i=f_i$ for $i=0,\ldots, 5$, and $(I_t^{(5)})$, $(\tilde I_t^{(5)})$ its associated $X$-RAP and $Y$-RAP, respectively. Although we are using the same driver and interpolating coefficients for both RAPs, and that the vectors $X$ and $Y$ have the same marginal distributions, the paths of $(I_t^{(5)})$ and $(\tilde I_t^{(5)})$, shown below, are different.
\begin{figure}[H]
\centering
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{EllipticRAP.png}
\caption{Paths simulation of $(I_t^{(5)})$ on $[0,10]$ using the equidistant partition $\{T_i = 2i \| i=0,1,\ldots, 5\}$.}
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{EllipticRAP2.png}
\caption{Paths simulation of $(\tilde I_t^{(5)})$ on $[0,10]$ using the equidistant partition $\{T_i = 2i \| i=0,1,\ldots, 5\}$.}
\end{minipage}
\end{figure}
\end{ex}
Besides its main purpose of interpolating in the strong sense, a RAP can also be used to mimic another stochastic process. Let $(Y_t)_{t \in [T_0,T_n]}$ be a sample-continuous stochastic process. For instance, if $\{T_0,T_n\}_*$ is the equidistant partition of an interval $[a,b]$, $X=(Y_{T_0}, \ldots, Y_{T_n})$, and $\{f_0,\ldots,f_n\}=\{g_0,\ldots,g_n\}$ are the piecewise linear interpolating coefficients, then for nearly all $\omega \in \Omega$, $\sup\limits_{t \in [a,b]} A_t^{(n)}(\omega) \rightarrow 0$, and $\sup\limits_{t \in [a,b]} S_t^{(n)}(\omega) \rightarrow Y_t(\omega)$ as $n \rightarrow \infty$. Hence, in that case,
\begin{equation}
\sup\limits_{t \in [a,b]} \abs{I_t^{(n)}-Y_t}=\sup\limits_{t \in [a,b]} \abs{S_t^{(n)}+ A_t^{(n)}-Y_t} \rightarrow 0,
\end{equation}
with probability one.
\begin{ex}
Let $[a,b]=[0,10]$, and $\{T_0,T_n\}_*$ its equidistant partition. If $(Y_t)$ is a fractional Brownian motion and $X=(Y_{T_0}, \ldots, Y_{T_n})$, then the paths of an $X$-randomized stitched Brownian arcade will be similar to the one of $(Y_t)$ when $n$ is large enough.
\begin{figure}[H]
\centering
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{FBM.png}
\caption{Path of the fractional Brownian motion $(Y_t)$.}
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{RAPapproxFBM.png}
\caption{Path of an $X$-RAP mimicking the same fractional Brownian motion $(Y_t)$.}
\end{minipage}
\end{figure}
\end{ex}
\begin{prop}
Let $(S_t^{(n)})$ and $(A_t^{(n)})$ have mean functions $\mu_S,\mu_A$, variance functions $\sigma_S^2,\sigma_A^2$ and covariance functions $K_S,K_A$, respectively. Then
\begin{equation}
\mu_I(t):=\mathbb{E} [I_{t}^{(n)}]= \mu_S(t) + \mu_A(t),
\end{equation}
\begin{equation}
\sigma_I^2(t):=\mathrm{Var}{[I_{t}^{(n)}]} = \sigma^2_S(t) + \sigma^2_A(t),
\end{equation}
\begin{align}
K_I(s,t):=\mathrm{Cov}[{I_{s}^{(n)}, I_{t}^{(n)}}] &= K_S(s,t) + K_A(s,t) \\
&= \sum\limits_{i=0}^n \sum\limits_{j=0}^n g_i(t)g_j(s)\mathrm{Cov}[X_{i},X_j] + K_A(s,t).
\end{align}
\end{prop}
We introduce terminologies similar to the ones from the previous section.
\begin{defn} \label{standardRAP}
Let $(I_t^{(n)})$ be a RAP.
\begin{enumerate}
\item $(I_t^{(n)})$ is said to be a Gaussian RAP if its stochastic driver $(D_t)$ is a Gaussian process.
\item $(I_t^{(n)})$ is said to be a standard RAP if its noise process $(A_t^{(n)})$ is a standard AP, and if for all $x\in[T_0,T_n]$ and $j=1,\ldots,n$ it holds that $g_j(x) \mathds{1}_{[T_0,T_{j-1}]}(x)=0$, and $g_j(x) \mathds{1}_{[T_{j-1},T_{j}]}(x)= f_j(x)\mathds{1}_{[T_{j-1},T_{j}]}(x)$.
\end{enumerate}
\end{defn}
We give a similar Markovianity result for Gaussian RAPs to the one of AP.
\begin{defn} \label{nearlymarkov}
Let $\mathcal I \subseteq \mathbb{R}^+$ be a real interval, and $ \tau_0<\tau_1 < \ldots < \infty$ such that $\tau=\{\tau_0, \tau_1, \ldots\} \subset \mathcal I$. The set $\tau$ may be finite, i.e., there exists a maximal element $\tau_n \in \tau$, or contain infinitely many elements. A stochastic process $(Y_t)_{t \in \mathcal I}$ is called $\tau$-nearly-Markov if
\begin{equation}
\mathbb{P}\left[Y_t \in \cdot \mid \mathcal{F}^Y_s \right] = \mathbb{P}\left[Y_{t} \in \cdot \mid Y_{\tau_0},\ldots,Y_{\tau_{m(s)}} , Y_s \right]
\end{equation}
for any $(s,t) \in \mathcal I^2$ such that $s \leqslant t$, and $\tau_{m(s)} = \max\limits_{i\in \mathbb{N}} \{\tau_i \| \tau_i \leqslant s\}$.
\end{defn}
\begin{thm} \label{thmsemimarkov}
Let $(I_t^{(n)})_{t \in [T_0,T_n]}=(S_t^{(n)} + A_t^{(n)})_{t \in [T_0,T_n]}$ be a Gaussian $X$-RAP on $\{T_0,T_n\}_*$. Then $(I_t^{(n)})$ is $\{T_0,T_n\}_*$-nearly-Markov if the following conditions are all satisfied:
\begin{enumerate}
\item The AP $(A_t^{(n)})$ is Markov, i.e., $K_A(s,t)=\sum\limits_{i=0}^{n-1} A_1(\min(s,t))A_2(\max(s,t)) \mathds{1}_{(T_i,T_{i+1})}(s,t)$.
\item For all $ j=1,\ldots, n$, and for all $ x \in [T_0,T_n]$,
\begin{equation} \label{subcond1}
g_j(x) \mathds{1}_{[T_0,T_{j-1}]}(x)=0,
\end{equation}
\begin{equation} \label{subcond2}
g_j(x)A_1(T_j) \mathds{1}_{[T_{j-1},T_{j}]}(x)= A_1(x)\mathds{1}_{[T_{j-1},T_{j}]}(x).
\end{equation}
\end{enumerate}
\end{thm}
\begin{proof}
Let $k>1$ and $(s_1 , s_2 , \ldots , s_k , t ) \in [T_0,T_n] ^{k+1}$ such that $s_1 < s_2 < \ldots < s_k < t$. Then, $(I_t^{(n)})$ is $\{T_0,T_n\}_*$-nearly-Markov if and only if
\begin{equation}
\mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m(s_k)}, I_{s_1}^{(n)}, \ldots, I_{s_{k}}^{(n)}\right] = \mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m(s_k)}, I_{s_k}^{(n)} \right].
\end{equation}
where $m(s_k):= \max\limits\{i \in \mathbb{N} \| T_i \leqslant s_k\}$. In the following, we will refer to $m(s_k)$ by $m$ since $s_k$ is fixed.
\bigskip
We first show that $s_1,\ldots, s_k$ can be picked to all be in the sub-interval $(T_m,T_{m+1})$. To see this, assume there is an integer $j \in \{1,\ldots, k\}$ such that $s_j <T_m$ and $T_m<s_{j+1}$. Then
\begin{equation}
\sigma( X_0, \ldots, X_{m}, I_{s_1}^{(n)}, \ldots, I_{s_{k}}^{(n)}) = \sigma( X_0, \ldots, X_{m}, A_{s_1}^{(n)}, \ldots, A_{s_j}^{(n)}, I_{s_{j+1}}^{(n)}, \ldots, I_{s_{k}}^{(n)})
\end{equation}
by Eq. (\ref{subcond1}). We also know that $(A_{s_1}^{(n)}, \ldots, A_{s_j}^{(n)}) \perp\!\!\!\perp (X_0, \ldots, X_{m})$ by the definition of the $X$-RAP, and $(A_{s_1}^{(n)}, \ldots, A_{s_j}^{(n)}) \perp\!\!\!\perp (I_{s_{j+1}}^{(n)}, \ldots, I_{s_{k}}^{(n)}, I_{t}^{(n)} )$ since $(A_t^{(n)})$ is Markov. We conclude that
\begin{equation}
\mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m}, I_{s_1}^{(n)}, \ldots, I_{s_{k}}^{(n)}\right] = \mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m}, I_{s_{j+1}}^{(n)}, \ldots, I_{s_{k}}^{(n)}\right],
\end{equation}
which means we can assume that $s_1,\ldots, s_k$ are all in the same sub-interval $(T_m,T_{m+1})$.
\bigskip
Let us define $a_m(\cdot):= \sum\limits_{i=0}^m g_i(\cdot) X_i$, and
\begin{equation}
\Delta_q:=\sum_{i=1}^{k} c_{i,q} I_{s_i}^{(n)} = \sum_{i=1}^{k} c_{i,q} a_m(s_i) + \sum_{i=1}^{k} c_{i,q} g_{m+1}(s_i) X_{m+1} + \sum_{i=1}^{k} c_{i,q} A_{s_i}^{(n)}, \quad q=1,\ldots,k-1,
\end{equation}
where the coefficients $(c_{i,q})$ are chosen such that
\begin{equation} \label{choice}
\sum\limits_{i=1}^{k} c_{i,q} K_A(s_i,t)=0 \quad \mbox{ and } \quad \mathbb \det
\begin{pmatrix}
c_{1,1} & \ldots & c_{1,k-1} \\
\vdots & & \vdots \\
c_{k-1,1} & \ldots & c_{k-1,k-1} \\
\end{pmatrix} \neq 0.
\end{equation}
This guarantees the following (where the notation "$\| (X_0, \ldots, X_{m})$" means conditional on $(X_0, \ldots, X_{m})$):
\begin{enumerate}
\item $\mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m}, I_{s_1}^{(n)}, \ldots, I_{s_{k}}^{(n)}\right] = \mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m}, \Delta_1, \ldots, \Delta_{k-1},I_{s_k}^{(n)} \right]$.
\item $(\Delta_1, \ldots, \Delta_{k-1}) \| (X_0, \ldots, X_{m})$ is a Gaussian vector. To see this, we observe that $\forall q =1,\ldots k-1$,
\begin{equation}
\sum\limits_{i=1}^{k} c_{i,q} K_A(s_i,t)=0 \implies \sum\limits_{i=1}^{k} c_{i,q} g_{m+1}(s_i)=0,
\end{equation}
where we used Eq. (\ref{subcond2}). Hence, $\Delta_q=\sum\limits_{i=1}^{k} c_{i,q} I_{s_i}^{(n)}=\sum\limits_{i=1}^{k} c_{i,q} a_m(s_i)+ \sum\limits_{i=1}^{k} c_{i,q} A_{s_i}^{(n)}$ for all $q=1,\ldots,k-1$, which implies that $(\Delta_1, \ldots, \Delta_{k-1}) \| (X_0, \ldots, X_{m})$ is a Gaussian vector.
\item $A_{t}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1}) \| (X_0, \ldots, X_{m})$, since $\sum\limits_{i=1}^{k} c_{i,q} K_A(s_i,t)=0$.
\item $A_{s_k}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1}) \| (X_0, \ldots, X_{m})$, since $(A_t^{(n)})$ is Markov.
\end{enumerate}
To conclude, we need to show
\begin{equation} \label{ind2}
I_{t}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1}) \| (X_0, \ldots, X_{m}) \text{ and } I_{s_k}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1}) \| (X_0, \ldots, X_{m}).
\end{equation}
Since $(I_t^{(n)})=(S_t^{(n)}+A_t^{(n)})$, and $(S_t^{(n)}) \perp\!\!\!\perp (A_t^{(n)})$, we have \ref{ind2} if
\begin{equation}
A_{t}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1})\| (X_0, \ldots, X_{m}) \text{ and } A_{s_k}^{(n)} \perp\!\!\!\perp (\Delta_1, \ldots, \Delta_{k-1})\| (X_0, \ldots, X_{m}),
\end{equation}
which is guaranteed by conditions \ref{choice}.
\end{proof}
\begin{rem}
If $(A_t^{(n)})$ is standard (see Definition \ref{standard}), then Eq. (\ref{subcond2}) is equivalent to
\begin{equation}
g_j(x) \mathds{1}_{[T_{j-1},T_{j}]}(x)= f_j(x)\mathds{1}_{[T_{j-1},T_{j}]}(x).
\end{equation}
This makes standard RAPs automatically nearly-Markov.
\end{rem}
\begin{rem}
Depending on the coupling of $X$, the probability $\mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m(s)}, I_{s}^{(n)} \right]$ might simplify further. For instance, if $X$ has continuous marginals and is distributed according to Kantorovich's coupling, then
\begin{equation}
\mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, \ldots, X_{m(s)}, I_{s}^{(n)} \right]= \mathbb{P}\left[I_{t}^{(n)} \in \cdot \mid X_0, I_{s}^{(n)} \right]
\end{equation}
because $X_1,\ldots, X_{m(s)}$ are all deterministic functions of $X_0$.
\end{rem}
\begin{rem}
It is important to notice that the nearly-Markov property is not symmetric in time. Define $\mathcal G_t^I=\sigma ( I_u^{(n)} \| t \leqslant u \leqslant T_n) $. Let $s < t$ in $[T_0,T_n]$. Then, to get
\begin{equation}
\mathbb{P}\left[I_s^{(n)} \in \cdot \mid \mathcal G^I_t \right] = \mathbb{P}\left[I_s^{(n)} \in \cdot \mid I_t^{(n)}, X_{k(t)}, X_{k(t)+1}, \ldots, X_n \right]
\end{equation}
where $k(t)=\min\limits\{i \in \mathbb{N} \| T_i \geqslant t\}$, one needs to replace Cond. \ref{subcond1} by
\begin{equation}
g_j(x) \mathds{1}_{[T_{j+1},T_{n}]}(x)=0.
\end{equation}
\end{rem}
\begin{ex} \label{exf}
We give an example of a non-standard $X$-RAP on $[T_0,T_2]$ that is $\{T_0,T_2\}_*$-nearly-Markov, where $X_0\overset{\mathcal L}{=} X_1 \overset{\mathcal L}{=} X_2 \overset{\mathcal L}{=} \textrm{Uni}\,(\{-1,1\})$ are pairwise independent. Consider the interpolating coefficients
\begin{align}
f_0(t)&= \frac{T_1-t}{T_1-T_0} \mathds{1}_{\left[T_0,\frac{T_1+T_2}{2}\right]}(t) - \frac{T_2-t}{T_1-T_0} \mathds{1}_{\left(\frac{T_1+T_2}{2},T_2\right]}(t), \\
f_1(t)&= \frac{t-T_0}{T_1-T_0} \mathds{1}_{[T_0,T_1]}(t) + \frac{T_2-t}{T_2-T_1} \mathds{1}_{(T_1,T_2]}(t), \\
f_2(t)&= \frac{t-T_1}{T_2-T_1} \mathds{1}_{\left[T_1,T_2\right]}(t).
\end{align}
Let $(A_t^{(2)})_{t \in [T_0,T_2]}$ be the arcade process with these interpolating coefficients driven by a standard Brownian motion. As shown in Example \ref{nonstarcade}, $(A_t^{(2)})$ is Markov. For Eq. (\ref{subcond1}) to be fulfilled, one only need to impose $g_2(t)\mathds{1}_{[T_0,T_1]}(t)=0$. For Eq. (\ref{subcond2}) to be fulfilled, one requires that
\begin{align}
g_1(t)\mathds{1}_{[T_0,T_1]}(t)&= \frac{t-T_0}{T_1-T_0}\mathds{1}_{[T_0,T_1]}(t),\\ g_2(t)\mathds{1}_{\left[T_1,\frac{T_1+T_2}{2}\right]}(t)&=\left( \frac{t-T_1}{T_2-T_1} + \frac{(t-T_1)T_0}{(T_1-T_0)^2}\right) \mathds{1}_{\left[T_1,\frac{T_1+T_2}{2}\right]}(t),\\
g_2(t)\mathds{1}_{\left(\frac{T_1+T_2}{2}, T_2 \right]}(t) &= \left(\frac{t-T_1}{T_2-T_1} + \frac{(T_2 -t )T_0}{(T_1-T_0)^2}\right) \mathds{1}_{\left(\frac{T_1+T_2}{2}, T_2 \right]}(t).
\end{align}
Outside of the considered intervals, the functions $g_i$ can take any values as long as they remain interpolating coefficients. Notice that the theorem does not impose a condition on $g_0$. For example, we could choose $g_i=f_i$ outside the above intervals. Hence, all three conditions are fulfilled and this $X$-RAP is $\{T_0,T_2\}_*$-nearly-Markov. As we can see from the paths simulation below, this process is visually different from a randomized stitched Brownian arcade on the second arc (the noise has been diminished to make the paths more informative). Simulating the signal function by itself highlights the following: $X_0$ will determine the fate of the signal function on $[T_1,T_2]$ since this RAP is not forgetting about previously matched random variables when changing arc. On the first arc, where the process is simply a randomized Brownian bridge: to go from $X_0=-1$ to $X_1=-1$ for instance, there is only one way, a straight line. On the second arc, to go from $X_1=-1$ to $X_2=-1$, there are two ways. The signal function will choose which way to use based on the value of $X_0$. This is illustrated by the paths of the signal function below: the blue path and the green path both take value $-1$ at $T_1$ and value $1$ at $T_2$, but have different values in $T_0$. Hence they differ on $[T_1,T_2]$, as observed.
\begin{figure}[H]
\centering
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{Puresignal.png}
\caption{Paths simulation of the signal function of a non-standard $X$-RAP, where $\{T_0,T_2\}_*=\{0,5,10\}$.}
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{SpecialRAP.png}
\caption{Paths simulation of a non-standard $X$-RAP, where the noise process was rescaled by 0.3, and $\{T_0,T_2\}_*=\{0,5,10\}$.}
\end{minipage}
\end{figure}
\end{ex}
\section{Filtered arcade martingales}
In this section, we construct martingales with respect to the filtration generated by an $X$-RAP, which interpolate between the components of the random vector $X=(X_0,\ldots,X_n)$. These martingales solve an underlying stochastic filtering problem, and extend the martingale class constructed within the information-based theory developed in \cite{BHM1}, see also \cite{BHM2}. We call such martingales \emph{filtered arcade martingales} (FAMs). We introduce the following notation:
\begin{enumerate}
\item $\mathcal P(\mathbb{R}^n)$ is the set of Borel probability measures on $\mathbb{R}^n$, for $n\in \mathbb{N}_0$.
\item $\mathcal P_1(\mathbb{R}^n)$ is the set of Borel probability measures on $\mathbb{R}^n$, for $n\in \mathbb{N}_0$, with finite first moment.
\item \begin{multline}\Pi(\mu_0,\mu_1,\ldots, \mu_n):= \{ \pi \in \mathcal P(\mathbb{R}^{n+1}) \| \mu_{i-1} \text{ is the ith marginal measure of } \pi, \text{ for all } \\ i=1,\ldots, n+1 \}\nonumber \end{multline} is the set of couplings of $(\mu_0,\mu_1,\ldots, \mu_n) \subseteq \mathcal P (\mathbb{R})$.
\item \begin{multline}\mathcal M (\mu_0,\mu_1,\ldots, \mu_n):=\{\pi \in \Pi(\mu_0,\mu_1,\ldots, \mu_n) \cap \mathcal P_1(\mathbb{R}^{n+1}) \| (X_0,X_1,\ldots, X_n) \sim \pi \implies \\ \mathbb{E}[X_n \| X_0,\ldots, X_m] = X_m \text{ for all } m=0,\ldots, n\}\nonumber \end{multline} is the set of martingale couplings of $(\mu_0,\mu_1,\ldots, \mu_n) \subseteq \mathcal P (\mathbb{R})$.
\end{enumerate}
We recall that $\mathcal M (\mu_0,\mu_1,\ldots, \mu_n) \neq \emptyset$ if and only if $\int_\mathbb{R} x \, \mathrm{d} \mu_0(x)= \int_\mathbb{R} x \, \mathrm{d} \mu_1(x)= \ldots = \int_\mathbb{R} x \, \mathrm{d} \mu_n(x)$ and $\int_\mathbb{R} f(x) \, \mathrm{d} \mu_0(x) \leqslant \int_\mathbb{R} f(x) \, \mathrm{d} \mu_1(x)\leqslant \ldots \leqslant \int_\mathbb{R} f(x) \, \mathrm{d} \mu_n(x)$ for any convex function $f$ on $\mathbb{R}$, i.e, the measures $\mu_0,\mu_1,\ldots, \mu_n$ are convexly ordered. In that case, we write $\mu_0\leqslant_{\mathrm{cx}}\mu_1\leqslant_{\mathrm{cx}}\ldots\leqslant_{\mathrm{cx}} \mu_n$.
\subsection{The one-arc FAM}
Let $X=(X_0,X_1)$ be a real random vector and $(\mu_0,\mu_1) \in \mathcal P (\mathbb{R}) \times \mathcal P (\mathbb{R})$ such that $X_0 \sim \mu_0$, $X_1 \sim \mu_1$, and $\mu_0\leqslant_{\mathrm{cx}}\mu_1$. Given a martingale coupling $\pi^X \in \mathcal M(\mu_0,\mu_1)$ for $X$ and an $X$-RAP $(I_t^{(1)})_{t \in [T_0,T_1]}$ on the partition $\{T_0,T_1\}_*$, we aim at constructing an $(\mathcal{F}^I_t)$-martingale $(M_t)_{t\in[T_0, T_1]}$ such that $M_{T_0}\overset{a.s.}{=} X_0$ and $M_{T_1}\overset{a.s.}{=} X_1$. The BHM framework developed in \cite{BHM1} is recovered when $X_0 =\mathbb{E}[X_1]=0$ and $T_0=0$.
\begin{defn} \label{1FAM}
Given an $X$-RAP $(I_t^{(1)})$ and a martingale coupling $\pi^X$, a one-arc FAM for $X$ on $[T_0,T_1]$ is a stochastic process of the form $M_t=\mathbb{E}[X_1 \| \mathcal{F}_t^I ]$.
\end{defn}
\begin{prop}
The FAM $(M_t)_{t\in[T_0, T_1]}$ is an $(\mathcal{F}_t^I)$-martingale that interpolates between $X_0$ and $X_1$, almost surely.
\end{prop}
\begin{proof}
The process $(M_t)$ is adapted to $(\mathcal F_t^I)$ by construction. We show that it is a martingale that interpolates between $X_0$ and $X_1$:
\begin{enumerate}
\item $\mathbb{E}[|M_t|] < +\infty$ for all $t\in [T_0,T_1]$ by Jensen inequality, since $\mathbb{E}[|X_1|] < +\infty$.
\item For $s<t$, $\mathbb{E}[M_t \| \mathcal{F}_s^I] = M_s$ by the tower property of the conditional expectation.
\item $M_{T_0} =\mathbb{E}[X_1 \| X_0] = X_0$, since $\pi^X \in \mathcal M (\mu_0,\mu_1)$ and $I_{T_0}^{(1)}=X_0$.
\item $M_{T_1}= \mathbb{E}[X_1 \| \mathcal{F}_{T_1}^I] = X_1$ by construction of $(I_t^{(1)})$.
\end{enumerate}
Hence, $(M_t)$ is an $(\mathcal{F}_t^I)$-martingale that interpolates almost surely between $X_0$ and $X_1$ on $[T_0,T_1]$.
\end{proof}
\begin{rem}
The process $(M_t)_{t\in [T_0,T_1]}$ is also a martingale with respect to its natural filtration, denoted $(\mathcal{F}_t^M)_{t\in[T_0, T_1]}$. For $s<t$, we have
\begin{align}
\mathbb{E} [M_t \| \mathcal{F}^M_s]= \mathbb{E}[\mathbb{E} [M_t \| \mathcal{F}^M_s] \| \mathcal{F}^I_s] = \mathbb{E}[\mathbb{E} [M_t \| \mathcal{F}^I_s] \| \mathcal{F}^M_s] = \mathbb{E}[ M_s\| \mathcal{F}^M_s] = M_s.
\end{align}
\end{rem}
A relationship that we will use later, between the RAP $(I_t^{(1)})$, its noise process $(A_t^{(1)})$, and its associated FAM $(M_t)$ is the following:
\begin{prop} \label{relations}
If $I_t^{(1)}= g_0(t)X_0 + g_1(t) X_1 + A_t^{(1)}$ and $M_t=\mathbb{E}[X_1 \| \mathcal{F}_t^I ]$, then
\begin{equation} \label{rel2}
\mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_s]=g_0(t) X_0 + g_1(t) M_s + \mathbb{E}[ \mathbb{E} [A_t^{(1)} \| \mathcal{F}^A_s] \| \mathcal{F}^I_s]
\end{equation}
for any pair $(s,t) \in [T_0,T_1]^2$ such that $s\leqslant t$. Furthermore, if $(A_t^{(1)})$ is Gauss-Markov with $K_A(x,y)= A_1(\min(x,y))A_2(\max(x,y))$, then
\begin{align}
\mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_s]&= \left(g_0(t) - \frac{A_2(t)}{A_2(s)} g_0(s) \right) X_0 + \left(g_1(t)- \frac{A_2(t)}{A_2(s)} g_1(s)\right) M_s\nonumber \\
& \quad + \frac{A_2(t)}{A_2(s)} I_s + \mu_A(t) +\frac{A_2(t)}{A_2(s)} \mu_A(s)
\end{align}
for any $(s,t) \in [T_0,T_1]^2$ such that $s\leqslant t$.
\end{prop}
\begin{proof}
Let $(s,t) \in [T_0,T_1]^2$ such that $s\leqslant t$. Notice that
\begin{align}
\mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_s]&= g_0(t) X_0 + g_1(t) M_s + \mathbb{E}[A_t^{(1)} \| \mathcal{F}^I_s] \nonumber \\
&= g_0(t) X_0 + g_1(t) M_s + \mathbb{E}[ \mathbb{E} [A_t^{(1)} \| \mathcal{F}^A_s] \| \mathcal{F}^I_s],
\end{align}
since $\mathbb{E}[A_t^{(1)} \| \mathcal{F}^I_s] = \mathbb{E}[ \mathbb{E} [A_t^{(1)} \| X_0,X_1, \mathcal{F}^A_s] \| \mathcal{F}^I_s] =\mathbb{E}[ \mathbb{E} [A_t^{(1)} \| \mathcal{F}^A_s] \| \mathcal{F}^I_s]$, where we used the fact that $(A_t^{(1)}) \perp\!\!\!\perp (X_0,X_1)$ by Definition \ref{RAP}. If $(A_t^{(1)})$ is Gauss-Markov, then
\begin{equation}
\mathbb{E} [A_t^{(1)} \| \mathcal{F}^A_s] = \mathbb{E} [A_t^{(1)} \|A_s^{(1)}] = \mu_A(t) + \frac{K_A(s,t)}{\sigma_A^2(s)} \left( A_s^{(1)} - \mu_A(s) \right).
\end{equation}
Hence, by linearity of the conditional expectation, we have
\begin{equation} \label{rel4}
\mathbb{E}[ \mathbb{E} [A_t^{(1)} \| \mathcal{F}^A_s] \| \mathcal{F}^I_s] = \mu_A(t) + \frac{K_A(s,t)}{\sigma_A^2(s)} \left( \mathbb{E}[A_s^{(1)} \| \mathcal{F}^I_s] - \mu_A(s) \right).
\end{equation}
Moreover,
\begin{equation}
I_t^{(1)} = \mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_t] = \mathbb{E}[A_t^{(1)} \| \mathcal{F}^I_t] + g_0(t) X_0+g_1(t) M_t,
\end{equation}
which implies
\begin{equation} \label{rel1}
\mathbb{E}[A_t^{(1)} \| \mathcal{F}^I_t]= I_t^{(1)} - g_0(t) X_0-g_1(t) M_t.
\end{equation}
Then, plugging Eq. (\ref{rel1}) in Eq. (\ref{rel4}) and recalling that $K_A(s,t)=A_1(s)A_2(t)$, gives
\begin{align}
\mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_s]&= \left(g_0(t) - \frac{A_2(t)}{A_2(s)} g_0(s) \right) X_0 + \left(g_1(t)- \frac{A_2(t)}{A_2(s)} g_1(s)\right) M_s\nonumber \\
& \quad + \frac{A_2(t)}{A_2(s)} I_s + \mu_A(t) +\frac{A_2(t)}{A_2(s)} \mu_A(s). \label{rel3}
\end{align}
\end{proof}
If $(I_t^{(1)})$ is $\{T_0,T_1\}$-nearly-Markov, then $M_t = \mathbb{E} [ X_1 \| X_0, I_t^{(1)} ] $. One can then derive the dynamics of $(M_t)$ using Bayes' rule and Itô's lemma under mild assumptions. In the remainder of this section, we assume that the driver of $(I_t^{(1)})$ has a density function for $t \in [T_0,T_1]$.
\begin{prop} \label{Bayes}
Let $M_t = \mathbb{E} [ X_1 \| X_0, I_t^{(1)} ]$ be a one-arc FAM restricted to $t\in(T_0,T_1)$. Then
\begin{equation} \label{Bayeseq}
M_t= \frac{\int_\mathbb{R} y f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \, \mathrm{d} F^{X_1 \| X_0} (y)}{\int_\mathbb{R} f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \, \mathrm{d} F^{X_1 \| X_0} (y)},
\end{equation}
where $f^{I_t^{(1)} \| X_0,X_1} (I_t^{(1)})$ is the conditional density of $(I_t^{(1))}$ given $X_0$ and $X_1$ and $F^{X_1 \| X_0}$ is the distribution function of $X_1$ given $X_0$. In particular,
\begin{enumerate}
\item If $(X_0,X_1)$ is a continuous random vector, then
\begin{equation}
M_t = \frac{\int_\mathbb{R} y f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) f^{X_1 \| X_0}(y) \, \mathrm{d} y}{\int_\mathbb{R} f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) f^{X_1 \| X_0}(y) \, \mathrm{d} y} = \frac{\int_\mathbb{R} y f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) f^{(X_0,X_1)}(X_0,y) \, \mathrm{d} y}{\int_\mathbb{R} f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) f^{(X_0,X_1)}(X_0,y) \, \mathrm{d} y}.
\end{equation}
\item If $(X_0,X_1)$ is a discrete random vector, then
\begin{equation}
M_t = \frac{\sum\limits_{y} y f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \mathbb{P}[X_1=y \| X_0] }{\sum\limits_{y} f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \mathbb{P}[X_1=y \| X_0]} .
\end{equation}
\end{enumerate}
\end{prop}
\begin{proof}
By the Bayes rule,
\begin{align}
\mathbb{P} [ X_1 \leqslant y \| X_0, z \leqslant I_t^{(1)} \leqslant z + \epsilon ] &= \frac{\mathbb{P} [ z \leqslant I_t^{(1)} \leqslant z + \epsilon \| X_0, X_1 \leqslant y ] \mathbb{P} [ X_1 \leqslant y \| X_0]}{ \mathbb{P} [ z \leqslant I_t^{(1)} \leqslant z + \epsilon \| X_0 ]}\nonumber\\
&= \frac{\mathbb{P} [ z \leqslant I_t^{(1)} \leqslant z + \epsilon \| X_0, X_1 \leqslant y ] \mathbb{P} [ X_1 \leqslant y \| X_0]}{ \int_\mathbb{R} \mathbb{P} [ z \leqslant I_t^{(1)} \leqslant z + \epsilon \| X_0 , X_1 =y] \, \mathrm{d} F^{X_1\| X_0}(y) }\nonumber \\
&= \frac{\mathbb{P} [ z \leqslant I_t^{(1)} \leqslant z + \epsilon, X_1 \leqslant y \| X_0] }{ \int_\mathbb{R} \mathbb{P} [ z \leqslant I_t^{(1)} \leqslant z + \epsilon \| X_0 , X_1 =y] \, \mathrm{d} F^{X_1\| X_0}(y) }.
\end{align}
This means that, by taking the limit when $\epsilon \rightarrow 0$,
\begin{equation}
F^{X_1\| X_0,I_t^{(1)} =z}(y) = \frac{ \frac{\, \mathrm{d}}{\, \mathrm{d} z}\mathbb{P} [I_t^{(1)} \leqslant z \| X_0, X_1 \leqslant y] F^{X_1\| X_0}(y)}{ \int_\mathbb{R} f^{I_t^{(1)} \| X_0,X_1 = y} (z)\, \mathrm{d} F^{X_1\| X_0}(y) },
\end{equation}
which implies
\begin{equation}
\, \mathrm{d} F^{X_1\| X_0,I_t^{(1)} =z}(y) = \frac{f^{I_t^{(1)} \| X_0,X_1 =y} (z)\, \mathrm{d} F^{X_1\| X_0}(y) }{ \int_\mathbb{R} f^{I_t^{(1)} \| X_0,X_1 = y} (z)\, \mathrm{d} F^{X_1\| X_0}(y) }.
\end{equation}
Inserting the expression for $\, \mathrm{d} F^{X_1\| X_0,I_t^{(1)} =z}(y)$ into $M_t$, we obtain
\begin{equation}
M_t= \int_{\mathbb{R}} y \, \mathrm{d} F^{X_1\| X_0,I_t^{(1)}}(y) = \frac{\int_\mathbb{R} y f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \, \mathrm{d} F^{X_1 \| X_0} (y)}{\int_\mathbb{R} f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \, \mathrm{d} F^{X_1 \| X_0} (y)}.
\end{equation}
\end{proof}
\begin{ex}
Let $X_0 \sim \mathcal U ([-1,1])$, and $X_1 \sim \mathcal U ([-2,2])$. These random variables are convexly ordered, i.e., $X_0 \leqslant_{\mathrm{cx}} X_1$, hence there exists at least one martingale coupling for $(X_0,X_1)$. We choose the coupling defined by
\begin{equation}
X_1 \| X_0 = \left\{
\begin{array}{ll}
\frac{3}{2} X_0 + \frac{1}{2} & \mbox{with probability } \frac{3}{4},\\ \\
-\frac{1}{2} X_0 - \frac{3}{2} & \mbox{with probability } \frac{1}{4}.
\end{array}
\right.
\end{equation}
This is a martingale coupling since $\mathbb{E}[X_1 \| X_0] = X_0$. In fact, it can be shown that this coupling is the solution to a martingale optimal transport problem, see \cite{Sester}. For any $\{T_0,T_1\}$-nearly-Markov $X$-RAP $(I_t^{(1)})$, we have
\begin{equation}
M_t =\mathbb{E} [ X_1 \| X_0, I_t^{(1)} ]= \frac{(9 X_0 + 3) f^{I_t^{(1)} \| X_0,X_1=\frac{3}{2} X_0 + \frac{1}{2}} (I_t^{(1)}) - ( X_0 + 3 ) f^{I_t^{(1)} \| X_0,X_1=-\frac{1}{2} X_0 - \frac{3}{2} } (I_t^{(1)})} { 6 f^{I_t^{(1)} \| X_0,X_1=\frac{3}{2} X_0 + \frac{1}{2}} (I_t^{(1)}) + 2f^{I_t^{(1)} \| X_0,X_1=-\frac{1}{2} X_0 - \frac{3}{2} } (I_t^{(1)}) } .
\end{equation}
\end{ex}
\begin{ex}
Let $X_0 \sim \mathcal N (0,1)$, and $X_1 \sim \mathcal N (0,2)$, where $X_1 \| X_0 \sim \mathcal N (X_0,1)$. For any $\{T_0,T_1\}$-nearly-Markov $X$-RAP $(I_t^{(1)})$, we have
\begin{equation}
M_t =\mathbb{E} [ X_1 \| X_0, I_t^{(1)} ]= \frac{\int_\mathbb{R} y f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \mathrm{e}^{\frac{-(y-X_0)^2}{2}} \, \mathrm{d} y} {\int_\mathbb{R} f^{I_t^{(1)} \| X_0,X_1=y} (I_t^{(1)}) \mathrm{e}^{\frac{-(y-X_0)^2}{2}} \, \mathrm{d} y} .
\end{equation}
\end{ex}
To simplify some of the expressions, we introduce the following notation:
\begin{equation}
u(t, z, X_0, y) := f^{I_t^{(1)} \| X_0,X_1=y} (z), \quad u_t(t, z, X_0, y) := \frac{\partial}{\partial t}u(t, z, X_0, y),
\end{equation}
\begin{equation}
u_z(t, z, X_0, y) := \frac{\partial }{\partial z}u(t, z, X_0, y), \quad u_{z z} (t, z, X_0, y) := \frac{\partial^2 }{\partial z^2}u(t, z, X_0, y),
\end{equation}
\begin{equation}
K_{\boldsymbol{\cdot}} (t,z,X_0):=\int_\mathbb{R} u_{\boldsymbol{\cdot}} (t, z, X_0, y) \, \mathrm{d} F^{X_1 \| X_0} (y), \quad V_{\boldsymbol{\cdot}} (t,z,X_0):=\int_\mathbb{R} y u_{\boldsymbol{\cdot}} (t, z, X_0, y)\, \mathrm{d} F^{X_1 \| X_0} (y).
\end{equation}
Thus, under the conditions in Proposition $\ref{Bayes}$, we may write Eq. (\ref{Bayeseq}) as
\begin{equation}
M_t = \frac{V(t,I_t^{(1)},X_0)}{K(t,I_t^{(1)},X_0)}.
\end{equation}
\begin{prop} \label{Ito}
If $(I_t^{(1)})$ is a semimartingale such that $(t,x) \rightarrow \frac{V(t,x,X_0)}{K(t,x,X_0)}$ is $C^2(((T_0,T_1) \setminus N) \times Im(I^{(1)}))$ where $N \subset (T_0,T_1)$ contains finitely many elements, and $M_t = \mathbb{E} [ X_1 \| X_0, I_t^{(1)} ]$ is a one-arc FAM, then
\begin{align}
\, \mathrm{d} M_t =& \frac{V_t(t,I_t^{(1)},X_0)-M_t K_t(t,I_t^{(1)},X_0) }{K(t,I_t^{(1)},X_0)} \, \mathrm{d} t + \frac{V_z(t,I_t^{(1)},X_0)-M_t K_z(t,I_t^{(1)},X_0) }{K(t,I_t^{(1)},X_0)} \, \mathrm{d} I^{(1)}_t \nonumber \\
&+ \Bigg( \frac{M_t K_z^2(t,I_t^{(1)},X_0) -K_z(t,I_t^{(1)},X_0)V_z(t,I_t^{(1)},X_0)}{K^2(t,I_t^{(1)},X_0)} \nonumber \\
& \hspace{5cm}+ \frac{V_{z z}(t,I_t^{(1)},X_0)-M_t K_{z z}(t,I_t^{(1)},X_0) }{2K(t,I_t^{(1)},X_0)} \Bigg) \, \mathrm{d} [I^{(1)}]_t.
\end{align}
for $t \in (T_0,T_1)$.
\end{prop}
\begin{proof}
This is verified by a straightforward application of Itô's lemma.
\end{proof}
\begin{rem}
The Itô condition, i.e., $(t,x) \rightarrow \frac{V(t,x,X_0)}{K(t,x,X_0)}$ is $C^2(((T_0,T_1) \setminus N) \times Im(I^{(1)})))$, where $N \subset (T_0,T_1)$ contains finitely many elements, imposes implicit integrability conditions on $(X_0,X_1)$.
\end{rem}
Since Gaussian RAPs were studied in detail above, we can specialize Proposition \ref{Ito} to this particular subclass.
\begin{coro}
Under the conditions in Proposition \ref{Ito}, if the conditional probability distribution of $(I_t^{(1)})$ given $(X_0,X_1)$ is $\mathcal N (g_0(t) X_0 + g_1(t) X_1 + \mu_A(t), \sigma_A^2(t))$, we have
\begin{align}
&\frac{V_t(t,I_t^{(1)},X_0)-M_t K_t(t,I_t^{(1)},X_0) }{K(t,I_t^{(1)},X_0)} = -\frac{(\mathbb{E}[X_1^3 \| X_0, I_t^{(1)}] - M_t^3) g_1(t) \left(\frac{g_1(t)}{\sigma_A(t)}\right)'}{\sigma_A(t)} \nonumber \\
&+ \Bigg( \frac{M_t g_1(t) \left(\frac{g_1(t)}{\sigma_A(t)}\right)' + (I_t^{(1)} - g_0(t) X_0- \mu_A(t))\left( \left(\frac{g_1(t)}{\sigma_A(t)}\right)' + g_1(t)\left(\frac{1}{\sigma_A(t)}\right)'\right) }{\sigma_A(t)} \nonumber \\
& \hspace{4cm}- \frac{(X_0 g_0'(t) + \mu_A'(t))g_1(t) }{\sigma_A^2(t)} \Bigg) \times \mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ],\\
&\frac{V_z(t,I_t^{(1)},X_0)-M_t K_z(t,I_t^{(1)},X_0) }{K(t,I_t^{(1)},X_0)} = \frac{g_1(t)}{\sigma_A^2(t)} \mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ],
\end{align}
\begin{align}
&\frac{M_t K_z^2(t,I_t^{(1)},X_0) -K_z(t,I_t^{(1)},X_0)V_z(t,I_t^{(1)},X_0)}{K^2(t,I_t^{(1)},X_0)} = \frac{I_t^{(1)} -g_0(t)X_0 - \mu_A(t)- g_1(t)M_t }{\sigma_A^4(t)} g_1(t) \nonumber \\
& \hspace{12cm} \times \mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ],\\
& \frac{V_{z z}(t,I_t^{(1)},X_0)-M_t K_{z z}(t,I_t^{(1)},X_0) }{2K(t,I_t^{(1)},X_0)} = \frac{\mathbb{E}[X_1^3 \| X_0, I_t] - M_t^3}{2\sigma_A^4(t)} g_1^2(t) \nonumber \\
&\hspace{6cm} - \frac{2(I_t^{(1)} -g_0(t)X_0 - \mu_A(t)) + g_1(t)M_t} {2\sigma_A^4(t)} g_1(t)\mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ].
\end{align}
\end{coro}
\begin{proof}
Denoting $Z_t= I_t^{(1)} -g_0(t)X_0 - \mu_A(t)$, and $J_t= X_0\, g_0'(t) + \mu_A'(t)$, the result follows from the following computations:
\begin{enumerate}
\item \begin{align}
V_t(t,I_t^{(1)},X_0) &= \frac{g_1(t) \sigma_A'(t) - g_1'(t) \sigma_A(t) }{ \sigma_A^3(t)} g_1(t) K (t,I_t^{(1)},X_0) \mathbb{E}[X_1^3 \| X_0, I_t^{(1)}] \nonumber \\
&\quad + \frac{Z_t J_t\sigma_A(t) + \sigma_A'(t)(Z_t^2 - \sigma_A^2(t) )}{\sigma_A^3(t)} V(t,I_t^{(1)},X_0) \nonumber \\
&\quad - \frac{g_1(t) (J_t \sigma_A(t) + 2 Z_t \sigma_A'(t)) - Z_t \sigma_A(t) g_1'(t) }{\sigma_A^3(t)} K(t,I_t^{(1)},X_0) \mathbb{E}[X_1^2 \| X_0, I_t^{(1)}],
\end{align}
\item \begin{align}
K_t(t,I_t^{(1)},X_0) &= \frac{g_1(t) \sigma_A'(t) - g_1'(t) \sigma_A(t) }{ \sigma_A^3(t)} g_1(t) K (t,I_t^{(1)},X_0) \mathbb{E}[X_1^2 \| X_0, I_t^{(1)}] \nonumber \\
&\quad + \frac{Z_t J_t\sigma_A(t) + \sigma_A'(t)(Z_t^2 - \sigma_A^2(t) )}{\sigma_A^3(t)} K(t,I_t^{(1)},X_0) \nonumber \\
&\quad - \frac{g_1(t) (J_t \sigma_A(t) + 2 Z_t \sigma_A'(t)) - Z_t \sigma_A(t) g_1'(t) }{\sigma_A^3(t)} K(t,I_t^{(1)},X_0) M_t,
\end{align}
\item \begin{equation}
V_z(t,I_t^{(1)},X_0) = \frac{g_1(t) \mathbb{E}[X_1^2 \| X_0, I_t^{(1)}] K (t,I_t^{(1)},X_0) - Z_t V (t,I_t^{(1)},X_0)} {\sigma_A^2(t)},
\end{equation}
\item \begin{equation}
K_z(t,I_t^{(1)},X_0) = \frac{g_1(t) V (t,I_t^{(1)},X_0) - Z_t K (t,I_t^{(1)},X_0)} {\sigma_A^2(t)},
\end{equation}
\item \begin{align}
V_{zz}(t,I_t^{(1)},X_0) &= \frac{g_1^2(t) E[X_1^3 \| X_0, I_t^{(1)}] - 2 Z_t g_1(t) \mathbb{E}[X_1^2 \| X_0, I_t^{(1)}]}{\sigma_A^4(t)} K (t,I_t^{(1)},X_0) \nonumber \\
&\quad + \frac{Z_t^2 - \sigma_A^2(t)}{\sigma_A^4(t)} V(t,I_t^{(1)},X_0),
\end{align}
\item \begin{align}
K_{zz}(t,I_t^{(1)},X_0) &= \frac{g_1^2(t) E[X_1^2 \| X_0, I_t^{(1)}]K(t,I_t^{(1)},X_0) - 2 Z_t g_1(t) V(t,I_t^{(1)},X_0)}{\sigma_A^4(t)} K (t,I_t^{(1)},X_0) \nonumber \\
&\quad + \frac{Z_t^2 - \sigma_A^2(t)}{\sigma_A^4(t)} K(t,I_t^{(1)},X_0).
\end{align}
\end{enumerate}
\end{proof}
Using the notation $Z_t= I_t^{(1)} -g_0(t)X_0 - \mu_A(t)$ and $J_t= X_0\, g_0'(t) + \mu_A'(t)$ from the previous proof, and introducing $U_t= \mathbb{E}[X_1^3 \| X_0, I_t] - M_t^3$, the SDE satisfied by $(M_t)$ can be rewritten as
\begin{align}
\, \mathrm{d} M_t =& U_t\left(\frac{g_1^2(t)}{2\sigma_A^4(t)} \, \mathrm{d} [I^{(1)}]_t - \frac{ g_1(t) \left(\frac{g_1(t)}{\sigma_A(t)}\right)'}{\sigma_A(t)} \, \mathrm{d} t \right) + \frac{g_1(t)}{\sigma_A^2(t)} \mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ] \nonumber \\
&\times \Bigg( \Bigg( M_t \sigma_A(t) \left(\frac{g_1(t)}{\sigma_A(t)}\right)' + Z_t \frac{\sigma_A(t)}{g_1(t)}\left( \left(\frac{g_1(t)}{\sigma_A(t)}\right)' + g_1(t)\left(\frac{1}{\sigma_A(t)}\right)'\right) - J_t \Bigg) \, \mathrm{d} t \nonumber \\
& \hspace{5cm}- \frac{3g_1(t)M_t }{2\sigma_A^2(t)} \, \mathrm{d} [I^{(1)}]_t +\, \mathrm{d} I^{(1)}_t \Bigg).
\end{align}
If, furthermore, $(I_t^{(1)})$ is a standard RAP (see Definition \ref{standardRAP}), its driver $(D_t)$ is Gauss-Markov with $K_D(x,y)=H_1(\min(x,y))H_2(\max(x,y))$, where $H_1$ and $H_2$ are continuous functions on $[T_0,T_1]$ such that $H_1/H_2$ is positive and non-decreasing on $[T_0,T_1]$. Then, as shown below,
\begin{equation}
[I^{(1)}]_t = [D]_t=\int_{T_0}^t H_2(s) \, \mathrm{d} H_1(s) - \int_{T_0}^t H_1(s) \, \mathrm{d} H_2(s),
\end{equation}
where the RHS is interpreted as a difference of Riemann-Stieltjes integrals. The RHS exists since
\begin{equation}
\int_{T_0}^t H_2(s) \, \mathrm{d} H_1(s) - \int_{T_0}^t H_1(s) \, \mathrm{d} H_2(s) =\int_{T_0}^t H_2^2(s) \, \mathrm{d} \left(\frac{H_1}{H_2}\right)(s),
\end{equation}
and $H_1/H_2$ is continuous and monotone, so of bounded variation and differentiable almost everywhere.
\begin{prop}
If the driver $(D_t)$ is a Gauss-Markov semimartingale with $K_D(x,y)=H_1(\min(x,y))H_2(\max(x,y))$, then
\begin{equation}
[D]_t=\int_{T_0}^t H_2(s) \, \mathrm{d} H_1(s) - \int_{T_0}^t H_1(s) \, \mathrm{d} H_2(s).
\end{equation}
\end{prop}
\begin{proof}
Let $t\in [T_0,T_1]$, $T_0=t_0 < t_1 < \ldots < t_n = t$ be a partition of $[T_0,t]$, and $\Delta_j=D_{t_{j+1}}-D_{t_{j}}$. Then,
\begin{align}
\mathbb{E} \left[ \sum_{j=0}^{n-1} \Delta_j^2 \right] &= \sum_{j=0}^{n-1} \mathbb{E} [D_{t_{j+1}}^2] + \mathbb{E}[D_{t_{j}}^2] - 2\mathbb{E} [D_{t_{j+1}}D_{t_{j}}]\nonumber \\
&=\sum_{j=0}^{n-1} H_1(t_{j+1})H_2(t_{j+1}) + H_1(t_{j})H_2(t_{j}) - 2 H_1(t_{j})H_2(t_{j+1}) \nonumber \\
&\hspace{7cm}+ \left( \mu_D(t_{j+1}) - \mu_D(t_{j}) \right)^2\nonumber \\
&= \sum_{j=0}^{n-1} H_2(t_{j+1})(H_1(t_{j+1}) - H_1(t_{j})) - H_1(t_{j})(H_2(t_{j+1}) - H_2(t_{j})) \nonumber \\
&\hspace{7cm}+ \left( \mu_D(t_{j+1}) - \mu_D(t_{j}) \right)^2.
\end{align}
Since $\mu_D$ is continuous and of bounded variation (see Proposition 1 in \cite{Stricker}), we have
\begin{equation}
\sum_{j=0}^{n-1} \left( \mu_D(t_{j+1}) - \mu_D(t_{j}) \right)^2 \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} 0.
\end{equation}
This shows that
\begin{equation}
\mathbb{E} \left[ \sum_{j=0}^{n-1} \Delta_j^2 \right] \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} \int_{T_0}^t H_2(s) \, \mathrm{d} H_1(s) - \int_{T_0}^t H_1(s) \, \mathrm{d} H_2(s).
\end{equation}
Furthermore,
\begin{align}
\mathrm{Var} \left[ \sum_{j=0}^{n-1} \Delta_j^2 \right] &= \sum_{j=0}^{n-1} \sum_{i=0}^{n-1} \mathrm{Cov} \left(\Delta_i^2, \Delta_j^2 \right) \nonumber \\
&= 2 \sum_{j=0}^{n-1} \sum_{i=0}^{n-1} \mathrm{Cov}^2 \left(\Delta_i, \Delta_j \right) + 4 \sum_{j=0}^{n-1} \sum_{i=0}^{n-1} \mathbb{E}[\Delta_i]\mathbb{E}[\Delta_j] \mathrm{Cov} \left(\Delta_i, \Delta_j \right) \label{fullsum},
\end{align}
and
\begin{equation}
\mathrm{Cov} \left(\Delta_i, \Delta_j \right) = \left\{
\begin{array}{ll}
(H_1(t_{i+1}) - H_1(t_{i}))(H_2(t_{j+1}) - H_2(t_{j})) & \mbox{if } i<j,\\ \\
(H_1(t_{j+1}) - H_1(t_{j}))(H_2(t_{i+1}) - H_2(t_{i})) & \mbox{if } j<i, \\ \\
H_2(t_{j+1})(H_1(t_{j+1}) - H_1(t_{j})) - H_1(t_{j})(H_2(t_{j+1}) - H_2(t_{j})) & \mbox{if } i=j.
\end{array}
\right.
\end{equation}
We split Eq. (\ref{fullsum}) into the cases $i<j$, $j<i$, and $i=j$. Considering the case $i=j$ first, we get
\begin{align}
\sum_{j=0}^{n-1} 2 \mathrm{Var}^2 [ \Delta_j] + 4 \mathbb{E}^2 [\Delta_j] \mathrm{Var} [ \Delta_j] &= \sum_{j=0}^{n-1} \mathrm{Var} [ \Delta_j] \left(2 \mathrm{Var} [ \Delta_j] +4 \left( \mu_D(t_{j+1}) - \mu_D(t_{j}) \right)^2 \right) \nonumber \\
&\leqslant \sum_{j=0}^{n-1} \left(2 \mathrm{Var} [ \Delta_j] +4 \left( \mu_D(t_{j+1}) - \mu_D(t_{j}) \right)^2 \right) \nonumber \\
&\hspace{5cm} \times \max_{k \in\{0,1,\ldots, n-1 \}} \mathrm{Var} [ \Delta_k].
\end{align}
Since $\max\limits_{k \in\{0,1,\ldots, n-1 \}} \mathrm{Var} [ \Delta_k] \xrightarrow[\max \, \abs{t_{j+1} - t_{j}} \rightarrow 0]{} 0$ by uniform continuity of $H_1$ and $H_2$, and
\begin{equation}
\lim\limits_{\max \, \abs{t_{j+1} - t_{j}} \rightarrow 0} \sum_{j=0}^{n-1} \left(2 \mathrm{Var} [ \Delta_j] +4 \left( \mu_D(t_{j+1}) - \mu_D(t_{j}) \right)^2 \right) < +\infty,
\end{equation}
we have $\sum\limits_{j=0}^{n-1} 2 \mathrm{Var}^2 [ \Delta_j] + 4 \mathbb{E}^2 [\Delta_j] \mathrm{Var} [ \Delta_j] \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{}0 $.
\bigskip
Next, for $i<j$, we have
\begin{align}
&2 \sum_{j=0}^{n-1} \sum_{i=0}^{j-1} (H_1(t_{i+1}) - H_1(t_{i}))^2(H_2(t_{j+1}) - H_2(t_{j}))^2 \nonumber \\
&\quad + 4 \sum_{j=0}^{n-1} \sum_{i=0}^{j-1} (\mu_D(t_{i+1}) - \mu_D(t_{i}))(\mu_D(t_{j+1}) - \mu_D(t_{j})) (H_1(t_{i+1}) - H_1(t_{i}))(H_2(t_{j+1}) - H_2(t_{j})) \nonumber \\
=&2 \sum_{j=0}^{n-1} (H_2(t_{j+1}) - H_2(t_{j}))^2 \left( \sum_{i=0}^{j-1} (H_1(t_{i+1}) - H_1(t_{i}))^2\right) \nonumber \\
&\quad + 4 \sum_{j=0}^{n-1}(\mu_D(t_{j+1}) - \mu_D(t_{j}))(H_2(t_{j+1}) - H_2(t_{j}))\left( \sum_{i=0}^{j-1} (\mu_D(t_{i+1}) - \mu_D(t_{i})) (H_1(t_{i+1}) - H_1(t_{i}))\right) \nonumber \\
&\xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} 0,
\end{align}
since
\begin{enumerate}
\item $\sum\limits_{i=0}^{j-1} (H_1(t_{i+1}) - H_1(t_{i}))^2 \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} [H_1]_{t_{j}}=0$,
\item $\sum\limits_{j=0}^{n-1} (H_2(t_{j+1}) - H_2(t_{j}))^2 \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} [H_2]_{t}=0$,
\item $\sum\limits_{i=0}^{j-1} (\mu_D(t_{i+1}) - \mu_D(t_{i})) (H_1(t_{i+1}) - H_1(t_{i})) \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} [\mu_D, H_1]_{t_j}=0$,
\item $\sum\limits_{j=0}^{n-1}(\mu_D(t_{j+1}) - \mu_D(t_{j}))(H_2(t_{j+1}) - H_2(t_{j})) \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} [\mu_D, H_2]_{t}=0$.
\end{enumerate}
The same argument can be applied to the case $j<i$. Hence,
\begin{equation}
\mathrm{Var} \left[ \sum_{j=0}^{n-1} \Delta_j^2 \right] \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{} 0,
\end{equation}
which means that
\begin{equation}
\sum_{j=0}^{n-1} \Delta_j^2 \xrightarrow[\max \, \abs{t_{u+1} - t_{u}} \rightarrow 0]{L^2} \int_{T_0}^t H_2(s) \, \mathrm{d} H_1(s) - \int_{T_0}^t H_1(s) \, \mathrm{d} H_2(s).
\end{equation}
\end{proof}
Recalling that, in the standard RAP case (see Definition \ref{standardRAP}), we have $g_1(x) = (H_1(x)H_2(T_0) - H_1(T_0)H_2(x))/(H_1(T_1)H_2(T_0) - H_1(T_0) H_2(T_1))$, $A_1(x)=g_1(x)H_2(T_1)$, as well as $ A_2(x)= (H_1(T_1)/H_2(T_1))H_2(x) - H_1(x)$, $\sigma_A^2(x)=A_1(x)A_2(x)$, we obtain the following expression for the SDE of the martingale $(M_t)_{t\in[T_0,T_1]}$.
\begin{coro}
Under the conditions of Proposition \ref{Ito}, if $(I_t^{(1)})$ is a standard $X$-RAP with driver covariance $K_D(x,y)=H_1(\min(x,y))H_2(\max(x,y))$, then
\begin{align}
&\, \mathrm{d} M_t= \frac{\mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ]}{H_1(T_1)H_2(t)-H_1(t)H_2(T_1)} \nonumber \\
&\times \left( \left( \frac{Z_t ( H_1'(t)H_2(T_1) - H_1(T_1) H_2'(t)) - M_t ( H_1'(t)H_2(t) - H_1(t) H_2'(t))}{H_1(T_1)H_2(t) - H_1(t) H_2(T_1) } - J_t \right) \, \mathrm{d} t + \, \mathrm{d} I^{(1)}_t \right),
\end{align}
where $Z_t= I_t^{(1)} -g_0(t)X_0 - \mu_A(t)$ and $J_t= X_0 g_0'(t) + \mu_A'(t)$.
\end{coro}
\begin{proof}
The result follows from the following calculations:
\begin{align}
&1.\quad\frac{g_1^2(t)(H_1'(t)H_2(t) - H_1(t) H_2'(t))}{2\sigma_A^4(t)} - \frac{ g_1(t) \left(\frac{g_1(t)}{\sigma_A(t)}\right)'}{\sigma_A(t)}=0,\\
&2.\quad M_t \sigma_A(t) \left(\frac{g_1(t)}{\sigma_A(t)}\right)' + Z_t \frac{\sigma_A(t)}{g_1(t)}\left( \left(\frac{g_1(t)}{\sigma_A(t)}\right)' + g_1(t)\left(\frac{1}{\sigma_A(t)}\right)'\right)\nonumber \\
&\hspace{8cm} - \frac{3g_1(t)M_t (H_1'(t)H_2(t) - H_1(t) H_2'(t))}{2\sigma_A^2(t)} \nonumber \\
&\quad\,\,\,=\frac{Z_t ( H_1'(t)H_2(T_1) - H_1(T_1) H_2'(t)) - M_t ( H_1'(t)H_2(t) - H_1(t) H_2'(t))}{H_1(T_1)H_2(t) - H_1(t) H_2(T_1) }.
\end{align}
\end{proof}
Introducing the notation
\begin{equation} \label{NSDE}
\, \mathrm{d} N_t:= \left( \frac{Z_t ( H_1'(t)H_2(T_1) - H_1(T_1) H_2'(t)) - M_t ( H_1'(t)H_2(t) - H_1(t) H_2'(t))}{H_1(T_1)H_2(t) - H_1(t) H_2(T_1) } - J_t \right) \, \mathrm{d} t + \, \mathrm{d} I^{(1)}_t,
\end{equation}
we can then write
\begin{equation}
M_t= X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ]}{H_1(T_1)H_2(t)-H_1(t)H_2(T_1)} \, \mathrm{d} N_t.
\end{equation}
As shown below, the process $(N_t)_{t\in [T_0,T_1]}$, defined by the above SDE \ref{NSDE} and initial condition $N_{T_0}=0$, is a martingale, and can be used to construct a \emph{standard Brownian motion on $[T_0,T_1]$}.
\begin{defn}
Let $I \subseteq \mathbb{R}^+$ be an interval or the real positive line. A stochastic process $(Y_t)_{t\in I}$ is said to be a standard Brownian motion on an interval $[a,b] \subseteq I$ if there exists a standard Brownian motion $(B_t)_{t \geqslant 0}$ such that $Y_t= B_{t-a}$ for all $t\in [a,b]$.
\end{defn}
\begin{prop} \label{propinnovations}
Under the conditions in Proposition \ref{Ito}, the process $(W_t)_{t \in [T_0,T_1]}$ defined by
\begin{equation} \label{innovations}
W_t= \int_{T_0}^t \frac{1}{H_1'(s)H_2(s) - H_1(s) H_2'(s)} \, \mathrm{d} N_s
\end{equation}
is standard Brownian motion on $[T_0,T_1]$, adapted to $(\mathcal{F}_t^I)$.
\end{prop}
\begin{proof}
We start by showing that $(N_t)_{t \in [T_0,T_1]}$ is an $(\mathcal{F}_t^I)$-martingale. We introduce the following notation: $h_1(t):= H_1'(t)H_2(T_1) - H_1(T_1) H_2'(t)$, $h_2(t):= H_1'(t)H_2(t) - H_1(t) H_2'(t)$, $h_3(t):= H_1(T_1)H_2(t) - H_1(t) H_2(T_1)$, and
$$
S(X_0,T_0,t):= \int_{T_0}^t \frac{(g_0(u) X_0 + \mu_A(u))h_1(u)}{h_3(u)} \, \mathrm{d} u.
$$
Then, it follows that
\begin{equation}
N_t = \int_{T_0}^t \frac{I^{(1)}_u h_1(u) - M_u h_2(u)}{h_3(u)} \, \mathrm{d} u- S(X_0,T_0,t) - g_0(t)X_0 - \mu_A(t) + I^{(1)}_t.
\end{equation}
Let $(s,t) \in [T_0,T_1] ^2$ such that $s < t$. We shall show that $\mathbb{E} [N_t \| \mathcal{F}_s^I ] = N_s$. By the linearity of the conditional expectation, we have
\begin{align} \label{master}
\mathbb{E} [N_t \| \mathcal{F}_s^I ]&= \int_{T_0}^s \frac{I^{(1)}_u h_1(u) - M_u h_2(u)}{h_3(u)} \, \mathrm{d} u + \int_{s}^t \frac{\mathbb{E}[I^{(1)}_u \| \mathcal{F}_s^I] h_1(u) }{h_3(u)} \, \mathrm{d} u - M_s\int_{s}^t \frac{ h_2(u)}{h_3(u)} \, \mathrm{d} u \nonumber \\
&\quad - (S(X_0,T_0,s) + S(X_0,s,t)) -g_0(t) X_0 - \mu_A(t) + \mathbb{E} [ I^{(1)}_t \| \mathcal{F}_s^I ].
\end{align}
By Proposition \ref{relations}, the last and the second terms in the above expression of $\mathbb{E} [N_t \| \mathcal{F}_s^I ]$ can be expressed in the following forms:
\begin{align}
\mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_s]&= \left(g_0(t) - \frac{A_2(t)}{A_2(s)} g_0(s) \right) X_0 + \left(g_1(t)- \frac{A_2(t)}{A_2(s)} g_1(s)\right) M_s\nonumber \\
& \quad \quad + \frac{A_2(t)}{A_2(s)} I_s^{(1)} + \mu_A(t) +\frac{A_2(t)}{A_2(s)} \mu_A(s),
\end{align}
\begin{align}
\int_{s}^t \frac{\mathbb{E}[I^{(1)}_u \| \mathcal{F}_s^I] h_1(u) }{h_3(u)} \, \mathrm{d} u &= X_0 \int_s^t \frac{g_0(u)h_1(u)}{h_3(u)} \, \mathrm{d} u + M_s\int_s^t \frac{g_1(u)h_1(u)}{h_3(u)} \, \mathrm{d} u + \int_s^t \frac{\mu_A(u) h_1(u)}{h_3(u)} \, \mathrm{d} u \nonumber \\
& \quad \quad+ \left( I_s^{(1)} - g_1(s)M_s - g_0(s)X_0 + \mu_A(s) \right) \frac{1}{A_2(s)}\int_s^t \frac{A_2(u) h_1(u)}{h_3(u)} \, \mathrm{d} u.
\end{align}
Moreover, we notice that
\begin{equation}
\frac{1}{A_2(s)}\int_s^t \frac{A_2(u) h_1(u)}{h_3(u)} \, \mathrm{d} u = \frac{H_1(t)-H_1(s) - \frac{H_1(T_1)}{H_2(T_1)} ( H_2(t) - H_2(s) )}{A_2(s)}
= 1- \frac{A_2(t)}{A_2(s)}.
\end{equation}
Hence,
\begin{align}
&\mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_s] + \int_{s}^t \frac{\mathbb{E}[I^{(1)}_u \| \mathcal{F}_s^I] h_1(u) }{h_3(u)} \, \mathrm{d} u \nonumber \\
&= (g_0(t)-g_0(s) ) X_0 + (g_1(t)-g_1(s)) M_s + X_0 \int_s^t \frac{g_0(u)h_1(u)}{h_3(u)} \, \mathrm{d} u + M_s\int_s^t \frac{g_1(u)h_1(u)}{h_3(u)} \, \mathrm{d} u\nonumber \\
&\quad \quad + \int_s^t \frac{\mu_A(u) h_1(u)}{h_3(u)} \, \mathrm{d} u +I_s^{(1)} + \mu_A(s) + \mu_A(t) \nonumber \\
&= (g_0(t)-g_0(s) ) X_0 + (g_1(t)-g_1(s)) M_s + S(X_0,s,t) + M_s\int_s^t \frac{g_1(u)h_1(u)}{h_3(u)} \, \mathrm{d} u +I_s^{(1)}\nonumber \\
&\quad \quad + \mu_A(s) + \mu_A(t).
\end{align}
Next, we observe that
\begin{align}
\int_s^t \frac{g_1(u)h_1(u)- h_2(u)}{h_3(u)} \, \mathrm{d} u &= \int_s^t \frac{H_1'(u)H_2(T_0) - H_1(T_0)H_2'(u)}{H_1(T_1)H_2(T_0) - H_1(T_0) H_2(T_1)}\, \mathrm{d} u= g_1(t)-g_1(s),
\end{align}
which allows one to write
\begin{align} \label{work}
&\mathbb{E}[I_t^{(1)} \| \mathcal{F}^I_s] + \int_{s}^t \frac{\mathbb{E}[I^{(1)}_u \| \mathcal{F}_s^I] h_1(u) }{h_3(u)} \, \mathrm{d} u \nonumber \\
&= (g_0(t)-g_0(s) ) X_0 + M_s\int_{s}^t \frac{ h_2(u)}{h_3(u)} \, \mathrm{d} u M_s + S(X_0,s,t) +I_s^{(1)} + \mu_A(s) + \mu_A(t).
\end{align}
Inserting Eq. (\ref{work}) into Eq. (\ref{master}) concludes the proof:
\begin{equation}
\mathbb{E} [N_t \| \mathcal{F}_s^I ] = \int_{T_0}^s \frac{I^{(1)}_u h_1(u) - M_u h_2(u)}{h_3(u)} \, \mathrm{d} u- S(X_0,T_0,s) - g_0(s)X_0 - \mu_A(s) + I^{(1)}_s=N_s.
\end{equation}
Moreover, $\mathbb{E} [ \, \abs{N_t} ] < \infty$ for all $t \in [T_0,T_1]$. Hence $(N_t)$ is a martingale with respect to $(\mathcal{F}_t^I)$. We now show that $(W_t)$ is an $(\mathcal{F}_t^I)$-adapted standard Brownian motion on $[T_0,T_1]$. We compute the quadratic variation of $(W_t)$ and obtain
\begin{equation}
[W]_t= \int_{T_0}^t \frac{1}{h_2(s)} \, \mathrm{d} [N]_s = \int_{T_0}^t \frac{1}{h_2(s)} \, \mathrm{d} [I^{(1)}]_s = \int_{T_0}^t \frac{h_2(s)}{h_2(s)} \, \mathrm{d} s = t-T_0.
\end{equation}
Since $(N_t)$ is an $(\mathcal{F}_t^I)$-martingale and $\mathbb{E}[[W]_t] < \infty$, $(W_t)$ is an $(\mathcal{F}_t^I)$-martingale. Hence, by Lévy's characterization theorem, $(W_t)$ is an $(\mathcal{F}_t^I)$-adapted standard Brownian motion on $[T_0,T_1]$.
\end{proof}
The process $(W_t)$ is referred to as the innovations process of $(M_t)$. It follows that we can write $(M_t)$ as an integral with respect its innovations process:
\begin{coro} \label{FAMSDEfinal}
Under the conditions in Proposition \ref{Ito}, it holds that
\begin{equation} \label{GMFAM}
M_t = X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_s^{(1)} ] \sqrt{H_1'(s)H_2(s) - H_1(s) H_2'(s) }}{H_1(T_1)H_2(s)-H_1(s)H_2(T_1)} \, \mathrm{d} W_s.
\end{equation}
\end{coro}
In the case that $(I^{(1)}_t)$ is an $X$-randomized anticipative Brownian bridge on $[T_0,T_1]$, that is,
\begin{equation}
I_{t}^{(1)}= B_t - \frac{T_1-t}{T_1-T_0} (B_{T_0}-X_0) - \frac{t-T_0}{T_1-T_0} (B_{T_1}-X_1),
\end{equation}
where $(B_t)$ is a standard Brownian motion, the expressions become significantly simpler.
\begin{coro} \label{Ito2}
If $(I^{(1)}_t)_{t \in [T_0,T_1]}$ is an $X$-randomized anticipative Brownian bridge on $[T_0,T_1]$, the following holds:
\bigskip
1.\quad The process $(M_t)_{t\in[T_0,T_1]}$ satisfies the SDE
\begin{equation}
M_t = X_0+ \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ]}{T_1-t} \left( \frac{ I_t^{(1)} - M_t }{T_1-t} \, \mathrm{d} t+ \, \mathrm{d} I^{(1)}_t \right).
\end{equation}
2.\quad The process $(W_t)_{t \in [T_0,T_1]}$, defined by
\begin{equation}
W_t := \int_{T_0}^t\frac{ I_u^{(1)} - M_u }{T_1-u} \, \mathrm{d} u + I^{(1)}_t - X_0,
\end{equation}
is an $(\mathcal{F}_t^I)$-adapted standard Brownian motion on $[T_0,T_1]$.
\end{coro}
\begin{proof}
It suffice to set $H_1(x)=x, H_2(x)=1, g_0(t)=\frac{T_1-t}{T_1-T_0}$ and $g_1(t)=\frac{t-T_0}{T_1-T_0}$ in Eqs. \ref{GMFAM} and \ref{innovations}.
\end{proof}
For the case where $X_0=0$, the random variable $X_1$ is centered around $0$, and $T_0=0$, then the randomized anticipative Brownian bridge gives rise to the martingale developed in \cite{BHM1}, i.e,
\begin{equation}
M_t= \int_{0}^t \frac{\mathrm{Var} [ X_1 \| I_s^{(1)} ]}{T_1-s} \, \mathrm{d} W_s.
\end{equation}
Using Eq. (\ref{GMFAM}), we give examples of FAMs where the underlying RAP is standard (see Definition \ref{standardRAP}).
\begin{ex}
Let $D_t=tB_t$ where $(B_t)_{t \geqslant0}$ is a standard Brownian motion. Then, $K_D(x,y)= \min(x,y)^2 \max(x,y)$, and so $H_1(x)=x^2$, $H_2(x)=x$. The standard AP driven by $(D_t)$ is given by
\begin{equation}
A_t^{(1)}= D_t - \frac{T_1t-t^2}{T_1T_0 - T_0^2} D_{T_0} - \frac{t^2-tT_0}{T_1^2 - T_1T_0} D_{T_1}.
\end{equation}
We have
\begin{equation}
K_A(x,y)= \frac{\min(x,y)(\min(x,y)-T_0) \max(x,y) (T_1-\max(x,y))}{T_1-T_0},
\end{equation}
hence $A_1(x)= \frac{x(x-T_0)}{T_1-T_0}$ and $A_2(x)=x(T_1-x)$. A standard $X$-RAP with noise process $(A_t^{(1)})$ is
\begin{equation}
I_t^{(1)}= D_t - \frac{T_1t-t^2}{T_1T_0 - T_0^2} (D_{T_0} -X_0) - \frac{t^2-tT_0}{T_1^2 - T_1T_0} (D_{T_1}-X_1).
\end{equation}
We recall that a different interpolating coefficient for $X_0$ could have been chosen without disrupting the standard property of $(I_t^{(1)})$. The quadratic variation of $(I_t^{(1)})$ is given by $\, \mathrm{d} [I^{(1)}]_t= H_1'(t)H_2(t) - H_1(t) H_2'(t) \, \mathrm{d} t= t^2 \, \mathrm{d} t$. Hence,
\begin{equation}
M_t = X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ] s }{T_1^2 s - s^2 T_1} \, \mathrm{d} W_s = X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_s^{(1)} ] }{T_1^2 - s T_1} \, \mathrm{d} W_s.
\end{equation}
\end{ex}
\begin{ex}
Let $(D_t)$ be the Ornstein-Uhlenbeck process satisfying the SDE
\begin{equation}
D_{t}=\int_0^t\theta\left(\mu-D_{s}\right) \, \mathrm{d} s+\int_0^t\sigma \, \mathrm{d} B_{s},
\end{equation}
where $\theta>0, \sigma>0, \mu \in \mathbb{R}$. Then $ K_D(s,t)=\frac{\sigma^2}{2 \theta} \mathrm{e}^{\theta \min(s,t)} \mathrm{e}^{-\theta \max(s,t)}$, and so $H_1(x)=\frac{\sigma^2}{2 \theta} \mathrm{e}^{\theta x}$, $H_2(x)=\mathrm{e}^{-\theta x}$. The standard AP driven by $(D_t)$ is
\begin{equation}
A_{t}^{(1)}=D_t-\frac{\mathrm{e}^{\theta (T_1-t)}-\mathrm{e}^{-\theta (T_1-t)}}{\mathrm{e}^{\theta (T_1-T_0)}-\mathrm{e}^{-\theta (T_1-T_0)}} D_{T_0} - \frac{\mathrm{e}^{\theta (t-T_0)}-\mathrm{e}^{-\theta (t-T_0)}}{\mathrm{e}^{\theta (T_1-T_0)}-\mathrm{e}^{-\theta (T_1-T_0)}} D_{T_1}.
\end{equation}
A standard $X$-RAP with noise process $(A_t^{(1)})$ is
\begin{equation}
I_t^{(1)}= D_t - \frac{\mathrm{e}^{\theta (T_1-t)}-\mathrm{e}^{-\theta (T_1-t)}}{\mathrm{e}^{\theta (T_1-T_0)}-\mathrm{e}^{-\theta (T_1-T_0)}} (D_{T_0} -X_0) - \frac{\mathrm{e}^{\theta (t-T_0)}-\mathrm{e}^{-\theta (t-T_0)}}{\mathrm{e}^{\theta (T_1-T_0)}-\mathrm{e}^{-\theta (T_1-T_0)}} (D_{T_1}-X_1).
\end{equation}
The quadratic variation of $(I_t^{(1)})$ is given by $\, \mathrm{d} [I^{(1)}]_t= H_1'(t)H_2(t) - H_1(t) H_2'(t) \, \mathrm{d} t= \sigma^2 \, \mathrm{d} t$. Hence,
\begin{equation}
M_t = X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ] \sigma }{\frac{ \sigma^2}{2 \theta} \left(\mathrm{e}^{\theta(T_1-s)} - \mathrm{e}^{\theta(s-T_1)} \right)} \, \mathrm{d} W_s = X_0 + \frac{2\theta}{\sigma}\int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_t^{(1)} ] }{ \mathrm{e}^{\theta(T_1-s)} - \mathrm{e}^{\theta(s-T_1)} } \, \mathrm{d} W_s.
\end{equation}
\end{ex}
In general, one can expect that for a given coupling $\pi^X$, $M_t=\mathbb{E}[X_1 \| \mathcal{F}_t^I]$ has no explicit analytical expression, even in the case $M_t=\mathbb{E}[X_1 \| X_0, I_t^{(1)}]$. Nonetheless, we give an example where $M_t=\mathbb{E}[X_1 \| X_0, I_t^{(1)}]$ and its SDE are explicit.
\begin{ex}
Let $X_0 \sim \textrm{Uni}\,(\{-1,1\})$ and $X_1 \sim \textrm{Uni}\,(\{-2,0,2\})$ such that
\begin{equation}
X_1 \| X_0 = \left\{
\begin{array}{ll}
X_0 +1 & \mbox{with probability } \frac{1}{2},\\ \\
X_0 - 1 & \mbox{with probability } \frac{1}{2}.
\end{array}
\right.
\end{equation}
Clearly, $\mathbb{E}[X_1 \| X_0 ] = X_0$. Let $(B_t)_{t \geqslant0}$ be a standard Brownian motion and
\begin{equation}
I_{t}^{(1)}= B_t - \frac{T_1-t}{T_1-T_0} (B_{T_0}-X_0) - \frac{t-T_0}{T_1-T_0} (B_{T_1}-X_1).
\end{equation}
Denoting by $\phi$ the density function of the measure $\mathcal N\left(0, \frac{(T_1-t)(t-T_1)}{T_1-T_0}\right)$, and recalling that $g_1(x)= \frac{t-T_0}{T_1-T_0}$, we have, by Proposition \ref{Bayes},
\begin{align}
M_t&= \frac{(X_0+1)\phi(I_t^{(1)}-X_0-g_1(t)) \frac{1}{2} + (X_0-1)\phi(I_t^{(1)}-X_0+g_1(t)) \frac{1}{2}}{\phi(I_t^{(1)}-X_0-g_1(t)) \frac{1}{2} + \phi(I_t^{(1)}-X_0+g_1(t)) \frac{1}{2}}, \\
&= X_0 + \frac{\phi(I_t^{(1)}-X_0-g_1(t)) - \phi(I_t^{(1)}-X_0+g_1(t))}{\phi(I_t^{(1)}-X_0-g_1(t)) + \phi(I_t^{(1)}-X_0+g_1(t))}, \\
&= X_0 + \tanh \left( \frac{I_t^{(1)}-X_0}{T_1-t}\right).
\end{align}
The associated SDE can be expressed by
\begin{equation}
M_t= X_0+ \bigintss_{T_0}^t \frac{ \sech^2 \left( \frac{I_s-X_0}{T_1-s} \right) }{T_1-s} \, \mathrm{d} W_s.
\end{equation}
\begin{figure}[h]
\centering
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{RAPdiscreteUnifrom.png}
\caption{Paths of the $X$-RAP $(I_{t}^{(1)})$.}
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{FAMdiscreteUniform.png}
\caption{Paths of the associated FAM $(M_t)$.}
\end{minipage}
\end{figure}
\end{ex}
\subsection{The n-arc FAM}
Let $X=(X_0,X_1,\ldots, X_n)$ be a real random vector and $(\mu_0,\mu_1, \ldots, \mu_n) \in \mathcal P (\mathbb{R}) \times \mathcal P (\mathbb{R}) \times \ldots \times \mathcal P (\mathbb{R})$ such that $X_i \sim \mu_i$ for $i=0,\ldots, n$ and $\mu_{i}\leqslant_{\mathrm{cx}}\mu_{i+1}$ for $i=0,\ldots, n-1$. Given a martingale coupling $\pi^X \in \mathcal M(\mu_0,\mu_1, \ldots, \mu_n)$ for $X$ and an $X$-RAP $(I_t^{(n)})_{t \in [T_0,T_n]}$ on the partition $\{T_0,T_n\}_*$, we aim at constructing an $(\mathcal{F}^I_t)$-martingale $(M_t)_{t\in[T_0, T_n]}$ such that $M_{T_i}\overset{a.s.}{=} X_i$, for $i=0,\ldots, n$. A major difference in the $n$-arc case is that $(I_t^{(n)})$ must be $\{T_0,T_n\}_*$-nearly-Markov to guarantee the interpolation of $(M_t)$.
\begin{defn} \label{nFAM}
Given an $X$-RAP $(I_t^{(n)})_{t \in [T_0,T_n]}$ that is $\{T_0,T_n\}_*$-nearly-Markov, an $n$-arc FAM for $X$ is a stochastic process of the form $M_t = \mathbb{E}[X_n \| \mathcal F_t^I] = \mathbb{E}[ X_n \| X_0, \ldots, X_{m(t)}, I_t^{(n)} ]$ where $m(t)= \max \{i \in \mathbb{N} \| T_i \leqslant t \}$.
\end{defn}
By the tower property of conditional expectation, $(M_t)_{t \in [T_0,T_n]}$ is a martingale with respect to $(\mathcal{F}_t^I)$. Moreover, $M_{T_i}= \mathbb{E}[ X_n \| X_0, \ldots, X_{i} ] = X_i $ since $\pi^X \in \mathcal M(\mu_0,\mu_1, \ldots, \mu_n)$. Hence $(M_t)$ is an interpolating martingale with respect to $(\mathcal{F}_t^I)$. We notice that, without the nearly-Markov property, $M_{T_i}= \mathbb{E}[X_n \| \mathcal F_{T_i}^I]$ is not necessarily equal to $X_i$. As for the case $n=1$, in the remainder of this section, we assume that the driver of $(I_t^{(n)})$ has a density function on $[T_0,T_n]$.
\begin{prop}
Let $M_t = \mathbb{E}[ X_n \| X_0, \ldots, X_{m(t)}, I_t^{(n)} ]$ be an $n$-arc FAM restricted to $(T_0,T_n)_*$. Then,
\begin{equation}
M_t = \frac{\int_\mathbb{R} y f^{I_t^{(n)} \| X_0,\ldots,X_{m(t)},X_n=y} (I_t^{(n)}) \, \mathrm{d} F^{X_n \| X_{0}, \ldots, X_{m(t)}}}{\int_\mathbb{R} f^{I_t^{(n)} \| X_0,\ldots,X_{m(t)},X_n=y} (I_t^{(n)}) \, \mathrm{d} F^{X_n \| X_{0}, \ldots, X_{m(t)}}},
\end{equation}
where $F^{X_n \| X_{0}, \ldots, X_{m(t)}}$ is the distribution function of $X_n$ conditional on random vector $(X_{0}, \ldots, X_{m(t)})$. In particular:
\begin{enumerate}
\item If $(X_{0}, \ldots, X_{m(t)},X_n)$ is a continuous random variable, with density function $f^{X_n \| X_{0}, \ldots, X_{m(t)}}$, then
\begin{equation}
M_t = \frac{\int_\mathbb{R} y f^{I_t^{(n)} \| X_0,\ldots,X_{m(t)},X_n=y} (I_t^{(n)}) f^{X_n \| X_{0}, \ldots, X_{m(t)}}(y) \, \mathrm{d} y}{\int_\mathbb{R} f^{I_t^{(n)} \| X_0,\ldots,X_{m(t)},X_n=y} (I_t^{(n)}) f^{X_n \| X_{0}, \ldots, X_{m(t)}}(y) \, \mathrm{d} y}.
\end{equation}
\item If $(X_{0}, \ldots, X_{m(t)},X_n)$ is a discrete random variable, then
\begin{equation}
M_t = \frac{\sum\limits_{y} y f^{I_t^{(n)} \| X_{0}, \ldots, X_{m(t)},X_n=y} (I_t^{(n)}) \mathbb{P}[X_n=y \| X_{0}, \ldots, X_{m(t)}] }{\sum\limits_{y} f^{I_t^{(n)} \|X_{0}, \ldots, X_{m(t)},X_n=y} (I_t^{(n)}) \mathbb{P}[X_n=y \| X_{0}, \ldots, X_{m(t)}]} .
\end{equation}
\end{enumerate}
\end{prop}
The proof follows from the one-arc case. Introducing the notation
\begin{equation}
u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y)= f^{I_t^{(n)} \| X_0,\ldots,X_{m(t)}, X_{m(t)+1}= x_{m +1}, \ldots, X_{n-1}=x_{n-1}, X_n=y} (I_t^{(n)}),
\end{equation}
it is often more convenient to write
\begin{equation}
M_t = \frac{\int_{\mathbb{R}^{n-m(t)}} y\, u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} F^{X_{m(t)+1}, \ldots, X_{n-1}, X_n \| X_{0}, \ldots, X_{m(t)}} ( x_{m +1}, \ldots,x_{n-1},y)}{\int_{\mathbb{R}^{n-m(t)}} u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} F^{X_{m(t)+1}, \ldots, X_{n-1}, X_n \| X_{0}, \ldots, X_{m(t)}} ( x_{m +1}, \ldots,x_{n-1},y)}.
\end{equation}
We can apply Itô's lemma in the same fashion as in the one-arc case. An interesting pattern appears if the signal function of $(I_t^{(n)})$ satisfies $g_j(x) \mathds{1}_{[T_0,T_{j-1}]}(x)=0$ for all $ j=1,\ldots, n$, and for all $ x \in [T_0,T_n]$, that significantly simplifies the $n$-arc case.
\begin{prop} \label{special}
Let $M_t = \mathbb{E}[ X_n \| X_0, \ldots, X_{m(t)}, I_t^{(n)} ]$ be an $n$-arc FAM, where the signal function of $(I_t^{(n)})$ satisfies $g_j(x) \mathds{1}_{[T_0,T_{j-1}]}(x)=0$ for all $ j=1,\ldots, n$, and for all $ x \in [T_0,T_n]$. Then, $M_t=\mathbb{E}[ X_{m(t)+1} \| X_0, \ldots, X_{m(t)}, I_t^{(n)} ]$.
\end{prop}
\begin{proof}
The result is trivial for $t \in \{T_0,T_n\}_*$, so we treat the case $t\in (T_0,T_n)_*$. Let $\varphi$ be the density function of $(A_t^{(n)})$, the noise process of $(I_t^{(n)})$. Then,
\begin{align}
u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) &= \varphi\left( I_t^{(n)} - \sum_{i=0}^{m(t)} g_i(t) X_i - g_{m(t)+1}(t) x_{m+1})\right),
\end{align}
since $g_j(x) \mathds{1}_{[T_0,T_{j-1}]}(x)=0$ for all $ j=1,\ldots, n$, and for all $ x \in [T_0,T_n]$. Denoting by $\tilde F$ the conditional distribution of $(X_{m(t)+1},\ldots, X_n)$ given $(X_0,\ldots, X_{m(t)})$, i.e.,
\begin{equation}
\tilde F(x_{m+1},\ldots, x_n)= F^{X_{m(t)+1}, \ldots, X_n | X_0,\ldots, X_{m(t)}} (x_{m+1},\ldots, x_n),
\end{equation}
it implies
\begin{align}
&\int_{\mathbb{R}^{n-m(t)}} u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} \tilde F (x_{m+1},\ldots, x_{n-1},y) \nonumber \\
&\hspace{3cm} =\int_\mathbb{R} u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} F^{X_{m(t)+1} | X_0,\ldots, X_{m(t)}} (x_{m+1}).
\end{align}
We also have
\begin{align}
&\int_{\mathbb{R}^{n-m(t)}} y \, \mathrm{d} \tilde F (x_{m+1},\ldots, x_n) \nonumber \\
&= \int_{\mathbb{R}^{n-m(t)}} y \, \mathrm{d} F^{X_n| X_0, \ldots, X_{m(t)}, X_{m(t)+1}=x_{m+1},\ldots, X_{n-1}=x_{n-1} } (y) \nonumber \\
& \hspace{5cm} \, \mathrm{d} F^{X_{m(t)+1},\ldots, X_{n-1} | X_0, \ldots, X_{m(t)} }(x_{m+1},\ldots, x_{n-1} ) \nonumber\\
&= \int_{\mathbb{R}^{n-m(t)-1}} \mathbb{E}[X_n \| X_0, \ldots, X_{m(t)}, X_{m(t)+1}=x_{m+1},\ldots, X_{n-1}=x_{n-1} ] \nonumber \\
& \hspace{5cm} \, \mathrm{d} F^{X_{m(t)+1},\ldots, X_{n-1} | X_0, \ldots, X_{m(t)} }(x_{m+1},\ldots, x_{n-1} ) \nonumber\\
&= \int_{\mathbb{R}^{n-m(t)-1}} x_{n-1} \, \mathrm{d} F^{X_{m(t)+1},\ldots, X_{n-1} | X_0, \ldots, X_{m(t)} }(x_{m+1},\ldots, x_{n-1} ),
\end{align}
where we used the martingale property of $(M_t)$. Applying the same argument $m(t)$ times, one obtains
\begin{align}
&\int_{\mathbb{R}^{n-m(t)}} y \, \mathrm{d} \tilde F (x_{m+1},\ldots, x_n) = \int_{\mathbb{R}} x_{m+1} \, \mathrm{d} F^{X_{m(t)+1} | X_0, \ldots, X_{m(t)}} (x_{m+1}).
\end{align}
Thus,
\begin{align}
&\int_{\mathbb{R}^{n-m(t)}} y u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} \tilde F (x_{m+1},\ldots, x_{n-1},y) \nonumber \\
&\hspace{2cm}= \int_{\mathbb{R}} x_{m+1}u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) d F^{X_{m(t)+1} | X_0, \ldots, X_{m(t)}} (x_{m+1}).
\end{align}
So, we have
\begin{align}
M_t &= \frac{\int_{\mathbb{R}^{n-m(t)}} y u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} \tilde F (x_{m+1},\ldots, x_{n-1},y)}{\int_{\mathbb{R}^{n-m(t)}} u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} \tilde F (x_{m+1},\ldots, x_{n-1},y)} \nonumber \\ \nonumber \\
&= \frac{\int_{\mathbb{R}} x_{m+1}u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} F^{X_{m(t)+1} | X_0, \ldots, X_{m(t)}} (x_{m+1})}{\int_\mathbb{R} u(I_t^{(n)},t,x_{m +1},\ldots,x_{n-1},y) \, \mathrm{d} F^{X_{m(t)+1} | X_0,\ldots, X_{m(t)}} (x_{m+1})} \nonumber \\ \nonumber \\
&= \int_{\mathbb{R}} x_{m+1} \, \mathrm{d} F^{X_{m(t)+1} | X_0, \ldots, X_{m(t)}, I_t^{(n)}} (x_{m+1})\nonumber \\ \nonumber \\
&= \mathbb{E}[ X_{m(t)+1} \| X_0, \ldots, X_{m(t)}, I_t^{(n)} ],
\end{align}
where we use the Bayes rule in the second-last step.
\end{proof}
This allows one to use the one-arc case to derive the SDE for the $n$-arc FAM.
\begin{prop}
Let $(I_t^{(n)})$ be a semimartingale standard RAP satisfying the condition in Proposition \ref{special}, with driver covariance $K_D(x,y)=H_1(\min(x,y))H_2(\max(x,y))$, such that $(t,x) \rightarrow \mathbb{E} [ X_n \| X_0, \ldots, X_{m(t)}, I_t^{(n)}=x ]$ is $C^2(((T_0,T_n)_* \setminus N) \times Im(I_t^{(n)}))$ where $N \subset (T_0,T_n)_*$ contains finitely many elements. Let $M_t = \mathbb{E} [ X_n \| X_0, \ldots, X_{m(t)}, I_t^{(n)} ]$ be an $n$-arc FAM. Then, the follwing holds:
\bigskip
1. \quad The process $(M_t)_{t\in[T_0,T_n]}$ satisfies the equation
\begin{equation}
M_t = X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_{m(s)+1} \| X_0, \ldots, X_{m(s)}, I_t^{(n)} ] \sqrt{H_1'(s)H_2(s) - H_1(s) H_2'(s))}}{H_1(T_{m(s)+1})H_2(s)-H_1(s)H_2(T_{m(s)+1})} \, \mathrm{d} W_s.
\end{equation}
where
\begin{equation}
W_t = \int_{T_0}^t \frac{1}{H_1'(s)H_2(s) - H_1(s) H_2'(s)} \, \mathrm{d} N_s,
\end{equation}
\begin{align}
\, \mathrm{d} N_t&=\bigg ( \frac{Z_t ( H_1'(t)H_2(T_{m(t)+1}) - H_1(T_{m(t)+1}) H_2'(t)) - M_t ( H_1'(t)H_2(t) - H_1(t) H_2'(t))}{H_1(T_{m(t)+1})H_2(t) - H_1(t) H_2(T_{m(t)+1}) } - J_t \bigg ) \, \mathrm{d} t \nonumber \\
&\quad+ \, \mathrm{d} I^{(1)}_t,
\end{align}
and $Z_t= I_t^{(1)} - \sum\limits_{i=0}^{m(t)} g_i(t)X_i - \mu_A(t)$, and $J_t=\sum\limits_{i=0}^{m(t)} g_i'(t)X_i + \mu_A'(t)$.
\bigskip
2. \quad The process $(W_t)_{t\in[T_0,T_n]}$ is an $(\mathcal F^I_t)$-adapted standard Brownian motion on $[T_0,T_n]$.
\end{prop}
The proof follows from the one-arc case. Moreover, we note that
\begin{align}
M_t &= X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_{m(s)+1} \| X_0, \ldots, X_{m(s)}, I_t^{(n)} ] \sqrt{H_1'(s)H_2(s) - H_1(s) H_2'(s))}}{H_1(T_{m(s)+1})H_2(s)-H_1(s)H_2(T_{m(s)+1})} \, \mathrm{d} W_s, \\
&= X_{m(t)} + \int_{T_{m(t)}}^t \frac{\mathrm{Var} [ X_{m(s)+1} \| X_0, \ldots, X_{m(s)}, I_t^{(n)} ] \sqrt{H_1'(s)H_2(s) - H_1(s) H_2'(s))}}{H_1(T_{m(s)+1})H_2(s)-H_1(s)H_2(T_{m(s)+1})} \, \mathrm{d} W_s.
\end{align}
\subsection{Information-based martingale optimal transport}
Let $(B_t)_{t\geqslant0}$ be a standard Brownian motion. For now, to compare with other interpolating methods, we start with the standard one-arc $X$-RAP $(I_t^{(1)})_{t\in[T_0,T_1]}$ given by
\begin{equation}
I_{t}^{(1)}= B_t - \frac{T_1-t}{T_1-T_0} (B_{T_0}-X_0) - \frac{t-T_0}{T_1-T_0} (B_{T_1}-X_1).
\end{equation}
We remark that the $n$-arc case, with a general standard RAP, can be treated similarly. We consider the FAM
\begin{equation}
M_t= \mathbb{E}[ X_1 \| X_0, I_t^{(1)}] = X_0+ \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_s^{(1)} ]}{T_1-s} \, \mathrm{d} W_s,
\end{equation}
where $(W_t)$ is the innovations process of $(M_t)$, see Eq. (\ref{innovations}), a standard Brownian motion $[T_0,T_1]$ adapted to the filtration $(\mathcal{F}_t^I)$ generated by $(I_{t}^{(1)})$. The FAM $(M_t)$ depends heavily on the coupling $\pi^X$ of $(X_0,X_1)$. In real-world problems, the coupling is usually not observed directly, only its marginals are. Choosing a coupling is then part of the modelling task determined by the problem at hand. Another approach is to go ``model-free'' by utilizing a least action principle, such as martingale optimal transport (MOT).
\bigskip
Optimal transport (OT) dates back to Gaspard Monge in 1781 \cite{Monge}, with significant advancements by Leonid Kantorovich in 1942 \cite{Kantorovich} and Yann Brenier in 1987 \cite{Brenier}. It provides a way of comparing two measures, $\mu$ and $\nu$, defined on the Borel sets of topological Hausdorff spaces $\mathcal{X}$ and $\mathcal{Y}$, respectively. We denote by $\mathcal P(\mathcal{X})$ and $\mathcal P(\mathcal{Y})$ the sets of Borel measures on $\mathcal{X}$ and $\mathcal{Y}$ respectively. A popular mental image in the context of optimal transport is a pile of sand, modelled by a measure $\mu$, and a hole, modelled by another measure $\nu$. One wishes to fill the hole with the sand at one's disposal in an optimal manner, by exercising the least amount of effort. To make this statement more precise, one needs a cost function $c: \mathcal{X} \times \mathcal{Y} \rightarrow [0,\infty]$ that measures the cost of transporting a unit mass from $x \in \mathcal{X}$ to $y\in\mathcal{Y}$. The optimal transport problem is concerned with how to transport $\mu$ to $\nu$ whilst minimizing the cost of transportation. That is, given $\mu \in \mathcal{P}(\mathcal{X})$ and $\nu \in \mathcal{P}(\mathcal{Y})$,
\begin{equation}
\inf_{\pi \in \Pi(\mu,\nu)} \textup{\textbf{K}}(\pi) := \inf_{\pi \in \Pi(\mu,\nu)} \int_{\mathcal{X} \times \mathcal{Y}} c(x,y) \, \mathrm{d} \pi(x,y).
\end{equation}
This problem enjoys many interesting properties. For instance, when $\mathcal{X}=\mathcal{Y}$ is a Polish space, and $(\mathcal{X},c)$ is a metric space, then
\begin{equation}
W_{p}(\mu, \nu):=\left(\inf _{\pi \in \Pi(\mu, \nu)} \int_{\mathcal{X} \times \mathcal{X}} c(x, y)^{p} \mathrm{d} \pi(x, y)\right)^{1 / p}
\end{equation}
defines a metric for any $p\geqslant 1$, the Wasserstein $p$-metric, on the space $\mathcal P_p(\mathcal{X})$ of probability measures on $\mathcal{X}$ with finite $p$th moment. Furthermore, if $\mathcal{X}$ is Euclidean, and $c(x,y)=\norm{x-y}$, there is a one-to-one correspondence between the minimizers $\pi^*$ of $\inf\limits_{\pi \in \Pi(\mu,\nu)} \int_{\mathcal{X} \times \mathcal{Y}} c(x,y)^p \, \mathrm{d} \pi(x,y)$, and the geodesics in $(\mathcal P_p(\mathcal{X}),W_{p})$: If $(X_0,X_1) \sim \pi^*$, the law of the process
\begin{equation}
\left(\frac{T_1-t}{T_1-T_0} X_0 + \frac{t-T_0}{T_1-T_0} X_1\right)_{t\in[T_0,T_1]}
\end{equation}
is the shortest paths from $\mu$ to $\nu$ in $(\mathcal P_p(\mathcal{X}),W_{p})$. This links optimal transport to interpolation on the space of random variables. The interpolating process is basic, since it is deterministic conditionally on $(X_0,X_1)$. This is because OT does not incorporate noise in its formulation. If one wishes to introduce noise in the system, a solution could be to add an independent Brownian bridge on $[T_0,T_1]$ to $\frac{T_1-t}{T_1-T_0} X_0 + \frac{t-T_0}{T_1-T_0} X_1$. This is trivial on some level, since the noise term, here the Brownian bridge, is not related with the initial optimization problem. We call such interpolator artificially noisy. If one wants to interpolate while a noise term is involved within the optimal transport framework, the entropic regularization of optimal transport provides a way forward. Given $\mu \in \mathcal{P}(\mathcal{X})$, $\nu \in \mathcal{P}(\mathcal{Y})$, and $\varepsilon >0$, one considers
\begin{equation}
\inf_{\pi_\varepsilon \in \Pi(\mu,\nu)} \textup{\textbf{K}}_\varepsilon(\pi_\varepsilon) := \inf_{\pi_\varepsilon \in \Pi(\mu,\nu)} \int_{\mathcal{X} \times \mathcal{Y}} c(x,y) \, \mathrm{d} \pi_\varepsilon(x,y) + \varepsilon \textup{ KL}(\pi_\varepsilon \| \mu \otimes \nu )
\end{equation}
where$\textup{ KL}$ is the Kullback-Leibler divergence
\begin{equation}
\textup{ KL}(\pi_\varepsilon \| \mu \otimes \nu ) = \left\{
\begin{array}{ll}
\displaystyle{\int_{\mathcal{X} \times \mathcal{Y}} \log \left(\frac{\, \mathrm{d}\pi_\varepsilon}{\, \mathrm{d} \mu \, \mathrm{d} \nu} (x,y)\right)} \, \mathrm{d}\pi_\varepsilon (x,y), & \mbox{if } \pi_\varepsilon<< \mu \otimes \nu, \\
\\
\infty & \mbox{otherwise.} \\
\end{array}
\right.
\end{equation}
In the case where $\mathcal{X}=\mathcal{Y}$ is Euclidean, $c(x,y)=\norm{x-y}^2$, and $\varepsilon = 2 (T_1-T_0)$, this problem is equivalent to Schrödinger's problem \cite{Schrodinger}. If $\pi_S^*$ is the solution to Schrödinger's problem, any stochastic process with cumulative distribution function
\begin{equation}
F_S(\cdot)=\int_{\mathcal{X}}\int_{\mathcal{X}} F(x,y,\cdot) \, \mathrm{d}\pi_S^* (x,y),
\end{equation}
where $F(x,y,\cdot)$ is the distribution function of a Brownian bridge starting at time $T_0$ in value $x$ and ending at time $T_1$ in value $y$, is called a Schrödinger bridge. Hence, $(I_{t}^{(1)})$ is an anticipative representation of a Schrödinger bridge (in dimension one, since we did not introduce a definition of RAPs in higher dimensions) as long as $(X_0,X_1) \sim \pi_S^*$. We call this interpolator genuinely noisy, as opposed to artificially noisy, because, in this case, the noise is taken into account in the optimization problem. We recap the three interpolators built using optimal transport in dimension one:
\begin{enumerate}
\item The default OT interpolator, or, the shortest path interpolator on $[T_0,T_1]$: $\frac{T_1-t}{T_1-T_0} X_0 + \frac{t-T_0}{T_1-T_0} X_1$ where $(X_0,X_1) \sim \pi^*$.
\item The artificially noisy OT interpolator on $[T_0,T_1]$: $(I_{t}^{(1)})$, where $(X_0,X_1) \sim \pi^*$.
\item The genuinely noisy OT interpolator, or, Schrödinger's bridge on $[T_0,T_1]$: $(I_{t}^{(1)})$, where $(X_0,X_1) \sim \pi_S^*$.
\end{enumerate}
To re-design these interpolators to suit a martingale setup, let us fix $\mathcal{X}=\mathcal{Y}=\mathbb{R}$, which is the FAM setting. We can adapt optimal transport to yield a martingale coupling instead: given $\mu,\nu \in \mathcal{P}_1(\mathbb{R})$ in convex order,
\begin{equation}
\inf_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}(\pi) = \inf_{\pi \in \mathcal M (\mu,\nu)} \int_{\mathbb{R}^2} c(x,y) \, \mathrm{d} \pi(x,y),
\end{equation}
We denote the minimizers of this problem by $\pi_m^*$. There are many differences between optimal transport and its martingale counterpart. For instance, the most popular cost function $c(x,y)=(x-y)^2$ for optimal transport cannot be used in the martingale context because, in this case, the objective function $\textup{\textbf{K}}(\pi)$ does not depend on $\pi$:
\begin{align}
\int_{\mathbb{R}^2} (x-y)^2 \, \mathrm{d} \pi(x,y) &= \int_{\mathbb{R}^2} x^2 + y^2 - 2xy \, \mathrm{d} \pi(x,y) \nonumber \\
&= \int_{\mathbb{R}^2} x^2 \, \mathrm{d} \mu(x) + \int_{\mathbb{R}^2} y^2 \, \mathrm{d} \nu(y) - 2 \int_{\mathbb{R}^2} y^2 \, \mathrm{d} \nu(y) \nonumber \\
&= \int_{\mathbb{R}^2} x^2 \, \mathrm{d} \mu(x) - \int_{\mathbb{R}^2} y^2 \, \mathrm{d} \nu(y).
\end{align}
Another difference is that we loose the geodesic interpretation since there is no counterpart to the Wasserstein distance in the martingale context. So, what would be the martingale counterparts to the shortest path, artificially noisy, and genuinely noisy OT interpolators? This is not trivial, since the interpolators must be martingales, and $\frac{T_1-t}{T_1-T_0} X_0 + \frac{t-T_0}{T_1-T_0} X_1$ is not a martingale, regardless of the coupling of $(X_0,X_1)$.
\bigskip
Since there is no longer something like a shortest path interpolator, we expect its counterpart in the martingale setup to be ``broken'' , i.e., not continuous. We propose the process that is equal to $X_0$ in $T_0$ and equal to $X_1$ for $t \in (T_0,T_1]$. For the artificially noisy MOT interpolator, the FAM $(M_t)$ with $(X_0,X_1) \sim \pi_m^*$ is a candidate. Indeed, it is a martingale, and its noise process is not taken into account in the selection of the optimal coupling $\pi_m^*$. Furthermore, if we remove the artificial noise, i.e., we set $A_t^{(1)}=0$ in the expression of $(M_t)$, and denote by $(\mathcal G_t)_{t\in[T_0,T_1]}$ the filtration generated by $(\frac{T_1-t}{T_1-T_0} X_0 + \frac{t-T_0}{T_1-T_0} X_1)_{t\in[T_0,T_1]}$, we obtain
\begin{equation}
M_t= \mathbb{E} [ X_1 \| \mathcal G_t] = \mathbb{E} \left[ X_1\, \bigg|\, X_0, \frac{T_1-t}{T_1-T_0} X_0 + \frac{t-T_0}{T_1-T_0} X_1 \right] = \left\{
\begin{array}{ll}
X_0 & \mbox{if } t=T_0, \\
X_1 & \mbox{if } t \in (T_0,T_1].
\end{array}
\right.
\end{equation}
Thus, we retrieve the broken MOT interpolator, similarly to how one retrieves the shortest path OT interpolator when setting the artificial noise to $0$ in the artificially noisy OT interpolator. For a genuinely noisy MOT interpolator, i.e., a sort of martingale counterpart to Schrödinger's bridge, we propose a new problem that we call the \emph{information-based martingale optimal transport problem}.
\begin{defn}
Let $X_0\sim\mu$ and $X_1 \sim \nu$, where $\mu$ and $\nu$ are $L^2$-probability measures on $\mathbb{R}$, in convex order. The information-based martingale optimal transport (IB-MOT) problem associated with the randomized anticipative Brownian bridge $(I_t^{(1)})$ is
\begin{equation}
\sup_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E} \left[ \int_{T_0}^{T_1} \frac{(X_1-M_t(X_0,I_t^{(1)}))^2}{T_1-t} \, \mathrm{d} t \right].
\end{equation}
For a general standard RAP $(I_t^{(1)})$ that satisfies the conditions in Proposition \ref{Ito}, the IB-MOT problem associated with $(I_t^{(1)})$ is
\begin{equation} \label{IB-MOT}
\sup_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) := \sup_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E} \left[ \int_{T_0}^{T_1} \frac{(X_1-M_t(X_0,I_t^{(1)}))^2 \sqrt{H_1'(t)H_2(t) - H_1(t) H_2'(t) }}{H_1(T_1)H_2(t)-H_1(t)H_2(T_1)} \, \mathrm{d} t \right],
\end{equation}
where the functions $H_1$ and $H_2$ are given by the covariance of the driver $(D_t)$ of $(I_t^{(1)})$: $K_D(x,y)= H_1 (\min(x,y)) H_2(\max(s,t))$ for all $(x,y) \in [T_0,T_1]^2$.
\end{defn}
As we shall see later, the supremum is always attained and the solution $\pi_I^*$ is unique. The FAM associated with $\pi_I^*$ is denoted $(M_t^*)$ and is also referred to as the solution to the IB-MOT problem.
\begin{rem}
IB-MOT is not simply MOT with a specific cost function. If one were to write $\textup{\textbf{K}}_I(\pi)$ as an integral of a function $c(x_0,x_1)$ against $\, \mathrm{d} \pi (x_0,x_1)$, the function would be
\begin{equation}
c(x_0,x_1)=\mathbb{E} \left[ \int_{T_0}^{T_1} \frac{(x_1-M_t(x_0,g_0(t) x_0 + g_1(t) x_1 + A_t^{(1)}))^2 \sqrt{H_1'(t)H_2(t) - H_1(t) H_2'(t) }}{H_1(T_1)H_2(t)-H_1(t)H_2(T_1)} \, \mathrm{d} t \right],
\end{equation}
where $g_0$ and $g_1$ are given by the signal function of $(I_t^{(1)})$. While $c(x_0,x_1)$ does indeed depend only on $x_0$ and $x_1$ explicitly, the expression of $(M_t)$ itself depends on $\pi$. Hence, $c(x_0,x_1)$ is not a suitable cost function for martingale optimal transport.
\end{rem}
We give another formulation of IB-MOT, by which we are able to show $\sup\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) < \infty$.
\begin{prop}
It holds that
\begin{equation}
\sup\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) = \sup\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E} \left[ \int_{T_0}^{T_1} \frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]\sqrt{H_1'(t)H_2(t) - H_1(t) H_2'(t) }}{H_1(T_1)H_2(t)-H_1(t)H_2(T_1)} \, \mathrm{d} t \right] < \infty.
\end{equation}
\end{prop}
\begin{proof}
Let $u_1(t)=\sqrt{H_1'(t)H_2(t) - H_1(t) H_2'(t) }$ and $u_2(t)= H_1(T_1)H_2(t)-H_1(t)H_2(T_1)$. By the definition of the FAM and Itô's isometry, we have
\begin{align}
\mathbb{E}[(X_1-X_0)^2] &= \mathbb{E}[(M_{T_1}-X_0)^2 ] \nonumber\\
&= \mathbb{E} \left[ \left(\int_{T_0}^{T_1} \frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)}{u_2(t)} \, \mathrm{d} W_t\right)^2 \right] \nonumber\\
&= \mathbb{E} \left[ \int_{T_0}^{T_1} \left(\frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)}{u_2(t)}\right)^2 \, \mathrm{d} t \right].
\end{align}
Since $\mathbb{E}[(X_1-X_0)^2 ]= \mathbb{E}[X_1^2 ] -\mathbb{E}[X_0^2 ] < \infty$, we have $\mathbb{E} [ \int_{T_0}^{T_1} (\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)/u_2(t))^2 \, \mathrm{d} t ] <\infty$, and because the product space $\Omega \times [T_0,T_1]$ is of finite measure, by Hölder's inequality, we also have $\mathbb{E} [ \int_{T_0}^{T_1} |\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)/u_2(t)| \, \mathrm{d} t ] <\infty$. Using the martingale property, we see that
\begin{align}
\mathbb{E}[\mathrm{Var}[X_1\| X_0, I_t^{(1)}]]&= \mathbb{E} [ \mathbb{E} [ X_1^2 \| X_0, I_t^{(1)} ] - M_t^2 ] \nonumber\\
&= \mathbb{E} [ X_1^2 ] - \mathbb{E}[M_t^2]\nonumber\\
&= \mathbb{E} [ X_1^2+M_t^2] - 2 \mathbb{E}[M_t\,\mathbb{E}[X_1 \| \mathcal{F}_t^I]] \nonumber\\
&= \mathbb{E} [ X_1^2+M_t^2] - 2 \mathbb{E}[\mathbb{E}[X_1 M_t \| \mathcal{F}_t^I]] \nonumber\\
&= \mathbb{E}[X_1^2+M_t^2 - 2X_1 M_t] \nonumber\\
&= \mathbb{E} [ (X_1-M_t)^2].
\end{align}
By Fubini's theorem, since $\mathbb{E} [ \int_{T_0}^{T_1} |\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)/u_2(t)| \, \mathrm{d} t ] <\infty$, we get
\begin{align}
\mathbb{E} \left[ \int_{T_0}^{T_1} \frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)}{u_2(t)} \, \mathrm{d} t \right] &= \int_{T_0}^{T_1} \frac{\mathbb{E} \left[ \mathrm{Var}[X_1\| X_0, I_t^{(1)}]\right]u_1(t)}{u_2(t)} \, \mathrm{d} t \nonumber \\
&= \int_{T_0}^{T_1} \frac{\mathbb{E} [ (X_1-M_t)^2]u_1(t)}{u_2(t)} \, \mathrm{d} t.
\end{align}
Furthermore, also by Fubini's theorem, since
\begin{equation}
\int_{T_0}^{T_1}\mathbb{E}\left[ \abs{\frac{ (X_1-M_t)^2u_1(t)} {u_2(t)}}\right] \, \mathrm{d} t =\int_{T_0}^{T_1}\mathbb{E} \left[ \abs{\frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)}{u_2(t)}}\right] \, \mathrm{d} t <\infty,
\end{equation}
one has $\int_{T_0}^{T_1} \mathbb{E} [ (X_1-M_t)^2]u_1(t)/u_2(t) \, \mathrm{d} t=\textup{\textbf{K}}_I(\pi)$. Hence,
\begin{equation}
\sup\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) = \sup\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E} \left[ \int_{T_0}^{T_1} \frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]u_1(t)}{u_2(t)} \, \mathrm{d} t \right] < \infty.
\end{equation}
\end{proof}
Since $$\left(\frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]\sqrt{H_1'(t)H_2(t) - H_1(t) H_2'(t) }}{H_1(T_1)H_2(t)-H_1(t)H_2(T_1)}\right)_{t\in [T_0,T_1]}$$ is exactly the volatility process of $(M_t)$, the IB-MOT problem can be seen as the martingale Benamou-Brenier problem (see \cite{Veraguas}) with the additional constraint that the volatility must be of the form
\begin{equation*}
\frac{\mathrm{Var}[X_1\| X_0, I_t^{(1)}]\sqrt{H_1'(t)H_2(t) - H_1(t) H_2'(t) }}{H_1(T_1)H_2(t)-H_1(t)H_2(T_1)}.
\end{equation*}
Eq. (\ref{IB-MOT}) can be written more concisely using the integration by parts formula.
\begin{coro}
It holds that
\begin{equation}
\sup\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) = \sup\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E}[X_1 W_{T_1}].
\end{equation}
\end{coro}
\begin{proof}
We recall that
\begin{equation}
M_t = X_0 + \int_{T_0}^t \frac{\mathrm{Var} [ X_1 \| X_0, I_s^{(1)} ] \sqrt{H_1'(s)H_2(s) - H_1(s) H_2'(s) }}{H_1(T_1)H_2(s)-H_1(s)H_2(T_1)} \, \mathrm{d} W_s,
\end{equation}
where $(W_t)$ is the innovations process of $(M_t)$, an $(\mathcal{F}^I)$-adapted standard Brownian motion on $[T_0,T_1]$. Integrating by parts, we obtain
\begin{equation}
\mathbb{E} \left[ \int_{T_0}^{T_1} \frac{\mathrm{Var} [ X_1 \| X_0, I_s^{(1)} ] \sqrt{H_1'(s)H_2(s) - H_1(s) H_2'(s) }}{H_1(T_1)H_2(s)-H_1(s)H_2(T_1)} \, \mathrm{d} t \right] = \mathbb{E}[M_{T_1}W_{T_1} - M_{T_0}W_{T_0}] = \mathbb{E} [X_1 W_{T_1}],
\end{equation}
and so, $\sup\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) = \sup\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E} [X_1 W_{T_1}]$.
\end{proof}
Because $W_{T_1}$ is a $\mathcal N(0,T_1-T_0)$-distributed random variable, regardless of the coupling $\pi$, one can complete the square to obtain an equivalent problem to Eq. (\ref{IB-MOT}):
\begin{align}
\mathbb{E}[X_1^2] - 2 \sup\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E} [X_1 W_{T_1}] + \mathbb{E}[W_{T_1}^2] &= \inf\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E}[X_1^2] - 2\mathbb{E} [X_1 W_{T_1}] + \mathbb{E}[W_{T_1}^2] \nonumber \\
&= \inf\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E}[(X_1 - W_{T_1})^2].
\end{align}
This equivalent formulation of IB-MOT highlights the similarities and differences with MOT. We retrieve the ability of using the difference squared as a cost, and noise has been introduced inside the objective function by replacing what in MOT is usually $X_0$ by $W_{T_1}$. Next, we show that a solution to an IB-MOT problem always exists in $\mathcal M (\mu,\nu)$, and it is unique.
\begin{prop} \label{IB-MOTsol}
It holds that
\begin{equation}
\sup\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) = \max\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi)
\end{equation}
and the solution $\pi^*_I :=\argmax\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi)$ is unique.
\end{prop}
\begin{proof}
Let $\mathcal M_x (\mu,\nu) = \{\gamma_x \in \mathcal P(\mathbb{R}) \| \, \mathrm{d}\pi(x,y) = \, \mathrm{d} \gamma_x (y) \, \mathrm{d} \mu (x), \pi \in \mathcal M (\mu,\nu)\}$ be the set of conditional martingale measures for $\mu$ and $\nu$. First, we show that
\begin{equation}
\inf\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E}\left[(X_1 - W_{T_1})^2\right] = \inf\limits_{\gamma_x \in \mathcal M_x (\mu,\nu)} \int_\mathbb{R} \int_0^1( Q_{\gamma_x}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x),
\end{equation}
where $Q_{\mathcal N_x}$ is the conditional quantile function of $W_{T_1}$ given $X_0=x$, and $Q_{\gamma_x}$ is the conditional quantile function of $X_1$ given $X_0=x$. We have:
\begin{enumerate}
\item For any $\gamma_x \in \mathcal M_x (\mu,\nu)$, there exists a $\pi \in \mathcal M(\mu,\nu)$, satisfying $\, \mathrm{d} \pi(x,y) = \, \mathrm{d} \gamma_x(y) \, \mathrm{d}\mu(x)$, such that $\int_\mathbb{R} \int_0^1( Q_{\gamma_x}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x) = \mathbb{E}_\pi[(X_1 - W_{T_1})^2]$. So,
\begin{equation}
\inf\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E}\left[(X_1 - W_{T_1})^2\right] \leqslant \inf\limits_{\gamma_x \in \mathcal M_x (\mu,\nu)} \int_\mathbb{R} \int_0^1( Q_{\gamma_x}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x).
\end{equation}
\item For any $\pi \in \mathcal M(\mu,\nu)$, there exists a $\tilde \pi_x \in \Pi(\gamma_x, \mathcal N_x)$ such that $\mathbb{E}_\pi[(X_1 - W_{T_1})^2] = \mathbb{E} [\mathbb{E}[(X_1 - W_{T_1})^2 \| X_0] ]= \int_\mathbb{R} \int_{\mathbb{R}^2} (x_1 - w)^2 \, \mathrm{d} \tilde \pi_x (x_1,w) \, \mathrm{d} \mu(x)$. So,
\begin{align}
\inf\limits_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E}\left[(X_1 - W_{T_1})^2\right] &\geqslant \inf\limits_{\gamma_x \in \mathcal M_x (\mu,\nu)} \int_\mathbb{R} \inf\limits_{\pi_x \in \Pi (\gamma_x, \, \mathcal N_x)} \int_{\mathbb{R}^2} (x_1 - w)^2 \, \mathrm{d} \pi_x (x_1,w) \, \mathrm{d} \mu(x) \nonumber \\
&=\inf\limits_{\gamma_x \in \mathcal M_x (\mu,\nu)} \int_\mathbb{R} \int_0^1( Q_{\gamma_x}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x).
\end{align}
\end{enumerate}
Hence, we have equality, and, if one of the infimums is attained, both are, while satisfying $\, \mathrm{d} \pi(x,y) = \, \mathrm{d} \gamma_x(y) \, \mathrm{d} \mu(x)$. Since $\mathcal M (\mu,\nu)$ is weakly compact in $\mathcal P(\mathbb{R}^2)$, any sequence admits a weakly converging subsequence with limit inside $\mathcal M (\mu,\nu)$. Let us consider a minimizing sequence, and furthermore denote a weakly converging subsequence by $(\pi_n) \subseteq \mathcal M (\mu,\nu)$. Using the relationship $\, \mathrm{d} \pi(x,y) = \, \mathrm{d} \gamma_x(y) \, \mathrm{d} \mu(x)$, we construct another weakly converging sequence $(\gamma_{x,n}) \subseteq \mathcal M_x (\mu,\nu)$ with limit inside $\mathcal M_x (\mu,\nu)$. We denote this limit by $\gamma_x^*$. We have
\begin{align}
\inf\limits_{\gamma_x \in \mathcal M_x (\mu,\nu)} \int_\mathbb{R} \int_0^1( Q_{\gamma_x}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x) &\leqslant \int_\mathbb{R} \int_0^1( Q_{\gamma_x^*}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x) \nonumber \\
& \leqslant \int_\mathbb{R} \liminf_{n \rightarrow \infty} \int_0^1( Q_{\gamma_{x,n}}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x),
\end{align}
where we used the lower semicontinuity of the Wasserstein metric on $\mathcal P (\mathbb{R})$, see Lemma 4.3 and Remark 6.12 in \cite{Villani2}. By Fatou's lemma applied to $x \mapsto \int_0^1( Q_{\gamma_{x,n}}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha$, we have
\begin{align}
\int_\mathbb{R} \liminf_{n \rightarrow \infty} \int_0^1( Q_{\gamma_{x,n}}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x) &\leqslant \lim_{n \rightarrow \infty}\int_\mathbb{R} \int_0^1( Q_{\gamma_{x,n}}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x)\nonumber \\
&= \inf\limits_{\gamma_x \in \mathcal M_x (\mu,\nu)} \int_\mathbb{R} \int_0^1( Q_{\gamma_x}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x).
\end{align}
Hence,
\begin{equation}
\lim_{n \rightarrow \infty}\int_\mathbb{R} \int_0^1( Q_{\gamma_{x,n}}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x)=\int_\mathbb{R} \int_0^1( Q_{\gamma_x^*}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha \, \mathrm{d} \mu(x),
\end{equation}
and the infimum is attained at $\gamma_x^* \in \mathcal M_x (\mu,\nu)$. Thus, the coupling $\pi^* \in \mathcal M (\mu ,\nu)$ defined by $\, \mathrm{d} \pi^* (x,y) = \, \mathrm{d} \gamma_x^*(y) \, \mathrm{d} \mu (x)$ is equal to $\argmin_{\pi \in \mathcal M (\mu,\nu)} \mathbb{E}[(X_1 - W_{T_1})^2]$ and $\argmax_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi)$. Uniqueness follows from the strict convexity of the map $\gamma_x \mapsto \int_0^1( Q_{\gamma_x}(\alpha)-Q_{\mathcal N_x}(\alpha))^2 \, \mathrm{d} \alpha$.
\end{proof}
\begin{ex}
Let $T=T_1-T_0, \sigma>0, X_0 \sim \mathcal N(0,\sigma^2), X_1 \sim \mathcal N(0,\sigma^2 +T )$. We show that $$\pi=\mathcal N \left( \begin{pmatrix} 0 \\ 0 \end{pmatrix}, \begin{pmatrix} \sigma^2 & \sigma^2\\ \sigma^2 & \sigma^2+T \end{pmatrix}\right),$$ which we call Brownian coupling, is the solution to the IB-MOT problem when
\begin{equation}
I_{t}^{(1)}= B_t - \frac{T_1-t}{T_1-T_0} (B_{T_0}-X_0) - \frac{t-T_0}{T_1-T_0} (B_{T_1}-X_1).
\end{equation}
We recall that $\sup\limits_{\pi \in \mathcal M (\mu,\nu)} \textup{\textbf{K}}_I(\pi) \leqslant \mathbb{E}[X_1^2]-\mathbb{E}[X_0^2] = T$. Assuming that $$ (X_0,X_1) \sim \mathcal N \left( \begin{pmatrix} 0 \\ 0 \end{pmatrix}, \begin{pmatrix} \sigma^2 & \sigma^2\\ \sigma^2 & \sigma^2+T \end{pmatrix}\right),$$
we obtain
\begin{enumerate}
\item $(X_0,X_1,I_s^{(1)})$ is Gaussian for all $s\in [T_0,T_1]$, and $(I_t^{(1)})$ is a standard Brownian motion on $[T_0,T_1]$.
\item $\mathrm{Cov}[X_0,I_s^{(1)}]= \sigma^2$.
\item $\mathrm{Cov}[X_1,I_s^{(1)}]= \frac{T_1-s}{T} \sigma^2 + \frac{s-T_0}{T} (\sigma^2+T) = \sigma^2 + s-T_0$.
\item \begin{align}
\mathrm{Var}[I_s^{(1)}]&= \frac{(T_1-s)^2 }{T^2}\sigma^2 + \frac{(s-T_0)^2}{T^2} (\sigma^2+T) + 2 \frac{(T_1-s)(s-T_0)}{T} \sigma^2 + \frac{(T_1-s)(s-T_0)}{T}, \nonumber \\
&= (s-T_0)^2 + \sigma^2T + \frac{(T_1-s)(s-T_0)}{T} \nonumber.
\end{align}
\item \begin{align}
M_s &= \begin{pmatrix} \sigma^2 & \sigma^2 + s-T_0 \end{pmatrix} \begin{pmatrix} \sigma^2 & \sigma^2 \\ \sigma^2 & (s-T_0)^2 + \sigma^2T + \frac{(T_1-s)(s-T_0)}{T} \end{pmatrix}^{-1}\begin{pmatrix} X_0 \\ I_s^{(1)} \end{pmatrix}, \nonumber \\
&= I_s^{(1)} \nonumber.
\end{align}
\item \begin{align}
\textup{\textbf{K}}_I \left(\mathcal N \left( \begin{pmatrix} 0 \\ 0 \end{pmatrix}, \begin{pmatrix} \sigma^2 & \sigma^2\\ \sigma^2 & \sigma^2+T \end{pmatrix}\right)\right) &= \mathbb{E} \left[ \int_{T_0}^{T_1} \frac{(X_1 - I_s^{(1)})^2}{T_1-s} \, \mathrm{d} s \right], \nonumber \\
&= \mathbb{E} \left[ \int_{T_0}^{T_1} \frac{\frac{(T_1-s)^2}{T^2}(X_1-X_0)^2 + \left(A_s^{(1)}\right)^2}{T_1-s} \, \mathrm{d} s \right], \nonumber \\
&= \int_{T_0}^{T_1} \frac{(T_1-s)^2 + (T_1-s)(s-T_0)}{T(T_1-s)} \, \mathrm{d} s, \nonumber \\
&=T\nonumber .
\end{align}
\end{enumerate}
Hence, the Brownian coupling is the optimal martingale coupling for $(X_0,X_1)$, and Brownian motion is the optimal FAM between $X_0$ and $X_1$, according to IB-MOT.
\end{ex}
\section{Conclusions}
Given a finite collection of convexly ordered random variables indexed by pre-specified fixed dates, we show how to construct continuous-time martingales, the so-called \textit{filtered arcade martingales} (FAMs, Definitions \ref{1FAM} and \ref{nFAM}), which match almost surely these random variables at the given dates. These interpolating martingales, defined as conditional expectations, filter noisy information about the last target random variable such that the interpolation through the collection of random variables is guaranteed. This martingale-interpolating method relies on a special class of stochastic processes, the \textit{randomized arcade processes} (RAPs, Definition \ref{RAP}). These processes, which interpolate strongly between the target random variables, generate information flows that reveal each target random variable at its indexing date, allowing a FAM to interpolate through the sequence of random variables. The FAM is also the solution to a filtering problem where the signal process is constant in time, and equal to the last target random variable, and the observation process is a RAP.
\bigskip
A randomized arcade process is the sum of two independent components: a signal function and a noise process, the latter is the so-called \textit{arcade process} (AP, Definition \ref{AP}). An AP is a functional of a stochastic driving process, and interpolates strongly between zeros. A well-known example of an AP over one period of time is the anticipative Brownian bridge (Example \ref{ABB}). Since APs play the role of noise, a natural class of APs are the ones driven by a Gaussian process. We give sufficient condition for such an AP to be Markovian (Corollary \ref{coromarkov}), and show that there are Markovian APs that are not anticipative representations of Markov bridges (Example \ref{nonstarcade}). As for the signal function, it interpolates between the target random variables, and it is deterministic conditional on these random variables. Since a RAP is a sum of a signal function and an AP, it does not possess the Markov property with respect to its natural filtration, apart from in trivial cases. Instead, it may have a similar property, the \textit{nearly-Markov property} (Definition \ref{nearlymarkov}), which plays a central role for the computational tractability of a FAM, and for guaranteeing the interpolation in the case with more than two random variables. We give sufficient conditions for a RAP to satisfy the nearly-Markov property (Theorem \ref{thmsemimarkov}), and we consider FAMs constructed with such RAPs. Using the Bayes rule and Itô’s lemma, we derive the SDEs satisfied by such FAMs (Propositions \ref{Bayes} and \ref{Ito}). If furthermore, the RAPs are \textit{standard} (Definition \ref{standardRAP}), the obtained FAMs can be expressed in terms of an integral with respect to a Brownian motion adapted to the filtration generated by the RAP underlying the FAM (Proposition \ref{propinnovations}). This Brownian motion corresponds to the innovations process in stochastic filtering theory.
\bigskip
The FAM theory relies on convexly ordered random variables, that is, the existence of a martingale coupling for the target variables. To select such a coupling based on a scheme (instead of fixing an arbitrary one), we propose to consider the \textit{information-based martingale optimal transport problem} (IB-MOT, Definition \ref{IB-MOT}). This method is inspired by martingale optimal transport and Schrödinger’s problem. IB-MOT aims at introducing noise in MOT in a similar fashion to how the entropic regularization of optimal transport introduces noise in optimal transport. In the IB-MOT setting, the noise contained in FAMs, which depends on the choice of the underlying RAP, is considered in the selection of an optimal coupling for the target random variables. Once a RAP is chosen, the IB-MOT problem admits a unique solution (Proposition \ref{IB-MOTsol}). For instance, according to IB-MOT, when the RAP is the anticipative randomized Brownian bridge, the optimal coupling for two Brownian marginals, is the Brownian coupling, that is, the optimal FAM that connects Brownian marginals is Brownian motion. Arcade processes and randomized arcade processes, alongside the filtered arcade martingales, appear thereof to suggest applications in, e.g., financial and insurance mathematics, mathematical biology, statistics, and climate science.
\section*{Acknowledgments}
A. Macrina is grateful for discussions on stochastic interpolation with Camilo A. Garcia Trillos and Jun Sekine, especially in the context of information-based modelling and stochastic filtering, during a research visit at Osaka University in March 2017 as part of a Daiwa Anglo-Japanese Foundation Small Grant, Ref: 3078/12370. Osaka University's hospitality is acknowledged and appreciated. A. Macrina is also grateful for the support by the Fields Institute for Research in Mathematical Sciences through the award of a Fields Research Fellowship in 2022. The contents of this manuscript are solely the responsibility of the authors and do not necessarily represent the official views of the Fields Institute. G. Kassis acknowledges the UCL Department of Mathematics for a Teaching Assistantship Award. Furthermore, the authors thank Lane P. Hughston for helpful comments and suggestions at an early stage of the work on arcade processes, Beatrice Acciaio for pointers to optimal transport, Julien Guyon, Gareth W. Peters, Ting-Kam Leonard Wong for useful conversations, and Daniel C. Schwarz for feedback. Attendees of the AIFMRM Research Seminar at the University of Cape Town (August 2021) and participants in the SIAM Conference on Mathematics of Data Science (MDS22, September 2022) are thanked for their comments and suggestions.
|
\section{Introduction}
Odor-guided navigation is common across the animal kingdom \citep{Baker2018-ys}. Olfactory cues inform an animal of its location in a natural environment \citep{Boie2018-eb}, and allow it to adjust its locomotion to navigate an odor landscape in a goal directed manner \citep{bargmann1991chemosensory,berg1972chemotaxis,aceves1979learning}. Odor guided navigation is an ethologically relevant task that is important for the animal's survival, and it has been a useful framework with which to study genes and circuits underlying sensory-motor transformations \citep{calhoun2017quantifying, clark2013mapping}.
Small model organisms navigate continuous gradients established by the spread of odorants from their sources due to diffusion and drift.
How animals interpret these gradients and use them to inform their actions remains an active and productive area of research especially in genetic model systems like \textit{C. elegans} and \textit{Drosophila melanogaster} \citep{bargmann1991chemosensory,aceves1979learning,Levy2020-oh,mattingly2021escherichia, Gomez-Marin2011-ok,Gepner2015-wm}
A major challenge is to quantitatively relate the animal's behavior to the precise olfactory cue that the animal experiences moment-by-moment.
Therefore it is critical to precisely control the odor environment and record the sensory cues experienced by these animals. The need to \emph{control} and \emph{measure} odorants still pose a formidable challenge. While many techniques exist to either present or measure odors in a lab environment, no technique currently exists for precise control and continuous monitoring of an odor landscape.
All approaches to generate odor landscapes in a lab environment must contend with the odor's diffusivity and interaction with other substrates. Early approaches to \textit{control} odor concentration relied on passive diffusion to construct a quasi-stationary spatial odor gradient, for example by adding a droplet of odorant in a petri dish in a ``droplet assay'' \citep{Louis2008-ju,Iino2009-al,Pierce-Shimomura1999-nt, monte1989characterization}. Diffusion places severe limits on the space of possible landscapes that can be created and on the timescales over which they are stable, and the created odor profile is sensitive to adsorption of odor to surfaces, absorption into the substrate, temperature gradients, and air currents, all parameters that are difficult to measure, model, or control. Microfulidics allow water-soluble odors to be continuously delivered to a chamber in order to provide spatiotemporal control \citep{chronis2007microfluidics, Albrecht2011-fj, lockery2008artificial}. Microfluidics devices, however, are limited in extent, require water-soluble odors and must be tailored-designed to the specific attributes of the animal's size and locomotion. While a post array has been shown to support \textit{C. elegans} locomotion, no microfluidic device has been demonstrated to support olfactory navigation of \textit{Drosophila} larvae, for example.
We previously reported a macroscopic gas-based active flow cell that uses parallel flow paths to construct temporally stable odor profiles \citep{Gershow2012-nt}. That approach allows for finer spatiotemporal control of the odor gradient, is compatible with \textit{Drosophila} larvae, and works with volatile airborne-based odor cues. This device used an array of solenoid valves to generate programmable odor profiles, but perhaps because of its complexity has not been widely adopted.
Most methods to create an odor landscape do not provide a means for knowing or specifying the spatiotemporal odor concentration. In other words while an experimenter may know that some regions of an area have higher odor concentrations, they cannot quantify the animal's behavior given a precise concentration of the odor. This limits the ability to quantitatively characterize sensorimotor processing. To address this shortcoming, various methods have been proposed to \textit{measure} odor concentration across space. For example, gas samples at specific locations could be taken and measured offline \citep{Yamazoe-Umemoto2018-nx}. In one of the most comprehensive measurements to date, Louis and colleagues
\citep{Louis2008-ju, tadres2022depolarization} used infra-red spectroscopy to measure the spatial profile of a droplet based odor gradient.
In all of these cases, measurements were performed offline, not during animal behavior, and the odor concentration was assumed to be the same across repeats of the same experiment, and when animals are present. But even a nominally stable odor landscape is subject to subtle but significant disruptions over time from small changes in airflow, from temperature variation, and from the odor's interaction with the substrate, which can include absorption, adhesion, and reemission \citep{Gorur-Shandilya2019-me, Yamazoe-Umemoto2018-nx,tanimoto2017calcium,Yamazoe-Umemoto2015-ru}.
This is challenging to account for and control within a single behavior experiment, and is even more difficult to account for across multiple instances of such experiments. Additional variability also arises across experiments as a result of the introduction of animals, changes to agar substrates, and alteration in humidity or other environmental conditions. To recover the odor concentration that an animal experienced, there is a need to measure odor concentration and animals' behaviors concurrently.
Our previously reported flow cell used a photo-ionization detector (PID) sensor moved across the lid before behavioral experiments to measure the odor concentration across space at a single point in time \citep{Gershow2012-nt}. During experiments, the total concentration of odor in the chamber was monitored concurrently with measurements of behavior. While this provided some assurances that the overall odor concentration was relatively stable, it did not provide any spatial information concurrently with behavior measurements.
Here we present a new flow chamber and a new multi-sensor odor array that addresses these prior limitations and can be used for measurement of the odor gradient with high spatial and temporal resolution. The array of sensors can be used two ways: the full array can be used to measure the generated gradient throughout the extent of the chamber, or parts of the array can be used on the borders to monitor, \textit{during behavioral experiments}, the odor profile in the chamber. By varying flow rates and the sites of odor introduction, we show a variety of odor profiles can be generated and stabilized.
To demonstrate the utility of the apparatus, we applied this instrument to quantitatively characterize the sensorimotor transformation underlying navigational strategies used by \textit{C. elegans} and \textit{D. melanogaster} larva to climb up a butanone odor gradient. Butanone is a water-soluble odorant found naturally in food sources \citep{worthy2018identification} that is often used in odor-guided navigation studies \citep{bargmann1993odorant, Levy2020-oh, Cho2016-is, Torayama2007-qi}. We show that the agar gel used during behavioral experiments greatly disrupts an applied butanone gradient, and we demonstrate a pre-equilibration protocol allowing generation of stable gradients taking into consideration the effects of agar. Moreover we monitor these gradients during ongoing behavior measurements via continuous measurements of the odor profile along the boundaries of the arena. %
Using these stable and continuously measured butanone gradients, we measure odor-guided navigation in animals by tracking their posture and locomotion as they navigate the odor landscape. We record chemotaxis behavior and identify navigation strategies in response to the changing odor concentration they experience.
In \textit{C. elegans}, we observe the presence of navigational strategies that were reported in other sensory-guided navigation conditions, such as salt chemotaxis \citep{Iino2009-al, Dahlberg2020-ip, Luo2014-pc}. These two strategies are: a biased random turn, known as a pirouette \citep{Pierce-Shimomura1999-nt}, and a gradual veering, known as weathervaning \citep{Iino2009-al, Izquierdo2015-la}. %
In \textit{Drosophila melanogaster} larvae, we identify runs followed by directed turns \citep{Gershow2012-nt, Louis2008-ju, Gomez-Marin2011-ok}. %
By using concurrent measurements of behavior and odor gradient we characterize olfactory navigation in these small animals on agar with known butanone odor concentrations, which for \textit{C. elegans} has not been reported before.
\section{Results}
We developed new methods both for generating and measuring odor gradients which we describe here. The systems are modular, scalable, and flexible. The components, which can be used independently of each other, can be fabricated directly from provided files using online machining services, or the provided plans can be modified for other geometries.
\subsection{Flow chamber for generating spatiotemporal patterns of airborne odors}
We first sought to develop a method of creating odor gradients that satisfied the following criteria:
\begin{enumerate}
\item The spatial odor profile should be \textit{controllable}. Varying control parameters (e.g. flow rates, tubing connections) should result in predictable changes to the resulting odor landscape.
\item The odor profile should be \textit{stable} and \textit{verifiable}. The same spatial profile should be maintained over the course of an experiment lasting up to an hour, and this should be verifiable via concurrent measurements during behavior experiments.
\item The apparatus should be \textit{straightforward} to construct and to use, and \textit{flexible} to adapt to various experimental configurations, including using with either \textit{C. elegans} or \textit{Drosophila} larva, and with agar arenas of various sizes.
\end{enumerate}
\begin{figure}
\begin{fullwidth}
\includegraphics[width=0.85\linewidth]{Odor_flow_Fig1_v2.png}
\caption{\textbf{Odor flow chamber with controlled and measured odor concentration.} \textbf{(a)} Schematic of airflow paths. Airflow paths for odor solution and water are controlled separately by mass flow controllers (MFCs) and spatially arranged into the odor chamber. The outflux from the chamber connects to a flow meter and photo-ionization detector (PID). \textbf{(b)} Flow chamber design. \textbf{(c)} Odor sensory array (OSA). Seven odor sensor bars are connected to a sensor hub. Each bar has 16 odor sensors (OS) and 8 temperature/humidity sensors (THS). Measured odor concentrations from the OSA of a spatially patterned butanone odor concentration shown in \textbf{(d)} for each sensor, \textbf{(e)} interpolated across the arena with square dashed line indicating the area where agar and animals are placed. \textbf{(f)} A two-parameter analytic flow model fit to measurement. \textbf{(g)} Cross-sections from (f) at at 4 different x-axis positions (show as colored arrows). Sensor readouts are overlaid points on the smooth curves from a 1D diffusion model. \label{fig:fig1}
}
\figsupp[Long-duration calibration confirms stable control and measurements.]{Long-duration calibration confirms stable control and measurements. (a, top) Odor flow rate driven by the MFC, (a, middle) raw reading of an odor sensor (OS) located in the chamber, (a, bottom) odor concentration readout from the photo ionization detector (PID) located at the outlet of the chamber. Note that the measurement is stable across the 90 minute recording. (b) We correct for the time lag between sensors and plot the mapping between PID measurements and OS readings.}{\includegraphics[width=10cm]{Odor_flow_SI1_2.png}}
\label{figsupp:figSI1-2}
\end{fullwidth}
\end{figure}
We constructed a flow chamber to control odor air flow across an arena (\FIG{fig1}a). Odor and humidified air are sourced from two bubblers, one containing pure water and the other an aqueous solution of odorant and water. Flow rates are controlled by separate mass flow controllers (MFCs) upstream of the bubblers. Downstream of the bubblers, the odor and air streams are divided into parallel sections of equal lengthed tubing. Each tube is connected to one input port of the flow chamber. The pattern of connections and the flow rates set by the two MFCs determines the shape of the produced odor profile. For instance, if odor is provided at a single central inlet, the resulting profile is a `cone' (\FIG{fig1}d-g) whose peak concentration and divergence are controlled with the MFCs (e.g. speeding up the odor flow while slowing down the air flow broadens the cone). Temporal gradients can be achieved by varying the odor flow in time, subject to constraints imposed by the odor's absorption into the agar gel.
The outflux from the flow chamber is connected to a flow meter and photo-ionization detector to monitor the overall flow rate and odor concentration respectively. The geometry of this flow chamber is shown in (\FIG{fig1}b), where parallel tubings are connected from the side and the chamber is vacuum sealed with a piece of acrylic on top during experiments. The chamber is designed for use with $\sim$100 mm square agar plates. The extra width (2.5 cm on either side) diminishes the influence of the chamber boundary on the odor profile over the arena. Interchangeable inserts allow for different agar substrates (e.g. circular plates) or for full calibration by odor sensor arrays (\FIG{fig1}c), discussed in the next section. Metal components are designed for low-cost fabrication by automated mechanisms (either laser cut-able or 3-axis CNC machinable). The fabrication plans for the flow chamber, design for the agar plate inserts, and required components to construct the flow path are publicly available in the \nameref{ssec:num1} section.
\subsection{Measuring the spatiotemporal odor distributions} A central difficulty in measuring animals' responses to olfactory cues is quantifying airborne odor concentrations that vary in space and time. This difficulty is exacerbated in turbulent environments where odor plumes carry abrupt spatial and temporal jumps in concentration far from the source with fundamentally unpredictable dynamics. But even in laminar flows, boundary conditions, slight changes in temperature, and the presence of absorbing substrates like agar make this challenging.
There is therefore a need, even for quasi-stationary gradients, to characterize the odor profiles in situ and to monitor these profiles during experiments. Various optical techniques, like laser induced fluorescence or optical absorption \citep{Louis2008-ju,tadres2022depolarization,demir2020walking}, %
exist to monitor concentration across planar arenas, but in general, these are incompatible with behavior experiments, expensive to construct, require specially designed arenas, or some combination of these disadvantages. Electronic chemical sensors can reveal the time-varying concentration at a particular point in space. A tiled array of these sensors acts as a `camera' forming a 2D spatiotemporal reading of the concentration. The gold-standard for measurement of odor concentration is the photo-ionization detector (PID), but even the smallest versions of these sensors are both too large ($\sim$ 2 cm in all dimensions) and too expensive ($\sim$ \$500 each) to make an array. Metal-oxide odor sensors, designed to be used in commercial air quality sensors, are available in inexpensive and compact integrated circuit packages. However, in general, commercial metal-oxide sensors are not designed for precision work - they tend to drift due to variations in heater temperature, humidity, adsorption of chemicals and ageing effects. Most such sensors are designed to detect the presence of gas above a particular concentration but not to precisely measure the absolute concentration. We became aware of a newer metal-oxide sensor, the Sensirion SGP30 that was designed for long-term stability and concentration measurement; we wondered if such a sensor could be calibrated for use in an odor sensor array.
To calibrate the sensor, we created a controllable concentration source by bubbling air through butanone. The odor reservoir contains butanone dissolved in water and is kept below the saturation concentration (11 mM or 110 mM odor sources).
We then mixed this odorized air flow into a carrier stream of pure air. We kept the carrier air flow rate constant ($\sim 400$ mL/min) and varied the flow rate through the odor source ($0-50$ mL/min); the odor flow rate was slow enough that the vapor remained saturated, so the concentration of butanone in the mixed stream was proportional to the flow rate through the butanone bubbler, as directly measured with a PID (\FIGSUPP[fig1]{figSI1-2}). We typically calibrated concentration with continuously ramped flow rate in triangle wave with 500 s period for 2-3 cycles.
We found a one-to-one correspondence between the odor sensor reading and the PID reading that persisted over time and showed no hysteresis. We reasoned that after applying this calibration procedure to an array of sensors, we could use the array to measure spatiotemporal odor concentration distributions with accuracy derived from the PID. Continuous calibration for 90 minutes showed that the odor sensors reliably reported concentration across durations (\FIGSUPP[fig1]{figSI1-2}) much longer than the typical behavioral experiment.
We constructed the sensor array from `odor sensor bars' (OSBs), printed circuit boards each containing 16 sensors in two staggered rows of 8. Each OSB also contained 8 temperature and humidity sensors to allow compensation of the odor sensor readings. The OSBs are mounted orthogonal to the direction of air flow; 7 OSBs fit inside our flow chamber (112 sensors total) allowing a full measurement of the odor profile. Taken together these 112 sensors formed an odor sensor array (OSA), capable of measuring odor concentrations with %
$\sim$ 1 cm spatial and 1 second temporal resolution. Prior to all experiments we calibrated the OSA in situ by varying the butanone concentration across the entire anticipated range of measurement while simultaneously recording the odor sensor and PID readings.
To verify the ability of the OSA to measure concentration gradients we created an artificially simple steady-state odor landscape by flowing odorized air ($\sim 30$ mL/min) through a central tube and clean air through the others ($\sim 400$ mL/min distributed into 14 surrounding tubings) in an environment without agar and without animals. This results in an air flow velocity $\sim 5$ mm/s in the flow chamber. As the flow rates and concentrations are all known and the flow is non-turbulent, and there is no agar or animals present, the concentration across the chamber should match a convection-diffusion model. After establishing the gradient, we recorded from the discrete odor sensors on the array (\FIG{fig1}d) and estimated the values in between sensors using spline interpolation with length scale equal to the inter-sensor distances (\FIG{fig1}e). We compared this stationary profile with a two-parameter convection-diffusion flow model (\FIG{fig1}f,g) fit to the data and described in the Methods section.
The measured concentrations in this artificially simplistic odor gradient show good agreement with the fit convection-diffusion model, especially in the central region where experiments are to be conducted, leading us to conclude the OSA can accurately report the odor concentration. We proceed to consider more complex odor environments. %
We demonstrate several examples of flexible control over the odor profile. In the configuration shown in \FIG{fig1} the center-most input provides odorized air and all surrounding inputs provide moisturized clean air to form a cone-shape stationary odor pattern. A narrower cone can be created by increasing the air flow (600 mL/min) of the surrounding inputs relative to the middle odorized flow (\FIG{fig2}a). %
The cone can be inverted by placing odorized air in the
two most distal inputs, and clean air in all middle inputs (\FIG{fig2}b). This inverse cone has lower concentration in the middle and higher on the sides. In later animal experiments, we restrict odorized air to one side to form a biased-cone odor landscapes, resulting in a cone with an offset from the middle line of the arena. Many more configurations are possible, demonstrating that the odor flow chamber enables the flexible control of airborne odor landscapes that are much more complex than a single odor point source. To show that the flow control and measurement methods are not restricted to any single odor molecule, we created and measured a cone profile using ethanol (\FIG{fig2}c).
\begin{figure}
\begin{fullwidth}
\includegraphics[width=1.0\linewidth]{Odor_flow_SI2.png}
\caption{
\textbf{Flexible control of a steady-state odor landscape.} \textbf{(a)} Configuration for a narrow cone with butanone. Colorbar shows interpolated odor concentration in ppm as measured by the odor sensor array. \textbf{(b)} An inverse cone landscape that has higher concentration of butanone on both sides and lower in the middle. \textbf{(c)}Another stationary odor landscape of a different shape, this time with ethanol.
\label{fig:fig2}
}
\end{fullwidth}
\end{figure}
\subsection{Odor-agar interactions dominate classical droplet assays}
\begin{figure}
\begin{fullwidth}
\includegraphics[width=1.0\linewidth]{Odor_flow_droplet_2.png}
\caption{\textbf{The presence of agar in the droplet assay alters the time-evolution of butanone odor landscape.}
\textbf{(a)} Concentration measured by odor sensor array is reported immediately after butanone droplet is introduced into the arena without agar. Red dot indicates the position of butanone droplet on the lid over the sensor array ($2\mu L$ of $10\%$ v/v butanone in water). \textbf{(b)} Same but three minutes later.
\textbf{(c)} Same measurement as in (a) but now droplet is added onto agar (gray). Two odor sensor bars have been removed to make space for the agar. \textbf{(d)} Same as (c) but three minutes later. Side view of the configurations with OSB sensors, butanone droplets, and agar gel are shown on the right.}
\label{fig:fig3}
\figsupp[Concentration measurements with an odor droplet and experimental perturbation.]{Concentration measurements with an odor droplet and experimental perturbation. (a) The initial concentration readout from a droplet of butanone on the lid near the middle of the arena. (b) Same condition as (a), but 20 minutes after the recording (left) and after shortly opening and closing the lid back (right) to mimic perturbation during worm experiments. (c) When there is agar in the chamber, the odor concentration is better maintained in the chamber. Location of the odor droplet is shown with a red dot. (d) After 20 minutes of recording (left) and after shortly opening and closing the lid back (right) to mimic experimental perturbation.}{\includegraphics[width=10cm]{Odor_flow_SI2-2.png}}\label{figsupp:figSI3}
\videosupp{Time-evolution of odor landscape from a butanone droplet with (right) and without (left) agar. The experimental conditions are the same as \FIG{fig3}. Blank sensor positions indicate sensors replaced by agar. The video updates every 2 seconds and measures butanone concentration from a droplet in the first 3 minutes.
Video available online at \href{https://figshare.com/articles/dataset/Continuous_odor_profile_monitoring_to_study_olfactory_navigation_in_small_animals/21737303}{10.6084/m9.figshare.21737303}
}\label{videosupp:sv1}
\end{fullwidth}
\end{figure}
Classic chemotaxis experiments in small animals commonly construct odor environments with odor droplets in a petri dish, usually with a substrate like agar. Our odor delivery instrument is designed to be compatible with a similar environment. To first better understand classical chemotaxis experiments, we sought to characterize the spatiotemporal odor profile from an odor droplet point source using our odor sensor array. We first considered the case without agar. In that case the odor concentration should be governed entirely by gas-phase diffusion. %
We placed a $2 \mu$L droplet of $10\%$ butanone in water %
on the lid of our instrument centered in the arena above the full OSA and without any airflow (\FIG{fig3}a,b). Butanone was observed to diffuse across the arena in the first three minutes (supplementary video (\VIDEOSUPP[fig3]{sv1})
and the equilibrium concentration is close to uniform across the odor sensors. We note that the final concentration of roughly 100 ppm, and the equilibration timescale both match what we would expect from first principals for $\sim 10^{-6}$ mol of butanone in a $\sim 225 \text{mL}$ arena that have a diffusion rate of $\sim 0.08 \text{cm}^2/$s in air.
A uniform odor landscape is not helpful for studying odor guided navigation, but most behavioral experiments are not conducted in a bare flow chamber but contain a biologically compatible substrate, such
as an gar gel, as is typically used in droplet assays. We therefore sought to investigate the role that agar plays in sculpting the odor landscape.
We introduced agar into the droplet assay by removing two sensor bars and replacing them with agar. We placed a butanone droplet directly on the agar, as done classically, and measured the odor landscape over time (\FIG{fig3}c,d). The odor concentration measured with agar is dramatically different from that measured without agar. Instead of quickly equilibrating to a uniform concentration, in the presence of agar there was instead a local maximum of butanone surrounding the droplet that persists even after 3 minutes. This difference in airborne odor concentration with and without agar persists after experimental perturbation such as removing and replacing the lid over the chamber (\FIGSUPP[fig3]{figSI3}). More broadly, the odor landscape we observed in the presence of agar would have been hard to predict ahead of time. An important consequence of this finding is that, to create a specific odor landscape (as in \FIG{fig1} or \FIG{fig2}) with agar, one will need to account for the effect of agar. We therefore sought to study odor-agar interactions more systematically and in the context of air flow. %
\subsection{Measuring and compensating odor-agar interactions with flow}
We first sought to measure whether the presence of agar changed the odor profiles generated due to flow in a bare chamber (\FIG{fig1}, \FIG{fig2}). As in \FIG{fig3}c,d, we replaced two odor sensor bars with a rectangular strip of agar gel or a metal plate as a control, and then measured airborne odor concentration upstream and downstream of the agar under odorized airflow that would normally produce a cone profile (\FIG{fig4}a-b).
While the agar had little effect on the odor landscape upstream of the agar, it drastically altered the downstream odor landscape (\FIG{fig4}b), suggesting that the agar absorbs
the airborne butanone molecules. This finding is consistent with the odor droplet experiments (\FIG{fig3}) and to be expected since butanone is highly soluble in water (275 g/L). Pulse-chase style experiments confirm that agar does indeed absorb and reemit butanone (\FIGSUPP[fig4]{figSI4-2}b). We also observed disruptions to the odor landscape when we used a full-sized 96 mm square agar plate intended for use with animals, \FIG{fig4}c. To accommodate the full sized agar plate we measured only the one dimensional odor profiles upstream and downstream of the agar \FIG{fig4}d. Taken together, these experiments suggest that agar-butanone interaction presents a challenge for setting up and maintaining stable odor landscapes.
We next sought a method to generate desired odor landscapes even in the presence of agar. We generate the odor profile by constant flow, which continuously replenishes the airborne odor. In principle, the disruption caused by agar should be overcome by constant flow of a sufficiently long duration, after which the agar and airborne odor would be in quasi-equilibrium at all spatial locations, with the concentration of odor dissolved in the gel proportional to the airborne concentration above it.
We measured odor concentration downstream of the agar and found that the airborne concentration failed to approach equilibrium on the timescales of single experiments \FIG{fig4}e,f. This suggests that it is not practical to simply wait for the agar and odor to reach equilibrium. Instead we developed a pre-equilibration protocol to more efficiently bring the agar and airborne odor into equilibrium before our experiments.
To more rapidly establish a desired airborne odor landscape, we briefly first exposed the agar to an airflow pattern corresponding to higher-than-desired odor concentration, created by replacing the odor reservoir with one containing a higher concentration of butanone.
We monitored the odor profile downstream of the agar until it reached the desired concentration and then switched to the original bubbler to maintain that concentration. Using this pre-equilibration protocol, we reached quasi-equilibrium quickly, typically after the order of ten minutes, \FIGSUPP[fig4]{figSI4-2},c. Note the spatial parameters of the two airflow patterns were the same, only the concentration of the odor source changes. Pre-equilibration allows the generation of airborne odor gradients in the presence of agar that match those in the absence of agar (\FIG{fig4}b,d,f right vs left column). We modeled the pre-equilibration protocol using a reaction-convection-diffusion model considering first order interactions between odor and agar. Under reasonable assumptions about the absorption rate, reemission rate, and capacity of the agar, simulations of this simple model provided qualitative agreement to our observations (\FIG{PE_model}a-c).
\subsection*{Monitoring the boundary determines the odor landscape at quasi-equilibrium}
\begin{figure}
\begin{fullwidth}
\includegraphics[width=0.95\linewidth]{Odor_flow_Fig2_v3.png}
\caption{\textbf{Under flow, agar interacts with odor to disrupt the downstream spatial odor profile, but a pre-equilibration protocol can coax the system into quasi-equilibrium and restore the odor profile.} %
\textbf{(a)} Two odor sensor bars are replaced with agar to observe the effect of introducing agar on the downstream spatial odor profile. %
\textbf{(b)} Measured odor profile is shown upstream and downstream of the removed odor sensor bars in the absence (left) and presence of agar (middle).
Transiently delivering a specific higher odor concentration ahead of time via a pre-equilibration (PE) protocol restores the downstream odor profile even in the presence of agar. \textbf{(c)} Additional odor sensor bars are removed and replaced with a larger agar, as is typical for animal experiments. \textbf{(d)} Measurements from the downstream sensor bar under the same three conditions in (b). The dots are sensor measurements and the smooth curve is a Gaussian fit. \textbf{(e)} The same experimental setup in (a), here focusing on time traces of only three downstream odor-sensors (colored circles for selected OS). \textbf{(f)} Concentration time series of three sensors color-coded in (e). Traces for three conditions are shown: time aligned to initial flow without agar (left), time aligned to initial flow with agar (middle), and traces after PE (right, with transparent line showing measurements another 20 min after the protocol). The dash-lines indicate the target steady-state concentration for each sensor.
\label{fig:fig4}
}
\figsupp[Time series of concentration change that capture effects of agar gel and the PE protocol]{Time series of concentration change that capture effects of agar gel and the PE protocol. (a) Concentration readout from the downstream PID (top) in response to the impulse of air flow rate though the odor bottle controlled by MFC (bottom), with no agar in the flow chamber. The background clean air flow is constant $\sim$ 400 mL/min throughout the recording. (b) Same as (a) but with agar plate in the flow chamber. Note that the response time scales to the same impulse are significantly different. (c) The time trace recorded from the PE protocol with agar plates. Odor reservoir with high butanone concentration (110 mM) is applied in the beginning, swapped back to the target concentration (11 mM) at $\sim$ 200 seconds, the odor concentration readout relaxes and stabilizes after $\sim$ 900 seconds, which enters steady-state for a duration longer than animal experiments.
}{\includegraphics[width=10cm]{Odor_flow_SI_PE2.png}}
\label{figsupp:figSI4-2}
\end{fullwidth}
\end{figure}
\begin{figure}
\begin{fullwidth}
\includegraphics[width=0.9\linewidth]{Odor_flow_SI_PE_model.png}
\caption{\textbf{Simulations from a reaction-convection-diffusion model of odor-agar interaction show that at quasi-equilibrium the airborne odor concentration is the same with or without agar.} \textbf{(a)} Simulation results of steady-state odor concentration in air without agar and with flow configured as in \FIG{fig1}. \textbf{(b)} The same simulation condition in (a) but now shortly after agar is introduced and before quasi-equilibrium is reached. A schematic of odor-agar interaction model is shown below. When agar is introduced, it absorbs the odor in air and decreases concentration measured downstream, producing a non-equilibrium (NE) concentration profile. \textbf{(c)} Odor concentration profile in air, with agar present, but after the pre-equilibration (PE) protocol brings this system to quasi-equilibrum. The PE protocol is shown in the schematic below, followed with steady-state (SS) with the stable odor concentration profile shown above. \textbf{(d)} The absolute difference of concentration profile without agar (a) and with agar after PE (c) is shown. \textbf{(e)} Upstream and \textbf{(f)} downstream odor concentrations along the agar boundary are shown for all three conditions.
}
\label{fig:PE_model}
\end{fullwidth}
\end{figure}
It is critical to monitor the odor landscape during animal experiments because the landscape is sensitive to environmental and experimental conditions which may fluctuate within and between experiments. But it is inconvenient to measure airborne concentration directly over the agar (e.g. because sensors impede optical access and also require heat management).
Fortunately, measuring the odor profile upstream and downstream of the agar places strong constraints on the airborne odor concentration over the agar such that in practice the spatial concentration can be confidently inferred.
If the airborne odor concentration upstream and downstream of the agar matches the profile in the absence of agar, one can infer that the airborne odor concentration landscape above the agar is also the same. The argument is straightforward: in the absence of sources or sinks, the fact that two concentration distributions obey the same differential equations and share the same conditions on all boundaries means that the the distributions are identical throughout the interior. Given identical measurements of the with-agar and without agar profiles at the inlet and outlets and reflecting boundary conditions on both walls, the only way for the with-agar distribution to differ from that without-agar is for sources and sinks of odor in the agar to be precisely arranged so that the all excess odor emitted from one point is exactly reabsorbed somewhere else before reaching the boundary. Not only is such an arrangement unlikely, it is inherently temporally unstable. A mathematical version of this argument is presented in the \hyperref[ssec:appendix]{Appendix}.
This quasi-equilibrium argument is supported by empirical concentration measurements shown in \FIG{fig4} and numerical results with the reaction-convection-diffusion model demonstrated in \FIG{PE_model}. The simulation results show that there is negligible difference between conditions with and without agar at quasi-equilibrium (\FIG{PE_model}d) when the odor concentration along the boundary is the same (\FIG{PE_model}e,f). Together, our numerical estimations, along with empirically observations allow us to safely infer that when measurements along the boundary indicate that they system is in quasi-equilibrium, the odor concentration experienced by animals on agar are the same as the concentrations measured in the absence of agar.
\subsection{Butanone chemotaxis in \emph{C. elegans}} %
We sought to directly quantify \textit{C. elegans}' navigation strategies for airborne butanone using our odor delivery system.
\textit{C. elegans} are known to climb gradients towards butanone \citep{bargmann1993odorant, Cho2016-is, Levy2020-oh}. Microfluidic environments suggest that they use a biased random walk strategy to navigate in a liquid butanone environment \citep{Levy2020-oh, Albrecht2011-fj}. Worms are also known to use weathervaning to navigate airborne odor gradients \citep{Iino2009-al, kunitomo2013concentration} although to our knowledge this has not been specifically investigated for butanone.
Worms were imaged crawling on agar in the flow chamber under an airborne butanone odor landscape illuminated by infrared light. Here 6 recording assays were presented, with approximately $50-100$ animals per assay, and two different odor landscapes were used.
\textit{C. elegans} navigated up the odor gradient towards higher concentrations of butanone, as expected (\FIG{fig5}a,b). Importantly, the odor concentration experienced by the animal at every point in time was inferred from concurrent measurements of the odor profile along the boundary of the agar, \FIG{fig5}c. On average, animals were more likely to travel in a direction up the local gradient than away from the local gradient, as expected for chemo-attraction \FIG{fig5}d. We use the term ``bearing to local gradient'' to describe the animal's direction of travel with respect to the local odor gradient that it experiences.
We find quantitative evidence that the worm exhibits both biased random walk and weathervaning strategies.
To investigate biased random walks, we measured the animal's probability of turning (pirouette) depending on its bearing with respect to the local airborne butanone gradient \FIG{fig5}e.
We find that the animal is least likely to turn when it navigates up the local gradient and most likely to turn when it navigates down the gradient, a key signature of the biased random walk strategy \citep{berg2018random, mattingly2021escherichia}.
To test for weathervaning, we measured how the curvature of the animal's trajectory depended on its bearing with respect to the local airborne butanone gradient, \FIG{fig5}f.
When the animal navigated up the butanone gradient (\FIG{fig5}f, blue) the distribution of the curvature of its trajectory was roughly symmetric and centered around 0 (straight line trajectory). By contrast, when the animal navigated perpendicular to the gradient (\FIG{fig5}f, yellow and red) the distribution of the curvature of its trajectories was skewed. The skew was such that it enriched for cases where the animal curved its trajectories towards the local gradient, a key signature of weathervaning.
Both the biased random walk and weathervaning behavior was absent in control experiments with flow but not odor, and we observed no evidence of anemotaxis at the $\sim 5$mm/s air velocities encountered by the animals (\FIGSUPP[fig5]{no_odor_control}).
We conclude that \textit{C. elegans} utilize both biased random walk and weathervaning strategies to navigate butanone airborne odor landscapes. We note that the quantitative analysis needed to make this conclusion relied on knowledge of the local airborne odor gradient experienced by the animal, which was provided by our odor profile measurements.
\begin{figure}
\begin{fullwidth}
\includegraphics[width=1.0\linewidth]{Odor_flow_Fig3_v3.png}
\caption{\textbf{\textit{C. elegans} use both biased random walk and weathervaning to navigate in a butanone odor landscape.} Animals on agar were exposed to butanone in the flow chamber. \textbf{(a,b)} Measured animal trajectories are shown overlaid on airborne butanone concentration for different odor landscapes. Green dots are each animal’s initial positions and red dots are the endpoints. %
\textbf{(c)} An animal's trajectory is shown colored by the butanone concentration it experiences at each position (top). Its turning behavior is quantified and plotted over time. Turning bouts are highlighted in gray. \textbf{(d)} Distribution of the animal's bearing with respect to the local airborne odor gradient is shown. Peak around zero is consistent with chemotaxis. \textbf{(e)}
Probability of observing a sharp turn per time is shown as a function of the absolute value of the bearing relative to the local gradient. Modulation of turning is a signature of biased random walk. Error bars show error for counting statistics. Data analyzed from over 9,000 tracks produced from $\sim$300 worms, resulting in 108 hours of observations. \textbf{(f)} Probability density of the curvature of the animal's trajectory is shown conditioned on bearing with respect to the local gradient. Weathervaning strategy is evident by a skew in the distribution of trajectory curvature when the animal travels perpendicular to the gradient (yellow and red). Three distributions are significantly different from each other according to two-sample Kolmogorov-Smirnov test ($p < 0.001$). Means are shown as vertical dashed lines.
\label{fig:fig5}}
\figsupp[Control measurements with air flow and without odor gradient.]{Control measurements with air flow but no odor. (a) Behavioral trajectories overlaid on the odor landscape that would have been expected had odor been present (mock odor landscape). No odor is presented, only moisturized airflow. (b) Distribution of bearing to the mock gradients under clean air flow. (c) Turn probability at different bearing conditions to the mock gradient. (d) Curvature conditioned on different bearing measurements.
}{\includegraphics[width=10cm]{Odor_flow_SI_noodor.png}}
\label{figsupp:no_odor_control}
\end{fullwidth}
\end{figure}
To quantify the overall navigational response with respect to local gradients, we further compute the animal's drift velocity as a function of local gradients (\FIG{fig6}). This captures the animal’s overall gradient climbing performance as a result of all the navigational strategies it uses, including the biased random walk and weathervaning. This calculation is only possible with a knowledge of the odor concentration experienced by the animal.
\begin{figure}
\begin{fullwidth}
\includegraphics[width=0.7\linewidth]{Odor_flow_Fig6_v3.png}
\caption{
\textbf{Tuning curve relating animal drift velocity to experienced odor concentration gradient.} \textbf{(a)} Schematic of a worm tracked in the odor landscape. The crawling velocity vector $V$, local concentration gradient $\nabla C$, bearing angle $\theta$, and drift velocity $V\cos(\theta)$ are shown. \textbf{(b)} Tuning curve shows the drift velocity $V\cos(\theta)$
as a function of the odor concentration gradient. Gray dash line indicates an unbiased performance with zero drift velocity, gray dots are the discrete measurements, and the black line shows the average value within bins. Error bar shows lower and upper quartiles of the measurements. %
}
\label{fig:fig6}
\end{fullwidth}
\end{figure}
\subsection{Butanone chemotaxis in \emph{Drosophila} larvae}
To further evaluate the utility of the flow chamber and gradient calibration for the study of small animal navigation, we investigated how larval \textit{Drosophila} navigate butanone. Although butanone is not as commonly used as a stimulus with \textit{Drosophila} as with \textit{C. elegans}, butanone is known to be attractive to larval flies \citep{dubin1995scutoid, dubin1998involvement} and has been variously reported to be attractive \citep{park2002inactivation} and aversive \citep{israel2022olfactory, lerner2020differential} to adult flies.
To investigate the larva's navigational strategy in a butanone gradient, we created a "cone" shaped butanone gradient over the agar substrate using the pre-equilibration protocol, as before, and we confirmed the presence and stability of the gradient by continuously measuring the spatial distribution of butanone upstream and downstream of the agar arena. We monitored the orientation and movement of 59 larvae over 6 separate 10 minute experiments ($\sim$ 10 larvae per experiment) with an average observation time of 7 min per larva (\FIG{fig7}).
Larvae moved towards higher concentration of butanone (\FIG{fig7}a). To analyze the strategy by which they achieved this, we first constructed a coordinate system in which 0 degrees was in the direction of the odor gradient (towards higher concentration) and 180 degrees was directly down-gradient; angles increased counterclockwise when viewed from above.
We found that larvae initiated turns at a higher rate when headed down-gradient ($\pm 180^\circ$ bearing with respect to the local gradient) than up gradient ($0^\circ$) (\FIG{fig7}b). When larvae turned, their reorientations tended to orient up gradient (negative angle changes from $+90^\circ$ bearing with respect to the local gradient, and positive angles changes from $-90^\circ$) (\FIG{fig7}c). Thus \textit{Drosophila} larvae use similar navigational strategies to \textit{C. elegans} to move towards butanone
\begin{figure}
\begin{fullwidth}
\includegraphics[width=0.9\linewidth]{Odor_flow_Fig7.png}
\caption{
\textbf{\emph{D. melanogaster} larvae chemotaxis in the odor flow chamber.} \textbf{(a)} Trajectories overlaid on the measured butanone odor concentration landscape. Example tracks are highlighted and the initial points are indicated with white dots. \textbf{(b)} Top: turn rate versus the bearing, which is the instantaneous heading relative to the gradient defined by quadrants shown on top. Error bar show counting statistics. Bottom: average heading change versus the bearing prior to all turns (re-orientation with at least one head cast). Error bars show standard error of the mean. Data analyzed from 6 experiments, 59 animals, with 620 turns over 6.8 hours of observation.
\label{fig:fig7}
} %
\end{fullwidth}
\end{figure}
\section{Discussion}
We present a custom-designed flow chamber and odor sensor array that enables us to measure navigation strategies of worms and fly larvae within the context of a controlled and measured odor environment.
The key features of this odor delivery system are that (1) the odor concentration profile through space is controlled in the flow chamber, (2) the odor sensor array provides a spatial readout to calibrate and measure the profile, and (3) the odor concentration profile is monitored during animal experiments. This last feature, the ability to monitor the spatial profile of odor concentration on the boundary during experiments, sets this method apart from previous approaches.
The ability to monitor spatial profile during experiments, along with a quantitative understanding of odor-agar interactions, provides confident knowledge of the odor experienced by the animal over time. This in turn allows us to extract tuning curves that describe the animal's behavioral response to the odor it experiences. In the future, such tuning curves may form the basis of investigations into neural mechanisms driving the sensorimotor transformations underlying navigation.
In contrast to liquid delivery of odor gradients via microfluidic chips \citep{Albrecht2011-fj,Larsch2015-xy}, our method allows worms to crawl freely on an agar surface. This allows our behavior measurements to be directly compared against classical chemotaxis assays \citep{bargmann1993odorant, Louis2008-ju, Pierce-Shimomura1999-nt}. Additionally, the macroscopic odor airflow chamber makes it straightforward to flexibly adjust the the spatial pattern between experiments without the need to redesign the chamber.
Our setup uses low flow rates corresponding to low wind speed velocities (5 mm/s) to avoid anemotaxis. Larger organisms, including adult flies, navigate towards odor sources by combining odor and wind flow measurements \citep{vergassola_infotaxis_2007, matheson_neural_2022}. Fly larvae exhibit negative anemotaxis at wind speeds 200 to 1000 times higher than those used here \citep{jovanic_neural_2019}, but previous work showed that they do not exhibit anemotaxis at lower windspeeds like the ones used here \citep{Gershow2012-nt}. For example, they do not exhibit anemotaxis at 12 mm/s which is still higher than the velocities they experience here \citep{Gershow2012-nt}. Therefore we do not expect \textit{Drosophila} larvae to exhibit anemotaxis under our flow conditions. \textit{C. elegans} are not thought to respond to airflow. In experiments in aqueous microfluidic chips under flow, \textit{C. elegans} move towards higher concentrations of attractant and do not respond to the flow of the liquid \citep{Albrecht2011-fj}. In agreement, we do not observe any evidence of \textit{C. elegans} anemotaxis in our chamber in response to control experiments without odor and with wind speeds of 5 mm/s \FIGSUPP[fig5]{no_odor_control}.
We focused on the odor butanone because it is important for a prominent associative learning assay \citep{Torayama2007-qi, kauffman_c._2011}. Butanone is soluble in water, and therefore it interacts strongly with agar. In this work we showed that this odor-agar interaction makes it challenging to \textit{a priori} infer an odor landscape experienced by the animal when agar is present, but that continuously monitoring the odor profile on the boundary overcomes this challenge. Other odors may instead have interactions with other substrates, such as glass, aluminum or plastic, which would also necessitate the use of our continuous monitoring approach. We show that our system is also compatible with less water-soluble odors, such as ethanol.
Here we have addressed the problem of creating airborne odor landscapes. The biophysical processes governing odor sensing in small animals such as \textit{C. elegans} are not fully understood. The worm carries a thin layer of moisture around its body as it moves on the agar substrate \citep{Bargmann2006-dy} and it is unclear to what extent the worm pays attention to the concentration of an odorant in the agar below it vs the air above it.
Our reaction-convection-diffusion model suggests that at the quasi-equilibrium conditions used in our experiments the odor concentration in agar is related to the airborne odor concentration directly above it up to a scalar that we predict to be constant across the agar. Although we have not measured this empirically, this suggests that even in the extreme case that the the animal only senses odor molecules in the agar, the odor concentration experienced by the animal in our experiments should differ by no more than a scaling factor compared to our estimates based on the airborne odor concentration.
Knowing the concentration experienced by the animal is not only useful for measuring navigational strategies more precisely than in classical assays, like the droplet chemotaxis assays. It will also be crucial for studying \textit{changes} in navigational strategy, such as those in the context of associative learning \citep{Cho2016-is, Torayama2007-qi}, sensory adaptation \citep{Levy2020-oh, itskovits2018concerted}, and long time scale behavioral states \citep{Calhoun2014-aa, Gomez-Marin2011-ok, klein2017exploratory}. In all those cases, it will be critical to disambiguate slight changes to the odor landscape from gradual changes in the navigational strategies. Continuously monitoring the odor landscape during behavior will remove this ambiguity.
\section{Methods and Materials}
\subsection{Odor flow chamber}
\subsubsection{Flow chamber setup}
The odor chamber (\FIG{fig1}b) was machined from
aluminum (CAD file in supplementary \nameref{ssec:num1} section). %
The chamber is vacuum sealed with an acrylic lid. The inner arena contains an aluminum insert that can hold the odor sensor array or a square petri dish lid (96x96 mm). The heading in which air flow can travel above the insert in the arena is 1 cm tall.
The whole setup is mounted on an optical breadboard and enclosed in a black box during imaging.
The airflow system is connected to a pressurized air source, passing through a particulate filter (Wilkerson F08) and a coalescing filter (Wilkerson M03), then regulated by mass flow controllers (MFCs, Aalborg GFC).
MFCs are controlled via a Labjack D/A board from a computer using custom Labview code.
We modulate the flow rate bubbling through liquid in enclosed bottles (Duran GL 45). The moisturized or odorized air is then passed into the flow chamber through inlet tubings. The outlets are connected to a copper manifold, then passed to a flow meter to assure that the inlet and outlet flow rate match. An optical flow sensor is fixed on the flow meter to make time stamps for opening and closing of the lid of the flow chamber during animal experiments. A photo-ionization detector (PID, piD-TECH 10.6 eV lamp) is connected to the outlet of airflow, providing calibration for the odor sensor array and detection of air leaks or odor residuals in the system.
Output readings from the PID, MFC, odor sensors described in the next section, and imaging camera, are all captured on the same computer sharing the same clock. Analog signals from the PID readout and MFC readback are digitized via a Labjack and recorded with the Labview program.
\subsubsection{Odor flow control}
To construct different odor landscapes tubes from the liquid-odor and water reservoirs are connected to the flow chamber in different configurations.
For a centered "cone-shape" odor landscape the tubing carrying odorized airflow is connected to the middle inlet. For the "biased-cone" landscape, the tubing for odorized air is connected to the inlet 4 cm off-center. For uniform patterns, all are connected to the same source through a manifold.
For all experiments the background airflow that carries moisturized clean air is set to $\sim$ 400 mL/min, except for \FIG{fig2} where this value was varied,
The odor reservoir contains either a 11mM or 110mM butanone solution in water with $\sim$ 30 mL/min airflow bubbling through the liquid.
Overall flow rates across the chamber in experiments were always around or less than $\sim$ 400 mL/min to avoid turbulence. We confirmed that this regime had no turbulence by visualizing flow in a prototype chamber using dry ice and dark field illumination. Our empirical observations matched theory: Given that the chamber is 15 cm wide and 1 cm deep, a flow rate up to 1 L/min corresponds to $\sim$ 1.1 cm/s. With kinematic viscosity of air $\sim 0.15$ cm$^2$/s, the Reynolds number is 7.3 times the flow rate in L/min, which is below the turbulence onset (Re=2000).
\subsection{Odor sensor array}
A spatial array of metal-oxide based gas sensors (Sensorion, SGP30) along with a relative humidity and temperature sensors (ams, ENS210) was used to measure the odor concentration field in the flow chamber.
Sensors are arranged together into groups of 16 odor sensors and 8 humidity sensors on a custom circuit board (MicroFab, Plano, TX) called an odor sensor bar (OSB). OSB's are in turn plugged into a second circuit board (OSH Park, Portland, OR) called the odor sensor hub (OSH). OSBs can be added or removed in different arrangements depending on the experiment, for example to make room for agar. Depending on the experiment, up to 112 odor sensors are arranged in a triangular grid such that no sensor directly blocks the flow from its downstream neighbor, accompanied by 56 humidity sensors in a rectangular grid.
Sensors are read out via the I2C protocol. Each SGP30 sensor has the same I2C address, as does each ENS210 sensor (different from the SGP30); to address multiple sensors of the same type we use an I2C bus multiplexer (NXP,
PCA9547
). Each OSB contains 2 multiplexers for its 16 sensors. The multiplexers are also addressed over I2C and can have one of 8 addresses (3 address bits). On each board, the two multiplexers share two bits (set by DIP switches); the remaining bit is hardwired to be opposite on the two multiplexers. Thus each OSB can have one of 4 addresses set by DIP switches, and 4 OSBs can be shared on one I2C bus.
To communicate with the sensors, we used a Teensy 4.0 microcontroller (PJRC, Sherwood, OR) running custom Arduino software. While the Teensy has two I2C busses, we found it more straightforward to use two micro-controllers instead. Both micro-controllers communicated via USB serial to a desktop computer running custom LabView software. Measurements from all sensors are saved to computer disk in real time. Readouts from the humidity sensors are also sent to their neighboring odor sensors in real time for an on-chip humidity compensation algorithm.
\subsubsection*{Heat management}
To avoid generating thermal gradients, the system has been designed to dissipate heat to the optics table. Each metal oxide odor sensor contains a micro hotplate which consumes 86 mW power during readings. To dissipate this heat the aluminum insert inside the flow chamber serves as a heat sink. Odor sensor bars are connected to the insert using heat conductive tape and thermal paste. The insert and chamber are in turn in direct thermal contact with the optics table. Temperature and humidity is constantly monitored at 8 locations per OSB via the on-board temperature and humidity sensors during experiments to confirm that there is no thermal or moisture gradient created in the environment.
\subsubsection{Measurements and calibration}
We measure from the odor sensors at 1 Hz for both calibration and behavior experiment modes.
We sample from the PID at up to 13 Hz. We synchronize and time align the measurements from the odor sensor array, MFC read-back, and PID recording with the same computer clock. %
To calibrate the odor sensors to the PID as in \FIGSUPP[fig1]{figSI1-2}, a spatial uniform flow was delivered in a triangle wave or a step pattern.
Time series from each odor sensor and the downstream PID were aligned by time shifting according to the peak location found via cross-correlation. The time shift was confirmed to be reasonable based on first principle estimates form the flow rate.
After measuring odor sensors' baseline response under clean moisturized air for 5 minutes, an odorized air was delivered.
To fit calibration curves, the raw sensor readout was fit to the PID measurements with an exponential of form:
\begin{equation}
\text{PID}(t) = A \exp(B*\text{OS}(t-\tau))
\end{equation}
where $\text{PID}(t)$ voltage is on the left hand side, the scale factor $A$ and sensitivity $B$ are fitted to match the raw sensor reading $\text{OS}(t-\tau)$ that is time shifted by time window $\tau$. This fitted curve maps from raw readings to odor concentration for each sensor. We validate the fitted curve across different recordings. The distribution of the coefficients $A$ and $B$ are relatively uniform across sensors in the middle of the arena. The sensor mapping are also reliable, so using $\pm$std of the fitted curve changes less than $10\%$ of the overall concentration scale of the landscape.
\subsection{Models for odor flow and odor-agar interaction}
We use two models in our work: (1) a convection-diffusion model that captures quasi-steady state odor concentration profile measured without agar used for the fits in \FIG{fig1}f,g and (2) a reaction-convection-diffusion model for odor-agar interaction shown in \FIG{PE_model}. A version of this second model is also used to justify the pre-equilibration protocol, as discussed in the \hyperref[ssec:appendix]{Appendix}.
\subsubsection{Convection-diffusion model for odor flow without agar}
To model odor flow without agar, for example for the fits in \FIG{fig1}f,g, we use a two-dimensional convection-diffusing model:
\begin{equation}
\frac{\partial C(x,y,t)}{\partial t} = -v\nabla C + D\nabla^2 C \label{eq:convection-diffusion}
\end{equation}
where the concentration across space and time is $C(x,y,t)$, flow velocity is $v$, and the diffusion coefficient of our odor is $D$. In our chamber, at steady state $(\frac{\partial C}{\partial t} =0)$ we have:
\begin{equation}
v\frac{\partial C}{\partial x} = D\frac{\partial^2 C}{\partial y^2}
\label{eq:steady-state}
\end{equation}
because with our configuration flow along the $x$ axis is dominated by convection while flow along the $y$ axis is dominated by diffusion, and therefore $\frac{\partial^2 C}{\partial x^2} \ll \frac{\partial^2 C}{\partial y^2}$.
The fit in \FIG{fig1}f is the solution to equation \ref{eq:steady-state}:
\begin{equation}
\label{eq:2dmodelfit}
C(x,y) = \frac{C_o}{2}(1-\erf(\frac{x}{2\sqrt{D\frac{x}{v}}})) \exp(-\frac{y^2}{4D\frac{x}{v}}),
\end{equation}
where $\erf$ is the error function and $C_o$ is the odor source concentration measured in air.
In \FIG{fig1}g we show a fit for a one dimensional slice along $y$ at various positions along $x_c$, for the situation in which there is an
odor-source at $(y=0,x=0)$:
\begin{equation}
C(y) = \frac{C(x_c,y=0)}{\sqrt{4\pi D \frac{x_c}{v}}} \exp(-\frac{y^2}{4D\frac{x_c}{v}})
\end{equation}
where $\frac{x_c}{v}$ is an analogy of time in non-stationary diffusion process at the cross-section at $x_c$. %
For the fits in \FIG{fig1} the air flow velocity is set to be $v\sim0.5$ cm/s based on the flow rate and geometry of the chamber (15 parallel tubes provide around 450 mL/min of flow into a $\sim$255 mL chamber with $\sim$15 cm$^2$ cross section). The diffusion coefficient $D$ is left as a free parameter and the value that minimizes the mean-squared error between the model and the empirical measurement is used. We chose to leave the diffusion coefficient as a free parameter instead of using butanone's nominal diffusion constant of $D\sim0.08$ cm$^2$/s, because we expect butanone's effective diffusion coefficient to be different in a confined chamber with background flow.
We note that the fitted profile shown in \FIG{fig1}g,f and the fitted value agrees with what is expected in a stable convection-diffusion process (Peclet number $\sim 80$).
\subsubsection{Reaction-convection-diffusion model for odor-agar interaction}
To justify the pre-equilibration protocol of \FIG{fig4} and to show that measurements of odor concentration along the agar's boundary allows us to infer the concentration on the agar, we propose a reaction-convection-diffusion model. This phenomenological model forms the basis of \FIG{PE_model}. Compared to the convection-diffusion model, we include the "reaction" term to account for odor-agar interactions.
The model used is a 2D generalization of this non-spatial model:
\begin{equation} \label{eqn_flow+agar}
\frac{dC}{dt} = -\frac{1}{\tau}(C - C_o) - w \frac{dA}{dt}
\end{equation}
\begin{equation} \label{eqn_agar}
\frac{dA}{dt} = k_a C (1-\frac{A}{M}) - k_d A
\end{equation}
where $C$ is a downstream concentration readout after the airflow has surface interaction with the agar gel. The influx odor concentration is $C_o$ and the odor concentration in agar is $A$. Without agar interaction, the flow chamber has its own timescale $\tau$ and the molecular flux into the agar is weighted by a scalar $w$ (so $w=0$ when there's no agar in the chamber). The association and dissociation constants are $k_a$ and $k_d$ and the maximum capacity of odor concentration that can be absorbed is $M$. This model is similar to the description of odorant pulse kinetics shown in \citep{Gorur-Shandilya2019-me}.
In \FIG{PE_model} we use the 2D generalization:
\begin{equation}
\frac{\partial}{\partial t} C(x,y) = \mathcal{L} C(x,y) - w \frac{\partial}{\partial t} A(x,y)
\end{equation}
\begin{equation}
\frac{\partial}{\partial t} A(x,y) = k_a C(x,y)(1-\frac{A(x,y)}{M(x,y)}) - k_d A(x,y)
\end{equation}
where $\mathcal{L} = -v\nabla + D\nabla^2 $ (\autoref{eq:convection-diffusion}) is a linear operator for the convection-diffusion process and the odor influx is at the boundary $C(x=0,y=0)=C_o$. We perform numerical analysis on the set of 2D equations and permit $A$ to be non-zero only in the region where agar is present. We use a target concentration $C_o$ that is lower than $M$ and $k_a \gg k_d$ to capture odor absorption into agar. In the simulated pre-equilibration protocol we temporarily increase $C_o$ above $M$ then switch back to the target concentration to efficiently reach a steady-state.
A slightly simplified version of this model forms the basis of the arguments in the \hyperref[ssec:appendix]{Appendix}.
\subsection{Animal handling}
\subsubsection{\emph{C. elegans}}
Wild type \emph{C. elegans} (N2) worms were maintained at 20 C on NGM agar plates with OP50 food patches. Before each chemotaxis experiments, we synchronized batches of worms and conducted measurements on young adults. Worms were rinsed with M9 solution and kept in S. Basal solution for around 30 min, while applying the pre-equilibration protocol to the flow chamber. Experiments were performed on $1.6\%$ agar pads with chemotaxis solution (5 mM phosphate buffer with pH 6.0, 1 mM CaCl$_2$, 1 mM MgSO$_4$) \citep{Bargmann1993-is, Bargmann2006-dy} formed in the lid of a 96x96 mm square dish. 50-100 worms were deposited onto the plate by pippetting down droplets of worms and removing excess solution with kimwipes. The plate was then placed in the odor flow chamber to begin recordings.
\subsubsection{\emph{D. melanogaster}}
Wild type \emph{D. melanogaster} (NM91) were maintained at 25 C incubator with 12 hr light cycle. Around 20 pairs of male and female flies were introduced into a 60 mm embryo-collection cage. A petri dish with apple juice and yeast paste was fixed at the bottom of the cage and replace every 3 hrs for two rounds during the day time. The collected eggs were kept in the petri dish in the same 25 C environment for another 48-60 hours to grow to second instars. We washed down and sorted out the second instar larva from the plate via $30\%$ sucrose in water around 10 min before each behavioral experiments. We used a 96x96 mm lid with $2.5\%$ agar containing $0.75\%$ activated charcoal for larval experiments \citep{Gepner2015-wm, Gershow2012-nt}. Around 10-20 larva were rinsed with water in a mesh and placed onto the agar plate with a paint brush. The same imaging setup and flow chamber configuration as the worm experiments were used for \textit{Drosophila} larva.
\subsection{Imaging and behavioral analysis}
\subsubsection{Image acquisition}
Animals are imaged via a CMOS camera (Baslar, acA4112-30, with Kowa LM16FC lens) suspended above the flow chamber and illuminated by a rectangular arrangement of 850 nm LED lights. The camera acquires $2,500 \times 3,000$ pixel images at 14 fps. A single pixel corresponded to 32 $\mu$m on the agar plate. Labview scripts acquired images during experiments.
\subsubsection{\emph{C. elegans} behavioral analysis}
To increase contrast for worm imaging, a blackout fabric sheet is placed underneath the agar plate. Custom Matlab scripts based on \citep{Liu2018-mv} were used to process acquired images after experiments, as linked in the \nameref{ssec:num2} section. Briefly, the centroid position of worms were found in acquired images via thresholding and binarization. The animal's centerline was found, and its body pose was estimated follwing \citep{Liu2018-mv}, but in this work only the position and velocity was used. The tracking parameters are adjusted for this imaging setup and we extract the centroid position and velocity of worm.
The analysis pipeline focuses on the trajectory of animal navigation in the arena. The trajectories are smoothed in space with a third order polynomial in a 0.5 s time window to remove tracking noise. We only consider tracks that appear in the recording for more than 1 minutes and produce displacement larger than 3 mm across the recordings. Trajectories starting at a location with odor concentration higher than $70 \%$ of the maximum odor concentration in space is removed, since these are likely tracks from animals that have performed chemotaxis already. We calculate the displacement of the center of the worm body in the camera space. The location in pixel space is aligned with the odor landscape constructed with the odor sensor array to compute concentration gradient given a position. To avoid double counting turns when the animal turns slowly, and to mitigate effects of small displacements from tracking noise, we measure the angle change between displacement vectors over 1 s time window and define turns as angle changes larger than 60 degrees. To quantify the curvature of navigation trajectories, we measure the angle between displacement vectors over 1 mm displacement in space.
\subsubsection{\emph{D. melanogaster} behavioral analysis}
Analysis of fly larvae is performed as previously \citep{Gepner2015-wm, Gershow2012-nt}.
\subsection{Data sharing}
\label{ssec:num1}
Recordings for odor flow control, concentration measurements, and behavioral tracking data are publicly available: \href{https://figshare.com/articles/dataset/Continuous_odor_profile_monitoring_to_study_olfactory_navigation_in_small_animals/21737303}{10.6084/m9.figshare.21737303}
\subsection{Software sharing}
\label{ssec:num2}
\begin{itemize}
\item Odor sensor array: \href{https://github.com/GershowLab/OdorSensorArray}{https://github.com/GershowLab/OdorSensorArray}
\item Worm imaging and analysis: \href{https://github.com/Kevin-Sean-Chen/leifer-Behavior-Triggered-Averaging-Tracker-new}{https://github.com/Kevin-Sean-Chen/leifer-Behavior-Triggered-Averaging-Tracker-new}
\item Larvae imaging: \href{https://github.com/GershowLab/Image-Capture-Software}{https://github.com/GershowLab/Image-Capture-Software}
\end{itemize}
\section{Acknowledgments}
Research reported in this work was supported by the National Institutes of Health National Institute of Neurological Disorders and Stroke under New Innovator award number DP2-NS116768 to AML and DP2-EB022359 to MHG; the Simons Foundation under award SCGB \#543003 to A.M.L.; by the National Science Foundation, through NSF 1455015 to MHG, an NSF CAREER Award to AML (IOS-1845137), under Grant No. NSF PHY-1748958 and through the Center for the Physics of Biological Function (PHY-1734030). This work was also supported in part by the Gordon and Betty Moore Foundation Grant No. 2919.02. We thank the Kavli Institute for Theoretical Physics at University of California Santa Barbara for hosting us during the completion of this work.
Strains from this work are being distributed by the CGC, which is funded by the NIH Office of Research Infrastructure Programs (P40 OD010440). We thank the Murthy Lab and Gregor Labs for flies.
|
\section{Introduction}
The one-dimensional (1D) Heisenberg model and its extension
1D Heisenberg-XY/Ising model were born many decades ago \cite{bethe1931,Kasteleijn_1952104,Sutherland_1970}.
Due to their quantum integrability and algebraic beauty \cite{Jimbo1995,Faddeev1979,slavnov_algebraic_2019,Baxter_1982_ExactlySM,caux_remarks_2011},
close-entwined physical and mathematical progress have been continuously surging since the birth of the models.
The studies not only significantly deepen and extend our understanding
on magnetic excitations and many-body physics,
but also expand frontiers of mathematics and even
create new mathematical directions.
From physical side, the studies reveal a series of charming
many-body excitations and rich emergent phenomena.
The revealed excitations include fractional
types such as spinon \cite{FADDEEV_spinwave_1981,faddeev_spectrum_1984,Muller_spinon_1981,Karbach_spinon_1997,Bougourzi_spinon_1996,Bougourzi_spinon_1998,caux2008}, (anti)psinon \cite{Karbach_pp_pap_2002,karbach2000III},
and exotic ones such as string excitations \cite{bethe1931,takahashi_1D_1971,Gaudin_XXZ_1971,Taka_suzuki_XXZ_1972,Takahashi1999},
where each of the former two excitations always appears in pairs while the later one always
contains bounded magnons.
Depending on the way to tune the anisotropy of the model and how the magnetic field is applied,
a variety of emergences appear:
Tomonaga-Luttinger liquid (without field
or with longitudinal field)
\cite{tomonaga_remarks_1950,Luttinger_1960,Luther_TLL_1975},
integrability with supersymmery (with a fine-tuned anisotropy without field) \cite{Fendley_SUSY_2003},
transverse-field Ising universality (with transverse field) \cite{Pfeuty1970,sachdev_2011,Jianda_crossover_2018}, and $E_8$
physics (with transverse and longitudinal fields)
\cite{a_b_zamolodchikov_integrals_1989,jianda_E8_2014,DELFINO_1995,xiao_cascade_2021,DELFINO199440}.
From mathematical side,
the Yang-Baxter equation and algebraic structure discovered in the model
significantly extend
our understandings on algebraic structure, and play an essential role
in building up the mathematical framework for quantum group and quantum algebra
\cite{Jimbo1995,Faddeev1979,slavnov_algebraic_2019,Baxter_1982_ExactlySM,drinfeld_hopf_1990,drinfeld_quantum_1988,drinfeld_Yangians_1988,drinfeld_QHA_1990,drinfeld_YBE_1983,Maillet_Drifeld_twists_2000}.
Along the course, experimental probes of the aforementioned exciting physics meet
great difficulties in early years due to limited techniques and lack of clean crystals.
With the experimental developments, the first progress is made for
the realization of spinon in the $\rm SrCuO_2$ \cite{kim_spinon_2006,Zaliznyak_spinon_2004}
and even higher-order spinon states are observed in $\rm Sr_2CuO_3$ \cite{mourigal_XXX_2013}.
Soon after the first observation of the
spinon, the magnetic excitations with clear Tomonaga-Luttinger liquid behavior
are also observed in the $\rm BaCo_2V_2O_8$ (BCVO) \cite{Kimura_TLL_2007}.
However, the demonstrations in material for the string excitations,
transverse field Ising chain (TFIC) universality along with the exotic $E_8$ physics
remain as a challenge until recently. The difficulty is largely due to
the challenge to figure out proper parameter regions for the
real material to realize those exotic physics.
Substantial progresses are made with a series of theoretical work
\cite{Jianda_crossover_2018,jianda_E8_2014,xiao_cascade_2021,Zou_universality_2019,yang_string_one-dimensional_2019} which
clearly identify proper parameters to realize those physics in corresponding
materials.
Meanwhile, experiments are carried out and unambiguously realize for
the first time the long-desired physics in corresponding quasi-1D materials
\cite{kimura_high_2006,bera_magnetic_2014,wang_string_experimental_2018,bera_string_dispersions_2020,cui_tfic_quantum_2019,wang_tfic_quantum_2018,zou_e_8_2021,zhang_e8_observation_2020,amelin_e8_experimental_2020,yang_Local_2022}.
\begin{figure}[h]
\centering
\includegraphics[width=0.6\textwidth]{figures/Phase_dia_XXZ.png}
\caption{
Phase diagram of the XXZ model Eq.~\eqref{eq:H_XXZ_lh}.
}
\label{fig:phase_XXZ}
\end{figure}
In this review, we focus on the 1D spin-1/2 Heisenberg-Ising model (XXZ model) with Ising anisotropy $\Delta>1$,
\begin{equation}
H=J\sum_{i=1}^N \left(S^x_iS^x_{i+1}+S^y_iS^y_{i+1}+\Delta S^z_iS^z_{i+1}
\right)
-h\sum_{i=1}^N S^z_i,
\label{eq:H_XXZ_lh}
\end{equation}
with the spin component $S_i^\alpha = \sigma_i^\alpha/2$ at site $i$ $(\alpha=x,y,z; \;{\rm and}\; \sigma_i^\alpha
\;{\rm labels\;the\; Pauli\; matrix})$,
the coupling $J>0$ between neighbouring spins,
and the external field $h$ along $z$ direction.
The phase diagram for the model is illustrated in Fig.~\ref{fig:phase_XXZ} \cite{CNYang_XXZ_1966,Johnson1991,Franchini2017}.
Since total spin along $z$-direction $S^z_{T}$ is conserved,
the Hilbert space can be divided into different subspaces according to
the quantum numbers of $S^z_{T}$ which is the magnetization along $z$ direction.
If $\Delta=1$, the Hamiltonian recovers the XXX model with an $SU(2)$ symmetry.
\begin{figure}[h]
\centering
\includegraphics[width=0.6\textwidth]{figures/spinon.png}
\caption{From the AFM ground state (a), a pair of spinons illustrated as two domain walls,
created by a single spin-flip (b), propagate along the chain (c). }
\label{fig:spinon}
\end{figure}
At zero field $h=0$, the 1D spin-1/2 XXZ model, has an antiferromagnetic (AFM) ground state.
Its elementary excitations referred to spinons are fractional spin-1/2 quasiparticles
appear in pairs in the spin dynamics.
The dynamical structure factor of two-spinon excitations can be calculated exactly
in both transverse and longitudinal channels \cite{caux2008,castillo_exact_2020}.
The two-spinon spectrum is directly observed in $\rm SrCo_2V_2O_8$ (SCVO) by inelastic neutron scattering (INS)
and agrees with theoretical prediction \cite{bera_spinon_2017}.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{figures/String.png}
\caption{An illustration for the 2 (3)-string excitation [red arrows in top (bottom) row] in spin configurations. }
\label{fig:2string}
\end{figure}
One appealing excitation in Heisenberg models is the string state which is a magnetic soliton state \cite{bethe1931,takahashi_1D_1971,Gaudin_XXZ_1971,Taka_suzuki_XXZ_1972}.
In the Ising limit ($\Delta \gg 1$) the magnetic soliton corresponds to
a bound object of flipped spins [Fig.~\ref{fig:2string}].
Despite of long-time theoretical recognition and continuously
theoretical studies of the
string excitations and their essentially compelling many-body nature \cite{bethe1931,Takahashi1999},
feasible proposals to
directly observe those exotic states remain a challenge \cite{ishimura_dynamical_1980,Imambekov_1D_2012,kohno_string_dynamically_2009,Pereira_edge_2008,Pereira_spectral_2009,Caux_comput_2005,Caux_computation_2005,Shashi_Nonuniversal_2011, Ganahl_Observation_2012}.
The difficulty lies in locating a proper parameter region
such that the string contribution to the spin dynamics is non-negligible.
A silver lining appears in 2017 that
the Ref.~\cite{yang_string_one-dimensional_2019}
proposes a promising
field-induced quantum critical region of the 1D spin-1/2 XXZ model Eq.~\eqref{eq:H_XXZ_lh},
to directly probe the string excitations, where
the string excitations are shown to dominate the dynamic spectrum in the
quantum critical region with small magnetization.
Following the proposal, the
string excitations are clearly detected and observed
in the zone center of SCVO for the first time
via the high-resolution terahertz (THz) spectroscopy measurement \cite{wang_string_experimental_2018}.
Under concrete theoretical guidance, the exotic 2- and 3-string excitations
as well as novel low energy fractional
magnetic excitations are identified in the field-induced quantum-critical region.
In 2020, with INS measurments on the same material,
complete dispersions of the string exciations
over the full Brillouin zone are obtained \cite{bera_string_dispersions_2020}.
In both experiments, the obtained excitation spectrum
and its magnetic field dependency
perfectly agree with theoretical calculations from low to high energy,
demonstrating a rare success in understanding strongly correlated magnetic systems.
When the longitudinal field is replaced by a transverse field
in Eq.~\eqref{eq:H_XXZ_lh}, a quantum phase transition arises with
field tuning. The corresponding quantum criticality around the
quantum critical point (QCP) falls into the class of
TFIC universality \cite{dmitriev_1D_2002}.
Near the QCP of the TFIC universality,
the Gr\"uneisen ratio,
the ratio of magnetic expansion coefficient to specific heat \cite{Yu_Gruneisen_2020},
is found to exhibit a unique singular behavior:
it can be either divergent or convergent when the system approaches
its QCP with field or
temperature tuning, respectively \cite{Jianda_crossover_2018}.
The exotic singular behavior
can serve as a smoking gun to
justify the TFIC universality in real materials.
Meanwhile, a relevant experiment is carried out for BCVO
in the transverse field \cite{wang_tfic_quantum_2018}, where
a 1D QCP appears at a strong field.
Near the 1D QCP, the obtained quantum critical behaviors
for the Gr\"uneisen ratio exactly
follow the aforementioned scaling relation in the class of TFIC universality,
which thus beautifully realizes the TFIC universality.
More surprisingly, when the quantum critical TFIC is perturbed
by a longitudinal field, an integrable massive relativistic quantum field theory,
dubbed as quantum $E_8$ integrable model, further emerges. In the model, it
contains eight types of massive particles whose scattering can be fully described
by the $E_8$ exceptional Lie algebra \cite{a_b_zamolodchikov_integrals_1989}.
In 2010, an INS experiment on the quasi-1D ferromagnetic
material $\rm CoNb_2O_6$ (CNO) observes
two lowest excitations whose energy ratio matches the mass ratio of the two lightest $E_8$
particles, providing a preliminary evidence for the existence of the exotic $E_8$ physics
in real material \cite{coldea_quantum_2010}.
However, the measured DSF in the continuum region is not accurate enough
to distinguish other $E_8$ particles,
which makes inconclusive a complete realization of the $E_8$
physics in the material CNO.
In 2020, a further THz spectroscopy experiment is performed on the same material,
which obtained a detailed spectrum in the continuum region
\cite{amelin_e8_experimental_2020}.
However, the obtained spectrum in the continuum region apparently deviates
from theoretically results \cite{xiao_cascade_2021},
implying more systematic and careful studies are needed
to conclude
the realization of the $E_{8}$ physics in the CNO \cite{morris_duality_2021,fava_glide_2020}.
In the same year,
by taking advantage of the delicate microscopic structure of BCVO,
combined efforts made from theoretical analysis,
numerical simulation,
THz spectroscopy, nuclear magnetic resonance (NMR), and INS experiments unambiguously reveal
beautiful $E_8$ physics for the first time in the material,
which not only confirms the eight single $E_8$
particles but also the excitations composed by multi-$E_8$ particles.
\cite{xiao_cascade_2021,zou_e_8_2021,zhang_e8_observation_2020}.
Above brief summary clearly shows rich physics in the Co-based materials, such as BCVO, SCVO, and CNO etc.,
where the $\rm Co^{2+}$ ion carries on the magnetism with $3d^7$ electrsons.
The ground state of the local $\rm Co^{2+}$ ion possesses
total orbital angular momentum $L=3$ and total spin $S_{tot} = 3/2$ \cite{Fazekas1999}.
Due to the superexchange mechanism the interaction of the neighbouring spin is isotropic, thus the effective model
becomes a spin-3/2 XXX model \cite{Anderson1959,Anderson1994,SAWATZKY1976}.
In those materials, the local environment and the spin-orbit coupling (SOC) are non-negligible,
which can cause further splitting of the local ground state manifold.
Since time-reversal symmetry is still preserved, the new ground state
in general at least have
a degeneracy of Kramers doublet. If we project the spin-3/2 XXX model to
the new ground state, we can obtain an effective 1D spin-1/2 XXZ model to
describe the low-energy physics, as manifested in
a large number of Co-based materials
\cite{shiba_exchange_2003,lines_magnetic_1963,Sarte2018}.
The remainder of the review is organized as follows.
Sec.~\ref{sec:Ogroup} introduces some basic properties of octahedral group and atomic basis for the $\rm CoO_6$ octahedron.
Sec.~\ref{sec:Hamiltonian} is devoted to derive an effective 1D spin-1/2 XXZ model for the Co-based materials whose ground state is a Kramers doublet, and gives a detailed discussion for the Zeeman effect in SCVO and BCVO.
Sec.~\ref{sec:excitations} reviews recent theoretical progresses on the magnetic excitations and emergent physics in the 1D spin-1/2 XXZ model without or with the presence various magnetic fields, and summarizes their experimental realizations.
Sec.~\ref{sec:conclusion} is the conclusions.
\section{The Octahedral Group and Atomic Basis}
\label{sec:Ogroup}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.7]{figures/Co_octa.png}
\caption{The local octahedron environment for a $\rm Co^{2+}$ with six oxygen ions surrounded.
}
\label{fig:Co_octa}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.4]{figures/oct_cube.png}
\caption{Equivalent axes belonging to $\mathcal{O}$ symmetry of the octahedron and the cube.}
\label{fig:Ogroup}
\end{figure}
For the local octahedron environment
the crystal field produced by six oxygen ions in the octahedron $\rm CoO_6$ has the octahedral symmetry described by $\mathcal{O}$ group,
consisting of all the rotations that keep an octahedron invariant [Fig.~\ref{fig:Co_octa}].
$\mathcal{O}$ group can also characterize the symmetry of
a crystal field induced by eight corner atoms of a cube [Fig.~\ref{fig:Ogroup}].
Thus, the crystal field of an octahedron is also referred to the ``cubic field''.
From Fig.~\ref{fig:Ogroup}, we can enumerate the elements of $\mathcal{O}$ group,
and distribute them to different classes
of inequivalent sets
of operations \cite{Dresselhaus2008,Zee2016,Tinkham1964}.
Rotations by the same angle about equivalent axes form a class
and two axes are equivalent if they are related to each other by a symmetry operation \cite{Fazekas1999}.
Thus, there are six $\pi/2$- and $3\pi/2$-rotations in the $\mathcal{C}_4$ class.
The $\pi$-rotations about the coordinate axis form the three-element class $\mathcal{C}_2$.
And there is another type of $\pi$-rotation $\mathcal{C}_2'$ about axes, where the axes pass through
the centre of the cube and are parallel to the face diagonals.
Rotations by $\pi/3$ and $2\pi/3$ about body diagonals of the cube generate the eight-element class $\mathcal{C}_3$.
Finally, the identity forms a class in itself.
In total, there are five different classes in $\mathcal{O}$ group with 24 group elements.
In group theory \cite{Dresselhaus2008,Zee2016}, the action of a symmetry operation $\hat R$ on a given
basis set can be described by a matrix $M({R})$.
The matrices of all symmetry operations form a group which is homomorphic to the corresponding symmetry group.
And the matrix group is referred to a
representation of the corresponding symmetry group.
The representation characterizes the symmetry properties which must be independent of the choice of the basis.
Furthermore,
the trace of a matrix representation is invariant under basis transformation,
which is introduced to serve as a character of the representation,
\begin{equation}
\chi(R)=\sum_i M_{ii}(R)=\text{Tr}(M(R)),
\end{equation}
For the identity $E$ of the group, $\chi(E)=d$, the
dimensionality of the representation.
For the $\mathcal{O}$ group, Table \ref{tab:cha_O} lists the characters of
all classes and irreducible representations.
The 1D representation $A_1$ is known as the identity representation,
and other four irreducible representations have dimensions 1, 2, 3, 3.
\begin{table}[t!]
\centering
\begin{center}
\begin{tabular}{c|c|c|c|c|c|c|c}
& &basis & $ E $ & $8\mathcal{C}_3$ & $3\mathcal{C}_2$ & $6\mathcal{C}_2'$ & $6\mathcal{C}_4$ \\
\hline
$A_1$ & $ \Gamma_1 $ & $\{x^2+y^2+z^2\}$ & 1 & 1 & 1 & 1 & 1 \\
$A_2$ & $ \Gamma_2 $ & $\{xyz\}$ & 1 & 1 & 1 & -1 & -1 \\
$E$ & $\Gamma_3$ & $\{x^2-y^2,3z^2-r^2\}$ & 2 & -1 & 2 & 0 & 0 \\
$T_1$ & $\Gamma_4$ & $\{x,y,z\}$ & 3 & 0 & -1 & -1 & 1 \\
$T_2$ & $\Gamma_5$ & $\{xy,yz,zx\}$ & 3 & 0 & -1 & 1 & -1
\end{tabular}
\end{center}
\caption{ The character table of the $\mathcal{O}$ group with two standard notations of the irreducible representations.
One commonly used basis set is also listed. }
\label{tab:cha_O}
\end{table}
A representation $\Gamma$ of a group is reducible if we can find a transformation $U$ which makes all matrices block-diagonal.
Then the reducible representation $\Gamma$ can be decomposed
into a direct sum of irreducible representations $\Gamma_j$,
\begin{equation}
\Gamma=\bigoplus_{j}a_j \Gamma_j.
\label{eq:decompose}
\end{equation}
where $a_j$ is the number of times that the irreducible representations occur.
Using the characters of different classes, $a_j$ can be determined as follows,
\begin{equation}
a_j=\frac{1}{\mathcal{N}}
\sum_{k=1}^r
\mathcal{N}_k \chi_j^*(\mathcal{C}_k)\chi(\mathcal{C}_k).
\end{equation}
where the summation is over the classes and $\mathcal{N}_k$ is the number of elements in the $k$th class $\mathcal{C}_k$ with $\sum_k\mathcal{N}_k=\mathcal{N}$, the total number of the group elements.
And the $\chi(\mathcal{C}_k)$ and $\chi_j(\mathcal{C}_k)$ are the characters of the
reducible representation $\Gamma$ and irreducible representation $\Gamma_j$
for the class $\mathcal{C}_k$, respectively.
\begin{table}[ht]
\centering
\begin{center}
\begin{tabular}{c|c|c|c|c|c}
& $ E $ & $8\mathcal{C}_3$ & $3\mathcal{C}_2$ & $6\mathcal{C}_2'$ & $6\mathcal{C}_4$
\\\hline
$\Gamma_{l=0}$ & 1 & 1 & 1 & 1 & 1 \\
$\Gamma_{l=1}$ & 3 & 0 & -1 & -1 & 1 \\
$\Gamma_{l=2}$ & 5 & -1 & 1 & 1 & -1 \\
$\Gamma_{l=3}$ & 7 & 1 & -1 & -1 & -1
\end{tabular}
\end{center}
\caption{A character table of the representations of the octahedral group
$\mathcal{O}$ with different orbital angular momentum $l = 0,1,2,3$.}
\label{tab:cha_spdf}
\end{table}
At the level of the atomic (ionic) Hamiltonian, the angular dependence of the atomic wave functions is described by the familiar spherical harmonics $Y_l^m(\theta,\varphi)$.
This comes from the fact that the pure atomic potential is spherically
symmetrical and the symmetry group of the Hamiltonian is the continuous group $SO(3)$
which contains all rotations in three-dimensional space.
As a result, the (total) angular momentum $l$ is a good quantum number
with eigenstates classified by $l$.
For an isotropic system, all rotation axes are equivalent thus
rotations around equivalent axes by a same angle $\alpha$ belong to a same class.
The corresponding character of an $\alpha$-rotation is
$\chi^l(\alpha)={\sin[(l+1/2)\alpha]}/{\sin(\alpha/2)}$ where $l$ is the angular momentum quantum number.
Since the octahedral group is a subgroup of the rotation
group $\mathcal{O}\in SO(3)$, every irreducible
representation of $SO(3)$ can provide a representation of $\mathcal{O}$, which usually is reducible.
In Table \ref{tab:cha_spdf}, we list the characters of four representations with different orbital angular momentum $l$.
As for the $\rm Co^{2+}$ with $L=3$, the seven-fold representation $\Gamma_{L=3}$ is reducible in $\mathcal{O}$ symmetry.
Following Eq.~\eqref{eq:decompose},
$\Gamma_{L=3}$ can be decomposed into a direct
sum of a singlet $A_2$ and two triplets $T_1$ and $T_2$, i.e.,
$\Gamma_{L=3}=A_2\oplus T_1\oplus T_2$.
Within this seven dimensional Hilbert space, we are free to choose
any seven linearly independent basis which follow the transformation of $\mathcal{O}$ group as a basis set.
These seven basis functions can be constructed after taking into account the cubic symmetry \cite{AABB1970,Bleaney_1953}:
\begin{equation}
\small
\begin{cases}
|T_1,2\rangle=\sqrt{\frac{3}{8}}Y_3^{-1}
+\sqrt{\frac{5}{8}}Y_3^3\\
|T_1, 0\rangle=Y_3^0\\
|T_1,-2\rangle=\sqrt{\frac{3}{8}}Y_3^{1}
+\sqrt{\frac{5}{8}}Y_3^{-3}
\end{cases},
\begin{cases}
|T_2,2\rangle=\sqrt{\frac{5}{8}}Y_3^{-1}
-\sqrt{\frac{3}{8}}Y_3^3\\
|T_2,0\rangle=\frac{1}{\sqrt{2}}(Y_3^2+Y_3^{-2})\\
|T_2,-2\rangle=\sqrt{\frac{5}{8}}Y_3^{1}
-\sqrt{\frac{3}{8}}Y_3^{-3}
\end{cases},
|A_2\rangle=\frac{1}{\sqrt{2}}(Y_3^2-Y_3^{-2}).
\label{eq:basis}
\end{equation}
It's easy to verify that the above basis functions are invariant under rotation within corresponding subspace and the trace of matrix representation of every rotation operation is consistent with the character table \ref{tab:cha_O}.
Group theory alone can not distinguish which irreducible representation is the lowest lying one.
To figure out this, the details of the Hamiltonian have to be specified.
{So next we shall find out the ground state manifold of $\rm CoO_6$ with the help of crystal field Hamiltonian.}
\section{Crystal Field and Effective Hamiltonian }
\label{sec:Hamiltonian}
In this section,
we show details to determine the lowest-lying Kramers doublet followed by
the crystal field Hamiltonian of the octahedron $\rm CoO_6$ via the
equivalent operator method \cite{AABB1970,Bleaney_1953,hutchings1964}.
Within the obtained ground-state doublet,
an effective XXZ model with pseudospin $S=1/2$ is built up for SCVO and BCVO.
\subsection{Crystal field Hamiltonian}
We start from the point charge model where the electrostatic potential
is generated from the surrounding anions, like the oxygen ions in Fig.~\ref{fig:Co_octa}.
This potential is referred to crystal field potential (in coordinate representation):
\begin{equation}
\mathcal{V} ( \mathbf{r} ) = \sum _ { j } \frac { q _ { j } } { \left| \mathbf { R } _ { j } - \mathbf { r } \right| }
\end{equation}
where $q_j=q=-Ze$ is the charge at the $j$th anion with a distance $|\mathbf { R } _ { j }| = a$ from the origin (e.g., Co$^{2+}$) and $\bf r$ is an arbitrary spatial point from the origin.
Since the crystal field potential $\mathcal{V}$ satisfies Laplace's equation $\Delta \mathcal{V}=0$,
it can be conveniently expanded as a sum of spherical harmonics $Y_n^m(\theta,\varphi)$:
\begin{equation}
\mathcal{V} ( r , \theta , \varphi ) =
\sum_{n=0}^{\infty}\mathcal{V}_n
=
\sum_{n=0}^\infty \sum_{m=-n}^n A_n^m Y_n^m(\theta,\varphi)r^n
\label{eq:Vexpansion}
\end{equation}
where $A_n^m$ are coefficients to be determined.
Thus, the crystal field potential energy of a magnetic ion at the origin, e.g., Co$^{2+}$ in Fig.~\ref{fig:Co_octa}, is
$W=\sum_i(-e)\mathcal{V}({\bf r}_i)$ where ${\bf r}_i$ is the position of an electron of the cobalt ion.
We need to calculate the matrix element of it between two basis functions, i.e. $\int{\rm d}{\bf r}\phi^*\mathcal{V}\psi$.
In order to calculate it efficiently, we can reduce the right hand side of Eq.~\eqref{eq:Vexpansion} by symmetry argument, since not all the terms in the sum have the same symmetry of surrounding anions.
For instance, all terms with $n>2L$ vanish,
because both $\psi$ and $\phi$ are $L=3$ basis functions Eq.~\eqref{eq:basis} for $\rm Co^{2+}$ ion.
Thus, the terms with $n>6$ vanish due to the orthogonality of spherical harmonics.
By a similar argument,
all terms with odd $n$
vanish
because $\psi$ and $\phi$ have the same parity, i.e. $\phi^*\psi$ has even parity.
Finally, due to the $C_4$ rotational symmetry along the $z$ direction of the octahedron, the potential $\mathcal{V}(r,\theta,\varphi)$ must be the same as $\mathcal{V}(r,\theta,\varphi+\pi/2)$.
And as the only $\varphi$ dependence is $\exp(im\varphi)$, this means the only non-vanishing $A_n^m$ are those terms with $m=0$ and $\pm4$.
After neglecting the $Y_0^0$ term, which is a constant, the remaining non-trivial contributions are $\mathcal{V}_4$ and $\mathcal{V}_6$ \cite{hutchings1964}:
\begin{equation}
\mathcal{V}(\mathbf{r})=\mathcal{V}_4(\mathbf{r})+\mathcal{V}_6(\mathbf{r}) \,,
\label{eq:V46}
\end{equation}
where
\begin{equation}
\begin{split}
\mathcal { V } _ { 4 }(\mathbf{r})
&=
D_4
\left[
Y_4^0(\theta,\varphi)
+\sqrt{\frac{5}{14}}
\left(
Y_4^4(\theta,\varphi)
+
Y_4^{-4}(\theta,\varphi)
\right)
\right]r^4
\\&=
A_4 \left( x ^ { 4 } + y ^ { 4 } + z ^ { 4 } - \frac { 3 } { 5 } r ^ { 4 } \right)
\end{split}\,,
\label{eq:V4}
\end{equation}
and
\begin{equation}
\begin{split}
\mathcal { V } _ { 6 } (\mathbf{r})
&=
D_6 \left[
Y_6^0(\theta,\varphi)
-\sqrt{\frac{7}{2}}
\left(
Y_6^4(\theta,\varphi)
+Y_6^{-4}(\theta,\varphi)
\right)
\right]r^6
\\&=
A_6 \left[ x ^ { 6 } + y ^ { 6 } + z ^ { 6 } + \frac { 15 } { 4 } \left( x ^ { 2 } y ^ { 4 } + x ^ { 2 } z ^ { 4 }
+ y ^ { 2 } x ^ { 4 } + y ^ { 2 } z ^ { 4 } + z ^ { 2 } x ^ { 4 } + z ^ { 2 } y ^ { 4 } \right) - \frac { 15 } { 14 } r ^ { 6 } \right]
\end{split}\,.
\label{eq:V6}
\end{equation}
The prefactors depend on the detailed geometry of the anions.
For instance, if the potential $\mathcal{V}(x,y,z)$ is generated by anions $q$ at the corners of an octahedron;
i.e., at $(a,0,0)$, $(0,a,0)$, $(0,-a,0)$, $(0,0,a)$, $(0,0,-a)$, one finds [{\it c.f.} appendix.~\ref{app:Oct_V}]
\begin{equation}
D_4=\frac{7\sqrt\pi}{3}
\frac{q}{a^5}
,\quad
D_6=\frac{3}{2}
\sqrt\frac{\pi}{13}
\frac{q}{a^7}
,\quad
A_4=\frac{35q}{4a^5}
,\quad A_6=\frac{-21q}{2a^7}.
\end{equation}
To further improve the efficiency of calculation, based on the Wigner-Eckart theorem \cite{Edmonds_2016},
Stevens proposed an equivalent operator approach \cite{Stevens_1952},
whose recipe is to replace
$\{x , y, z\}$ and $ ( x \pm i y )$ by $\{ L _ { x } ,\ L _ { y } ,\ L _ { z } \}$
and $L ^ { \pm }$, respectively.
Although $x$ and $y$ commute, $L_x$ and $L_y$ do not.
Thus, the symmetrical replacement is needed, for instance, $xy$ is replaced by the $(L_xL_y+L_yL_x)/2$.
Following the recipe,
the crystal field Hamiltonian Eq.~\eqref{eq:V46} can be converted to,
\begin{equation}
H = B _ { 4 } ^ { 0 } \left[ O _ { 4 } ^ { 0 } + 5 O _ { 4 } ^ { 4 } \right]
+
B _ { 6 } ^ { 0 } \left[ O _ { 6 } ^ { 0 } - 21 O _ { 6 } ^ { 4 } \right],
\label{eq:cry_H}
\end{equation}
where
\begin{equation}
O_4^0=\left[35 L _ { z } ^ { 4 } - 30 L ( L + 1 ) L _ { z } ^ { 2 } + 3 L ^ { 2 } ( L + 1 ) ^ { 2 } + 25 L _ { z } ^ { 2 } - 6 L ( L + 1 )\right]\,,
\end{equation}
\begin{equation}
O_4^4=
\frac{1}{2}\left[\left( L ^ { + } \right) ^ { 4 } + \left( L ^ { - } \right) ^ { 4 }
\right]\,,
\end{equation}
\begin{equation}
\begin{split}
O_6^0&=
231L_z^6-(315L(L+1)
-735)L_z^4
+(105 (L(L+1))^2 - 525 L(L+1) + 294) L_z^2
\\& \quad
- 5 (L(L+1))^3 + 40 (L(L+1))^2 - 60 L(L+1)\,,
\end{split}
\end{equation}
\begin{equation}
\begin{split}
O_6^4&=
\frac{1}{4}
\left[ (11L_z^2-L(L+1)-38)(L_+^4+L_-^4)+(L_+^4+L_-^4)(11L_z^2-L(L+1)-38) \right]\,.
\end{split}
\end{equation}
And the coefficients
\begin{equation}
B_4^0=\frac{7}{16}\frac{-eq}{a^5}\cdot\beta\langle r^4\rangle,\quad
B_6^0=\frac{3}{64}\frac{-eq}{a^7}\cdot\gamma\langle r^6\rangle\,,
\end{equation}
where
\begin{equation}
\beta = \mp \frac { 3 ( 2 l + 1 - 4 S ) [ - 7 ( l - 2 S ) ( l - 2 S + 1 ) + 3 ( l - 1 ) ( l + 2 ) ] } { ( 2 l - 3 ) ( 2 l - 1 ) ( 2 l + 3 ) ( 2 l + 5 ) ( L - 1 ) ( 2 L - 1 ) ( 2 L - 3 ) }\,.
\end{equation}
The minus (plus) sign is used for a shell less (more) than half full.
The values of $\gamma $ in different configurations are listed in TABLE 20 of \cite{AABB1970}.
Because $\beta<0$ and $q<0$, $B^0_4<0$ for $3d^7$ electrons case.
And for the $d$-electrons, only $\mathcal{V}_4$ needs to be taken into account since $\mathcal{V}_6$ has much weaker effect on $d$-electrons.
Thus, the crystal field Hamiltonian for $d$-electrons follows
\begin{equation}
H = B _ { 4 } ^ { 0 } \left[ O _ { 4 } ^ { 0 } + 5 O _ { 4 } ^ { 4 } \right]
\,.
\label{eq:cry_Hd}
\end{equation}
\subsection{ Effective 1D spin-1/2 XXZ model }
For the octahedron $\rm CoO_6$, the crystal field Hamiltonian Eq.~\eqref{eq:cry_Hd} has already been diagonalized in the $\Gamma_{L=3}$ manifold Eq.~\eqref{eq:basis}.
As a result, the triplet $T_1$ has $360B_4^0$, the triplet $T_2$ has $-120B_4^0$ and the singlet $A_2$ has $-720B_4^0$.
{Because $B_4^0 < 0$ for $3d^7$ electrons,} the ground manifold of $\rm CoO_6$ is triplet $T_1$.
If the energy gap between the ground state and the excited state is large,
we can use a pseudo angular momentum to describe the ground manifold \cite{Fazekas1999,AABB1970}.
One can find that the matrix representation of ${L}$ in the $T_1$ manifold {can be effectively considered as it in} $p$-basis (i.e. $Y^1_1,\ Y_1^0,\ Y_1^{-1}$) with projection coefficient $-3/2$.
Thus, we can replace ${L}$ by $-3/2 \tilde L$ where $\tilde L=1$.
In this sense, $3d^7$ electrons in $T_1$ manifold behave as if they were $p$-electrons carrying $\tilde L=1$.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{figures/Six_doublets.pdf}
\caption{
The six Kramers doublets due to crystal field splitting.
}
\label{fig:six_doublets}
\end{figure}
Recalling the total spin $S_{tot}=3/2$, the ground state is $3\times4=12$-fold degenerate,
which can be labeled as $|m_l, m_s\rangle$ where $m_l = -1, 0, 1 $ and $m_s = -3/2, -1/2, 1/2, 3/2 $
are components of $\tilde L$ and $S_{tot}$, respectively.
The SOC effect and the distortion play non-negligible roles in the materials we discuss,
which can lift the 12-fold degeneracy to certain degree.
The perturbation Hamiltonian follows
\begin{equation}
H'=-\frac{3}{2}\lambda {\bf\tilde L}\cdot\mathbf{S}_{tot}
-\delta(\tilde L_z^2-2/3)
\label{eq:Hper_SOC_dist}
\end{equation}
where $\lambda$ and $\delta$ are coupling coefficient and distortion strength, respectively.
The Hamiltonian $H'$ has a diagonal form in states $|m_j\rangle$
where $m_j=m_l+m_s$ is the component of the total angular momentum $J=\tilde L+S_{tot}$.
Due to the time reversal symmetry of $H'$, $|\pm m_j\rangle$ are degenerate and form a Kramers doublet.
In total, there are $2\times5/2+1=6$ Kramers doublets whose energies as a function of $\delta/\lambda$ are shown in Fig.~\ref{fig:six_doublets}.
The ground state doublet with energy $E_1$ is given by:
\begin{equation}
|\pm 1/2\rangle=
c_1|\mp1,\pm3/2\rangle
+c_2|0,\pm1/2\rangle
+c_3|\pm1,\mp1/2\rangle
\label{eq:gs_doublet}
\end{equation}
where $c_1$, $c_2$ and $c_3$ are coeffecients dependent on $\delta/\lambda$.
When the ground state doublet is well separated from the second lowest one,
we can construct the matrix representation for the total spin $S_{tot}=3/2$
within the ground state doublet
\begin{equation}
S^x_{tot}=\frac{\hbar}{2}
\begin{pmatrix}
0 & q\\
q & 0
\end{pmatrix}
,\quad
S^y_{tot}=\frac{\hbar}{2}
\begin{pmatrix}
0 & -iq\\
iq & 0
\end{pmatrix}
,\quad
S^z_{tot}=\frac{\hbar}{2}
\begin{pmatrix}
p & 0\\
0 & -p
\end{pmatrix}
\label{eq:S-1/2}
\end{equation}
where $q=c_2^2+\sqrt{3}c_1c_3$ and $p=({3}c_1^2+c_2^2-c_3^2)/2$.
Then, in the lowest-lying Kramers doublet,
the $S_{tot}=3/2$ can be effectively described by the pesudospin $S=1/2$,
\begin{equation}
S_{tot}^x=qS^x, \quad
S_{tot}^y=qS^y, \quad
S_{tot}^z=pS^z.
\label{eq:Smapping}
\end{equation}
In light of this, the pesudospin S=1/2 exhibits an anisotropy in $S^z$ component.
Similar results for trigonal distortion can also be obtained \cite{lines_magnetic_1963,shiba_exchange_2003,abragam_1951,achiwa_linear_1969}.
Based on superexchange mechanism
\cite{Anderson1959,Anderson1994,SAWATZKY1976},
the interaction of the neighbouring total spin is isotropic, thus the
effective spin-3/2 Hamiltonian is the XXX model,
\begin{equation}
H_{XXX}
=J\sum_{i=1}^N \left(S^x_iS^x_{i+1}+S^y_iS^y_{i+1}
+S^z_iS^z_{i+1}
\right).
\label{eq:H_XXX}
\end{equation}
However, for SCVO and BCVO, the low-energy physics can be effectively described by pseudospin $S=1/2$ instead of the total spin $S_{tot}=3/2$ [{\it c.f.} Eq.~\eqref{eq:Smapping}].
Thus, the effective spin-1/2 Hamiltonian for SCVO and BCVO follows
\begin{equation}
H_{XXZ}=J\sum_{i=1}^N \left(S^x_iS^x_{i+1}+S^y_iS^y_{i+1}+\Delta S^z_iS^z_{i+1}
\right),
\label{eq:H_XXZ}
\end{equation}
where $J > 0$ and
$\Delta={p^2}/{q^2}$ represents the Ising anisotropy.
In the BCVO and SCVO the energy gap between the ground state doublet and second lowest one is large \cite{bera_magnetic_2014,bera_string_dispersions_2020},
thus the effective Hamiltonian Eq.~\eqref{eq:H_XXZ} provides
a very good theoretical description for those materials,
which is detailedly discussed in Sec.~\ref{sec:excitations}.
With similar crystal field environment,
Co-based materials usually exhibits an Ising-anisotropy in the spin-spin interaction,
such as
the quasi-2D material $\rm Ba_3CoSb_2O_9$
\cite{Kamiya2018,Ghioldi2015}.
But this is not always the case \cite{Kim_2021}, since the origin of the spin-spin interaction is rather
complicated compared with the simple superexchange mechanism.
\subsection{ the Land\'{e} \texorpdfstring{$g$}{} tensor of \texorpdfstring{$\rm CoO_6$}{} screw chains }
\label{sec:Lande_g}
\begin{figure}[h]
\centering
\includegraphics[width=0.6\textwidth]{figures/Combine_Screw_Coordinate.png}
\caption{Schematic view of
(a) $\rm CoO_6$ screw chain and
(b) the relation between the $xyz$- and $\xi \psi \zeta$-coordinate system.
}
\label{fig:ScrewCoOchain}
\end{figure}
In the following we shall analyze the Land\'{e} $g$ factor for the SCVO and BCVO materials \cite{bera_magnetic_2014,He_SCVO_2006,kimura_collapse_2013,faure_topological_2018}.
The SCVO and BCVO are quasi-1D material
with four-fold periodicity along the $\rm CoO_6$ screw chain [Fig.~\ref{fig:ScrewCoOchain} (a)].
Due to local slight inclinations, the local $\xi \psi \zeta$-coordinate system for a local
$\rm CoO_6$ octahedron has
a tilted angle from the right-handed $xyz$-coordinate system for the lab frame [Fig.~\ref{fig:ScrewCoOchain} (b)].
As such, the $\tilde g$ factor becomes a tensor from the view of $xyz$ frame where
in experiments the $z$-axis corresponds to the $c$-axis,
$x$-axis is set along the transverse field, and $y$-axis is perpendicular to the $xz$-plane.
The local $\xi \psi \zeta$-coordinate system is given by rotating the $xyz$-coordinate system around the $y$-axis by $\theta$ and then around the $z$-axis by $\phi$.
Thus, the effective $\tilde g$ factor expressed in the $xyz$-coordinate system follows
\begin{equation}
\tilde g_{xyz}=
\begin{pmatrix}
{g_{xx}} & {g_{xy}} & {g_{xz}} \\
{g_{xy}} & {g_{yy}} & {g_{yz}} \\
{g_{xz}} & {g_{yz}} & {g_{zz}}
\end{pmatrix},
\label{eq:g_tensor}
\end{equation}
where
\begin{equation}
\begin{aligned}
g_{xx} &=\left(g_{\xi} \cos ^{2} \theta+g_{\zeta} \sin ^{2} \theta\right) \cos ^{2} \phi+g_{\psi} \sin ^{2} \phi, \\
g_{yy} &=\left(g_{\xi} \cos ^{2} \theta+g_{\zeta} \sin ^{2} \theta\right) \sin ^{2} \phi+g_{\psi} \cos ^{2} \phi, \\
g_{zz} &=g_{\xi} \sin ^{2} \theta+g_{\zeta} \cos ^{2} \theta, \\
g_{xy} &=\left(g_{\xi} \cos ^{2} \theta-g_{\psi}+g_{\zeta} \sin ^{2} \theta\right) \frac{\sin 2 \phi}{2}, \\
g_{yz} &=\left(g_{\zeta}-g_{\xi}\right) \sin \theta \cos \theta \sin \phi, \\
g_{xz} &=\left(g_{\zeta}-g_{\xi}\right) \sin \theta \cos \theta \cos \phi.
\end{aligned}
\end{equation}
Here, $g_\xi$, $g_\psi$ and $g_\zeta$ are the values of $g$ factor in the $\xi\psi\zeta$-coordinate system.
In the screw chain, the angle $\phi$ shifts by $\pi/2$ if we change
$\rm Co^{2+}$ site to the next one along the chain, reflecting the four-fold periodicity of the screw structure.
We denote $\phi_1$ the angle between $y$-axis and $\psi$-axis for the site 1 in the screw chain.
Then the four-fold periodicity condition is $\phi=\phi_1+(j-1)\pi/2$ where $j=1,2,\ldots,N$ is the site index.
Thus, along the chain, the Zeeman effect of a transverse field $\bf H_0$ follows
\begin{equation}
\begin{split}
{H}_{Zeeman}&=
\mu_B\sum_j \mathbf{S}_j\cdot \tilde{g}_{xyz}\cdot\mathbf{H}_0
\\&=
\mu_B\sum_j\left[ S_j^x g_{xx} \text{H}_0 +S_j^y g_{xy} \text{H}_0 +S_j^z g_{xz} \text{H}_0 \right],
\end{split}
\label{eq:zeeman_trans}
\end{equation}
where
\begin{equation}
\begin{aligned}
g_{xx}&=\left(g_{\xi} \cos ^{2} \theta+g_{\zeta} \sin ^{2} \theta\right) \cos ^{2} \left[ \phi_1+\frac{\pi}{2}(j-1) \right]+g_{\psi} \sin ^{2} \left[ \phi_1+\frac{\pi}{2}(j-1) \right],
\\
g_{xy}&=\frac{1}{2}\left(g_{\xi} \cos ^{2} \theta-g_{\psi}+g_{\zeta} \sin ^{2} \theta\right) \sin \left[2 \phi_1+\pi(j-1) \right],
\\
g_{x z}&=\frac{1 }{2}\left(g_{\zeta}-g_{\xi}\right) \sin 2\theta \cos \left[ \phi_1+\frac{\pi}{2}(j-1) \right].
\end{aligned}
\end{equation}
\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{figures/Two_H_direction.png}
\caption{Two cases with different directions of magnetic field.
}
\label{fig:2_H_direction}
\end{figure}
Next we present two cases with different directions of transverse field, i.e. $\phi_1=0^\circ$ and $\phi_1=45^\circ$.
We should bear in mind that $\hat{x}//\mathbf{H}_0$.
In $\phi_1=0^\circ$ case, namely, $\mathbf{H}_0$ is along $[110]$ direction [Fig.~\ref{fig:2_H_direction} (a)], then
the components of $\tilde g$ factor become
\begin{equation}
\begin{aligned}
g_{xx}&=\left(g_{\xi} \cos ^{2} \theta+g_{\zeta} \sin ^{2} \theta\right) \cos ^{2} \left[ 0+\frac{\pi}{2}(j-1) \right]+g_{\psi} \sin ^{2} \left[ 0+\frac{\pi}{2}(j-1) \right],
\\
g_{xy}&=\frac{1}{2}\left(g_{\xi} \cos ^{2} \theta-g_{\psi}+g_{\zeta} \sin ^{2} \theta\right) \sin \left[0+\pi(j-1) \right]=0,
\\
g_{xz}&=\frac{1 }{2}\left(g_{\zeta}-g_{\xi}\right) \sin 2\theta \cos \left[ 0+\frac{\pi}{2}(j-1) \right],
\end{aligned}
\label{eq:g_110}
\end{equation}
which indicates that
the transverse field along [110] direction can induce a four-periodic field but without staggered field.
In $\phi_1=45^\circ$ case, namely, $\mathbf{H}_0$ is along $[100]$ direction [Fig.~\ref{fig:2_H_direction} (b)], then
the components of $\tilde g$ factor become
\begin{equation}
\begin{aligned}
g_{xx}&=
\frac{1}{2}
\left[g_{\xi}\cos^2\theta+g_{\zeta}\sin^2\theta+g_{\psi}\right],
\\
g_{xy}&=
(-1)^{j} \frac{1}{2}\left(g_{\xi} \cos ^{2} \theta-g_{\psi}+g_{\zeta} \sin ^{2} \theta\right),
\\
g_{xz}&=\frac{1}{2}\left(g_{\zeta}-g_{\xi}\right) \sin 2\theta \cos \left[ \frac{\pi}{4}(2j-1) \right],
\end{aligned}
\label{eq:g_100}
\end{equation}
which indicates that
the transverse field along [100] direction can induce staggered and four-periodic fields.
One can obtain a similar result for the [010] direction.
Now we are ready to discuss magnetic excitations
in the 1D spin-1/2 XXZ model with various external fields,
as well as their
experimental realization in the BCVO and SCVO materials.
\section{Magnetic Excitations in the 1D Spin-1/2 XXZ Model with Various External Fields} \label{sec:excitations}
Depending on the directions of the applied external field, the magnetic excitations
in the spin-1/2 XXZ chain exhibit rich emergent phenomena and exotic physics,
e.g., Tomonaga-Luttinger liquid \cite{tomonaga_remarks_1950,Luttinger_1960,Luther_TLL_1975},
spinon \cite{FADDEEV_spinwave_1981,faddeev_spectrum_1984,Muller_spinon_1981,Karbach_spinon_1997,Bougourzi_spinon_1996,Bougourzi_spinon_1998,caux2008},
string \cite{bethe1931,takahashi_1D_1971,Gaudin_XXZ_1971,Taka_suzuki_XXZ_1972,kohno_string_dynamically_2009},
and $E_8$ particles \cite{a_b_zamolodchikov_integrals_1989,jianda_E8_2014,DELFINO_1995}.
To help measure or probe those exotic magnetic excitations,
a study on the spin dynamical structure factor (SDSF) becomes a must,
which can reveal weight distributions and dispersions of
various magnetic excitations on one hand, and give a direct guidance for their experimental
realization on the other hand \cite{Negele1988,chaikin_lubensky_1995,Zhu_MTCMM_2005}.
At zero temperature, the SDSF follows the Fourier transform of the
spin correlation function in space and time,
\begin{equation}
S^{a\bar{a}}(q,\omega)
=
\frac{1}{N}\sum_{jj'}^N e^{- iq(j-j')}
\int^{+\infty}_{-\infty}dt e^{i\omega t}
\langle GS | S_j^a(t)S_{j'}^{\bar{a}}(0)| GS \rangle
\end{equation}
where $a\in \{-,+,z\}$ and $|GS\rangle$ is the ground state.
Using the Lehmann representation,
it reduces to a single summation over a complete basis set,
\begin{equation}
S^{a\bar{a}}(q,\omega)=
2\pi\sum_{\alpha}
|\langle GS |S_q^a|\alpha \rangle|^2\delta(\omega-\omega_\alpha)
\label{eq:SDSF}
\end{equation}
where $\omega_\alpha=E_\alpha-E_0$ is the energy difference between
excited state $|\alpha\rangle$ and ground state $|GS\rangle$.
In Eq.~\eqref{eq:SDSF},
the spectral weight
$|\langle GS |S_q^a|\alpha \rangle|^2$
describes the transition probability between $|GS\rangle$ and $|\alpha\rangle$ after an external perturbation (the incoming photon or neutron) which couples linearly to the $S^a_q$ operator \cite{Negele1988,chaikin_lubensky_1995,Zhu_MTCMM_2005}.
Therefore, based on
spectral weight distribution $|\langle GS |S_q^a|\alpha \rangle|^2$
from different excitations $|\alpha\rangle$
in the energy-momentum space,
we can carry out detailed comparison
between theoretical calculation and experimental measurement
for recognizing possible realizations of various types
of magnetic excitations.
In the following, we will discuss the magnetic excitations in the spin-1/2 XXZ chain with
various external fields and their experimental realizations.
\subsection{ Zero field }
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{figures/twp_spinon_tDSF.png}
\caption{
The density plots for $S^{-+}$ at different anisotropy $\Delta$. From \cite{caux2008}. }
\label{fig:spinon_mp}
\end{figure}
At zero field, ground state of the 1D spin-1/2 XXZ model [Eq.~\eqref{eq:H_XXZ}]
accommodates a long-range AFM ordering
with an energy gap dependent on the Ising anisotropic parameter $\Delta$.
Above the gap, spin-flip excitations fractionalize
into pairs of spinons with fractional quantum number $S=1/2$ \cite{FADDEEV_spinwave_1981}.
By studying the SDSF of the system, we are able to
explore how the spinons influence the dynamics of the system.
For the $S^{-+}(q,\omega)$ [{\it c.f.} Eq.~(\ref{eq:SDSF})], the two-spinon
excitations provide a gapped continuum spectrum \cite{caux2008}, as shown in Fig.~\ref{fig:spinon_mp}.
The two-spinon excitations are
a small fraction of the total number of states,
however, they carry almost {100\%} spectral weight in $S^{-+}$ channel if $\Delta\ge2$.
When $\Delta < 2$ multi-spinon states like four-spinons excitations
play a more and more important role with decreasing $\Delta$ toward
the isotropic limit ($\Delta =1$) \cite{caux2008,Caux_2006}.
In $S^{zz}(q,\omega)$ channel, two-spinon excitations have similar results as
in $S^{-+}(q,\omega)$ \cite{castillo_exact_2020}.
The analytical SDSF provides a concrete ground
for the experimental realization.
Indeed, INS measurements
on the quasi-1D material SCVO directly
observe the expected excitation spectrum.
In SCVO, the long-range AFM order appears below
its N\'eel temperature $T_N=5.2$ K.
1D physics is expected to take over dominancy when $T > T_N $,
where SCVO can be effectively
described by the 1D spin-1/2 XXZ model Eq.~\eqref{eq:H_XXZ}.
The INS measurements at 6 K reveal a gapped scattering spectrum [Fig.~\ref{fig:spinon_INS_BA}] with
excellent agreement between the theoretical and experimental results \cite{bera_spinon_2017}.
Additionally,
Fig.~\ref{fig:spinon_INS_BA}(a)
also exhibits resonance peaks
caused by thermal fluctuation
and referred to as the Villian modes \cite{Villain_1975}.
These modes are firstly observed in $\rm CsCoBr_3$ by INS experiments
and can be explained by the scattering between two domain-wall states
\cite{Nagler_villian_PRL_1982,Nagler_villian_PRB_1983,James_XXZ_FiniteT_2009}.
Next, we shall discuss the magnetic excitations in 1D spin-1/2 XXZ model
with the presence of various external magnetic fields.
\begin{figure}[h]
\centering
\includegraphics[width=0.55\textwidth]{figures/spinon_INS_BA.png}
\caption{Dynamical spectrum of spinons. (a) INS measurement in SCVO and (b) Bethe ansatz calculation.
The theoretical dispersion of the Villian mode is represented by the dashed line
and the yellow points are used to highlight experimental peaks in (a).
From \cite{bera_spinon_2017}.}
\label{fig:spinon_INS_BA}
\end{figure}
\subsection{Longitudinal field}
With a longitudinal field $h$, the Hamiltonian becomes
\begin{equation}
H=
H_{XXZ}
-h\sum_{i=1}^N S^z_i,
\label{eq:H_XXZ_lh_sec4}
\end{equation}
where the time reversal symmetry is broken but the $U(1)$ symmetry is preserved.
If the longitudinal field is small such that the gap holds, i.e.,
the AFM ground state maintains,
there is no magnetization.
The magnetization starts to develop when $h$ is tuned above the field threshold $h_c(\Delta)$ where
the gap closes, and then the system enters the quantum critical (gapless) regime.
In this region,
the spinon can no longer serve as a good quasi-particle to describe the excitations
in the model.
Instead, the dynamics is dominated by
fractional excitations known as
``psinon-psinon'' (PP) and ``psinon-antipsinon'' (PAP)
\cite{Karbach_pp_pap_2002,karbach2000III},
as well as exotic string excitations \cite{yang_string_one-dimensional_2019}.
For observing those exotic states, an SDSF study needs to be carried out
in order to probe the excitations and quantify their spectra contributions.
In the $S^{-+}(q,\omega)$ channel, fractional excitations PP have a similar SDSF shape
as that of spinon [Fig.~\ref{fig:DSF_XXZ}] at small magnetization and disappears
when the magnetization becomes saturated \cite{yang_string_one-dimensional_2019}.
In both zero and nonzero field cases, these fractional excitations dominate almost the whole spectrum
implying negligibly small contribution from the
string excitations in the $S^{-+}(q,\omega)$ channel.
In a sharp contrast, for the $S^{+-}(q,\omega)$ channel [Fig.~\ref{fig:DSF_XXZ}],
contributions from low-energy factional excitations PAP
diminish in the zero field limit and recover the spin-wave excitations in the
limit of full polarization.
String excitations take over the dominance at the low-magnetization regions,
and tend to vanish when the system approaches full polarization.
At relatively low magnetization,
three well-separated continuums indicate three
different contributions from fractional PAP,
2-string and 3-string excitations.
The theoretical progress provides a concrete guidance to directly probe the
string excitations in real materials.
\begin{figure}[h]
\centering\includegraphics[scale=0.6]{figures/XXZ_DSF.png}
\caption{
The density plots for $S^{-+}$ (top row) and $S^{+-}$ (bottom row) of
the 1D spin-1/2 XXZ model,
where the magnetization is 20\% (a1, a2), 50\% (b1, b2) and 80\% (c1, c2) of full polarization.
From \cite{yang_string_one-dimensional_2019}.}
\label{fig:DSF_XXZ}
\end{figure}
In spite of the intriguing many-body nature of the excitations, especially the string excitations,
it is a great challenge to directly detect or
realize these exotic excitations in real materials
\cite{Imambekov_1D_2012,kohno_string_dynamically_2009,Pereira_edge_2008,Pereira_spectral_2009,
Caux_comput_2005,Caux_computation_2005,Shashi_Nonuniversal_2011,Ganahl_Observation_2012}.
In 2017, a silver lining appears, which suggests that the field-induced quantum critical region
of the 1D spin-1/2 XXZ model Eq.~\eqref{eq:H_XXZ_lh_sec4} is
a promising region to directly observe the string excitations \cite{yang_string_one-dimensional_2019}.
Following the suggestion, a high-resolution THz spectroscopy
measurement observes the string excitations in SCVO at the zone cneter
for the first time \cite{wang_string_experimental_2018}.
Under the concrete theoretical guidance, 2- and 3-string excitations
as well as low-energy fractional excitations (PP and PAP) are probed [Fig.~\ref{fig:THz_INS} (a)].
In 2020, the dispersion relation of the string excitations
over the full Brillouin zone is obtained via the INS measurement on the same material \cite{bera_string_dispersions_2020}.
Fig.~\ref{fig:THz_INS} (b) shows that the dispersion for the 2- and 3-string excitations
appears in the intermediate (2 - 15 meV at 6 T) and high (4 - 15.5 meV at 9 T)
energy regions, respectively.
In both experiments, the obtained excitation spectra and their magnetic field dependencies
are perfectly consistent with theoretical predictions,
closing the long-time open problem to directly observe the string excitations.
\begin{figure}[h]
\centering
\includegraphics[width=0.75\textwidth]{figures/Str_THz_INS.png}
\caption{(a) THz spectroscopy; (b1, b2) INS measurement of SCVO;
and (b3, b4) Bethe ansatz calculations.
From \cite{wang_string_experimental_2018,bera_string_dispersions_2020}.
}
\label{fig:THz_INS}
\end{figure}
\subsection{Transverse field}
With presence of a transverse field the XXZ Hamiltonian
becomes
\begin{equation}
H=
H_{XXZ}+H_x\sum_{i=1}^N S^x_i\,.
\label{eq:H_XXZ_Hx}
\end{equation}
With increasing field, a quantum phase transition emerges, which
falls in the class of TFIC universality \cite{dmitriev_1D_2002}.
For realizing the TFIC universality in real material,
distinctive criticality is desired for discerning the TFIC universality from others.
It is found that the Gr\"uneisen ratio, directly related to the magnetocaloric effect,
exhibits a very unique quantum critical scaling. The Gr\"uneisen ratio either
approaches to a constant or divergence when the QCP
of the TFIC universality is accessed via decreasing temperature or tuning field,
respectively ~\cite{Jianda_crossover_2018}.
The critical behaviors are different
from general quantum criticality where the Gr\"uneisen ratio in general approaches to
divergence regardless how to access the QCP \cite{Zhu_Universally_2003}.
As such measurements of Gr\"uneisen ratio near a QCP of the real material
can serve as a smoking gun to justify whether the universality falls
in the TFIC universality.
As discussed in Sec.~\ref{sec:Lande_g}, in the BCVO and SCVO,
when the applied field is along [100] direction,
there are two field-induced terms [{\it c.f.} Eq.~\eqref{eq:zeeman_trans}] in the Hamiltonian,
\begin{equation}
\begin{split}
H&=
H_{XXZ}
- H_x \sum_{i=1}^N\left\{ S_{i}^{x}
+h_y(-1)^{i}S_{i}^{y}
+h_zS^{z}_{i}
\cos\left[{\pi(2i-1)}/{4}
\right]
\right\}
\,,
\end{split}
\label{eq:H_XXZ_Hxzy}
\end{equation}
where $h_y \approx 0.4$ and $h_z \approx 0.14$ are the
internal-induced staggered and four-periodic fields (reduced by $H_x$), respectively.
And if the applied field is along [110] direction, only the four-periodic field with
slightly different form is present [{\it c.f.} Eq.~\eqref{eq:g_110}].
Although the Hamiltonian Eq.~\eqref{eq:H_XXZ_Hxzy} becomes more involved compared with Eq.~\eqref{eq:H_XXZ_Hx},
it still preserves the TFIC universality when the system is tuned to its QCP.
It is found that
the four-periodic perturbation term only slightly changes the location of
the QCP,
while
the staggered-field can significantly reduce
the field strength to access the QCP \cite{Zou_universality_2019,faure_topological_2018}.
Thus, from above theoretical analysis,
both BCVO and SCVO effectively described by 1D spin-1/2 XXZ model,
can accommodate the TFIC universality
via tuning the transverse field properly.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\textwidth]{figures/TFIC_Gruneisen_phadiag.png}
\caption{
The Gr\"uneisen parameter for BCVO vs. transverse field (a) and temperature (b).
Phase diagram of SCVO vs. transverse field (c).
From \cite{cui_tfic_quantum_2019, wang_tfic_quantum_2018}.
}
\label{fig:TFIC}
\end{figure}
As expected, by applying a magnetic field along [110] direction,
the TFIC universality is realized in BCVO where the critical field is around 40 T \cite{wang_tfic_quantum_2018}.
The measured Gr\"uneisen ratio is divergent when the transverse field approaches 40 T
but converges when the temperature decreases at 40 T [Fig.~\ref{fig:TFIC} (a) and (b)],
consistent with the theoretical analysis~\cite{Jianda_crossover_2018,wang_tfic_quantum_2018}.
If the magnetic field is along [100] direction, much lower critical fields can be obtained in both SCVO and BCVO
\cite{Zou_universality_2019,cui_tfic_quantum_2019,zou_e_8_2021}.
For instance, in SCVO, two QCPs at $H_{C1} = 7$ T and $H_{C2} = 7.7$ T are determined by NMR experiment \cite{cui_tfic_quantum_2019}. The former one is a (3+1)D Ising critical point
while the later one falls in the class of TFIC universality [Fig.~\ref{fig:TFIC} (c)].
Near the critical field $H_{C2}$, the critical exponent of transverse field is found to be consistent with the 1D TFIC universality class \cite{sachdev_2011,cui_tfic_quantum_2019}.
\subsection{More field and the \texorpdfstring{$E_8$}{} physics}
\begin{figure}[h]
\centering
\includegraphics[width=0.72\textwidth]{figures/E8region.png}
\caption{
The quantum $E_{8}$ integrable model emerges in the region of blue solid line.
Masses of the eight particles are expressed in units of the lightest two particles' masses $m_{1}$ and $m_{2}$.
From~\cite{xiao_cascade_2021}.
}
\label{fig:E8_phasedia}
\end{figure}
In the vicinity of QCPs, the low energy excitations in the infrared (IR) limit could be described by conformal field theory (CFT) due to the scaling invariant and large conformal symmetry. Moreover, with perturbation deformation from relevant primary field of corresponding CFT, the original physics in the IR limit will be greatly influenced. For instance, it may turn to massive field theory associated with certain Lie algebra, dubbed as affine toda field theory (ATFT)~\cite{BRADEN1990689,MUSSARDO1992215}.
In the scaling limit, the TFIC universality
can be described by a CFT with central charge $c=1/2$.
With an additional small longitudinal-field perturbation,
an integrable massive quantum field theory (QFT)
dubbed as quantum $E_{8}$ model emerges.
The quantum $E_8$ integrable model contains eight different types of massive particles
with their scattering fully described by the $E_8$ exceptional Lie algebra~\cite{a_b_zamolodchikov_integrals_1989,jianda_E8_2014,DELFINO_1995,xiao_cascade_2021}.
The $E_8$ Hamiltonian follows
\begin{equation}
H_{E_{8}}=H_{c=1/2}+\lambda\int\sigma(x)dx,
\label{eq:E8}
\end{equation}
with the $c=1/2$ CFT Hamiltonian $H_{c=1/2}$ and
the intensity of the small longitudinal field in the scaling limit $\lambda$.
The mass of the lightest quasiparticle $m_{1}=C\lambda^{8/15}$, $C\approx 4.40490858$~\cite{DELFINO_1995,fateev}.
And $\sigma(x)$ is the spin density operator,
a relevant field in the perturbed CFT~\cite{a_b_zamolodchikov_integrals_1989,DELFINO_1995}.
Fig.~\ref{fig:E8_phasedia} shows the phase diagram and massive excitations of the $E_{8}$ model.
The eigenstates of ATFTs satisfy Faddeev-Zamolodchikov algebra and can be described by exact S-matrix theory~\cite{ZAMOLODCHIKOV1979253,BRADEN1990689}. For solving the spectrum of ATFTs, a programmable method dubbed as conformal bootstrap is proposed by Cardy and Mussardo~\cite{CARDY1990387}, which could generate expressions of form factor for certain operator.
Later, the form factor theory of the $E_{8}$ model is further developed by Zamolodchikov, Mussardo and Delfino~\cite{YUROV1991,DELFINO_1995,DELFINO1996327,DELFINO1996469,Delfino_2004}.
A relevant and complete process of the conformal bootstrap can be found in the appendix of
Ref.~\cite{xiao_cascade_2021}.
With another relevant field energy density (corresponding to $\sigma^{x}$ in the lattice model),
we can use $E_{8}$ form factor theory and bootstrap approach to
determine the SDSFs $S^{xx}(\omega)$, $S^{zz}(\omega)$.
And $S^{yy}(\omega)$ can be obtained via an exact relation
$S^{yy}(\omega) = S^{zz}(\omega) \omega^2/(2gJ)^2 $ \cite{jianda_E8_2014}.
Fig.~\ref{fig:E8_DSF_analytic} shows
the analytical SDSF with transfer momentum $q=0$
(corresponding to the Brillouin zone center in lattice),
which includes contributions from single- and multi-$E_{8}$ particle channels
up to total energy $5m_{1}$.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{figures/E8spectra.png}
\caption{Analytical calculation of DSF with boradening of $0.05m_{1}$ has been exhibited. (a), (b) and (c) are corresponding to $S^{xx}(\omega)$, $S^{yy}(\omega)$ and $S^{zz}(\omega)$ respectively. All the contributions of different peaks have been highlighted inside the figures, from ~\cite{xiao_cascade_2021}.}
\label{fig:E8_DSF_analytic}
\end{figure}
The existence of the TFIC universality in the BCVO and SCVO materials brings
in a hope to realize the $E_{8}$ model.
In order to realize the model in the AFM materials,
the perturbed longitudinal field needs to be a stagger field
along $z$ direction (the Ising spin direction) from site to site along the chain.
It is extremely difficult to directly apply such a field for condensed-matter experiments.
Different from SCVO where the 1D QCP is always outside its 3D ordering dome,
for BCVO when applying a field along [100] (or [010]) direction the field strength for
the corresponding 1D QCP is at $H_C^{1D} \approx 4.7$T $< H_C^{3D} \approx 10 $T, deep inside the BCVO's
3D ordering AFM dome [Fig.~\ref{fig:E8_phase_NMR_TFIC}].
Inside the 3D ordering phase, due to the weak interchain coupling,
we can conveniently apply chain mean-field theory to effectively
describe the corresponding physics,
whose effective Hamiltonian now becomes
\begin{equation}
\small
\begin{split}
H&=H_{XXZ}
- H_x \sum_{i=1}^N\left\{ S_{i}^{x}
+h_y(-1)^{i}S_{i}^{y}
+h_zS^{z}_{i}
\cos\left[{\pi(2i-1)}/{4}
\right]
\right\}
-H'\sum_{i}(-1)^{i}S_{i}^{z},
\label{eq:XXZ_screw1}
\end{split}
\end{equation}
where $J=5.8$~meV in $H_{XXZ}$ and the staggered perturbation $-H'\sum_{i}(-1)^{i}S_{i}^{z}$ with $H'=0.018J$ comes from the chain mean-field
of the interchain coupling, which can not be neglected in the 3D AFM ordering region.
The effective staggered field along $z$ direction
now can serve as the key perturbation field to realize the quantum $E_{8}$ model.
Near $H_C^{\rm 1D}$, the INS is carried out
to probe the magnetic excitations in the BCVO,
and the long-desired $E_8$ spectrum consists of all the eight single-$E_8$ particle peaks
and multi-particle continuum are observed for the first time~\cite{zou_e_8_2021}, as shown in Fig.~\ref{fig:E8_DSF}.
This result is fully consistent with the analytical calculation and the iTEBD calculation ~\cite{zou_e_8_2021}.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{figures/E8_phadiag_NMR1DQCP.png}
\caption{
(a) Phase diagram of the BCVO with a transverse field along [010] direction;
(b) the quantum criticality of TFIC universality verified by NMR experiment.
From~\cite{zou_e_8_2021}.
}
\label{fig:E8_phase_NMR_TFIC}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{figures/BCVOdsf.png}
\caption{The $E_{8}$ spectrum. (a) INS data, (b) analytical results, and (c) numerical results, where $0.1m_{1}$ energy resolution
broadening is applied to the latter two results
for comparison with INS data. (d), (e), (f) and (g) reveal the analytical
contributions of the spectrum from different $E_{8}$ channels. From~\cite{zou_e_8_2021}.}
\label{fig:E8_DSF}
\end{figure}
It is worth discussing a little more about how the comparison is completed, since it is a nontrivial task to compare continuous results (of field theory) with discrete results (of lattice calculations and experiments).
In the field theory the mass of lightest $E_8$
particle $m_1$ is related to the longitudinal field
as $m_1 = C \lambda^{8/15}$, which plays the role of
IR cutoff energy scale. When down to lattice
we have a simple estimation $m_1^{lattice} = C' H'^{8/15} \approx 1.20$ meV where
$C' \approx 0.91C$ \cite{CASELLE2000667}.
The estimated $m_1^{lattice}$ is close to the
real experimental observation,
therefore, we can safely pick up the first peak in the experimental data as the corresponding $m_1$
in the $E_8$ field theory.
Now we can make a detailed comparison of
the analytical and experimental (numerical) zero-transfer-momentum DSF by using $m_1$ and the first peak of
experimental (and numerical data) to re-scale their DSF spectra, respectively.
The comparison as shown in Fig.~\ref{fig:E8_DSF},
shows excellent agreement between discrete and analytical results,
which allows us to confidently claim that the
$E_8$ physics is indeed observed in real experiments.
In the comparison, we should also note that when taking the space continuum limit,
$J$ for the nearest spin interaction in the lattice model indeed should serve as a UV energy cutoff for the applicability of the corresponding quantum field theory.
However, since $J$ is the energy scale for local spin interaction in the lattice, then, after taking the space continuum limit,
it actually appears as a cutoff scale for the energy of local or short-range fluctuations in the quantum field theory.
In the experiment (Fig.~\ref{fig:E8_DSF}),
the observed excitations at the zone center are coherent excitations involving large amounts of spins at long-range scale (zone center implies zero transfer momentum thus corresponding to long-range scale).
The excitation energy at such long-range scale cannot be simply considered
as the same energy for the local or short-range spin interaction.
Although it is highly nontrivial to accurately determine the energy
of local fluctuation (here it is the strength of nearest spin interaction)
from a many-body wavefunction,
it should be safe to claim that even the energy cost to get
a many-body coherent excitation at long-range scale is close to $J$,
on average the corresponding energy of local fluctuation should be still far smaller than $J$.
This should provide an underlying physical reason for the excellent comparison up to one $J$ as
shown in Fig.~\ref{fig:E8_DSF}.
Actually, even the energy cost is beyond $J$ for the long-range-scale
(corresponding to small (transfer) momentum) coherent excitation,
the field-theory prediction can still agree well with the lattice result,
as demonstrated in Ref.~\cite{Keselman_Dynamical_2020} which carefully
compares field theory results based on Bosonization with lattice calculation of the Heisenberg chain,
and in Ref.~\cite{Wybo_Quantum_2022} which carefully compares sine-Gordon field theory results
with lattice calculation of ladder XXZ chain.
Therefore, when considering the applicability of quantum field theory
to the corresponding lattice model, it is not simply about energy scale (or time scale),
the momentum scale (or length scale) also plays a crucial role.
\section{Conclusion}
\label{sec:conclusion}
This review details rich physics in the 1D spin-1/2 XXZ model with
the presence of various external fields and their experimental realization.
Because the experimental realization is
highly relevant to the quasi-1D AFM materials,
we first provide a systematic and pedagogical analysis for
setting up the effective Hamiltonian for the BCVO and SCVO,
which can also be applied to other
Co-based materials with cautions on possible
changes of symmetry for the local environment.
Then, we review recent theoretical progress on a variety of
magnetic excitaions of the 1D spin-1/2 XXZ model,
such as spinon (without magnetization),
(anti)psinon and string excitations (with finite magnetization),
and their experimental observations.
With the presence of transverse field, we further discuss
how the TFIC universality and the
exotic $E_8$ physics
(with additional small longitudinal field)
emerge near the corresponding 1D QCP, as well as their material realizations.
The 1D spin-1/2 XXZ model with various external fields contains
rich magnetic physics, whose exotic excitations
from low to high energy
are now successfully realized in real materials.
Those concrete progresses
can further inspire and bridge many research fields
such as quantum statistical field theory,
cold atom, and AdS/CFT etc..
Moreover, it is worth to further explore physical properties of those
exotic magnetic excitations, which will be an important step
toward practical control and application of those exotic magnetic
excitations.
\vskip 1.0 truecm
\noindent {\bf Acknowledgements}\\
The work is support by the Innovation Program for Quantum Science and Technology
No. 2021ZD0301900, the Natural Science Foundation of Shanghai with grant No. 20ZR1428400, and Shanghai Pujiang Program with grant No. 20PJ1408100. J.W. acknowledges additional support from a Shanghai talent program.
|
\section{Introduction}
In this paper, we consider solving the following distributed optimization problem by a group of agents $[n] := \crk{1,2,\cdots, n}$ connected over a network:
\begin{equation}
\label{eq:P}
\min_{x\in\mathbb{R}^p} f(x) = \frac{1}{n}\sum_{i=1}^n f_i(x),
\end{equation}
where $f_i$ is the local cost function known by agent $i$ only. Problem \eqref{eq:P} appears naturally in many machine learning and signal processing problems, where each $f_i$ represents an expected or empirical risk function corresponding to the local dataset of agent $i$. Solving problem \eqref{eq:P} in a decentralized manner over a multi-agent network has gained great interests in recent years, partly because
decentralization helps relieve the high latency at the central server in a centralized computing architecture \cite{nedic2018network}.
However, decentralized optimization algorithms may still suffer from the communication bottleneck under the huge size of modern machine learning models and/or limited bandwidth. For example, the gradients used for training LSTM models can be as large as $110.43$ MB \cite{koloskova2019decentralized}.
Communication compression techniques, such as quantization \cite{alistarh2017qsgd,bernstein2018signsgd,seide20141} and sparsification \cite{wangni2018gradient,stich2018sparsified}, are one of the most effective means for reducing the communication cost in distributed computation. The performance of several commonly used communication compression methods has been well studied under the centralized master-worker architecture, and the algorithms have been shown to enjoy comparable convergence rates with their non-compressed counterparts when equipped with various techniques such as error compensation \cite{seide20141,richtarik2021ef21,alistarh2018convergence,stich2018sparsified} and gradient difference compression \cite{mishchenko2019distributed}.
It is thus natural to integrate decentralization and communication compression to benefit from both techniques. The work in \cite{tang2018communication} combined distributed stochastic gradient descent (DSGD) with quantization for the first time, followed by \cite{chocosgd,koloskova2019decentralized} which extended the compressed DSGD method to fit a richer set of compression operators. The paper \cite{liu2021linear} equipped NIDS \cite{nids} with communication compression and demonstrated its linear convergence rate when the objective functions are smooth and strongly convex. More recently, the papers \cite{liao2021compressed,song2021compressed} considered a class of linearly convergent decentralized gradient tracking method with communication compression that applies to general directed graphs; see also \cite{zhang2021innovation,yi2022communication} for more related works.
It is worth noting that introducing decentralization may slow down the algorithmic convergence as the network size grows when compared to the centralized master-worker architecture \cite{pu2020asymptotic}. Therefore, an important question arises when considering decentralized optimization with communication compression, that is,
\textit{can a compressed decentralized (stochastic) gradient method achieve similar convergence rate compared to its centralized counterpart?}
Such a question has been answered positively in \cite{tang2018communication} first with several works following \cite{chocosgd,koloskova2019decentralized,singh2022sparq,singh2021squarm,vogels2020practical}. However, these methods are all variants of DSGD which suffer from the data heterogeneity \cite{nedic2009distributed,pu2021sharp,lian2017can}. As a result, significant amount of transient times are often required for achieving comparable convergence rate with centralized SGD.
In this paper, we consider a novel method for solving Problem \eqref{eq:P}, termed ``compressed exact diffusion algorithm with adaptive stepsizes (CEDAS)", which is adapted from EXTRA \cite{shi2015extra}, exact diffusion \cite{yuan2018exact} and the LEAD algorithm \cite{liu2021linear}. We analyze CEDAS under an unbiased compression operator for both smooth strongly convex objective functions and smooth nonconvex objective functions.
We are able to show that under both scenarios, the CEDAS method enjoys linear speedup similar to a centralized SGD method without communication compression.
In particular, the performance of CEDAS outperforms the state-of-the-art algorithms in terms of the transient time (with respect to the graph specifics) to achieve the convergence rate of centralized SGD.
These results suggest that CEDAS benefits from both communication compression and decentralization better than the existing algorithms.
\subsection{Related Works}
There is a vast literature on solving Problem \eqref{eq:P} under the communication restricted settings. For example, under the centralized master-worker architecture, the works in \cite{seide20141,strom2015scalable,alistarh2017qsgd,de2017understanding,wangni2018gradient,bernstein2018signsgd,konevcny2018randomized,tang2019doublesqueeze,mishchenko2021intsgd,richtarik2021ef21,huang2022lower} have considered transmitting quantized or sparse information to the master node to save the communication costs.
In the decentralized setting, existing compressed gradient methods can be classified into three main categories: (1) D(S)GD variants \cite{chocosgd,tang2018communication,singh2021squarm,koloskova2019decentralized,singh2022sparq,tang2019texttt}; (2) primal-dual like methods \cite{liu2021linear,li2021decentralized,lan2020communication,kovalev2021linearly,zhang2021innovation,yi2022communication}; (3) gradient tracking based algorithms \cite{liao2021compressed,zhao2022beer,song2021compressed,xu2022quantized,xiong2021quantized}. Compared to D(S)GD variants, the latter two types of methods can relieve from data heterogeneity. Specifically, D(S)GD variants cannot achieve exact convergence to the optimal solution under a constant stepsize even when the gradient variance goes to zero. By contrast, the methods considered in \cite{xiong2021quantized,michelusi2022finite,liu2021linear,li2021decentralized,liao2021compressed,song2021compressed,kovalev2021linearly,zhang2021innovation,xu2022quantized} enjoy linear convergence rate under smooth strongly convex objective functions with deterministic gradients.
In previous works, several decentralized stochastic gradient methods \cite{lian2017can,pu2021sharp,huang2021improving,pmlr-v80-tang18a,spiridonoff2020robust,pu2021distributed,alghunaim2021unified,yuan2021removing,koloskova2021improved}, including the compressed D(S)GD type methods \cite{chocosgd,tang2018communication,koloskova2019decentralized,singh2022sparq,singh2021squarm,vogels2020practical,tang2019texttt}, have been shown to achieve linear speedup and enjoy the so-called ``asymptotic network independent property'' \cite{pu2020asymptotic}. In other words, the error term related to the network is negotiable after a finite number of iterations (transient time).
Such a property is desirable since it guarantees the number of iterations to reach a high accuracy level will not grow as fast as the size of the network increases.
Currently, the shortest transient times achieved by decentralization stochastic gradient methods without communication compression are $\order{n/(1-\lambda_2)}$ for smooth strongly convex objective functions \cite{huang2021improving,yuan2021removing} and $\order{n^3/(1-\lambda_2)^2}$ for smooth nonconvex objective functions \cite{alghunaim2021unified}, respectively, where $(1-\lambda_2)$ defines the spectral gap of the mixing matrix.
Regarding compressed decentralized algorithms based on stochastic gradients, we compare the performance of those that achieve ``asymptotic network independent property'' in Table \ref{tab:kt} assuming that unbiased compressors are utilized. Note that in the literature, there are two commonly considered conditions on the compression compressors: (i) the signal noise ratio (SNR) between the compressed value and the original value is bounded by some constant $C> 0$, and the compressor is unbiased (see Assumption \ref{ass:ub_op} for details). We refer to such compressors as unbiased compressors; (ii) the corresponding SNR is bounded by $(1-\delta), \delta \in (0,1)$ (see Assumption \ref{ass:b_op}). We refer to such compressors as biased ones. Examples of the compressors that satisfy the above two assumptions can be found in \cite{beznosikov2020biased,safaryan2022uncertainty,xu2020compressed} and the references therein. Unbiased and biased compressors can be transformed into the other type under certain mechanisms \cite{horvath2021a}.
Assuming biased compressors, Choco-SGD achieves a transient time that behaves as $\order{n/((1-\lambda_2)^4\delta^2}$ when the objective function $f$ is smooth strongly convex, and its transient time for smooth nonconvex objective function is $\order{n^3/((1-\lambda_2)^8\delta^4}$. SPARQ-SGD \cite{singh2022sparq} shares the same transient time with Choco-SGD as the method reduces to Choco-SGD without event-triggering and communication skip. The results listed in Table \ref{tab:kt} are based on an unbiased compressor satisfying Assumption \ref{ass:b_op} that has been transformed into a biased one following the mechanism in \cite{horvath2021a}.
However, among the communication compressed algorithms, only variants of D(S)GD have been shown to achieve the asymptotic network independent property. For example, LEAD \cite{liu2021linear} converges in the order of $\order{1/[(1-\lambda_2)^2 k]}$ (where $k$ counts the iteration number) when decreasing stepsizes are employed. Therefore, when $(1-\lambda_2)$ is small, more communication rounds are required for LEAD to achieve the same accuracy level compared to the centralized SGD method.
\subsection{Main Contribution}
The main contribution of this work is four-fold.
Firstly, we develop a new compressed decentralized stochastic gradient method, termed ``compressed exact diffusion with adaptive stepsizes (CEDAS)", and show the method asymptotically achieves comparable convergence rate as centralized SGD for both smooth strongly convex objective functions and smooth nonconvex objective functions under unbiased compression operators.
Secondly, we characterize the transient time of CEDAS to reach the convergence rate of centralized SGD under unbiased compressors, which behaves as $\order{nC^3/(1-\lambda_2)^{2}}$ for smooth strongly convex objective functions and $\order{n^3C^6/(1-\lambda_2)^4}$ for smooth nonconvex objective functions (see Theorem \ref{thm:total1} and Theorem \ref{thm:ncvx_ng}).
Particularly, the derived transient times for CEDAS are the shortest (with respect to the graph specifics) compared to the state-of-the-art works to the best of our knowledge (see Table \ref{tab:kt}).
Thirdly, when no compression is performed, the derived transient times for CEDAS are consistent with those given in \cite{huang2021improving,alghunaim2021unified,yuan2021removing}, which are the shortest transient times so far among the decentralized stochastic gradient methods without communication compression.
Finally, compared to the closely related algorithm LEAD \cite{liu2021linear} which has been shown to be successful for minimizing smooth strongly convex objectives,
CEDAS is shown to work with both smooth nonconvex objective functions and smooth strongly convex objective functions.
Moreover, the convergence results for CEDAS are superior to LEAD under stochastic gradients. It is worth noting that obtaining the improved results is nontrivial because: 1) even without compression, analyzing CEDAS can be much more involved compared to studying DSGD variants; 2) the compression-related terms are nonlinear and require careful treatment; 3) dealing with nonconvexity is challenging. We have utilized various techniques involving constructing proper Lyapunov functions to demonstrate the results.
\begin{table}
\centering
\begin{tabular}{@{}cccc@{}}
\toprule
Algorithm & $f$ & Transient Time \\ \midrule
Choco-SGD \cite{koloskova2019decentralized} & NCVX & $\frac{n^3C^4}{(1-\lambda_2)^8}$ \\
Choco-SGD \cite{chocosgd} & SCVX & $\frac{nC^2}{(1-\lambda_2)^4}$ \\
DeepSqueeze \cite{tang2019texttt} & NCVX & $\frac{n^3C^6}{(1-\lambda_2)^{12}}$ \\
SQuARM-SGD \cite{singh2021squarm} & NCVX & $\frac{n^3C^4}{(1-\lambda_2)^8}$ \\
SPARQ-SGD \cite{singh2022sparq} & $f_i$ SCVX & $\frac{nC^2}{(1-\lambda_2)^4}$ \\
SPARQ-SGD \cite{singh2022sparq} & NCVX & $\frac{n^3C^4}{(1-\lambda_2)^8}$\\ \bottomrule
This work & SCVX & $\boldsymbol{\frac{nC^3}{(1-\lambda_2)^2}}$\\
This work & NCVX & $\boldsymbol{\frac{n^3 C^6}{(1-\lambda_2)^4}}$
\\ \bottomrule
\end{tabular}
\caption{Comparison among the existing compressed decentralized stochastic gradient methods with the asymptotic network independent property. The transient times are presented by hiding some constants independent of $(1-\lambda_2)$, n, and $C$, where $C$ is the compressor-related parameter (see Subsection \ref{subsec:ass}) and $(1-\lambda_2)$ defines the spectral gap of the mixing matrix. Each $f_i$ is assumed to be a smooth function. NCVX stands for nonconvex functions and SCVX represents strongly convex functions.
For the works that assume biased compression operators, we transform the unbiased compressors into biased ones and compute the compressor-related constant correspondingly according to \cite{horvath2021a}.}
\label{tab:kt}
\end{table}
\subsection{Notation}
\label{sec:notations}
In this paper, we use column vectors by default. Let $x_{i,k}\in\mathbb{R}^p$ denote the local copy of agent $i$ at the $k-$th iteration. For the ease of presentation, we use bold lowercase letters and capital letters to denote stacked variables. For example,
\begin{align*}
\mathbf{x}_k&:= \prt{x_{1,k}, x_{2,k},\cdots, x_{n,k}}^{\intercal} \in \mathbb{R}^{n\times p}\\
H_k &:= \prt{h_{1,k}, h_{2,k},\cdots, h_{n,k}}^{\intercal} \in\mathbb{R}^{n\times p}\\
\nabla F(\mathbf{x}_k) &:= \prt{\nabla f_1(x_{1,k}), \nabla f_2(x_{2,k}), \cdots, \nabla f_n(x_{n,k})}^{\intercal}\in\mathbb{R}^{n\times p}.
\end{align*}
We use $\bar{x}\in\mathbb{R}^p$ to denote the averaged (among agents) variables, e.g., $\bar{x}_k = \sum_{i=1}^n x_{i,k}/n$ is defined as the average of all the $n$ agents' solutions at the $k-$iteration. Notation $\inpro{a,b}$ represents the inner product for two vectors $a, b\in\mathbb{R}^p$, while the inner product $\inpro{A,B}$ for two matrices $A,B\in\mathbb{R}^{n\times p}$ is defined as $\inpro{A,B}:= \sum_{i=1}^n \inpro{A_i,B_i}$, where $A_i$ stands for the $i-$row of $A$.
\subsection{Organization}
The remaining parts of this paper are organized as follows. In Section \ref{sec:setup}, we introduce the standing assumptions and the CEDAS algorithm. We then conduct the convergence analysis under smooth strongly convex objective functions in Section \ref{sec:scvx}. In Section \ref{sec:ncvx}, we perform the analysis for smooth nonconvex objective functions. Section \ref{sec:exp} provides numerical examples that corroborate the theoretical results.
\section{Setup}
\label{sec:setup}
In this section, we first introduce the standing assumptions in Subsection \ref{subsec:ass} with some necessary discussions.
Then we present the new algorithm, termed ``compressed exact diffusion with adaptive stepsizes (CEDAS)", in Subsection \ref{subsec:lead_ds} along with some preliminary analysis.
\subsection{Assumptions}
\label{subsec:ass}
In this part, we introduce the standing assumptions for this work. Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:b_op} are related to the network structure, stochastic gradients and compression operators, respectively. Regarding the objective functions, we consider Assumption \ref{ass:fi} and Assumptions \ref{ass:smooth} separately in Section \ref{sec:scvx} and Section \ref{sec:ncvx}.
We start with stating the assumption regarding the multi-agent network structure. Suppose the agents are connected over a network $\mathcal{W} = (\mathcal{V}, \mathcal{E})$, where $\mathcal{V} = [n]$ represents the sets of nodes (agents), and $\mathcal{E}\subset \mathcal{V}\times \mathcal{V}$ denotes the set of edges linking different nodes in the network. We also denote $\mathcal{N}_i=\crk{j| (i,j)\in \mathcal{E}}$ the set of neighbors of agent $i$. The matrix $W = (w_{ij})\in\mathbb{R}^{n\times n}$ is the mixing matrix compliant to the network $\mathcal{W}$. In particular, we assume the following condition on the network $\mathcal{W}$ and matrix $W$.
\begin{assumption}\label{ass:W}
The graph $\mathcal{W}$ is undirected and strongly connected. There exists a link from $i$ and $j$ ($i\neq j$) in $\mathcal{W}$ if and only if $w_{ij},w_{ji}>0$; otherwise, $w_{ij}=w_{ji}=0$. The mixing matrix $W$ is nonnegative, symmetric and stochastic ($W\mathbf{1} =\mathbf{1}$).
\end{assumption}
Assumption \ref{ass:W} is common in the decentralized optimization literature; see, e.g., \cite{lian2017can,pu2021sharp,yuan2018exact}. The conditions imply that the eigenvalues of $W$, denoted as $\lambda_1\geq \lambda_2\geq \cdots\geq\lambda_n$, lie in the range of $(-1,1]$. The term $(1-\lambda_2)$ is called the spectral gap of the graph/mixing matrix, which generally gets closer to $0$ when the connectivity of $\mathcal{W}$ is worse \cite{nedic2018network}.
Regarding the stochastic gradients, we consider the following standard assumption.
\begin{assumption}
\label{ass:sgrad}
For all iteration $k\ge 0$, each agent $i$ is able to obtain noisy gradient $\nabla f_i(x_{i,k};\xi_{i,k})$ given $x_{i,k}$, where
each random vector $\xi_{i,k}\in\mathbb{R}^q$ is independent across $i\in\mathcal{N}$. In addition,
\begin{equation}
\label{condition: gradient samples}
\begin{split}
& \mathbb{E}[\nabla f_i(x_{i,k};\xi_{i,k})\mid x_{i,k}] = \nabla f_i(x_{i,k}),\\
& \mathbb{E}[\|\nabla f_i(x_{i,k};\xi_{i,k})-\nabla f_i(x_{i,k})\|^2\mid x_{i,k}] \le \sigma^2,\quad\hbox{\ for some $\sigma>0$}.
\end{split}
\end{equation}
\end{assumption}
Stochastic gradients are common in machine learning problems. For example, when each agent $i$ randomly samples a minibatch of data points from its local dataset with replacement at every iteration and evaluate the gradient on the minibatch, an unbiased noisy gradient that is independent across the agents can be obtained.
\begin{remark}
The bounded variance assumption may be hard to verify in practice, in which case a relaxed condition can be considered as follows (see, e.g., \cite{pu2021sharp,huang2021improving}):
\begin{equation*}
\condE{\norm{\nabla f_i(x_{i,k};\xi_{i,k}) - \nabla f_i(x_{i,k})}^2}{x_{i,k}} \leq A\sigma^2 + B\norm{\nabla f_i(x_{i,k})}^2, \ \text{for some }A, B, \sigma\in\mathbb{R}_{++}.
\end{equation*}
We expect that such a relaxation does not affect the main results, since under $L-$smoothness, we have
\begin{align*}
\sum_{i=1}^n \norm{\nabla f_i(x_{i,k})}^2 &\leq 2L^2\sum_{i=1}^n \norm{x_{i,k} - \hat{x}}^2 + \sum_{i=1}^n\norm{\nabla f_i(\hat{x})}^2, \text{ for some } \hat{x},
\end{align*}
and the term $\norm{x_{i,k} - \hat{x}}^2$
can be shown to converge to $0$ while the term $\norm{\nabla f_i(\hat{x})}^2$ can be controlled by the stepsize ( the choice of $\hat{x}$ depends on whether the objective function is convex or not).
Thus, we consider Assumption \ref{ass:sgrad} to simplify the presentation of the paper.
Note that Assumption \ref{ass:sgrad} is more general than assuming bounded stochastic gradients \cite{chocosgd,koloskova2019decentralized}.
\end{remark}
We now introduce the conditions on the compression operators. If a compressor $\mathcal{C}$ satisfies Assumption \ref{ass:ub_op}, we denote $\mathcal{C}\in \mathbb{U}(C)$ for simplicity. Similarly, if $\mathcal{C}$ satisfies Assumption \ref{ass:b_op}, we write $\mathcal{C}\in\mathbb{B}(\delta)$.
\begin{assumption}
\label{ass:ub_op}
The compression operator $\compress{\cdot}:\mathbb{R}^p\rightarrow\mathbb{R}^p$ is unbiased, i.e., $\condE{\compress{x}}{x} = x$ and satisfies
\begin{equation}
\label{eq:com_C}
\condE{\norm{\compress{x} - x}^2}{x}\leq C\norm{x}^2,\quad C>0,\forall x\in\mathbb{R}^p.
\end{equation}
\end{assumption}
\begin{assumption}
\label{ass:b_op}
The compression operator $\compress{\cdot}:\mathbb{R}^p\rightarrow\mathbb{R}^p$ satisfies
\begin{equation}
\label{eq:com_delta}
\condE{\norm{\compress{x} - x}^2}{x} \leq (1-\delta)\norm{x}^2, \quad \delta\in(0,1],\forall x\in\mathbb{R}^p.
\end{equation}
\end{assumption}
The above two types of compressors can be transformed between each other in the following way.
On one hand, note that given $\mathcal{C}\in\mathbb{U}(C)$, we can construct $\mathcal{C}'\in \mathbb{B}(1/(C+1))$ with $\mathcal{C}':= \mathcal{C}/(C+1)$. Therefore, an unbiased compressor can be transformed into a biased one.
On the other hand, Lemma \ref{lem:b2ub} below introduces a mechanism for constructing an unbiased compressor from any biased compressor.
Such an idea first appears in \cite{horvath2021a}.
\begin{lemma}
\label{lem:b2ub}
For any compressor $\mathcal{C}_1\in\mathbb{B}(\delta_1)$, we can choose a compressor $\mathcal{C}_2\in\mathbb{U}(C_2)$ so that an introduced compressor $\mathcal{C}: \mathbb{R}^p \rightarrow \mathbb{R}^p$ defined by
\begin{equation}
\label{eq:in_op}
\compress{x} := \mathcal{C}_1(x) + \mathcal{C}_2\prt{x - \mathcal{C}_1(x)}
\end{equation}
satisfies Assumption \ref{ass:ub_op} with $C = C_2(1-\delta_1)$.
\end{lemma}
\begin{remark}
The mechanism in Lemma \ref{lem:b2ub} allows users to apply algorithms that were not compatible with Assumption \ref{ass:b_op} originally. The price to pay is more computations related to compressing $(x - \mathcal{C}_1(x))$ and sending more bits. However, the compression parameter is decreased from $C_2$ to $C_2(1-\delta_1)$ compared to using $\mathcal{C}_2$ directly. For the choice of $\mathcal{C}_2$, it is preferable to choose $\mathcal{C}_2$ with similar compression complexity as $\mathcal{C}_1$ according to \cite{horvath2021a}. This would at most double the bits to send per iteration. The proof of Lemma \ref{lem:b2ub} is deferred to Appendix \ref{app:b2ub}.
\end{remark}
Assumptions \ref{ass:fi} and \ref{ass:smooth} below formally define smooth strongly convex objective functions and smooth nonconvex objective functions, respectively.
\begin{assumption}\label{ass:fi}
Each $f_i:\mathbb{R}^p\rightarrow \mathbb{R}$ has $L$-Lipschitz continuous gradients, and the average function $f:= \frac{1}{n}\sum_{i=1}^n f_i:\mathbb{R}^p\rightarrow \mathbb{R}$ is $\mu$-strongly convex, i.e., $\forall x, x'\in\mathbb{R}^p$,
\begin{align*}
\Vert \nabla f_i(x)-\nabla f_i(x')\Vert\leq L\Vert x-x'\Vert,\quad \forall i,
\end{align*}
\begin{align*}
\langle\nabla f(x)-\nabla f(x'), x-x'\rangle\geq \mu \Vert x-x'\Vert^2.\\
\end{align*}
\end{assumption}
\begin{assumption}
\label{ass:smooth}
Each $f_i:\mathbb{R}^p\rightarrow \mathbb{R}$ has $L$-Lipschitz continuous gradients, i.e., $\forall x, x'\in\mathbb{R}^p$,
\begin{align*}
\Vert \nabla f_i(x)-\nabla f_i(x')\Vert\leq L\Vert x-x'\Vert,\quad \forall i.
\end{align*}
In addition, each $f_i$ is bounded from below by $\underline{f}_i$. Define $f^{\text{inf}}:= \frac{1}{n}\sum_{i=1}^n \underline{f}_i$, then $f(x)\geq f^{\text{inf}}, \forall x\in\mathbb{R}^p$.
\end{assumption}
In summary, the above assumptions are common and standard. In particular, Lemma \ref{lem:b2ub} provides a mechanism that generalizes the applicability of the proposed CEDAS algorithm introduced in the next section.
\subsection{Algorithm}
\label{subsec:lead_ds}
In this part, we introduce the CEDAS algorithm and discuss the strategies for analyzing CEDAS in light of the previous work \cite{huang2021improving}.
The procedures of CEDAS are stated in Algorithm \ref{alg:lead_ds}.
\begin{algorithm}
\caption{Compressed exact diffusion with adaptive stepsizes (CEDAS)}
\label{alg:lead_ds}
\begin{algorithmic}[1]
\Require Stepsizes ${\eta_k}$, parameters $\gamma$ and $\alpha$, and initial values $x_{i,-1}, h_{i,0}, i\in[n]$.
\For{Agent $i$ in parallel}
\State $\ul{d}_{i,0} = 0$
\State $(h_w)_{i, 0} = \sum_{j\in \mathcal{N}_i\cup\crk{i}}w_{ij}h_{j,0}$
\State Compute $\nabla f_i(x_{i,-1};\xi_{i, -1})$ and $x_{i,0} = x_{i,-1} - \eta_{-1} \nabla f_i(x_{i,-1};\xi_{i, -1})$
\EndFor
\For{$k = 0,1,\cdots, K-1$, agent $i$ in parallel}
\State Compute $\nabla f_i(x_{i,k};\xi_{i,k})$
\State $y_{i,k} = x_{i,k} - \eta_k\nabla f_i(x_{i,k};\xi_{i,k}) - \ul{d}_{i,k}$ \label{line:y}
\State $\hat{y}_{i,k}, (\hat{y}_w)_{i,k}, h_{i,k+1}, (h_w)_{i,k + 1}$ = COMM($y_{i,k}, h_{i,k}, (h_w)_{i,k},\alpha$) \label{line:comm}
\State $\ul{d}_{i,k + 1} = \ul{d}_{i,k} + \frac{\gamma}{2}\prt{\hat{y}_{i,k} - (\hat{y}_w)_{i,k}}$\label{line:d}
\State $x_{i,k + 1} = x_{i,k} - \eta_k\nabla f_i(x_{i,k};\xi_{i,k}) - \ul{d}_{i, k + 1}$\label{line:x}
\EndFor
\State Output $x_{i, K}$.
\end{algorithmic}
\end{algorithm}
{
\floatname{algorithm}{Procedure}
\setcounter{algorithm}{0}
\begin{algorithm}
\caption{COMM($y_i, h_i, (h_w)_i, \alpha$)}
\label{alg:comm}
\begin{algorithmic}[1]
\State $q_i = \compress{y_i - h_i}$
\State $\hat{y}_i^+ = h_i + q_i$
\For{Neighbors $j\in \mathcal{N}_i$}
\State Send $q_i$ and receive $q_j$
\State $(\hat{y}_w)_i^+ = (h_w)_i + \sum_{j\in \mathcal{N}_i\cup\crk{i}}w_{ij} q_j$
\State $h_i^+ = (1-\alpha) h_i + \alpha \hat{y}_i^+$
\State $(h_w)_i^+ = (1-\alpha)(h_w)_i + \alpha (\hat{y}_w)_i^+$
\EndFor
\State Output $\hat{y}_i^+, (\hat{y}_w)_i^+, h_i^+, (h_w)_i^+$
\end{algorithmic}
\end{algorithm}
}
We first discuss the intuitive idea for the CEDAS algorithm.
With the notations in Subsection \ref{sec:notations}, Problem \eqref{eq:P} can be equivalently written as
\begin{equation}
\label{eq:P_cons}
\min_{\mathbf{x}\in\mathbb{R}^{n\times p}} \sum_{i=1}^nf_i(x_i), \ (I-W)\mathbf{x} = \mathbf{0}, \ \mathbf{x} = (x_1, x_2,\cdots, x_n)^{\intercal}.
\end{equation}
As mentioned, for example, in \cite{huang2021improving,yuan2018exact,xu2021distributed}, we can perform the following primal-dual-like update to solve \eqref{eq:P_cons}:
\begin{subequations}
\label{eq:pd}
\begin{align}
x_{i, k + 1} &= \sum_{j=1}^nw_{ij} \prt{x_{j,k} - \eta_k \nabla f_i(x_{i,k};\xi_{i,k})} - \tilde{d}_{i,k}\label{eq:pd1}\\
\tilde{d}_{i,k+1} &= \tilde{d}_{i,k} + x_{i,k+1} - \sum_{j=1}^n w_{ij}x_{i,k +1}.\label{eq:pd2}
\end{align}
\end{subequations}
The CEDAS algorithm can be viewed as equipping \eqref{eq:pd} with communication compression, where Procedure \ref{alg:comm} combines communication and compression for communication efficiency. More specifically, Line \ref{line:y} in Algorithm \ref{alg:lead_ds} is corresponding to the update \eqref{eq:pd1} without mixing among the agents. Line \ref{line:comm} produces the compressed version of $y_{i,k}$ as $\hat{y}_{i,k}$ and the mixed version of $\hat{y}_{i,k}$ as $(\hat{y}_w)_{i,k}$ in CEDAS. This procedure, which previously appears in \cite{liu2021linear}, indicates that $(\hat{y}_w)_{i,k} = \sum_{j=1}^n w_{ij}\hat{y}_{j,k}$ and $(h_w)_{i,k} = \sum_{j=1}^nw_{ij}h_{j,k}$. Then, Line \ref{line:d} performs the same step as \eqref{eq:pd2} based on the compressed information with an additional parameter $\gamma$. Such a parameter is introduced to control the so-called consensus error. In particular, Line \ref{line:d} is essentially performing
\begin{align*}
\ul{d}_{i, k + 1} = \ul{d}_{i,k} + \frac{\gamma}{2}\prt{\hat{y}_{i,k} - \sum_{j=1}^nw_{i,j}\hat{y}_{j,k}},
\end{align*}
which is similar to \eqref{eq:pd2}. Finally, Line \ref{line:x} (together with Line \ref{line:y}) performs an update similar to \eqref{eq:pd1} with mixed information. Note in procedure \ref{alg:comm}, $q_i = \compress{y_i - h_i}$ is the only variable to be transmitted by agent $i$.
Compared with the LEAD algorithm (Algorithm \ref{alg:lead} in Appendix \ref{app:lead}),
CEDAS employs diminishing stepsizes so that the expected error $\frac{1}{n}\sum_{i=1}^n\mathbb{E}\brk{\norm{x_{i,k} - x^*}^2}$ decreases to $0$ at an order optimal $\mathcal{O}_k\prt{\frac{1}{k}}$ rate with stochastic gradients under strongly convex objective functions. Note that the update of the term $a_{i,k}$ in LEAD involves computing $1/\eta$ which is not compatible with diminishing stepsizes $\crk{\eta_k}$ as $\eta_k$ goes to zero. Therefore, we consider a different update for the corresponding term $d_{i,k}$ in Algorithm \ref{alg:lead_ds}.
\subsection{Preliminary Analysis}
The compact form of Algorithm \ref{alg:lead_ds} is given in \eqref{eq:lead_ds} below, based on which we perform some preliminary analysis on CEDAS.
\begin{subequations}
\label{eq:lead_ds}
\begin{align}
\mathbf{y}_k &= \mathbf{x}_k - \eta_k G_k - \ul{D}_k\\
\ul{D}_{k + 1} &= \ul{D}_k + \frac{\gamma}{2}\prt{I-W}\hat{\mathbf{y}}_k = \ul{D}_k + \frac{\gamma}{2}(I-W)\prt{H_k + \compress{\mathbf{y}_k - H_k}}\label{eq:Dk0}\\
\mathbf{x}_{k + 1} &= \mathbf{x}_k - \eta_k G_k - \ul{D}_{k + 1}\label{eq:xk0}.
\end{align}
\end{subequations}
Inspired by \cite{huang2021improving,liu2021linear}, we introduce the compression error $E_k := \compress{\mathbf{y}_k - H_k} -(\mathbf{y}_k - H_k)$ and a new mixing matrix $\tilde{W} := I - \frac{\gamma}{2}(I-W)$. It follows that
\begin{align}
\ul{D}_{k + 1} &= \ul{D}_k + \frac{\gamma}{2}(I-W)E_k + \frac{\gamma}{2}(I-W)\mathbf{y}_k\label{eq:Dk2}\\
&= \ul{D}_k + \frac{\gamma}{2}(I-W)E_k + \frac{\gamma}{2}(I-W)(\mathbf{x}_k - \eta_k G_k - \ul{D}_k)\nonumber\\
&= \tilde{W} \ul{D}_k + (I-\tilde{W}) E_k + (I-\tilde{W})(\mathbf{x}_k - \eta_k G_k)\label{eq:Dk1}.
\end{align}
From \eqref{eq:Dk0}, we have $\ul{D}_k\in \Span{I-W}$ for any $k\geq 0$ with initialization $\ul{D}_0 \in \Span{I-W}$. Noting the relation $I-\tilde{W} = \frac{\gamma}{2}(I-W)$, we obtain Lemma \ref{lem:tW} which directly results from the definition of $\tilde{W}$ for $\gamma\in(0,1)$. The proof is omitted.
\begin{lemma}
\label{lem:tW}
Let Assumption \ref{ass:W} hold and $\gamma \in (0, 1)$. We have
\begin{enumerate}[label=\alph*)]
\item $\tilde{W} = I -\frac{\gamma}{2}(I-W)\in\mathbb{R}^{n\times n}$ is positive definite, symmetric, and stochastic;
\item $\Span{I-\tilde{W}} = \Span{I-W}$;
\item Let $\lambda_1\geq \lambda_2\geq \cdots\geq \lambda_n$ be the eigenvalues of $W$, then $\tilde{\lambda}_i = 1 - \frac{\gamma}{2}(1-\lambda_i), i=1,2,\cdots, n$ are the eigenvalues of $\tilde{W}$ and $1=\tilde{\lambda}_1>\tilde{\lambda}_2\geq \tilde{\lambda}_3\geq \cdots\geq \tilde{\lambda}_n>0$;
\item $\tilde{W}$ and $(I-\tilde{W})$ commute.\label{ite:commute}
\end{enumerate}
\end{lemma}
We now introduce new iterates $\crk{\mathbf{s}_k}$ to facilitate the analysis. Similar techniques can be found, e.g., in \cite{xu2021distributed,huang2021improving,yuan2018exact}.
\begin{equation}
\label{eq:Dksk}
\ul{D}_k = (I-\tilde{W})^{\frac{1}{2}} \mathbf{s}_k.
\end{equation}
With \eqref{eq:Dksk} and Lemma \ref{lem:tW} in hand, we discuss how to obtain the updates in terms of $\mathbf{s}_k$ and $\mathbf{x}_k$. Invoking \ref{ite:commute} in Lemma \ref{lem:tW}, equation \eqref{eq:Dk1} becomes
\begin{equation}
\label{eq:sk}
(I-\tilde{W})^{\frac{1}{2}} \mathbf{s}_{k + 1} = (I-\tilde{W})^{\frac{1}{2}}\tilde{W} \mathbf{s}_k + (I-\tilde{W}) E_k + (I-\tilde{W})(\mathbf{x}_k - \eta_k G_k).
\end{equation}
Combining \eqref{eq:xk0} and \eqref{eq:Dk1} leads to
\begin{align}
\mathbf{x}_{k + 1} &= \mathbf{x}_k - \eta_k G_k -\tilde{W} \ul{D}_k- (I-\tilde{W}) E_k - (I-\tilde{W})(\mathbf{x}_k - \eta_k G_k)\nonumber\\
&= \tilde{W} (\mathbf{x}_k - \eta_k G_k) -\tilde{W} \ul{D}_k - (I-\tilde{W})E_k\nonumber\\
&= \tilde{W} (\mathbf{x}_k - \eta_k G_k)-\tilde{W} (I-\tilde{W})^{\frac{1}{2}}\mathbf{s}_k - (I-\tilde{W}) E_k\label{eq:xk1}.
\end{align}
Let $\tilde{V} := (I-\tilde{W})^{1/2}$. We are ready to derive the following recursions for the iterates $\crk{\mathbf{x}_k, \mathbf{s}_k}$ according to \eqref{eq:sk} and \eqref{eq:xk1}:
\begin{subequations}
\label{eq:edas_c}
\begin{align}
\mathbf{x}_{k + 1} &= \tilde{W} (\mathbf{x}_k - \eta_k G_k)-\tilde{W} \tilde{V}\mathbf{s}_k - (I-\tilde{W}) E_k\label{eq:edas_cx}\\
\mathbf{s}_{k + 1} &= \tilde{W} \mathbf{s}_k + \tilde{V} \prt{\mathbf{x}_k - \eta_k G_k} + \tilde{V} E_k\label{eq:edas_cs}.
\end{align}
\end{subequations}
\begin{remark}
Relation \eqref{eq:edas_c} resembles that of EDAS \cite{huang2021improving} when the compression error $E_k = 0$ with the new mixing matrix $\tilde{W}$. However, analyzing the compression error term $E_k$ is not trivial as the compression operator is nonlinear. In particular, it prevents us from applying the results in \cite{huang2021improving} and calls for additional procedures; see Lemma \ref{lem:yh0_sim} for the case of smooth strongly convex objective functions and Lemma \ref{lem:ncvx_yh} for the case of smooth nonconvex objective functions.
\end{remark}
Following the update \eqref{eq:edas_c}, we have the following optimality condition for solving Problem \eqref{eq:P}. It will also guide us to define the error terms for analyzing the performance of CEDAS.
\begin{lemma}
\label{lem:opt_cond}
Let Assumption \ref{ass:W} hold, if there exists some points $(\mathbf{x}^*, \mathbf{s}^*_k)$ that satisfies:
\begin{subequations}
\begin{align}
\eta_k \nabla F(\mathbf{x}^*) + \tilde{V} \mathbf{s}^*_k &= 0\label{eq:x_opt}\\
\tilde{V} \mathbf{x}^* &=0,\label{eq:s_opt}
\end{align}
\end{subequations}
where $\mathbf{x}^* = (x_1^*, x_2^*,\cdots, x_n^*)^{\intercal}\in\mathbb{R}^{n\times p}$ and $\tilde{V}^2 = I-\tilde{W}$. Then, it holds that $x_1^* = x_2^* = \cdots = x^*_n = x^*$, where $x^*\in\mathbb{R}^p$ is a stationary point to Problem \eqref{eq:P}.
\end{lemma}
\begin{proof}
See Appendix \ref{app:fixed_scvx_C}.
\end{proof}
\section{Convergence Analysis: Strongly Convex Case}
\label{sec:scvx}
In this section, we present the convergence analysis for CEDAS when the objective function is strongly convex and smooth, in which case Problem \eqref{eq:P} has an optimal solution $x^*\in\mathbb{R}^p$.
We first introduce the transformed error dynamics stated in Lemma \ref{lem:error_edas_ct}. Since the derivations are similar to those in \cite{huang2021improving}, we included them in Appendix \ref{app:transformed_error} for completeness.
Then, we present the coupled recursions corresponding to three sources of errors in Lemmas \ref{lem:cons0}-\ref{lem:opt0}. In particular, compared with the related works, the compression error term $\mathbb{E}[\norm{\mathbf{y}_k - H_k}^2]$ requires careful treatment.
In light of these preliminary results, Subsection \ref{subsec:sublinear} demonstrates the sublinear convergence rate $\order{1/(nk)}$ for the total expected error of Algorithm \ref{alg:lead_ds} with decreasing stepsize $\eta_k = \order{1/k}$. Such a result shows that CEDAS achieves the asymptotic network independent property, that is, its performance is comparable with centralized SGD in the long run. In addition, we derive the transient time of CEDAS in Subsection \ref{subsec:KT} consistent with the discussions in Subsection \ref{subsec:lead_ds}.
We highlight the technical challenges of analyzing CEDAS for smooth strongly convex objective functions compared to the uncompressed method in \cite{huang2021improving} and the compressed cohorts \cite{liu2021distributed,chocosgd,liao2021compressed}. The compression error term $\mathbb{E}[\norm{\mathbf{y}_k - H_k}^2]$ is the major cause of difficulties for the analysis due to the nonlinearity of the compression operators. Constructing the recursion of such a term is nontrivial due to the complicated update of CEDAS and Procedure \ref{alg:comm}. In addition, the extra compression error term imposes further challenges for constructing two novel Lyapunov functions in \eqref{eq:lya} to decouple the three recursions in Lemmas \ref{lem:cons0}-\ref{lem:opt0}.
\subsection{Preliminary Results}
\label{subsec:scvx_pre}
In this subsection, we first introduce the transformed error dynamics for CEDAS in Lemma \ref{lem:error_edas_ct} and then derive three recursions for the error terms $\mathbb{E}[\norm{\check{z}_k}^2]$, $\mathbb{E}[\norm{\mathbf{y}_k - H_k}^2]$, and $\mathbb{E}[\norm{\bar{z}_k}^2]$ in Lemmas \ref{lem:cons0}-\ref{lem:yh0_sim} ($\check{z}_k$ and $\bar{z}_k$ are defined in Lemma \ref{lem:error_edas_ct}). Generally speaking, Lemma \ref{lem:cons0} implies the expected consensus error $\mathbb{E}[\norm{\mathbf{x}_k - \mathbf{1}\bar{x}_k^{\intercal}}^2]$ is contractive (in itself), and Lemma \ref{lem:opt0} indicates the expected optimization error $\mathbb{E}[\norm{\bar{x}_k - x^*}^2]$ is contractive.
These two Lemmas are similar to those in \cite{huang2021improving} but with an additional term $\mathbb{E}[\norm{\mathbf{y}_k-H_k}^2]$, which is corresponding to the compression error. We show in Lemma \ref{lem:yh0_sim} that it is also contractive.
Lemma \ref{lem:error_edas_ct} introduces the transformed error dynamics of $\crk{\mathbf{x}_k, \mathbf{s}_k}$ in terms of the new iterates $\crk{\bar{z}_k, \check{z}_k}$.
\begin{lemma}
\label{lem:error_edas_ct}
Let Assumptions \ref{ass:W} and \ref{ass:fi} hold, the transformed error dynamics of \eqref{eq:edas_c} is given by
\begin{align*}
\bar{z}_{k + 1} &= \bar{z}_k -\frac{\eta_k}{n}\sum_{i=1}^n g_{i,k} ,\\
\check{z}_{k + 1} &= P_1\check{z}_k + \eta_kP_1 U_{L,l}\prt{\nabla F(\mathbf{x}^*) - \nabla F(\mathbf{x}^k)} + \eta_k P_1 U_{L,l}\hat\mathbf{g}_k\\
&\quad + \prt{P_1 - I}U_{L,l}E_k + U_{L,r}\prt{\mathbf{s}_k^* - \mathbf{s}_{k + 1}^*},
\end{align*}
where $\mathbf{s}_k^*$ is defined in Lemma \ref{lem:opt_cond}, $x^*$ is the optimal solution to Problem \eqref{eq:P}, the diagonal matrix $P_1\in\mathbb{R}^{2(n-1)\times 2(n-1)}$ and matrices $U_{R,u},U_{R,l}\in\mathbb{R}^{n\times 2(n-1)}$, $U_{L,l},U_{L,r}\in\mathbb{R}^{2(n-1)\times n}$ are defined in Lemma \ref{lem:Bk} in Appendix \ref{app:transformed_error}. In particular,
\begin{equation*}
\bar{z}_k^{\intercal} := \frac{\mathbf{1}^{\intercal}}{n}\hat{\mathbf{x}}_k, \ \check{z}_k:= U_{L,l}\hat{\mathbf{x}}_{k} + U_{L,r}\hat{\mathbf{s}}_k, \forall k,
\end{equation*}
with $\hat{\mathbf{x}}_k:= \mathbf{x}_k - \mathbf{1}(x^*)^{\intercal}, \hat{\mathbf{s}}_k:= \mathbf{s}_k - \mathbf{s}_k^*,\ \hat{\mathbf{g}}_k:= \nabla F(\mathbf{x}_k) - G_k$.
\end{lemma}
\begin{proof}
See Appendix \ref{app:error_edas_ct}.
\end{proof}
\begin{remark}
\label{rem:cons_cz}
We have from Lemma \ref{lem:ulur}
that $\mathbf{x}_k - \mathbf{1}\bar{x}_k^{\intercal} = U_{R,u}\check{z}_k,\forall k$. Hence, the consensus error $\norm{\mathbf{x}_k - \mathbf{1}\bar{x}_k^{\intercal}}^2$ is upper bounded by $(2\norm{\check{z}_k}^2)$.
\end{remark}
Lemma \ref{lem:cons0} below states the recursion for $\mathbb{E}\brk{\norm{\check{z}_{k + 1}}^2}$ which relates to the expected consensus error.
\begin{lemma}
\label{lem:cons0}
Let Assumptions \ref{ass:W}, \ref{ass:sgrad} and \ref{ass:ub_op} hold and set
\begin{align*}
\eta_k \leq \frac{\gamma(1-\lambda_2)}{8\sqrt{2}L},\
\gamma \leq \frac{1}{2}.
\end{align*}
We then have
\begin{align*}
\mathbb{E}\brk{\norm{\check{z}_{k + 1}}^2}&\leq \frac{3 + \tilde{\lambda}_2}{4} \mathbb{E}\brk{\norm{\check{z}_k}^2} + \frac{4\eta_k^2L^2 n}{1-\tilde{\lambda}_2}\mathbb{E}\brk{\norm{\bar{z}_k}^2} + \eta_k^2n\sigma^2\\
&\quad + \gamma C \mathbb{E}\brk{\norm{\mathbf{y}_k - H_k}^2} + \frac{4\norm{\nabla F(\mathbf{x}^*)}^2}{(1-\tilde{\lambda}_2)^2}\prt{\eta_k - \eta_{k + 1}}^2.
\end{align*}
\end{lemma}
\begin{proof}
See Appendix \ref{app:cons0}.
\end{proof}
\begin{remark}
If there is no compression, i.e., $C = 0$, then Lemma \ref{lem:cons0} reduces to Lemma 10 in \cite{huang2021improving}.
\end{remark}
Lemma \ref{lem:yh0_sim} states the recursion for the expected compression error $\mathbb{E}\brk{\norm{\mathbf{y}_k - H_k}^2}$.
\begin{lemma}
\label{lem:yh0_sim}
Suppose Assumptions \ref{ass:W} and \ref{ass:ub_op} hold and let
\begin{equation*}
\alpha\leq \frac{1}{12C},\ \gamma\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}}, \ \eta_k\leq \min\crk{\frac{2\sqrt{\alpha}}{3 \sqrt{6} L},\frac{\gamma(1-\lambda_2)}{8\sqrt{2}L}}.
\end{equation*}
We have
\begin{align*}
\mathbb{E}\brk{\norm{\mathbf{y}_{k + 1} - H_{k + 1}}^2} &\leq \prt{1 - \frac{\alpha}{3}}\mathbb{E}\brk{\norm{\mathbf{y}_k - H_k}^2} + \frac{9\gamma}{\alpha}\mathbb{E}\brk{\norm{\check{z}_k}^2}\\
&\quad + \frac{22\eta_k^2n L^2}{\alpha}\mathbb{E}\brk{\norm{\bar{z}_k}^2} + \frac{11\eta_k^2n\sigma^2}{\alpha} + \frac{18(\eta_k - \eta_{k + 1})^2}{\alpha}\norm{\nabla F(\mathbf{x}^*)}^2.
\end{align*}
\end{lemma}
\begin{proof}
See Appendix \ref{app:yh0}.
\end{proof}
Regarding $\mathbb{E}\brk{\norm{\bar{z}_k}^2}$, Lemma \ref{lem:error_edas_ct} shows that $\bar{x}_{k + 1} - x^* = \bar{x}_k - x^* - \frac{\eta_k}{n} \sum_{i=1}^n g_{i,k}$, which is the same as in \cite{huang2021improving} and \cite{pu2021sharp}. Therefore, we can apply those results directly to obtain Lemma \ref{lem:opt0} below.
\begin{lemma}
\label{lem:opt0}
Let Assumptions \ref{ass:W}, \ref{ass:sgrad}, and \ref{ass:fi} hold. Suppose $\eta_k\leq \min\crk{2/(\mu + L), 1/(3\mu)}$, then
\begin{align*}
\mathbb{E}\brk{\norm{\bar{z}_{k + 1}}^2} &\leq \prt{1 - \frac{3}{2}\eta_k\mu}\mathbb{E}\brk{\norm{\bar{z}_k}^2} + \frac{3\eta_k L^2}{n\mu}\mathbb{E}\brk{\norm{\check{z}_k}^2} + \frac{\eta_k^2\sigma^2}{n}.
\end{align*}
\end{lemma}
In the next subsection, we construct two Lyapunov functions to derive the explicit convergence results given specific stepsize $\crk{\eta_k}$.
\subsection{Convergence}
\label{subsec:sublinear}
In this subsection, we construct two Lyapunov functions $\mathbb{E}\mathcal{L}_k^a$ and $\mathbb{E}\mathcal{L}_k^b$ to decouple the recursions in Lemmas \ref{lem:cons0}-\ref{lem:opt0}. We will show in Lemma \ref{lem:lya} that $\mathbb{E}\mathcal{L}_k^a\sim\order{1/k}$ and $\mathbb{E}\mathcal{L}_k^b\sim\order{1/k^2}$ when using decreasing stepsizes $\eta_k$ specified in \eqref{eq:etak}:
\begin{equation}
\label{eq:etak}
\eta_k = \frac{\theta}{\mu(k + m)}, \text{ for some }\theta > 0.
\end{equation}
The two Lyapunov functions take the form stated below which guide us to bound $\mathbb{E}[\norm{\bar{z}_k}^2]$ and $\mathbb{E}[\norm{\check{z}_k}^2]$ respectively:
\begin{subequations}
\label{eq:lya}
\begin{align}
\mathcal{L}_k^a &:= \norm{\bar{z}_k}^2 + a_1\eta_k \norm{\check{z}_k}^2 + a_2\eta_k\norm{\mathbf{y}_k - H_k}^2\label{eq:lya1_cd},\\
\mathcal{L}_k^b &:= \norm{\check{z}_k}^2 + b\norm{\mathbf{y}_k - H_k}^2\label{eq:lya2_cd},
\end{align}
\end{subequations}
where
\begin{align*}
a_1:= \frac{36L^2}{n\mu(1-\tilde{\lambda}_2)}, \ a_2:= \frac{216\gamma CL^2}{\alpha n\mu(1-\tilde{\lambda}_2)}, b:= \frac{4\gamma C}{\alpha}.
\end{align*}
In what follows, we first derive an $\order{1/k}$ bound for $\mathbb{E}\mathcal{L}_k^a$ in Lemma \ref{lem:lya}, which serves as a loose bound on $\mathbb{E}[\norm{\bar{z}_k}^2]$.
Substituting the result into Lemmas \ref{lem:cons0} and \ref{lem:yh0_sim} helps us construct a bound for the Lyapunov function $\mathbb{E}\mathcal{L}_k^b$ which relates to $\mathbb{E}[\norm{\bar{z}_k}^2]$. Based on these results, we demonstrate the improved $\order{1/(nk)}$ convergence result for $\mathbb{E}[\norm{\bar{z}_k}^2]$ in Lemma \ref{lem:opt1}. The way to determine the proper coefficients $a_1,a_2$, and $b$ in \eqref{eq:lya} can be seen from the proof of Lemma \ref{lem:lya}.
\begin{lemma}
\label{lem:lya}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:fi} hold. Let the stepsize policy be \eqref{eq:etak} and
\begin{align*}
\alpha&\leq \frac{1}{12C},\ \gamma\leq\min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}, \frac{(1-\lambda_2)\alpha^2}{1296C}, \frac{\alpha}{1-\lambda_2}},\ \theta > 18,\\
m&= \max\crk{\frac{32\theta}{\gamma (1-\lambda_2)}, \frac{8\sqrt{2} L}{\mu\gamma(1-\lambda_2)}, \frac{3\theta L\sqrt{6}}{2\sqrt{\alpha}\mu}, \frac{8\theta}{\alpha}}.
\end{align*}
Denote the constants as
\begin{align}
\label{eq:cLs}
\hat{\mathcal{L}}^a_1:= \frac{12\theta^2\sigma^2}{\mu^2(4\theta - 3)},\ \hat{\mathcal{L}}^a_2:= \frac{40\theta^2\norm{\nabla F(\mathbf{x}^*)}^2}{n\mu^2(1-\tilde{\lambda}_2)^2(4\theta - 9)},\ \hat{\mathcal{L}}^a_3:= m^{\frac{4\theta}{3}}\mathbb{E}\mathcal{L}_0^a,\ \mathcal{B}_1:= \frac{6n L^2\theta^2}{\mu^2(1-\tilde{\lambda}_2)}.
\end{align}
Then
\begin{align*}
\mathbb{E}\mathcal{L}_k^a \leq \frac{\hat{\mathcal{L}}_1^a}{k + m} + \frac{\hat{\mathcal{L}}_2^a}{(k + m)^3} + \frac{\hat{\mathcal{L}}_3^a}{(k + m)^{\frac{4\theta}{3}}}.
\end{align*}
In addition,
\begin{align}
&\mathbb{E}\brk{\norm{\check{z}_k}^2}\leq \mathbb{E}\mathcal{L}_k^b\leq \prt{1 - \frac{1-\tilde{\lambda}_2}{6}}^k \mathbb{E}\mathcal{L}_0^b\nonumber\\
&\quad + \frac{12}{1-\tilde{\lambda}_2}\brk{\frac{2n\theta^2\sigma^2}{\mu^2(k + m)^2} + \frac{\mathcal{B}_1\hat{\mathcal{L}}_1^a}{(k + m)^3} + \frac{5\theta^2\norm{\nabla F(\mathbf{x}^*)}^2}{\mu^2(1-\tilde{\lambda}_2)^2(k + m)^4} + \frac{\mathcal{B}_1 \hat{\mathcal{L}}^a_2}{(k + m)^5} + \frac{\mathcal{B}_1\hat{\mathcal{L}}^a_3}{(k + m)^{\frac{4\theta}{3} + 2}}}.\label{eq:cons1}
\end{align}
\end{lemma}
\begin{proof}
See Appendix \ref{app:lem_lya}.
\end{proof}
In light of \eqref{eq:lya1_cd} and Lemma \ref{lem:lya}, we have $\mathbb{E}\brk{\norm{\bar{z}_k}^2}\sim \order{1/k}$. However, such a convergence rate does not characterize the linear speedup property of CEDAS. Lemma \ref{lem:opt1} below refines the convergence rate for $\mathbb{E}\brk{\norm{\bar{z}_{k}}^2}$ from $\order{1/k}$ to $\order{1/(nk)}$.
\begin{lemma}
\label{lem:opt1}
Let the conditions in Lemma \ref{lem:lya} hold. We have
\begin{align*}
& \mathbb{E}\brk{\norm{\bar{z}_{k}}^2} \leq \prt{\frac{m}{k + m}}^{\frac{3\theta}{2}}\mathbb{E}\brk{\norm{\bar{z}_0}^2} + \frac{72\theta L^2m^{\frac{3\theta}{2} - 1}\mathbb{E}\mathcal{L}_0^b}{n\mu^2(1-\tilde{\lambda}_2)(k + m)^{\frac{3\theta}{2}}} + \frac{6\theta^2\sigma^2}{(3\theta -2)n\mu^2(k + m)}\\
&\quad + \frac{216\theta^2 L^2}{\mu^2(3\theta-10)}\brk{\frac{\mathcal{B}_2}{(k + m)^2} + \frac{\mathcal{B}_3}{(k + m)^3} + \frac{\mathcal{B}_4}{(k + m)^4} + \frac{\mathcal{B}_5}{(k + m)^5} + \frac{\mathcal{B}_6}{(k + m)^{\frac{4\theta}{3} + 2}}},
\end{align*}
where
\begin{align*}
\mathcal{B}_2 &:= \frac{2\theta^2\sigma^2}{\mu^2(1-\tilde{\lambda}_2)}, \mathcal{B}_3:= \frac{6\theta^2 L^2\hat{\mathcal{L}}_1^a}{\mu^2(1-\tilde{\lambda}_2)^2}, \mathcal{B}_4:= \frac{5\theta^2\norm{\nabla F(\mathbf{x}^*)}^2}{n\mu^2(1-\tilde{\lambda}_2)^3}, \mathcal{B}_5:= \frac{6\theta^2 L^2\hat{\mathcal{L}}_2^a}{\mu^2(1-\tilde{\lambda}_2)^2},
\mathcal{B}_6:= \frac{12\theta^2L^2(3\theta -10)\hat{\mathcal{L}}_3^a}{\mu^2(1-\tilde{\lambda}_2)^2(\theta-12)}.
\end{align*}
\end{lemma}
\begin{proof}
See Appendix \ref{app:lem_opt1}.
\end{proof}
Note that the total expected error has the following decomposition \cite{huang2021improving,pu2021sharp}, and the inequality holds according to Remark \ref{rem:cons_cz}:
\begin{align}
\label{eq:decomp}
\frac{1}{n}\sum_{i=1}^n \mathbb{E}\brk{\norm{x_{i,k} - x^*}^2} = \mathbb{E}\brk{\norm{\bar{x}_k - x^*}^2} + \frac{1}{n}\sum_{i=1}^n\mathbb{E}\brk{\norm{x_{i,k} - \bar{x}_k}^2} \leq \mathbb{E}\brk{\norm{\bar{z}_k}^2} + \frac{2}{n}\mathbb{E}\brk{\norm{\check{z}_k}^2}.
\end{align}
We are now able to derive the convergence rate for the total expected error in Theorem \ref{thm:total1} below.
\begin{theorem}
\label{thm:total1}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:fi} hold. Let the stepsize policy be \eqref{eq:etak} and
\begin{align*}
\alpha&\leq \frac{1}{12C},\ \gamma\leq\min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}, \frac{(1-\lambda_2)\alpha^2}{1296C}, \frac{\alpha}{1-\lambda_2}},\ \theta > 18,\\
m&= \max\crk{\frac{32\theta}{\gamma (1-\lambda_2)}, \frac{8\sqrt{2} L}{\mu\gamma(1-\lambda_2)}, \frac{3\theta L\sqrt{6}}{2\sqrt{\alpha}\mu}, \frac{8\theta}{\alpha}}.
\end{align*}
We have for Algorithm \ref{alg:lead_ds} that
\begin{align*}
&\frac{1}{n}\sum_{i=1}^n \mathbb{E}\brk{\norm{x_{i,k} - x^*}^2} \leq \prt{1 - \frac{1-\tilde{\lambda}_2}{6}}^k \frac{2\mathbb{E}\mathcal{L}_0^b}{n} + \frac{6\theta^2\sigma^2}{(3\theta - 2)n\mu^2(k + m)} + \prt{\frac{m}{k + m}}^{\frac{3\theta}{2}}\mathbb{E}\brk{\norm{\bar{z}_0}^2}\\
&\quad + \frac{72\theta L^2m^{\frac{3\theta}{2} - 1}\mathbb{E}\mathcal{L}_0^b}{n\mu^2(1-\tilde{\lambda}_2)(k + m)^{\frac{3\theta}{2}}} + \frac{24L^2(9\theta^2 + 3\theta-10)}{\mu^2(3\theta-10)}\brk{\frac{\mathcal{B}_2}{(k + m)^2} + \frac{\mathcal{B}_3}{(k + m)^3} + \frac{\mathcal{B}_4}{(k + m)^4} + \frac{\mathcal{B}_5}{(k + m)^5}}\\
&\quad + \frac{\mathcal{B}_7}{(k + m)^{\frac{4\theta}{3} + 2}},
\end{align*}
where
\begin{align*}
\mathcal{B}_7:= \frac{144\theta^2 L^2\hat{\mathcal{L}}_3^a}{\mu^2 (1-\tilde{\lambda}_2)^2}\prt{\frac{18\theta^2 L^2}{\mu^2(\theta - 12)}+1}.
\end{align*}
\end{theorem}
\begin{proof}
Combining \eqref{eq:cons1} and Lemma \ref{lem:opt1} yields the desired results.
\end{proof}
\begin{remark}
Theorem \ref{thm:total1} gives the convergence result of CEDAS under smooth and strongly convex objective functions assuming the compression operator is unbiased. Compared to the $\order{1/[(1-\lambda_2)^2 k]}$ convergence rate in \cite{liu2021linear}, we are able to show the linear speedup property, i.e., $\order{1/(nk)}$ asymptotic convergence rate of CEDAS.
\end{remark}
According to Theorem \ref{thm:total1}, the convergence rate of CEDAS behaves as $\order{1/(nk)}$ when the number of iteration is sufficiently large. This result is comparable to the convergence rate of centralized SGD, i.e., $\order{1/(nk)}$. Then a natural question arises: \textit{how long does it take CEDAS to perform similarly as centralized SGD?} This question is answered by estimating the transient time of the algorithm in the next subsection.
\subsection{Transient Time}
\label{subsec:KT}
In this part, we estimate how long it takes CEDAS to achieve the convergence rate of centralized SGD method, i.e., the \emph{transient time} of CEDAS. We first give the formal definition of the transient time in \eqref{def:transient}.
The transient time (for smooth strongly convex objective functions) is formally defined as the following:
\begin{align}
K_T&:=\inf_K\left\{\frac{1}{n}\sum_{i=1}^n\mathbb{E}\left[\Vert x_{i,k}-x^*\Vert^2\right]\le \order{\frac{1}{nk}},\ \forall k\ge K. \right\}.\label{def:transient}
\end{align}
In other words, the transient time denotes the least time for the convergence rate of Algorithm \ref{alg:lead_ds} to be dominated by the convergence rate of centralized SGD. In the following, we hide the $\order{\cdot}$ notation for ease of presentation when referring to the transient time.
Theorem \ref{thm:KT} states the transient time of Algorithm \ref{alg:lead_ds}.
\begin{theorem}
\label{thm:KT}
Let the conditions in Theorem \ref{thm:total1} hold. Then it takes
\begin{align*}
K_T= \max&\crk{\frac{n}{\gamma(1-\lambda_2)}, \frac{1}{\alpha}, \frac{\norm{\nabla F(\mathbf{x}^*)}^{\frac{2}{3}}}{\gamma(1-\lambda_2)},\prt{\frac{n\mathbb{E}\mathcal{L}_0^a}{(1-\tilde{\lambda}_2)^{\frac{4\theta}{3} +2}}}^{\frac{1}{\frac{4\theta}{3} + 1}}, \frac{\max\crk{\log\prt{\mathbb{E}\mathcal{L}_0^b}, -\log(1-\tilde{\lambda}_2)}}{1-\tilde{\lambda}_2}}
\end{align*}
iterations for Algorithm \ref{alg:lead_ds} to reach the asymptotic network independent convergence rate, that is, when $k\geq K_T$, we have
\begin{align*}
\frac{1}{n}\sum_{i=1}^n \mathbb{E}\brk{\norm{x_{i,k} - x^*}^2} \leq \order{\frac{1}{nk}}.
\end{align*}
The constant $\mathbb{E}\mathcal{L}_0^a$ and $\mathbb{E}\mathcal{L}_0^b$ are in the order of
\begin{align*}
\mathbb{E}\mathcal{L}_0^a &= \order{\frac{(C+1)(\norm{\mathbf{x}_{-1} - \mathbf{x}^*}^2 + \norm{H_0 - \mathbf{x}^*}^2 + \norm{\nabla F(\mathbf{x}^*)}^2/m^2)}{n}},\\
\mathbb{E}\mathcal{L}_0^b &= \order{\frac{(C+1)(\norm{\mathbf{x}_{-1} - \mathbf{x}^*}^2 + \norm{H_0 - \mathbf{x}^*}^2 + \norm{\nabla F(\mathbf{x}^*)}^2 / m^2)}{\alpha}}.
\end{align*}
\end{theorem}
\begin{proof}
See Appendix \ref{app:thm_KT}.
\end{proof}
Under some mild additional conditions, we can obtain a cleaner expression for the transient times of CEDAS with unbiased and biased compression operators respectively in the following corollaries.
\begin{corollary}
\label{cor:KT_unbiased}
Let the conditions in Theorem \ref{thm:total1} hold and initiate $h_{i,0} = x_{i,-1}, \forall i\in[n]$. If we further choose $\alpha = 1/(12C)$ and assume for some $q>0$,
\begin{align}
\label{eq:graph_p}
\norm{\nabla F(\mathbf{x}^*)}^2 = \order{n^3},\ \norm{\mathbf{x}_{-1} - \mathbf{x}^*}^2 = \order{n^2},\ 1/(1-\lambda_2) = \order{n^q}.
\end{align}
Then
\begin{align}
\label{eq:KT_ub_p}
K_T = \frac{nC^3}{(1-\lambda_2)^2}.
\end{align}
\end{corollary}
\begin{remark}
$\norm{\nabla F(\mathbf{x}^*)}^2 = \order{n^3}$ holds for many problem settings, including linear regression and logistic regression. The restriction on the initial values $x_{i,-1}, i\in[n]$ is also mild. It aims to simplify the formula of the transient time. Indeed, we can initialize all the agents with the same solution to satisfy $\norm{\mathbf{x}_{-1} - \mathbf{x}^*}^2 = \order{n^2}$.
According to \cite{nedic2018network}, $1/(1-\lambda_2) = \order{n^2}$ is satisfied for any connected undirected graph if we use the Lazy Metropolis rule \eqref{eq:lazy} for constructing the mixing matrix $W$, i.e.,
\begin{equation}
\label{eq:lazy}
w_{ij} = \begin{cases}
\frac{1}{2\max\crk{\deg(i), \deg(j)}}, & i\in \mathcal{N}_i\\
1 - \sum_{j\in\mathcal{N}_i} w_{ij}, & i = j\\
0, & \text{ otherwise}
\end{cases}.
\end{equation}
\end{remark}
When the compression operator is biased (satisfying Assumption \ref{ass:b_op}), we can also derive the transient time for CEDAS using the technique in Lemma \ref{lem:b2ub}.
\begin{corollary}
\label{cor:KT_biased}
Let the conditions in Theorem \ref{thm:total1} hold except Assumption \ref{ass:ub_op} and initiate $h_{i,0} = x_{i,-1},\forall i\in[n]$. In addition, assume the compression operator satisfies Assumption \ref{ass:b_op}, and we apply the technique in Lemma \ref{lem:b2ub} with some unbiased compressor with parameter $C$. Under condition \eqref{eq:graph_p}, letting $\alpha = 1/(12C(1-\delta))$ yields the transient time of Algorithm \ref{alg:lead_ds}:
\begin{equation}
\label{eq:KT_b}
K_T = \frac{nC^3(1-\delta)^3}{(1-\lambda_2)^2}.
\end{equation}
\end{corollary}
We can also calculate the transient time for CEDAS when there is no compression.
\begin{corollary}
\label{cor:KT_noC}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, and \ref{ass:fi} hold. We further assume there is no compression, i.e., $C=0$. Let the stepsize policy be \eqref{eq:etak} and condition \eqref{eq:graph_p} hold. We set
\begin{align*}
\alpha&= 1,\ \gamma = \frac{1}{2},\ \theta > 18,\\
m&= \max\crk{\frac{64\theta}{1-\lambda_2}, \frac{16\sqrt{2} L}{\mu(1-\lambda_2)}, \frac{3\theta L\sqrt{6}}{2\mu}}.
\end{align*}
Then, the transient time of CEDAS reduces to
\begin{align*}
K_T = \frac{n}{1-\lambda_2}.
\end{align*}
\end{corollary}
\begin{remark}
When there is no compression error, the transient time of Algorithm \ref{alg:lead_ds} is similar to that of EDAS \cite{huang2021improving}.
If we further assume condition \eqref{eq:graph_p} holds, the transient time becomes $K_T = \order{n/(1-\lambda_2)}$.
This is consistent with our discussions in Section \ref{subsec:lead_ds}.
\end{remark}
\section{Convergence Analysis: Nonconvex Case}
\label{sec:ncvx}
In this section, we no longer assume the objective function $f$ is strongly convex but only consider Assumption \ref{ass:smooth}. Then, the error dynamics involving the optimal solution $x^*$ does not hold. However, Lemma \ref{lem:opt_cond} inspires us to introduce a term similar to $\hat{\mathbf{s}}_k$. Specifically, noting that $\tilde{V} \mathbf{s}_k^* + \eta\nabla F(\mathbf{x}^*) = \mathbf{0}$ and $\bar{x}_k\in\mathbb{R}^p$ should converge to a stationary point of Problem \eqref{eq:P}, we consider instead
\begin{align}
\label{eq:rk}
\hat{\r}_k := \tilde{V}\mathbf{s}_k + \eta\nabla F(\mathbf{1}\bar{x}_k^{\intercal}).
\end{align}
The new variable $\hat{\r}_k$ is a major difference compared to Section \ref{sec:scvx}.
In the nonconvex case, in addition to the challenges stated in Section \ref{sec:scvx}, the absence of (strong) convexity poses extra difficulties. One major obstacle is that $\tilde{V}$ has a zero eigenvalue which leads to $\tilde{V}\tV^{-} -I\neq 0$. Such an issue was previously handled by Lemma \ref{lem:opt_cond} in Section \ref{sec:scvx} due to the existence of the unique optimal solution $x^*$. In the nonconvex case, we consider a different decomposition in Lemma \ref{lem:hk} which resembles Lemma \ref{lem:error_edas_ct}. A similar idea has appeared in \cite{alghunaim2021unified}.
\subsection{Preliminaries}
\label{subsec:ncvx_pre}
In this subsection, we present the supporting lemmas corresponding to the expected consensus error and the expected compression error. The main procedures are similar to those in Subsection \ref{subsec:scvx_pre}, whereas the differences come from the nonexistence of an optimal solution.
In Lemma \ref{lem:hk}, we introduce the new transformed recursions corresponding to the consensus error. Lemma \ref{lem:descent0} constructs the ``approximate" descent property of $\mathbb{E} f(\bar{x}_k)$ and guides the further steps.
\begin{lemma}
\label{lem:hk}
Let Assumptions \ref{ass:W} and \ref{ass:smooth} hold, and denote $\check{h}_k:= \prt{J_{L,l}Q_1^{\intercal}\mathbf{x}_k + J_{L,r}(I-\tilde{\Lambda}_1)^{-1/2}Q_1^{\intercal}\hat{\r}_k}$. Then we have
\begin{align*}
\check{h}_{k + 1} &= P_1 \check{h}_k + P_1 J_{L,l}Q_1^{\intercal}\eta \prt{\nabla F(\mathbf{1}\bar{x}_k^{\intercal}) - \nabla F(\mathbf{x}_k)} + (J_{L,l}\tilde{\Lambda}_1 + J_{L,r}\sqrt{I-\tilde{\Lambda}_1})Q_1^{\intercal} \eta\hat{\mathbf{g}}_k\\
&\quad + \prt{J_{L,r} - J_{L,l}\sqrt{I-\tilde{\Lambda}_1}}\sqrt{I-\tilde{\Lambda}_1}Q_1^{\intercal}E_k + J_{L,r}(I-\tilde{\Lambda}_1)^{-\frac{1}{2}}\eta Q_1^{\intercal}\prt{\nabla F(\mathbf{1}\bar{x}_{k + 1}^{\intercal}) - \nabla F(\mathbf{1}\bar{x}_k^{\intercal})},
\end{align*}
where $Q_1$ is defined such that
\begin{align*}
\tilde{W} = \begin{pmatrix}
\mathbf{1} & Q_1
\end{pmatrix}
\begin{pmatrix}
1 & \\
& \tilde{\Lambda}_1
\end{pmatrix}
\begin{pmatrix}
\frac{1}{n}\mathbf{1}^{\intercal}\\
Q_1^{\intercal}
\end{pmatrix},
\end{align*}
with
\begin{align*}
\tilde{\Lambda}_1 := \mathrm{diag}\prt{\tilde{\lambda}_2, \tilde{\lambda}_3, \cdots, \tilde{\lambda}_n}\in\mathbb{R}^{(n-1)\times (n-1)}.
\end{align*}
The matrices $J_{L,l}, J_{L,r}\in\mathbb{R}^{2(n-1)\times n}$ are defined in \eqref{eq:js}, and $\norm{\mathbf{x}_k - \mathbf{1}\bar{x}_k}^2\leq 2\norm{\check{h}_k}^2$.
\end{lemma}
\begin{proof}
See Appendix \ref{app:lem_hk}.
\end{proof}
\begin{remark}
It is critical to introduce $\hat{\r}_k$. Suppose instead we consider $\mathbf{s}_k$ and define $\check{h}_k':= (U_{L,l}\mathbf{x}_k + U_{L,r}\mathbf{s}_k)$, then similar derivation to Lemma \ref{lem:error_edas_ct} would lead to
\begin{align*}
\check{h}_{k + 1}' = P_1\check{h}_k' + \eta P_1U_{L,l}\d_k + \eta P_1 U_{L,l}\nabla F(\mathbf{x}_k) + (P_1 - I)U_{L,l}E_k,
\end{align*}
which requires bounding $\eta^2\norm{\nabla F(\mathbf{x}_k)}^2$. This would make the convergence result suffer from data heterogeneity. By contrast, Lemma \ref{lem:hk} only requires considering $\eta^4\norm{\nabla F(\mathbf{x}_k)}^2$ when dealing with $(\nabla F(\mathbf{1}\bar{x}_{k + 1}^{\intercal}) - \nabla F(\mathbf{1}\bar{x}_k^{\intercal}))$. This argument will be made more clearly in the proof of Lemma \ref{lem:ncvx_cons0}.
\end{remark}
\begin{lemma}
\label{lem:descent0}
Let Assumptions \ref{ass:W}, \ref{ass:sgrad}, and \ref{ass:smooth} hold and suppose $\eta\leq 1/L$, we have
\begin{align*}
\mathbb{E} f(\bar{x}_{k + 1}) - f^{\text{inf}} &\leq \mathbb{E} f(\bar{x}_k) - f^{\text{inf}} - \frac{\eta}{2}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2} + \frac{\eta L^2}{n}\mathbb{E}\brk{\norm{\check{h}_k}^2} + \frac{L\sigma^2\eta^2}{2n}.
\end{align*}
\end{lemma}
\begin{proof}
See Appendix \ref{app:ncvx_lem_descent0}.
\end{proof}
In light of Lemma \ref{lem:hk}, we study the recursion of $\mathbb{E}[\norm{\check{h}_k}^2]$, which gives rise to Lemma \ref{lem:ncvx_cons0}. Then, we can utilize the ``approximate" descent property of $\mathbb{E} f(\bar{x}_k)$ in Lemma \ref{lem:descent0}.
\begin{lemma}
\label{lem:ncvx_cons0}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:smooth} hold, and let
\begin{equation*}
\gamma\leq \frac{1}{2}, \ \eta \leq \frac{\gamma(1-\lambda_2)}{16 L}.
\end{equation*}
We have
\begin{align*}
\mathbb{E}\brk{\norm{\check{h}_{k + 1}}^2} &\leq \frac{3 + \tilde{\lambda}_2}{4}\mathbb{E}\brk{\norm{\check{h}_k}^2} + 4\gamma C\mathbb{E}\brk{\norm{\mathbf{y}_k - H_k}^2} + 5n\eta^2\sigma^2 + \frac{8n\eta^4L^2}{(1-\tilde{\lambda}_2)^2}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2}.
\end{align*}
\end{lemma}
\begin{proof}
See Appendix \ref{app:ncvx_lem_cons0}.
\end{proof}
We next derive the recursion for $\mathbb{E}[\norm{\mathbf{y}_k - H_k}^2]$ in Lemma \ref{lem:ncvx_yh} below.
\begin{lemma}
\label{lem:ncvx_yh}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:smooth} hold, let
\begin{equation*}
\alpha\leq \frac{1}{12C},\ \gamma\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}},\ \eta\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{3}L},\frac{\gamma(1-\lambda_2)}{16L}},
\end{equation*}
we have
\begin{align*}
\mathbb{E}\brk{\norm{\mathbf{y}_{k + 1} - H_{k + 1}}^2} &\leq \prt{1-\frac{\alpha}{3}}\mathbb{E}\brk{\norm{\mathbf{y}_k - H_k}^2}+ \frac{10n\eta^2\sigma^2}{\alpha}+ \frac{18\gamma}{\alpha}\mathbb{E}\brk{\norm{\check{h}_k}^2} + \frac{6n\eta^2}{\alpha}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2}.
\end{align*}
\end{lemma}
\begin{proof}
See Appendix \ref{app:ncvx_lem_yh0}.
\end{proof}
Now we are ready to give the preliminary convergent result of CEDAS by constructing a Lyapunov function $\mathcal{L}_k^v$:
\begin{equation}
\label{eq:lya_ncvx}
\mathcal{L}_k^v:= f(\bar{x}_k) - f^{\text{inf}} + v_1\eta \norm{\check{h}_k}^2 + v_2\eta \norm{\mathbf{y}_k - H_k}^2, \ v_1:= \frac{8L^2}{n(1-\tilde{\lambda}_2)}, \ v_2 := \frac{96\gamma C L^2}{n\alpha(1-\tilde{\lambda}_2)}.
\end{equation}
\begin{lemma}
\label{lem:lya_ncvx}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:smooth} hold, and let
\begin{equation*}
\alpha\leq \frac{1}{12C},\ \gamma\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}, \frac{\alpha^2(1-\lambda_2)}{1728C}},\ \eta\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{3}L}, \frac{\gamma(1-\lambda_2)}{16L}}.
\end{equation*}
We have
\begin{align*}
\mathbb{E}\mathcal{L}_{k + 1}^v &\leq \mathbb{E}\mathcal{L}_k^v - \frac{\eta}{4}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2} + \frac{L\eta^2\sigma^2}{2n} + \frac{42L^2\eta^3\sigma^2}{1-\tilde{\lambda}_2}.
\end{align*}
\end{lemma}
\begin{proof}
See Appendix \ref{app:lya_ncvx}.
\end{proof}
\subsection{Convergence}
With the help of Lemma \ref{lem:lya_ncvx}, we can characterize the asymptotic network independent behavior of CEDAS for minimizing smooth nonconvex objective functions in Theorem \ref{thm:ncvx_ng}.
\begin{theorem}
\label{thm:ncvx_ng}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:smooth} hold. Let
\begin{equation*}
\alpha\leq \frac{1}{12C},\ \gamma\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}, \frac{\alpha^2(1-\lambda_2)}{1728C}},\ \eta\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{3}L}, \frac{\gamma(1-\lambda_2)}{16L}}.
\end{equation*}
We have
\begin{align*}
\frac{1}{K}\sum_{k=-1}^{K-1}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2} &\leq \frac{4\prt{f(\bar{x}_{-1}) - f^{\text{inf}}}}{\eta K} + \frac{32L^2\mathbb{E}\brk{\norm{\check{h}_0}^2}}{n(1-\tilde{\lambda}_2)K} + \frac{384\gamma C L^2\mathbb{E}\brk{\norm{\mathbf{y}_0 - H_0}^2}}{n\alpha (1-\tilde{\lambda}_2)K}\\
&\quad + \frac{4L\eta\sigma^2}{n} + \frac{168\eta^2L^2\sigma^2}{1-\tilde{\lambda}_2} + \frac{4L^2\norm{\mathbf{x}_{-1}}^2}{nK} + \frac{2L\eta \sigma^2}{nK}.
\end{align*}
\end{theorem}
\begin{proof}
We first consider the term $\mathbb{E} f(\bar{x}_0) - f^{\text{inf}}$. Similar to the proof of Lemma \ref{lem:descent0}, let $\eta \leq 1/L$. We have
\begin{align}
\label{eq:f0}
&\mathbb{E} f(\bar{x}_0) - f^{\text{inf}} \leq f(\bar{x}_{-1}) - f^{\text{inf}} - \frac{\eta}{2}\norm{\nabla f(\bar{x}_{-1})} + \frac{\eta L^2 \norm{\mathbf{x}_{-1} - \mathbf{1}\bar{x}_{-1}^{\intercal}}^2}{n} + \frac{L\eta^2 \sigma^2}{2n}.
\end{align}
Taking the average among $k=0,1,\cdots, K-1$ on both sides of the inequality in Lemma \ref{lem:lya_ncvx} yields
\begin{align*}
&\frac{1}{K}\sum_{k=0}^{K-1}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2} \leq \frac{4\prt{\mathbb{E}\mathcal{L}_{0}^v - \mathbb{E}\mathcal{L}_K^0}}{\eta K} + \frac{2L\eta\sigma^2}{n} + \frac{168\eta^2L^2\sigma^2}{1-\tilde{\lambda}_2}\\
&\leq \frac{4\prt{\mathbb{E} f(\bar{x}_0) - f^{\text{inf}}}}{\eta K} + \frac{32L^2\mathbb{E}\brk{\norm{\check{h}_0}^2}}{n(1-\tilde{\lambda}_2)K} + \frac{384\gamma C L^2\mathbb{E}\brk{\norm{\mathbf{y}_0 - H_0}^2}}{n\alpha (1-\tilde{\lambda}_2)K}+ \frac{2L\eta\sigma^2}{n} + \frac{168\eta^2L^2\sigma^2}{1-\tilde{\lambda}_2}.
\end{align*}
Combining the above relations yields the desired result.
\end{proof}
By choosing a specific stepsize $\eta$ and setting the total iteration number $K$ to be large enough, we obtain the convergence rate of CEDAS stated in Corollary \ref{cor:ncvx_rate} which behaves as $\order{1/\sqrt{nK}}$. Such a result is comparable to the centralized SGD method.
\begin{corollary}
\label{cor:ncvx_rate}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:smooth} hold. Let
\begin{equation*}
\alpha\leq \frac{1}{12C},\ \gamma\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}, \frac{\alpha^2(1-\lambda_2)}{1728C}},\ \eta = \sqrt{\frac{n}{K}},\ K\geq \max\crk{\frac{12L^2 n}{\alpha}, \frac{256nL^2}{\gamma^2 (1-\lambda_2)^2}}.
\end{equation*}
We have
\begin{align*}
\frac{1}{K}\sum_{k=-1}^{K-1}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2} &= \order{\frac{1}{\sqrt{nK}} + \frac{n}{(1-\tilde{\lambda}_2)K}+ \frac{\norm{\mathbf{x}_{-1}}^2 + \norm{H_0}^2}{n(1-\tilde{\lambda}_2)K} + \frac{\norm{\nabla F(\mathbf{x}_{-1})}^2}{(1-\tilde{\lambda}_2)^2 K^2}}.
\end{align*}
\end{corollary}
\begin{proof}
Note that for the terms $\mathbb{E}\brk{\norm{\check{h}_0}^2}$ and $\mathbb{E}\brk{\norm{\mathbf{y}_0 - H_0}^2}$, we have
\begin{align*}
\mathbb{E}\brk{\norm{\check{h}_0}^2} & = \mathbb{E}\brk{\norm{J_{L,l}Q_1^{\intercal}\mathbf{x}_0 + J_{L,r}(I-\tilde{\Lambda}_1)^{-\frac{1}{2}}\eta\nabla F(\mathbf{1}\bar{x}_0^{\intercal})}^2}\\
&= \order{\norm{\mathbf{x}_{-1}}^2 + \frac{\eta^2}{1-\tilde{\lambda}_2}\norm{\nabla F(\mathbf{x}_{-1})}^2},
\end{align*}
and
\begin{align*}
\mathbb{E}\brk{\norm{\mathbf{y}_0 - H_0}^2} &= \norm{\mathbf{x}_{-1} - \eta G_{-1} - \eta G_0 - H_0}^2\\
&= \order{\norm{\mathbf{x}_{-1}}^2 + \eta^2\norm{\nabla F(\mathbf{x}_{-1})}^2 + \norm{H_0}^2}.
\end{align*}
Substituting the above results into Theorem \ref{thm:ncvx_ng} and noting the choice of $K$ yield the desired result.
\end{proof}
\subsection{Transient Time}
This part introduces the transient time of CEDAS without (strong) convexity, where the formal definition of transient time is as the following:
\begin{align}
K_T^{(ncvx)}&:= \inf_K\left\{\frac{1}{K}\sum_{t=0}^{K - 1}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_t)}^2}\leq \order{\frac{1}{\sqrt{nK}}},\ \forall k\ge K. \right\}\label{def:transient_ncvx}
\end{align}
We state the transient time of CEDAS in Theorem \ref{thm:ncvx_KT} below.
\begin{theorem}
\label{thm:ncvx_KT}
Suppose Assumptions \ref{ass:W}, \ref{ass:sgrad}, \ref{ass:ub_op}, and \ref{ass:smooth} hold. Let
\begin{equation*}
\alpha\leq \frac{1}{12C},\ \gamma\leq \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}, \frac{\alpha^2(1-\lambda_2)}{1728C}},\ \eta = \sqrt{\frac{n}{K}},\ K\geq \max\crk{\frac{12L^2 n}{\alpha}, \frac{256nL^2}{\gamma^2 (1-\lambda_2)^2}}.
\end{equation*}
It takes
\begin{align*}
K_T^{(ncvx)} = \max&\crk{\frac{n^3}{\gamma^2 (1-\lambda_2)^2}, \frac{\norm{\mathbf{x}_{-1}}^{4}}{n\gamma^2 (1-\lambda_2)^2}, \frac{\norm{H_0}^4}{n\gamma^2 (1-\lambda_2)^2}, \prt{\frac{n \norm{\nabla F(\mathbf{x}_{-1})}^4}{\gamma^4(1-\lambda_2)^4}}^{\frac{1}{3}}}
\end{align*}
for Algorithm \ref{alg:lead_ds} to achieve the asymptotic network independent convergence rate.
\end{theorem}
\begin{proof}
We have from Corollary \ref{cor:ncvx_rate} that
\begin{align*}
\frac{1}{K}\sum_{k=-1}^{K-1}\mathbb{E}\brk{\norm{\nabla f(\bar{x}_k)}^2} &=
\order{1 + \frac{n^{\frac{3}{2}}}{(1-\tilde{\lambda}_2)\sqrt{K}}+ \frac{\norm{\mathbf{x}_{-1}}^2 + \norm{H_0}^2}{\sqrt{nK}(1-\tilde{\lambda}_2)} + \frac{\sqrt{n}\norm{\nabla F(\mathbf{x}_{-1})}^2}{(1-\tilde{\lambda}_2)^2 K^\frac{3}{2}}}\frac{1}{\sqrt{nK}}.
\end{align*}
According to the definition of $K_T^{(ncvx)}$ in \eqref{def:transient_ncvx}, we obtain the desired result.
\end{proof}
We next introduce a mild additional condition \eqref{eq:ncvx_initial} similar to \eqref{eq:graph_p} to simplify the transient time in Theorem \ref{thm:ncvx_KT}.
\begin{corollary}
\label{cor:ncvx_ub}
Let the conditions in Theorem \ref{thm:ncvx_KT} hold and further assume that for some $q>0$,
\begin{align}
\label{eq:ncvx_initial}
\norm{\mathbf{x}_{-1}}^2 = \norm{H_0} = \order{n^2},\ \norm{\nabla F(\mathbf{x}_{-1})}^2 = \order{n^2},\ 1/(1-\lambda_2) = \order{n^q}.
\end{align}
Choose
\begin{equation*}
\alpha= \frac{1}{12C},\ \gamma= \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C}},\frac{1}{2}, \frac{\alpha^2(1-\lambda_2)}{1728C}},\ \eta = \sqrt{\frac{n}{K}},\ K\geq \max\crk{\frac{12L^2 n}{\alpha}, \frac{256nL^2}{\gamma^2 (1-\lambda_2)^2}}.
\end{equation*}
Then the transient time of Algorithm \ref{alg:lead_ds} becomes
\begin{align*}
K_T^{(ncvx)} = \frac{n^3C^6}{(1-\lambda_2)^4}.
\end{align*}
\end{corollary}
\begin{remark}
Condition \eqref{eq:ncvx_initial} is mild. Indeed, we can initialize $x_{i,-1} = h_{i,0}=x_{-1}$ for any $i\in[n]$, then $\norm{\mathbf{x}_{-1}}^2 = \norm{H_0} = n\norm{x_{-1}}^2$ and $\norm{\nabla F(\mathbf{x}_{-1})}^2 = \sum_{i=1}^n\norm{\nabla f_i(x_{-1})}^2$. Such an initialization satisfies condition \eqref{eq:ncvx_initial}.
\end{remark}
Similar to Corollary \ref{cor:KT_biased}, we can derive the transient time for CEDAS when the compression operator is biased.
\begin{corollary}
\label{cor:ncvx_b}
Let the conditions in Theorem \ref{thm:ncvx_KT} hold. In particular, assume the compression operator satisfies Assumption \ref{ass:b_op}. If we further assume condition \eqref{eq:ncvx_initial} holds and apply the technique in Lemma \ref{lem:b2ub} with some unbiased compressor with parameter $C$, let
\begin{align*}
&\alpha = \frac{1}{12C(1-\delta)},\ \gamma= \min\crk{\frac{\sqrt{\alpha}}{2\sqrt{C(1-\delta)}},\frac{1}{2}, \frac{\alpha^2(1-\lambda_2)}{1728C(1-\delta)}},\\
&\eta= \sqrt{\frac{n}{K}}, K\geq \max\crk{\frac{12L^2 n}{\alpha}, \frac{256nL^2}{\gamma^2 (1-\lambda_2)^2}}.
\end{align*}
Then the transient time of Algorithm \ref{alg:lead_ds} becomes
\begin{align*}
K_T^{(ncvx)} = \frac{n^3C^6(1-\delta)}{(1-\lambda_2)^4}.
\end{align*}
\end{corollary}
Finally, we can obtain the transient time for CEDAS when there is no compression.
\begin{corollary}
\label{cor:ncvx_KT_noC}
Let the conditions in Theorem \ref{thm:ncvx_KT} hold and further assume condition \eqref{eq:ncvx_initial}. In addition, suppose there is no compression, i.e., $C = 0$. Choose
\begin{align*}
&\alpha = 1,\ \gamma= \frac{1}{2},\ \eta= \sqrt{\frac{n}{K}}, K\geq \frac{512nL^2}{(1-\lambda_2)^2}.
\end{align*}
Then the transient time of Algorithm \ref{alg:lead_ds} becomes
\begin{align*}
K_T^{(ncvx)} = \frac{n^3}{(1-\lambda_2)^2}.
\end{align*}
\end{corollary}
\begin{remark}
Such a result is consistent to those given in \cite{alghunaim2021unified} when the objective functions are smooth and nonconvex.
\end{remark}
\section{Numerical Experiments}
\label{sec:exp}
In this section, we present the numerical results regarding logistic regression and neural network training. For the compression schemes, we consider Tok-$K$ and Random-$K$ as the biased compressor $\mathcal{C}_1$ and $\mathcal{C}_2$ respectively and scaled Random-$K$ as the unbiased compressor $\mathcal{C}_3$.
We set $K = \lfloor 5\% \cdot p\rfloor$ by default for compressors $\mathcal{C}_1$, $\mathcal{C}_2$, and $\mathcal{C}_3$, where $p$ is the dimension. We also consider the unbiased $b-$bit quantization $\mathcal{C}_4$ in \cite{liu2021linear}:
\begin{align*}
\mathcal{C}_4(x) =\prt{\norm{x}_{\infty}2^{-(b-1)}\mathrm{sign}(x)}\cdot \left\lfloor\frac{2^{(b-1)}|x|}{\norm{x}_{\infty}} + \mu\right\rfloor,
\end{align*}
where $\cdot$ is the Hadamard product. Both $\mathrm{sign}(\cdot)$ and $|\cdot|$ operate element-wisely. The vector $\mu$ is random and uniformly distributed in $[0,1]^p$. In the following, we choose $b$ such that the bits sent are $5\%$ of the uncompressed schemes.
The network topologies we consider can be found in figure \ref{fig:network}. Each node in the exponential network is connected to its $2^0, 2^1, 2^2, \cdots$ hops neighbors. The mixing matrices compliant with these two networks are constructed under the Lazy Metropolis rule \cite{nedic2018network}.
\begin{figure}[htbp]
\centering
\subfloat[Exponential network, $n = 16$.]{\includegraphics[width = 0.3\textwidth]{figs/exp16.pdf}}
\hspace{0.2\textwidth}
\subfloat[Grid network, $n = 16$.]{\includegraphics[width = 0.3\textwidth]{figs/grid16.pdf}}
\caption{Illustration of two network topologies.}
\label{fig:network}
\end{figure}
\subsection{Logistic Regression}
\label{subsec:logistic}
We consider a binary classification problem using logistic regression \eqref{eq:logistic}. Each agent possesses a distinct local dataset $\mathcal{S}_i$ selected from the whole dataset $\mathcal{S}$. The classifier can then be obtained by solving the following convex optimization problem using all the agents' local datasets $\mathcal{S}_i, i=1,2,...,n$:
\begin{subequations}
\label{eq:logistic}
\begin{align}
&\min_{x\in\mathbb{R}^{785}} f(x) = \frac{1}{n}\sum_{i=1}^n f_i(x),\\
&f_i(x) := \frac{1}{|\mathcal{S}_i|} \sum_{j\in\mathcal{S}_i} \log\left[1 + \exp(-x^{\intercal}u_jv_j)\right] + \frac{\rho}{2}\norm{x}^2.
\end{align}
\end{subequations}
For Problem \eqref{eq:logistic}, we first consider the MNIST dataset \cite{mnist}. We set $\rho = 1/5$ and use decreasing stepsize $5 /(k + 100)$ for all the algorithms. The parameters of Algorithm \ref{alg:lead_ds} are chosen as $\alpha = 0.1$ and $\gamma = 0.004$. As shown in Table \ref{tab:kt}, it is sufficient to compare CEDAS with Choco-SGD, which previously enjoys the shortest transient time. The parameter of Choco-SGD is also set as $\gamma = 0.004$. The performance of those uncompressed decentralized methods is presented as the baseline.
We first illustrate the performance of different algorithms via the residual error $\sum_{i=1}^n\mathbb{E}\brk{\norm{x_{i,k} - x^*}^2} / n$ against the number of iterations in Figure \ref{fig:iter_logistic}. It can be seen that, regarding the asymptotic network independent property of CEDAS, the results are consistent with our theoretical finding, that is, it takes more iterations for CEDAS to achieve comparable performance with centralized SGD when the connectivity of the network becomes worse (from an exponential network to a grid network). Moreover, the performance of CEDAS is better than that of Choco-SGD for both graphs (under different compressors). The superiority of CEDAS is more evident in a grid network.
\begin{figure}[htbp]
\centering
\subfloat[Grid network, $n = 100$, $1-\lambda_2 = 0.013$]{\includegraphics[width = 0.5\textwidth]{figs/grid100nt10introduced_iter.pdf}}
\subfloat[Exponential network, $n = 100$, $1-\lambda_2 = 0.133$.]{\includegraphics[width = 0.5\textwidth]{figs/exponential100nt10introduced_iter.pdf}}
\caption{Residual against the number of iterations. The results are averaged over $10$ repeated runs.}
\label{fig:iter_logistic}
\end{figure}
Then we check the variation of the residual error while fixing the total transmitted bits of each node in Figure \ref{fig:bits_logistic}.
As shown in Figure \ref{fig:bits_logistic}, the compressed decentralized methods achieve better accuracy compared to their uncompressed counterparts under the same trasmitted bits. In particular, CEDAS achieves better performance compared to Choco-SGD. Such a difference becomes more evident when the graph connectivity is worse (in a grid network).
\begin{figure}[htbp]
\centering
\subfloat[Grid network, $n = 100$, $1-\lambda_2 = 0.013$]{\includegraphics[width = 0.5\textwidth]{figs/grid100bits.pdf}}
\subfloat[Exponential network, $n = 100$, $1-\lambda_2 = 0.133$.]{\includegraphics[width = 0.5\textwidth]{figs/exponential100bits.pdf}}
\caption{Residual against the communicated bits. The results are averaged over $5$ repeated runs.}
\label{fig:bits_logistic}
\end{figure}
\subsection{Neural Network}
For the nonconvex case, we consider training a neural network with one hidden layer of $64$ neurons for a $10$-class classification problem on the MNIST dataset. We use constant stepsize $\eta = 0.1$ and set the parameter $\alpha=0.1$ and $\gamma = 0.004$. The dimension of the problem is $p=51675$. Figure \ref{fig:bits_nn} illustrates the performance of Choco-SGD and CEDAS with different compressors and their uncompressed counterparts DSGD and EDAS. When fixing the total transmitted bits, the compressed decentralized methods are preferable compared to their uncompressed counterparts. Particularly, the proposed CEDAS algorithm enjoys the best performance as illustrated in Figure \ref{fig:bits_nn}.
\begin{figure}[htbp]
\centering
\subfloat[Grid network, $n = 25$, $1-\lambda_2 = 0.054$]{\includegraphics[width = 0.5\textwidth]{figs/nn_grid25_bits.pdf}}
\subfloat[Exponential network, $n = 25$, $1-\lambda_2 = 0.305$.]{\includegraphics[width = 0.5\textwidth]{figs/nn_exponential25_bits.pdf}}
\caption{Loss against communicated bits. The results are averaged over $2$ repeated runs.}
\label{fig:bits_nn}
\end{figure}
\newpage
|
\section{Introduction}
A study of $\mathbb{N}$-graded vertex algebras $V=\oplus_{n=0}^{\infty}V_n$ when $\dim V_0\geq 2$ is far from being completed. For this case, $V_0$ is a unital commutative associative algebra, and $V_1$ is a Leibniz algebra. The skew symmetry and Jacobi identity of vertex algebra give rise to compatible extra relations. Indeed, these additional relations on $V_0\oplus V_1$ are summarized in the notion of a vertex $V_0$-algebroid $V_1$ (see \cite{GMS, MS1, MS2, MSV}). It was shown in \cite{GMS} that for a given vertex $A$-algebroid $B$, one can construct an $\mathbb{N}$-graded vertex algebra $V=\oplus_{n=0}^{\infty}V_n$ such that $V_0=A$ and $V_1=B$. Also, the classification of graded simple twisted and non-twisted modules of vertex algebras associated with vertex algebroids were studied in \cite{LiY, LiY2} by Li and the last author of this paper.
In \cite{JY}, among many things, Jitjankarn and the last author of this paper show that for an $\mathbb{N}$-graded vertex algebra $V=\oplus_{n=0}^{\infty}V_n$ such that $2\leq \dim V_0<\infty$ and $1<\dim V_1<\infty$, if $V_0$ is a local algebra then $V$ is an indecomposable vertex algebra. Also, if $V$ is generated by $V_0$ and $V_1$, and $V_0$ is a local algebra and is not a simple module for a Lie $V_0$-algebroid $V_1/(V_0)_{-1}\partial(V_0)$, then $V$ is an indecomposable non-simple vertex algebra. In \cite{JY2}, Jitjankarn and the last author of this paper examined an algebraic structure of an indecomposable non-simple $\mathbb{N}$-graded vertex algebra associated with a vertex $A$-algebroid $B$ such that $B$ is (semi-)simple Leibniz algebra that has $sl_2$ as its Levi factor. The one-to-one correspondence between the set of representatives of equivalence classes of simple $sl_2$-modules and the set of representatives of equivalence classes of the $\mathbb{N}$-graded simple $V_B$-modules was established. In addition, Jitjankarn and the last of author of this paper showed that a certain quotient space of $V_B$ is an indecomposable non-simple vertex algebra that satisfies the $C_2$-condition and has only two irreducible modules. These two irreducible modules are in fact isomorphic to irreducible modules of the rational $C_2$-cofinite $CFT$-type vertex operator algebra $V_{\mathbb{Z}\alpha}$ associated with a rank one lattice $\mathbb{Z}\alpha$ equipped with a bilinear form $(~,~)$ such that $(\alpha,\alpha)=2$. In \cite{BuY}, Bui and the last author of this paper generalized the results in \cite{JY2} to the case when a vertex $A$-algebroid $B$ is an arbitrary finite-dimensional simple Leibniz algebras. Indecomposable non-simple $C_2$-cofinite $\mathbb{N}$-graded vertex algebras $U_B$ were established. A one-to-one correspondence between these irreducible $U_B$-modules and irreducible modules of certain types of rational $C_2$-cofinite $CFT$-type affine vertex algebras was exhibited as well. Note that the results in \cite{BuY, JY2} provided new families of irrational $C_2$-cofinite vertex algebras.
The work in this paper is a stepping stone of our attempt to understand an algebraic structure $\mathbb{N}$-graded vertex algebras $V_B$ associated to vertex $A$-algebroids $B$ when $B$ are no longer (semi) simple Leibniz algebras and their connections to many well known $\mathbb{N}$-graded vertex algebras $V=\oplus_{n=0}^{\infty}V_n$ such that $\dim V_0=1$. We are also interested in roles of the commutative associative algebras $A$ that play on a study of representation theory of $\mathbb{N}$-graded vertex algebras.
In this paper, we construct indecomposable non-simple $\mathbb{N}$-graded vertex algebras $V_B$ associated to vertex $A$-algebroids $B$ when $B$ are non-Lie cyclic left Leibniz algebra. In addition we investigate relationships between the constructed indecomposable non-simple vertex algebras $V_B$ and a rank one Heisenberg vertex operator algebra. As mentioned above, to study indecomposable non-simple properties of $\mathbb{N}$-graded vertex algebras $V=\oplus_{n=0}^{\infty}V_n$ that are generated by $V_0$ and $V_1$, we only need to study the algebraic structure of vertex $V_0$-algebroid $V_1$. Therefore, it is natural to begin our study by classifying and investigating an algebraic structure of vertex $A$- algebroids $B$ associated to non-Lie cyclic Leibniz algebras $B$. This classification of vertex $A$-algebroids $B$ help us identify connections between the indecomposable non-simple vertex algebra $V_B$ and a rank one Heisenberg vertex operator algebra. In fact, commutative associative algebras $A$ are the main tools for pining down possible relations between the indecomposable non-simple vertex algebras $V_B$ and a rank one Heisenberg vertex operator algebra. We show that appropriate quotient spaces of a certain type of indecomposable, non-simple vertex algebras associated to vertex $A$-alebroids $B$ are in fact a rank one Heisenberg vertex operator algebras.
In Section 2, we provide background on cyclic left Leibniz algebras and vertex algebroids. In Section 3, we classify vertex algebroids associated to cyclic left Leibniz algebras. Precisely, in subsections 3.1 and 3.2, we classify vertex algebroids associate to 2-dimensional non-Lie cyclic left Leibniz algebras and 3-dimensional non-Lie cyclic left Leibniz algebra, respectively. A review for a construction of vertex algebras associated to vertex algebroids and their modules is provided in subsection 4.1. Discussions about representation theory of vertex algebras associated to vertex $A$-algebroids and their modules when $B$ are cyclic non-Lie left Liebniz algebras are in subsection 4.2. Exploration of relationships between these vertex algebras and the vertex operator algebra associated with a rank one Heisenberg algebra is in subsection 4.3.
\section{A review on cyclic Leibniz algebras and vertex algebroids}
We begin by providing background on non-Lie cyclic Leibniz algebras that we will use in this paper. After that we recall definitions of a 1-truncated conformal algebra, a vertex algebroid, a Lie algebroid, and discuss about modules of Lie algebroids.
\begin{dfn}\cite{DMS, FM} A {\em left Leibniz algebra} $\mathfrak{L}$ is a $\mathbb{C}$-vector space equipped with a bilinear map $[~,~]:\mathfrak{L}\times\mathfrak{L}\rightarrow\mathfrak{L}$ satisfying the Leibniz identity $$[a,[b,c]]=[[a,b],c]+[b,[a,c]]$$ for all $a,b,c\in\mathfrak{L}$.
Let $I$ be a subspace of $\mathfrak{L}$. $I$ is a {\em left} (respectively, {\em right}) {\em ideal} of $\mathfrak{L}$ if $[\mathfrak{L}, I]\subseteq I$ (respectively, $[I,\mathfrak{L}]\subseteq I$). $I$ is an {\em ideal} of $\mathfrak{L}$ if it is both a left and a right ideal.
\end{dfn}
Let $\mathfrak{L}$ be a left Leibniz algebra and let $u\in \mathfrak{L}$. Now, we fix the notation $u^1=u$, $u^2=[u,u]$, and in general, $u^{n+1}=[u,u^n]$ for $n\geq 1$. Clearly, $[[u,u],y]=0$, and more generally $[u^n,y]=0$ for all $y\in \mathfrak{L}$, $n\geq 2$.
\begin{dfn} Let $\mathfrak{L}$ be a Leibniz algebra. $\mathfrak{L}$ is cyclic if and only if there exists some $u\in\mathfrak{L}$ such that $\mathfrak{L}=\langle u\rangle=Span\{u^k~|~k=1,2,....\}$. If $\mathfrak{L}=\langle v\rangle$, we call $v$ a generator of $\mathfrak{L}$.
\end{dfn}
\begin{prop}\label{cyclicclass}\cite{BuHLSS, DMS}
\ \
\begin{enumerate}
\item For a non-Lie left Leibniz algebra $B$ such that $\dim B=2$, $B$ is isomorphic to a cyclic left Leibniz algebra generated by $b$ with either $[b,b^2]=0$ or $[b,b^2]=b^2$.
\item A complete list of cyclic non-Lie left Leibniz algebras $B$ such that $\dim B=3$ with non-zero bracket $[~,~]$ is the following:
\begin{enumerate}
\item $[x,x]=y$, $[x,y]=z$.
\item $[z,y]=y$; $[z,x]=\alpha x$, $\alpha\in\mathbb{C}\backslash\{0\}$ and $\alpha\neq 1$.
\item $[z,x]=x+y$; $[z,y]=y$.
\item $[z,x]=y$; $[z,y]=y$; $[z,z]=x$.
\end{enumerate}
\end{enumerate}
\end{prop}
\begin{cor}\label{cyclicclass3}\cite{BuHLSS}
Let $B$ be a non-Lie cyclic Leibniz algebra $B$ such that $\dim B=3$.
\begin{enumerate}
\item If $B$ is of type $(b)$ in Proposition \ref{cyclicclass}, then $B$ has a basis $\{b,b^2,b^3\}$ such that $b^4=b^2-\frac{\alpha+1}{\sqrt{\alpha}}ib^3$ ($\alpha\neq 1$).
\item If $B$ is of type $(c)$ in Proposition \ref{cyclicclass}, then $B$ has a basis $\{b,b^2,b^3\}$ such that $b^4=b^2+2ib^3$.
\item If $B$ is of type $(d)$ in Proposition \ref{cyclicclass}, then $B$ has a basis $\{b,b^2,b^3\}$ such that $b^4=b^3$.
\end{enumerate}
\end{cor}
\begin{dfn}\cite{GMS} A {\em 1-truncated conformal algebra} is a graded vector space $C=C_0\oplus C_1$ equipped with a linear map $\partial:C_0\rightarrow C_1$ and bilinear operations $(u,v)\mapsto u_iv$ for $i=0,1$ of degree $-i-1$ on $C=C_0\oplus C_1$ such that the following axioms hold:
\medskip
\noindent(Derivation) for $a\in C_0$, $u\in C_1$,
\begin{equation}
(\partial a)_0=0,\ \ (\partial a)_1=-a_0,\ \ \partial(u_0a)=u_0\partial a;
\end{equation}
\noindent(Commutativity) for $a\in C_0$, $u,v\in C_1$,
\begin{equation}
u_0a=-a_0u,\ \ u_0v=-v_0u+\partial(u_1v),\ \ u_1v=v_1u;
\end{equation}
\noindent(Associativity) for $\alpha,\beta,\gamma\in C_0\oplus C_1$,
\begin{equation}
\alpha_0\beta_i\gamma=\beta_i\alpha_0\gamma+(\alpha_0\beta)_i\gamma.
\end{equation}
\end{dfn}
\begin{dfn}\cite{Br1, Br2, GMS} Let $(A,*)$ be a unital commutative associative algebra over $\mathbb{C}$ with the identity $1_A$. A {\em vertex $A$-algebroid} is a $\mathbb{C}$-vector space $\Gamma$ equipped with
\begin{enumerate}
\item a $\mathbb{C}$-bilinear map $A\times \Gamma\rightarrow \Gamma, \ \ (a,v)\mapsto a\cdot v$ such that $1\cdot v=v$ (i.e. a nonassociative unital $A$-module),
\item a structure of a Leibniz $\mathbb{C}$-algebra $[~,~]:\Gamma\times \Gamma\rightarrow\Gamma$,
\item a homomorphism of left Leibniz $\mathbb{C}$-algebra $\pi:\Gamma\rightarrow Der(A)$,
\item a symmetric $\mathbb{C}$-bilinear pairing $\langle ~,~\rangle:\Gamma\otimes_{\mathbb{C}}\Gamma\rightarrow A$,
\item a $\mathbb{C}$-linear map $\partial :A\rightarrow \Gamma$ such that $\pi\circ \partial =0$ which satisfying the following conditions:
\begin{eqnarray*}
&&a\cdot (a'\cdot v)-(a*a')\cdot v=\pi(v)(a)\cdot \partial(a')+\pi(v)(a')\cdot \partial(a),\\
&&[u,a\cdot v]=\pi(u)(a)\cdot v+a\cdot [u,v],\\
&&[u,v]+[v,u]=\partial(\langle u,v\rangle),\\
&&\pi(a\cdot v)=a\pi(v),\\
&&\langle a\cdot u,v\rangle=a*\langle u,v\rangle-\pi(u)(\pi(v)(a)),\\
&&\pi(v)(\langle v_1,v_2\rangle)=\langle [v,v_1],v_2\rangle+\langle v_1,[v,v_2]\rangle,\\
&&\partial(a*a')=a\cdot \partial(a')+a'\cdot\partial(a),\\
&&[v,\partial(a)]=\partial(\pi(v)(a)),\\
&&\langle v,\partial(a)\rangle=\pi(v)(a)
\end{eqnarray*}
for $a,a'\in A$, $u,v,v_1,v_2\in\Gamma$.
\end{enumerate}
\end{dfn}
\begin{prop}\cite{LiY} Let $(A,*)$ be a unital commutative associative algebra and let $B$ be a module for $A$ as a nonassociative algebra . Then a vertex $A$-algebroid structure on $B$ exactly amounts to a 1-truncated conformal algebra structure on $C=A\oplus B$ with
\begin{eqnarray*}
&&a_ia'=0,\\
&&u_0v=[u,v],~u_1v=\langle u,v\rangle,\\
&&u_0a=\pi(u)(a),~ a_0u=-u_0a
\end{eqnarray*} for $a,a'\in A$, $u,v\in B$, $i=0,1$ such that
\begin{eqnarray*}
&&a\cdot(a'\cdot u)-(a*a')\cdot u=(u_0a)\cdot \partial a'+(u_0a')\cdot \partial a,\\
&&u_0(a\cdot v)-a\cdot (u_0v)=(u_0a)\cdot v,\\
&&u_0(a*a')=a*(u_0a')+(u_0a)*a',\\
&&a_0(a'\cdot v)=a'*(a_0v),\\
&&(a\cdot u)_1v=a*(u_1v)-u_0v_0a,\\
&&\partial(a*a')=a\cdot \partial(a')+a'\cdot \partial(a).
\end{eqnarray*}
\end{prop}
\begin{dfn} Let $I$ be a subspace of a vertex $A$-algebroid $B$. The vector space $I$ is called an {\em ideal } of the vertex $A$-algebroid $B$ if $I$ is a left ideal of the left Leibniz algebra $B$ and $a\cdot u\in I$ for all $a\in A$, $u\in I$
\end{dfn}
Let $(A,*)$ be a unital commutative associative algebra. Let $B$ be a vertex $A$-algebroid. We set $A\partial(A):=Span\{a\cdot\partial(a')~|~a,a'\in A\}$. The vector space $A\partial(A)$ is an ideal of the vertex $A$-algebroid $B$. Moreover, $A\partial(A)$ is an abelian Lie algebra.
Observe that for $a,a',a''\in A$, $u\in B$, we have
\begin{eqnarray*}
(a\cdot \partial(a'))_0a''&&=0,\text{ and }\\
(a\cdot \partial(a'))_0u&&=a\cdot (u_0\partial(a'))+(u_0a)\cdot\partial(a')+\partial(u_1( a\cdot \partial(a'))\in A\partial(A).
\end{eqnarray*}
\begin{dfn} Let $A$ be a commutative associative algebra. A Lie $A$-algebroid is a Lie algebra $\mathfrak{g}$ equipped with an $A$-module structure and a module action on $A$ by derivation such that
\begin{eqnarray*}
[u,av]&=&a[u,v]+(ua)v,\\
a(ua')&=&(au)a'
\end{eqnarray*}
for $u,v\in\mathfrak{g}$, $a,a'\in A$.
A module for a Lie $A$-algebroid $\mathfrak{g}$ is a vector space $W$ equipped with a $\mathfrak{g}$-module structure and an $A$-module structure such that
\begin{eqnarray*}
&&u(aw)-a(uw)=(ua)w,\\
&&a(uw)=(au)w
\end{eqnarray*}
for all $a\in A$, $u\in\mathfrak{g}$, $w\in W$.
\end{dfn}
\begin{prop}\cite{Br2} Let $A$ be a unital commutative associative algebra over $\mathbb{C}$ and let $B$ be a vertex $A$-algebroid . Then $B/A \partial(A)$ is a Lie $A$-algebroid.
\end{prop}
\begin{prop}\cite{LiY} Let $\mathfrak{g}$ be a Lie $A$-algebroid. Then $\mathfrak{g}$ is a Lie algebra with $A$ a $\mathfrak{g}$-module. By adjoining the $\mathfrak{g}$-module $A$ to $\mathfrak{g}$ we have a Lie algebra $A\oplus\mathfrak{g}$ with $A$ as an abelian ideal. Denote by $J$ the 2-sided ideal of the universal enveloping algebra $U(A\oplus\mathfrak{g})$ generated by the vectors $$1_A-1,~a\cdot a'-aa',~ a\cdot b-ab$$ for $a,a'\in A$, $b\in\mathfrak{g}$ where $1_A$ is the identity of $A$, and $\cdot $ denotes the product in the universal enveloping algebra. Set $$\overline{U}(A\oplus \mathfrak{g})=U(A\oplus \mathfrak{g})/J.$$ Then a (simple) module structure for the Lie $A$-algebroid $\mathfrak{g}$ on a vector space $W$ exactly amounts to a (simple) $\overline{U}(A\oplus\mathfrak{g})$-module structure.
\end{prop}
\section{Classification of vertex algebroids associated with non-Lie cyclic left Leibniz algebras}
In this section, we classify all vertex $A$-algebroids $B$ when $B$ are cyclic non-Lie left Leibniz algebras such that $B\neq A\partial(A)$ and $\dim B$ is either 2 or 3. In subsection 3.1, we will find all vertex $A$-algebroids $B$ when $B$ are 2-dimensional non-Lie cyclic left Leibniz algebras such that $B\neq A\partial(A)$. In subsection 3.2, we investigate and classify all vertex $A$-algebroids $B$ when $B$ are 3-dimensional cyclic non-Lie left Leibniz algebras such that $B\neq A\partial(A)$.
It is worth mentioning that the results in this section will play a crucial role in a study of representation theory of vertex algebras $V_B$ associated to vertex $A$-algebroids $B$ when $B$ is a non-Lie cyclic left Leibniz algebra such that $B\neq A\partial(A)$ and $2\leq \dim B\leq 3$, and an investigation on the relationships between vertex algebras $V_B$ and a rank one Heisenberg vertex operator algebra in Section 4.
Now, we let $A$ be a unital commutative associative algebra, and let $B$ be a vertex $A$-algebroid such that $B$ is a cyclic non-Lie left Leibniz algebra, $B\neq A\partial(A)$, and $\dim (B)=n$. Hence, $B$ is of the form $Span\{b,b_0b,....., (b_0)^{n-1}b\}$. We set $a=b_1b$. Since $b_0b=-b_0b+\partial(b_1b)$ and $B$ is a cyclic non-Lie left Leibniz algebra, we can conclude that $\partial(a)=2b_0b\neq 0$ and $a\neq 0$. Using the fact that $\{b, b_0b,.....,(b_0)^{n-1}b\}$ is a basis of $B$, and $(b_0)^i\partial(a)=\partial((b_0)^ia)$, we can conclude that $$\{b, \partial(a),...,\partial((b_0)^{n-2}a)\}$$ is a basis of $B$ as well. Since $\partial(A)\subseteq A\partial(A)\subseteq B$, and $1=\dim B/\partial(A)\geq\dim B/A\partial(A)\geq 1 $, we can conclude that $A\partial(A)=\partial(A)$, and
\begin{equation*}\{\partial(a),...,\partial((b_0)^{n-2}a) \}\end{equation*} is a basis for $\partial(A)$.
\begin{prop}\cite{JY} Let $A$ be a finite dimensional unital commutative associative algebra with the identity $1_A$ and let $B$ be a finite dimensional non-Lie left Leibniz algebra. Assume that $B$ is a vertex $A$-algebroid. If $Ker(\partial)=\mathbb{C}1_A$ then $A$ is a local algebra.
\end{prop}
\begin{prop} Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid such that $B$ is a cyclic non-Lie left Leibniz algebra, $B\neq A \partial(A)$, and $\dim (B)=n$. Then, there exists $b\in B$ such that $\{b,b_0b,....., (b_0)^{n-1}b\}$ is a basis for $B$. We set $a=b_1b$. Assume that $Ker(\partial)=\mathbb{C}1_A$. Then $A$ is a local algebra with a basis $\{1_A, a,b_0a,....,(b_0)^{n-2}a\}$.
\end{prop}
\begin{proof} First, we will show that $\{1_A,a,b_0a,....,(b_0)^{n-2}a\}$ is linearly independent. We set
\begin{equation}\label{Aindependent}\alpha 1_A+\alpha_0 a+\alpha_1 b_0a+.....+\alpha_{n-2}(b_0)^{n-2}a=0.\end{equation} Applying $\partial$ to (\ref{Aindependent}), we have $$\alpha_0\partial(a)+\alpha_1\partial(b_0a)+....+\alpha_{n-2}\partial((b_0)^{n-2}a)=0.$$ Since $\{\partial(a),\partial(b_0a),....,\partial((b_0)^{n-2}a)\}$ is linearly independent, we can conclude that
$\alpha_0=\alpha_1=....=\alpha_{n-2}=0$. In addition, we have $\alpha 1_A=0$. This implies that $\alpha=0$, and $\{1_A,a,b_0a,....,(b_0)^{n-2}a\}$ is linearly independent.
Next, we show that $A=Span\{1_A,a,b_0a,....,(b_0)^{n-2}a\}$. Let $a'\in A$. Then $\partial(a')\in \partial(A)$. Hence, there exist $\beta_0,....,\beta_{n-2}\in\mathbb{C}$ such that $$\partial(a')=\beta_0\partial(a)+\beta_1\partial(b_0a)+....+\beta_{n-2}\partial((b_0)^{n-2}a).$$ Consequently, we have $a'-(\beta_0a+\beta_1 b_0a+....+\beta_{n-2}(b_0)^{n-2}a) \in Ker(\partial).$ Since $Ker(\partial)=\mathbb{C}1_A$, this implies that there exists $\chi\in \mathbb{C}$ such that $$a'=\beta_0a+\beta_1 b_0a+....+\beta_{n-2}(b_0)^{n-2}a+\chi 1_A.$$ Therefore, $A=Span\{1_A,a,b_0a,....,(b_0)^{n-2}a\}$.
\end{proof}
\subsection{Classification of Vertex Algebroids associated to $2$-dimensional non-Lie left Leibniz algebras}
\ \
Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid such that $B$ is a non-Lie left Leibniz algebra, $\dim(B)=2$, and $B\neq A\partial(A)$. Hence, there exists $b\in B$ such that $\{b,b_0b\}$ is a basis of $B$.
We set $a=b_1b$. Assume that $Ker(\partial)=\mathbb{C}1_A$. Then $A$ is a local algebra with a basis $\{1_A,a\}$, and $\{b,\partial(a)\}$ is a basis for $B$. For convenience, we also set $$a*a=\alpha_1 1_A+\alpha_2 a.$$ Here, $\alpha_1,\alpha_2\in\mathbb{C}$. Clearly, we have $$a\cdot \partial(a)=\frac{1}{2}\alpha_2\partial(a).$$ Since $B$ is a 2-dimensional left Leibniz algebra, by Proposition \ref{cyclicclass}, we know that either $b_0(b_0b)=0$ or $b_0(b_0b)=b_0b$. In Theorem \ref{Bcyclicnilpotent} and Theorem \ref{Bcyclicsolvable}, we study the algebraic structure of the vertex $A$-algebroid $B$ for the case when $b_0(b_0b)=0$ and for the case when $b_0(b_0b)=b_0b$, respectively.
\begin{thm}\label{Bcyclicnilpotent} Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid such that $B$ is a cyclic non-Lie left Leibniz algebra, $\dim B=2$, $B\neq A \partial(A)$, and $Ker(\partial)=\mathbb{C}1_A$. Clearly, there exists $b\in B$ such that $\{ b,b_0b\}$ is a basis for $B$. When we set $a=b_1b$, the set $\{b,\partial(a)\}$ is a basis of $B$ and the set $\{1_A,a\}$ is a basis for $A$.
Assume that $b_0(b_0b)=0$. Then
\noindent(i) $b_0a=0$,
\noindent(ii) $a\cdot b\in\partial(A)$,
\noindent(iii) $a*a=0$, $A\cong \mathbb{C}[x]/(x^2)$, and
\noindent(iv) $a\cdot\partial(a)=0$.
\noindent (v) In addition, $B$ is a module of $A$ as a commutative associative algebra.
\noindent (vi) The ideal $(a)$ is the unique maximal ideal of $A$. Moreover, for $u\in (a)$, $w\in B/A\partial(A)$, $w_0u=0$ and $u\cdot w=0$.
\end{thm}
\begin{proof} Assume that $b_0b_0b=0$. Because $\partial(a)=2b_0b$, we have $b_0\partial(a)=0$. First, we will prove statement (i). Since $\partial(b_0a)=b_0\partial(a)=0$, we then have that $b_0a\in Ker(\partial)=\mathbb{C}1_A$. So, there exists $\lambda\in\mathbb{C} $ such that $b_0a=\lambda 1_A$. Moreover, $b_0b_0a=0.$
We set $a\cdot b=\beta_1 b+\beta_2 \partial(a)$. Recall that $a\cdot \partial(a)=\frac{1}{2}\alpha_2\partial(a) $. Since
\begin{eqnarray*}
b_0(a\cdot b)&&=a\cdot b_0b+(b_0a)\cdot b=\frac{1}{2}a\cdot \partial(a)+\lambda b=\frac{1}{4}\alpha_2\partial(a)+\lambda b\text{ and}\\
b_0(a\cdot b)&&=b_0(\beta_1 b+\beta_2 \partial(a) )=\beta_1 \frac{1}{2}\partial(a)+\beta_2 \partial(b_0a)=\beta_1\frac{1}{2}\partial(a),
\end{eqnarray*}
we have $\beta_1\frac{1}{2}\partial(a)=\frac{1}{4}\alpha_2\partial(a)+\lambda b$.
Because $\{b,\partial(a)\}$ is linearly independent, we can conclude that $\lambda=0$ and $\frac{1}{4}\alpha_2=\frac{1}{2}\beta_1$. In addition, we have
\begin{equation}\label{adotbB2}
b_0a=0,~\beta_1=\frac{1}{2}\alpha_2\text{ and }a\cdot b=\frac{1}{2}\alpha_2 b+\beta_2 \partial(a).\end{equation} This proves $(i)$.
Next, we will proof statement $(v)$. Recall that for $a',a''\in A$, $u\in B$, $$a''\cdot (a'\cdot u)-(a''*a')\cdot u=(u_0a'')\cdot \partial(a')+(u_0a')\cdot \partial(a'').$$ If we set $u=b$, $a'=a''=a$, then
$$a\cdot(a\cdot b)-(a*a)\cdot b=(b_0a)\cdot \partial(a)+(b_0a)\cdot \partial(a).$$ Since $b_0a=0$, we have
\begin{equation}\label{aab} a\cdot (a\cdot b)-(a*a)\cdot b=0.\end{equation}
Similarly, if we set $u=\partial(a)$, $a'=a''=a$, then
\begin{equation}\label{aapartiala} a\cdot (a\cdot \partial(a))-(a*a)\cdot \partial(a)=((\partial(a))_0a)\cdot \partial(a)+((\partial(a))_0a)\cdot\partial(a)=0. \end{equation}
By (\ref{aab}), (\ref{aapartiala}), we can conclude that $B$ is a module of $A$ as a commutative associative algebra. This proves $(v)$.
Now, we will prove statements $(ii)-(iv)$. Recall that for $u,v\in B$, $a\in A$, $$(a\cdot u)_1v=a*(u_1v)-u_0v_0a.$$ When we set $u=v=b$, we have
$(a\cdot b)_1b=a*(b_1b)-b_0b_0a$. Since $b_1b=a$ and $b_0a=0$, we obtain that
\begin{equation}\label{abaa}
(a\cdot b)_1b=a*a.
\end{equation}
By (\ref{adotbB2}), we have $$a*a=(a\cdot b)_1b=(\frac{1}{2}\alpha_2 b+\beta_2 \partial(a) )_1b=\frac{1}{2}\alpha_2 a+\beta_2(b_0a)=\frac{1}{2}\alpha_2 a.$$
Since $a*a=\alpha_1 1_A+\alpha_2 a$ and $a*a=\frac{1}{2}\alpha_2 a$, we can conclude that
$\alpha_1 1_A+\frac{1}{2}\alpha_2 a=0$. This implies that $\alpha_1=0$, and $\alpha_2=0$. In addition, $$a*a=0,~a\cdot b=\beta_2\partial(a)\in \partial(A),~a\cdot\partial(a)=\frac{1}{2}\alpha_2\partial(a)=0.$$ This proves $(ii)-(iv)$. The statements $(v)$, $(vi)$ are consequences of $(i)$-$(iv)$.
\end{proof}
\begin{thm}\label{Bcyclicsolvable} Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid such that $B$ is a cyclic non-Lie left Leibniz algebra, $\dim B=2$, $B\neq A\partial(A)$, and $Ker(\partial)=\mathbb{C}1_A$. There exists $b\in B$ such that $\{ b,b_0b\}$ is a basis for $B$. We set $a=b_1b$. Then $\{b,\partial(a)\}$ is a basis of $B$ and $\{1_A,a\}$ is a basis for $A$.
Assume that $b_0(b_0b)=b_0b$. If we set $a*a=\alpha_1 1_A +\alpha_2 a$, then we have
\noindent(i) $\alpha_1=-\frac{1}{4}\alpha_2^2$,
\noindent(ii) $b_0a=a-\frac{1}{2}\alpha_2 1_A$,
\noindent(iii) $a\cdot b=\frac{1}{2}\alpha_2 b+(\frac{1}{2}\alpha_2-1)\partial(a)$,
\noindent(iv) $a*a=\alpha_2 a-\frac{1}{4}\alpha_2^2 1_A$, $a\cdot\partial(a)=\frac{\alpha_2}{2}\partial(a)$ and $A\cong \mathbb{C}[x]/(x-\frac{1}{2}\alpha_2)^2$.
\noindent (v) The vector space $(a-\frac{1}{2}\alpha_2 1_A )$ is the unique maximal ideal of $A$. In addition, for $u\in (a-\frac{1}{2}\alpha_2 1_A)$, $w\in B/A\partial(A)$, $w_0u\in (a-\frac{1}{2}\alpha_2 1_A)$ and $u\cdot w=0$.
\end{thm}
\begin{proof} Since $\partial(b_0a)=\partial(a)$, we then have that $\partial(b_0a-a)=0$, and $b_0a-a\in Ker(\partial)$. Since $Ker(\partial)=\mathbb{C} 1_A$, we can conclude that $b_0a=a+\rho 1_A$ for some $\rho\in\mathbb{C}$.
Now, we set $a\cdot b=\beta_1 b+\beta_2\partial(a)$. Since $b_0(a\cdot b)=a\cdot (b_0b)+(b_0a)\cdot b$, and $b_0a=a+\rho 1_A$, we then have that $b_0(a\cdot b)=a\cdot \frac{1}{2}\partial(a)+(a+\rho 1_A )\cdot b.$
Since
\begin{eqnarray*}
b_0(a\cdot b)&&=b_0(\beta_1 b+\beta_2\partial(a) )\\
&&=\beta_1\frac{1}{2}\partial(a)+\beta_2\partial(b_0a)\\
&&=\frac{1}{2}\beta_1\partial(a)+\beta_2\partial(a)\\
&&=(\frac{1}{2}\beta_1+\beta_2)\partial(a)
\end{eqnarray*}
and
\begin{eqnarray*}
b_0(a\cdot b)&&=a\cdot \frac{1}{2}\partial(a)+(a+\rho 1_A )\cdot b\\
&&=\frac{1}{4}\alpha_2\partial(a)+a\cdot b+\rho b\\
&&=\frac{1}{4}\alpha_2\partial(a)+\beta_1 b+\beta_2\partial(a)+\rho b\\
&&=(\beta_1+\rho)b+(\frac{1}{4}\alpha_2+\beta_2)\partial(a),
\end{eqnarray*}
we can conclude that $(\frac{1}{2}\beta_1+\beta_2)\partial(a) = (\beta_1+\rho)b+(\frac{1}{4}\alpha_2+\beta_2)\partial(a).$ Because $\{b,\partial(a)\}$ is a basis,
we have $\beta_1=-\rho\text{ and }\beta_1=\frac{1}{2}\alpha_2.$ These imply that $\rho=-\frac{1}{2}\alpha_2.$
In summary, we have
\begin{eqnarray*}
b_0a&&=a-\frac{1}{2}\alpha_2 1_A,\\
a\cdot b&&=\frac{1}{2}\alpha_2 b+\beta_2\partial(a).
\end{eqnarray*} This proves $(ii)$.
Since $b_0(a*a)=b_0(\alpha_1 1_A+\alpha_2 a)=\alpha_2b_0a=\alpha_2(a-\frac{1}{2}\alpha_2 1_A)=\alpha_2 a-\frac{1}{2}\alpha_2^2 1_A$ and
$$b_0(a*a)=a*(b_0a)+(b_0a)*a=2 a*(b_0a)=2 a*(a-\frac{1}{2}\alpha_2 1_A)= 2 a*a-\alpha_2 a,$$
we have
$$a*a=\alpha_2 a-\frac{1}{4}\alpha_2^2 1_A.$$ This proves $(iv)$.
Since $a*a=\alpha_1 1_A+\alpha_2 a$ and $\{1_A, a\}$ is a basis of $A$, we can conclude that
$$\alpha_1=-\frac{1}{4}\alpha_2^2.$$ This proves $(i)$.
Now, we will solve for $\beta_2$. Recall that $(a\cdot b)_1b=a*a-b_0(b_0a) $. Since
\begin{eqnarray*}
(a\cdot b)_1b&=&(\frac{1}{2}\alpha_2 b+\beta_2\partial(a) )_1b\\
&=&\frac{1}{2}\alpha_2 a+\beta_2b_0a,\\
&=&\frac{1}{2}\alpha_2 a+\beta_2(a-\frac{1}{2}\alpha_2 1_A )\text{ and }\\
(a\cdot b)_1b&=&a*a-b_0(b_0a)\\
&=&\alpha_2a-\frac{1}{4}\alpha_2^2 1_A-b_0(a-\frac{1}{2}\alpha_2 1_A )\\
&=&\alpha_2a-\frac{1}{4}\alpha_2^2 1_A-b_0a\\
&=&\alpha_2a-\frac{1}{4}\alpha_2^2 1_A-(a-\frac{1}{2}\alpha_2 1_A )\\
\end{eqnarray*}
we have
$\frac{1}{2}\alpha_2 a+\beta_2(a-\frac{1}{2}\alpha_2 1_A )=\alpha_2a-\frac{1}{4}\alpha_2^2 1_A-(a-\frac{1}{2}\alpha_2 1_A ).$
Equivalently, we have
$$(\frac{1}{2}\alpha_2+(-1-\beta_2))a+(-\frac{1}{4}\alpha_2^2+(1+\beta_2)(\frac{1}{2}\alpha_2))1_A=0.$$
Using the fact that $\{1_A,a\}$ is linearly independent, we can conclude that
$(\frac{1}{2}\alpha_2+(-1-\beta_2))=0$ and $(-\frac{1}{4}\alpha_2^2+(1+\beta_2)(\frac{1}{2}\alpha_2))=0.$ This implies that $$\beta_2=\frac{1}{2}\alpha_2-1.$$
Therefore, $a\cdot b=\frac{1}{2}\alpha_2 b+(\frac{1}{2}\alpha_2-1)\partial(a).$ This proves $(iii)$. The statement $(v)$ follows immediately from $(ii)$-$(iv)$.
\end{proof}
\subsection{Classification of Vertex Algebroids associated with $3$-dimensional cyclic non-Lie Leibniz algebras}
\ \
There are 4 types of cyclic non-Lie left Leibniz algebras (cf. Proposition \ref{cyclicclass} and Corollary \ref{cyclicclass3}). In this subsection, we will classify vertex algebroids associated with these 4 types of 3-dimensional cyclic non-Lie left Leibniz algebras. The results of these classification are in Theorem \ref{3nilcase}-Theorem \ref{3solvablecase3}.
\begin{lem}\label{dim3case} Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid such that $B$ is a cyclic non-Lie left Leibniz algebra, $\dim B=3$, $B\neq A\partial(A)$, and $Ker(\partial)=\mathbb{C}1_A$. Then there is $b\in B$ such that $\{b, b_0b, (b_0)^2b\}$ is a basis of $B$. In addition, if we set $a=b_1b$, then $\{1_A, a, b_0a\}$ is a basis of $A$, and $\{b,\partial(a),\partial(b_0a)\}$ is a basis of $B$.
Assume that
\begin{eqnarray}
&&a\cdot b=\beta b+\gamma_0\partial(a)+\gamma_1\partial(b_0a),\label{3relab}\\
&&b_0\partial(b_0a)=c_0 \partial(a)+c_1 \partial(b_0a).\label{3relb0b0a}
\end{eqnarray}
Here, $\beta, \gamma_0, \gamma_1, c_0, c_1\in \mathbb{C}$. Then we have the following statements:
\begin{eqnarray*}
&&b_0b_0a=\chi 1_A+c_0a+c_1b_0a\text{ for some }\chi\in\mathbb{C},\\
&&a*(b_0a)= \beta b_0a,\\
&&(b_0a)*(b_0a)=0,\\
&&(b_0a)\cdot \partial(b_0a)=0,\\
&&(b_0a)\cdot b\in A\partial(A),\\
&&a*a=(\gamma_1+1)\chi 1_A+(\beta +(\gamma_1+1)c_0)a+(\gamma_0+(\gamma_1+1)c_1)b_0a,\\
&&a\cdot \partial(a)=\frac{1}{2} ((\beta +(\gamma_1+1)c_0)\partial(a)+(\gamma_0+(\gamma_1+1)c_1)\partial(b_0a)).
\end{eqnarray*}
Moreover, the following statements hold
\begin{eqnarray*}
&&(\gamma_0+(\gamma_1+1)c_1)\chi =0,\\
&&(\gamma_0+(\gamma_1+1)c_1 )c_0=0,\\
&&\beta=(\gamma_1+1)c_0+(\gamma_0+(\gamma_1+1)c_1)c_1 .
\end{eqnarray*}
\end{lem}
\begin{proof} Because $\partial((b_0)^2a-(c_0a+c_1b_0a))=0\text{ (cf. equation (\ref{3relb0b0a})), and }Ker(\partial)=\mathbb{C}1_A,$ there exists $\chi\in \mathbb{C}$ such that $(b_0)^2 a=\chi 1_A+c_0a+c_1 b_0a$. Since $(b+A\partial(A))_0(a\cdot (b+A\partial(A))=a\cdot((b+A\partial(A))_0(b+A\partial(A)))+(b_0a)\cdot (b+A\partial(A)),$ and $a\cdot b=\beta b+\gamma_0\partial(a)+\gamma_1\partial(b_0a)$ we have
$$(b_0a)\cdot (b+A\partial(A))=0+A\partial(A),\text{ and }(b_0a)\cdot b\in A\partial(A).$$ Using the fact that $B/A\partial(A)$ is a Lie $A$-algebroid, we have
$$a*(b_0a)=a*((b+A\partial(A))_0a)=(a\cdot (b+A\partial(A))_0a=(\beta b+A\partial(A))_0a=\beta b_0a.$$
Because $(b_0a)\cdot b\in A\partial(A)$, we then obtain that $$(b_0a)*(b_0a)=(b_0a)*((b+A\partial(A))_0a)=((b_0a)\cdot (b+A\partial(A)))_0a=0.$$
Since $(a\cdot b)_1b=a*(b_1b)-b_0b_0a$, we have
\begin{eqnarray*}
a*a&=&a*(b_1b)\\
&=&(a\cdot b)_1 b+(b_0)^2a\\
&=&(\beta b+\gamma_0\partial(a)+\gamma_1\partial(b_0a) )_1b+(b_0)^2a\\
&=&\beta a+\gamma_0b_0a+\gamma_1(b_0)^2a+(b_0)^2a\\
&=&\beta a +\gamma_0 b_0a+(\gamma_1+1)(b_0)^2a\\
&=&\beta a +\gamma_0 b_0a+(\gamma_1+1)(\chi 1_A+c_0 a+c_1 b_0a )\\
&=&(\gamma_1+1)\chi 1_A+(\beta +(\gamma_1+1)c_0)a+(\gamma_0+(\gamma_1+1)c_1)b_0a.
\end{eqnarray*}
Also, since $b_0(a*a)=2a*(b_0a)=2\beta b_0a$, we can conclude that
\begin{eqnarray*}
&&2\beta b_0a\\
&&=b_0(a*a)\\
&&=(\beta +(\gamma_1+1)c_0)b_0a+(\gamma_0+(\gamma_1+1)c_1)(b_0)^2a\\
&&=(\beta +(\gamma_1+1)c_0)b_0a+(\gamma_0+(\gamma_1+1)c_1)(\chi 1_A+c_0a+c_1 b_0a )\\
&&=(\gamma_0+(\gamma_1+1)c_1)\chi 1_A+(\gamma_0+(\gamma_1+1)c_1 )c_0a+(\beta+(\gamma_1+1)c_0+(\gamma_0+(\gamma_1+1)c_1)c_1 )b_0a.
\end{eqnarray*}
Using the fact that $\{1_A, a, b_0a\}$ is a basis for $A$, we have
\begin{eqnarray*}
&&(\gamma_0+(\gamma_1+1)c_1)\chi =0,\\
&&(\gamma_0+(\gamma_1+1)c_1 )c_0=0,\\
&&(\gamma_1+1)c_0+(\gamma_0+(\gamma_1+1)c_1)c_1=\beta.
\end{eqnarray*}
\end{proof}
\begin{thm}\label{3nilcase}
Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid that has properties as in Lemma \ref{dim3case}.
Assume that $b_0\partial(b_0a)=0$. Then
\begin{eqnarray*}
&&\beta=0,~ a*(b_0a)=0,~ a*a=(\gamma_1+1)\chi 1_A+\gamma_0 b_0a,~a\cdot b=\gamma_0\partial(a)+\gamma_1\partial(b_0a),\\
&&a\cdot \partial(a)=\frac{1}{2}\gamma_0\partial(b_0a),\text{ and }\gamma_0\chi=0.
\end{eqnarray*}
\noindent (i) When $\gamma_0=0$, we have \begin{eqnarray*}
&&(b_0a)\cdot b=0,~\chi=0,~a*a=0,~a\cdot\partial(a)=0,~a\cdot b=\gamma_1\partial(b_0a),\\
&&A\cong \mathbb{C}[x,y]/(x^2,y^2,xy).
\end{eqnarray*}
The vector space $(a, b_0a)$ is the unique maximal ideal of $A$. In addition, for $u\in (a, b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a, b_0a)$ and $u\cdot w=0$.
\noindent (ii) When $\gamma_0\neq 0$, we have \begin{eqnarray*}
&&a*a=\gamma_0 b_0a,~a*a*a=0,~(b_0a)\cdot b=\frac{3}{4}\gamma_0\partial(b_0a),~A\cong \mathbb{C}[x]/(x^3).
\end{eqnarray*}
The vector space $(a )$ is the unique maximal ideal of $A$. For $u\in (a)$, $w\in B/A\partial(A)$, $w_0u\in (a)$ and $u\cdot w=0$.
\end{thm}
\begin{proof} By setting $c_0$, $c_1$ in Lemma \ref{dim3case} to be $0$, we have \begin{eqnarray*}
&&\beta=0,~ a*(b_0a)=0,~ a*a=(\gamma_1+1)\chi 1_A+\gamma_0 b_0a,~a\cdot b=\gamma_0\partial(a)+\gamma_1\partial(b_0a),\\
&&a\cdot \partial(a)=\frac{1}{2}\gamma_0\partial(b_0a),\text{ and }\gamma_0\chi=0.
\end{eqnarray*}
Now, we first consider the case when $\gamma_0=0$. Recall that for $u,v\in B, a'\in A$, we have $u_0(a'\cdot v)-a'\cdot(u_0v)=(u_0a')\cdot v.$ When we set $u=v=b$ and $a'=a$, we have
$$b_0(a\cdot b)-a\cdot (b_0b)=(b_0a)\cdot b.$$ Since $a\cdot b_0b=\frac{1}{2}a\cdot\partial(a)=0$ and $a\cdot b=\gamma_1\partial(b_0a)$, we have
\begin{equation}(b_0a)\cdot b=b_0(\gamma_1\partial(b_0a))=\gamma_1\partial(b_0b_0a)=0.\end{equation} Recall that for $a',a''\in A$, $u\in B$, we have $a'\cdot(a''\cdot u)-(a'*a'')\cdot u=(u_0a')\cdot \partial(a'')+(u_0a'')\cdot\partial(a').$ Now, if we set $a'=a$, $a''=b_0a$, $u=b$, then
$$a\cdot(b_0a\cdot b)-(a*(b_0a))\cdot b=(b_0a)\cdot \partial(b_0a)+(b_0b_0a)\cdot \partial(a). $$ Because $b_0a\cdot \partial(b_0a)=0$, $(b_0a)\cdot b=0$ and $a*(b_0a)=0$, we have $(b_0b_0a)\cdot \partial(a)=0$.
Also, because $b_0b_0a=\chi 1_A+c_0a+c_1b_0a$, and $c_0=c_1=0$, we have $b_0b_0a=\chi 1_A$. Since $(b_0b_0a)\cdot \partial(a)=0$, we have $\chi \partial(a)=0$. Therefore, \begin{equation}\chi =0,\text{ and }a*a=0.\end{equation} Hence, $A\cong\mathbb{C}[x,y]/(x^2,y^2,xy)$. In addition, the vector space $(a, b_0a)$ is the unique maximal ideal of $A$. For $u\in (a, b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a, b_0a)$ and $u\cdot w=0$ This completes the case when $\gamma_0=0$.
Next, we consider the case when $\gamma_0\neq 0$. Because $\gamma_0\chi=0$ and $\gamma_0\neq 0$, we have $\chi=0$. Since $\chi=0$ and $a*a=(\gamma_1+1)\chi 1_A+\gamma_0b_0a$, we have \begin{equation}a*a=\gamma_0b_0a.\end{equation} Since $a*(b_0a)=0$, we can conclude that \begin{equation}a*(a*a)=a*(\gamma_0b_0a)=0.\end{equation}
Recall that for $u,v\in B$, $a'\in A$, we have $u_0(a'\cdot v)-a'\cdot (u_0v)=(u_0a')\cdot v$. When we set $u,v=b$ and $a'=a$, we have
$$b_0(a\cdot b)-a\cdot b_0b=(b_0a)\cdot b.$$
Since $a\cdot b=\gamma_0\partial(a)+\gamma_1\partial(b_0a)$ and $b_0b=\frac{1}{2}\partial(a)$, we then have that $$b_0(\gamma_0\partial(a)+\gamma_1\partial(b_0a))-a\cdot \frac{1}{2}\partial(a)=(b_0a)\cdot b.$$ Equivalently,
$\gamma_0\partial(b_0a)+\gamma_1\partial(b_0b_0a)-\frac{1}{2}a\cdot\partial(a)=(b_0a)\cdot b$. Since $\partial(b_0b_0a)=0$ and $a\cdot\partial(a)=\frac{1}{2}\gamma_0\partial(b_0a)$, we have
\begin{equation}(b_0a)\cdot b=\gamma_0\partial(b_0a)-\frac{1}{4}\gamma_0\partial(b_0a)=\frac{3}{4}\gamma_0\partial(b_0a).\end{equation}
Hence, $A\cong \mathbb{C}[x]/(x^3)$. Moreover, $(a )$ is the unique maximal ideal of $A$. For $u\in (a)$, $w\in B/A\partial(A)$, $w_0u\in (a)$ and $u\cdot w=0$. This completes the case when $\gamma_0\neq 0$.
\end{proof}
\begin{thm}\label{3solvablecase1}
Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid that has properties as in Lemma \ref{dim3case}.
Assume that
$$b_0\partial(b_0a)=\partial(a)-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i \partial(b_0a)~(\alpha\notin\{0,1\}).$$ Then the following statments hold:
\begin{eqnarray*}
&&(\gamma_0+(\gamma_1+1)(-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i ))=0,~\beta=\gamma_1+1,~a*(b_0a)=(\gamma_1+1)b_0a,\\
&&a\cdot \partial(a)=(\gamma_1+1)\partial(a),\\
&&(b_0a)\cdot b=\gamma_1\partial(a)+\frac{\alpha+1}{\sqrt{\alpha}}i\partial(b_0a),\\
&&a\cdot b=(\gamma_1+1) b+(\gamma_1+1)\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i\partial(a)+\gamma_1\partial(b_0a).
\end{eqnarray*}
\noindent (i) If $\gamma_1+1\neq 0$ then
\begin{eqnarray*}
&&\chi=-\gamma_1-1,\\
&&b_0b_0a=(-\gamma_1-1 )1_A+a-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i b_0a,\\
&&a*a=-(\gamma_1+1)^2 1_A+2(\gamma_1+1)a,\\
&&A\cong \mathbb{C}[x,y]/( (x-(\gamma_1+1))^2, (x-(\gamma_1+1))y, y^2).
\end{eqnarray*}
The vector space $(a-(\gamma_1+1) ) 1_A )$ is the unique maximal ideal of $A$. In addition, for $u\in (a-( \gamma_1+1_A) 1_A)$, $w\in B/A\partial(A)$, $w_0u\in (a-(\gamma_1+1 ) 1_A)$ and $u\cdot w=0$.
\vspace{0.2cm}
\noindent (ii) If $\gamma_1+1=0$, then
\begin{eqnarray*}
&&\chi=0,\\
&&b_0b_0a=a-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i b_0a,\\
&&a*a=0,\\
&&a\cdot b=-\partial(b_0a),\\
&&A\cong \mathbb{C}[x,y]/( x^2, xy, y^2).
\end{eqnarray*}
The vector space $(a, b_0a )$ is the unique maximal ideal of $A$. Also, for $u\in (a,b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a, b_0a)$ and $u\cdot w=0$.
\end{thm}
\begin{proof} When we set $c_0=1$, $c_1=-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i $ in Lemma \ref{dim3case}, we have
\begin{eqnarray*}
&&(\gamma_0+(\gamma_1+1)(-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i ))\chi=0,\\
&&(\gamma_0+(\gamma_1+1)(-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i ))=0,\\
&&\beta=\gamma_1+1+(\gamma_0+(\gamma_1+1)( -\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i ))(-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i ).
\end{eqnarray*}
These imply that
\begin{eqnarray*}
&&\beta=\gamma_1+1,\\
&&a*(b_0a)=(\gamma_1+1)b_0a,\\
&&a*a=(\gamma_1+1)\chi 1_A+2(\gamma_1+1)a,\\
&&a\cdot \partial(a)=(\gamma_1+1)\partial(a).
\end{eqnarray*}
Moreover, $A\cong \mathbb{C}[x,y]/( x^2-2(\gamma_1+1)x-(\gamma_1+1)\chi, xy-(\gamma_1+1)y, y^2)$.
Since $\beta=\gamma_1+1$, $\gamma_0=(\gamma_1+1)\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i$, we have that
$$a\cdot b=(\gamma_1+1) b+(\gamma_1+1)\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i\partial(a)+\gamma_1\partial(b_0a). $$
Because
$b_0(a\cdot b)-a\cdot (b_0b)=(b_0a)\cdot b$, and $a\cdot \partial(a)=(\gamma_1+1)\partial(a)$, we have
\begin{eqnarray*}
&&(b_0a)\cdot b\\
&&=b_0((\gamma_1+1) b+(\gamma_1+1)\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i\partial(a)+\gamma_1\partial(b_0a) )-\frac{1}{2}a\cdot \partial (a)\\
&&=b_0((\gamma_1+1) b+(\gamma_1+1)\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i\partial(a)+\gamma_1\partial(b_0a) )-\frac{1}{2}(\gamma_1+1)\partial(a)\\
&&=\gamma_1\partial(a)+\frac{\alpha+1}{\sqrt{\alpha}}i\partial(b_0a).
\end{eqnarray*}
Since $c_0=1$, and $c_1=-\frac{\alpha+1}{\sqrt{\alpha}}i$, we have
$b_0b_0a=\chi 1_A+a-\frac{\alpha+1}{\sqrt{\alpha}}i b_0a$. Because $$a\cdot (b_0a\cdot b)-(a*b_0a)\cdot b=(b_0a)\cdot \partial(b_0a)+(b_0b_0a)\cdot \partial(a),$$
$a*(b_0a)=(\gamma_1+1)b_0a$, and $(b_0a)\cdot \partial(b_0a)=0$, we have
$$a\cdot (\gamma_1\partial(a)+\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i\partial(b_0a))-((\gamma_1+1)b_0a)\cdot b=(\chi 1_A+a-\frac{(\alpha+1)}{\sqrt{\alpha}}i b_0a)\cdot \partial(a).$$
This implies that
\begin{eqnarray*}
&&\chi \partial(a)\\
&&=(\gamma_1-1)a\cdot \partial(a)+\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i (a\cdot \partial(b_0a)+(b_0a)\cdot \partial(a))-(\gamma_1+1)(b_0a)\cdot b\\
&&=(\gamma_1-1)(\gamma_1+1) \partial(a)+\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i \partial(a*(b_0a))-(\gamma_1+1)(b_0a)\cdot b\\
&&=(\gamma_1^2-1)\partial(a)+\left(\frac{\alpha+1}{\sqrt{\alpha}}\right) i \partial((\gamma_1+1)b_0a)-(\gamma_1+1)(b_0a)\cdot b\\
\end{eqnarray*}
The above statement is equivalent to the following:
$$(\gamma_1^2-\chi-1)\partial(a)+\left(\frac{\alpha+1 }{\sqrt{\alpha}}\right) i(\gamma_1+1)\partial(b_0a)=(\gamma_1+1)(b_0a)\cdot b.$$
If $(\gamma_1+1)\neq 0$ then $(b_0a)\cdot b=\frac{\gamma_1^2-1-\chi}{\gamma_1+1}\partial(a)+\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i\partial(b_0a).$
Since
\begin{eqnarray*}
&&(b_0a)\cdot b=\gamma_1\partial(a)+\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i\partial(b_0a)\text{ and }\\
&&(b_0a)\cdot b=\frac{\gamma_1^2-1-\chi}{\gamma_1+1}\partial(a)+\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i\partial(b_0a),
\end{eqnarray*} we can conclude that
$\chi=-\gamma_1-1$ and $$b_0b_0a=(-\gamma_1-1 )1_A+a-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i b_0a.$$ In addition, $A\cong \mathbb{C}[x,y]/( (x-(\gamma_1+1))^2, (x-(\gamma_1+1))y, y^2)$. The vector space $(a-(\gamma_1+1) ) 1_A )$ is the unique maximal ideal of $A$. In addition, for $u\in (a-( \gamma_1+1_A) 1_A)$, $w\in B/A\partial(A)$, $w_0u\in (a-(\gamma_1+1 ) 1_A)$ and $u\cdot w=0$.
If $\gamma_1+1=0$, then $\chi \partial(a)=0$. This implies that $\chi=0$ and $b_0b_0a=a-\left(\frac{\alpha+1}{\sqrt{\alpha}}\right)i b_0a.$ In addition, we have $\beta=0$, $a*b_0a=0$, $a*a=0$, $a\cdot\partial(a)=0$, $\gamma_0=0$ and $a\cdot b=-\partial(b_0a)$. The unital commutativ associative algebra $A\cong \mathbb{C}[x,y]/( x^2, xy, y^2)$. The vector space $(a, b_0a )$ is the unique maximal ideal of $A$. Also, for $u\in (a,b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a, b_0a)$ and $u\cdot w=0$.
\end{proof}
\begin{thm}\label{3solvablecase2} Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid that has properties as in Lemma \ref{dim3case}.
Assume that
$b_0\partial(b_0a)=\partial(a)+2i \partial(b_0a).$ Then the following statments hold:
\begin{eqnarray*}
&&\gamma_0+(\gamma_1+1)2i=0,~\beta=\gamma_1+1,\\
&&a*(b_0a)=(\gamma_1+1)b_0a,~a*a=-(\gamma_1+1)^21_A+2(\gamma_1+1)a,\\
&&a\cdot \partial(a)=(\gamma_1+1)\partial(a),~(b_0a)\cdot b=\gamma_1\partial(a)-2i\partial(b_0a).
\end{eqnarray*} In addition, $A\cong \mathbb{C}[x,y]/( (x-(\gamma_1+1))^2, xy-(\gamma_1+1)y, y^2).$
The vector space $(a-(\gamma_1+1 ) 1_A, b_0a )$ is the unique maximal ideal of $A$. In addition, for $u\in (a-( \gamma_1+1) 1_A, b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a-(\gamma_1+1 )1_A)$ and $u\cdot w=0$.
\end{thm}
\begin{proof} In Lemma \ref{dim3case}, when we set $c_0=1$, and $c_1=2i$, the following statements hold
\begin{eqnarray*}
&&\gamma_0=-(\gamma_1+1)2i,~\beta=\gamma_1+1,\\
&&a*(b_0a)=(\gamma_1+1)b_0a,~a*a=(\gamma_1+1)\chi 1_A+2(\gamma_1+1)a,\\
&&a\cdot \partial(a)=(\gamma_1+1)\partial(a),~a\cdot b=(\gamma_1+1)b-(\gamma_1+1)2i\partial(a)+\gamma_1\partial(b_0a).
\end{eqnarray*}
Since $(b_0a)\cdot b=b_0(a\cdot b)-a\cdot (b_0b)$, we have
\begin{equation}(b_0a)\cdot b=\gamma_1\partial(a)-2i\partial(b_0a).\end{equation}
Because
\begin{eqnarray*}
&&b_0b_0a=\chi 1_A+a+2i b_0a,~a\cdot ((b_0a)\cdot b)-(a*(b_0a))\cdot b=(b_0a)\cdot \partial(b_0a)+(b_0b_0a)\cdot \partial(a),\\
&&a*(b_0a)=(\gamma_1+1)b_0a,~(\gamma_1+1)\partial(b_0a)=\partial(a*b_0a)=a\cdot\partial(b_0a)+(b_0a)\cdot \partial(a),
\end{eqnarray*}
we have
\begin{eqnarray*}
&&a\cdot (\gamma_1\partial(a)-2i\partial(b_0a))-(\gamma_1+1)(b_0a)\cdot b=(\chi 1_A+a+2i b_0a)\cdot \partial(a).
\end{eqnarray*}
Equivalently,
$(\chi+\gamma_1+1)\partial(a)=0.$ Therefore, $\chi=-(\gamma_1+1)$, and $$a*a=-(\gamma_1+1)^21_A+2(\gamma_1+1)a.$$ The rest follows immediately.
\end{proof}
\begin{thm}\label{3solvablecase3}
Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid that has properties as in Lemma \ref{dim3case}.
Assume that
$b_0\partial(b_0a)= \partial(b_0a).$ Then the following statments hold:
\begin{eqnarray*}
&&\chi=0,~\beta=\gamma_0+\gamma_1+1,\\
&&a*(b_0a)=\beta b_0a,~a*a=\beta a+\beta b_0a,\\
&&a\cdot \partial(a)=\frac{1}{2}(\beta\partial(a)+\frac{1}{2}\beta \partial(b_0a)).
\end{eqnarray*}
\begin{enumerate}
\item If $\gamma_0+\gamma_1+1=0$, then
\begin{eqnarray*}
&&a\cdot b=(-1-\gamma_1)\partial(a)+\gamma_1\partial(b_0a),~b_0b_0a=b_0a,\\
&&a*(b_0a)=0,~a*a=0,~a\cdot \partial(a)=0,~(b_0a)\cdot b=-\partial(b_0a),\\
&&A\cong \mathbb{C}[x,y]/( x^2, y^2,xy).
\end{eqnarray*}
The vector space $(a, b_0a )$ is the unique maximal ideal of $A$. For $u\in (a,b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a,b_0a)$ and $u\cdot w=0$.
\item If $\gamma_0+\gamma_1+1\neq 0$, then
\begin{eqnarray*}
&&b_0b_0a=b_0a,~ a*a=(\gamma_0+\gamma_1+1)(a+b_0a),\text{ and }\\
&&A\cong \mathbb{C}[x,y]/( x^2-(\gamma_0+\gamma_1+1)(x+y), xy-(\gamma_0+\gamma_1+1)y, y^2).
\end{eqnarray*}
The vector space $(a-(\gamma_0+\gamma_1+1) 1_A, b_0a )$ is the unique maximal ideal of $A$. In addition, for $u\in (a-(\gamma_0+\gamma_1+1 ) 1_A, b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a-(\gamma_0+\gamma_1+1 ) 1_A,b_0a)$ and $u\cdot w=0$.
\end{enumerate}
\end{thm}
\begin{proof} If we set $c_0=0$, $c_1=1$ in Lemma \ref{dim3case}, we then have that
\begin{eqnarray*}
&&(\gamma_0+(\gamma_1+1))\chi=0,~\beta=\gamma_0+\gamma_1+1,\\
&&a*a=(\gamma_1+1)\chi 1_A+(\gamma_0+\gamma_1+1)a+(\gamma_0+\gamma_1+1)b_0a,\\
&&a\cdot \partial(a)=\frac{1}{2}((\gamma_0+\gamma_1+1)\partial(a)+(\gamma_0+\gamma_1+1)\partial(b_0a)).
\end{eqnarray*}
Because $(b_0a)\cdot b=b_0(a\cdot b)-a\cdot (b_0b)$, we have
\begin{eqnarray*}
&&(b_0a)\cdot b\\
&&=b_0(\beta b+\gamma_0\partial(a)+\gamma_1\partial(b_0a))-\frac{1}{2}a\cdot \partial(a)\\
&&=\frac{1}{2}\beta \partial(a)+\gamma_0\partial(b_0a)+\gamma_1\partial(b_0a)-\frac{1}{2}(\frac{1}{2}\beta\partial(a)+\frac{1}{2}\beta\partial(b_0a))\\
&&=\frac{1}{4}\beta\partial(a)+(\gamma_0+\gamma_1-\frac{1}{4}\beta)\partial(b_0a)\\
&&=\frac{1}{4}\beta\partial(a)+(\frac{3}{4}\beta-1)\partial(b_0a).
\end{eqnarray*}
Now, assume that $\gamma_0+\gamma_1+1=0$. For this case, we have
\begin{eqnarray*}
&&\beta=0,~b_0b_0a=\chi 1_A+b_0a,~a*b_0a=\beta b_0a=0,\\
&&(b_0a)\cdot \partial(b_0a)=0,~a\cdot \partial(a)=0,~(b_0a)\cdot b=-\partial(b_0a).
\end{eqnarray*}
Recall that $ a\cdot(b_0a\cdot b)-(a*b_0a)\cdot b=(b_0a)\cdot \partial(b_0a)+(b_0b_0a)\cdot \partial(a)$. Since
$$a\cdot ((b_0a)\cdot b)-(a*(b_0a))\cdot b=a\cdot (-\partial(b_0a))$$ and
$(b_0a)\cdot \partial(b_0a)+(b_0b_0a)\cdot \partial(a)=(\chi 1_A+b_0a)\cdot \partial(a)$, we can conclude that
$$\chi\partial(a)+(b_0a)\cdot \partial(a)=-a\cdot \partial(b_0a).$$ Equivalently, we have $\chi \partial(a)+\partial(a*b_0a)=0$. Since $a*b_0a=0$, we then have that $\chi \partial(a)=0$. This implies that $\chi=0$, and $a*a=0$. The unital commutative associative algebra $A\cong \mathbb{C}[x,y]/( x^2, y^2,xy)$. The vector space $(a, b_0a )$ is the unique maximal ideal of $A$. For $u\in (a,b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a,b_0a)$ and $u\cdot w=0$.
Next, we assume that $\gamma_0+\gamma_1+1\neq 0$. Then $\chi=0$, $b_0b_0a=b_0a$, and $$a*a=(\gamma_0+\gamma_1+1)(a+b_0a).$$ The unital commutative associative algebra $A\cong \mathbb{C}[x,y]/( x^2-(\gamma_0+\gamma_1+1)(x+y), xy-(\gamma_0+\gamma_1+1)y, y^2)$. The vector space $(a-(\gamma_0+\gamma_1+1) 1_A, b_0a )$ is the unique maximal ideal of $A$. In addition, for $u\in (a-(\gamma_0+\gamma_1+1 ) 1_A, b_0a)$, $w\in B/A\partial(A)$, $w_0u\in (a-(\gamma_0+\gamma_1+1 ) 1_A,b_0a)$ and $u\cdot w=0$.
\end{proof}
\section{On vertex algebras associated to non-Lie cyclic left Leibniz algebras}
For this section, we discuss representation theory of vertex algebras $V_B$ associated to vertex $A$-algebroids $B$ when $B$ are non-Lie cyclic left Leibniz algebras, and $2\leq \dim B\leq 3$ from various aspects. Precisely, in subsection \ref{generaltheory}, we review a general construction of a vertex algebra $V_B$ associated to an arbitrary vertex $A$-algebroid $B$ and its modules. In subsection \ref{moduletheory} and subsection \ref{relationtoHeisenberg}, we focus our attention on vertex algebras $V_B$ when $B$ are vertex $A$-algebroids that are non-Lie cyclic left Leibniz algebras, and $2\leq \dim B\leq 3$. Also, we discuss representation theory of vertex algebras $V_B$ associated to vertex $A$-algebroids $B$ from representation theory of Lie algebroid point of view, and describe relations between specific types of vertex algebras $V_B$, and a rank one Heisenberg vertex algebra.
\begin{prop}\cite{JY}Let $V=\oplus_{n=0}^{\infty}V_{(n)}$ be a $\mathbb{N}$-graded vertex algebra such that $V_{(0)}$ is a finite dimensional unital commutative associative algebra and $\dim V_{(0)}\geq 2$. If $V_{(0)}$ is a local algebra then $V$ is indecomposable.
\end{prop}
\begin{prop}\label{indecomposableVB}\cite{JY} Let $V=\oplus_{n=0}^{\infty}V_{(n)}$ be a $\mathbb{N}$-graded vertex algebra that satisfies the following properties:
\vspace{0.2cm}
\noindent (a) $2\leq \dim V_{(0)}<\infty$, $1\leq dim V_{(1)}<\infty$, $V$ is generated by $V_{(0)}$ and $V_{(1)}$;
\vspace{0.2cm}
\noindent (b) $V_{(0)}$ is a local algebra.
\vspace{0.2cm}
\noindent If $V_{(0)}$ is not a simple module for a Lie $V_{(0)}$-algebroid $V_{(1)}/{V_{(0)}}D(V_{(0)})$, then $V$ is an indecomposable non-simple vertex algebra.
\end{prop}
\begin{thm}\label{v0v1cyclic} Let $V=\oplus_{n=0}^{\infty}V_{(n)}$ be an $\mathbb{N}$-graded vertex algebra that satisfies the following properties:
\vspace{0.2cm}
\noindent (a) $2\leq \dim V_{(0)}<\infty$, $1\leq dim V_{(1)}<\infty$, $V$ is generated by $V_{(0)}$ and $V_{(1)}$;
\vspace{0.2cm}
\noindent (b) $V_{(1)}$ is a cyclic left Leibniz algebra such that $\dim V_{(1)}$ is either 2 or 3;
\noindent (c) $Ker(D|_{V_{(0)}})=\mathbb{C}{\bf 1}$
\vspace{0.2cm}
\noindent Then $V$ is an indecomposable non-simple $\mathbb{N}$-graded vertex algebra.
\end{thm}
\begin{proof} Since $V$ is an $\mathbb{N}$-graded vertex algebra, we can conclude immediately that $V_1$ is a vertex $V_0$-algebroid. Since $Ker(D|_{V_{(0)}})=\mathbb{C}{\bf 1}$, we can conclude immediately that $V_{(0)}$ is a local algebra. By Theorem \ref{Bcyclicnilpotent}, Theorem \ref{Bcyclicsolvable}, Theorem \ref{3nilcase}-Theorem \ref{3solvablecase3}, we have that $V_0$ is not a simple module for the Lie $V_{(0)}$-algebroid $V_{(1)}/V_{(0)}D(V_{(0)})$.
By Proposition \ref{indecomposableVB}, we can conclude that $V$ is an indecomposable non-simple $\mathbb{N}$-graded vertex algebra.
\end{proof}
\subsection{A review on vertex algebras $V_B$ associated with vertex algebroids $B$ and their modules}\label{generaltheory}
\noindent In this subsection, we recall a construction of vertex algebras associated with vertex algebroids in \cite{LiY}.
\vspace{0.2cm}
\noindent Let $A$ be a commutative associative algebra with identity $1_A$ and let $B$ be a vertex $A$-algebroid. We set $L(A\oplus B)=(A\oplus B)\otimes \mathbb{C}[t,t^{-1}].$
Subspaces $L(A)$ and $L(B)$ of $L(A\oplus B)$ are defined in the obvious way.
\vspace{0.2cm}
\noindent We set $\hat{\partial}=\partial\otimes 1+1\otimes \frac{d}{dt}:L(A)\rightarrow L(A\oplus B).$ We define
$deg(a\otimes t^n)=-n-1$, $deg(b\otimes t^n)=-n$ for $a\in A, ~b\in B,~n\in\mathbb{Z}$.
Then $L(A\oplus B)$ becomes a $\mathbb{Z}$-graded vector space:
$$L(A\oplus B)=\oplus_{n\in\mathbb{Z}}L(A\oplus B)_{(n)}$$ where
$L(A\oplus B)_{(n)}=A\otimes \mathbb{C}t^{-n-1}+B\otimes \mathbb{C}t^{-n}$. Clearly, the subspaces $L(A)$ and $L(B)$ are $\mathbb{Z}$-graded vector spaces as well. In addition, for $n\in \mathbb{N}$, $L(A)_{(n)}=A\otimes \mathbb{C}t^{-n-1}.$
The linear map $\hat{\partial}:L(A)\rightarrow L(A\oplus B)$ is of degree 1. We define a bilinear product $[\cdot,\cdot]$ on $L(A\oplus B)$ as follow:
\begin{eqnarray}
&&[a\otimes t^m,a'\otimes t^n]=0,\label{aa'}\\
&&[a\otimes t^m, b\otimes t^n]=a_0b\otimes t^{m+n},\\
&&[b\otimes t^n,a\otimes t^m]=b_0a\otimes t^{m+n},\\
&&[b\otimes t^m,b'\otimes t^n]=b_0b'\otimes t^{m+n}+m(b_1b')\otimes t^{m+n-1}\label{bb'}
\end{eqnarray}
for $a,a'\in A$, $b,b'\in B$, $m,n\in\mathbb{Z}$. For convenience, we set
$$\mathcal{L}:=L(A\oplus B)/\hat{\partial}L(A).$$ It was shown in \cite{LiY} that $\mathcal{L}=\oplus_{n\in\mathbb{Z}}\mathcal{L}_{(n)}$ is a $\mathbb{Z}$-graded Lie algebra. Here,
$$\mathcal{L}_{(n)}=L(A\oplus B)_{(n)}/\hat{\partial}(L(A)_{(n-1)})=(A\otimes \mathbb{C}t^{-n-1}+B\otimes \mathbb{C}t^{-n})/\hat{\partial}(A\otimes\mathbb{C}t^{-n}).$$ In particular, $\mathcal{L}_{(0)}=A\otimes \mathbb{C}t^{-1}+B/\partial A$.
\vspace{0.2cm}
\noindent Let $\rho:L(A\oplus B)\rightarrow\mathcal{L}$ be a natural linear map defined by $$\rho( u\otimes t^n)=u\otimes t^n+\hat{\partial}L(A).$$ For $u\in A\oplus B$, $n\in\mathbb{Z}$, we set $u(n)=\rho(u\otimes t^n)$ and $u(z)=\sum_{n\in\mathbb{Z}}u(n)z^{-n-1}$. Let $W$ be a $\mathcal{L}$-module. We use $u_W(n)$ or sometimes just $u(n)$ for the corresponding operator on $W$ and we write $u_W(z)=\sum_{n\in\mathbb{Z}}u(n)z^{-n-1}\in ({\rm End}\, W)[[z,z^{-1}]]$. The commutator relations in terms of generating functions are the following:
\begin{eqnarray}
&&[a(z_1),a'(z_2)]=0,\label{vaaa'}\\
&&[a(z_1), b(z_2)]=z_2^{-1}\delta\left(\frac{z_1}{z_2}\right)(a_0b)(z_2),\label{vaab}\\
&&[b(z_1),b'(z_2)]=z_2^{-1}\delta\left(\frac{z_1}{z_2}\right)(b_0b')(z_2)+(b_1b')(z_2)\frac{\partial}{\partial z_2}z_2^{-1}\delta\left(\frac{z_1}{z_2}\right)\label{vabb'}
\end{eqnarray}
for $a,a'\in A$, $b,b'\in B$.
\vspace{0.2cm}
\noindent Next, we define $\mathcal{L}^{\geq 0}=\rho((A\oplus B)\otimes \mathbb{C}[t])\subset \mathcal{L}$, and $\mathcal{L}^{<0}=\rho((A\oplus B)\otimes t^{-1}\mathbb{C}[t^{-1}])\subset \mathcal{L}$. As a vector space, $\mathcal{L}=\mathcal{L}^{\geq 0}\oplus \mathcal{L}^{<0}$. The subspaces $\mathcal{L}^{\geq 0}$ and $\mathcal{L}^{<0}$ are graded sub-algebras of $\mathcal{L}$.
\vspace{0.2cm}
\noindent We now consider $\mathbb{C}$ as the trivial $\mathcal{L}^{\geq 0}$-module and form the following induced module $$V_{\mathcal{L}}=U(\mathcal{L})\otimes_{U(\mathcal{L}^{\geq 0})}\mathbb{C}.$$
In view of the Poincare-Birkhoff-Witt theorem, we have $V_{\mathcal{L}}=U(\mathcal{L}^{<0})$ as a vector space. We set ${\bf 1}=1\in V_{\mathcal{L}}$. We may consider $A\oplus B$ as a subspace: $$A\oplus B\rightarrow V_{\mathcal{L}}, \ \ a+b\mapsto a(-1){\bf 1}+b(-1){\bf 1}.$$
We assign $deg~\mathbb{C}=0$. Then $V_{\mathcal{L}}=\oplus_{n\in\mathbb{N}}(V_{\mathcal{L}})_{(n)}$ is a restricted $\mathbb{N}$-graded $\mathcal{L}$-module.
\begin{prop} \cite{FKRW, MeP}
\vspace{0.2cm}
\noindent There exists a unique vertex algebra structure on $V_{\mathcal{L}}$ with $Y(u,x)=u(x)$ for $u\in A\oplus B$. In fact, the vertex algebra $V_{\mathcal{L}}$ is a $\mathbb{N}$-graded vertex algebra and it is generated by $A\oplus B$. Furthermore, any restricted $\mathcal{L}$-module $W$ is naturally a $V_{\mathcal{L}}$-module with $Y_W(u,x)=u_W(x)$ for $u\in A\oplus B$. Conversely, any $V_{\mathcal{L}}$-module $W$ is naturally a restricted $\mathcal{L}$-module with $u_W(x)=Y_W(u,x)$ for $u\in A\oplus B$.
\end{prop}
\vspace{0.2cm}
\noindent Now, we set
\begin{eqnarray*}
&&E_0=Span\{1_A-{\bf 1},a(-1)a'-a*a'~|~a,a'\in A\}\subset (V_{\mathcal{L}})_{(0)},\\
&&E_1=Span\{a(-1)b-a\cdot b~|~a\in A,b\in\mathfrak{L}\}\subset (V_{\mathcal{L}})_{(1)},\\
&&E=E_0\oplus E_1.
\end{eqnarray*}
We define
$$I_{B}=U(\mathcal{L})\mathbb{C}[D]E.$$ The vector space $I_{B}$ is an $\mathcal{L}$-submodule of $V_{\mathcal{L}}$. We set $$V_{B}=V_{\mathcal{L}}/I_{B}.$$
\begin{prop}\cite{GMS, LiY} \ \
\vspace{0.2cm}
\noindent (i) $V_{B}$ is a $\mathbb{N}$-graded vertex algebra such that $(V_{B})_{(0)}=A$ and $(V_{B})_{(1)}=B$ (under the linear map $v\mapsto v(-1){\bf 1}$) and $V_{B}$ as a vertex algebra is generated by $A\oplus B$. Furthermore, for any $n\geq 1$,
\begin{eqnarray*}
&&(V_{B})_{(n)}\\
&&=span\{b_1(-n_1).....b_k(-n_k){\bf 1}~|~b_i\in B,n_1\geq...\geq n_k\geq 1, n_1+...+n_k=n\}.
\end{eqnarray*}
\vspace{0.2cm}
\noindent (ii) A $V_{B}$-module $W$ is a restricted module for the Lie algebra $\mathcal{L}$ with $v(n)$ acting as $v_n$ for $v\in A\oplus B$, $n\in\mathbb{Z}$. Furthermore, the set of $V_{B}$-submodules is precisely the set of $\mathcal{L}$-submodules.
\end{prop}
\begin{prop}\label{prop30}\cite{LiY}
Let $W=\oplus_{n\in\mathbb{N}}W_{(n)}$ be a $\mathbb{N}$-graded $V_{B}$-module with $W_{(0)}\neq\{ 0\}$. Then $W_{(0)}$ is an $A$-module with $a\cdot w=a_{-1}w$ for $a\in A$, $w\in W_{(0)}$ and $W_{(0)}$ is a module for the Lie algebra $B/A\partial(A)$ with $b\cdot w=b_0w$ for $b\in B$, $w\in W_{(0)}$. Furthermore, $W_{(0)}$ equipped with these module structures is a module for Lie $A$-algebroid $B/A\partial A$. If $W$ is graded simple, then $W_{(0)}$ is a simple module for Lie $A$-algebroid $B/A\partial A$.
\end{prop}
\vspace{0.2cm}
\noindent Now, we set $\mathcal{L}_{\pm }=\oplus_{n\geq 1}\mathcal{L}_{(\pm n)}$ and $\mathcal{L}_{\leq 0}=\mathcal{L}_{-}\oplus \mathcal{L}_{(0)}$. Let $U$ be a module for the Lie algebra $\mathcal{L}_{(0)}(=A\oplus B/\partial (A))$. Then $U$ is an $\mathcal{L}_{(\leq 0)}$-module under the following actions:
$$a(n-1)\cdot u=\delta_{n,0}a\cdot u,\ \ b(n)\cdot u=\delta_{n,0}b\cdot u\text{ for }a\in A,b\in B, n\geq 0.$$
Next, we form the induced $\mathcal{L}$-module $M(U)={\rm Ind}\,_{\mathcal{L}_{(\leq 0)}}^{\mathcal{L}}U$. Endow $U$ with degree 0, making $M(U)$ a $\mathbb{N}$-graded $\mathcal{L}$-module. In fact, $M(U)$ is a $V_{\mathcal{L}}$-module.
We set $$W(U)=span\{v_nu~|~v\in E, ~n\in\mathbb{Z}, ~u\in U\}\subset M(U),$$ and
$$M_{B}(U)=M(U)/U(\mathcal{L})W(U).$$
\begin{prop}\label{simplemodulerelations}\cite{LiY} \ \
\vspace{0.2cm}
\noindent (i) Let $U$ be a module for the Lie algebra $\mathcal{L}_{(0)}$. Then $M_{B}(U)$ is a $V_{B}$-module. If $U$ is a module for the Lie $A$-algebroid $B/A\partial A$ then $(M_{B}(U))_{(0)}=U$.
\vspace{0.2cm}
\noindent (ii) Let $U$ be a module for the Lie $A$-algebroid $B/A\partial A$. Then there exists a unique maximal graded $U(\mathcal{L})$-submodule $J(U)$ of $M(U)$ with the property that $J(U)\cap U=0$. Moreover, $L(U)=M(U)/J(U)$ is a $\mathbb{N}$-graded $V_{B}$-module such that $L(U)_{(0)}=U$ as a module for the Lie $A$-algebroid $B/A\partial A$. If $U$ is a simple $B/A\partial A$, $L(U)$ is a $\mathbb{N}$-graded simple $V_{B}$-module.
\vspace{0.2cm}
\noindent (iii) Let $W=\coprod_{n\in\mathbb{N}}W_{(n)}$ be an $\mathbb{N}$-graded simple $V_B$-module with $W_{(0)}\neq 0$. Then $W\cong L(W_{(0)})$.
\vspace{0.2cm}
\noindent (iv) For any complete set $H$ of representatives of equivalence classes of simple modules for the Lie $A$-algebroid $B/A\partial A$, $\{L(U)~|~U\in H\}$ is a complete set of representatives of equivalence classes of $\mathbb{N}$-graded simple $V_B$-modules.
\end{prop}
\subsection{On representation theory of vertex algebra $V_B$ when $B$ is a cyclic Leibniz algebra}\label{moduletheory}
The following theorem is an immediate consequence of Theorem \ref{v0v1cyclic}.
\begin{thm} Let $B$ be a vertex $A$-algebroid as in Theorem \ref{Bcyclicnilpotent}-Theorem \ref{Bcyclicsolvable}, Theorem \ref{3nilcase}- Theorem \ref{3solvablecase3}. Then $V_B$ is an indecomposable non-simple vertex algebra.
\end{thm}
Now, we will classify irreducible modules of the vertex algebra $V_B$ when $B$ are vertex $A$-algebroids as in Theorem \ref{Bcyclicnilpotent}-Theorem \ref{Bcyclicsolvable}, Theorem \ref{3nilcase}- Theorem \ref{3solvablecase3}. By Proposition \ref{simplemodulerelations}, it is enough to classify simple modules for the Lie $A$-algebroids $B/A\partial(A)$.
The following lemma is a consequence of the definitions of module of a Lie algebroid, decomposable module and non-irreducible module.
\begin{lem}\label{indecomposableirreducible} Let $U$ be a Lie $A$-algebroid $B/A\partial(A)$.
\noindent (i) If $U$ is a decomposable Lie $A$-algbroid $B/A\partial(A)$-module then $U$ is a decomposable module for the Lie algebra $B/A\partial(A)$ and $U$ is a decomposable $A$-module. Equivalently, if $U$ is either an indecomposable module for the Lie algebra $B/A\partial(A)$ or an indecomposable $A$-module, then $U$ is an indecomposable Lie $A$-algbroid $B/A\partial(A)$-module.
\noindent (ii) If $U$ is not an irreducible Lie $A$-algbroid $B/A\partial(A)$-module then $U$ is not an irreducible module for the Lie algebra $B/A\partial(A)$ and $U$ is not an irreducible $A$-module. Likewise, if $U$ is either an irreducible module for the Lie algebra $B/A\partial(A)$ or an irreducible $A$-module, then $U$ is an irreducible Lie $A$-algbroid $B/A\partial(A)$-module.
\end{lem}
\begin{lem}\label{solvabledimension} Let $(\mathfrak{a}, *)$ be a unital commutative associative algebra with the identity $1_{\mathfrak{a}}$. Let $\mathfrak{g}$ be a Lie $\mathfrak{a}$-algebroid such that $\mathfrak{g}$ is solvable. Assume that
\noindent (i) $\mathfrak{a}$ is a local algebra and $\mathfrak{a}=\mathbb{C}1_{\mathfrak{a}}\oplus \mathfrak{a}'$. Here $\mathfrak{a}'$ is a unique maximal ideal of $\mathfrak{a}$;
\noindent (ii) for $u\in \mathfrak{a}'$, $w\in \mathfrak{g}$, $w_0u\in \mathfrak{a}'$, and $u\cdot w=0$.
\noindent Then every finite dimensional irreducible Lie $\mathfrak{a}$-algebroid $\mathfrak{g}$-module $U\neq 0$ is one dimensional. In addition, $\mathfrak{a}'$ acts as zero on $U$.
\end{lem}
\begin{proof} Let $U\neq 0$ be a finite dimensional irreducible Lie $\mathfrak{a}$-algebroid $\mathfrak{g}$-module. Then $U$ is a module for the solvable Lie algebra $\mathfrak{g}$. In addition, $U$ contains a common eigenvector $v$ for all endomorphisms in $\mathfrak{g}$, and $K=\mathbb{C}v$ is a $\mathfrak{g}$-submodule of $U$. For $g\in \mathfrak{g}$, there exists $k_g\in\mathbb{C}$ such that $g\cdot v=k_gv$.
We set $\mathfrak{a}'\cdot v=Span\{\alpha'\cdot v~|~\alpha'\in \mathfrak{a}'\}\subset U$. Clearly, $\mathfrak{a}'\cdot v$ is a $\mathfrak{a}$-module.
\begin{itemize}
\item Case 1: $\mathfrak{g}$ acts trivially on $K$.
Suppose that $\mathfrak{a}'\cdot v\neq \{0\}$. Notice that for $g\in\mathfrak{g}$, $\alpha\in \mathfrak{a}'$, we have
$g\cdot (\alpha\cdot v)=\alpha \cdot (g\cdot v)+(g\cdot \alpha)\cdot v=(g\cdot \alpha)\cdot v\in \mathfrak{a}'\cdot v$. Hence, $\mathfrak{a}'\cdot v$ is a $\mathfrak{g}$-module. Since
\begin{eqnarray*}
&&g\cdot (a\cdot (\alpha\cdot v))-a\cdot (g\cdot (\alpha\cdot v))\\
&&=(g\cdot a)\cdot (\alpha\cdot v)
\end{eqnarray*}
and
\begin{eqnarray*}
&&a\cdot (g\cdot (\alpha v))=(a\cdot g)\cdot (\alpha\cdot v)
\end{eqnarray*}
for $g\in\mathfrak{g}, a\in\mathfrak{a},\alpha\in\mathfrak{a}'$, we can conclude that $\mathfrak{a}'\cdot v$ is a Lie $\mathfrak{a}$-algebroid $\mathfrak{g}$-module. Since $U$ is irreducible, $\mathfrak{a}'\cdot v\neq \{0\}$, we can conclude that $U=\mathfrak{a}'\cdot v$. This implies that $v\in\mathfrak{a}'\cdot v$. So, there exists $\lambda\in \mathfrak{a'}$ such that $v=\lambda v$. Equivalently, $(1_{\mathfrak{a}}-\lambda)v=0$. We know that $\lambda$ is not invertible since $\lambda\in \mathfrak{a'}$. Because $\mathfrak{a}$ is local, we can conclude that $1_{\mathfrak{a}}-\lambda$ is invertible. Since $1_{\mathfrak{a}}-\lambda$ is invertible and $(1_{\mathfrak{a}}-\lambda)v=0$, we have $v=0$. This is a contradiction because $v$ is an eigenvector. Hence, $\mathfrak{a}'\cdot v= \{0\}$. Also, $K$ is a module of the Lie $\mathfrak{a}$-algebroid $\mathfrak{g}$. This implies that $U=K$.
\item Case 2: $\mathfrak{g}$ does not trivially on $K$.
There exists $g'\in \mathfrak{g}$ such that $g'v=k_{g'}v\neq 0$. Notice that $k_{g'} a'\cdot v=a'(g'\cdot v)=(a'\cdot g')\cdot v=0$ for all $a'\in\mathfrak{a}'$. This implies that $K$ is closed under the action of $\mathfrak{a}$, and $a'\cdot v=0$ for all $a'\in\mathfrak{a}'$. In addition, $K$ is a module for the Lie $\mathfrak{a}$-algebroid $\mathfrak{g}$. Since $K$ is a Lie $\mathfrak{a}$-algebroid $\mathfrak{g}$-submodule of $U$ and $U$ is an irreducible Lie $\mathfrak{a}$-algebroid $\mathfrak{g}$-module, we can conclude that $U=K$ is one-dimensional.
\end{itemize}
\end{proof}
\begin{thm} Let $B$ be a vertex $A$-algebroid as in Theorem \ref{Bcyclicnilpotent}-Theorem \ref{Bcyclicsolvable}, Theorem \ref{3nilcase}-Theorem \ref{3solvablecase3}. If $U$ is a finite dimensional irreducible Lie $A$-algebroid $B/A\partial(A)$-module then there exists $v\in U$ such that $U=\mathbb{C}v$, the unique maximal ideal of $A$ acts as zero on $v$ and $b\cdot v=\lambda v$. Here, $\lambda\in\mathbb{C}$. In addition, $$\{\mathbb{C}v_{\lambda}~|~\lambda\in\mathbb{C},~ b\cdot v_{\lambda}=\lambda v_{\lambda},~\text{ the unique maximal ideal of $A$ acts as zero on $v_\lambda$}\}$$ is the complete set of representatives of equivalence classes of finite dimensional irreducible Lie $A$-algebroid $B/A\partial(A)$-modules, and
$$\{L(\mathbb{C}v_{\lambda})~|~\lambda\in\mathbb{C},~ b\cdot v_{\lambda}=\lambda v_{\lambda},~\text{ the unique maximal ideal of $A$ acts as zero on $v_\lambda$}\}$$ is the complete set of representatives of equivalence classes of irredudible $\mathbb{N}$-graded $V_B$-modules whose degree zero sub-spaces are finite dimensional.
\end{thm}
\begin{proof} Let $B$ be a vertex $A$-algebroid as in Theorem \ref{Bcyclicnilpotent}-Theorem \ref{Bcyclicsolvable}, Theorem \ref{3nilcase}-Theorem \ref{3solvablecase3}. Then $A$ is a local algebra with unique maximal ideal $A'$ such that $A=\mathbb{C}1_A\oplus A'$ (as a vector space). $B/A\partial(A)$ is a one dimensional abelian Lie algebra. Moreover, $B/A\partial(A)$ is a Lie $A$-algebroid. For $u\in A'$, $w\in B/A\partial(A)$, $w_0u\in A'$ and $u\cdot w=0+A\partial(A)$. By Lemma \ref{solvabledimension}, we can conclude that every finite dimensional irreducible Lie $A$-algebroid $B/A\partial(A)$-module $U\neq 0$ is one dimensional. In addition, $A'$ acts as zero on $U$. In addition, $$\{\mathbb{C}v_{\lambda}~|~\lambda\in\mathbb{C},~ b\cdot v_{\lambda}=\lambda v_{\lambda},~\text{ the unique maximal ideal of $A$ acts as zero on $v_\lambda$}\}$$ is the complete set of representatives of equivalence classes of finite dimensional irreducible Lie $A$-algebroid $B/A\partial(A)$-modules. By Proposition \ref{simplemodulerelations}, we can conclude further that $$\{L(\mathbb{C}v_{\lambda})~|~\lambda\in\mathbb{C},~ b\cdot v_{\lambda}=\lambda v_{\lambda},~\text{ the unique maximal ideal of $A$ acts as zero on $v_\lambda$}\}$$ is the complete set of representatives of equivalence classes of irredudible $\mathbb{N}$-graded $V_B$-modules whose degree zero sub-spaces are finite dimensional.
\end{proof}
\subsection{On a connection between vertex algebras $V_B$ and a vertex algebra associated with a rank one Heisenberg}\label{relationtoHeisenberg}
In this section, we connect the vertex algebra $V_B$ that we study with a well-known vertex operator algebra associated with a rank one Heisenberg. First, we briefly review construction of vertex algebra associated with rank one Heisenberg algebra. Next, we show that one can construct a Heisenberg vertex operator algebra from certain vertex algebras $V_B$.
We denote by $\mathfrak{h}$ a one-dimensional abelian Lie algebra spanned by $h$ with a bilinear form $\langle\cdot,\cdot\rangle$ such that $\langle h,h\rangle=1$. We set $$\hat{\mathfrak{h}}=\mathfrak{h}\otimes \mathbb{C}[t,t^{-1}]\oplus\mathbb{C}{\bf{k}}$$ the affinization of $\mathfrak{h}$ with bracket relations: for $a,b\in\mathfrak{h},m, n\in\mathbb{Z}$,
$$[a(m),b(n)]=m\langle a,b\rangle \delta_{m+n,0}{\bf{k}},~[{\bf{k}},a(m)]=0.$$ Here we define $a(m)=a\otimes t^m$ for $a\in\mathfrak{h}$, $m\in\mathbb{Z}$. We let $$\hat{\mathfrak{h}}^+=\mathfrak{h}\otimes t\mathbb{C}[t]\text{ and }\hat{\mathfrak{h}}^-=\mathfrak{h}\otimes t^{-1}\mathbb{C}[t^{-1}].$$ The vector spaces $\hat{\mathfrak{h}}^+$ and $\hat{\mathfrak{h}}^-$ are abelian subalgebras of $\hat{\mathfrak{h}}$.
Now, we consider the induced $\hat{\mathfrak{h}}$-module given by $$M(1)={\mathcal{U}}(\hat{\mathfrak{h}})\otimes_{{\mathcal{U}}(\mathbb{C}[t]\otimes \mathfrak{h}\oplus{\mathbb{C}}{\bf k})}\mathbb{C}{\bf 1}\cong S(\hat{\mathfrak{h}}^-).$$ Here $\mathcal{U}(\cdot)$ and $S(\cdot)$ denote the universal enveloping algebra and symmetric algebra, respectively. Also, $\mathfrak{h}\otimes \mathbb{C}[t]$ acts trivially on $\mathbb{C}{\bf 1}$ and $\bf{k}$ acts as multiplication by 1. Then $M(1)$ is a vertex algebra that is often called the vertex operator algebra associated to the rank one Heisenberg algebra or simply the rank one Heisenberg vertex operator algebra, or the one free boson vertex operator algebra.
Any element of $M(1)$ can be expressed as a linear combination of elements of the form $$h(-k_1).....h(-k_j){\bf 1}\text{ with }k_1\geq...\geq k_j\geq 1,\text{ for }j\in\mathbb{N}.$$
It is known that $M(1)$ is simple and has infinitely many non-isomorphic irreducible modules (cf. \cite{FLM2, Gu}). In addition, the indecomposable modules have been completely determined in \cite{Mil}.
\begin{thm} Let $A$ be a unital commutative associative algebra with the identity $1_A$. Let $B$ be a vertex $A$-algebroid that satisfies assumptions in either Theorem \ref{Bcyclicsolvable} when $\alpha_2\neq 0$ or Theorem \ref{3solvablecase1} when $\gamma_1+1\neq 0$ or Theorem \ref{3solvablecase2} when $\gamma_1+1\neq 0$ or Theorem \ref{3solvablecase3} when $\gamma_0+\gamma_1+1\neq 0$ hold. If $\mathfrak{a}'$ is the unique maximal ideal of the local algebra $A$ then the ideal $(\mathfrak{a}')$ is the maximal ideal of $V_B$. In addition, $V_B/(\mathfrak{a}')$ is isomorphic to the rank one Heisenberg vertex algebra $M(1)$.
\end{thm}
\begin{proof} Assume that $B$ is a vertex $A$-algebroid that satisfies assumptions in either Theorem \ref{Bcyclicsolvable} when $\alpha_2\neq 0$ or Theorem \ref{3solvablecase1} when $\gamma_1+1\neq 0$ or Theorem \ref{3solvablecase2} when $\gamma_1+1\neq 0$ or Theorem \ref{3solvablecase3} when $\gamma_0+\gamma_1+1\neq 0$ hold.
Let $\mathfrak{a}'$ be the unique maximal ideal of $A$. Recall that $A=\mathbb{C}1_A\oplus \mathfrak{a}'$. In addition, $B=\mathbb{C} b\oplus\partial(A)$. Let $\bar{b}\in\mathbb{C}^{\times} b$ such that $\bar{b}_1\bar{b}=1_A+a'$ where $a'\in \mathfrak{a}'$. Notice that $(V_B/(\mathfrak{a}'))_0=\mathbb{C}(1_A+(\mathfrak{a}')$ and for $n\geq 1$ $(V_B/(\mathfrak{a}'))_n=Span_{\mathbb{C}}\{\bar{b}(-n_1)...\bar{b}(-n_k){\bf 1}+(\mathfrak{a}')~|~n_1\geq ....\geq n_k\geq 1,~n_1+...+n_k=n\}$. Note ${\bf 1}=1_A$.
Let $\mathfrak{h}$ be a one-dimensional abelian Lie algebra spanned by $h$ with a linear form $\langle\cdot,\cdot\rangle$ such that $\langle h,h\rangle=1$. Let $f:M(1)\rightarrow V_B/(\mathfrak{a}')$ be a linear map defined by
\begin{eqnarray*}
f({\bf 1})&=&1_A+(\mathfrak{a}'),\\
f(h(-n_1)...h(-n_k){\bf 1})&=&\bar{b}(-n_1)...\bar{b}(-n_k)1_A+(\mathfrak{a}'),\text{ and }\\
fY(h(-1){\bf 1},x)&=&Y(f(h(-1){\bf 1}),x)f.
\end{eqnarray*}
Notice that we have $ fY({\bf 1},x)=Y(1_A,x)f$ and $f$ is onto. Now, we will show that $f$ is a vertex algebra homomorphism. We set
$C(f)=\{v\in M(1)~|~Y(f(v),x)f=fY(v,x)\}$. To show that $f$ is a vertex algebra homomorphism. It is enough to show that $C(f)=M(1)$. Clearly, ${\bf 1}, h(-1){\bf 1}\in C(f)$. Let $u,v\in C(f)$ and $n\in\mathbb{Z}$. Using the fact that $Y(u_nv,x)={\rm Res}\,_{x_1}((x_1-x)^nY(u,x_1)Y(v,x)-(-x+x_1)^nY(v,x)Y(u,x_1))$, we then have that
\begin{eqnarray*}
&&fY(u_nv,x)\\
&=&{\rm Res}\,_{x_1}((x_1-x)^nfY(u,x_1)Y(v,x)-(-x+x_1)^nfY(v,x)Y(u,x_1))\\
&=&{\rm Res}\,_{x_1}((x_1-x)^nY(f(u),x_1)fY(v,x)-(-x+x_1)^nY(f(v),x)fY(u,x_1))\\
&=&{\rm Res}\,_{x_1}((x_1-x)^nY(f(u),x_1)Y(f(v),x)f-(-x+x_1)^nY(f(v),x)Y(f(u),x_1)f)\\
&=&Y(f(u)_nf(v),x)f.
\end{eqnarray*}
Hence, $u_nv\in C(f)$. In addition, $C(f)=M(1)$ and $f$ is a homomorphism. Since $M(1)$ is simple, we can conclude further that $f$ is an isomorphism and $(\mathfrak{a}')$ is a maximal ideal of $V_B$.
\end{proof}
|
\section{Introduction}
\label{sec:intro}
Among the most commonly used slicing conditions in numerical relativity is the {\em Bona-Mass\'o condition}
\begin{equation} \label{bona_masso}
(\partial_t - \beta^i \partial_i) \alpha = - \alpha^2 f(\alpha) K,
\end{equation}
where $\alpha$ is the lapse function, $\beta^i$ the shift vector, $K$ the mean curvature (i.e.~the trace of the extrinsic curvature), and the {\em Bona-Mass\'o function} $f(\alpha)$ is a function of the lapse that has yet to be determined (see \cite{BonMSS95}). The properties of the resulting slices depend, of course, on the choice for $f(\alpha)$; for $f(\alpha) = 1$, for example, the slicing condition (\ref{bona_masso}) is equivalent to the lapse condition in harmonic coordinates.
A particularly successful choice for the Bona-Mass\'o function is
\begin{equation} \label{1+log}
f(\alpha) = \frac{2}{\alpha},
\end{equation}
especially for simulations of black-hole spacetimes. In the absence of a shift vector, Eq.~(\ref{bona_masso}) can then be integrated to yield $\alpha = 1 + \log(\gamma)$, where $\gamma$ is the determinant of the spatial metric, which lends this slicing condition its name {\em 1+log slicing} (see \cite{Alc08,BonPB09,BauS10,Gou12,Shi16,BauS21} for textbook discussions). Dynamical simulations with 1+log slicing render black holes in a trumpet geometry, which, in the static limit, have been analyzed by a number of different authors \cite{HanHPBM07,HanHOBO08,Bru09}. These studies, together with those of similar trumpet geometries (e.g., \cite{HanHBGSO07,BauN07,DenB14,BaudeO22}) have helped explain the remarkable numerical properties of these slicing conditions, especially in the context of black-hole simulations.
Even in the context of vacuum evolution calculations, however, 1+log slicing is known to lead to coordinate shocks in some circumstances (see \cite{Alc97,AlcM98,Alc05}). Alcubierre \cite{Alc97,Alc03} therefore suggested an alternative {\em shock-avoiding} Bona-Mass\'o slicing condition with
\begin{equation} \label{shock}
f(\alpha) = 1 + \frac{\kappa}{\alpha^2},
\end{equation}
where $\kappa > 0$ is a constant. While this condition has indeed been found to avoid some coordinate pathologies that arise in 1+log slicing, it also has some very unusual properties---in particular, it allows the lapse to become negative (see the discussion in \cite{Alc03}, as well as Fig.~\ref{fig:lapse} below for an example), which may explain why it has been adopted only rarely (see, e.g., \cite{JimVA21}).
Despite the appearance of negative values for the lapse, shock-avoiding slicing has recently been shown to perform very similarly to 1+log slicing in terms of stability and accuracy for a number of test calculations involving black holes, neutron stars, and gravitational collapse (see \cite{BauH22}). One of these tests considered a Schwarzschild black hole initially represented on a slice of constant Schwarzschild time, i.e.~in a wormhole geometry. These data are then evolved with the Bona-Mass\'o slicing condition (\ref{bona_masso}), which results in a coordinate transition to a trumpet geometry. In Fig.~\ref{fig:lapse} we reproduce results from this test and show the values of the lapse at the black-hole puncture, i.e.~at the center of the isotropic coordinate system.
Evidently, the behavior of the lapse at the black-hole puncture for 1+log versus shock-avoiding slices shows not only quantitative but also qualitative differences. For 1+log slices the lapse appears to decay approximately exponentially after a brief dynamical period, while, for shock-avoiding slices, the lapse appears to perform harmonic oscillations. At early times these oscillations appear to be damped, but at later times the amplitude remains approximately constant. We also observe that the period of the oscillations is larger for a smaller value of the constant $\kappa$ in (\ref{shock}).
We caution that neither the exponential decay nor the harmonic oscillation is exact. We also note that, because of the lack of differentiability at the center of the black hole, numerical error arising from finite-differencing across the black-hole puncture is large and prevents pointwise convergence. Using a completely independent code based on a multi-domain spectral method (see \cite{deO22}) we found some quantitative differences resulting from the different treatment of the puncture, but the same qualitative behavior as with the finite-difference code: exponential decay, typically associated with a first-order ordinary differential equation, for 1+log slicing, versus harmonic oscillation, pointing to a second-order equation, for shock-avoiding slicing. Since both slicing conditions are imposed by the same equation, the Bona-Mass\'o condition (\ref{bona_masso}), the origin of this qualitatively different behavior is, a priori, not clear at all.
Our goal in this paper is to gain analytical insight into what causes these qualitative differences. We employ a dynamical height-function approach to describe time-dependent coordinate transformations of Schwarzschild black holes, and to explore the behavior of the lapse at the black-hole puncture. We introduce this formalism in Sec.~\ref{sec:height_function}, and review results for static slices in Sec.~\ref{sec:static}. In Sec.~\ref{sec:perturbation} we then consider dynamical slices in the limit that they can be considered small perturbations of static slices. At large distances from the black hole, the Bona-Mass\'o condition (\ref{bona_masso}) results in well-known wave equations for the lapse, as expected. At the black-hole puncture, however, the resulting equation depends on whether or not the (unperturbed) mean curvature $K$ vanishes at the puncture. Typically, including for 1+log slicing, $K$ is positive at the puncture, in which case one obtains exponential damping. Shock-avoiding slices, however, form an exception in that $K$ vanishes at the puncture, in which case one obtains harmonic oscillation. We briefly summarize in Sec.~\ref{sec:summary}, concluding that the vanishing of $K$ at the black-hole puncture results in the qualitative differences observed.
\section{Dynamical height functions}
\label{sec:height_function}
\begin{figure}
\centering
\includegraphics[width = 0.45 \textwidth]{Fig_1.png}
\caption{The lapse $\alpha$ at the black-hole puncture in the evolution of a single black hole with the Bona-Mass\'o slicing condition (\ref{bona_masso}). All simulations start with wormhole initial data and transition to a trumpet geometry determined by the choice of the Bona-Mass\'o function $f(\alpha)$. Note the qualitatively different behavior of the lapse for different functions $f(\alpha)$ after the initial perturbation: for 1+log slicing with (\ref{1+log}), the lapse appears to decay approximately exponentially, while for shock-avoiding slicing conditions (\ref{shock}) it appears to perform harmonic oscillations, with a period that appears to depend on $\kappa$. (Figure adopted from Fig.~2 of \cite{BauH22}, to where the reader is referred for numerical details.)}
\label{fig:lapse}
\end{figure}
We start with the Schwarzschild line element in Schwarzschild coordinates,\footnote{We adopt geometrized units with $G = 1 = c$ unless noted otherwise.}
\begin{equation} \label{lineelement1}
ds^2 = - F \dd{\bar{t}}^2 + F^{-1} \dd{R}^2 + R^2 \dd{\Omega}^2,
\end{equation}
where $R$ is the areal radius, $F = F(R) = 1 - 2M/R$, and $M$ is the black-hole mass.\footnote{We focus on uncharged Schwarzschild black holes here, but note that our calculation generalizes to charged Reissner-Nordstr\"om black holes simply by letting $F = 1 - 2M/R + Q^2 / R^2$, where $Q$ is the black-hole charge.} We then transform to a new time coordinate\footnote{Unlike in \cite{LiBDdeO22}, where we denoted the Schwarzschild time as $t$ and the new time coordinate as $\bar{t}$, we here adopt the opposite convention in order to reduce notational clutter for dynamical slices.} $t$ by introducing a {\em height-function} $h(t,R)$ that measures how far the new time slices lift off the old time slices,
\begin{equation} \label{heightfunction}
t = \bar{t} + h(t,R)
\end{equation}
(see, e.g.,~\cite{rei73,BeiO98,MalO03,GarGH08}, as well as \cite{BauS10} for a textbook treatment). Unlike in many previous applications, we allow the height-function to depend on time in order to study dynamical coordinate transitions. Inserting (\ref{heightfunction}) into the line element (\ref{lineelement1}) results in
\begin{align} \label{lineelement2}
ds^2 = & - F\qty(1 - \dot h)^2 dt^2 + 2 F \qty(1 - \dot h) h' dt dR \nonumber \\
& + \qty(F^{-1} - F h'^2) dR^2 + R^2 d\Omega^2,
\end{align}
where the dot denotes a partial derivative with respect to time and a prime with respect to areal radius $R$. From (\ref{lineelement2}) we can identify the $RR$-component of the spatial metric $\gamma_{ij}$ as
\begin{subequations} \label{metric}
\begin{equation} \label{gamma1}
\gamma_{RR} = F^{-1} (1 - F^2 h'^2),
\end{equation}
the lapse function $\alpha$ as
\begin{equation} \label{alpha1}
\alpha^2 = \frac{F(1 - \dot h)^2}{1 - F^2 h'^2},
\end{equation}
and the $R$-component of the shift vector $\beta^R$ as
\begin{equation} \label{beta1}
\beta^R = \frac{F^2 (1 - \dot h) h'}{1 - F^2 h'^2}.
\end{equation}
\end{subequations}
Finally we compute the mean curvature from
\begin{equation} \label{K1}
K = - \nabla_a n^a = - |g|^{-1/2} \partial_a \qty(|g|^{1/2} n^a),
\end{equation}
where $\nabla_a$ is the covariant derivative associated with the spacetime metric, $n^a$ the future-oriented normal to the spatial hypersurface, $n^a = \alpha^{-1}(1, - \beta^i)$, and $g$ the determinant of the spacetime metric, $g = - \alpha^2 \gamma_{RR} R^4 \sin^2 \theta$.
While the height-function approach has been adopted to study the Schwarzschild spacetime in many different coordinate systems, we focus here on transformations to trumpet geometries that satisfy the Bona-Mass\'o slicing condition (\ref{bona_masso}).
\section{Static slices}
\label{sec:static}
The construction of static trumpet geometries using a time-independent height function $h = h_0(R)$ has been discussed by a number of authors (see, e.g., \cite{HanHBGSO07,HanHPBM07,BauN07,HanHOBO08,Bru09,DenB14,BaudeO22,LiBDdeO22}), and we therefore review only some important results that are relevant for our discussion in the following sections.
For static slices, the Bona-Mass\'o condition (\ref{bona_masso}) is
\begin{equation} \label{bona_masso_static}
\beta^i \partial_i \alpha = \alpha^2 f(\alpha) K,
\end{equation}
the expression (\ref{gamma1}) for the $RR$-component of the spatial metric remains unchanged, Eq.~(\ref{alpha1}) for the lapse reduces to\footnote{We note a typo in the corresponding Eq.~(7) of \cite{BaudeO22}, where the factor $F$, denoted $f_0$ there, should be squared in the denominator.}
\begin{subequations} \label{metric_static}
\begin{equation} \label{alpha_static}
\alpha^2 = \frac{F}{1 - F^2 h_0'^2} = \gamma_{RR}^{-1},
\end{equation}
and Eq.~(\ref{beta1}) for the shift becomes
\begin{equation} \label{beta_static}
\beta^R = \frac{F^2 h_0'}{1 - F^2 h_0'^2} = \alpha \sqrt{\alpha^2 - F}.
\end{equation}
\end{subequations}
Inserting the above expressions together with (\ref{K1}) into (\ref{bona_masso_static}) then yields an ordinary differential equation that, for many choices of the Bona-Mass\'o function $f(\alpha)$, can be integrated in closed form. A constant of integration can be determined by imposing regularity across a singular point, making the solution unique. For some choices of $f(\alpha)$, this solution can be expressed as an explicit function $\alpha = \alpha(R)$, but for others the solution can be written in implicit form for $\alpha$ only.
In either case we may find the location $R_0$ of the root of the lapse, $\alpha(R_0) = 0$, which must be inside the horizon, i.e.~$R_0 < 2M$, for horizon-penetrating slices. Defining
\begin{equation} \label{a1}
a_1 \equiv \eval{\left( \frac{d\alpha}{dR} \right)}_{R = R_0}
\end{equation}
we see from (\ref{alpha_static}) that $\gamma_{RR} \simeq a_1^{-2} (R - R_0)^{-2}$ close to the root of the lapse. Assuming that $a_1$ is positive and finite, we may integrate $ds = \gamma_{RR}^{1/2} dR$ to find that the root is located at an infinite proper distance from all points $R > R_0$. We therefore refer to this location as the {\em puncture} and note that the scaling of $\gamma_{RR}$ with $(R - R_0)^{-2}$ is characteristic of a {\em trumpet geometry} (see \cite{HanHBGSO07,HanHOBO08}). In the vicinity of the puncture, the height function diverges according to
\begin{equation} \label{h_prime}
h_0' \simeq - \frac{1}{\sqrt{-F(R_0)} \, a_1 (R - R_0)},
\end{equation}
where we have adopted a negative sign in taking a square root (note also that $F(R_0) < 0$ since $R_0 < 2M$). Even for the time-dependent slices in the following sections, we will identify the puncture with a divergence of the metric component $\gamma_{RR}$, which, according to (\ref{gamma1}), coincides with a divergence of the height function $h$ at $R < 2M$. In terms of an isotropic radius $r$, which is typically employed in numerical simulations, the puncture corresponds to the origin $r = 0$. For static slices the divergence of $h_0$ automatically coincides with a root of the lapse, but this need not be the case for time-dependent slices (see Sec.~\ref{sec:perturbation} below). Also note that, for horizon-penetrating slices, $\alpha$ is nonzero and finite on the horizon, where $F = 0$, so that (\ref{alpha_static}) indicates that the height-function necessarily has to diverge there, too.\footnote{This divergence could have been avoided by starting with a horizon-penetrating coordinate system in (\ref{lineelement1}), rather than with Schwarzschild coordinates.} Finally we note that we can compute the static mean curvature at the puncture from
\begin{equation} \label{K_static}
K_0(R_0) = \frac{\beta^R \partial_R \alpha}{\alpha^2 f(\alpha)} = \frac{\sqrt{- F(R_0)} \, a_1}{\alpha f({\alpha)}},
\end{equation}
where we have used (\ref{beta_static}) and (\ref{a1}) in the second equality. Evidently, whether or not $K_0(R_0)$ is finite depends on the behavior of $\alpha f(\alpha)$ as $\alpha \rightarrow 0$, which we have not yet evaluated in (\ref{K_static}).
\begin{table}[t]
\centering
\begin{tabular}{c|c|c|c|c|c}
$f(\alpha)$ & Ref. & $R_0 / M$ & $F(R_0)$
& $K_0(R_0) M$ & $K'(R_0)M^2$ \\
\hline \hline
$2/\alpha$ & \cite{BonMSS95} & 1.312 & $-0.524$
& 0.301 & $-1.730$ \\
$(1 - \alpha)/\alpha$ & \cite{DenB14} & 1 & $-1$
& 1 & $-3$ \\
$1 + \kappa / \alpha^2$ & \cite{Alc97} & 3/2 & $-1/3$ & 0 & $-8\sqrt{3}/9$
\end{tabular}
\caption{Summary of properties of static slices for three different choices of the Bona-Mass\'o function $f(\alpha)$. For each $f(\alpha)$ we list the areal radius of the black-hole puncture $R_0$ (which, for static slices, coincides with a root of the lapse function $\alpha$), as well as the values of $F$,
the mean curvature $K_0$, and the derivative $d K / dR$ as computed from (\ref{K4}), all evaluated at the puncture.}
\label{tab:static}
\end{table}
After this general discussion, we consider some examples for specific choices of the Bona-Mass\'o function $f(\alpha)$ and summarize the key results for these static slices in Table \ref{tab:static}.
By far the most common choice for $f(\alpha)$ is (\ref{1+log}), which leads to {\em 1+log} slices \cite{BonMSS95}. For 1+log slicing the integral for the lapse $\alpha$ cannot be solved for $\alpha$ directly, so that the resulting equations are usually solved numerically. In particular, this yields $R_0 \simeq 1.312 M$ for the root of the lapse and $a_1 \simeq 0.832 M^{-1}$ for the derivative of the lapse at the root (see \cite{HanHBGSO07,HanHOBO08,Bru09}). Finally, we use $\alpha f(\alpha) = 2$ in (\ref{K_static}) to find $K_0(R_0) = 0.301 M^{-1}$ for the mean curvature at the puncture.
\begin{figure}
\centering
\includegraphics[width = 0.45 \textwidth]{Fig_2.png}
\caption{Graph of the static height function $h_0(R)$, Eq.~(\ref{analytical_h}), for the analytical trumpet slice (solid line), together with a hypothetical perturbation (dashed line). As discussed in Section \ref{sec:perturbation}, these perturbations can be described by changes $\eta$ in $h$ far from the puncture, and changes $\rho$ in the radius $R$ close to the puncture at $R_0 = M$.}
\label{fig:h}
\end{figure}
As a second example we consider the choice
\begin{equation} \label{analytical_f}
f(\alpha) = \frac{1 - \alpha}{\alpha},
\end{equation}
which results in a completely {\em analytical trumpet slicing} of the Schwarzschild spacetime (see \cite{DenB14}). In this case the integral for the lapse can be solved explicitly, yielding
\begin{equation} \label{analytical_alpha}
\alpha = \frac{R - M}{R}.
\end{equation}
We evidently have $R_0 = M$, from which we compute $a_1 = M^{-1}$ and, using $\alpha f(\alpha) = 1 - \alpha$, the mean curvature at the puncture, $K_0(R_0) = M^{-1}$, in agreement with Eq.~(17) in \cite{DenB14}. Because of the simplicity of the above expressions, it is also straightforward to insert (\ref{analytical_alpha}) into (\ref{alpha_static}), solve for $h_0'$ and integrate to obtain
\begin{equation} \label{analytical_h}
h_0(R) = M \ln \frac{2 (R - 2M)^2}{(R - M) M}~~~~~~~~~~~~~~(R > M),
\end{equation}
where we have arbitrarily chosen a constant of integration so that $h_0 = 0$ at $R = 3M$. Note that $h_0$ diverges logarithmically both at the puncture $R_0 = M$ as well as the horizon $R = 2M$, as expected from our discussion above. In the vicinity of the puncture at $R = M$ we have $h_0' \simeq - M/(R - M)$, in agreement with (\ref{h_prime}). We graph (\ref{analytical_h}) in Fig.~\ref{fig:h} together with a hypothetical time-dependent perturbation as considered in Sec.~\ref{sec:perturbation}.
We next consider the {\em shock-avoiding} slicing condition with $f(\alpha)$ given by (\ref{shock}). In this case the lapse function $\alpha$ can again be expressed explicitly in terms of $R$,
\begin{equation}
\alpha = \left( \frac{R^4 - 2MR^3 + C}{R^4 - C/\kappa} \right)^{1/2},
\end{equation}
where the constant of integration $C$ is given by $C = 3^3 M^4/2^4$ (see \cite{BaudeO22}). The puncture is now located at $R_0 = 3 M/2$, independently of $\kappa$, and we can determine $a_1 = 2 \,(6\kappa /(3 \kappa - 1))^{1/2}/(3 M)$. In a key difference from the other slicing conditions, however, we now observe that $\alpha f(\alpha) = \alpha + \kappa / \alpha$ diverges at the puncture, so that the mean curvature (\ref{K_static}) vanishes there, $K_0(R_0) = 0$.
\section{Dynamical slices: perturbative treatments}
\label{sec:perturbation}
We now consider dynamical slices in the limit that they may be considered linear perturbations of the static slices of Sec.~\ref{sec:static}. Inspecting Fig.~\ref{fig:h} we note that these perturbations need to be described in different ways in different regimes. Far from the black hole, where the slope of $h$ becomes increasingly small, the perturbed slice can be described in terms of a perturbation $\eta$ of the height function $h_0$ itself, so that $h(t,R) = h_0(R) + \eta(t,R)$. We will briefly discuss this approach in Sec.~\ref{subsec:minkowski}, recovering well-known wave equations for the lapse function $\alpha$. Close to the puncture, however, $h_0$ and its derivative diverge, so that changes in $h$ may also diverge. In this region it is more natural to describe the perturbation in terms of a small shift $\rho$ in the radius $R$. As we will show in Sec.~\ref{subsec:puncture}, this approach will yield our main result concerning the dynamical behavior of different slicing conditions at the black-hole puncture.
\subsection{Perturbations in the far limit}
\label{subsec:minkowski}
We first consider perturbations to the height function $h$ in the far limit $R \gg M$, where we assume $h_0' \rightarrow 0$. As discussed above we describe the perturbation as
\begin{equation} \label{perturb_far}
h(t,R) = h_0(R) + \eta(t,R)
\end{equation}
in this regime. Given our assumption $h_0' \ll 1$ we have $\dot h = \dot \eta$ and $h' \simeq \eta'$. Since we also have $F \simeq 1$ in this limit, we obtain, to leading order in $\eta$,
\begin{equation}
\alpha \simeq 1 - \dot \eta,~~~~~\beta^R \simeq \eta',~~~~~K \simeq \eta'' +\frac{2}{R}\eta'=\nabla^{2}\eta.
\end{equation}
Inserting these expressions into the Bona-Mass\'o condition (\ref{bona_masso}) we obtain the wave equation
\begin{equation}
- \ddot \eta + f(1) \, \nabla^{2}\eta \simeq 0,
\end{equation}
where $f(1)$ denotes the Bona-Mass\'o function $f(\alpha)$ evaluated for $\alpha = 1$. We may now take a time derivative of this equation to see that, in this limit, the lapse function $\alpha$ satisfies a wave equation, and that perturbations in the lapse travel at speeds $v = \sqrt{f(1)} \, c$, where we have inserted the speed of light $c$ for clarity. For 1+log slicing with (\ref{1+log}) we have $f(1) = 2$, confirming the well-known result that gauge modes travel at a speed $v = \sqrt{2} \, c$, while for shock-avoiding slices with (\ref{shock}), gauge modes travel at a speed $v = \sqrt{1 + \kappa} \, c$ (see, e.g., \cite{Alc97,Alc03}).\footnote{Recall that the above waves describe pure gauge modes, so that a wave speed $v$ exceeding the speed of light $c$ does not violate causality.}
\subsection{Perturbations at the puncture}
\label{subsec:puncture}
We now turn to perturbations close to the puncture. In this regime, where the static height function and its derivative diverge, a perturbative ansatz like (\ref{perturb_far}) cannot describe a perturbation like the one sketched in Fig.~\ref{fig:h}, i.e.~one that shifts the puncture to a different radius, with finite $\eta$. Instead, we describe perturbations in the vicinity of the puncture in terms of a perturbation $\rho = \rho(t,R)$ of the radius $R$. Specifically, we will equate the (perturbed) height function $h(t,R)$ with the static height function $h_0$ at a radius
\begin{equation}
\bar R = R + \rho(t,R)
\end{equation}
as sketched in Fig.~\ref{fig:h}, with $\rho$ defined by
\begin{equation}
h(t,R) = h_0\qty(\bar R) = h_0\qty(R + \rho).
\end{equation}
Derivatives of $h$ are then given by
\begin{subequations} \label{derivs_h}
\begin{align}
\dot h(t,R) & = h_0'\qty(\bar R) \, \dot \rho(t,R) \qq{and} \\
h'(t,R) & = h_0'\qty(\bar R) \, \left[ 1 + \rho'(t,R) \right].
\end{align}
\end{subequations}
Inserting (\ref{derivs_h}) into (\ref{metric}) we obtain
\begin{subequations} \label{metric2}
\begin{equation} \label{gamma2}
\gamma_{RR} = F^{-1} \left[ 1 - F^2 h_0'^2 (1 + \rho')^2 \right]
\end{equation}
for the radial metric component,
\begin{equation} \label{alpha2}
\alpha^2 = \frac{F \qty(1 - h_0' \dot \rho)^2}{1 - F^2 h_0'^2 \qty(1 + \rho')^2}
\end{equation}
for the lapse, and
\begin{equation} \label{beta2}
\beta^R = \frac{F^2 \qty(1 - h_0' \dot \rho) h_0' \qty(1 + \rho')}{1 - F^2 h_0'^2\qty(1 + \rho')^2}
\end{equation}
\end{subequations}
for the radial shift component.
We observe from (\ref{gamma2}) that $\gamma_{RR}$ diverges when $h_0'(\bar R)$ diverges. As in the static case, we may therefore identify the puncture with a divergence of $h_0'$ at $\bar R = R_0$, except that it is now located at $R = R_0 - \rho$ (as suggested in the sketch of Fig.~\ref{fig:h}). In the following we will assume that $\gamma_{RR}$ diverges at the puncture at the same rate as in the static case, so that a trumpet geometry is maintained even for small perturbations. According to (\ref{metric2}) this implies that $\rho'$ remains finite as the puncture is approached.
Evaluating the lapse (\ref{alpha2}) at the puncture, we obtain
\begin{subequations} \label{tempalphabeta}
\begin{equation} \label{alpha3a}
\alpha = (-F)^{-1/2} \frac{\dot \rho}{1 + \rho'},
\end{equation}
while the shift (\ref{beta2}) becomes
\begin{equation} \label{beta3a}
\beta^R = \frac{\dot \rho}{1 + \rho'}.
\end{equation}
\end{subequations}
Requiring that both lapse and shift are finite at the puncture, and having assumed that $\rho'$ is finite there, we see that we may assume that $\dot \rho$ remains finite at the puncture as well. Defining the derivative along the normal vector $n^a$ as
\begin{equation} \label{partial_n}
\partial_n \equiv \alpha \, n^a \partial_a = \partial_t - \beta^R \partial_R,
\end{equation}
we may rewrite Eqs.~(\ref{tempalphabeta}) in the more compact form
\begin{equation} \label{alphabeta}
\alpha = (-F)^{-1/2} \, \partial_n \rho \qq{and} \beta^R = \partial_n \rho.
\end{equation}
Unlike in the static case (see Eq.~\ref{alpha_static}), the lapse function does {\em not} necessarily vanish at the puncture for time-dependent slices, as has been observed in numerous numerical simulations (see Fig.~\ref{fig:lapse} for an example).
We next evaluate the Bona-Mass\'o condition (\ref{bona_masso}) at the puncture. On the left-hand side we use the definition (\ref{partial_n}) together with (\ref{alphabeta}) to obtain
\begin{equation} \label{dn_alpha}
(\partial_t - \beta^R \partial_R) \alpha = \partial_n \left[(-F)^{-1/2} \, \partial_n \rho \right],
\end{equation}
while on the right-hand side we insert the expressions (\ref{metric2}) into the mean curvature (\ref{K1}) and evaluate the result at the puncture, where $h_0' \rightarrow \infty$ and $\bar R = R_0$. Many terms, including $\rho' / (h_0')^2$ and $\rho'' / (h_0')^2$, disappear at the puncture by virtue of the scaling (\ref{h_prime}) for $h_0'$ and our assumption that $\rho'$ remains finite there, resulting in
\begin{equation} \label{K3}
K = - \frac{4 F + R F'}{2 R (-F)^{1/2}}.
\end{equation}
In both (\ref{dn_alpha}) and (\ref{K3}) the function $F$ and its derivative are evaluated at $R = R_0 - \rho$. For $\rho \ll R_0$ we may expand the above expressions about $\bar R = R_0$ and rewrite (\ref{K3}) as
\begin{equation} \label{K4}
K(R) \simeq K(R_0) - \rho K'(R_0),
\end{equation}
where $K' = dK / dR$. As expected, evaluating $K(R_0)$ from (\ref{K3}) yields the values listed in Table \ref{tab:static}, where we also list the values of $K'(R_0)$ for the different slicing conditions. We now insert (\ref{dn_alpha}) and (\ref{K4}) into the Bona-Mass\'o condition (\ref{bona_masso}) and obtain
\begin{equation} \label{bona_masso2}
\partial_n \left[(-F)^{-1/2} \, \partial_n \rho \right] = - \alpha^2 f(\alpha) \left[ K(R_0) - \rho K'(R_0) \right],
\end{equation}
where we have not yet evaluated the term $\alpha^2 f(\alpha)$. Remarkably, we have obtained an {\em ordinary} differential equation for $\rho$ at the puncture. As we will explore in the next two subsections, even the qualitative behavior of solutions to this equation depends on the choice of $f(\alpha)$ and hence $K(R_0)$, because it determines whether (\ref{bona_masso2}) acts as a first- or a second-order equation.
\subsubsection{1+log slicing}
\label{subsubsec:1+log}
For almost all slicing conditions, the leading-order mean curvature term $K(R_0)$ on the right-hand side of (\ref{bona_masso2}) is nonzero. One such slicing condition is 1+log slicing with $f(\alpha) = 2 / \alpha$ (see \ref{1+log}), for which (\ref{bona_masso2}) becomes
\begin{equation} \label{bm_1+log}
\partial_n \left[(-F)^{-1/2} \, \partial_n \rho \right] = 2 (-F)^{-1/2} \, \partial_n \rho \left[ K(R_0) - \rho K'(R_0) \right]
\end{equation}
after inserting (\ref{alphabeta}) for $\alpha$ on the right-hand side. We now observe that, to leading order in $\rho$, the term $\rho K'(R_0)$ on the right-hand side disappears and with it the only appearance of $\rho$ itself (rather than its derivatives).
To linear order in $\rho$, we may therefore replace the term $(-F)^{-1/2} \, \partial_n \rho$ with $\alpha$ to obtain a first-order equation for the lapse alone,
\begin{equation}
\partial_n \alpha = - 2 \alpha K(R_0).
\end{equation}
This equation is identical to our starting point (\ref{bona_masso}), of course, except that now, in the linear limit, we assume the mean curvature $K$ to be given by a positive and constant value. In this case we may integrate to obtain
\begin{equation} \label{lapse_1+log}
\alpha = C e^{-2 K(R_0) t}
\end{equation}
where $C$ is a constant of integration, demonstrating that, to linear order, we should expect the lapse function at the puncture to decay exponentially for 1+log slicing.
As one might expect from the discussion in Sec.~\ref{sec:intro}, a quantitative comparison of (\ref{lapse_1+log}) with the numerical data of Fig.~\ref{fig:lapse} shows some differences. During the time around $10 M \lesssim t \lesssim 15 M$, when Fig.~\ref{fig:lapse} suggests an approximately exponential decay, the numerical data fall off more rapidly than predicted by (\ref{lapse_1+log}). A rough fit to the numerical data suggests a time constant $\tau$ that is smaller than $[2 K(R_0)]^{-1}$ by about a factor of two. However, rather than being constant, $K$ also changes rapidly during this period, as it transitions from its initial value of zero to the equilibrium value of $K(R_0) \simeq 0.301$, indicating that nonlinear terms are still important during this time. At later times numerical error becomes important; in particular, the lapse settles down to a small but nonzero value (that depends on the numerical resolution), so that exponential decay can no longer be observed.
\subsubsection{Shock-avoiding slices}
\label{subsubsec:shockavoiding}
For shock-avoiding slices the unperturbed puncture is located at $R_0 = 3 M/2$ so that the mean curvature $K(R_0)$ on the right-hand side of (\ref{bona_masso2}) vanishes. Inserting $f(\alpha) = 1 + \kappa / \alpha^2$ [see (\ref{shock})] into (\ref{bona_masso2}) we now obtain
\begin{equation} \label{bm_sa}
\partial_n \left[(-F)^{-1/2} \, \partial_n \rho \right] = \kappa K'(R_0) \rho,
\end{equation}
where we have already neglected a term quadratic in $\alpha$ on the right-hand side. In contrast to 1+log slicing, the term proportional to $\rho$ now dominates the right-hand side, so we cannot rewrite this second-order equation as a first-order equation for $\alpha$. We instead expand to linear order in $\rho$ to obtain the harmonic-oscillator equation
\begin{equation} \label{bm_sa2}
\partial^2_n \rho = - \omega^2 \rho,
\end{equation}
with the angular frequency $\omega$ given by
\begin{equation}
\omega^2 = - (-F)^{1/2} \kappa K'(R_0) = \frac{8 \kappa}{9 M^2}.
\end{equation}
Accordingly, $\rho$, and hence $\alpha$, performs harmonic oscillations with period
\begin{equation} \label{period}
P = \frac{3 \pi M}{\sqrt{2 \kappa}}.
\end{equation}
Note that we have assumed $\kappa > 0$ in the above arguments, in accordance with our original assumption in (\ref{shock}) (see also \cite{Alc97}).
As in Sec.~\ref{subsubsec:1+log} we attempt a quantitative comparison with the numerical data with some caution. Measuring the period of the oscillations observed for the shock-avoiding slices around $130 M \lesssim t \lesssim 170 M$, we find $P_1 \simeq 8 M$ for $\kappa = 1$ and $P_{2/3} \simeq 11 M$ for $\kappa = 2/3$ (even though the latter, in particular, shows some variation). Evaluating (\ref{period}), on the other hand, we obtain $P_1 \simeq 6.7 M$ and $P_{2/3} \simeq 8.2 M$. While we again do not find accurate quantitative agreement, we see that our analysis does explain the origin of the observed harmonic oscillation and correctly predicts that the period increases with decreasing $\kappa$.
We suspect that nonlinear terms cause the damping of the oscillations at early times, as seen in Fig.~\ref{fig:lapse}. Once the amplitude is sufficiently small, however, the oscillations should be governed by (\ref{bm_sa2}), which does not have a damping term. Accordingly, one would expect these oscillations to persist at a small amplitude, which is consistent with the numerical results.
\section{Summary}
\label{sec:summary}
Motivated by recent numerical experiments with shock-avoiding slicing conditions as alternatives to 1+log slicing, we explore the origins of a qualitative difference observed in these simulations: while, for the latter, the lapse function at the black-hole puncture appears to decay in an approximately exponential fashion, the former leads to approximately harmonic oscillations in the lapse. We apply a dynamical height-function approach to Schwarzschild black holes to describe time-dependent coordinate transitions, impose the Bona-Mass\'o condition (\ref{bona_masso}) with different choices for the function $f(\alpha)$, evaluate the resulting equation at the black-hole puncture, and finally consider small perturbations of a background trumpet solution.
Describing these perturbations in terms of the displacement $\rho$ of the location of the puncture, the Bona-Mass\'o equation becomes a second-order equation for $\rho$ (see \ref{bona_masso2}). The key difference between 1+log slices and shock-avoiding slices then arises from the behavior of the (unperturbed) mean curvature $K(R_0)$ at the location of the puncture. For 1+log slicing, $K(R_0)$ takes a nonzero, positive value, in which case $\rho$ itself drops out of the equation to linear order, resulting in a {\em first-order} equation for $\partial_n \rho$ that governs exponential decay. For shock-avoiding slices, on the other hand, $K(R_0)$ vanishes, and the right-hand side of (\ref{bona_masso2}) ends up being dominated by $\rho$ at linear order. The equation therefore remains a {\em second-order} equation for $\rho$, resulting in harmonic oscillation. We further observe that the period of the oscillations depends on the constant $\kappa$ in (\ref{shock}), with larger $\kappa$ resulting in a shorter period.
While a quantitative comparison of our analytical results with the numerical findings of \cite{BauH22} shows some discrepancies as discussed in Sect.~\ref{subsec:puncture}, we believe that these can be explained in terms of nonlinear effects and numerical error resulting from the lack of differentiability at the black-hole puncture. It is also possible, of course, that our assumption concerning the derivative of $\rho$ at the puncture is overly restrictive. Despite these discrepancies, our findings provide analytical insight into the dynamical behavior of the lapse at the black-hole puncture, point to the origin of qualitative differences between different slicing conditions, and predict the dependence of decay constants and oscillation periods on the given parameters.
\begin{acknowledgements}
It is a pleasure to thank David Hilditch for helpful conversations as well as comments on a previous version of this manuscript. SEL acknowledges support through an undergraduate research fellowship at Bowdoin College.
This work was supported in part by National Science Foundation (NSF) grant PHY-2010394 to Bowdoin College and the Coordena\c c\~ao de Aperfei\c coamento de Pessoal de N\'ivel Superior - Brasil (CAPES) - Finance Code 001.
\end{acknowledgements}
|
\section{\label{sec:level1}Introduction}
Density functional theory (DFT) is an important constituent of the modern approach to quantum many-body physics \cite{parr1980density}. Consider systems with electrons under an external potential $V$, with the Hamiltonian
\begin{equation}\label{eq:DFT}
\hat H = \hat H_0 + \sum_{i} V_i \hat n_i,
\end{equation}
where $\hat H_0$ is the Hamiltonian involving kinetic and pair interaction terms, and $\hat n_i$ and $V_i$ are the density operator and external potential at site $i$. Two celebrated Hohenberg and Kohn's (HK's) theorems \cite{hohenberg1964inhomogeneous} state that the system energy satisfies
\begin{equation}\label{eq:HK}
E[n] = F[n] + \sum_i V_i n_i,
\end{equation}
as a unique functional of particle density $n_i \equiv \langle \hat n_i \rangle$, where $F[n]$ is the HK functional of the external potential. Furthermore, the energy functional $E[n]$ is bounded and uniquely saturated by the ground state. The HK functional $F[n]$ depends only on $\hat H_0$, and thus is universal and shared by any electron systems.
HK's original proof was non-constructive. Therefore, the exact form of the density functional $F[n]$ is unknown. Despite computational advances over more than half a century, the search for the exact functional is largely limited. A numerical algorithm for finding exact density functional was initially proposed by Levy in 1979 \cite{levy1979universal,levy1982electron} and later by Lieb \cite{lieb1983density}. However, the Levy-Lieb constraint search algorithm does not provide an analytical solution to the exact density functional. In fact, a common belief in the community is that the explicit construction of HK functional will never be found.
The bounded nature of Eq.~(\ref{eq:HK}) reminds us of Heisenberg's uncertainty principle introduced in the early days of quantum mechanics \cite{Heisenberg1927},
\begin{equation}\label{eq:HUP}
\Delta x^2 \Delta p^2 \geq \hbar^2/4,
\end{equation}
where $\Delta x^2 \equiv \langle x^2 \rangle - \langle x \rangle^2 $ and $\Delta p^2 \equiv \langle p^2 \rangle - \langle p \rangle^2$. The uncertainty principle~(\ref{eq:HUP}) provides a non-linear bound of expectation values of physical observables and is saturated by the ground state of the harmonic oscillator. Inspired by this analogy, one wonders about the existence of a similar relation between $\langle \hat H_0 \rangle$ and $\langle \hat n_i \rangle$ for Eq.~(\ref{eq:DFT}), which in turn determines the density functional $F[n]$ explicitly.
In this Letter, we provide a definite answer to this postulation by developing a comprehensive theory for systems with a family of Hamiltonians, which was recently introduced by the author as the eigenstate moduli problem \cite{song2023reduced}. We show the correspondence between any Hamiltonian family and its moduli space of operator expectation values, where the ground states are uniquely determined by the boundary of this moduli space. Furthermore, all eigenstates correspond to singularity sets of the mapping and are determined by a unique equation. This finding unifies Heisenberg's uncertainty principle and the DFT as special cases of our general framework. As a result, we provide constructive proof of HK theorems with an explicit construction of $F[n]$. We demonstrate the explicit HK functional for systems with several particles and investigate their geometry in detail. Finally, we show our framework does not rely on the introduction of a Hilbert space and provides a potential alternative quantum formulation as a vast generalization of Heisenberg's uncertainty principle.
{\it Moduli problem.--}
We provide a comprehensive theory of the moduli problem. Consider a Hilbert space $\mathcal{H}$, and the corresponding real vector space of self-adjoint operators $\mathcal{L}_{s}(\mathcal{H})$. For simplicity, we assume the Hilbert space $\mathcal{H}$ has a finite dimension $N$. We are interested in the operator subspace $O \subset \mathcal{L}_{s}(\mathcal{H})$ of dimension $M$, which includes the identity operator $\hat I \in O$ as one of its elements. The operator subspace $O$ defines a family of systems where Hamiltonians $H(\lambda) \in O$ are parameterized by a $M$-dimensional real vector $\lambda$. One may express $\hat H(\lambda) = \sum_{i} \lambda_i \hat H_i$ in terms of a linear combination of a set of linearly-independent basis $\mathbf{\hat H} \equiv \{\hat H_i\}$. The moduli problem aims to address the solution space of the Hamiltonian family $H(\lambda)$. We start with the Schr\"odinger equation
\begin{equation}\label{eq:schrodinger}
\hat H(\lambda) \psi = 0,
\end{equation}
where the eigenvalue term $-E \hat I$ has been absorbed by a linear combination of basis. Instead of solving the eigenstate $\psi$ of a fixed parameter set $\lambda$, we consider the dual problem of solving $\lambda$ for a fixed $\psi$. We rewrite the Schr\"odinger equation~({\ref{eq:schrodinger}}) as
\begin{equation}
\lambda^t \mathbf{\hat H} \psi = 0.
\end{equation}
Since $\lambda$ is real, we also require its conjugate equation, leading to
\begin{equation} \label{eq:coker}
\lambda^t \mathcal{J}(\psi, \psi^\dag) = 0,
\end{equation}
where the $M \times 2N$ rectangular matrix
\begin{equation}\label{eq:J}
\mathcal{J}(\psi, \psi^\dag) \equiv \left [ \psi^{\dag}\hat H_i , \psi^t \hat H_i ^t \right],
\end{equation}
is the row concatenation over $1\leq i \leq M$. For technical convenience, we will assume $M < 2N$ from now on. Equation~({\ref{eq:coker}) implies the existence of a non-trivial cokernel $\lambda$ for the matrix $\mathcal{J}$, reflecting the fact that it is dual to the Schr\"odinger equation~(\ref{eq:schrodinger}). This cokernel condition has a geometrical interpretation. Consider a mapping $\rho$ from $\mathcal{H}$ to a $M$-dimensional moduli space $\mathcal{M}$ consisting of the expectation values of the operator subspace $O$, or
\begin{equation}
\rho(\psi, \overline \psi) \equiv \left( \langle \hat H_1 \rangle , \ldots, \langle \hat H_{M} \rangle \right),
\end{equation}
by choosing a set of coordinates $\{\langle H_i\rangle\}$, where $ \langle \cdot \rangle \equiv \langle \psi | \cdot| \psi \rangle$ denotes the expectation value of an operator. The moduli space $\mathcal{M}$ is generally a semi-algebraic set (see Fig.~\ref{fig:map}). Consider the energy functional
\begin{equation}\label{eq:E}
E[\psi] = \langle \hat H(\lambda) \rangle = \lambda^t \rho .
\end{equation}
The variational principle suggests that the stationary points of Eq.~(\ref{eq:E}) correspond to the eigenstates $\psi$ of $\hat H$, satisfying
\begin{equation}\label{eq:var}
0 = \lambda^t \frac{\delta \rho}{\delta \psi} =
\lambda^t \mathcal{J},
\end{equation}
that recovers the cokernel condition~(\ref{eq:coker}), where the Jacobian $\mathcal{J} = \delta \rho /\delta \psi $ is precisely the one defined in Eq.~(\ref{eq:J}). The existence of non-trivial cokernel of the Jacobian implies the eigenstate family corresponds to the singularity set of the moduli space $\mathcal{M}$ \cite{hartshorne2013algebraic}, where the parameter $\lambda$ is the normal vector at the point $\rho(\psi_{eigen})$. Moreover, all eigenstates form a algebraic variety. To see this, we notice that the cokernel condition~(\ref{eq:coker}) is equivalent to the vanishing of all $M \times M$ minors, as
\begin{equation}\label{eq:minor}
\det M_{I_1,\ldots, I_k, J_1, \ldots, J_{k'}} = 0
\end{equation}
where $ M_{I_1,\ldots, I_k, J_1, \ldots, J_{k'}}$ represents the submatrix formed by selecting $\{I_1,\ldots, I_k \}$-th $\psi$ columns and $\{J_1,\ldots, J_{k'} \}$-th $\overline\psi$ columns in Eq.~(\ref{eq:J}), with $k+k' = M$ and $k,k' \leq N$. The minor condition~({\ref{eq:minor}) generates a homogeneous ideal $\mathcal{I}$ over the polynomial ring $\mathbb{C}[\{\psi, \overline{\psi}\}]$. Consequently, it determines the eigenstate spaces $\mathcal{M} \equiv \mathbb{C}[\{\psi, \overline{\psi}\}]/\mathcal{I}$ as the corresponding homogeneous coordinate ring. Consequently, the minor condition~(\ref{eq:minor}) also determines the geometry of their images, i.e., the eigenstate moduli. The explicit equations in terms of the expectation values $\{\rho_i\}$ can be constructed explicitly as
\begin{equation} \label{eq:elimination}
\langle \det M_{I_1,\ldots, I_k, J_1, \ldots, J_{k'}}, \langle \psi | \hat H_i | \psi \rangle - \rho_i \rangle \cap \mathbb{C}[\rho_i] ,
\end{equation}
which eliminates $\psi$ with the affine coordindates $\rho_i$. Equation~(\ref{eq:elimination}) can be solved explicitly by Buchberger's algorithm using the Gr\"obner basis. Moreover, the eigenstate moduli have a codimension one, implying that Eq.~(\ref{eq:elimination}}) reduces to a single algebraic relation
\begin{equation} \label{eq:ae}
f(\rho_1, \ldots, \rho_M) = 0,
\end{equation}
in terms of the expectation values $\rho_i$. The ground state (and the largest eigenvalue state) family of Eq.~(\ref{eq:schrodinger}), if exists, corresponds to the boundary $\partial \mathcal{M}$ as a partial solution of Eq.~(\ref{eq:ae}). This is due to the fact that the smallest and largest eigenvalues are the global lower and upper bounds of the energy functional~(\ref{eq:E}). In fact, the projective nature of the parameter set $\lambda$ for $M > 2$ implies that the lower and upper bounds are connected, corresponding to a single closed boundary. Figure~{\ref{fig:map} illustrates the mapping $\rho$ from the Hilbert space $\mathcal{H}$ to the moduli space $\mathcal{M}$. The boundary $\partial\mathcal{M}$ offers a bound of the algebraic equation~ (\ref{eq:ae}), and can be viewed as a generalization of Heisenberg uncertainty principle~(\ref{eq:HUP}), as we will demonstrate in next section.
\begin{figure}
\includegraphics[width=1\linewidth]{fig1.pdf}
\caption{ Mapping $\rho$ from the Hilbert space $\mathcal{H}$ (grey domain) to the moduli space $\mathcal{M}$ (dark grey domain). The solid and dashed curves in $\mathcal{H}$ correspond to ground and excited states, where the former maps to the boundary $\partial \mathcal{M}$.
}.
\label{fig:map}
\end{figure}
The mapping $\rho$ is not injective, implying that a point in the moduli space $\mathcal{M}$ might correspond to multiple $\psi$ in $\mathcal{H}$. However, if there is no degeneration, the ground state is uniquely determined by the boundary $\partial M$ and vice versa. For the degenerate case, there exists a center of the operator subspace $\mathfrak{z}(O)$, where any self-adjoint operator $\hat L \in \mathfrak{z}(O)$ commutes with Hamiltonian,
\begin{equation}\label{eq:L}
[\hat L, \hat H(\lambda)] = 0,
\end{equation}
which defines a symmetry group $Z \equiv \exp(\i \mathfrak{z}(O)) \subset U(N)$. The moduli space $\mathcal{H}$ is the image of $\mathcal{H}$ modulus $Z$. Nevertheless, one may include $\mathfrak{z}(O)$ into the operator subspace $O$ to distinguish degenerate ground states.
{\it Uncertainty Principle. --}
Our approach can be further generalized to infinite dimensional Hilbert spaces. Consider a simple example of the harmonic oscillator with the Hamiltonian family,
\begin{equation}\label{eq:HO}
\hat H(\lambda) = \lambda_{p^2} \hat p^2 + \lambda_{x^2} \hat x^2 + \lambda_{p} \hat p + \lambda_{x} \hat x + \lambda_{-E} \hat I.
\end{equation}
The moduli space $\mathcal{M}_{ho}$ of Eq.~(\ref{eq:HO}) is a four-dimensional semi-algebraic set for expectation values $\langle p^2 \rangle$, $\langle x^2 \rangle$, $\langle p \rangle$ and $\langle x \rangle$, whose geometry is determined by Heisenberg's uncertainty principle~(\ref{eq:HUP}), as shown in Fig.~\ref{fig:HO}.The ground state of Eq.~(\ref{eq:HO}) saturates this inequality, corresponding to the boundary of the moduli space $\partial\mathcal{M}_{ho}$, as expected. The eigenstates correspond to the singularity set of the mapping $\rho$, determined by the analytical equation $f = \prod_{n=0}^\infty \left(1-\frac{\Delta x^2 \Delta p^2}{((n+1/2)\hbar)^2}\right)= \cos \left( \frac{\pi \sqrt{\Delta x^2 \Delta p^2}}{\hbar}\right) = 0$, an infinite-dimensional analogy to the algebraic equation~({\ref{eq:ae}). By further including operator $\hat x \hat p + \hat p \hat x$ to the Hamiltonian family, we end up with Schr\"odinger-Robertson uncertainty relation \cite{robertson1929uncertainty,schrodinger1930sitzungsberichte}.
\begin{figure}
\includegraphics[width=1\linewidth]{fig2.pdf}
\caption{The moduli space $\mathcal{M}_{ho}$ for the harmonic oscillator is represented by the grey domain and bounded by Heisenberg's uncertainty principle. The solid curve represents the ground state moduli, saturating the uncertainty principle. Dashed curves correspond to excited states. }
\label{fig:HO}
\end{figure}
In general, given a Hilbert space $\mathcal{H}$ and the corresponding real vector space $\mathcal{L}_{s}(\mathcal{H})$ of self-adjoint operators, we introduce the operator space category $\textbf{Op}$, whose object is subspace $O_i \subset \mathcal{L}_{s}(\mathcal{H})$, including the identity operator $\hat I \in O_i$. Morphisms are inclusions between subspaces. The category $\textbf{Op}$ contains all physical relevant information. It has the initial object $\mathcal{I}_{Op} = \mathcal{L}_{s}(\mathcal{H})$ and terminal object $\mathcal{T}_{Op} = \hat I$. We next introduce the category $\textbf{Ev}$ of the expectation value spaces, whose object $E_i = \langle O_i \rangle$ is space formed by the expectation values of $O_i$, generally a real semi-algebraic set. Morphisms are linear projections between objects. There is a natural functor from the operator category $\textbf{Op}$ to the expectation value category $\textbf{Ev}$. Note that the construction of $\textbf{Ev}$ does not necessarily rely on the introduction of the Hilbert space $\mathcal{H}$. To see this, we first construct the initial object $\mathcal{I}_{Ev}$. After choosing the coordinates, $\mathcal{I}_{Ev}$ is an affine variety formed by a Veronese-type embedding $\psi \rightarrow \overline{\psi}_i \psi_j \equiv x_{ij}$, whose ideal is generated by a set of quadratic equations $x_{ij}x_{kl} = x_{il} x_{kj}$ without involving $\psi$. An object $E_i$ in $\textbf{Ev}$ is determined by the linear projection $\pi_i$ from the initial object $\mathcal{I}_{Ev}$ to it, since the map $\rho$ factors through $\pi_i$ as $\rho(\mathcal{H}\rightarrow E_i) =\pi_i \circ \rho(\mathcal{H}\rightarrow \mathcal{I}_{Ev}) $. All physical measurements can be read directly from $\textbf{Ev}$, either from the geometry of an object (ground state expectations) and/or the morphism from $\mathcal{I}_{Ev}$ to it (excited state expectations). Thus, the category $\textbf{Ev}$ contains all time-independent physics without involving the Hilbert space $\mathcal{H}$. Thus, our approach provides a potential new quantum formulation as a vast generalization of Heisenberg's uncertainty principle. Our framework can be further generalized to thermal ensembles by considering the following partition function
\begin{equation}
\mathcal{Z}(\lambda) \equiv \int e^{-\lambda^t \rho } d \mu(f[\rho]),
\end{equation}
where $\mu(f)$ is the invariant measure of Eq.~(\ref{eq:ae}). We leave the development of the time-dependent theory and generalizations to quantum field theory for future research.
{\it Density functional.--}
We demonstrate the DFT as a special case of our theory. Consider a system of $n$ identical particles with the single-particle Hilbert space $\mathcal{H}^1$. For simplicity, we require $\mathcal{H}^1$ to be finite-dimensional with $q$ states. The Hilbert space of the system $\mathcal{H}^\pm = \mathrm{Sym}^\pm \otimes^n \mathcal{H}^1$, where $\mathrm{Sym}^\dag$ and $\mathrm{Sym}^-$ are the symmetrization and anti-symmetrization projections, with dimension $N = \binom{q+n-1}{n}$ and $\binom{q}{n}$for bosons and fermions, respectively.
The DFT Hamiltonian family satisfies Eq.~(\ref{eq:DFT}), with $ \hat H_0 = \sum_{ij} \left( -t_{ij} a_i^\dag a_j + U_{ij} \hat n_i \hat n_j \right) $, where $a_i^\dag$ and $a_i$ are creation and annihilation operators, and $\hat n_i \equiv a_i^\dag a_i $. Since $\sum \hat n_i = n \hat I$. Note that the energy term $E$ can be absorbed into the parameter set by $V_i \rightarrow V_i - E/n$. The corresponding parameter $\lambda = (1, V_1, \ldots, V_q )$ and Hamiltonian basis $\mathbf{\hat H} \equiv (\hat H_0, \hat n_1, \ldots, \hat n_q )$ with $M = q+1$, leading to the Jacobian
\begin{equation}\label{eq:JDFT}
\mathcal{J}_{DFT}(\psi, \psi^\dag) \equiv
\begin{bmatrix}
\psi^{\dag}\hat H_0 , & \psi^t \hat H_0 ^t \\
\psi^{\dag}\hat n_1 , & \psi^t \hat n_1 ^t \\
\ldots, & \ldots \\
\psi^{\dag}\hat n_q , & \psi^t \hat n_q ^t
\end{bmatrix}.
\end{equation}
Introducing $F \equiv \rho_0 = \langle \hat H_0 \rangle$ and $n_i \equiv \rho_i = \langle \hat n_i \rangle $ with $1\leq i \leq q$, Eq.~(\ref{eq:ae}) defines the algebraic equation connecting $F$ to $n_i$, providing an explicit construction of the HK functional.
\begin{figure}
\includegraphics[width=1\linewidth]{fig3.pdf}
\caption{The density functional $F[n]$ in unit of $t$ for the two-boson system~(\ref{eq:toy}) for a fixed $U/t = 1$ and $U'/t = 0$. Curves are determined by Eq.~(\ref{eq:functional}), where colors correspond to three different eigenstates. Grey scatter points generate a random sampling of the moduli space $\mathcal{M}$, which is bounded by $F[n]$. The black arrow represent the normal vector $(2, V_1-V_2)$.}
\label{fig:DFT2}
\end{figure}
To demonstrate a concrete example of this construction, we analytically solve Eq.~(\ref{eq:elimination}) for small $(n,q)$-systems. We start with the simplest case $q=2$ for $n=2$ bosons, with the Hamiltonian
\begin{equation}\label{eq:toy}
\hat H_0 = -t \left( a_1^\dag a_2 + a_2^\dag a_1 \right) + \frac{U}{2} \sum_{i=1,2} \hat n_i (\hat n_i-1) + U' \hat n_1 \hat n_2,
\end{equation}
for fixed parameters $t$, $U$ and $U'$. This toy model has $N=3$ states, implying there are three eigenstates. We find the explicit form of Eq.~(\ref{eq:ae}) as a homogeneous algebraic equation,
\begin{equation}\label{eq:functional}
f(F,n_1 \ldots, n_q) = \sum_{r_0 + \ldots + r_q = r} c_{r_0,\ldots, r_q} n_1^{r_1} \ldots n_{q}^{r_q} F^{r_0} = 0,
\end{equation}
with the degree $r = 6$, where $c_{r_0,\ldots, r_q}$ are real constants depending only on $\hat H_0$ (see SM Eq.~S1). Note that for this toy model, the density functional $F[n]$ is a one-dimensional function since the sum $n_1 + n_2 = 2$ is fixed. Figure~\ref{fig:DFT2} visualizes Eq.~(\ref{eq:functional}) by plotting the density functional $F[n]$ as a function of $\Delta n = n_2 - n_1$. The geometry of $F[n]$ contains three curves, each corresponding to one of the three eigenstates. Note that the ground state and highest excited state correspond to the boundary and are connected smoothly as predicted by our theory. The first excited state corresponds to the inner singular curve with three cusps. The normal vector corresponds to $(2, V_1-V_2)$. To test our theory, we generate 10,000 $(n_1, n_2, F)$ points by randomly sampling the wavefunction $\psi$ and calculating numerically the corresponding expectation values. This allows us to visualize the moduli space $\mathcal{M}$ numerically, as shown by the grey scatter points in Fig. \ref{fig:DFT2}. We find that $\mathcal{M}$ is tightly bounded by the ground state moduli (blue curve).
To understand better the ground state moduli, we find analytically the lowest functional $F_0 = U - \frac{1}{2} \left(\Delta U + \sqrt{\Delta U ^2+(4t)^2}\right)$ with $\Delta U = U - U'$ occurs at $n_1 = n_2$, corresponding to the normal vector $(1,0)$, i.e., $V_1 = V_2$. Increasing $V_1$ or $V_2$ drives the ground state to a higher value of $F[n]$. Expanding $F[n] = F_0 + F_1 n_1 n_2 + O((n_1 n_2)^2)$ around $F_0$ in terms of the power series of $n_1n_2$, we find
\begin{equation}
F_1 = \frac{4 t^2 \sqrt{\Delta{U}^2+16 t^2}}{16 t^2-2 \Delta{U} \left(\sqrt{\Delta{U}^2+16 t^2}-\Delta{U}\right)}.
\end{equation}
For a weak coupling $\Delta U/t \ll 1$, we find $F_1 = - t^2 - \Delta U/2 + O(\Delta U ^2)$, which can be interpreted as the kinetic and the mean-field interaction term. However, the exact interaction expectation $U_1-\Delta U \langle \hat n_1 \hat n_2 \rangle$ has a different prefactor, reflecting a hidden correlation $\langle \hat n_1 \hat n_2 \rangle \neq \langle \hat n_1 \rangle \langle \hat n_2 \rangle$. Compared to the ground state, the excited state has a more interesting geometry, as shown by the green curve in Fig.~\ref{fig:DFT2}. There are three cusps where the middle one at $n_1=n_2$ corresponds to limiting cases $V_1 \rightarrow \infty$ or $V_2 \rightarrow \infty$. The other two cusps are due to $F =\langle H_0 \rangle$ and $\langle n_1 - n_2 \rangle$ of excited states not changing monotonically with the external potential $V_1-V_2$. In fact, there a critical values of $V_1 -V_2 = \Delta V_c$ when both $F$ and $\Delta n$ reach their maximum value (see SM Fig.~S1). Expanding around $\Delta V_c$ to the third order, we find $F$ and $\Delta n$ satisfy the normal cusp equation $y^2 = x^3$ after a linear transformation.
What about larger systems? Equation~(\ref{eq:functional}) still holds, however, its degree $r$ increases rapidly with system size. For example, by adding one more particle into the toy model~(\ref{eq:toy}), we find an equation of $r=12$ degree (see SM Eq.~S2 and Fig.~S2). This observation implies that an explicit form of Eq.~(\ref{eq:functional}) may not be of practical use. Not only because it is computationally infeasible to solve Eq.~(\ref{eq:elimination}) for large systems, but also because of an exponential number of terms occurring in the analytical form of the density functional~(\ref{eq:functional}). Nonetheless, our analytical solution will provide promising guidance for future developments in density functional approximations.
In conclusion, we develop a general theory for relations of expectation values among an arbitrary set of physical observables, unifying Heisenberg’s uncertainty principle and HK theorems into a general framework. As a result of our theory, we provide an explicit construction of exact density functional. Our approach leaves many challenges and opportunities for future research. For instance, developing the time-dependent theory will complete our framework as an alternative formulation of quantum theory. One also wonders about reconstructing Hilbert space from the proposed expectation value category, a question closely related to the recently developed geometrical formulation of quantum theory \cite{kibble1979geometrization, ashtekar1995geometry}. In all, our approach provides a new perspective on quantum theory, especially for many-body systems, with potential implications for strongly correlated systems, quantum computational chemistry, and quantum field theory.
|
\section{Introduction}\label{introduction}
We assume that the reader is familiar with the concepts of polar spaces, projective embeddings, subspaces, which we will briefly recall in Section~\ref{Prelim}.
Let $\cP:=(P,L)$ be a non-degenerate polar space, regarded as a point-line geometry. Given two points $p,q$ of $\cP$, we will say that they are
{\it collinear}, and write $p\perp q,$ if there exists a line $\ell\in L$
incident with both $p$ and $q$. For any $x\in P$ we shall write
$x^{\perp}$ for the set of all points of $\cP$ collinear with $x$ and if $X\subseteq P, $ then $X^{\perp}:=\bigcap_{x\in X}x^{\perp}.$
A \emph{subspace} of $\cP$ is a subset $X\subseteq P$ such that every line containing at least two points of $X$ is entirely contained in $X$.
The intersection of all subspaces of $\cP$ containing a given subset
$S\subseteq P$ is a subspace called the \emph{span} of $S$ and denoted by $\langle S\rangle.$ A {\em hyperplane} of $\cP$ is a proper subspace of $\cP$ meeting every line of $\cP$ non-trivially.
A subspace $X$ is {\it singular} if for all $x,y\in X$ we have $x\perp y$ (equivalently $X\subseteq X^{\perp}$).
The subspace $\cP^{\perp}: =\{ p\in P \colon p^\perp = P \}$ is called the \emph{radical} of $\cP$ and denoted by $\mathrm{Rad}(\cP)$. Obviously, $\mathrm{Rad}(\cP)$ is a singular subspace and is contained in all maximal singular subspaces of $\cP$. The polar space $\cP$ is \emph{degenerate} precisely when $\mathrm{Rad}(\cP) \neq \emptyset$.
Every subspace $S$ of $\cP$ can be naturally endowed with the structure of a polar space by taking as points the elements of $S$ and as lines all lines of $\cP$ fully contained in $S$.
In particular a singular subspace, regarded as a polar space, coincides with its own radical.
In this paper we shall always assume that $\cP$ is non-degenerate.
It is well-known that all singular subspaces of a non-degenerate polar
space are projective spaces; see~\cite[Theorem 7.3.6]{S}.
A hyperplane $\cH$ is called \emph{singular} if $\cH=p^\perp$ for $p\in \cP.$
If $\cP$ is non-degenerate, then any degenerate hyperplane $\cH$ of
$\cP$ is singular, i.e. $\cH=p^{\perp}$ where $\mathrm{Rad}(\cH)=\{p\}$.
The rank of a polar space $\cP$, usually denoted by $\rank(\cP)$, is
defined as the least upper bound of the lengths of the
well ordered chains of singular subspaces contained in it with $\emptyset$
regarded as the smallest singular subspace. We recall that the \emph{length} of a
chain is its cardinality, diminished by $1$ when the chain is finite.
In particular, the rank of an infinite well ordered chain is just its cardinality, namely the cardinality of the ordinal number representing
the isomorphism class of the chain itself.
The rank of a projective space is its generating rank, namely its dimension augmented by $1$. The \emph{rank} $\rank(X)$ of a singular subspace $X$ of $\cP$ is its rank as a projective space.
If the maximal singular subspaces of $\cP$ have all finite rank then they all have the same rank. This common rank coincides with $\rank(\cP)$ according to the definition above.
For the sake of completeness, when $\cP$ admits singular subspaces of infinite rank we put $\rank(\cP) = \infty$ even if we shall not deal with this case in the present paper.
Polar spaces of rank $2$ are called
\emph{generalized quadrangles}. Polar spaces of rank $1$ are just sets
of pairwise non-collinear points.
Henceforth we shall always assume that $\cP$ has finite rank, say $n$.
The hyperplanes of $\cP$ have rank either $n-1$ or $n$.
Since the rank of $\cP$ is finite, for any maximal singular subspace $M$ there exists another maximal singular subspace $M'$ disjoint from $M$.
Then, fixed a basis $(p_1,\dots, p_n)$ of $M$ (by {\it basis} of a projective space we mean a minimal generating set), there is a unique basis
$(p'_1,\dots, p'_n)$ of $M'$ such that $p_i\perp p'_j$ if and only if $i\neq j$. Such a pair $\{(p_1,\dots, p_n),(p'_1,\dots, p'_n)\}$ of generating
sets is called a {\em frame} of $\cP$ (see \cite{PasDG},~\cite{BC}).
In general, if $\cP$ is embeddable of rank $n$, a set spanning $\cP$ must
contain at least $2n$ points, the cardinality of a frame. In this case,
the frame is a minimal set spanning a subspace of $\cP$ with the same
rank as $\cP$.
Note that there are non-embeddable subspaces of rank $3$ which are spanned by a number of points smaller then the cardinality of a frame (\cite{P-VM}).
In~\cite{ILP21b}, we introduced an intrinsic parameter of $\cP$ called the \emph{anisotropic gap} (there under the name of {\it anisotropic defect}) of $\cP$ as follows. Let $\mathfrak{N}(\cP)$ denote the family, ordered by inclusion, of the well-ordered chains of subspaces of $\cP$ containing a frame. Then
the {\em anisotropic gap} $\mathrm{gap}({\cP})$ of $\cP$ is the least upper bound of the lengths of the elements of $\mathfrak{N}(\cP)$, i.e
\begin{equation}\label{anisotropic gap}
\mathrm{gap}(\cP)=\sup\{|C|-1\colon C\in \mathfrak{N}(\cP)\}.
\end{equation}
In other words,
the anisotropic gap of $\cP$ tells how ``far'' $\cP$ is from any of its subspaces spanned by frames.
A polar space is \emph{classical} if it is non-degenerate and
it admits the universal embedding; see Section~\ref{Prelim}.
Suppose that $\cP$ is a classical polar space and let $\varepsilon:\cP\rightarrow\PG(V)$ be its universal embedding.
Call $\KK$ the underlying division ring of $V.$
In this case, $\varepsilon(\cP) = \cP(f)$ with $f$ a non-degenerate alternating form or $\varepsilon(\cP) = \cP(\phi)$ for a non-degenerate pseudoquadratic form $\phi$.
We recall that the {\it sesquilinearization} $f_\phi$ of $\phi$ (henceforth denoted by just $f$ for simplicity) can be degenerate only if $char(\KK)=2.$
When $\phi$ is a non-degenerate quadratic form, $\cP(\phi)$ is called a {\it non-degenerate orthogonal polar space}; in this case $\cP(\phi)$ is defined over a field.
Clearly, when considering two orthogonal vectors or subspaces of $V$, we refer to orthogonality with respect to $f$ or with respect to the sesquilinearization of $\phi$, according to the case.
Given a frame $A$ of $\cP$, its image $\varepsilon(A)$ spans a $2n$-dimensional subspace $H$ of $V$ which splits as the direct sum $V_1\oplus V_2\oplus\dots\oplus V_n$ of mutually orthogonal $2$-dimensional subspaces $V_1, V_2,\dots, V_n$. These subspaces bijectively correspond to the $n$ pairs of non-collinear points of $A$ and appear as lines in $\PG(V)$.
Denote by $[V_i]$ the projective line corresponding to the vector
subspace $V_i$ for $i=1,\dots,n$.
The preimages $\varepsilon^{-1}([V_1]),\dots, \varepsilon^{-1}([V_n])$ are hyperbolic lines of $\cP$ (a {\em hyperbolic line} of a polar space being the double perp $\{p,q\}^{\perp\perp}$ of two non-collinear points $p$ and $q$).
If $V_0$ denotes an orthogonal complement of $H=V_1\oplus\dots \oplus V_n$ in $V$ then $V_0$ is $\phi$-anisotropic, i.e. $\phi(x)\neq 0,\, \forall x\in V_0\setminus\{\mathbf{0}\}$.
If $f$ is degenerate but $\cP$ is not,
then $V_0 \supseteq \mathrm{Rad}(f).$ In this case, suppose $V_0'$ is a complement of $\mathrm{Rad}(f)$ in $V_0$, that is $V_0=V_0'\oplus \mathrm{Rad}(f).$
We have the following orthogonal decomposition
\begin{equation}\label{decomposition}
V ~ = ~ H\oplus V_0 ~ = ~ (V_1 \oplus V_2 \oplus \dots \oplus V_n)\oplus V_0 ~ = ~ (V_1 \oplus V_2 \oplus \dots \oplus V_n)\oplus V_0'\oplus \mathrm{Rad}(f).
\end{equation}
In~\cite{ILP21b} we proved the following
\begin{theorem}\label{plain}
Let $\cP$ be a classical non-degenerate polar space. Then the anisotropic gap of $\cP$ is precisely the codimension of $V_1 \oplus V_2 \oplus \dots \oplus V_n$ in $V$, i.e. $\mathrm{gap}(\cP) =\dim(V_0)$. Moreover, every well ordered chain of $\mathfrak{N}({\cP})$ is contained in a maximal well ordered chain and all maximal well ordered chains of $\mathfrak{N}({\cP})$ have the same length, namely $\mathrm{gap}(\cP)$.
\end{theorem}
In the same paper, for $\ch(\KK)=2$, we mentioned two more parameters called \emph{parabolic} and \emph{elliptic gaps} of $\cP$ (there called parabolic and elliptic defects).
The parabolic gap corresponds to the dimension of the
radical $\mathrm{Rad}(f)$ and the elliptic gap is
defined as $\dim(V_0')$ (see~\eqref{decomposition}). We point out that our definition of
\emph{parabolic gap} corresponds to the definition of \emph{corank}
of the form $\phi$ in~\cite{T} when the form $\phi$ is non-degenerate.
It turns out that the notion of parabolic gap is also closely
related to the existence of projective embeddings of a polar space
different from the universal one (see Section~\ref{section finale}).
In the present paper, assuming that $\cP$ admits the universal embedding, we provide an intrinsic characterization of
the notions of parabolic and elliptic gap of $\cP$ along the lines of Theorem~\ref{plain}, thus answering Problem 5.2 of~\cite{ILP21b}.
We will also give a characterization of orthogonal polar spaces
(see Definition~\ref{oeh})
without explicit reference to the quadratic form describing them.
Once more, we only require the existence of the universal embedding.
The main motivation of this paper is to continue on the project started in~\cite{ILP21b} aimed to offer an embedding-free definition for
the notion of ``dimension'' of a classical polar space $\cP$.
The most natural way is to identify the dimension of $\cP$ with the vector dimension of its universal embedding, i.e. its embedding rank. Hence $\dim(\cP)=2n + \fd$, where $n$ is the rank of $\cP$ (for which we have an intrinsic definition) and $\fd$ is the dimension of a complement of the space spanned by a frame. In~\cite{ILP21b} we proved that $\fd$ can be defined in an embedding-free way by the length of well-ordered chains of suitable subspaces of $\cP$ thus leading to the definition of anisotropic gap. Here, we continue the job focusing on orthogonal polar spaces in characteristic $2$, where $\fd$ is in turn the sum of two terms (the elliptic and the parabolic gap) which we characterize intrinsically respectively in Theorems~\ref{m:1} and~\ref{m:2}.
Our characterizations provide some insight on the
requirements for two orthogonal polar spaces to be isomorphic and
provide the groundwork for some future research.
Indeed,
in many cases, even in characteristic $0$, it is not sufficient for two
orthogonal polar spaces having both the same rank and the same
anisotropic gap to
be isomorphic; for instance this is the case for polar spaces over
the rational field $\mathbb Q$; on the other hand, two orthogonal
polar spaces with the same rank and anisotropic gap over the real field $\mathbb R$ are isomorphic.
We aim to further investigate the relationship between
the fields involved, the gaps we defined, and isomorphisms in
future works.
Furthermore, the notions we introduce in the present paper
can be extended also to polar spaces described by
hermitian forms in characteristic $2$ (over non-commutative
division rings) as well as to some non-embeddable cases. We leave this
study to a future work.
We warn the reader that the results of the present paper all
encompass
the case in which the gaps are possibly infinite
where a Witt-like decomposition of the quadratic form describing
the universal embedding might not be easily manageable.
None the less, as
mentioned above, we restrict our current analysis to spaces of finite
rank $n$.
Throughout the paper we use greek letters to denote ordinals related to
chains and blackletter characters for cardinal numbers which might possibly
be infinite.
\medskip
We will say that an embedded non-degenerate orthogonal polar space $\cP(\phi)$ of rank $n$ is \emph{$(\fe,\fp)$-orthogonal} if in the decomposition~\eqref{decomposition}
$\dim(V_0')=\fe$ and $\dim(\mathrm{Rad}(f))=\fp$.
More in particular, we say that $\cP(\phi)$ is
of \emph{hyperbolic type} if $\fe=\fp=0$; so
$\cP(\phi)$ is spanned by a frame and
\[V ~ = V_1 \oplus V_2 \oplus \dots \oplus V_n;\]
it is \emph{elliptic} if $\fe>0$ but $\fp=0$, that is
$\cP(\phi)$ is not hyperbolic and the bilinear form $f$ polarizing $\phi$ is non-degenerate, i.e. $\mathrm{Rad}(f)=\{\mathbf{0}\}$ and $V_0=V_0'$ in \eqref{decomposition}:
\[V ~ = (V_1 \oplus V_2 \oplus \dots \oplus V_n)\oplus V_0';\]
it is \emph{parabolic} if $\fe=0$ and $\fp>0$, that is it is not hyperbolic and $\mathrm{Rad}(f)=V_0$, i.e. $\mathrm{Rad}(f)$ is a non-trivial direct complement of
$H=\bigoplus_{i=1}^{n}V_i$ in $V$, i.e.
\[V ~ = (V_1 \oplus V_2 \oplus \dots \oplus V_n)\oplus \mathrm{Rad}(f).\]
The aim of this paper is to provide an
intrinsic description of $(\fe,\fp)$-orthogonal spaces without mentioning
the embedding; for this, we refer to Corollary~\ref{d112}.
We observe that in general there
might be $(\fe,\fp)$-orthogonal spaces where both $\fe>0$ and $\fp>0$.
The following results, to be proved in Section~\ref{hyperb lines}, characterize hyperbolic and elliptic orthogonal polar spaces relying on the cardinality of hyperbolic lines.
\begin{theorem}\label{thm hyperb lines}
Let $\cP$ be an embeddable polar space and $\varepsilon$ a relatively universal embedding of $\cP$. The embedded polar space
$\varepsilon(\cP) $ is $(\fe,0)$-orthogonal (i.e. either of hyperbolic or
elliptic type) if and only if the hyperbolic lines of $\cP$ contain exactly $2$ points.
\end{theorem}
In light of Theorem~\ref{thm hyperb lines} it is possible to provide
an intrinsic characterization encompassing all orthogonal polar spaces,
as shown by the following corollary.
\begin{corollary}
\label{pr:orth}
Let $\cP$ be an embeddable polar space and $\varepsilon$ a relatively universal embedding of $\cP$. The embedded polar space $\varepsilon(\cP)$ is orthogonal i.e. $\varepsilon(\cP)=\cP(\phi)$ where
$\phi$ is a quadratic form, if and only if any subspace $\cF$ of $\cP$
generated by a frame has the property that all its hyperbolic lines consist of exactly $2$ points.
\end{corollary}
Using these results we can formulate the following definition of
orthogonal polar space which does not explicitly mention the
universal embedding.
\begin{definition}
\label{oeh}
A non-degenerate
embeddable polar space $\cP$ is \emph{orthogonal} if the hyperbolic lines
of any subspace $\cF$ of $\cP$ generated by a frame contain exactly $2$ points.
A non-degenerate polar (sub)space $\cP$
is \emph{hyperbolic} if it is orthogonal and generated by a
frame;
it is \emph{elliptic} if it is orthogonal but
not hyperbolic and each of its hyperbolic lines consists of $2$
points.
\end{definition}
According to this definition, we warn the reader that if $\mathrm{char}(\KK)\neq 2$, then the quadrics
of rank $n$ in dimension $2n+1$, usually called parabolic, will be here
named elliptic; see Remark~\ref{rem13} for more details on this choice.
\begin{definition}
\label{dec}
Suppose that $\cP$ is an orthogonal non-degenerate polar space.
We say that a well ordered chain of subspaces
\begin{equation}\label{elliptic chain}
\fE: \cF=\cE_0\subset\cE_1\subset\dots\subset\cE_\mu\subset\cdots \subset \cP
\end{equation}
is an \emph{elliptic chain} of $\cP$ if any $\cE_i$ for $i=1,\dots,\mu,\dots$
is elliptic and $\cF$ is a hyperbolic subspace of $\cP$.
\end{definition}
\begin{remark}
It might be possible to extend Definition~\ref{dec} of \emph{elliptic subspace} and
\emph{elliptic chain} also to polar spaces which are not orthogonal, by stating that
a subspace $\cE$ of $\cP$ containing a frame $F$ of $\cP$ is ``\emph{elliptic}'' if, called $\cF$ the subspace of $\cE$
generated by $F$, we have $\cF\neq\cE$ and $\forall p,q\in\cF$ with
$p\not\perp q$
\[ \{p,q\}^{\perp\perp}\cap\cF=\{p,q\}^{\perp\perp}\cap\cE, \]
that is the hyperbolic lines of $\cF$ are also hyperbolic lines of $\cE$.
As the group of a classical polar space of finite rank is transitive on the hyperbolic lines,
this new definition for orthogonal polar spaces is equivalent to Definition~\ref{oeh}.
On the other hand, non-elliptic subspaces might occur for non-orthogonal
classical polar spaces $\cP$ of finite rank only if $\KK$ is a
non-commutative division ring in characteristic $2$. We leave considering
these cases to a future work.
\end{remark}
Our main result is the following theorem.
\begin{theorem}
\label{m:1}
Let $\cP$ be a non-degenerate orthogonal polar space of rank at least $2$ embeddable over a field $\KK.$ Then all well ordered maximal elliptic chains of $\cP$ have the same cardinality $\fd$. In particular, if $\mathrm{char}(\KK)\neq 2$, then
$\fd$ is precisely the anisotropic gap of $\cP$ while if $\mathrm{char}(\KK)=2$ then all maximal
subspaces of $\cP$ having the property that all their hyperbolic lines contain exactly $2$ points
have the same
anisotropic gap equal to $2\fd$ and $2\fd$ is precisely the codimension of $\mathrm{Rad}(f)$ in $V_0$ where $f$ is the bilinearization of a (relatively) universal embedding of $\cP$ and $V_0$ is as in \eqref{decomposition}.
\end{theorem}
We prove in Corollary~\ref{e:max} that any elliptic chain $\fE$
admits a maximal element $\cE_{\omega}$ which is an elliptic subspace.
In light of this and Theorem~\ref{m:1}, the \emph{elliptic gap}
of $\cP,$ already introduced before, can be defined as follows.
\begin{definition}\label{elliptic gap}
Let $\cP$ be an
orthogonal polar space of rank at least $2$ and
$\fE: \cF=\cE_0\subset\cE_1\subset\dots\subset\cE_{\omega}$
be a maximal well-ordered
elliptic chain of $\cP$ where $\cF$ is the subspace spanned by a frame
and $\cE_{\omega}$ is a maximal element of $\fE$.
A \emph{maximal enrichment} of $\fE$ is a (possibly non-elliptic)
well ordered chain
$\fE^E: \cX_0=\cE_0\subseteq\cX_1\subseteq\dots\subseteq\cX_{\theta}=\cE_{\omega}$,
containing all elements of $\fE$,
starting at $\cE_0$ and ending at $\cE_{\omega}$ and maximal with
respect to these properties.
The \emph{elliptic gap} of $\cP$ is the length of
a maximal enrichment $\fE^E$ of a maximal elliptic chain $\fE$ of $\cP$.
\end{definition}
Observe that for any $0\leq i<\theta$, the space $\cX_{i}$ in $\fE^E$
is necessarily a maximal subspace of $\cX_{i+1}$.
Note that if $\mathrm{char}(\KK)\not=2$ then $\fE\equiv \fE^E$, i.e. a maximal
elliptic chain $\fE$ admits no proper enrichment, namely it is maximal as a chain of subspaces containing the span of a given frame; also,
in this case the elliptic gap coincides with the anisotropic gap of $\cP.$
In contrast, if $\mathrm{char}(\KK)=2$ then $\fE$ always admits proper refinements and
in this case the elliptic gap of $\cP$ is twice the length of $\fE$ (of course this makes sense if the elliptic gap is finite).
The anisotropic gap of $\cP$ is possibly larger.
In any case, the elliptic gap of $\cP$ is the anisotropic gap of
a maximal subspace of $\cP$ with the property that all its hyperbolic lines contain exactly $2$ points, i.e. it is the anisotropic gap of a maximal elliptic subspace of $\cP$.
A consequence of Theorem~\ref{m:1} is the following.
\begin{theorem}
\label{m:2}
Let $\cP$ be a non-degenerate orthogonal
polar space of rank at least $2$ and $\varepsilon\colon \cP\rightarrow \PG(V)$ its universal embedding.
Let $\fE^E$ be an enrichment of a maximal elliptic chain $\fE$ of $\cP$.
Then,
$\dim({\mathrm{Rad}}(f))=\dim\left(V/\left\langle\varepsilon\left(
\bigcup_{X\in \fE^E}X\right)\right\rangle\right)$.
\end{theorem}
\begin{remark}
\label{r:111}
If we call $\fr$ the anisotropic gap of $\cP$ and
$\fd$ its elliptic gap and both $\fr$ and $\fd$ are
finite, then
the statement of Theorem~\ref{m:2} reads as
$\dim(\mathrm{Rad}(f))=\fr-\fd$.
When $\fr$ and $\fd$ are possibly infinite cardinals
the statement
should be read as follows: denote by $\fr$ the anisotropic gap
of $\cP$ and
let $\cE$ be a maximal elliptic subspace of $\cP$ of
anisotropic gap $\fd$; then there exists a maximal chain ${\mathfrak M}$
of
subspaces of $\cP$ containing a frame of $\cP$ as well as $\cE$
and this chain has length $\fr$.
Let ${\mathfrak z}$ be the length
of the subchain of $\mathfrak M$ from $\cE$ to $\cP$; then
${\mathfrak z}+\fd=\fr$.
\end{remark}
\begin{definition}
By Theorem~\ref{m:2} the
\emph{parabolic gap} of $\cP$ can be defined as the
cardinality of any maximal well ordered chain $\mathfrak M$
of subspaces of $\cP$
all containing a space $\cE:=\bigcup_{X\in \fE}(X)$, where $\fE$ is any
maximal well ordered elliptic chain of $\cP$.
\end{definition}
In light of what we have said so far, the following corollary,
shows that the notion of $(\fe,\fp)$-orthogonal space can be
formulated without any explicit
mention of its embeddings and, as such, is intrinsic to the space.
\begin{corollary}
\label{d112}
Let $\cP$ be a non-degenerate embeddable
polar space such that the hyperbolic lines
of a subspace $\cF$
generated by a frame of $\cP$ consist of only two points.
Then, $\varepsilon(\cP)$ is an $(\fe,\fp)$-orthogonal polar space
where $e$ and $p$ are respectively the elliptic and the parabolic
gaps of $\cP$.
In particular, if $\fe=\fp=0$, the space $\cP$ is hyperbolic;
if $\fe>0$ and $\fp=0$, then $\cP$ is elliptic and if $\fe=0$ and $\fp>0$
the space is parabolic.
\end{corollary}
\begin{remark}
\label{rem13}
The terminology we have chosen for elliptic and parabolic polar spaces is motivated by the terminology in use for quadrics over finite fields of characteristic $2.$ Indeed, if $\KK=\mathrm{GF}(2^r)$ and $\cQ$ is a non-degenerate orthogonal polar space in $\PG(V)$, $V=V(n,\KK)$, described by a quadratic form $\phi$ with bilinearization $f$, then we have only two possibilities for $\dim(\mathrm{Rad}(f))$, i.e. $\dim(\mathrm{Rad}(f))=0$ or $\dim(\mathrm{Rad}(f))=1.$ In this case, the anisotropic space $V_0$ has dimension at most $2.$ If $\dim(\mathrm{Rad}(f))=0$ and $V_0=0$ then $\cQ$ is usually called {\it hyperbolic quadric}; if $\dim(\mathrm{Rad}(f))=0$ and $\dim(V_0)=2$ then $\cQ$ is usually called {\it elliptic quadric}; if $\dim(\mathrm{Rad}(f))=1$ and $V_0=\mathrm{Rad}(f)$ then $\cQ$ is usually called {\it parabolic quadric}.
When $\mathrm{char}(\KK)\neq 2$ our terminology is not the same as
what is usually found in literature for quadrics.
In our case, \emph{all} orthogonal polar spaces are either
hyperbolic (i.e. generated by a frame) or elliptic and
their anisotropic gap is the same as their elliptic gap.
It is worth to mention that if $\KK$ is a non-perfect field of characteristic $2$, then there can exist non-degenerate $(\fe,\fp)$-orthogonal polar spaces which are neither hyperbolic, nor elliptic or parabolic, i.e. for them $0\subset\mathrm{Rad}(f)\subset V_0$ and both $\fe,\fp>0$.
Note that if $\cQ$ has $\dim(\mathrm{Rad}(f))=[\KK:\KK^2]<\infty$, where $\KK^2=\{a^2\colon a\in \KK\}$, then $\cQ$ is parabolic, see Corollary~\ref{c:cpp}, but the converse is not true, i.e. there exist parabolic spaces (i.e. such that $\mathrm{Rad}(f)=V_0$) with $\dim(\mathrm{Rad}(f))<[\KK:\KK^2].$
\end{remark}
\paragraph{Structure of the paper}
In Section~\ref{Prelim} we will give some basics on embeddable polar spaces recalling the definitions and the relevant results from the literature. We will also set the notation.
In Section~\ref{hyperb lines} we will focus on hyperbolic lines of $\cP$ thus proving Theorem~\ref{thm hyperb lines}. In Section ~\ref{sec 3} we will characterize the elliptic gap and we will prove Theorem~\ref{m:1}.
In Section~\ref{parabolic gap} we will focus on the parabolic gap proving Theorem~\ref{m:2} and we will characterize the parabolic polar spaces.
\section{Preliminaries}\label{Prelim}
Let $\cP =(P, L)$ be a non-degenerate polar space and let $V$ be a vector space
over some division ring $\KK$.
A \emph{projective embedding} of $\cP$ is an injective map
$\varepsilon:\cP\rightarrow \PG(V)$ with the property that $\langle \eps(P)\rangle=\PG(V)$
and every line of $\cP$ is mapped onto a projective line of $\PG(V)$. So, $\varepsilon(L):=\{\varepsilon(\ell)\}_{\ell\in L}$ is a set of lines of $\PG(V)$ and $\varepsilon(\cP) := (\varepsilon(P), \varepsilon(L))$ is a full subgeometry of (the point-line space of) $\PG(V)$.
If $\eps_1:\cP \rightarrow \PG(V_1)$ and $\eps_2:\cP \rightarrow \PG(V_2)$ are two projective embeddings of $\cP$
we say that $\eps_1$ \emph{covers} $\eps_2$ (in symbols, $\eps_2\leq \eps_1$) if $V_1$ and $V_2$ are defined over the same division ring $\KK$ and there exists a $\KK$-semilinear mapping $\pi\colon V_1\rightarrow V_2$ such that $\eps_2= \pi\circ \eps_1$ (actually, writing $\pi\circ\eps_1$ is an abuse, since morphisms of projective spaces are involved here rather than their underlying semilinear maps, but this is a harmless abuse).
The map $\pi$ such that $\eps_2=\pi\circ\eps_1$, if it exists, is unique up to
rescaling. It is called the projection of $\eps_1$ onto $\eps_2$.
We say that $\eps_1$ and $\eps_2$ are
\emph{equivalent} (in symbols $\eps_1\simeq\eps_2$) if $\eps_1\leq \eps_2\leq \eps_1$; this is the same as to say that the projection $\pi:V_1\to V_2$ of $\eps_1$ onto $\eps_2$ is an isomorphism.
An embedding $\tilde{\varepsilon}$ is said to be \emph{relatively universal} if $\bar{\varepsilon}\geq\tilde{\varepsilon}$ implies
$\bar{\varepsilon}\simeq\tilde{\varepsilon}$.
Note that every embedding $\varepsilon$ is covered by a relatively universal embedding (Ronan \cite{Ron}), unique modulo equivalence. This relatively universal embedding is called the \emph{hull} of $\varepsilon$.
An embedding $\tilde{\varepsilon}$ of $\cP$ is \emph{absolutely universal} if it covers all projective embeddings of $\cP$. Clearly, when it exists, the universal embedding is unique up to equivalence and it is the hull of all embeddings of $\cP$. In this case all embeddings of $\cP$ are necessarily defined over the same division ring.
The \emph{embedding rank} $\er(\cP)$ of $\cP$ is the least upper bound of the dimensions of the embeddings of $\cP$ (the maximal dimension of an embedding of $\cP$ when all these dimensions are finite and range in a finite set). Suppose that $\cP$ admits the absolutely universal embedding, say $\tilde{\varepsilon}$. Then $\dim(\tilde{\varepsilon}) = \er(\cP)$ and, for any embedding $\varepsilon$ of $\cP$, if $\er(\cP) < \infty$ then $\varepsilon\simeq\tilde{\varepsilon}$ if and only if $\dim({\varepsilon})=\er(\cP)$.
As proved by Tits \cite[chp. 7-9]{T} (also Buekenhout and Cohen \cite[chp. 7-11]{BC}), all thick-lined non-degenerate polar spaces of rank at least $3$ are embeddable except for the following two exceptions, both of rank $3$: the line-grassmannian of $\PG(3,\KK)$ with $\KK$ non-commutative and a family of polar spaces of rank $3$ with non-desarguesian planes, described in \cite[chp. 9]{T} (also Freudenthal \cite{Fr}), which we call {\em Freudenthal-Tits polar spaces}; see also~\cite{P-VM} for this geometry.
Let $\cP$ be now an embeddable non-degenerate polar space of finite rank $n \geq 2$. By Tits~\cite[chp. 8]{T}, the polar space $\cP$ admits the universal embedding but for the following two exceptions of rank $2$:
\begin{enumerate}[(E1)]
\item\label{E1} $\cP$ is a grid with lines of size $s+1 > 4$, where $s$ is a prime power if $s < \infty$. If $\varepsilon:\cP\rightarrow\PG(V)$ is an embedding of $\cP$, then $V \cong V(4,\KK)$ for a field $\KK$ and $\varepsilon(\cP)$ is a hyperbolic quadric of $\PG(V)$. The field $\KK$ is uniquely determined by $\cP$ only if $s < \infty$.
\item\label{E2} $\cP$ is a generalized quadrangle admitting just two non-isomorphic embeddings $\varepsilon_1, \varepsilon_2:\cP\rightarrow \PG(3,\KK)$ for a quaternion division ring $\KK$ (the same for $\varepsilon_1$ and $\varepsilon_2$). We refer to \cite[8.6]{T} for more on these examples
which we call \emph{bi-embeddable quaternion quadrangles}.
\end{enumerate}
Many polar spaces admit just a unique embedding, which is the absolutely universal one. The grids of case (E\ref{E1}) admit several
non-equivalent embeddings which are all relatively universal and
each of them is described by a quadratic form; as grids are generated
by frames, these turn out to be $(0,0)$-orthogonal polar spaces.
The bi-embeddable quaternion quadrangles of case (E\ref{E2}) are
generated by frames but their hyperbolic lines contain more than $2$
points, so they do not play any further role in our study here.
We recall that in characteristic $2$ there are more possibilities
for the embeddings of polar spaces (see Section~\ref{sesquilinear and pseudoquadratic form}).
\subsection{Sesquilinear and pseudoquadratic forms}\label{sesquilinear and pseudoquadratic form}
\label{s2}
Sesquilinear and pseudoquadratic forms are essential in order to describe
the projective embeddings of polar spaces. In particular, the universal embedding of a polar space
(when it exists) can always be described by either a sesquilinear or a quadratic form.
Given an anti-automorphism $\sigma$ of $\KK$, a $\sigma$-{\em sesquilinear form} is a function $f:V\times V\to\KK$ such that $f(\sum_ix_it_i, \sum_jy_js_j) = \sum_{i,j}t_i^\sigma f(x_i, y_j)s_j$ for any choice of vectors $x_i, y_j \in V$ and scalars $t_i, s_j \in \KK$. A sesquilinear form $f$ is said to be {\em reflexive} when for any two vectors $x, y\in V$ we have $f(x,y) = 0$ if and only if $f(y,x) = 0$.
Let $f$ be a non-trivial (i.e. not null) $\sigma$-sesquilinear form.
Then $f$ is reflexive if and only if there exists a scalar $\epsilon\in \KK^\ast$ such that $f(y,x) = f(x,y)^\sigma\epsilon$ for any choice of $x, y\in V$; this condition forces $\epsilon^\sigma = \epsilon^{-1}$ and $t^{\sigma^2} = \epsilon t\epsilon^{-1}$ for any $t\in \KK$ (see Bourbaki \cite{Bour}). With $\epsilon$ as above, $f$ is called a $(\sigma,\epsilon)$-{\em sesquilinear form}. Clearly, $\epsilon \in \{1, -1\}$ if and only if $\sigma^2 = \mathrm{id}_\KK$; also, $\sigma = \mathrm{id_\KK}$ only if $\KK$ is commutative.
Let $\sigma= \mathrm{id}_\KK$. If $\epsilon = 1$ then $f$ is said to be {\em symmetric}; when $\mathrm{char}(\KK) \neq 2$ then $\epsilon = -1$ if and only if $f(x,x) = 0$ for any $x\in V$. In this case $f$ is said to be {\em alternating}. When $\mathrm{char}(\KK) = 2$, an {\em alternating} form is a $(\mathrm{id}_{\KK}, 1)$-form $f$ such that $f(x,x) = 0$ for every $x\in V$. A $(\sigma, \epsilon)$-form with $\sigma\neq\mathrm{id}_\KK$ and $\epsilon = 1$ (or $\epsilon = -1$) is called {\em hermitian} (respectively {\em antihermitian}).
Two vectors $v, w\in V$ are {\em orthogonal} with respect to $f$ (in symbols $v\perp_f w$) if $f(v,w) = 0$. A vector $v\in V$ is {\em isotropic} if $v\perp_f v$. A subspace $X$ of $V$ is {\em totally isotropic} if $X\subset X^{\perp_f}$. In contrast, if $\mathbf{0}$ is the unique isotropic vector of a subspace $X$ of $V$ then we say that $f$ is {\em anisotropic} over $X$ and that $X$ is anisotropic for $f$. The same terminology is adopted for points and subspaces of $\PG(V)$, in an obvious way. The subspace $\mathrm{Rad}(f) = V^{\perp_f}$ is the {\em radical} of $f$. The form $f$ is {\em degenerate} if $\mathrm{Rad}(f) \neq \{\mathbf{0}\}$.
The isotropic points of $\PG(V)$ together with the totally isotropic lines of $\PG(V)$ form a polar space $\cP(f)$. The singular subspaces of $\cP(f)$ are precisely the totally isotropic subspaces of $\PG(V)$ and the radical of $\cP(f)$ is the (subspace of $\PG(V)$ corresponding to) $\mathrm{Rad}(f)$. In particular, $\cP(f)$ is non-degenerate if and only if $f$ is non-degenerate.
If we are interested in the polar space $\cP(f)$ rather than in peculiar properties of the underlying form $f$, we can always assume that $f$ is either alternating, symmetric or hermitian (or antihermitian, if we prefer).
Lets us turn to pseudoquadratic forms.
Let $\sigma$ and $\epsilon$ be as above but with $\epsilon\neq-1$ when $\sigma = \mathrm{id}_\KK$ and $\mathrm{char}(\KK)\neq 2$. Put
$\KK_{\sigma,\epsilon}:= \{t-t^{\sigma}\epsilon\colon t\in\KK\}$. The set $\KK_{\sigma, \epsilon}$ is a subgroup of the additive group of $\KK$. Moreover $\KK_{\sigma,\epsilon}\subset \KK$ in view of the hypotheses we have assumed on $\sigma$ and $\epsilon$. Hence the quotient group $\KK/\KK_{\sigma,\epsilon}$ is non-trivial.
A function $\phi:V\to\KK/\KK_{\sigma,\epsilon}$ is a \emph{$(\sigma,\varepsilon)$-pseudoquadratic
form} if there exists a $\sigma$-sesquilinear form $g:V\times V\to\KK$ such that $\phi(x)=g(x,x)+\KK_{\sigma,\varepsilon}$ for
all $x\in V$.
The sesquilinear form $g$ is not uniquely determined by $\phi$ and needs not
to be reflexive. In contrast, the form $f_\phi:V\times V\rightarrow \KK$ defined by the clause $f_\phi(x,y) := g(x,y) + g(y,x)^\sigma\epsilon$ only depends on $\phi$ and is $(\sigma,\epsilon)$-sesquilinear (hence reflexive). Moreover $\phi(x+y) = \phi(x) + \phi(y) + (f_\phi(x,y) + \KK_{\sigma,\epsilon})$ for any choice of $x, y \in V$. We call $f_\phi$ the {\em sesquilinearization} of $\phi$.
In general there is no way to recover $\phi$ from its sesquilinearization $f_\phi$. However, suppose that the center of $\KK$ contains an element $\mu$ such that $\mu+\mu^{\sigma}\neq 0$ (as it is always the case when $\mathrm{char}(\KK) \neq 2$: just choose $\mu = 1$). Then
\[\phi(x) ~ = ~ \frac{\mu}{\mu+\mu^{\sigma}}\cdot f_\phi(x,x) + \KK_{\sigma,\epsilon}.\]
A vector $v\in V$ is \emph{singular} for $\phi$ if $\phi(v) = 0$. A subspace $X$ of $V$ is \emph{totally singular} if all of its nonzero vectors are singular while, if $\mathbf{0}$ is the unique singular vector of $X$, then we say that $\phi$ is {\em anisotropic} over $X$. The same terminology is used for points and subspaces of $\PG(V)$. The singular points of $\PG(V)$ together with the totally singular lines of $\PG(V)$ form a polar space $\cP(\phi)$. The singular subspaces of $\cP(\phi)$ are precisely the totally singular subspaces of $\PG(V)$. If the singular points of $\phi$ span $V$ (as it is always the case when $\phi$ admits at least one singular vector not in $\mathrm{Rad}(\phi)$) then the inclusion mapping of $\cP(\phi)$ in $\PG(V)$ is a projective embedding.
The set of singular vectors of $\mathrm{Rad}(f_\phi)$ is a subspace of $V$. It is called the {\em radical} of $\phi$ and denoted by $\mathrm{Rad}(\phi)$. The form $\phi$ is {\em non-degenerate} if $\mathrm{Rad}(\phi) = \{\mathbf{0}\}$, namely $\phi$ is anisotropic over $\mathrm{Rad}(f_\phi)$. The radical of the polar space $\cP(\phi)$ is (the subspace of $\PG(V)$ corresponding to) $\mathrm{Rad}(\phi)$. So, $\cP(\phi)$ is non-degenerate if and only if $\phi$ is non-degenerate.
All vectors that are singular for $\phi$ are isotropic for $f_\phi$ and the span $\langle v, w\rangle$ of two singular vectors $v, w\in V$ is totally singular if and only if $v\perp_{f_\phi} w$. It follows that $\cP(\phi)$ is a subspace of $\cP(f_\phi)$, but possibly different from $\cP(f_\phi)$. In particular, it can happen that $\cP(\phi)$ is non-degenerate and $\cP(f_\phi)$ is degenerate. However, when $\phi$ can be recovered from $f_\phi$ then $\cP(\phi) = \cP(f_\phi)$. If this is the case, then all we can do with $\phi$ can be done with $f_\phi$ as well.
When focusing on $\cP(\phi)$ regardless of peculiar properties of the form $\phi$, we can always assume that $\phi$ is $(\sigma,1)$-pseudoquadratic.
A $(\sigma,1)$-pseudoquadratic form is called \emph{quadratic} or
\emph{hermitian} according to whether $\sigma=\mathrm{id}_{\KK}$ or
$\sigma\neq\mathrm{id}_{\KK}$.
\medskip
Let now $\varepsilon:{\cP}\rightarrow \PG(V)$ be the universal embedding of $\cP$, where $V = V(N,\KK)$ and $\KK$ a division ring.
By Tits \cite[Chapter 8]{T}, $\varepsilon(\cP)=\cP(f)$ for a non-degenerate sesquilinear form $f$ or $\varepsilon(\cP)=\cP(\phi)$ for a non-degenerate pseudoquadratic form $\phi$.
By Tits \cite[Chapter 8]{T}, since $\varepsilon$ is universal, non-degenerate and trace-valued, $f$ cannot be alternating when $\ch(\KK) = 2$.
If either $\ch(\KK)\neq 2$ or $\ch(\KK)=2$ but $\sigma$ acts non-trivially
on the center of $\KK$ (more generally, $\phi$ is uniquely determined by
$f_{\phi}$), then $\varepsilon$ is the unique embedding of $\cP$.
Suppose $\ch(\KK) = 2$ and $\eps(\cP)=\cP(\phi)$. Then $R := \mathrm{Rad}(f_\phi)$ may be non-trivial. Suppose $R \neq \{\mathbf{0}\}.$ However no nonzero vector of $R$ is singular for $\phi$, since $\phi$ is non-degenerate. Hence $R$ contains no point of $\cP(\phi)$. Moreover, every projective line of $\PG(V)$ containing at least two $\phi$-singular points misses $R$. Therefore, for every subspace $X$ of $R$, the projection $\pi_X:\PG(V)\rightarrow \PG(V/X)$ induces an injective mapping on $\cP(\phi)$. Accordingly, the composite $\varepsilon_X = \pi_X\circ\eps$ is an embedding of $\cP$. All embeddings of $\cP$ arise in this way, by factorizing $V$ over subspaces of $R$. The embedding $\varepsilon_R$, obtained by factorizing $V$ over $R$, is the minimal one.
All embeddings $\varepsilon_X$ (included the ones above where $\KK_{\sigma,1}=\{0\}$) can be described by means of {\em generalized pseudoquadratic forms} (see \cite{P}), except the minimal one when $\phi(R) = \KK/\KK_{\sigma,1}$. If $\phi(R) = \KK/\KK_{\sigma,1}$ then $\sigma = \mathrm{id}_\KK$, the form $\phi$ is quadratic and it induces the null form on $V' := V/R$ and $\varepsilon_R(\cP) = \cP(f_\phi)$ for a non-degenerate alternating form $f:V'\times V'\rightarrow \KK$ (hence $\dim(V') = 2n$). Accordingly, if $\ch(\KK) = 2$ and $\cP$ admits an embedding $\eps':\cP\rightarrow\PG(V')$ such that $\eps'(\cP)$ is a symplectic variety of $\PG(V')$, then the universal embedding $\eps:\cP\rightarrow\PG(V)$ embeds $\cP$ as a quadric $\cP(\phi)$ of $\PG(V)$ and $V' = V/R$. Moreover, if
$\cP$ has finite rank $n$, then $\dim(\eps) = 2n+\fd$ ($= \fd$ when $\fd$ is infinite) where $\fd$ is the degree of $\KK$ over its subfield $\KK^2 = \{t^2\}_{t\in \KK}$ (see \cite{P}). In particular, if $\KK$ is perfect then
$\fd=1$.
\subsection{Subspaces of a polar space}
Suppose $\cP$ is an embeddable non-degenerate polar space of finite rank $n\geq 2$ and let $\varepsilon:{\cP}\rightarrow \PG(V)$ be an embedding of $\cP$.
For any $Z\leq V$, we shall denote by $[Z]$ the projective subspace
of $\PG(V)$ with underlying vector space $Z$.
If $\cS$ is a subspace of $\cP$, we say that $\cS$ \emph{arises} from $\eps$ if $\cS=\varepsilon^{-1}([X])$, where $[X]$ is a (projective) subspace of $\PG(V).$
In ~\cite{ILP21a} we proved the following (see also \cite[Remark 2.1]{P22}
for case (E\ref{E2})):
\begin{theorem}\label{thm subspaces}
All subspaces of non-degenerate rank at least $2$ of an embeddable polar space
of finite rank $n\geq 2$ arise from at least one of its embedding
(hence
from its universal embedding if the polar space is neither a grid as in
(E\ref{E1}) nor a bi-embeddable quaternion quadrangle as in (E\ref{E2})).
\end{theorem}
When $\cP$ is a grid or a bi-embeddable quadrangle (mentioned at the beginning of Section~\ref{sesquilinear and pseudoquadratic form}
there
are no proper subspaces of non-degenerate rank $2$, so the statement of the
theorem is empty, hence trivially true, in those cases.
In the Introduction we already defined a frame of $\cP$ as a pair $F=\{A,B\}$ where $A$ and $B$ are mutually disjoint sets of points of $\cP$, both of size $n$, such that $A\subseteq A^\perp$, $B\subseteq B^\perp$ and
for any $a\in A$ there exists a unique $b\in B$ such that $a\not\perp b$,
and conversely for any $b\in B$ there exists a unique $a\in A$ with
$b\not\perp a$.
We say that a subspace of $\cP$ is \emph{nice} if it contains a frame of
$\cP$.
In particular any subspace of $\cP$ containing a frame
is non-degenerate and
has the same rank as $\cP$.
The following property is immediate yet useful.
\begin{remark}
\label{nice hyperplane}
No frame of $\cP$ is contained in the perp of a point; consequently,
no nice subspace is contained in a singular hyperplane of $\cP$.
\end{remark}
Let $N({\cP})$ be the poset of the nice subspaces of $\cP$,
ordered by inclusion.
Following
the notation of the Introduction, we shall denote by $\fN(\cP)$
the family of all well-ordered chains of elements of $N(\cP)$, again
ordered by inclusion.
Clearly, $\cP$ is the greatest element of $N({\cal P})$ and the minimal elements of ${N}({\cal P})$ are exactly the subspaces spanned
by the frames of $\cP$.
The poset $N(\cP)$ always contains finite chains and, trivially, every finite chain is well ordered; so $\fN(\cP)\neq\emptyset$.
This being said, in general, not all chains of ${N}(\cP)$ are well ordered. However, when ${N}(\cP)$ has finite length, namely all of its chains are finite,
then all chains of ${N}(\cP)$ are well ordered.
\begin{lemma}
\label{ln}
Let $\cP$ be a non-degenerate embeddable polar space of
rank at least $2$ with relatively universal embedding $\varepsilon$ and let $\cX$
be a nice subspace of $\cP$. Put $[X]=\langle\varepsilon(\cX)\rangle$.
\begin{enumerate}
\item
If $\cX$ is not generated by a frame
then there
is $Y<X$ such that $\varepsilon^{-1}(Y)$ is nice, $\dim(X/Y)=1$ and
$\langle\varepsilon(\varepsilon^{-1}([Y]))\rangle=[Y]$.
\item
If $\cX\subset\cP$ is a proper nice subspace of $\cP$,
then there is $Z$ such that $X<Z$, $\dim(Z/X)=1$ and $\langle\varepsilon(\varepsilon^{-1}([Z]))\rangle=[Z]$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $(p_1,p_1',\dots,p_n,p_n')$ be a frame $\cF$ of $\cP$ contained in $\cX$. Then $X$ admits a basis
$B:=(\varepsilon(p_1),\varepsilon(p_1'),\dots,\varepsilon(p_n),\varepsilon(p_n'),\varepsilon(e_1),\dots,\varepsilon(e_{\lambda-1}),
\varepsilon(e_{\lambda}))$.
\begin{enumerate}
\item
Put $Y:=\langle
\varepsilon(p_1),\varepsilon(p_1'),\dots,\varepsilon(p_n),\varepsilon(p_n'),\varepsilon(e_1),\dots,\varepsilon(e_{\lambda-1})\rangle$.
By construction $\dim(X/Y)=1$ and $\cY:=\varepsilon^{-1}(Y)$ is nice, as it contains the frame $\cF$.
Clearly, $\cY\neq\cX$ (as $e_{\lambda}\in\cX\setminus\cY$), so $\cY$ is a proper hyperplane of $\cX$ and, by construction,
$\langle\varepsilon(\cY)\rangle=[Y]$.
\item
Suppose now $\cX$ is a proper nice subspace of $\cP$ and $V$ is the space hosting $\eps(\cP)$. Since $\cX$ arises from the embedding $\eps$
we have $X\neq V$; so there is at least one point $z\in\cP\setminus\cX$ such that $\varepsilon(z)\not\in X$. Put $Z=\langle X,\varepsilon(z)\rangle$.
\end{enumerate}
The thesis follows in both cases.
\end{proof}
The following lemma, to be used in Section~\ref{parabolic gap}, shows that for any nice subspace $\cX$ of $\cP$ there is at least one maximal well ordered
chain of nice subspaces of $\cP$ which has $\cX$ as an element.
\begin{lemma}
\label{parabolic defect}
Let $\cP$ be a non-degenerate embeddable polar space of rank at least $2$ and let $\cX$
be a nice subspace of $\cP$. Then there is a maximal chain
of nice subspaces of $\cP$ having $\cX$ as an element.
\end{lemma}
\begin{proof}
We may assume that $\cP$ is neither a grid nor a bi-embeddable quaternion
quadrangle, otherwise $\cP$ is its unique nice subspace and there is nothing
to prove.
By Theorem~\ref{thm subspaces}, all nice subspaces of $\cP$
arise from the universal embedding of $\cP$. Let $\cF$ be a frame
of $\cX$ and put $[F]:=\langle\varepsilon(\cF)\rangle$.
Let also $[X]:=\langle\varepsilon(\cX)\rangle$.
Let $B_0=(\mathbf{f}_1,\dots,\mathbf{f}_{2n})$ be a basis of $F$
where $[\mathbf{f}_i]=\varepsilon(f_i)$ with $f_i\in\cF$.
As $[X]$ is spanned by $\varepsilon(\cX)$, it is possible to
complete $B_0$ to a basis of $X$ by adding vectors from
$B_1:=(\mathbf{x}_i)_{i<\delta}$ where $\delta$ is a suitable ordinal number and again with
$[\mathbf{x}_i]=\varepsilon(x_i)$ with $x_i\in\cX$.
Finally, the basis $B_0\cup B_1$ can be completed to a basis of
$V$ by adding further vectors from
$B_2=(\mathbf{v}_j)_{j<\psi}$
with $[\mathbf{v}_i]=\varepsilon(v_i)$ with $v_i\in\cP$ and $\psi$ a suitable ordinal number.
The sets $B_0,B_1$ and $B_2$ admit a well ordering by
the axiom of choice. Consequently, the set
$B=B_0\cup B_1\cup B_2$ is also well ordered if we put
each element of $B_0$ before each element of
$B_1$ and each element of $B_1$ before each element of $B_2$.
For any $\gamma<\delta$ and $\xi<\psi$, define
\[ \mathfrak V: F\subset L_1\subset L_2\dots \subset X\subset
U_1\subset U_2\subset\dots\subset V \]
be the chain of subspaces of $V$ where
\[ L_{\gamma}:=\langle F\cup \{\mathbf{x}_{i}\}_{i<\gamma}\rangle,
\qquad
U_{\xi}:=\langle X\cup\{\mathbf{v}_{j}\}_{j<\xi}\rangle.
\]
We have
$L_0=F$, $L_{\delta}=X=U_0$ and $U_{\psi}=V$.
By Lemma~\ref{ln} we can always construct $\mathfrak V$ to be a maximal chain since
for any two successive spaces, say $V_{i+1}$ and $V_i$ we can
have $\dim(V_{i+1}/V_i)=1$.
Also, the chain $\mathfrak V$ contains $X$.
Put now $\cL_i:=\varepsilon^{-1}([L_i])$
and $\cU_j:=\varepsilon^{-1}([U_j])$.
We then obtain a well-ordered chain of subspaces:
\[ \cF\subset\cL_1\subset\dots\subset\cL_{\delta}=\cX=\cU_0\subset\dots
\subset\cU_{\psi}. \]
Observe that any two terms of this chain are (by construction) different
and each of them is a hyperplane in the term which follows
(since all subspaces arise from $\varepsilon$ by Theorem
\ref{thm subspaces}).
So, this chain is maximal
and it contains $\cX$. In particular, its length is $
|\delta|+|\psi|+1$.
\end{proof}
\section{Hyperbolic lines}\label{hyperb lines}
Suppose $\cP$ to be a non-degenerate embeddable polar space of rank $n\geq2$.
For any pair of non-collinear points $a,b\in\cP,$ the set $\{a,b\}^{\perp\perp}$ is the \emph{hyperbolic line} determined by $a$ and $b.$ Clearly, $a,b\in\{a,b\}^{\perp\perp}$
for any $a,b\in\cP$ and any two points of $\{a,b\}^{\perp\perp}$ are non-collinear. In particular a hyperbolic line of $\cP$ always
contains at least $2$ points.
Let now $\varepsilon:\cP\to\PG(V)$ be the universal embedding of $\cP$
if $\cP$ is not of type (E\ref{E1}) or
(E\ref{E2}) or one of its relatively universal embeddings in the latter cases.
If $\varepsilon(\cP)$ is described by a hermitian or quadratic
form $ \phi\colon V\rightarrow \KK$, so that $\varepsilon(\cP)=:\cP(\phi)$,
then let $f:V\times V\rightarrow \KK$ be the sesquilinearization of $\phi$; when $\varepsilon (\cP)$ is described by an alternating
form, let $f$ be that form;
in either case $\perp_{f}$ is the polarity
induced by $f$ in $\PG(V)$.
For $a,b\in\cP$, put $[\mathbf{a}]=\varepsilon(a)$ and
$[\mathbf{b}]=\varepsilon(b)$, with $\mathbf{a},\mathbf{b}\in V$.
We have $a\perp b$ if and only if $[\mathbf{a}]\perp_f[\mathbf{b}]$ if and only if $f(\mathbf{a}, \mathbf{b})=0.$
More explicitly, two given points
$[\mathbf{a}],[\mathbf{b}]\in\cP(\phi)$ with
$[\mathbf{a}]\not\perp[\mathbf{b}]$ determine the hyperbolic line
of $\cP(\phi)$ consisting of
$\{[\mathbf{a}],[\mathbf{b}]\}^{\perp_f\perp_f}\cap \cP(\phi).$
So any hyperbolic line $\{a,b\}^{\perp\perp}$ of $\cP$ is in correspondence with the unique hyperbolic line $\{\varepsilon(a),\varepsilon(b)\}^{\perp_f\perp_f}\cap \cP(\phi)$ of $\cP(\phi),$ through $\varepsilon$:
\begin{equation}\label{hyperbolic lines-1}
\{a,b\}^{\perp\perp}=\varepsilon^{-1}(\{\varepsilon(a),\varepsilon(b)\}^{\perp_f\perp_f}\cap\varepsilon(\cP)).
\end{equation}
Hence, with the notation introduced before, we have
\begin{lemma}\label{lemma1}
The embedding $\varepsilon$ induces a bijection between the set of
hyperbolic lines of $\cP$ and those of $\cP(\phi)$.
\end{lemma}
If $\eps(\cP)$ is a symplectic polar space, then, necessarily,
$\ch(\KK)\neq 2$ and any hyperbolic line of $\eps(\cP)$ is a line of $\PG(V)$.
If $\eps(\cP)=\cP(\phi)$ with $\phi$ hermitian with non-degenerate
sesquilinearization $f$,
then any of its hyperbolic lines
is a subline; explicitly if $a$ and $b$ are non-collinear points of $\cP$
and the representative vectors $\mathbf{a}$ and $\mathbf{b}$ of
$\varepsilon(a)$ and $\varepsilon(b)$ are chosen in such a way that
$f(\mathbf{a},\mathbf{b})=1$, then $\{\varepsilon(a),\varepsilon(b)\}^{\perp_f\perp_f}$ is the set
$\{[\mathbf{b}]\}\cup\{ [\mathbf{a}+\mathbf{b}t] \colon t^{\sigma}+t=0\}$,
where $\sigma$ is the anti-automorphism of $\KK$ associated to $\phi$.
When $\varepsilon(\cP)=\cP(\phi)$ for a quadratic form
$\phi$ with non-degenerate bilinearization, then every hyperbolic line of
$\varepsilon(\cP)$ consists of just $2$ points.
When $\varepsilon(\cP)=\cP(\phi)$ is hermitian or quadratic with sesquilinearization $f$ and $R=\mathrm{Rad}(f)\neq\{\mathbf{0}\}$, then
every hyperbolic line $\{\varepsilon(a),\varepsilon(b)\}^{\perp_f\perp_f}$
of $\varepsilon(\cP)$ spans a $(\dim(R)+2)$--dimensional subspace
$\langle\varepsilon(a),\varepsilon(b),R\rangle$ of $V$.
\begin{lemma}
\label{lemb}
With $\cP$ and $\varepsilon$ as above,
if all the hyperbolic lines of $\cP$ contain exactly two points, then
$\varepsilon (\cP)$
is a non-degenerate
orthogonal polar space of either hyperbolic or elliptic type.
\end{lemma}
\begin{proof}
According to the notation introduced at the beginning of this section, let $a$ and $b$ be two non-collinear points of $\cP.$
By hypothesis, we have that
$\{a,b\}^{\perp\perp}=\{a,b\}.$
By Equation~\eqref{hyperbolic lines-1} it follows that
$\{[\mathbf{a}],[\mathbf{b}]\}^{\perp_f\perp_f}\cap\cP(\phi)=
\{[\mathbf{a}],[\mathbf{b}]\}$. So, by Lemma~\ref{lemma1},
we have that all hyperbolic lines of $\cP(\phi)$ contain exactly two points.
Case A: $f$ is non-degenerate. Then
for
any two distinct non-orthogonal points $[\mathbf{a}],[\mathbf{b}]\in \eps(\cP)$,
we have that $\{\mathbf{a},\mathbf{b}\}^{\perp_f\perp_f}$ spans the projective line through $[\mathbf{a}]$ and $[\mathbf{b}]$, i.e.
$\dim\{\mathbf{a},\mathbf{b}\}^{\perp_f\perp_f}=2.$
So, a projective line is a $2$-secant for $\eps(\cP)$
if and only if it is spanned by a hyperbolic line of $\eps(\cP)$
(equivalently, it is spanned by the image of a hyperbolic line of $\cP$).
Let $\ell$ be a line of $\PG(V)$ containing at least three distinct points $[\mathbf{u}],[\mathbf{v}],[\mathbf{w}]$ of $\eps(\cP)$
and suppose by way of contradiction that $\ell$ is not contained in $\eps(\cP).$
Then $\mathbf{u},\mathbf{v},\mathbf{w}$ are pairwise non-orthogonal
vectors and
the hyperbolic line determined by $\mathbf{u}$ and $\mathbf{v}$
coincides with the hyperbolic line determined by
$\mathbf{u}$ and $\mathbf{w}$, i.e. $\{[\mathbf{u}],[\mathbf{v}]\}^{\perp_f \perp_f}=\{[\mathbf{u}],[\mathbf{w}]\}^{\perp_f \perp_f}.$ This is a contradiction because the hyperbolic line
$\{u,v\}^{\perp\perp}=\varepsilon^{-1}(\{[\mathbf{u}],[\mathbf{v}]\}^{\perp_f \perp_f})$ would contain at least three points, namely $u,v,w$.
So, $\ell\subseteq \eps(\cP).$
In other words,
any projective line intersecting $\eps(\cP)$ in at least three points must be contained in $\eps(\cP)$. This property together with the fact that $\eps(\cP)$
spans $\PG(V)$, proves that $\eps(\cP)$ fulfills the definition of
a {\it Tallini set}. By~\cite[Theorem 3.8]{d02} (see also \cite[Theorem 1.1]{B69}), we have that $\varepsilon(\cP)$ is described by a non-degenerate quadratic form, i.e. $\cP$ is an orthogonal polar space.
Moreover, since we are assuming that $f$ is non-degenerate, $\eps(\cP)$ is an orthogonal polar space of either hyperbolic or elliptic type, according
to our definitions.
Case B: $f$ is degenerate.
In this case we have that $[X]:=\{[\mathbf{u}],[\mathbf{v}]\}^{\perp_f\perp_f}=[\langle \mathbf{u},\mathbf{v}\rangle+\mathrm{Rad}(f)]$ where $[\mathbf{u}]\not\perp_f
[\mathbf{v}]$ and $\dim(\mathrm{Rad}(f))\geq 1.$
Put $R:=\mathrm{Rad}(f)$.
Since $R\cap\langle\mathbf{u},\mathbf{v}\rangle=\{\mathbf{0}\}$,
we have
$\dim(\{\mathbf{u},\mathbf{v}\}^{\perp_f\perp_f})>2$, i.e. $\{[\mathbf{u}],[\mathbf{v}]\}^{\perp_f\perp_f}$
contains at least a projective plane of $\PG(V)$.
Let
$\phi_X$ be the form induced by $\phi$ on $X$.
Since $R\subseteq X$ we have $\langle \mathbf{u},\mathbf{v}\rangle\subset X$.
Also $\mathbf{u},\mathbf{v}\not\in R$ (as $\mathbf{u}\not\perp_f\mathbf{v}$)
and $\phi_X(\mathbf{u})=0=\phi_X(\mathbf{v})$. It follows by~\cite[8.2.7]{T}
that $X$ is generated by its $\phi_X$-singular vectors. On the other
hand, the singular points of $[X]$ are those of $[X]\cap\cP(\phi)$
and $|[X]\cap\cP(\phi)|=2$. This is a contradiction.
This forces $f$ not to be degenerate. By Case A, the lemma is proved.
\end{proof}
\begin{lemma}
\label{l1}
Let $\cP$ and $\varepsilon:\cP\to\PG(V)$ be as above.
If $\eps(\cP)$ is orthogonal of hyperbolic type or elliptic then every hyperbolic line of $\cP$ contains exactly $2$ points.
\end{lemma}
\begin{proof}
According to the notation introduced at the beginning of this section, since $\eps(\cP)$ is
either hyperbolic or elliptic, we have
$\mathrm{Rad}(f)=\{\mathbf{0}\}.$
Take two arbitrary distinct points $a,b$ of $\cP$ which are not collinear in $\cP$ and let $\varepsilon(a)=[\mathbf{a}]$, $\varepsilon(b)=[\mathbf{b}]$ with
$\mathbf{a},\mathbf{b}\in V$.
Then
$\phi(\mathbf{a})=\phi(\mathbf{b})=0$ and $f(\mathbf{a},\mathbf{b})\neq0$.
Since $f$ is non-degenerate,
$\dim \{\mathbf{a}, \mathbf{b}\}^{\perp_f\perp_f}=2$ and this space
contains $\{\mathbf{a},\mathbf{b}\}.$
Since $\eps(\cP)$ is a hypersurface of degree $2$, either the line $[
\langle\mathbf{a}, \mathbf{b}\rangle]$ is a $2$-secant $\cP(\phi)$ or it is contained in $\cP(\phi).$
If
$\{[\mathbf{a}],[\mathbf{b}]\}^{\perp_f\perp_f}\subset\cP(\phi)$, then
$f(\mathbf{a},\mathbf{b})=0$,
which is a contradiction for this would imply $a\perp b$ in
$\cP.$
Hence $\{[\mathbf{a}],[\mathbf{b}]\}^{\perp_f\perp_f}\cap\cP(\phi)=
\{[\mathbf{a}],[\mathbf{b}]\}$.
In particular $\{a,b\}^{\perp\perp}=\varepsilon^{-1}(\{[\mathbf{a}],[\mathbf{b}]\}^{\perp_f\perp_f}\cap\cP(\phi))=\{a,b\}$
and all hyperbolic lines of $\cP$ consist of $2$ points.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm hyperb lines}]
The theorem follows directly from Lemma~\ref{lemb} and Lemma~\ref{l1}.
\end{proof}
\begin{proof}[Proof of Corollary~\ref{pr:orth}]
By Lemma~\ref{lemb}, there exists a quadratic form $\bar{\phi}\colon \langle \varepsilon(\cF)\rangle\rightarrow \KK$ such that $\eps|_{\cF}(\cF)=\cF(\bar{\phi})$, where $\eps$ is the universal embedding of $\cP$.
As $\eps|_{\cF}$ is the restriction to $\cF$ of $\eps,$ it follows that
$\eps(\cP)$ must be described by
a suitable quadratic form $\phi$ whose restriction to $\langle \varepsilon(\cF)\rangle$
is exactly $\bar{\phi}$.
Conversely, suppose $\varepsilon(\cP)=\cP(\phi)$ with $\phi$ a non-degenerate
quadratic form. Let $F$ be a frame of $\cP$ and $\cF$ be
the subspace of $\cP$ spanned by $F$. Then,
$\varepsilon(\cF)=[\langle\varepsilon(F)\rangle]\cap\varepsilon(\cP)$
and the latter is
a hyperbolic quadric in $[\langle\varepsilon(F)\rangle]$.
In particular, all hyperbolic lines of $\varepsilon(\cF)$
consist of exactly $2$ points. By Lemma~\ref{lemma1} we now have
that $\cF$ is hyperbolic and so all of its hyperbolic lines consist of
$2$ points.
\end{proof}
\section{Elliptic gap}\label{sec 3}
Recall from Definition~\ref{dec} that an elliptic chain of a non-degenerate polar space $\cP$ of finite rank is
a well-ordered chain
\[ \fE: \cF=\cE_0\subset\cE_1\subset\dots\subset\cE_{\delta}\subset\dots\]
where $\cE_{i}\subseteq\cP$ is a nice subspace of $\cP$ with the property that any of its hyperbolic lines contains exactly $2$ points.
The following Lemma deals with arbitrary non-degenerate polar spaces.
\begin{lemma}
\label{l:hs}
Let $\cP$ be a non-degenerate polar space and $(\cS_i)_{i\in I}$
be a chain of non-degenerate
polar subspaces of $\cP$ ordered with respect to the inclusion relation $\subseteq.$
Then
$\cS_I:=\bigcup_{i\in I}\cS_i$ is a non-degenerate polar space
and $\cS_i$ is a subspace of $\cS_I$ for all $i$.
For any $a,b\in\cS_I$ with $a\not\perp_I b$ we have
\[ \{a,b\}^{\perp_I\perp_I}\subseteq
\bigcup_{i:a,b\in\cS_i}\{a,b\}^{\perp_i\perp_i} \]
where $\perp_x$ denotes the collinearity relation in $\cS_x$
for $x=I$ or $x\in I$.
\end{lemma}
\begin{proof}
By definition of subspace,
$a\perp_I b$ in $\cS_I$ if and only if for all
$i\in I$ such that $a,b\in\cS_i$ we have $a\perp_i b$.
Let $a,b\in\cS_I$ with $a\not\perp_Ib$.
Take $c\in\{a\}^{\perp_I\perp_I}$.
For any $j$ such that $a,b,c\in\cS_j$ and any $x\in\{a,b\}^{\perp_j}$
we get $x\in\{a,b\}^{\perp_I}$; so $c\perp_I x$ and consequently $c\perp_i x$;
this implies $c\in\{a,b\}^{\perp_j\perp_j}$.
\end{proof}
The following corollary is a direct consequence of the lemma.
\begin{corollary}
\label{e:max}
Let $\cP$ be a non-degenerate orthogonal polar space.
\begin{enumerate}
\item\label{max1}
If $\fE$ be a chain of elliptic subspaces of $\cP$ then $\cE':=\bigcup_{\cE}\fE$ is
either an elliptic space or it is a subspace generated by a frame.
\item Let $\cE$ be an elliptic polar subspace of $\cP$, then $\cE$ is contained in a maximal
elliptic polar subspace $\cE'$ of $\cP$.
\end{enumerate}
\end{corollary}
\begin{proof}
\begin{enumerate}
\item\label{ec1}
By Lemma~\ref{l:hs} the hyperbolic lines of $\cE'$ consists of just two
points. Then Case 1 follows from Theorem~\ref{thm hyperb lines}.
\item
Consider the set of all elliptic polar subspaces of $\cP$
containing $\cE.$ By Case~\ref{max1} every chain in it has an upper bound. Then the result now follows from Zorn's lemma.
\end{enumerate}
\end{proof}
\begin{lemma}
Let $\cP$ be a non-degenerate orthogonal
polar space either defined over a field
$\KK$ with $\ch(\KK)\neq 2$ or a hyperbolic orthogonal polar space.
Then there are maximal elliptic chains of $\cP$
and these are exactly the maximal well ordered chains of nice subspaces of $\cP$.
\end{lemma}
\begin{proof}
If $\cP$ is hyperbolic, then it is generated by a frame and
there is nothing to prove.
Suppose now $\cP$ orthogonal and $\ch(\KK)\neq2$; then
all nice subspaces $\cE$ of $\cP$ have hyperbolic lines of size
$2$ so they are \emph{elliptic} according to our definition.
It follows that
maximal elliptic chains of $\cP$ and
maximal well ordered chains of nice subspaces of $\cP$ are
the same thing.
So the lemma follows.
\end{proof}
Note that, when $\ch(\KK)=2$, a \emph{proper} (i.e. containing more than
one term) elliptic chain is not a
maximal well ordered chain of nice subspaces of $\cP.$
\begin{setting}\label{notazione}
\rm{Throughout this section we fix the following notation:
$\cP$ is a non-degenerate
embeddable polar space defined over a field of characteristic $2$
and $\varepsilon\colon \cP \rightarrow \PG(V)$, where $V$ is defined over a field $\KK$, is
either its universal embedding if $\cP$ is not a grid or
any of its relatively universal embeddings if $\cP$ is a grid
(case (E\ref{E1})).
The image of $\varepsilon$ is $\varepsilon(\cP)=\cP(\phi)$,
where
$\phi\colon V\rightarrow \KK$ is a quadratic form
having bilinearization $f\colon V\times V\rightarrow \KK$
(see Section~\ref{sesquilinear and pseudoquadratic form}).}
Observe that the bi-embeddable quaternion quadrangles mentioned in case (E\ref{E2}) are excluded because $\KK$ is a field.
\end{setting}
Suppose that $\cS$ is a proper subspace of $\cP.$ Put $[W]:=\langle\varepsilon(\cS)\rangle.$
By Theorem~\ref{thm subspaces} $\cS=\varepsilon^{-1}([W])$ and
each subspace $\cS$ of non-degenerate rank at least $2$ arises
from a subspace $[W]$ of $\PG(V)$; see~\cite{ILP21a}.
Observe that since $\cP$ is non-degenerate
of rank $n\geq 2$, all subspaces of $\cP$ containing a frame
have also non-degenerate rank $n\geq2$ so they arise from subspaces of
$\PG(V)$.
Let $R_W=\mathrm{Rad}(f_W)$ be the radical of the bilinearization
$f_W\colon W\times W\rightarrow \KK$ of the quadratic form $\phi_W\colon W\rightarrow \KK$ induced by $\phi$ on $W.$
\begin{lemma}
\label{l4}
Let $\cP$ be elliptic and $\cS$ be a
maximal proper nice subspace of $\cP$.
Then $\dim(R_W)=1$.
\end{lemma}
\begin{proof}
Suppose $F=\{F_1,F_2\}$ is a frame of $\cP$ contained in $\cS$. Hence $F_1$ and $F_2$ are maximal singular subspaces of $\cP$ and $\varepsilon(F_1)$ and $\varepsilon (F_2)$ are maximal singular subspaces of $\varepsilon(\cP).$
Since $\cS$ is a maximal subspace of $\cP,$ it is indeed a hyperplane of $\cP$ and since it contains a frame, by Remark~\ref{nice hyperplane}, we have that $\cS$ is a non-singular hyperplane of $\cP.$ Hence $[W]$ is a hyperplane of $\PG(V).$
Since $\cP$ is elliptic, $\mathrm{Rad}(f)=\{\mathbf{0}\}$ and
the polarity induced by $f$ is non-degenerate; so, for
every hyperplane $\Sigma$ of $\PG(V)$ there is a point $[\mathbf{p}]$ of $\PG(V)$ such that $\Sigma=[\mathbf{p}]^{\perp_f}$.
So, $[W]=[\mathbf{p}]^{\perp_f}$ for some point $[\mathbf{p}]\in \PG(V).$
The point $[\mathbf{p}]$ cannot be in $\varepsilon(\cP)$, otherwise the
frames of $\varepsilon(\cP)$ contained in $\varepsilon(\cE)$,
which exist since $\cS$ is nice by assumption, would be
contained in the perp $[\mathbf{p}]^{\perp}$ but no singular hyperplane
contains a frame, by Remark~\ref{nice hyperplane}.
Thus, either $[\mathbf{p}]\in\PG(V)\setminus([W]\cup\varepsilon(\cP))$ or
$[\mathbf{p}]\in [W]\setminus\varepsilon(\cP)$.
Suppose the former.
Then
$[\mathbf{p}]$ (note that $[\mathbf{p}]\not\in \varepsilon (\cS)$) is orthogonal to every point of $[W]=[\mathbf{p}]^{\perp_f}$, hence $V=W\oplus \langle\mathbf{p}\rangle.$
Take $s\in \cP.$ Then $\varepsilon (s)=[\mathbf{w}+\alpha\mathbf{p}]$
where $\mathbf{w}\in W$ and $\alpha\in \KK.$ We have
\[f(\mathbf{p},\mathbf{w}+\alpha \mathbf{p})=
f(\mathbf{p},\mathbf{w})+\alpha f(\mathbf{p},\mathbf{p})=0\]
being
$f(\mathbf{p},\mathbf{p})=0$ (recall $\ch(\FF)=2$) and $f(\mathbf{p},\mathbf{w})=0$ because
$[\mathbf{p}]\perp_f[\mathbf{w}].$
So, $\varepsilon (s)\in [\mathbf{p}]^{\perp_f}$ hence $\varepsilon (\cP)\subseteq[W].$ This is clearly not possible since $\varepsilon (\cP)$ spans $\PG(V)$ and the equality
$\cS=\varepsilon^{-1}([W])$ forces
$\cP\subseteq\cS$, a contradiction.
So, only the latter case remains to consider. Hence, $[\mathbf{p}]\in [W]=[\mathbf{p}]^{\perp_f}.$ This means that the bilinear form $f_W\colon W\times W\rightarrow \FF$ induced by $f$ on $W$ is degenerate, implying $\dim(R_W)\geq 1$.
Suppose $\dim(R_W)\geq 2.$ Then, there exist $\mathbf{r_1},\mathbf{r_2}\in
R_W$ such
that $\varepsilon(\cS)\subseteq
[\mathbf{r_1}]^{\perp_f}\cap[\mathbf{r_2}]^{\perp_f}$.
This means that the codimension of $W$ in $V$ is at least $2,$
since $f$ is non-degenerate on $V.$
So, $[W]$ is not a hyperplane, which is a contradiction.
Thus, $\dim(R_W)=1$.
\end{proof}
\begin{lemma}
\label{l5}
Let $\cP$ be orthogonal.
If $\dim(\mathrm{Rad}(f))=1$ then there exists a
nice elliptic hyperplane $\cS$ of $\cP$.
\end{lemma}
\begin{proof}
Since $\langle \varepsilon (\cP)\rangle=\PG(V)$,
there is a basis $B$ of $V$ consisting of elements
\[ B=(\mathbf{e_1},\mathbf{e_2},\dots,\mathbf{e_{\delta}},\dots) \]
with $e_1,\dots,e_{\delta },\dots\in\cP$ and $\varepsilon(e_i)=[\mathbf{e_i}]$
and such that
the first $2n$ vectors of $B$ determine a frame of $\cP$.
Let $\mathbf{r}$ be a vector such that $\mathrm{Rad}(f)=\langle\mathbf{r}\rangle$.
Let $\phi'$ be the quadratic form induced by $\phi$ on
$X:=\langle\mathbf{e}_1,\dots,\mathbf{e}_{2n}\rangle$ and $\cF(\phi')$
be the polar space defined by $\phi'$ on $[X]$.
Then $\cF(\phi')=\varepsilon(\varepsilon^{-1}([X]))$ is a hyperbolic
quadric and consequently the bilinearization $f'$ of $\phi'$ (which is
the restriction to $X\times X$ of the bilinearization $f$ of $\phi$)
is non-degenerate. Therefore $\mathbf{r}\not\in X$.
Since $B$ is a basis of $V$,
$\mathbf{r}$ can be written in an unique way as a linear
combination of a finite number of elements (at least $2$) of $B$
with non-zero coefficients, at least one of which has index $i>2n$ in
$B$; up to reordering the elements of $B$
we can assume without loss of generality
$\mathbf{r}=\mathbf{e_{2n+1}}+\sum_{i\neq 2n+1}\alpha_i\mathbf{e_i}$ where $\alpha_i\neq 0$ for at most
a finite number of indexes $i$.
Put then $\overline{V}=\langle B\setminus\{\mathbf{e_{2n+1}}\}\rangle$.
Clearly we have
\[ V=\overline{V}\oplus\mathrm{Rad}(f), \]
the space $[\overline{V}]$ is a hyperplane in $\PG(V)$ and the hyperplane of
$\cP$ given by
$\cS:=\varepsilon^{-1}([\overline{V}])$ contains a frame and is such that
$\langle\varepsilon(\cS)\rangle=[\overline{V}]$.
So, $\cS$ is a nice hyperplane of $\cP$ and by
Remark~\ref{nice hyperplane} it is non-singular.
We claim that $\cS$ is elliptic.
Given the above considerations, we already know that
$\varepsilon(\cS)=\cS(\phi|_{\overline{V}})$ and the quadratic form
$\phi|_{\overline{V}}$ is non-degenerate. It is
straightforward to see that the bilinearization of $\phi|_{\overline{V}}$ is $f|_{\overline{V}\times\overline{V}}$.
By way of contradiction suppose now that $\mathrm{Rad}(f|_{\overline{V}\times\overline{V}})$
is not trivial.
Since $\varepsilon(\cS)\subseteq[\overline{V}]$, there exists a non-null vector $\mathbf{w}\in\mathrm{Rad}(f|_{\overline{V}\times\overline{V}})\subseteq \overline{V}$ such that
$\varepsilon(\cS)\subseteq [\mathbf{w}]^{\perp_f}$. Since $\mathbf{w}\in \overline{V}$ and $\overline{V}\cap\mathrm{Rad}(f)=\{\mathbf{0}\}$,
$\mathbf{w}\not\in\mathrm{Rad}(f)$, hence $[\mathbf{w}]^{\perp_f}$ is a hyperplane of $[V].$
Also, $[\mathrm{Rad}(f)]\in [\mathbf{w}]^{\perp_f}$, which gives
$[\mathbf{w}]^{\perp_f}\neq[\overline{V}]$.
On the other hand $\varepsilon(\cS)\subseteq[\overline{V}]\cap [\mathbf{w}]^{\perp_f}$, implying that $\langle\varepsilon(\cS)\rangle$ is not a hyperplane of $\PG(V)$. Contradiction.
Hence $\mathrm{Rad}(f|_{\overline{V}\times\overline{V}})=\{\mathbf{0}\}$
and $\cS$ is elliptic.
\end{proof}
\begin{corollary}
\label{c1}
Let $\cP$ be elliptic and $\cS$ be a maximal proper nice subspace.
Then there exists a nice elliptic hyperplane of $\cS.$
\end{corollary}
\begin{proof}
By Lemma~\ref{l4}, $\dim(\mathrm{Rad}(f|_{\langle \varepsilon(\cS)\rangle\times
\langle\varepsilon(\cS)\rangle})=1.$
Now, applying Lemma~\ref{l5} to $\cS$ we have that there exists nice elliptic hyperplane of $\cS.$
\end{proof}
Corollary~\ref{c1} shows
that if $\cE_{i}$ and $\cE_{i+1}$ are both elliptic polar spaces
(or for $i=0$, $\cE_i$ is the subspace generated by a frame)
with
$\cE_{i}\subset\cE_{i+1}$ and there are no elliptic subspaces
between $\cE_{i}$ and $\cE_{i+1}$
then there exists a nice polar space $\cP_i$ sitting between $\cE_{i}$ and $\cE_{i+1}$: $\cE_{i}\subset \cP_i \subset\cE_{i+1}$ such that $\cP_i$ is a hyperplane of $\cE_{i+1}$ and
$\cE_{i}$ is a hyperplane of $\cP_i$.
Consequently, $\dim(\langle\varepsilon(\cE_{i+1})\rangle/\langle\varepsilon(\cE_{i})\rangle)=2.$
\begin{theorem}
\label{t:exists}
Any non-degenerate orthogonal polar space $\cP$ admits
well ordered maximal elliptic chains $\fE$ of subspaces.
\end{theorem}
\begin{proof}
Let $\mathbf{C}$ be the set of all well-ordered chains
$\fE=(\cE_{\gamma})_{\gamma<\omega}$ of nice subspaces of $\cP$ such that the
following hold:
\begin{enumerate}[(C1)]
\item\label{cc1} The first element of $\fE$ is generated by a frame and all the other elements $\cE_{\gamma}$ are elliptic subspaces of $\cP.$
\item\label{cc2} If $\cE_{\gamma}$ and
$\cE_{\gamma+1}$ are consecutive elements
of $\fE$, then $\langle(\varepsilon(\cE_{\gamma})\rangle$ has
codimension $2$ in $\langle\varepsilon(\cE_{\gamma+1})\rangle$.
\item\label{cc3} If $\gamma<\omega$ is a limit ordinal, then
$\bigcup_{\xi<\gamma}\cE_{\xi}=\cE_{\gamma}$.
\end{enumerate}
By construction, the set $\mathbf{C}$ is non-empty.
Suppose $\fE,\fE'\in\mathbf{C}$ with $\fE\subset\fE'$.
Let $\delta$ be such that $\cE'_{\gamma}\in\fE$ for all
$\gamma<\delta$ and $\cE'_{\delta}\in\fE'\setminus\fE$.
Using
Lemmas~\ref{l4} and \ref{l5} and conditions (C\ref{cc1}), (C\ref{cc2}) and
(C\ref{cc3})
we see that $\cE_{\mu}'=\cE_{\mu}$ for all $\mu<\delta$; so
$\fE$ is an initial segment for $\fE'$ i.e. $\fE\subseteq\fE'$ and
for all $\cE'\in\fE'\setminus\fE$ and for all $\cE\in\fE$ we
have $\cE\subseteq\cE'$.
We need to show that
the set $\mathbf{C}$ contains maximal elements with respect to
inclusion.
Let $(\fE_i)_{i\in I}$ be a chain of elements
of $\mathbf{C}$ and put $\fE_I:=\bigcup_{i\in I}\fE_i$. We first
claim that $\fE_I$ is well ordered. Indeed, let $X\subseteq\fE_I$
be non-empty. For any $\cE_0\in X$ take $i\in I$ such
that $\cE_0\in\fE_i$; clearly the minimum of $X$ cannot properly contain
$\cE_0$, so it is enough to show that
${X}_0:=\{\cE\in X:
\cE\subseteq\cE_0\}$ admits minimum. On the other hand,
${X}_0\subseteq\fE_i$ and $\fE_i$ is well ordered,
so the minimum exist.
We now prove that $\fE_I\in\mathbf{C}$. Indeed,
$\fE_I$ satisfies by construction
(C\ref{cc1}). On the other
hand if $\cE\in\fE_i$ and $\cE'\in\fE_j$ with $i,j\in I$ and
$\cE'\subseteq\cE$, then $\cE'\in\fE_i$. In particular
$\fE_i$ is an initial segment to $\fE_I$ for all $i$. Since
(C\ref{cc2}) and (C\ref{cc3}) hold in any $\fE_i$ we get
that they must also hold in $\fE_I$.
Thus, we have that $\mathbf{C}$ contains maximal chains by Zorn's Lemma.
We now prove that a maximal chain $\fE$ in $\mathbf{C}$ is also
a maximal elliptic chain.
First observe that $\cE_{\fE}:=\bigcup_{\cE\in\fE}\cE$ is an
elliptic subspace; also $\cE_{\fE}\in\fE$, since $\fE$ is
maximal. So any maximal chain in $\mathbf{C}$ has a
maximum.
Let us suppose that $\fE$ is not a maximal elliptic chain;
then there is an elliptic space $\mathcal T$ which can be
added to $\fE$ and such that $\cE_{\fE}\subset{\mathcal T}$.
By Lemma~\ref{l4}, $\cE_{\fE}$ cannot be a hyperplane of $\mathcal T$.
Let $\langle\varepsilon(\cT)\rangle=[T]$. Then
$[W'']:=\langle\varepsilon(\cE_{\fE})\rangle$ has codimension at least $2$
in $[T]$. We show that under these assumptions there would be
an elliptic subspace $\cE$ such that
$\cE_{\fE}\subset\cE\subseteq{\mathcal T}$ and that
$[W'']$ has exactly codimension $2$ in $\langle\varepsilon(\cE)\rangle$.
This would imply ${\mathfrak G}:=\fE\cup\{\cE\}\in{\mathbf C}$ and
${\fE}\subset{\mathfrak G}$
against the maximality of ${\fE}$ leading to a contradiction.
So, let $W'\leq T$ be a subspace such that $W''$ is a hyperplane
in $W'$. Since $\cE_{\fE}$ is elliptic, by Lemma~\ref{l4} the form
$f_{W'}:W'\times W'\to\KK$ is degenerate with
$\mathrm{Rad}(f_{W'})\cap W''=\{\mathbf{0}\}$. Thus,
$\dim(\mathrm{Rad}(f_{W'}))=1$ and
$\mathrm{Rad}(f_{W'})=\langle\mathbf{r}\rangle$ for a
suitable $\mathbf{r}\in W'\setminus W''$. On the other hand,
$\mathcal T$ is elliptic; so $\mathbf{r}^{\perp_f}$ is a hyperplane of
$T$ and since $\varepsilon(\mathcal{T})$ spans $[T]$ there is
$\mathbf{p}\not\in\mathbf{r}^{\perp_f}$ which is singular for
$\psi_T$. Let $W=\langle W',\mathbf{p}\rangle$ and
$\cE:=\varepsilon^{-1}([W])$. We claim that $f_W$ is non-degenerate
(and consequently $\cE$ is elliptic). Suppose
$\mathrm{Rad}(f_W)\neq\{\mathbf{0}\}$. If $\mathrm{Rad}(f_w)\subseteq W'$,
then
$\mathrm{Rad}(f_W)\subseteq\mathrm{Rad}(f_{W'})=\langle\mathbf{r}\rangle$;
this would imply $\mathbf{r}\perp\mathbf{p}$, against the assumption on
$\mathbf{p}$. So $\mathrm{Rad}(f_W)\not\subseteq W'$ and there is
$\mathbf{x}\in W'$ such that $\mathbf{x}+\mathbf{p}\in\mathrm{Rad}(f_W)$.
It follows $(\mathbf{x}+\mathbf{p})\perp\mathbf{r}$; on the other hand
$\mathbf{r}\perp\mathbf{x}$ since $\mathbf{r}\in\mathrm{Rad}(f_{W'})$
and $\mathbf{x}\in W'$. This implies again $\mathbf{r}\perp\mathbf{p}$
which is a contradiction. This proves the theorem.
\end{proof}
\begin{lemma}
\label{lem pur elliptic}
Let $\cP$ have rank $n$ and anisotropic gap $2\fd$
such that all hyperbolic lines of $\cP$ have cardinality $2.$
Then there exists a maximal elliptic chain
\begin{equation}
\label{e2}
\fE:\cF=\cE_0\subset\cE_1\subset\dots\subset\cE_{\delta},
\end{equation}
of length $\fd=|\delta|+1$.
Moreover, the chain $\fE$ can be enriched to a maximal chain
$\fE^E$
of length $2\fd$
of nice subspaces of $\cP$ as follows
\begin{equation}
\label{eqn}
\fE^E:\cF=\cE_0\subset\cP_0\subset\cE_1\subset \dots \subset \cE_i\subset \cP_i\subset \cE_{i+1}\subset \dots \subset
\cE_{\delta}.
\end{equation}
\end{lemma}
\begin{proof}
By Theorem~\ref{thm hyperb lines}, $\cP$ is either a hyperbolic or an elliptic orthogonal polar space.
If $\cP$ is of hyperbolic type than $\fd=0$ and there is nothing to prove.
Suppose $\cP$ to be an elliptic polar space.
Then $\dim(V)=2n+2\fd$.
By Theorem~\ref{t:exists}, there are maximal well-ordered chains of
elliptic polar spaces in $\cP$ and their maximum element is $\cP$ itself.
Using Corollary~\ref{c1}, we can
enrich these chains to
well-ordered chains of nice subspaces of $\cP$
like in~\eqref{eqn} where the $\cE_{i}$ are all elliptic,
and the subspaces $\varepsilon (\cP_i)$ are described
by a quadratic form having bilinearization with radical of dimension $1.$ Moreover, $\cE_{i}$ is a hyperplane
in $\cP_i$ and $\cP_i$ is a hyperplane in $\cE_{i+1}$.
As a hyperplane of a polar subspace is indeed a maximal subspace,
it follows that
the chain~\eqref{eqn} is a maximal well ordered chain of nice subspaces
of $\cP$ with $2\fd$ terms.
\end{proof}
We keep on using the notation of Setting~\ref{notazione}.
\begin{lemma}
\label{lem:key}
Let $\cE$ be a nice subspace of $\cP$ maximal with respect to the property that all its hyperbolic lines contain exactly two points (i.e. $\cE$ is a maximal elliptic subspace).
Put $[T]:=\langle\varepsilon(\cE)\rangle$.
Then, $T$ is a direct complement
of $\mathrm{Rad}(f)$ in $V$, that is
\[ V:=T\oplus \mathrm{Rad}(f). \]
\end{lemma}
\begin{proof}
By Corollary~\ref{e:max} $\cP$ contains maximal subspaces $\cE$ with the
required properties. In light of
Theorem~\ref{thm hyperb lines}, $\varepsilon(\cE)$ is
either elliptic or hyperbolic. Then $\mathrm{Rad}(f_T)=\{\mathbf{0}\}$, where $f_T\colon T\times T\rightarrow \KK$ is the bilinearization of the quadratic form $\phi_T\colon T\rightarrow \KK$ describing $\varepsilon (\cE)$, induced by $\phi.$
In particular $T\cap\mathrm{Rad}(f)\subseteq \mathrm{Rad}(f_T)=\{\mathbf{0}\}.$
Let us consider the subspace $\langle T, \mathrm{Rad}(f)\rangle\subseteq V.$
Suppose that $\langle T, \mathrm{Rad}(f)\rangle\neq V.$
Then there exists a non-null vector $\mathbf{w}\in V\setminus \langle T, \mathrm{Rad}(f)\rangle$ (equivalently $\mathbf{w}+\langle T, \mathrm{Rad}(f)\rangle\in V/\langle T, \mathrm{Rad}(f)\rangle$).
As $[V]$ is spanned by $\cP(\phi)$, we can assume
without loss of generality $\phi(\mathbf{w})=0$. Put $T':=\langle T,\mathbf{w}\rangle$. Note that $T'\cap\mathrm{Rad}(f)=\{\mathbf{0}\}$,
since otherwise $T'\subseteq\langle T,\mathrm{Rad}(f)\rangle$ which would
be a contradiction.
Let now $\cE':=\varepsilon^{-1}([T'])$. Since $\cE\subseteq\cE' $ and
$\cE$ is a maximal elliptic or hyperbolic subspace of $\cP$, the space
$\cE'$ can be neither elliptic nor hyperbolic. Hence, the radical of the bilinear form $f_{T'}$ induced by $f$ on $T'\times T'$ is not null,
so there exists a non-null vector $\mathbf{u}\in\mathrm{Rad}(f|_{T'})$, i.e. a
vector $\mathbf{u}$ such that
$f(\mathbf{u},\mathbf{t})=0$ for all $\mathbf{t}\in T'$.
Since
$T'\cap\mathrm{Rad}(f)=\{\mathbf{0}\}$, there is at least one element
$\mathbf{v}\notin \mathbf{u}^{\perp_f}$ which is $\phi$-singular by~\cite[8.2.7]{T}
Put now $T'':=\langle T,\mathbf{u}, \mathbf{v}\rangle$ and $\cE'':=\varepsilon^{-1}([T'']).$ We have $\cE\subset \cE'\subset \cE''.$
We show that
$\cE''$ is elliptic or hyperbolic, thus contradicting the maximality of $\cE.$
Indeed, suppose that $\cE''$ is neither elliptic nor hyperbolic, so $\mathrm{Rad}(f|_{T''})\not=\{\mathbf{0}\}.$
Then, there exists a non-null vector $\mathbf{y}\in T''$ such that $f(\mathbf{y},\mathbf{x})=0$ for all $\mathbf{x}\in T''$.
Clearly, it cannot be $\mathbf{y}\in T$, for no point in $[T]$ is orthogonal to all points of $\varepsilon(\cE)$, being $\cE$ elliptic.
Then, $\mathbf{y}=\mathbf{t}+\alpha\mathbf{u}+\beta\mathbf{v}$
with $\mathbf{t}\in T$ and $(\alpha,\beta)\neq(0,0)$.
If $\mathbf{x}$ is a generic vector in $T''$, we can write
$\mathbf{x}=\mathbf{t}'+\alpha' \mathbf{u}+\beta'\mathbf{v}$ for
arbitrary $\mathbf{t}'\in T$, $\alpha',\beta'\in \KK$.
Since $f(\mathbf{y},\mathbf{x})=0$ (and $f(\mathbf{u},\mathbf{u})=0$ because
$\mathbf{u}\in\mathrm{Rad}(f|_{T'})$)
we have $\forall \mathbf{t}'\in T,\alpha',\beta'\in\KK$:
\begin{multline}
\label{ee}
f(\mathbf{t}+\alpha\mathbf{u}+\beta\mathbf{v},
\mathbf{t}'+\alpha'\mathbf{u}+\beta'\mathbf{v})=\\
f(\mathbf{t},\mathbf{t}')+\alpha' f(\mathbf{t},\mathbf{u})+
\beta' f(\mathbf{t},\mathbf{v})+
\alpha f(\mathbf{u},\mathbf{t}')+\alpha\beta' f(\mathbf{u},\mathbf{v})+
\beta f(\mathbf{v},\mathbf{t}')+\beta\alpha' f(\mathbf{v},\mathbf{u})= \\
f(\mathbf{t},\mathbf{t}')+
\beta' f(\mathbf{t},\mathbf{v})+
(\alpha\beta'+\alpha'\beta)f(\mathbf{u},\mathbf{v})+
\beta f(\mathbf{v},\mathbf{t}')=0.
\end{multline}
Take now $\mathbf{t}'\in \mathbf{t}^{\perp_f}\cap\mathbf{v}^{\perp_f}$, $\beta'=0$ and $\alpha'=1$.
Then Equation~\eqref{ee} becomes
$\beta (f(\mathbf{u},\mathbf{v}))=0.$
Since $f(\mathbf{u},\mathbf{v})\neq 0$ we have $\beta=0$;
so $\mathbf{y}=\mathbf{t}+\alpha\mathbf{u}$, hence
$\mathbf{t}=\mathbf{y}-\alpha\mathbf{u}.$ But
$\mathbf{y}^{\perp_f}\supseteq T''\supset T'\supset T$ and
$\mathbf{u}^{\perp_f}\supseteq T'\supset T$. Hence $\mathbf{t}^{\perp_f}
\supseteq T$, so $\mathbf{t}\in\mathrm{Rad}(f_T)$.
Since $\cE$ is elliptic, $f_T$ is non-degenerate; so
$\mathbf{t}=\mathbf{0}$.
Equation~(\ref{ee}) becomes $\alpha \beta' f(\mathbf{u},\mathbf{v})=0.$ Choose $\beta'=1$, so $\alpha=0$.
This is a contradiction. So the space $\cE''$ is elliptic.
This is again a contradiction because $\cE\subset \cE''$ and $\cE$ is a maximal elliptic subspace of $\cP$.
Hence we have that $\langle T, \mathrm{Rad}(f)\rangle$ cannot be properly contained in $V.$ This implies the lemma.
\end{proof}
We now conclude the proof of the Theorem~\ref{m:1} by
proving that all the maximal elliptic chains of a classical orthogonal polar space have the same length.
By Corollary~\ref{e:max} $\cP$ contains a maximal nice subspace with respect
to the property that its hyperbolic lines consist of just $2$ points;
this is either the subspace generated by a frame or a maximal elliptic
subspace of $\cP$; denote it by $\cE$
and let $[T]=\langle\varepsilon(\cE)\rangle$.
Then,
$\varepsilon|_{\cE}:\cE\to [T]$ is the universal
embedding of $\cE$. By Lemma~\ref{lem:key}, $T$ is a direct complement of the radical of $f.$ The subspace $\mathrm{Rad}(f)$, and hence $T\cong V/\mathrm{Rad}(f)$, is uniquely
determined by $\cP$ and $\varepsilon$ and does
not depend on the maximal elliptic subspace
of $\cP$ we have chosen.
So, fix any maximal elliptic subspace $\cE$ of $\cP$.
By Lemma~\ref{lem pur elliptic} with $\cE$ in the role of $\cP$ there exists a maximal elliptic chain of length $\delta$ in $\cE$ which refines to a maximal chain of nice
subspaces as in~\eqref{eqn} of length $2\delta$. In particular, the length of the
chain~\eqref{eqn} is the same as the anisotropic gap of $\cE$ which is the dimension of
$\dim(V/R)-2n$, or equivalently the codimension in $V_0$ of $\mathrm{Rad}(f)$.\\
\medskip
\noindent {\bf This ends the proof of Theorem~\ref{m:1}}. \hfill$\Box$
\begin{remark}
It is a consequence of Theorem~\ref{lem:key} that the codimension
in $\PG(V)$ of
the image of the embedding of a maximal elliptic subspace $\cE$ of $\cP$ does
not depend on the choice of $\cE$.
In particular the image of the universal embedding of any maximal
elliptic subspace of $\cP$ is hosted in $\PG(V/\mathrm{Rad}(f))$
where $\PG(V)$ is the codomain of the universal embedding of
$\cP$.
When the dimension of $V/\mathrm{Rad(f)}$ is finite, it is possible
to use the standard theory of quadratic forms (Witt's theorem) to
prove that all maximal elliptic subspaces of $\cP$ must be
isomorphic to each other (even if there might be elliptic subspaces
of $\cP$ which are non-isomorphic).
We leave the case of infinite dimension (in particular
when the dimension of this space is larger that $\aleph_0$ to future
investigation)
\end{remark}
\section{Parabolic gap}\label{parabolic gap}
As in Section~\ref{sec 3}, assume that $\cP$ is an orthogonal polar space defined over a field of characteristic $2$. If $\cP$ is a grid, then
let $\varepsilon\colon\cP\rightarrow\PG(V)$ be any of its relatively universal
embeddings. Otherwise, let
$\varepsilon\colon \cP \rightarrow \PG(V)$ be the universal embedding of
$\cP$.
In any case can write $\varepsilon(\cP)=\cP(\phi)$ where $\phi\colon V\rightarrow \KK$ is a quadratic form having bilinearization $f\colon V\times V\rightarrow \KK$ (see Section~\ref{sesquilinear and pseudoquadratic form}).
\subsection{Proof of Theorem~\ref{m:2}}
Suppose $\cP$ has elliptic gap $\fd.$
Let
\[\fE: \cF=\cE_0\subset\cE_1\subset\dots\subset\cE_{\delta}\]
be a maximal elliptic chain of $\cP$ (of length $\fd=|\delta|+1$) and put
$\cE':=\bigcup_{\cE\in\fE}\cE$.
Then $\cE'$ is a maximal subspace of $\cP$ with the property that all its hyperbolic lines have $2$ points and by Lemma~\ref{lem:key}
$V_0=\langle\varepsilon(\cE')\rangle\oplus\mathrm{Rad}(f)$.
So,
$\fd=\dim(V_0/\mathrm{Rad}(f))$
with $\fr:=\dim(V_0)$ being the anisotropic gap of $\cP$.
Using
Lemma~\ref{parabolic defect} the chain $\fE$ can be extended to a maximal well ordered chain
of nice subspaces of $\cP$, say $\fN$. By Theorem~\ref{plain}, the length of $\fN$ (being an invariant of $\cP$) is precisely the anisotropic gap $\fr$ of $\cP$ and $\fr=\dim (V_0).$ Hence $\dim({\mathrm{Rad}}(f))=\fr-\fd$
in the sense of Remark~\ref{r:111}.
\hfill$\Box$
\subsection{Parabolic polar spaces}\label{section finale}
In light of Corollary~\ref{d112} and our definitions $\cP$ is parabolic
if it is $(0,\fp)$-orthogonal with $\fp>0$; that is
$\cP$
properly contains a hyperbolic polar space as a subspace
and its elliptic gap is $0$.
As usual let $\varepsilon:\cP\to\PG(V)$ be the universal embedding
of $\cP$. Then, $\cP$ is parabolic if we have the
following decomposition of $V:$
\begin{equation}\label{pp}
V=(V_1\oplus V_2\oplus\dots\oplus V_n)\oplus\mathrm{Rad}(f).
\end{equation}
Note that the quotient embedding $\varepsilon/\mathrm{Rad}(f):\cP\to\PG(V/\mathrm{Rad}(f))$
has dimension $2n$ and it is described by a generalized quadratic form (see~\cite{P}) or by an alternating form.
By~\cite[Theorem 1.11]{P22} the condition of admitting an embedding of dimension $2n$ for $\cP$ can be formulated as follows:
\begin{itemize}
\item[(A)] for any two non-collinear points $a,b\in\cP$, if $M$ and $N$
are respectively maximal singular subspaces of
$\cP$ and $\{a,b\}^\perp$ and $N\subseteq M$, then $M\cap\{a,b\}^{\perp\perp}\neq\emptyset$.
\end{itemize}
\begin{remark}
\label{rr}
Observe that a hyperbolic polar space is an embeddable polar
space whose hyperbolic lines consist of just $2$ points
which satisfies Condition~(A).
Indeed, Condition~(A) by itself
assures the existence of an embedding of minimum
dimension (i.e. $2n$).
If all the hyperbolic lines of $\cP$ have
just $2$ points, then such an embedding cannot be obtained as
a quotient of an embedding of higher dimension.
Hence it is unique
(except for the case of grids). Moreover, the
condition on cardinality
of the hyperbolic lines guarantees that this embedding realizes
$\cP$ as the point set of
an hyperbolic quadric; this also in the case of the grids even if the embedding is not unique. So, a grid is always
a hyperbolic polar space, even if it does not admit
a universal embedding.
\end{remark}
Thanks to Remark~\ref{rr},
we can rephrase the definition of parabolic spaces already given in the Introduction as follows.
\begin{definition}
A \emph{parabolic polar space} $\cP$ is an embeddable polar
space properly containing a hyperbolic polar subspace and satisfying
Condition~(A).
\end{definition}
\begin{theorem}
Let $\cP$ be an orthogonal polar space.
Then its parabolic gap is at most
$[\KK:\KK^2]$.
\end{theorem}
\begin{proof}
Suppose $\cP$ to have rank $n$.
If $\cP$ is hyperbolic or elliptic, then there is nothing to prove.
Otherwise, put
$R:=\mathrm{Rad}(f)$. We claim
$\dim(R)\leq [\KK:\KK^2]$.
Indeed, let $(\mathbf{r}_i)_{i\in I}$ be a basis of $R$.
As $R$ is anisotropic, for any $x_i\in \KK\setminus\{0\}$
we have
\[ \sum_{i\in I} \phi(\mathbf{r_i})x_i^2\neq 0. \]
In particular, the values $\xi_i:=q(\mathbf{r_i})\in\KK$ must be linearly
independent over $\KK^2$, where $\KK$ is regarded as a vector space over its subfield of squares $\KK^2$. It follows that we have $\dim(R)\leq[\KK:\KK^2]$. So, by Theorem~\ref{m:2}, the parabolic gap of $\cP$ is at most $[\KK:\KK^2]$.
\end{proof}
\begin{remark}
\label{c:cpp}
Let $\cP$ be an orthogonal polar space
and suppose $[\KK:\KK^2]<\infty$ with $\mathrm{char}(\KK)=2$.
By our results, the image of the universal embedding of $\cP$ decomposes
as in~\ref{decomposition} and the parabolic and elliptic gaps of $\cP$
are respectively $\fp=\dim(\mathrm{Rad}(f))$ and $\fe=\dim(V_0')$.
By standard results in the theory of quadratic forms, see~\cite[Lemma 36.8]{EKM}
we have $\frac{1}{2}\fe+\fp\leq[\KK:\KK^2]$ (when $\fe$ is infinite we have $\frac{1}{2}\fe=\fe$).
In particular if
the parabolic gap $\fp$ of $\cP$ is
exactly $[\KK:\KK^2]$, then $\fe=0$.
In this case
the
(minimum)
embedding of $\cP$ (obtained by quotienting its universal
embedding over $\mathrm{Rad}(f)$ realizes $\cP$ as a symplectic space.
The converse is also true:
if $\cP$ is a polar space that admits an embedding over
a field $\KK$ with $\mathrm{char}(\KK)=2$ described by an alternating
bilinear form, then $\cP$ is parabolic of gap $[\KK:\KK^2]$.
\end{remark}
\section*{Thanks}
The authors wish to express their sincere
gratitude to Prof. Antonio Pasini for his
precious advice and remarks on this work.
Both authors are affiliated with GNSAGA of INdAM (Italy) whose support
they acknowledge.
|
\section{Introduction}
We continue the discussion from \cite{hhs21} on $\Orth(\q,k)$-conjugacy classes of involutions of $\Orth(\q,k)$, where $k$ is a field of characteristic 2, and turn our focus to the fixed point groups of such involutions. The goal of this research is to describe generalized symmetric spaces of the form $G(k)/H(k)$ where $G$ is an algebraic group defined over a field $k$, $H$ is the fixed point group of some automorphism of order $2$ on $G$ and $G(k)$ (respectively $H(k)$) denotes the $k$-rational points of $G$ (resp. $H$). In particular we want to extend Helminck's study of $k$-involutions and symmetric $k$-varieties \cite{he00} to include fields of characteristic $2$. This has been studied for groups of type $\G_2$ and $\mathrm{A}_n$ in \cite{hs18, sc18} and over fields of characteristic not $2$ in \cite{do06, bhw15, he02, bdhw16, hu14, hu15,hu16}. We also extend the results of Aschbacher and Seitz \cite{as76} who studied similar structures for finite fields of characteristic $2$. Symmetric spaces were first studied by Gantmacher in \cite{ga39} in order to classify simple real Lie groups. This was continued in \cite{be57} by Berger who provides a complete classification of symmetric spaces for simple real Lie algebras.
We refer to \cite{hl04} for notation and vocabulary concerning quadratic forms over fields of characteristic $2$. Orthogonal and symplectic groups over fields of characteristic $2$ have been studied extensively. In \cite{ch66} Cheng Hao discusses automorphisms of the orthogonal group over perfect fields of characteristic $2$ when the quadratic form is nondefective. Pollak discusses orthogonal groups over global fields of characteristic $2$ in the case the quadratic form is nondefective in \cite{po70} and Connors writes about automorphism groups of orthogonal groups over fields of characteristic $2$ in \cite{co73,co74,co75,co76} for a nondegenerate quadratic form. In \cite{wi78} Wiitala uses the Jordan canonical form to study involutions of the orthogonal group over fields of characteristic 2 in the nondefective case. In \cite{di51} Jean Dieudonn\'e discusses algebraic homogeneous spaces over fields of characteristic $2$.
In the following we establish isomorphism classes for the fixed point groups of involutions of orthogonal groups for quadratic spaces with a totally singular quadratic form and a non-singular quadratic form. We also provide a characterization of the fixed point group for a general quadratic form defining an orthogonal group over a field of characteristic 2. When possible, this is done without assuming the form is nondegenerate.
Over fields of characteristic not $2$ the $G(k)$-conjugacy classes of automorphisms of order $2$ are in bijection with $G(k)$-conjugacy classes of fixed point groups of involutions on $G$ as seen in \cite{hw93}. However, in characteristic $2$ there are examples of involutions that are not $G(k)$-conjugate with isomorphic fixed point groups, see Theorem \ref{orth_diag_fixed_pts}, as well as Theorem 5.27 in \cite{hs18}.
In the Preliminaries section we recall definitions, notation, and useful results concerning quadratic forms and bilinear forms over fields of characteristic $2$ and associated algebraic groups. This section also includes notation from the theory of symmetric $k$-varieties. Next we recall results concerning involutions of $\Orth(\q,k)$ when $k$ is a field of characteristic $2$ in Section $3$. This is followed by describing isomorphism classes of fixed point groups for different types of vector spaces. In Section $4$ we consider the fixed point group for involutions of $\Orth(\q,k)$ where $\q$ is the quadratic form of a totally singular vector space. Then in Section $5$ we discuss the fixed point group of involutions where $\q$ is the quadratic form of a nonsingular vector space. Finally, Section 6 has an overview of the implications for a general vector space over a field of characteristic $2$.
\section{Preliminaries}
The following definitions can be found in \cite{hl04}. We also refer to \cite{gr02} often, which uses some of the vocabulary for orthogonal groups differently than \cite{hl04}. When this happens, we default to the definitions of \cite{hl04}.
Let $k$ be a field of characteristic 2 and $W$ a vector space defined over $k$. We call $\q:W \rightarrow k$ a \emph{quadratic form} if it satisfies $\q(aw) = a^2 \q(w)$ for all $a \in k$, $w \in W$. There exists a corresponding symmetric bilinear form $\mathrm{B}: W \times W \rightarrow k$ such that $\q(w+w') = \q(w)+\q(w') + \mathrm{B}(w,w')$ for all $w,w' \in W$. Over fields of characteristic $2$ nonsingular symmetric bilinear forms are also symplectic.
The pair $(W,\q)$ is called a \emph{quadratic space}. The subspace of $W$
\[ \mathrm{rad}(W) = \{ x \in W \ | \ \mathrm{B}(x,w)= 0,\text{ for all } w \in W\}. \]
is called the \emph{radical} of $W$.
Given a quadratic form, there exists a basis of $W$, consisting of $x_i, y_i, g_j$, where $i \in \{1, 2, \hdots, r\}$ and $j \in \{1, 2, \hdots, s\}$ and field elements $a_i, b_i, c_j \in k$ such that
\[\q(w) = \displaystyle \sum_{i=1}^r (a_i e_i^2 + e_if_i + b_if_i^2) + \sum_{j=1}^s c_j z_j^2\]
when $w = \sum_{i=1}^r (e_ix_i + f_iy_i) + \sum_{j=1}^s z_j g_j$. We denote this quadratic form by
\begin{equation}
\label{quadratic_signature}
\q = [a_1,b_1] \perp [a_2,b_2] \perp \cdots \perp [a_r,b_r]\perp \langle c_1, c_2, \hdots, c_s \rangle
\end{equation}
and so $\rad(W)= \spa \{ g_1, g_2, \hdots, g_s\} $. The representation of $\q$ given in \ref{quadratic_signature} is called the \emph{quadratic signature} of $\q$. We say that such a quadratic form is of type $(r,s)$. A nonzero vector $w\in W$ is an \emph{isotropic vector} if $\q(w)=0$, $V$ is an \emph{isotropic vector space} if it contains isotropic elements and \emph{anisotropic} otherwise. The space $W$ is called \emph{nonsingular} if $s=0$, and \emph{totally singular} if $r=0$. Furthermore, $W$ is called \emph{nondefective} if $s=0$ or $\rad(W)$ is anisotropic. A \emph{hyperbolic plane} has a quadratic form isometric to the form $[0,0]$ and will be denoted by $\mathbb{H}$. We will call $\q'$ a \emph{subform} of $\q$ if there exists a form $\mathrm{p}$ such that $\q \cong \q' \perp \mathrm{p}$. In this case, $p$ is called a completion of $\q'$ with respect to $\q$.
Sometimes we will need to refer to a diagonal bilinear form
\[ \mathrm{B}(w,w') = a_1w_1w'_1 + a_2w_2w'_2 + \cdots + a_lw_lw'_l, \]
which we denote $\langle a_1, a_2, \cdots, a_l \rangle_{\mathrm{B}}$ following \cite{hl04}. Throughout the paper we will use the calligraphy math font to denote sets of vectors in $W$ such as $\mathcal{U} = \{u_1, u_2, \ldots, u_l\}$ and $\overline{\mathcal{U}} = \spa \mathcal{U}$. The next result is Proposition 2.4 from \cite{hl04}.
\begin{prop}\label{hofflaghdecomp}
Let $\q$ be a quadratic form over $k$. Then
\[ \q \cong m \times \mathbb{H} \perp \widetilde{\q_r} \perp \widetilde{\q_s} \perp d \times \langle 0 \rangle, \]
with $\widetilde{q_r}$ nonsingular, $\widetilde{\q_s}$ totally singular and $\widetilde{\q_r} \perp \widetilde{q_s}$ anisotropic. The form $\widetilde{\q_r} \perp \widetilde{\q_s}$ is uniquely determined up to isometry. In particular $m$ and $d$ are uniquely determined.
\end{prop}
We call $m$ the \emph{Witt index} and $d$ the \emph{defect} of $\q$. If
\[ \q \cong m \times \mathbb{H} \perp \widetilde{\q_r} \perp d \times \langle 0 \rangle \perp \widetilde{\q_s}, \]
with respect to the basis
\[ \{ x_1,y_1, \ldots x_m,y_m, \ldots , x_r,y_r, g_1, \ldots g_d, g_{d+1}, \ldots , g_{s} \}, \]
we will call
\[ \mathrm{def}(W) = \mathrm{span}\{g_1,\ldots g_d \}, \]
the \emph{defect} of $W$. If $\mathcal{U}$ is a basis for a subspace $\overline{\mathcal{U}}$ of $W$, we will refer to the restriction of $\q$ to $\overline{\mathcal{U}}$ with respect to this basis by $\q_{\mathcal{U}}$.
Let $G$ be an algebraic group, then an automorphism $\theta: G \rightarrow G$ is an \emph{involution} if $\theta^2 = \id$, $\theta \neq \id$. The automorphism $\theta$ is a $k$-\emph{involution} if in addition $\theta(G(k)) = G(k)$, where $G(k)$ denotes the $k$-rational points of $G$. We define the {\em fixed point group} of $\theta$ in $G(k)$ by
\[ G(k)^{\theta} = \{ \gamma \in G(k) \ | \ \theta \gamma \theta^{-1} = \gamma \}, \]
this is often denoted $H(k)$ or $H_k$ in the literature when there is no ambiguity with respect to $\theta$. Let $\mathcal{I}_{\tau}$ denote conjugation by $\tau$.
We often consider groups that leave a bilinear form or a quadratic form invariant. If $\mathrm{B}$ is a bilinear form on a nonsingular vector space $W$ we will denote the {\em symplectic group} of $(W,\q)$ over a field $k$ by
\[ \Sp(\mathrm{B},k) = \{ \varphi \in \GL(W) \ | \ \mathrm{B}(\varphi(w), \varphi(w')) = \mathrm{B}(w,w') \text{ for } w,w' \in W \}. \]
The classification of involutions for $\Sp(\mathrm{B},k)$ for a field $k$ such that $\mathrm{char}(k) \neq 2$ has been studied in \cite{bhw15}. For any quadratic space $W$ over a field $k$ we will denote the {\em orthogonal group} of $(W,\q)$ by
\[ \Orth(\q,k) = \{ \varphi \in \GL(W) \ | \ \q(\varphi(w)) = \q(w) \text{ for all } w \in W \}. \]
When $W$ is nonsingluar we have $\Orth(\q,k) \subset \Sp(\mathrm{B},k)$ if $\mathrm{B}$ is the bilinear form associated with $\q$.
We will need to make use of some facts about quadratic spaces stated in the following lemmas. The first statement outlines some standard isometries for quadratic forms over a field of characteristic $2$, and the second allows us to express $\q$ using a different completion of the nonsingular space. These results and more like them appear in \cite{hl04}.
\begin{lemma}
Let $\q$ be a quadratic form on a vector space $W$ and suppose $\alpha \in k$. Then the following are equivalent representations of $q$ on $W$:
\begin{enumerate}
\item $[a,a'] = [a,a+a'+1] = [a',a] = [ \alpha^2 a, \alpha^{-2} a'] $
\item $[a,a'] \perp [b,b'] = [a+b,a'] \perp [b,a'+b'] = [b,b'] \perp [a,a']$
\end{enumerate}
\end{lemma}
\begin{lemma} Let $a_i, b_i, c_i \in k$ for $1 \leq i \leq n$. Suppose $\{a_1, ... , a_n \}$ and $\{b_1, ... , b_n \}$ span the same vector space over $k^2$ and $\q = [a_1,c_1] \perp \hdots \perp [a_n , c_n]$. Then there exist $c_i^{\prime} \in k$, $1 \leq i \leq n$, such that $\q = [b_1,c_1^{\prime}] \perp \hdots \perp [b_n , c_n^{\prime}]$.
\end{lemma}
For a map $\theta \in \GL(W)$ we define the \emph{residual space} of $\theta$ to be the image of $\theta + \id$ and the \emph{residue} of $\theta$ to be the dimension of the residual space.
\section{Summary of involution results}
Here we summarize the results from \cite{hhs21} that we need to talk about the classification of the fixed points groups of each $\Orth(\q,k)$-conjugacy class of an involution.
If $W$ is a nonsingular vector space with respect to $\q$ we can define a \emph{transvection} on $W$ induced by $w \in W$ and $a \in k$
\[ \tau_{w,a}(z) = z + a \mathrm{B}(w,z)w, \]
and an \emph{orthogonal transvection} is of the form $\tau_{w, \frac{1}{\q(w)}}$ for $w \in W$ with $\q(w) \neq 0$. In this case we denote the transvection by $\tau_w$.
If $\mathcal{U}$ is the set of mutually orthogonal vectors inducing a product of orthogonal transvections $\tau$ we call $\mathcal{U}$ the \emph{inducing set} for
\[ \tau = \tau_{\mathcal{U}} = \tau_{u_l} \cdots \tau_{u_2}\tau_{u_1}. \]
Let $\mathcal{U} = \{u_1, \ldots, u_l\}$ and $\mathcal{X} = \{ x_1, \ldots, x_l\}$ be sets of mutually orthogonal vectors in a nonsingular subspace of $W$. In what follows we will factor involutions into a product of involutions of a given type (orthogonal transvection, basic null, basic radical), such a factorization is called \emph{reduced} if it uses the minimum number of factors. We refer to the number of factors in a reduced factorization as the \emph{length} of an involution.
\begin{lemma}\label{equal_tr}
Two orthogonal involutions given by reduced products of orthogonal transvections are equal, $\tau_{u_l} \cdots \tau_{u_2} \tau_{u_1} = \tau_{x_l} \cdots \tau_{x_2} \tau_{x_1}$, if and only if $\overline{\mathcal{U}} = \overline{\mathcal{X}}$.
and
\[ \left\langle \q(u_1), \q(u_2), \ldots, \q(u_l) \right\rangle_{\mathrm{B}} \cong \left\langle \q(x_1), \q(x_2), \ldots, \q(x_l) \right\rangle_{\mathrm{B}}. \]
\end{lemma}
The classification of $\Orth(\q,k)$-conjugacy classes of involutions given by a product of transvections in a nonsingular space can then be described.
\begin{thm} \label{conj_transv}
Let $\tau_{u_l} \cdots \tau_{u_2} \tau_{u_1}$ and $\tau_{x_l} \cdots \tau_{x_2} \tau_{x_1}$ be orthogonal involution written as a product of transvections on $W$ such that $\phi \in \Orth(\q,k)$. Then
\[ \phi \tau_{u_l} \cdots \tau_{u_2} \tau_{u_1}\phi^{-1} = \tau_{x_l} \cdots \tau_{x_2} \tau_{x_1} \]
if and only if
\[ \left\langle \q(u_1), \q(u_2), \ldots, \q(u_l) \right\rangle_{\mathrm{B}} \cong \left\langle \q(x_1), \q(x_2), \ldots, \q(x_l) \right\rangle_{\mathrm{B}}, \]
and $\dim(\overline{\mathcal{U}}) = \dim(\overline{\mathcal{X}})$.
\end{thm}
The last stipulation, $\dim(\overline{\mathcal{U}}) = \dim(\overline{\mathcal{X}})$, distinguishes between elements $\tau$ whose reduced length is $\mathrm{res}(\tau) + 1$, called hyperbolic involutions, and those whose reduced length is $\mathrm{res}(\tau)$, called diagonal.
There is one other type of nonsingular involution called a null involution.
\begin{definition} A plane $P = \spa\{ x, y \}$ is hyperbolic (or Artinian) if both of the following are satisfied:
\begin{enumerate}
\item $\q(x) = \q(y) = 0$
\item $B(x,y) \neq 0.$
\end{enumerate}
\end{definition}
If $x, y$ span a hyperbolic plane, we can rescale to assume $B(x,y) =1$. Proposition 188.2 of \cite{sn89} guarantees that every nonsingular nonzero isotropic vector is contained in a hyperbolic plane.
\begin{definition}
Let $\eta$ be an involution of $\Orth(\q,k)$ where $(W,\q)$ is a quadratic space, and let $\mathbb{P}$ be the orthogonal sum of two hyperbolic planes. Then an involution $\eta$ is called a basic null involution in $\mathbb{P}$ if all of the following are satisfied:
\begin{enumerate}
\item $\eta$ leaves $\mathbb{P}$ invariant
\item $\eta$ fixes a 2-dimensional subspace of $\mathbb{P}$ where every vector has norm zero
\item $\eta \vert_{\mathbb{P}^C} = \id_{\mathbb{P}^C}$, where $\mathbb{P}^C$ is the complement of $\mathbb{P}$ in $W$.
\end{enumerate}
\end{definition}
A \emph{null involution} is a product of basic null involutions. Note that if $\eta$ is a null involution and $\tau$ is a transvection, then $\tau \eta$ is a product of transvections \cite{sn89}, so we only need to consider the conjugacy of null involutions themselves.
\begin{prop} \label{conj_null}
Two null involutions are $\Orth(\q,k)$-conjugate if and only if they have the same length.
\end{prop}
Finally we consider involutions acting only on the radical of $W$.
\begin{prop} \label{Oqk_radV}
$\Orth(\q_{\rad(W)},k) \cong \GL_d(k) \ltimes \Mat_{d,s-d}(k)$ where $d$ is the defect of $\rad(W)$.
\end{prop}
\begin{proof}
Let $d$ be the defect of $W$. We can choose a basis of the form
\[ \{g_1, \ldots, g_d, g_{d+1}, \ldots, g_s\} \]
where $\q(g_i)=0$ for $1\leq i \leq d$
\[ \mathcal{G}_{\mathrm{is}} = \{g_{1}, \ldots, g_d\}, \]
and the subspace spanned by
\[ \mathcal{G}_{\mathrm{an}} = \{g_{d+1}, \ldots, g_s\} \]
is anisotropic. We claim we can represent an isometry on $\mathrm{rad}(W)$ by
\[ \Psi= \begin{bmatrix}
\psi & M \\
0 & \id
\end{bmatrix}. \]
We must have $\Psi|_{\mathcal{G}_{\mathrm{an}}} = \id$ otherwise there would exist two elements in $\mathcal{G}_{\mathrm{an}}$ with the same norm, and since the space is totally singular the sum of the two vectors would have norm zero. The restriction $\Psi|_{\mathcal{G}_{\mathrm{is}}}= \psi \in \GL_d(k)$ since $\psi$ must be invertible but has no other contraints due to the bilinear form and the norm being identically zero for $\overline{\mathcal{G}_{\mathrm{is}}}$. The isometric image of an element from $\overline{\mathcal{G}_{\mathrm{is}}}$ cannot include any components from $\overline{\mathcal{G}_{\mathrm{an}}}$ since that would imply that $\overline{\mathcal{G}_{\mathrm{an}}}$ was not anisotropic. Lastly the $M$ block is free since we can add any isotropic vector from the the radical to any vector in $W$ and leave the quadratic form (and bilinear form) invariant.
We can factor
\[ \Psi= \begin{bmatrix}
\psi & M \\
0 & \id
\end{bmatrix}=
\begin{bmatrix}
\psi & 0 \\
0 & \id
\end{bmatrix}
\begin{bmatrix}
\id & M \\
0 & \id
\end{bmatrix} \]
where $\bigl[\begin{smallmatrix} \psi & M \\ 0 & \id\end{smallmatrix} \bigr]\bigl[\begin{smallmatrix} \id & B \\ 0 & \id\end{smallmatrix} \bigr]\bigl[\begin{smallmatrix} \psi & M \\ 0 & \id\end{smallmatrix} \bigr]^{-1} = \bigl[\begin{smallmatrix} \id & \psi B \\ 0 & \id\end{smallmatrix} \bigr]$.
\end{proof}
A \emph{radical involution} is an element $\rho \in \Orth(\q,k)$ or order $2$ defined on vectors in the radical of $W$. Notice that a nontrivial orthogonal transformation on $\rad(W)$ detects a defective vector in $W$, since if $\rho(g) = g'$ then $\q(g+g') = \q(g) + \q(g') = 0$. A \emph{basic radical involution} is a radical involution $\rho_i \in \Orth(\q,k)$ such that $\rho_i(g_i) = g_i'$, where $g_i, g_i'$ are linearly independent vectors in $\rad(W)$ with $\q(g_i) = \q(g_i')$, and $\rho_i$ acts trivially on the complement of $\overline{\{g_i, g_i'\}}$. All radical involutions can be written as a finite product of basic radical involutions.
\begin{prop}
Two basic radical involutions $\rho_1, \rho_2$ are $\Orth(\q,k)$-conjugate if and only if $\rho_1$ and $\rho_2$ act non-trivially on isometric vectors.
\end{prop}
\begin{corollary} \label{conj_rad}
All radical involutions of length $m$ with quadratic signature isometric to
\[ \langle \q(g_1), \q(g_2), \ldots, \q(g_m) \rangle, \]
are $\Orth(\q,k)$-conjugate.
\end{corollary}
\section{Totally singular spaces}
We are now ready to discuss the isomorphism classes of the fixed point groups of our involutions of $\Orth(\q,k)$, which can be used to study the corresponding symmetric $k$-varieties. First, we discuss the fixed point groups of involutions acting on a totally singular vector space. This in turn leads to a discussion about fixed point groups of diagonal involutions on nonsingular subspaces.
Let us consider a vector space $W$ such that $W= \rad(W)$. Recall that in such a vector space any quadratic form will be of the form
\[ \q = \langle \q(g_1), \q(g_2), \ldots, \q(g_s) \rangle, \]
and by Proposition \ref{hofflaghdecomp} we have an isometry
\[ \q \cong \langle 0, \ldots, 0, \q(g_1), \q(g_2), \ldots, \q(g_{s-d}) \rangle, \]
where $d$ is the defect of $\q$.
By Proposition \ref{Oqk_radV} we have $\Orth(\q_{\rad(W)}, k) \cong \GL_d(k) \ltimes \Mat_{d,s-d}(k)$. For an element $\delta \in \Orth(\q,k)$ to be fixed by conjugation with the involution $\rho$ we need $\rho \delta \rho = \delta$.
\begin{prop}
If $\q$ is a quadratic form on $W = \mathrm{rad}(W)$ and $\rho \in \Orth(\q,k)$ is an involution on $W$ such that $\rho(g_i) = g_i'$ and quadratic signature
\[ \langle \q(g_1),\q(g_2), \cdots, \q(g_n) \rangle \]
then $\Orth(\q,k)^{\mathcal{I}_{\rho}} \cong \Orth(\q_{\mathcal{G} \cup \mathcal{H}}, k) \ltimes \Mat_{n,s-n}(k)$
where
\begin{align*}
\mathcal{G}' &= \{g_1',\ldots, g_n' \} \\
\mathcal{G} &= \{g_1,\ldots, g_n \} \\
\mathcal{H} &= \{h_1, \ldots, h_{s-2n}\}
\end{align*}
with $\rho(h_i) = h_i$ and $\overline{ \mathcal{G}' \cup \mathcal{G} \cup \mathcal{H} } = W$.
\end{prop}
\begin{proof}
Let $\langle \q(g_1), \q(g_2), \ldots, \q(g_n) \rangle$ be the signature of a radical involution $\rho$ such that $\rho(g_i) = g_i'$ and $g_{i}'\neq g_i$ for any $i$ or $i'$. Notice that $\rho(g_i + g_i') = g_i + g_i'$ and further that $\rho(g_i) = g_i + (g_i + g_i')$. Now we order a basis for $W$,
\[ \mathcal{W} = \{ g_1+g_1', g_2+g_2', \ldots, g_n+g_n', h_1, h_2, \ldots, h_{s-2n}, g_1,g_2, \ldots, g_n \} \]
such that $\overline{ \mathcal{W} } = W$ and $\rho(h_i) = h_i$. With respect to this basis we can represent $\rho$ as the block matrix
\[ \rho = \begin{bmatrix}
\id & 0 & \id \\
0 & \id & 0 \\
0 & 0 & \id
\end{bmatrix}, \]
and so any $\delta \in \Orth(\q, k)^{\mathcal{I}_{\rho}}$
\[ \delta = \begin{bmatrix}
\delta_{11} & \delta_{12} & \delta_{13} \\
\delta_{21} & \delta_{22} & \delta_{23} \\
\delta_{31} & \delta_{32} & \delta_{33} \\
\end{bmatrix}, \]
such that $\delta\rho = \rho \delta$ is of the form
\[ \delta = \begin{bmatrix}
\delta_{11} & \delta_{12} & \delta_{13} \\
0 & \delta_{22} & \delta_{23} \\
0 & 0 & \delta_{11} \\
\end{bmatrix}=
\begin{bmatrix}
\delta_{11} & 0 & 0 \\
0 & \delta_{22} & \delta_{23} \\
0 & 0 & \delta_{11} \\
\end{bmatrix}
\begin{bmatrix}
\id & \delta_{12}' & \delta_{13}' \\
0 & \id & 0 \\
0 & 0 & \id \\
\end{bmatrix} \]
where $\delta_{ij}' = \delta_{11}^{-1}\delta_{ij}$. Notice that $\delta_{12}'$ and $\delta_{13}'$ can take any matrix value, since $\overline{ \{ g_1+g_1', \ldots, g_n+g_n'\}}$ is contained in the defect of $W$ and the span of the remaining basis vectors would be
\[ \mathcal{H} \cup \mathcal{G} = \{ h_1, \dots, h_{s-2n}, g_1, \ldots, g_n \}, \]
Now the subgroup
\[ \mathsf{M}=\left\{ \begin{bmatrix}
\delta_{11} & 0 & 0 \\
0 & \delta_{22} & \delta_{23} \\
0 & 0 & \delta_{11} \\
\end{bmatrix} \bigg| \bigl[\begin{smallmatrix} \delta_{22} & \delta_{23} \\ 0 & \delta_{11} \end{smallmatrix} \bigr] \in \Orth(\q_{\mathcal{H} \cup \mathcal{G} },k, \overline{\mathcal{H}} ) \right\}, \]
is such that $\mathsf{M} \cong \Orth(\q_{\mathcal{H} \cup \mathcal{G} },k, \overline{\mathcal{H}} )$ which is the subgroup of $\Orth(\q_{\mathcal{H} \cup \mathcal{G} },k)$ that leaves the subspace $\overline{\mathcal{H}}$ invariant.
The subgroup
\[ \mathsf{M}'= \left\{ \begin{bmatrix}
\id & \delta_{12}' & \delta_{13}' \\
0 & \id & 0 \\
0 & 0 & \id \\
\end{bmatrix} \bigg| \delta_{12}' \in \Mat_{n, d'}(k), \delta_{13}' \in \Mat_{n,s-n-d'}(k) \right\}, \]
is such that $\mathsf{M}' \cong \Mat_{n,s-n}(k)$ and $\mathsf{M}$ is a normal subgroup of $\Orth(\q,k)^{\mathcal{I}_{\rho}}$,
so we have that $\Orth(\q,k)^{\mathcal{I}_{\rho}} \cong \Orth(\q_{\mathcal{H} \cup \mathcal{G} },k, \overline{\mathcal{H}} ) \ltimes \Mat_{n,s-n}(k)$.
\end{proof}
\begin{rmk}
Notice that if $W = \mathrm{rad}(W)$ is anisotropic then $\Orth(\q,k) = \{id\}$ so there are no involutions, and on the other hand if $W$ is totally isotropic we have that $\Orth(\q,k) \cong \GL_s(k)$.
\end{rmk}
\section{Nonsingular spaces}
Now we look at fixed point groups of diagonal involutions on nonsingular spaces.
\begin{prop} \label{tranv_fixU}
Let $\tau = \tau_{u_l} \cdots \tau_{u_2}\tau_{u_1}$ be a reduced diagonal orthogonal involution with inducing set $\mathcal{U}$, then $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ if and only if $\varphi( \overline{ \mathcal{U}})=\overline{\mathcal{U}}$ .
\end{prop}
\begin{proof}
Let $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ notice that $\tau \varphi \tau^{-1} = \varphi$ if and only if $\varphi \tau \varphi^{-1} = \tau$ and so
\[ \varphi \tau_{u_l} \cdots \tau_{u_2}\tau_{u_1} \varphi^{-1} = \tau_{\varphi(u_l)} \cdots \tau_{\varphi(u_2)}\tau_{\varphi(u_1)} = \tau_{u_l} \cdots \tau_{u_2}\tau_{u_1}. \]
By Lemma \ref{equal_tr} $\varphi(\overline{ \mathcal{U} }) = \overline{ \mathcal{U}}$.
Now assume $\varphi \in \Orth(\q,k)$ and $\varphi( \overline{ \mathcal{U}}) = \overline{ \mathcal{U} }$ then
\[ \langle \q( \varphi(u_1) ), \q( \varphi(u_2)), \ldots, \q( \varphi(u_l)) \rangle_B = \langle \q(u_1), \q( u_2), \ldots, \q( u_l) \rangle_B. \]
So by Theorem \ref{conj_transv} we have $\varphi \tau \varphi^{-1} = \tau$.
\end{proof}
We begin with the case that the length of $\tau$ is maximized in the nonsingular space. We denote by $\varphi_{\mathcal{U}}$ the matrix representation of the resriction of $\varphi$ to the subspace $\overline{ \mathcal{U}}$ with respect to the basis $\mathcal{U}$. Recall that
\[ \q_{\mathcal{U}} \cong \langle \q(u_1), \q(u_2), \ldots, \q(u_l) \rangle, \]
since $\overline{ \mathcal{U}}$ is a totally singular subspace of $W$. In this case, Proposition \ref{hofflaghdecomp} yields
\[ \q_{\mathcal{U}} \cong m' \times \langle 0 \rangle \perp \langle c_{m'+1}, \ldots, c_{l} \rangle. \]
This corresponds to a basis of $\mathcal{U}' = \{u_1', u_2', \ldots, u_l' \}$ such that $\q(u_i^{\prime})=c_i$ for $1 \leq i \leq l$ and $c_i=0$ for $1\leq i \leq m'$. Since the vectors in $\overline{\mathcal{U}}$ are not in $\rad(W)$ there exist non-singular completions within $W$ that form a basis for a nonsingular subspace. A nonsingular completion of $\q$ is of the form
\[ [0,\q(v_1)] \perp [ 0, \q(v_2) ] \perp \ldots \perp [0, \q(v_{m'}) ] \perp [ c_{m'+1}, \q(v_{m'+1})] \perp \ldots \perp [c_l, \q(v_l)], \]
where $\mathrm{B}(u_i',v_i)=1$, $\mathrm{B}(u_i',v_j)=0$ when $j\neq i$, and $\mathcal{V} = \{v_1, \ldots, v_l\}$ which spans a totally singular subspace. Notice that for $1\leq i \leq m'$ we can choose $v_i' = v_i + \q(v_i)u_i'$. We will refer to the vectors of this completion as $\mathcal{V}'$. Now
\[ \mathrm{B}(u_i', v_i') = \mathrm{B}(u_i', v_i) + \q(v_i)\mathrm{B}(u_i', u_i') = \mathrm{B}(u_i', v_i) = 1, \]
and
\[ \q(v_i') = \q(v_i) + \q(v_i)^2\q(u_i')+\q(v_i)B(v_i,u_i') = \q(v_i) + \q(v_i) = 0. \]
So there is a basis $\mathcal{W}$ of this nonsingular space of the form
\[\mathcal{W}'=\mathcal{U}'\cup \mathcal{V}'= \{u_1',v_1', u_2', v_2', \ldots, u_l',v_l' \}, \]
with the corresponding quadratic form
\[ \q_{\mathcal{W}} \cong m' \times [0,0] \perp [c_{m'+1}, d_{m'+1}] \perp \ldots \perp [c_l,d_l], \]
where $\q(v_i') = d_i$ for $1 \leq i \leq l$.
Let us consider $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$. Since $\varphi$ leaves $\overline{ \mathcal{U}}$ invariant we have
\[ \varphi=\begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}}A \\
0 & \varphi_{\mathcal{V}}
\end{bmatrix}, \]
where $\varphi_{\mathcal{V}} :\overline{ \mathcal{V}} \to \overline{ \mathcal{V}}$ and $A: \overline{ \mathcal{V}} \to \overline{ \mathcal{U}}$. Considering the bilinear form we have
\[ \mathrm{B}(\varphi(u_i), \varphi(v_j) ) = \mathrm{B}(\varphi_{\mathcal{U}}(u_i), \varphi_{\mathcal{V}}(v_j) + \varphi_{\mathcal{U}}(Av_j))= \mathrm{B}(\varphi_{\mathcal{U}}(u_i), \varphi_{\mathcal{V}}(v_j) ) = \mathrm{B}(u_i,v_j). \]
So, we have shown the following
\begin{prop} \label{phiUstar}
The map $\varphi_{\mathcal{V}}$ is completely determined by $\varphi_{\mathcal{U}}$.
\end{prop}
We will now refer to $\varphi_{\mathcal{V}} = \varphi_{\mathcal{U}}^*$ and using the matrix representation of $\varphi_{\mathcal{U}}$ we have $\varphi_{\mathcal{U}}^*=(\varphi_{\mathcal{U}}^{-1})^T$. We can represent $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ with the block matrix
\[ \varphi= \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}}A \\
0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} . \]
We define the additive group
\[ \mathsf{A}(\q_{\mathcal{U}},k) = \{ A \in \text{Lin}( \overline{ \mathcal{V}},\overline{ \mathcal{U}}) \ | \ \q(Av) = \mathrm{B}(v,Av) \}. \]
Notice that $\mathsf{A}(\q_{\mathcal{U}},k)$ embeds into $\Orth(\q,k)^{\mathcal{I}_{\tau}}$ as an additive group since
\[ \q( (A+C)v) = \q(Av) + \q(Cv) = \mathrm{B}(v,Av) + \mathrm{B}(v,Cv) = \mathrm{B}(v,(A+C)v), \]
for $A,C \in \mathsf{A}(\q_{\mathcal{U}},k)$. We will write $\varphi^{-*}$ for $(\varphi^{-1})^*$. The following results give us some insight into the structure of the fixed point groups.
\begin{lemma} \label{adjoint}
If $u \in \overline{ \mathcal{U}}$ and $v \in \overline{ \mathcal{V}}$, then $\mathrm{B}(\varphi_{\mathcal{U}}(u),v) = \mathrm{B}(u,\varphi_{\mathcal{U}}^{-*}(v))$.
\end{lemma}
\begin{proof}
\begin{align*}
\mathrm{B}(u,\varphi_{\mathcal{U}}^{-*}(v)) &= \mathrm{B}(\varphi (u), \varphi \varphi_{\mathcal{U}}^{-*}(v)) \\
&= \mathrm{B}(\varphi_{\mathcal{U}}(u), (\varphi_{\mathcal{U}}^* + \varphi_{\mathcal{U}}A)\varphi_{\mathcal{U}}^{-*}(v))\\
&=\mathrm{B}(\varphi_{\mathcal{U}}(u), v + \varphi_{\mathcal{U}}A\varphi_{\mathcal{U}}^{-*}(v) ) \\
&=\mathrm{B}(\varphi_{\mathcal{U}}(u), v),
\end{align*}
since $\varphi_{\mathcal{U}}A\varphi_{\mathcal{U}}^{-*}(v) \in \mathcal{U}$.
\end{proof}
In the next result we see that this gives us an isomorphism of the additive group $\mathsf{A}(\q_{\mathcal{U}},k)$.
\begin{lemma}
For $\varphi_{\mathcal{U}} \in \Orth(\q_\mathcal{U},k)$ and $A \in \mathsf{A}(\q_{\mathcal{U}},k)$ we have that
\[ \Phi_{\varphi_{\mathcal{U}}}(A) = \varphi_{\mathcal{U}} A \varphi_{\mathcal{U}}^{-*}, \]
is an isomorphism of $\mathsf{A}(\q_{\mathcal{U}},k)$.
\end{lemma}
\begin{proof}
The map $\Phi_{\varphi_{\mathcal{U}}}(A)$ is a homomorphism with respect to addition in $\mathsf{A}(\q_{\mathcal{U}},k)$ and
\begin{align*}
\q(\varphi_{\mathcal{U}} A \varphi_{\mathcal{U}}^{-*}(v)) &= \q(A \varphi_{\mathcal{U}}^{-*}(v)) \\
&= B(\varphi_{\mathcal{U}}^{-*}(v),A \varphi_{\mathcal{U}}^{-*}(v)) \\
&= B(v,\varphi_{\mathcal{U}} A \varphi_{\mathcal{U}}^{-*}(v)),
\end{align*}
by Lemma \ref{adjoint} and the fact that $\varphi_{\mathcal{U}} \in \Orth(\q_{\mathcal{U}},k)$, we have the result.
\end{proof}
\begin{prop}\label{tranv_fix_rel}
For $\varphi= \bigl[\begin{smallmatrix} \varphi_{\mathcal{U}} & \varphi_\mathcal{U}A \\ 0 & \varphi_{\mathcal{U}}^*\end{smallmatrix} \bigr]$ we have $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ if and only if $\varphi_{\mathcal{U}} \in \Orth(\q,k)$ and $\q(v+Av) = \q(\varphi_{\mathcal{U}}^*(v))$ for $v \in \overline{ \mathcal{V}}$ a nonsingular completion of $\overline{\mathcal{U}}$ in $W$.
\end{prop}
\begin{proof}
This can be shown with the following computation. Let us compute
\begin{align*}
\q(\varphi(v)) &= \q(\varphi_{\mathcal{U}}^*(v) + \varphi_{\mathcal{U}}(Av) ) \\
\q(v) &= \q(\varphi_{\mathcal{U}}^*(v)) +\q( \varphi_{\mathcal{U}}(Av) ) + \mathrm{B}(\varphi_{\mathcal{U}}^*(v), \varphi_{\mathcal{U}}(Av) ) \\
\q(v) &= \q(\varphi_{\mathcal{U}}^*(v)) +\q( Av ) + \mathrm{B}(v, Av ) \\
\q(v) +\q( Av ) + \mathrm{B}(v, Av ) &= \q(\varphi_{\mathcal{U}}^*(v)) \\
\q(v+Av) &= \q(\varphi_{\mathcal{U}}^*(v)).
\end{align*}
\end{proof}
\begin{rmk}
We note here that it is not necessarily the case that for every $\varphi_{\mathcal{U}} \in \Orth(\q_{\mathcal{U}},k)$ that there exists an element of the form $\bigl[\begin{smallmatrix} \varphi_{\mathcal{U}} & 0 \\ 0 & \varphi_{\mathcal{U}}^*\end{smallmatrix} \bigr] \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$. It is true, however, that for every $\varphi_{\mathcal{U}} \in \Orth(\q_{\mathcal{U}},k)$ there exists $A_{\varphi}: \overline{ \mathcal{V}} \to \overline{ \mathcal{U}}$ such that $\bigl[\begin{smallmatrix} \varphi_{\mathcal{U}} & \varphi_\mathcal{U}A_{\varphi} \\ 0 & \varphi_{\mathcal{U}}^*\end{smallmatrix} \bigr] \in \Orth(\q,k)^{\mathcal{I}_{\tau}} \subset \Orth(\q,k)$, which we state next.
\end{rmk}
\begin{lemma} \label{tranv_fix_exist}
If $\tau= \tau_{u_l} \cdots \tau_{u_2}\tau_{u_1}$ is a diagonal involution on $W$ of length $l$, $\dim (W)=2l$, and $\varphi_{\mathcal{U}} \in \Orth(\q_{\mathcal{U}},k)$, then there exists $\varphi=\bigl[\begin{smallmatrix} \varphi_{\mathcal{U}} & \varphi_\mathcal{U}A \\ 0 & \varphi_{\mathcal{U}}^*\end{smallmatrix} \bigr] \in \Orth(\q_{\mathcal{U}},k)^{\mathcal{I}_{\tau}}$.
\end{lemma}
\begin{proof}
We know that elements $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ leave $\overline{ \mathcal{U} }$ invariant and must be of the form
\[ \varphi= \begin{bmatrix}
\varphi_{\mathcal{U}} & A \\
0 & \varphi_{\mathcal{U}}^*
\end{bmatrix}. \]
We now consider $\varphi_{\mathcal{U'}}$ where $\mathcal{U}'$ is the basis of $\overline{\mathcal{U}}$ with respect to
\[ \q'=m\times [0,0] \perp [\q(u'_{m+1}),\q(v'_{m+1})] \perp \ldots \perp [\q(u'_l),\q(v'_l) ] \]
and $[\q(u'_{m+1}),\q(v'_{m+1})] \perp \ldots \perp [\q(u'_l),\q(v'_l) ]$ is anisotropic, so that the Witt index of $\q$ is $m$. Then we can write
\[ \varphi= \begin{bmatrix}
\varphi_{\mathcal{U'}} & A' \\
0 & \varphi_{\mathcal{U'}}^*
\end{bmatrix}, \]
where
\[ \varphi_{\mathcal{U'}} = \begin{bmatrix}
P_1 & P_2 \\
0 & \id
\end{bmatrix} \text{ which give us }
\varphi_{\mathcal{U'}}^* = \begin{bmatrix}
P_1^* & 0 \\
P_2^TP_1^* & \id
\end{bmatrix}. \]
Notice that we can choose $P_1 \in \GL_m(k)$ and $P_2 \in \Mat_{m,l-m}(k)$ since $P_1$ is acting on $\overline{ \{u_1',\ldots,u_m'\}}$ all of the vectors are orthogonal to one another and all have norm zero. The map $P_2$ adds norm zero vectors to the anisotropic vectors, and again all vectors in the space $\mathcal{U}'$ are mutually orthogonal, so there are no further constraints.
Let us define $\mathcal{U'}_{\mathbb{H}} = \{u_1',\ldots, u_m'\}$ and $\mathcal{V'}_{\mathbb{H}} = \{v_1',\ldots, v_m'\}$, the defective subspaces of $\mathcal{U}$ and $\mathcal{V}$, where $\overline{\mathcal{U}_{\mathbb{H}}' \cup \mathcal{V}_{\mathbb{H}}'}$ forms the hyperbolic subspace of $W = \overline{ \mathcal{W}}$.
We need to establish that $A'$ exists where
\[ A' = \begin{bmatrix}
A_1' & A_2' \\
A_3' & A_4'
\end{bmatrix}. \]
We can choose $A_2'=A_3'=A_4'=0$ and $A_1': \overline{\mathcal{V'}_{\mathbb{H}} } \to \overline{ \mathcal{V'}_{\mathbb{H}}}$ such that $A_1' v'_j = \q(P_2^TP_1^*v'_j)P_1 u'_j$. First notice that for $v_j'$
\begin{align*}
\q(\varphi(v_j')) &= \q(A_1'v_j' + P_1^* v_j' + P_2^TP_1^* v_j') \\
&= \q(A_1'v_j' ) + \q(P_1^*v_j' + P_2^TP_1^* v_j') + B(A_1'v_j', P_1^* v_j' + P_2^TP_1^* v_j') \\
&= \q(P_1^*v_j') + \q( P_2^TP_1^* v_j') + B(A_1'v_j', P_1^* v_j' ) \\
&= \q( P_2^TP_1^* v_j') + B(A_1'v_j', P_1^* v_j' ) \\
&= \q( P_2^TP_1^* v_j') + B(\q(P_2^TP_1^*v'_j)P_1 u'_j, P_1^* v_j' ) \\
&=0,
\end{align*}
since $A_1'v_j', P_1^* v_j' \in \mathcal{V}_{\mathbb{H}}$ and $B(u_j',v_j') = 1$. We use Proposition \ref{tranv_fix_rel} noting that $\varphi=\left[ \begin{smallmatrix} \varphi_{\mathcal{U}} & A \\ 0 & \varphi_{\mathcal{U}}^* \end{smallmatrix} \right]=\left[ \begin{smallmatrix} \varphi_{\mathcal{U}} & \varphi_{\mathcal{U}} \varphi_{\mathcal{U}}^{-1} A \\ 0 & \varphi_{\mathcal{U}}^* \end{smallmatrix} \right]$. Now if $v_j' \in \mathcal{V'}_{\mathbb{H}}$
\begin{align*}
\q(v_j' + \varphi_{\mathcal{U}}^{-1}(Av_j')) &= \q(v_j') + \q(\varphi_{\mathcal{U}}^{-1}(Av_j')) + \mathrm{B}(v_j', \varphi_{\mathcal{U}}^{-1}(Av_j')) \\
&= 0 + \q(\q(P_2 P_1^* v_j')u_j') + \mathrm{B}(v_j',\q(P_2 P_1^* v_j')u_j') \\
&= 0 + \q(P_2 P_1^* v_j')\mathrm{B}(v_j',\q(P_2 P_1^* v_j')u_j') \\
&=\q(P_2 P_1^* v_j') \\
&=\q(\varphi_{\mathcal{U}}^*(v_j')).
\end{align*}
If $v_j$ is a basis vector not in $\overline{ \mathcal{V'}_{\mathbb{H}}}$ then then $Av_j'=0$ and Proposition \ref{tranv_fix_rel} is trivially satisfied.
\end{proof}
We define $A_{\varphi} \in \Mat_{l,l}(k)$ as the diagonal matrix with $\q(\varphi_{\mathcal{U}}^* (v_j'))$ in the $j$-th diagonal entry $1\leq j \leq m$ and $0$ elsewhere, i.e. $A_{\varphi}v_j = \q(\varphi_{\mathcal{U}}^* (v_j'))u_j$ for $1\leq j \leq m$ and $A_{\varphi}v_j' = 0$ for $m+1 \leq j \leq l$.
\begin{prop} \label{AforP}
$A_{\varphi \theta} = \theta_{\mathcal{U}}^{-1} A_{\varphi} \theta_{\mathcal{U}}^* + A_{\theta}$
\end{prop}
\begin{proof}
If we let
\[ \theta_{\mathcal{U}} = \begin{bmatrix}
Q_1 & Q_2 \\
0 & \id
\end{bmatrix},\]
then we have
\[\varphi = \begin{bmatrix}
P_1 & P_2 & P_1 A_{\varphi 1} & 0 \\
0 & \id & 0 & 0 \\
0 & 0 & P_1^* & 0 \\
0 & 0 & P_2^TP_1^* & \id
\end{bmatrix} \text{ and }
\ \theta = \begin{bmatrix}
Q_1 & Q_2 & Q_1 A_{\theta 1} & 0 \\
0 & \id & 0 & 0 \\
0 & 0 & Q_1^* & 0 \\
0 & 0 & Q_2^T Q_1^* & \id
\end{bmatrix}.
\]
This gives us the product
\[\varphi \theta= \begin{bmatrix}
P_1Q_1 & P_1Q_2 + P_2 & P_1 A_{\varphi 1}Q_1 + P_1Q_1A_{\theta 1} & 0 \\
0 & \id & 0 & 0 \\
0 & 0 & P_1^*Q_1^* & 0 \\
0 & 0 & (P_2^TP_1^*+Q_2^T)Q_1^* & \id
\end{bmatrix} \]
and computing $(\theta_{\mathcal{U}}^{-1} A_{\varphi} \theta_{\mathcal{U}}^* + A_{\theta})v_j' $ for $1\leq j \leq m$
\begin{align*}
(\theta_{\mathcal{U}}^{-1} A_{\varphi} \theta_{\mathcal{U}}^* + A_{\theta})v_j' &= P_1Q_1(Q_1^{-1} A_{\varphi} Q_1^* + A_{\theta})v_j' \\
&=Q_1^{-1} A_{\varphi} Q_1^*v_j' + A_{\theta}v_j' \\
&= Q_1^{-1} \q(P_2^T P_1^* Q_1^* v_j) Q_1 u_j' + \q(Q_2^T Q_1^* v_j) u_j' \\
&= \q(P_2^T P_1^* Q_1^* v_j)u_j' + \q(Q_2^T Q_1^* v_j) u_j' \\
&= \q( (P_2^T P_1^* + Q_2^T)Q^*v_j ) u_j' \\
&= A_{\varphi \theta} v_j'.
\end{align*}
Notice that $(\theta_{\mathcal{U}}^{-1} A_{\varphi} \theta_{\mathcal{U}}^* + A_{\theta})v_j' =0$ for $m \leq j \leq l$.
\end{proof}
We make the identification
\[ \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}}A_{\varphi} \\
0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} \mapsto (\varphi_{\mathcal{U}}, 0), \]
which gives us the product
\begin{equation} \label{product_nonsing}(\varphi_{\mathcal{U}}, A)(\theta_{\mathcal{U}}, C) = (\varphi_{\mathcal{U}} \theta_{\mathcal{U}}, \theta_{\mathcal{U}}^{-1} A \theta_{\mathcal{U}}^* + C)
\end{equation}
and
\begin{equation}\label{inv_nonsing}
(\varphi_{\mathcal{U}}, A)^{-1} = (\varphi_{\mathcal{U}}^{-1}, \varphi_{\mathcal{U}} A \varphi_{\mathcal{U}}^{-*}).
\end{equation}
\begin{corollary}
$(\varphi_{\mathcal{U}}, 0 ) ( \theta_{\mathcal{U}},0) = ( \varphi_{\mathcal{U}}\theta_{\mathcal{U}}, 0)$
\end{corollary}
\begin{lemma} \label{tranv_fix_form}
If $\tau$ is a diagonal involution on $W$ of length $l$ and $\dim (W)=2l$, then $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ can be written as $(\varphi_{\mathcal{U}},A)$ where $\varphi_{\mathcal{U}} \in \Orth(\q_{\mathcal{U}},k)$ and $A \in \mathsf{A}(\q_{\mathcal{U}},k)$.
\end{lemma}
\begin{proof}
By Proposition \ref{phiUstar} we have determined that elements in $\Orth(\q,k)^{\mathcal{I}_{\tau}}$ are of the form
\[ \varphi= \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}}M \\
0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} \]
and we have made the identification
\[ \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}}A_{\varphi} \\
0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} \mapsto (\varphi_{\mathcal{U}}, 0). \]
This gives us by (\ref{product_nonsing}) that $(\varphi,A) = (\varphi,0)(\id,A)$ such that $A \in \mathsf{A}(\q_{\mathcal{U}},k)$ and we can set $M=A_{\varphi}+A$. For $v \in \overline{\mathcal{V}}$, a nonsingular completion of $\overline{\mathcal{U}}$, we check
\begin{align*}
\q(v + (A_{\varphi} + A)v) &= \q(v) + \q(A_{\varphi}v + Av) + \mathrm{B}(v, A_{\varphi}v) + \mathrm{B}(v,Av) \\
&= \q(v) + \q(A_{\varphi}v) + \q(Av) + \mathrm{B}(v, A_{\varphi}v) + \mathrm{B}(v,Av)\\
&= \q(v) + \q(A_{\varphi}v) + \mathrm{B}(v, A_{\varphi}v) \\
&=\q(v+A_{\varphi}v) \\
&= \q(\varphi^*_{\mathcal{U}}(v)) \\
\end{align*}
and so, by Proposition \ref{tranv_fix_rel} we have the result.
\end{proof}
Finally we show that all elements of the fixed point group can be decomposed as the semi-direct product of the two subgroups of $\Orth(\q,k)^{\mathcal{I}_{\tau}}$ described above.
\begin{thm}
When $\tau = \tau_{u_l} \cdots \tau_{u_2}\tau_{u_1}$ is a diagonal involution of length $l$ on a nonsingular space of dimension $2l$
\[ \Orth(\q,k)^{\mathcal{I}_{\tau}} \cong \Orth(\q_{\mathcal{U}},k) \ltimes \mathsf{A}(\q_{\mathcal{U}},k). \]
\end{thm}
\begin{proof}
By Proposition \ref{tranv_fixU} we know that $\varphi$ must leave $\overline{ \mathcal{U}}$ invariant. By Proposition \ref{phiUstar} we know that the action of $\varphi$ leaving $\overline{\mathcal{V}}$ invariant is completely determined by $\varphi_{\mathcal{U}}$. Lemma \ref{tranv_fix_exist} tells us that for any $\varphi_\mathcal{U} \in \Orth(\q_{\mathcal{U}}, k)$ there exists $(\varphi_{\mathcal{U}}, 0) \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$. Now consider any $(\varphi_{\mathcal{U}}, C) \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ then we can factor
\[ (\varphi_{\mathcal{U}},C) = (\varphi_{\mathcal{U}},0)( \id, C)\]
and we know $(\varphi_{\mathcal{U}}, 0), (\varphi_{\mathcal{U}},C) \in \Orth(\q,k)$ so we must have $( \id, C) \in \mathsf{A}(\q_{\mathcal{U}},k)$.
Now $\mathsf{A}(\q_{\mathcal{U}},k)$ is a normal subgroup of $\Orth(\q,k)^{\mathcal{I}_{\tau}}$, since
\begin{align*}
(\theta_{\mathcal{U}},D)(\id, C)(\theta_{\mathcal{U}},D)^{-1} &= (\theta_{\mathcal{U}},D)(\id, C)(\theta_{\mathcal{U}}^{-1},\theta_{\mathcal{U}}D \theta_{\mathcal{U}}^{-*}) \\
&= (\theta_{\mathcal{U}}, D+C)(\theta_{\mathcal{U}}^{-1},\theta_{\mathcal{U}}D \theta_{\mathcal{U}}^{-*}) \\
&= (\theta_{\mathcal{U}}\theta_{\mathcal{U}}^{-1}, \theta_{\mathcal{U}}(D+C)\theta_{\mathcal{U}}^{-*} + \theta_{\mathcal{U}} D \theta_{\mathcal{U}}^{-*}) \\
&= (\id, \theta_{\mathcal{U}}C \theta_{\mathcal{U}}^{-*}).
\end{align*}
\end{proof}
Here we look at the fixed point group of a diagonal involution on a nonsingular vector space $W$ with $\dim(W) = 2r$ with $2l< 2r$.
\begin{prop}
\label{nonsing_diag_fixed_pts}
An element $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ where $\tau = \tau_{\mathcal{U}}$ is a product of orthogonal transvections can be written in the form
\[ \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}} C & \varphi_{\mathcal{U}} A \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} D \\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix}, \]
where $\varphi_{\mathcal{U}} \in \Orth(\q_{\mathcal{U}},k)$, $\varphi_{\mathcal{X}} \in \mathrm{Sp}(\mathrm{B}_{\mathcal{X}},k)$ with $\varphi_{\mathcal{U}}^*$ is the unique linear map such that $\mathrm{B}(\varphi_{\mathcal{U}}(u),\varphi_{\mathcal{U}}^*(v))=\mathrm{B}(u,v)$.
\end{prop}
\begin{proof}
The map $\varphi$ acts on $W$ using ordered basis
\[ \mathcal{W} = \{u_1, \ldots, u_l, x_1,\ldots, x_{r-l}, y_1, \ldots, y_{r-l}, v_1, \ldots, v_l \} \]
where $\mathcal{U}$ is the inducing set for $\tau$, $\mathcal{V}$ is a nonsingular completion of $\mathcal{U}$ and $\tau(x) = x$ for all $x \in \overline{\mathcal{X}}$ such that
\[ \mathcal{X} = \{ x_1,\ldots, x_{r-l}, y_1, \ldots, y_{r-l} \}. \]
Notice that the blocks are zero under $\varphi_{\mathcal{U}}$ in the first column since by Proposition \ref{tranv_fixU} the subspace $\overline{ \mathcal{U}}$ is left invariant by any element in $\Orth(\q,k)^{\mathcal{I}_{\tau}}$ where
\[ \tau = \tau_{\mathcal{U}} = \tau_{u_l} \cdots \tau_{u_2} \tau_{u_1}. \]
In the second column the third block is zero since $\overline{\mathcal{U}}$ is left invariant by $\varphi$ and $\overline{\mathcal{X}} \subset \overline{\mathcal{U}}^{\perp}$.
\end{proof}
We have established $\mathrm{B}(\varphi_{\mathcal{U}}(u), \varphi_{\mathcal{U}}^*(v) ) = \mathrm{B}(u,v)$, and we will see that $C$ and $D$ have a similar relationship.
\begin{prop} \label{fix_prop_1}
Let $\mathcal{X}$ be a basis for $\left(\overline{ \mathcal{U} \cup \mathcal{V} } \right)^{\perp}$ such that $\tau(x) = x$ for all $x\in \overline{\mathcal{X}}$. The map $\varphi_{\mathcal{X}} : \overline{ \mathcal{X}} \to \overline{ \mathcal{X}}$ leaves the bilinear form invariant.
\end{prop}
\begin{proof}
Let $x,x' \in \overline{ \mathcal{X}}$, then
\begin{align*}
\mathrm{B}(x,x') &= \mathrm{B}(\varphi(x), \varphi(x')) \\
&= \mathrm{B}(\varphi_{\mathcal{X}}(x) + \varphi_{\mathcal{U}}(Cx), \varphi_{\mathcal{X}}(x') + \varphi_{\mathcal{U}}(Cx')) \\
&= \mathrm{B}(\varphi_{\mathcal{X}}(x), \varphi_{\mathcal{X}}(x'))
\end{align*}
since $\overline{ \mathcal{U}} \subset \overline{ \mathcal{U}}^{\perp}$.
\end{proof}
The group $\Orth(\q,k)^{\mathcal{I}_{\tau}}$ has a subgroup isomorphic to $\Orth(\q_{\mathcal{U}},k)$ with elements of the form
\[ \begin{bmatrix}
\varphi_{\mathcal{U}} & 0 & \varphi_{\mathcal{U}}A_{\varphi} \\
0 & \id & 0 \\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix}. \]
Next, we include Theorem 14.1 from \cite{gr02}, which is used several times. The following results lead to a description of the fixed point groups of involutions on $\Orth(\q,k)$.
\begin{thm}
\label{uniqueC}
Let $(W,\q)$ be a quadratic space with maximal nonsingular subspace $W_1$ where $2r= \mathrm{dim}(W_1)$ and $s=\mathrm{dim}(\mathrm{rad}(W))$ with $\q|_{\mathrm{rad}(W)}$ anisotropic. If $\sigma \in \Orth(W)$ then $\sigma|_{\mathrm{rad}{(W)}} = \id_{\mathrm{rad}(W)}$, $\sigma|_{W_1} = \sigma_0 + \sigma_1$, where $\sigma_0:W_1 \to \mathrm{rad}(W)$ and $\sigma_1 \in \Sp(W_1)$. If $\tau \in \mathrm{Sp}(W_1)$ then there exists $\sigma \in \Orth(W)$ with $\sigma_1 = \tau$ if and only if $\q(\tau(w)) + \q(w) \in \q(\mathrm{rad}(W))$ for all $w \in W_1$, moreover if $\sigma$ exists it is unique.
\end{thm}
\begin{prop} \label{fix_prop_2}
$\mathrm{B}(v,Cx) = \mathrm{B}(Dv,x)$.
\end{prop}
\begin{proof}
Consider
\begin{align*}
0 &= \mathrm{B}(v,x) \\
&= \mathrm{B}(\varphi(v), \varphi(x)) \\
&= \mathrm{B}(\varphi_{\mathcal{U}}(Av) + \varphi_{\mathcal{U}}^*(v) + \varphi_{\mathcal{X}}(Dv), \varphi_{\mathcal{U}}(Cx) + \varphi_{\mathcal{X}}(x)) \\
&= \mathrm{B}(\varphi_{\mathcal{X}}(Dv),\varphi_{\mathcal{X}}(x)) + \mathrm{B}(\varphi_{\mathcal{U}}^*(v), \varphi_{\mathcal{U}}(Cx)) \\
&= \mathrm{B}(Dv, x) + \mathrm{B}(v, Cx).
\end{align*}
\end{proof}
Since $D$ is completely determined by $C$ we will denote $D$ by $C^{\dagger}$. So, we can rewrite the result in Proposition \ref{fix_prop_2}
\begin{equation}
\mathrm{B}(v,Cx) = \mathrm{B}(C^{\dagger}v,x)
\end{equation}
In fact we can choose a basis for $W$ such that $\mathrm{B}_{\mathcal{U}\mathcal{V}}: \overline{\mathcal{U}} \times \overline{\mathcal{V}} \to k$ is represented by $\id$ and $\mathrm{B}_{\mathcal{X}}: \overline{\mathcal{X}} \times \overline{\mathcal{X}} \to k$ is represented by a diagonal block matrix of the form
\[ J = \begin{bmatrix}
0 & \id \\
\id & 0
\end{bmatrix} , \]
which gives us the equation $C = C^{\dagger T}J$ or in other words $C^\dagger = JC^T $. From this we can see that $(C + C')^{\dagger} = C^{\dagger} + C'^{\dagger}$
We choose the basis
\[ \mathcal{W} = \{ u_1,u_2, \ldots, u_l, x_1, x_2, \ldots, x_{r-l}, y_1, y_2, \ldots, y_{r-l
},v_1, v_2, \ldots, v_l \} \]
and write $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ as
\[ \varphi = \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}} C & \varphi_{\mathcal{U}} A \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} C^{\dagger} \\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix}. \]
The image of $x \in \overline{ \mathcal{X}}$ under $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ gives us the following relations
\begin{align*}
\q(x) &= \q(\varphi(x)) \\
&= \q(\varphi_{\mathcal{U}}(Cx) + \varphi_{\mathcal{X}}(x)) \\
&= \q(\varphi_{\mathcal{U}}(Cx)) + \q(\varphi_{\mathcal{X}}(x)) \\
&= \q(Cx) + \q(\varphi_{\mathcal{X}}(x)).
\end{align*}
So further we see that
\[ \q(x+Cx) = \q(\varphi_{\mathcal{X}}(x)). \]
From here we begin building an argument that for every pair $(\varphi_{\mathcal{X}}, C)$ such that $\q(x+Cx) = \q(\varphi_{\mathcal{X}}(x))$, we have a linear transformation $M$ such that the triple $(\varphi_{\mathcal{U}}^*, C^{\dagger}, M)$ leaves $\q$ invariant on the image of $\overline{\mathcal{V}}$ while preserving $\mathrm{B}$ on $\varphi(W)$.
\begin{lemma} \label{tranv_Sp_M}
Let $\q_{\mathcal{U}}$ be anisotropic and $T_{x,\omega} \in \GL(W)$ be of the form
\[ T_{x,\omega}= \begin{bmatrix}
\id_{\mathcal{U}} & C & M \\
0 & \tau_{x,\omega} & \tau_{x,\omega} C^{\dagger} \\
0 & 0 & \id_{\mathcal{V}}
\end{bmatrix} \]
with $\tau_{x,\omega} \in \Sp(\mathrm{B}_{\mathcal{X}},k)$ and $\q(\tau_{x,\omega} (z) + Cz) = \q(z)$ for all $z\in \mathcal{X}$. Then there exists an $M:\overline{\mathcal{V}} \to \overline{\mathcal{U}}$ such that $T_{x,\omega} \in \Orth(\q,k)$.
\end{lemma}
\begin{proof}
let $x \in \overline{ \mathcal{X}}$. To show that we have such an $M$ for the pair $(\tau_{x,\omega}, C)$, consider
\[ \begin{bmatrix}
\id & C & M \\
0 & \tau_{x,\omega} & \tau_{x,\omega} C^{\dagger} \\
0 & 0 & \id
\end{bmatrix} \]
We can assume that $x=x_1$ is a basis vector such that there exists $y_1 \in \overline{\mathcal{X}}$ with $\mathrm{B}(x_1,y_1)=1$.
Since $C: \overline{\mathcal{X}} \to \overline{\mathcal{U}}$ and $\overline{\mathcal{U}}$ is a totally singular anisotropic subspace of $W$ we know that the image of $\q_{\mathcal{U}}$ is $k^2[\q(u_1),\q(u_2),\ldots,\q(u_l) ]$, i.e. the $k^2$ vector space spanned by the norms of the elements of $\mathcal{U}$.
Assume $\q(z) = \q( \tau_{x,\omega}(z) + Cz)$, then we have $\q(z) + \q( \tau_{x,\omega}(z)) = \q(Cz)$ with $\q(Cz) \in k^2[\q(u_1),\q(u_2),\ldots,\q(u_l) ]$.
We extend $\{x_1,y_1\}$ to a basis $\{x_1,y_1, \dots, x_n,y_n\}$ for $\mathcal{X}$ such that
\[ \q_{\mathcal{X}} = [\q(x_1),\q(y_1)] \perp \cdots \perp [\q(x_n),\q(y_n)]. \]
We can now be more explicit about the image of $C$. If $\q(z) = \q(\tau_{x_1,\omega}(z))$, then $\q(Cz)=0$ and thus $Cz=0$ since $\overline{\mathcal{U}}$ is anisotropic. Now suppose $\q(z) \neq \q(\tau_{x_1,\omega}(z))$ and let $c_{jj'}$ be the coefficient of $u_j$ of $Cx_{j'}$ and $d_{jj'}$ be the coefficient of $u_j$ for $Cy_{j'}$. Since $\tau_{x_1,\omega}(x_{j'}) = x_{j'}$, $\tau_{x_1,\omega}(y_{j'}) = y_{j'}$ for $j'\neq 1$ and $\tau_{x_1,\omega}(y_{1})= y_1 + \omega x_1$ we have
\begin{align}
\displaystyle\sum_{j=1}^l c_{jj'}^2\q(u_j) &= 0 \text{ for all } j,j' \label{czero} \\
\displaystyle\sum_{j=1}^l d_{jj'}^2\q(u_j) &= 0 \text{ for all } j \text{ and all } j' \neq 1 \label{dzero} \\
\displaystyle\sum_{j=1}^l d_{j1}^2\q(u_j) &= \q(y_1) + \q(y_1 + \omega x_1) \label{dnotzero}
\end{align}
By (\ref{czero}) we have $c_{jj'}=0$ for all $j$ and $j'$ and by (\ref{dzero}) $d_{jj'}=0$ for all $j' \neq 1$. Solving for $\q(x_1)$ in (\ref{dnotzero}) we have
\[ \q(x_1) = \dfrac{1}{\omega} + \displaystyle\sum_{j=1}^l \left( \dfrac{d_{j1}}{\omega} \right)^2 \q(u_j). \]
Now we consider the image of $v_i$
\[ T_{x_1,\omega}(v_i) = v_i + \tau_{x_1,\omega}(C^{\dagger}v_i) + Mv_i, \]
and computing the norm of both sides of this equation we end up with
\begin{align*}
\q(v_i) &= \q(v_i) + \q(Mv_i) + \mathrm{B}(v_i,Mv_i) + \q(\tau_{x_1,\omega}(C^{\dagger}v_i) )\\
\q(Mv_i) + \mathrm{B}(v_i,Mv_i) &= \q(\tau_{x_1,\omega}(C^{\dagger}v_i) ) \\
\q(Mv_i) + \mathrm{B}(v_i,Mv_i) &= \q( C^{\dagger}v_i + \omega \mathrm{B}(x_1,C^{\dagger}v_i)x_1 ) \\
\q(Mv_i) + \mathrm{B}(v_i,Mv_i) &= \q( C^{\dagger}v_i)+ \q(\omega \mathrm{B}(x_1,C^{\dagger}v_i)x_1 ) + \mathrm{B}(C^{\dagger}v_i, \omega \mathrm{B}(x_1,C^{\dagger}v_i)x_1) \\
\q(Mv_i) + \mathrm{B}(v_i,Mv_i) &= \q(C^{\dagger}v_i ) + \omega^2\mathrm{B}(x_1,C^{\dagger}v_i)^2\q(x_1) + \omega \mathrm{B}(x_1,C^{\dagger}v_i)^2 \\
\end{align*}
Recall $C^{\dagger} = JC^T$ and we have that
\[ C^{\dagger}v_i = \sum d_{ij'}x_{j'} + \sum c_{ij'} y_{j'} = d_{i1} x_1, \]
since all other coefficients are zero. With the computation above we need $Mv_i$ such that
\begin{align*}
\q(Mv_i) + \mathrm{B}(v_i,Mv_i) &= \q(C^{\dagger}v_i ) + \omega^2\mathrm{B}(x_1,C^{\dagger}v_i)^2\q(x_1) + \omega \mathrm{B}(x_1,C^{\dagger}v_i)^2\\
&= \q(d_{i1}x_1 ) + \omega^2\mathrm{B}(x_1,d_{i1}x_1)^2\q(x_1) + \omega \mathrm{B}(x_1,d_{i1}x_1)^2 \\
&= \q(d_{i1}x_1 ) \\
&= \dfrac{d_{i1}^2}{\omega} + \displaystyle\sum_{j=1}^l \left(\dfrac{d_{i1}d_{j1}}{\omega}\right)^2 \q(u_j)
\end{align*}
Setting
\[ Mv_i = \displaystyle\sum_{j=1}^l \dfrac{d_{i1}d_{j1}}{\omega} u_j, \]
we see that
\begin{align*}
\q(Mv_i) + \mathrm{B}(v_i,Mv_i) &= \q\left(\displaystyle\sum_{j=1}^l \dfrac{d_{i1}d_{j1}}{\omega} u_j\right) + \mathrm{B}\left(v_i, \displaystyle\sum_{j=1}^l \dfrac{d_{i1}d_{j1}}{\omega} u_j \right) \\
&= \dfrac{d_{i1}^2}{\omega} + \displaystyle\sum_{j=1}^l \left(\dfrac{d_{i1}d_{j1}}{\omega}\right)^2 \q(u_j).
\end{align*}
Notice that this also preserves the bilinear form since
\begin{align*}
\mathrm{B}( T_{x_1,\omega}(v_i), T_{x_1,\omega}(v_{i'}) ) &= \mathrm{B}(v_i + \tau_{x_1,\omega}(C^{\dagger}v_i) + Mv_i, v_{i'} + \tau_{x_1,\omega}(C^{\dagger}v_{i'}) + Mv_{i'} ) \\
&= \mathrm{B}\left( v_i, \sum \frac{d_{i'1}d_{j1}}{\omega} u_j \right) \\
&+ \mathrm{B}(\tau_{x_1,\omega}(C^{\dagger}v_i), \tau_{x_1,\omega}(C^{\dagger}v_{i'}) ) + \mathrm{B}\left(\sum \frac{d_{i1}d_{j1}}{\omega} u_j, v_{i'} \right) \\
&=\dfrac{d_{i'1}d_{i1}}{\omega} \mathrm{B}(v_i,u_i) + \mathrm{B}(C^{\dagger}v_i, C^{\dagger}v_{i'}) + \dfrac{d_{i1}d_{i'1}}{\omega} \mathrm{B}(u_{i'},v_{i'}) \\
&= \dfrac{d_{i'1}d_{i1}}{\omega} + \mathrm{B}(d_{i1}x_1, d_{i'1}x_1) + \dfrac{d_{i1}d_{i'1}}{\omega} \\
&=0.
\end{align*}
\end{proof}
\begin{prop}
Let
\[ \varphi = \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}} C & \varphi_{\mathcal{U}} A \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} C^{\dagger} \\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} \]
such that $C: \overline{\mathcal{X}} \to \overline{\mathcal{U}}$ for $\varphi_{\mathcal{X}} \in \Sp(\mathrm{B}_{\mathcal{X}},k)$ such that $\q(x+Cx) = \q(\varphi_{\mathcal{X}})$ for all $x \in \overline{\mathcal{X}}$ then there exists an $M: \overline{\mathcal{V}} \to \overline{\mathcal{U}}$ such that
\[ \q( \varphi(v)) = \q(\varphi_{\mathcal{U}}(Mv) + \varphi_\mathcal{X}(C^{\dagger}v) + \varphi_{\mathcal{U}}^*(v)). \]
Moreover $\varphi \in \Orth(\q,k)$.
\end{prop}
\begin{proof}
Let us order the basis as before and consider
\begin{equation}
\label{proof_PXC}
\begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}} C & \varphi_{\mathcal{U}} A \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} C^{\dagger} \\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} = \begin{bmatrix}
\varphi_{\mathcal{U}} & 0 & \varphi_{\mathcal{U}} (A + M) \\
0 & \id & 0\\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix}
\begin{bmatrix}
\id & C & M \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} C^{\dagger} \\
0 & 0 & \id
\end{bmatrix}
\end{equation}
First we choose $v_j$ with $1 \leq j \leq m$ an isotropic basis vector $\q(v_j) = 0$ in $\overline{\mathcal{V}}$. We choose $Mv_j = (\q( \varphi_\mathcal{X}(C^{\dagger}v_j) ) + \q(\varphi_{\mathcal{U}}^*(v_j))) u_j$
\begin{align*}
\q(\varphi_{\mathcal{U}}(Mv_j) + \varphi_\mathcal{X}(C^{\dagger}v_j) + \varphi_{\mathcal{U}}^*(v_j)) &= \q(\varphi_{\mathcal{U}}(Mv_j)) + \q( \varphi_\mathcal{X}(C^{\dagger}v_j) ) \\
&\ \hspace{1cm} + \q(\varphi_{\mathcal{U}}^*(v_j)) + \mathrm{B}(\varphi_{\mathcal{U}}(Mv_j),\varphi_{\mathcal{U}}^*(v_j))) \\
&= \q(Mv_j) + \q( \varphi_\mathcal{X}(C^{\dagger}v_j) ) \\
&\ \hspace{2cm} + \q(\varphi_{\mathcal{U}}^*(v_j)) + \mathrm{B}(Mv_j,v_j) = 0
\end{align*}
Now let us choose $v_j$ such that $m < j \leq l$. Then $\q(v_j)\neq 0$ and $\varphi_{\mathcal{U}}^*(v_j) = v_j$. We can assume without loss of generality that $\overline{\mathcal{U}}$ is anisotropic, since any isotropic vector from $\overline{\mathcal{U} }$ would not contribute to the overall norm of $\varphi(v_j)$ being both norm zero and having a trivial bilinear form with all anisotropic basis vectors.
Since $\varphi_{\mathcal{X}} \in \Sp(\mathrm{B}_{\mathcal{X}},k)$ and $\mathcal{U}$ is anisotropic we have by Theoreom 2.1.9 in \cite{om78} that $\varphi_{\mathcal{X}}$ can be written as a product of symplectic transvections
\[ \varphi_{\mathcal{X}} = \tau_{x_N,\omega_N} \cdots \tau_{x_2,\omega_2} \tau_{x_1,\omega_1} \]
and by Theorem \ref{uniqueC} $C$ is unique. So we can factor
\[ \begin{bmatrix}
\id & C & M \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} C^{\dagger} \\
0 & 0 & \id
\end{bmatrix} =
\begin{bmatrix}
\id & C_N & M_N \\
0 & \tau_{z_N,\omega_N} & \tau_{z_N,\omega_N} C_N^{\dagger} \\
0 & 0 & \id
\end{bmatrix}\cdots
\begin{bmatrix}
\id & C_1 & M_1 \\
0 & \tau_{z_1,\omega_1} & \tau_{z_1,\omega_1} C_1^{\dagger} \\
0 & 0 & \id
\end{bmatrix}
\]
where $\varphi_{\mathcal{X}} = \tau_{z_N,\omega_N} \cdots \tau_{z_2,\omega_2}\tau_{z_1,\omega_1}$. By Lemma \ref{tranv_Sp_M} we have that $M_i$ exists for $1\leq i \leq N$. Since each $C_i$ is unique, the product on the right hand side gives us the unique $C$ and $M$ on the left hand side. The bilinear form is preserved by the product since it is preserved by each $T_{x_i,\omega_i}$.
By Lemma \ref{tranv_fix_exist} we can choose $A + M$ such that the first factor on the right hand side of (\ref{proof_PXC}) is an element of $\Orth(\q,k)$. We have constructed the second factor to be an element of $\Orth(\q,k)$, and so $\varphi \in \Orth(\q,k)$.
\end{proof}
In order to establish the isomorphism classes of the fixed point groups of involutions on nonsingular vector spaces over fields of characteristic $2$ we need a basis with the vectors organized by specific properties. Let
\[ \mathcal{U} = \{u_1,u_2, \ldots, u_l \} \]
be the set of anisotropic mutually orthogonal vectors in $W$ inducing a diagonal involution $\tau_{\mathcal{U}}$. We choose a new basis of $\overline{\mathcal{U}}$ indicating the Witt index $m$ of $\overline{\mathcal{U}}$. We call this basis
\[ \mathcal{U}' =\{ u_1', u_2', \ldots, u_m', u_{m+1}', \ldots, u_l' \}, \]
so that $\overline{\mathcal{U}} = \overline{\mathcal{U}'}$ and $\mathrm{dim}_k(\overline{\mathcal{U}}) = l$ and $\q(u_i')=0$ for $1\leq i \leq m$. We address the case when $\mathrm{dim}_k(\overline{\mathcal{U}}) = l-1$ at the end of this section.
Now we can choose $\mathcal{V}'$ to be the corresponding hyperbolic pairs making up the nonsingular subspace $\overline{\mathcal{U} \cup \mathcal{V}} = \overline{\mathcal{U}' \cup \mathcal{V}'} \subset W$ such that $\q(v_i') = 0$ for $1\leq i \leq m$ with $\dim(\overline{\mathcal{U} \cup \mathcal{V}}) = 2l$. We choose a new ordered basis $\mathcal{W}'$ for $W= \overline{\mathcal{W}}$
\[ \mathcal{W} = \{ u_1,u_2, \ldots, u_l, v_1, v_2, \ldots, v_l , x_1, x_2, \ldots, x_{r-l}, y_1, y_2, \ldots, y_{r-l} \} \]
with $\mathcal{W}' = \mathcal{U}' \cup \mathcal{X} \cup \mathcal{V}'$ with the order below
\[ \mathcal{U}' = \{u_1', u_2', \ldots u_m', u_{m+1}', \ldots, u_l' \}, \]
\[ \mathcal{X} = \{x_1, x_2, \ldots, x_{r-l}, y_1, y_2, \ldots, y_{r-l} \} \]
and
\[ \mathcal{V}'=\{v_{m+1}', \ldots, v_l', v_1', v_2', \ldots, v_m' \}. \]
so that the isotropic vectors in $\mathcal{U}'$ and $\mathcal{V}'$ are ordered first and last, respectively. We can represent an element in $\Orth(\q,k)^{\mathcal{I}_{\tau_{\mathcal{U}}}}$ as the matrix
\[ P_{\varphi}X_{\varphi}C_{\varphi} = \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}} C & \varphi_{\mathcal{U}} M \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} C^{\dagger} \\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} \]
\[ =\begin{bmatrix}
P_1 & P_2 & P_1C_1+P_2C_3 & P_1C_2+P_2C_4 & P_1M_1+P_2M_3 & P_1M_2+P_2M_4\\
0 & \id & C_3 & C_4 & M_3 & M_4 \\
0 & 0 & X_1 & X_2 & X_1C_4^T + X_2C_3^T & X_1C_2^T + X_2C_1^T \\
0 & 0 & X_3 & X_4 & X_3C_4^T + X_4C_3^T & X_3C_2^T + X_4C_1^T \\
0& 0 & 0 & 0 & \id & P_2^T P_1^* \\
0& 0 & 0 & 0 & 0 & P_1^* \\
\end{bmatrix}
\]
\begin{align} P_{\varphi} &=\begin{bmatrix}
P_1 & P_2 & 0 & 0 & 0 & A_{\varphi} \\
0 & \id & 0 & 0 & 0 & 0 \\
0 & 0 & \id & 0 & 0 & 0 \\
0 & 0 & 0 & \id & 0 & 0 \\
0& 0 & 0 & 0 & \id & P_2^T P_1^* \\
0& 0 & 0 & 0 & 0 & P_1^* \\
\end{bmatrix} \label{P} \\
X_{\varphi} &= \begin{bmatrix}
\id & 0 & 0 & 0 & 0 & 0 \\
0 & \id & C_3 & C_4 & M_{\varphi} & 0 \\
0 & 0 & X_1 & X_2 & X_1C_4^T + X_2C_3^T & 0 \\
0 & 0 & X_3 & X_4 & X_3C_4^T + X_4C_3^T & 0 \\
0& 0 & 0 & 0 & \id & 0 \\
0& 0 & 0 & 0 & 0 & \id \\
\end{bmatrix} \label{X} \\
C_{\varphi} &= \begin{bmatrix}
\id & 0 & C_1 & C_2 & M_1 & M_2 \\
0 & \id & 0 & 0 & 0 & M_4 \\
0 & 0 & \id & 0 & 0 & C_2^T \\
0 & 0 & 0 & \id & 0 & C_1^T \\
0& 0 & 0 & 0 & \id & 0 \\
0& 0 & 0 & 0 & 0 & \id \\
\end{bmatrix} \label{C}
\end{align}
We use $\mathcal{U}_{is}$ to denote the isotropic vectors in $\mathcal{U}$ and $\mathcal{U}_{an}$ to denote the anisotropic vectors in $\mathcal{U}$.
\begin{prop}
If $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ where $\tau= \tau_{\mathcal{U}}$ then $\varphi = P_{\varphi}X_{\varphi}C_{\varphi}$ from (\ref{P}),(\ref{X}) and (\ref{C}).
\end{prop}
\begin{proof}
The matrix $P_{\varphi}$ is isomorphic to an element in $\Orth(\q_{\mathcal{U}},k)$ with $A_{\varphi}$ unique. The matrix $X_{\varphi}$ is isomorphic to an element in $\Orth(\q_{\mathcal{U}_{\mathrm{an}} \cup X}, k) \cong H \subset \Sp(\mathrm{B}_{\mathcal{X}},k)$, where $[C_3 | C_4]$ is the unique linear map from Theorem 14.1 in \cite{gr02}. The maps $C_1, C_2: \overline{X} \to \overline{\mathcal{U}}_{is}$ have isotropic images in $\mathrm{rad}(\overline{\mathcal{U} \cup \mathcal{X}}) = \overline{\mathcal{U}}$ and so have zero norm and zero bilinear form on $\overline{\mathcal{U} \cup \mathcal{X}}$. Since $P_{\varphi},X_{\varphi} \in \Orth(\q,k)$ and $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ then $C_{\varphi} \in \Orth(\q,k)$ and $P_{\varphi}X_{\varphi}C_{\varphi} = \varphi$ which leaves $\overline{\mathcal{U}}$ invariant.
\end{proof}
We show that each of the matrices $P_{\varphi},X_{\varphi}$ and $C_{\varphi}$ corresponds to a subgroup of $\Orth(\q,k)^{\mathcal{I}_{\tau}}$, and we denote these subgroups by $\mathsf{P}, \mathsf{X}$ and $\mathsf{C}$ respectively. Next we see what further restrictions are necessary for $\mathsf{P},\mathsf{C}$ and $\mathsf{X}$ to be subgroups of $\Orth(\q,k)$.
For the following characterizations of these subgroups we assume that the basis for $\overline{\mathcal{U} \cup \mathcal{V}}$ has the following order
\begin{equation}
\{u_1', u_2', \ldots u_m', u_{m+1}', \ldots, u_l' , v_{m+1}', \ldots, v_l', v_1', v_2', \ldots, v_m'\}
\label{orderedbasis}
\end{equation}
where $m$ is the Witt index of $\overline{\mathcal{U} \cup \mathcal{V}}$ and $\q(u_i') = \q(v_i') = 0$ for $1 \leq i \leq m$.
\begin{prop} \label{groupC}
An element $C_{\varphi} \in \mathsf{C} \subset \Orth(\q,k)$ has $C_1, C_2 \in \Mat_{m,s-l}(k)$ and $M_1 \in \Mat_{m,l-m}(k)$, $M_4=M_1^T$ and $M_2 \in \Mat_{m,m}(k)$ such that
\[ M_2 + M_2^T = C_1C_2^T + C_2C_1^T. \]
\end{prop}
\begin{proof}
Using the ordered basis (\ref{orderedbasis}) we see that the matrix representation for $\mathrm{B}_{\mathcal{U} \cup \mathcal{V}}$ is of the form
\[ [\mathrm{B}_{ \mathcal{U} \cup \mathcal{V}}] = \begin{bmatrix}
0 & \id_{l-m} \\
\id_{m} & 0
\end{bmatrix}. \]
Now notice that $M_1: \overline{\mathcal{V}_{\mathrm{is}}} \to \overline{ \mathcal{U}_{\mathrm{an}} }$ and $M_4: \overline{ \mathcal{V}_{\mathrm{an}} } \to \overline{ \mathcal{U}_{\mathrm{is}} }$. We will compute $\mathrm{B}(Mv_i,v_j)$ for different values of $i$ and $j$. First, we notice
\begin{align*}
\mathrm{B}(Mv_i,v_j) &= (Mv_i)^T [\mathrm{B}_{ \mathcal{U} \cup \mathcal{V}}] v_j \\
&= v_i^T \begin{bmatrix}
M_1 & M_2 \\
0 & M_4
\end{bmatrix}^T
\begin{bmatrix}
0 & \id_{l-m} \\
\id_{m} & 0
\end{bmatrix} v_j \\
&= v_i^T \begin{bmatrix}
M_1^T & 0 \\
M_2^T & M_4^T
\end{bmatrix}
\begin{bmatrix}
0 & \id_{l-m} \\
\id_{m} & 0
\end{bmatrix} v_j \\
&= v_i^T
\begin{bmatrix}
0 & \id_{l-m} \\
\id_{m} & 0
\end{bmatrix}
\begin{bmatrix}
M_4^T & M_2^T \\
0 & M_1^T
\end{bmatrix} v_j
\end{align*}
so, we have the following relations
\begin{align*}
\mathrm{B}(M v_i, v_j ) &= \mathrm{B}(M_1 v_i, v_j) \\
&= \mathrm{B}(v_i, M_2^T v_j + M_1^T v_j) \\
&= \mathrm{B}(v_i, M_1^Tv_j)
\end{align*}
or
\begin{align*}
\mathrm{B}( v_i, M v_j ) &= \mathrm{B}(M_2v_j + M_4v_j, v_i) \\
&= \mathrm{B}(M_4v_j, v_i) \\
&= \mathrm{B}(v_j, M_4^Tv_i)
\end{align*}
when $1\leq j <m+1\leq i \leq l$
and
\begin{align*}
\mathrm{B}(M v_i, v_j ) &= \mathrm{B}(M_2 v_i + M_4v_i, v_j) \\
&= \mathrm{B}(M_2 v_i, v_j) \\
&= \mathrm{B}(v_i, M_2^T v_j + M_1^Tv_j) \\
&= \mathrm{B}(v_i, M_2^T v_j )
\end{align*}
when $1\leq i,j \leq m$.
Now we compute $\mathrm{B}(C_{\varphi} v_i, C_{\varphi} v_j)$ for $1\leq i,j \leq m$
\begin{align*}
0 &= \mathrm{B}(C_{\varphi} v_i, C_{\varphi} v_j) \\
&= \mathrm{B}(v_i + (C_1^T + C_2^T + M_4+M_2)v_i, v_j + (C_1^T + C_2^T + M_4+M_2)v_j) \\
&= \mathrm{B}(v_i,M_2v_j) + \mathrm{B}(C_1^Tv_i,C_2^Tv_j) + \mathrm{B}(C_2^Tv_i,C_1^Tv_j) + \mathrm{B}(M_2v_i,v_j)
\end{align*}
and so, we have
\[ \mathrm{B}(v_i, (M_2 + M_2^T) v_j) = \mathrm{B}(v_i, (C_1C_2^T + C_2C_1^T)v_j) \]
and since $\mathrm{B}$ is nondegenerate on this nonsingular space we have $M_2 + M_2^T = C_1C_2^T + C_2C_1^T$.
On the other hand, if $1 \leq i \leq m < j \leq l$ we have
\begin{align*}
0 &= \mathrm{B}(C_{\varphi} v_i, C_{\varphi} v_j) \\
&= \mathrm{B}(v_i + (C_1^T + C_2^T + M_4+M_2)v_i, v_j + M_1v_j) \\
&= \mathrm{B}(v_i,M_1v_j)+B(M_4v_i,v_j) \\
\mathrm{B}(v_i, M_1 v_j) &= \mathrm{B}(v_i, M_4^T v_j)
\end{align*}
and so, we have $M_1 = M_4^T$. Further notice that the images of $C_1,C_2$ and $M_1$ are in the totally isotropic subspace of $\overline{\mathcal{U}}$ so have zero norm and zero bilinear form values on $\overline{\mathcal{U} \cup \mathcal{X}}$.
There is a further constraint on the map $M_2$. To see this we consider the norm of the image under $C_{\varphi}$ on $v_i$ with $1 \leq i \leq m$
\begin{align*}
\q(C_{\varphi}v_i) &= \q(v_i + (C_1^T + C_2^T + M_4+M_2)v_i ) \\
\q(v_i) &= \q(v_i) + \q((C_1^T + C_2^T) v_i) + \q(M_4v_i+ M_2v_i) + \mathrm{B}(v_i, M_4v_i + M_2v_i) \\
\mathrm{B}(v_i,M_2v_i) &= \q((C_1^T + C_2^T) v_i) + \q(M_1^Tv_i)
\end{align*}
this computation shows us that the diagonal of $M_2$ consists of entries of the form $\q((C_1^T + C_2^T) v_i) + \q(M_1^Tv_i))$.
\end{proof}
\begin{lemma}
The subgroup $\mathsf{P}$ is isomorphic to $\Orth(\q_{\mathcal{U}}, k)$.
\end{lemma}
\begin{proof}
This the same group from Lemma \ref{tranv_fix_exist} and Proposition \ref{AforP}.
\end{proof}
\begin{lemma}
The subgroup $\mathsf{X}$ is isomorphic to $\Orth(\q_{\mathcal{U}_{\mathrm{an}}' \cup \mathcal{X}},k)$
\end{lemma}
\begin{proof}
This is the group described within \cite{gr02} by Theorem 14.1 and the related discussion. Theorem 14.1 also appears in this document as Theorem \ref{uniqueC}.
\end{proof}
\begin{lemma} \label{CnormalX}
$\mathsf{X} \mathsf{C} = \mathsf{C} \mathsf{X} \cong \mathsf{X} \ltimes \mathsf{C}$.
\end{lemma}
\begin{proof}
Notice that $\mathsf{XC}$ is closed under multiplication. We show that $\mathsf{XC} = \mathsf{CX}$ by verifying that $\mathsf{C}$ is a normal subgroup of $\mathsf{XC}$. Consider $X_{\varphi} \in \mathsf{X}$ from (\ref{X}) and compute the inverse
\[ X_{\varphi}^{-1} = \begin{bmatrix}
\id & 0 & 0 & 0 & 0 & 0 \\
0 & \id & C_3X_4^T + C_4X_3^T & C_3X_2^T + C_4 X_1^T & C_3C_4^T + C_4C_3^T + M_{\varphi} & 0 \\
0 & 0 & X_4^T & X_2^T & C_4^T & 0 \\
0 & 0 & X_3^T & X_1^T & C_3^T & 0 \\
0& 0 & 0 & 0 & \id & 0 \\
0& 0 & 0 & 0 & 0 & \id \\
\end{bmatrix} \]
We define $\varphi_{\mathcal{X}}$ as the symplectic matrix $\bigl[\begin{smallmatrix} X_1 & X_2 \\ X_3 & X_4 \end{smallmatrix} \bigr]$ with the standard symplectic basis for $\overline{\mathcal{X}}$. Now $X_{\varphi} C_{\varphi} X_{\varphi}^{-1} \in \mathsf{C}$.
\end{proof}
We define the group $\mathsf{C}$ described in Proposition \ref{groupC} to be $\mathsf{A}(\q_{\mathcal{U}},k)$.
\begin{thm}
\label{orth_diag_fixed_pts}
The involution $\tau=\tau_{\mathcal{U}}$ written as a product of transvections induced by vectors $\mathcal{U}$ that are mutually orthogonal has a fixed point group $\Orth(\q,k)^{\mathcal{I}_{\tau}} \cong \Orth(\q_{\mathcal{U}},k) \ltimes ( \Orth(\q_{ \mathcal{U}_{\mathrm{an}} \cup \mathcal{X} }, k) \ltimes \mathsf{A}(\q_{\mathcal{U}},k)$.
\end{thm}
\begin{proof}
We choose a basis for $\overline{\mathcal{U}}$ such that the basis vectors corresponding to the maximal defect in $\overline{\mathcal{U}}$ are listed first and the corresponding defect in $\overline{\mathcal{V}}$, the nonsingular completion of $\overline{\mathcal{U}}$ in $W$, has its basis vectors listed last. Then by Proposition \ref{nonsing_diag_fixed_pts} $\varphi \in \Orth(\q,k)^{\mathcal{I}_{\tau}}$ has the following form
\[ \varphi = \begin{bmatrix}
\varphi_{\mathcal{U}} & \varphi_{\mathcal{U}} C & \varphi_{\mathcal{U}} M \\
0 & \varphi_{\mathcal{X}} & \varphi_{\mathcal{X}} C^{\dagger} \\
0 & 0 & \varphi_{\mathcal{U}}^*
\end{bmatrix} \]\small
\[ =\begin{bmatrix}
P_1 & P_2 & P_1C_1+P_2C_3 & P_1C_2+P_2C_4 & P_1M_1+P_2M_3 & P_1(M_2+A_{\varphi})+P_2M_4\\
0 & \id & C_3 & C_4 & M_3 & M_4 \\
0 & 0 & X_1 & X_2 & X_1C_4^T + X_2C_3^T & X_1C_2^T + X_2C_1^T \\
0 & 0 & X_3 & X_4 & X_3C_4^T + X_4C_3^T & X_3C_2^T + X_4C_1^T \\
0& 0 & 0 & 0 & \id & P_2^T P_1^* \\
0& 0 & 0 & 0 & 0 & P_1^* \\
\end{bmatrix}
\] \normalsize
Notice that this factors as $\varphi = P_{\varphi}X_{\varphi}C_{\varphi}$ from (\ref{P}), (\ref{X}) and (\ref{C}) and since $\varphi \in \Orth(\q,k)$ and we have shown that $P_{\varphi}, X_{\varphi} \in \Orth(\q,k)$ so then $C_{\varphi} \in \Orth(\q,k)$. By Lemma \ref{CnormalX} we have shown the semi-direct product relationship between $\mathsf{X}$ and $\mathsf{C}$ and in a similar way one can verify $\mathsf{XC}$ is a normal subgroup of $\mathsf{P}(\mathsf{XC})$ so we have
\[ \Orth(\q,k)^{\mathcal{I}_{\tau}} \cong \Orth(\q_{\mathcal{U}},k) \ltimes ( \Orth(\q_{ \mathcal{U}_{\mathrm{an}} \cup \mathcal{X} }, k) \ltimes \mathsf{A}_{\mathcal{U}}(\q,k)). \]
\end{proof}
The only change when the involution is hyperbolic is that the dimension of $\overline{\mathcal{U} \cup \mathcal{V}}$ is $2l-2$ instead of $2l$. We can see from this fact that there are involutions that are not $\Orth(\q,k)$-conjugate that have the same fixed point group. This provides other examples of two distinct $G(k)$-conjugacy classes with the same fixed point group over a field of characteristics $2$.
\begin{rmk}
In the case that the inducing set $\mathcal{U}$ for the involution $\tau = \tau_{\mathcal{U}}$ spans an anisotropic vector space $\overline{\mathcal{U}}$, we have $\Orth(\q_{\mathcal{U}},k) = \{ \id \}$ and the matrices $C$ and $M$ are unique up to a given $\varphi_{\mathcal{X}}$. So we have $\Orth(\q,k)^{\mathcal{I}_{\tau}} \cong \Orth(\q_{\mathcal{U} \cup \mathcal{X}}, k)$.
If $\overline{\mathcal{U}}$ is totally isotropic then $\Orth(\q_{\mathcal{U}},k) \cong \GL(\overline{\mathcal{U}})$ and $\Orth(\q_{\mathcal{U}_{\mathrm{an}} \cup \mathcal{X}},k) \cong \Orth(\q_{\mathcal{X}},k)$.
\end{rmk}
\section{General quadratic spaces}
We now remove the restriction that $W$ is nonsingular. Let $W_1$ denote a nonsingular subspace of $W$ of maximal dimension, then $W = W_1 \perp \rad(W)$. A triple $(\rho, Y,\tau)$ will denote an element in $\Orth(\q,k)$ where $\tau$ is a product of symplectic transvections, $\q(\tau(w) + \rho(Yw)) = \q(w)$ and $\rho$ is a product of basic radical involutions.
\begin{prop}
An involution in $\Orth(\q,k)$ is a map of the form
\[ (\rho, Y, \tau) = \begin{bmatrix}
\rho & \rho Y \\
0 & \tau
\end{bmatrix}, \]
where $\tau = \tau_{\mathcal{U}} \in \Sp(B_{\mathcal{U} \cup \mathcal{V}}, k)$, $\overline{\mathcal{U} \cup \mathcal{V}} \subset W_1$, $Y:W_1 \to \mathrm{rad}(W)$, $\rho \in \Orth(\q_{\mathrm{rad}(W)},k)$ and $\q(\tau(w) + \rho(Yw)) = \q(w)$ with $\tau^2 = \id$, $\rho^2 = id$ and $Y = \rho Y \tau$.
\end{prop}
\begin{proof}
The map $(\rho, Y, \tau) \in \Orth(\q,k)$ must leave $\mathrm{rad}(W)$ invariant. If you square $(\rho, Y, \tau)$ and set it equal to the identity we need $\rho^2 = \id_{\mathrm{rad}(W)}$ and $\tau^2 = \id$. In \cite{om78} O'Meara shows us that $\tau$ can be written as a product of symplectic transvections induced by mutually orthogonal vectors in $W_1$. We call this set of vectors $\mathcal{U} = \{ u_1, \ldots, u_l\}$ and consider a non-singular completion of this totally singular subspace of $W_1$ completed by the vectors $\mathcal{V}=\{ v_1, \ldots, v_l\}$ such that $\mathrm{B}(u_i,v_j) = 1$ if $i=j$ and is zero otherwise. So $\tau \in \Sp(\mathrm{B}_{\mathcal{U} \cup \mathcal{V}}, k)$. We have $\rho \in \Orth(\q_{\mathrm{rad}(W)},k)$ since the radical is left invariant by elements in $\Orth(\q,k)$. The linear transformation $Y : W_1 \to \mathrm{rad}(W)$ can exist since adding on radical vectors leaves the bilinear form invariant. The quadratic form is left invariant as long as $\q(\tau(w) + \rho(Yw)) = \q(w)$.
\end{proof}
\begin{corollary}
If $w \in W$ is such that $\tau(w) = w$ then $\rho(Yw) = Yw$.
\end{corollary}
\begin{prop}
For an involution $(\rho, Y, \tau) \in \Orth(\q,k)$ of the form
\[ (\rho, Y, \tau) = \begin{bmatrix}
\rho & \rho Y \\
0 & \tau
\end{bmatrix}, \]
where $\tau = \tau_{\mathcal{U}} \in \Sp(\mathrm{B}_{\mathcal{U} \cup \mathcal{V}}, k)$, $\overline{\mathcal{U} \cup \mathcal{V}} \subset W_1$, $Y:W_1 \to \mathrm{rad}(W)$, $\rho \in \Orth(\q_{\mathrm{rad}(W)},k)$ and $\q(\tau(w) + \rho(Yw)) = \q(w)$ we can choose a basis $\mathcal{U}'$ such that with respect to the new basis $(\rho, Y, \tau)$ is of the form
\[ \begin{bmatrix}
\rho & Y_0\\
0 & \tau_Y
\end{bmatrix} \]
with $\tau_Y \in \Orth(\q,k)$ as a product of orthogonal transvections, $Y_0v_i=0$, and $\q(Y_0 w) = 0$.
\end{prop}
\begin{proof}
Let $\tau = \tau_{a_l,u_l} \cdots \tau_{a_2,u_2}\tau_{a_1,u_1} \in \Sp(\mathrm{B}_{ \mathcal{U} \cup \mathcal{V} }, k)$. So that $l$ is the length of the transvection we will choose $a_i \neq 0$ for $1 \leq i \leq l$. Let $\mathcal{X}$ be a basis for $(\overline{ \mathcal{U} \cup \mathcal{V} })^{\perp} \cap W_1$ and notice that if $w \in \overline{ \mathcal{U} \cup \mathcal{X} }$ that $(\rho, Y, \tau)(w) = \tau(w) + \rho(Yw) = w + \rho(Yw)$ and that $\q(Yw) = 0$ since $\mathrm{B}(w,\rho(Yw)) = 0$.
Next for $v_i \in \mathcal{V}$
\begin{align*}
\q((\rho, Y, \tau)(v_i) ) &= \q(\tau(v_i) + \rho(Yv_i)) \\
\q(v_i) &= \q\left( v_i + a_i u_i + \rho(Yv_i) \right) \\
\q(v_i) &= \q(v_i) + a_i^2 \q(u_i) + a_i\mathrm{B}(v_i,u_i) + \q(Yv_i) \\
\q(Yv_i) &= a_i^2 \q(u_i) + a_i.
\end{align*}
Now we see that
\begin{align*}
(\rho, Y, \tau)(v_i) &= \tau(v_i) + \rho(Yv_i) \\
&= v_i + a_iu_i + \rho(Yv_i) \\
&= v_i + a_i\mathrm{B}\left(v_i,u_i + \frac{1}{a_i}\rho(Yv_i)\right)\left(u_i+\frac{1}{a_i}\rho(Yv_i)\right) \\
&= \tau_{a_i, u_i+\frac{1}{a_i}\rho(Yv_i)}(v_i) \\
&= \tau_{u_i+\frac{1}{a_i}\rho(Yv_i)}(v_i)
\end{align*}
since
\begin{align*}
\q\left( u_i+\frac{1}{a_i}\rho(Yv_i) \right) &= \q(u_i) + \frac{1}{a_i^2} \q(Yv_i) \\
&= \q(u_i) + \frac{1}{a_i^2} (a_i^2 \q(u_i) + a_i) \\
&= \frac{1}{a_i}.
\end{align*}
Now we define
\[ \mathcal{U}_Y = \left\{u_1 + \frac{1}{a_1}\rho(Yv_1), u_2 + \frac{1}{a_2}\rho(Yv_2), \ldots, u_l + \frac{1}{a_l}\rho(Yv_l) \right\} \]
and consider the maximal nonsingular subspace of $W$, $W_1' = \overline{\mathcal{U}_Y \cup \mathcal{X} \cup \mathcal{V} }$. Further we define
\[ \tau_Y = \tau_{u_l + \frac{1}{a_l}\rho(Yv_l)} \cdots \tau_{ u_2 + \frac{1}{a_2}\rho(Yv_2)}\tau_{ u_1 + \frac{1}{a_1}\rho(Yv_1) } \]
and $Y_0w = Yw$ for $w \in \overline{ \mathcal{U} \cup \mathcal{X} }$ and $Y_0v = 0$ for $v \in \overline{\mathcal{V}}$.
Using this basis $(\rho, Y, \tau)$ is of the form
\[ \begin{bmatrix}
\rho & Y_0\\
0 & \tau_Y
\end{bmatrix} \]
and we have $\tau_Y \in \Orth(\q,k)$ is a product of orthogonal transvections, $Y_0v_i=0$, and $\q(Y_0 w) = 0$ for all $w\in W_1'$.
\end{proof}
\begin{prop}
The involution $\left[\begin{smallmatrix}\rho & \rho Y \\ 0 & \tau \end{smallmatrix} \right]$ fixes pointwise elements of the form $u_i + \frac{1}{a_i} \rho( Y v_i)$. Moreover $Y_0 \left(u_i + \frac{1}{a_i} \rho(Yv_i) \right) = 0$.
\end{prop}
\begin{proof}
Let $\tau = \tau_{\mathcal{U}}$ and $u_i \in \mathcal{U}$ with $v_i \in \mathcal{V}$ such that $\mathrm{B}(u_i,v_i) = 1$. Since $(\rho, Y, \tau)$ is an involution we have that $\rho^2 = \id$, $\tau^2 = \id$ and $Y = \rho Y \tau$.
First, we note
\[ Yv_i = \rho( Y \tau(v_i)) = \rho (Y (v_i + a_i u_i)). \]
We compute
\begin{align*}
(\rho, Y, \tau)\left(u_i + \frac{1}{a_i} \rho( Y v_i) \right) &= \tau(u_i) + \rho(Yu_i) + \rho\left( \frac{1}{a_i} \rho(Yv_i) \right) \\
&= u_i + \rho(Yu_i) + \frac{1}{a_i} Yv_i \\
&= u_i + \rho(Yu_i) + \frac{1}{a_i}\rho (Y (v_i + a_i u_i)) \\
&= u_i + \frac{1}{a_i} \rho( Y v_i). \\
\end{align*}
Since
\[ (\rho, Y, \tau) \left( u_i + \frac{1}{a_i} \rho( Y v_i) \right) = \tau_Y \left(u_i + \frac{1}{a_i} \rho( Y v_i) \right) \]
we have $Y_0\left(u_i + \frac{1}{a_i} \rho( Y v_i) \right) = 0$.
\end{proof}
We conclude with a statement about the conjugacy classes and the fixed point groups for involutions of $\Orth(\q,k)$ in the general case for $\q$ along with remarks on special cases for a general quadratic space.
\begin{prop} \label{conj_gen_inv}
Let $(\rho_1, Y_1, \tau_1) \in \Orth(\q,k)$ be an involution such that $\tau_1 \in \Orth(\q_{W_1'}, k)$ and $\rho_1 \in \Orth(\q_{\mathrm{rad}(W)},k)$, then $(\rho_1,Y_1, \tau_1)$ is $\Orth(\q,k)$-conjugate to $(\rho_2,Y_2, \tau_2) \in \Orth(\q,k)$ if and only if
\begin{enumerate}
\item $\rho_1$ is conjugate to $\rho_2$ by $\psi \in \Orth(\q_{\mathrm{rad}(W)}, k)$
\item $\tau_1$ is conjguate to $\tau_2$ by $\mu \in \Sp(\mathrm{B}_{W_1'},k)$
\item there exists $Z:W_1' \to \mathrm{rad}(W)$ such that $\q(w)= \q(\mu(w)) + \q(Zw)$ for all $w \in W_1'$ and
\[ \psi( \rho_1 Z + Z \tau_1 + Y_1)\mu^{-1} = Y_2. \]
\end{enumerate}
\end{prop}
\begin{proof}
The list of properties in the proposition is equivalent to $\mathcal{I}_{(\psi, Z, \mu)} ( (\rho_1, Y_1, \tau_1)) = (\rho_2, Y_2, \tau_2)$. To see that we compute
\[ \begin{bmatrix}
\psi & \psi Z \\
0 & \mu
\end{bmatrix}
\begin{bmatrix}
\rho_1 & \rho_1 Y_1 \\
0 & \tau_1
\end{bmatrix}
\begin{bmatrix}
\psi^{-1} & Z\mu^{-1} \\
0 & \mu^{-1}
\end{bmatrix} =
\begin{bmatrix}
\psi \rho_1 \psi^{-1} & \psi( \rho_1 Z + Z\tau_1 + Y_1) \mu^{-1} \\
0 & \mu \tau_1 \mu^{-1}
\end{bmatrix}. \]
\end{proof}
\begin{prop}
Let $(\rho, Y, \tau) \in \Orth(\q,k)$ be an involution such that $\tau \in \Orth(\q_{W_1'}, k)$ and $\rho \in \Orth(\q_{\mathrm{rad}(W)},k)$, then $(\psi, D,\mu) \in \Orth(\q,k)$ is fixed pointwise by $\mathcal{I}_{(\rho,Y, \tau)}$ if and only if
\begin{enumerate}
\item $\psi$ is fixed by conjugation by $\rho$ in $\Orth(\q_{\mathrm{rad}(W)},k)$
\item $\mu$ is fixed by conjugation by $\tau \in \Orth(\q_{W_1'},k)$
\item there exists $Z:W_1' \to \mathrm{rad}(W)$ such that $\q(w)= \q(\mu(w)) + \q(Zw)$ for all $w \in W_1'$ and
\[ Y + \psi^{-1} Y \mu = Z + \rho Z \tau \]
\end{enumerate}
\end{prop}
\begin{proof}
The list of properties above is equivalent to an element of the form $(\psi, Z, \mu) \in \Orth(\q,k)$ being fixed by $\mathcal{I}_{(\rho, Y, \tau)}$. We can compute
\[ \begin{bmatrix}
\rho & \rho Y \\
0 & \tau
\end{bmatrix}
\begin{bmatrix}
\psi & \psi Z \\
0 & \mu
\end{bmatrix}
\begin{bmatrix}
\rho & \rho Y \\
0 & \tau
\end{bmatrix} =
\begin{bmatrix}
\rho \psi \rho & \rho \psi Y \tau + (\rho \psi Z + \rho Y \mu)\tau \\
0 & \tau \mu \tau
\end{bmatrix}. \]
So we have $\rho \psi \rho = \psi$ and $\tau \mu \tau = \mu$ right away. If we consider
\begin{align*}
\psi Z &= \rho \psi Y \tau + (\rho \psi Z + \rho Y \mu)\tau \\
&= \rho \psi Y \tau + \rho \psi Z \tau + \rho Y \mu\tau \\
&= \psi \rho Y \tau + \psi \rho Z \tau + \rho Y \tau \mu \\
&= \psi Y + \psi \rho Z \tau + Y \mu \\
Z + \rho Z \tau &= Y + \psi^{-1} Y \mu
\end{align*}
we have the result, recalling that in order for $(\rho, Y , \tau)$ to be order $2$ we need $Y= \rho Y \tau$.
\end{proof}
\begin{rmk}
When $\mathrm{rad}(W)$ is anisotropic $\Orth(\q_{\mathrm{rad}(W)},k) = \{\id \}$ and $Z$ is unique to a given $\mu$.
When $\mathrm{rad}(W)$ is totally isotropic we have $\Orth(\q_{\mathrm{rad}(W)},k) \cong \GL_s(k)$, $Z \in \Mat_{s,2r}(k)$ and $\mu \in \Orth(\q_{W_1'},k)$.
\end{rmk}
\bibliographystyle{plain}
|
\section{Introduction}
Spherical robots are mobile, ball-like robots. These bots consist of an outer shell, which rolls, with the actuation system and electrical circuitry enclosed inside this shell protected from extreme environments and external disturbances. When disturbed, these robots can reattain their equilibrium state due to their spherical shape. Due to these advantages, spherical robots are suitable for applications such as surveillance, reconnaissance, hazardous environment assessment, search, rescue, and planetary exploration as a rover. Many novel Unmanned Aerial Vehicle (UAV) operations can be facilitated by carrying a spherical robot and dropping it at a suitable location. \cite{antol2005new} discuss deployment concepts for such a surface exploration rover.
In this paper, we consider a spherical bot with a yoke-pendulum arrangement for actuation, as shown in Fig. \ref{fig:bot}. The pendulum-actuated spherical robot has a major advantage of securing its equilibrium configuration while being dropped by a UAV. The driving mechanisms determine the robot's dexterity in terms of its ability to roll in multiple directions. A variety of such mechanisms have been developed, starting from the wheel drive by \cite{509415} to the pendulum actuated drives by \cite{mahboubi2013design,Michaud,yoon2011spherical}. \citep{raura2017spherical, kalita2020dynamics} present a spherical planetary robot with a hopping mechanism to jump over rugged terrain. To address power supply limitations faced by internal-driven spherical robot to explore large region of planetary surfaces, \cite{li2011design} propose a wind-driven spherical robot with multiple shapes capable of rolling, bouncing, and flying by adjusting its external shape. Several other driving mechanisms can are discussed by \cite{crossley2006literature}, \cite{hajos2005overview} and \cite{tomik2012design}.
The forward and backward motion of the robot presented in this work is made possible by a motor attached to the yoke, which rotates the hull at the required speed. The pendulum is used to perturb the center of mass of the bot to provide steering motion. A camera, other sensors, and necessary electronics are mounted on the yoke.
Several modeling approaches have been proposed for representing the dynamics of the spherical robots in literature. First-order mathematical models of the spherical robots are based on the principle of the pure rolling constraint and conservation of angular momentum \citep{897794,844763}. The dynamics of a rolling sphere on a smooth surface have been modeled using the Lagrangian method and Euler angles \citep{rosen,ambloch}.
The dynamics of the sphere with a pendulum-based actuation have been studied in different works. \cite{liu2008family} derive a simplified dynamic model of driving ahead motion with respect to the drive motor torque as input.
A unified model for all common decoupled dynamic models of pendulum-driven spherical robots have been developed \citep{ylikorpi2014unified}. \cite{li2012dynamic} present a model using Maggi’s equations without Lagrange’s multipliers from a view of the nonholonomic constraint. \cite{decoupled} model the pendulum and yoke-based spherical robot using a decoupling approach, in which the dynamics of the rolling and the steering motion are decoupled. \cite{ylikorpi2014gyroscopic} develop the gyroscopic precision model with no-slip conditions and simulate circular motion trajectories for GimBall spherical robot. The models presented by \cite{897794}, \cite{decoupled} and \cite{cai2012path} demonstrate straight line and circular motion executed by spherical bots in simulation. \cite{lee2013design} present a probabilistic approach for generating a path between predefined start and destination locations considering the stochastic motions of spherical robots.
Results presented in these works provide a bird-eye view of the motion. Still, they do not capture the small amplitude oscillations exhibited by the spherical robots while moving in simple paths like straight or a circle. \cite{8794742} demonstrate experimental results highlighting the shaky nature of pendulum-based spherical robots for various pendulum angles. Such oscillations become crucial when a robot houses a camera/sensors to capture its surroundings. Specifically, we are interested in the bot's wobbling or sideways fluctuations (perpendicular to the heading direction) as it leads to deviations in the bot's trajectory and unsteady video feedback from the mounted camera. Hence, the objective of this paper is to model the bot's dynamics and analyze the bot's behavior when it executes linear or circular motions. This analysis provided insights into the coupling within the robot's dynamics, which results in oscillatory behaviors. Such oscillations have been generally neglected in the literature due to their relatively smaller amplitude.
\begin{figure}[t!]
\begin{subfigure}{0.2\textwidth}
\centering
\includegraphics[width=\linewidth]{Figures/bot_w_hull.eps}
\label{fig:bot1}
\caption{Assembled robot}
\end{subfigure}
\begin{subfigure}{0.25\textwidth}
\centering
\includegraphics[width=\linewidth]{Figures/bot_wo_hull.eps}
\caption{Components of the spherical robot}
\end{subfigure}
\caption{The spherical robot \label{fig:bot}}
\end{figure}
The paper is organized as follows. In the next section, we model the robot's dynamics using Euler angles and the Lagrange D'Alembert equation, considering the non-holonomic constraints. The following section shows the results of the simulation of the dynamics of the spherical bot using a numerical solver. These results are analyzed to highlight the problem of oscillations in motion. Conclusion and future work involved in this project are stated towards the end.
\section{Modeling and Dynamics}
Fig. \ref{fig:bot} shows the pendulum actuated spherical bot. The bot comprises a hollow sphere called a hull, a platform housing all components called a yoke and a pendulum. The pendulum is mounted on the yoke at the geometric center of the hull through a motor. This motor rotates the pendulum with respect to the yoke resulting in sideways motion of the bot by providing a torque $T_p$. The yoke and the hull are connected through a motor mounted on the yoke. This motor rotates the hull with respect to the yoke resulting in a forward motion of the bot by providing a torque $T_s$. The center of mass of the hull and the yoke are at the geometric center of the hull, while the pendulum's center of mass is at a distance $R_p$ from the geometric center. We assume that the bot rolls without slipping and use Lagrange D'Alembert equations to find the equations of motion.
\subsection{Reference frames and Euler Angles}
Four reference frames are introduced as shown in Fig. \ref{fig:yxz}: a global inertial frame fixed to the ground ($\mathbf{G}$), and the three frames attached to the yoke ($\mathbf{Y}$), pendulum ($\mathbf{P}$) and the hull ($\mathbf{H}$) respectively with their origins at the geometric center of the hull. The yoke frame $\mathbf{Y}$ is defined to meet the following constraints: (1) The $z$-axis of the yoke is always aligned with the $z$-axis of the hull, and (2) The $x$-axis of the yoke frame lies in the global $XZ$-plane, i.e., it always remains parallel to the ground.
The orientation of the hull at any given instant is characterized by three $YXZ$ Euler angles $\phi$, $\theta$, and $\psi$. The orientation of the pendulum requires another angle $\beta$ due to an additional degree of freedom. The transformation between frames happens as follows :
\begin{figure}[!t]
\centering
\includegraphics[width=0.48\textwidth]{Figures/figure2_new.eps}
\caption{Frames and Euler angles (YXZ)}
\label{fig:yxz}
\end{figure}
\begin{figure}[t!]
\centering
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\linewidth]{Figures/figure3a_new.eps}
\caption{Rotation about Y axis (Top view)}
\label{fig:fig3a_new}
\end{subfigure}\hfil
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\linewidth]{Figures/figure3b_new.eps}
\caption{Rotation about $x_i$ axis (Front view)}
\label{fig:fig3b_new}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\linewidth]{Figures/figure3c_new.eps}
\caption{Rotation about $x_y$ axis (Front view)}
\label{fig:fig3c_new}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\linewidth]{Figures/figure3d_new.eps}
\caption{Rotation about $z_y$ axis (Side view)}
\label{fig:fig3d_new}
\end{subfigure}
\caption{Steps for obtaining the frames}
\label{fig:fig3_new}
\end{figure}
\begin{itemize}
\item Intermediate Frame ($\mathbf{I}$): $\mathbf{G}$ is rotated along its Y-axis by angle $\phi$ to obtain $\mathbf{I}$ as shown in Figure \ref{fig:fig3a_new}. $\phi$ represents the heading angle of the robot measured with respect to the global X-axis. This angle characterises the precession of the bot.
\item Yoke frame ($\mathbf{Y}$): $\mathbf{I}$ is rotated along its local $x$-axis $x_i$ by angle $\theta$ to obtain $\mathbf{Y}$ as shown in Figure \ref{fig:fig3b_new}. $\theta$ represents the lateral tilt of the robot perpendicular to the heading direction. This angle characterises the wobbling of the robot.
\item Pendulum frame ($\mathbf{P}$): $\mathbf{Y}$ is rotated along its local $x$-axis $x_y$ by angle $\beta$ to obtain $\mathbf{P}$ as shown in Figure \ref{fig:fig3c_new}. This angle characterises the pendulum angle relative to the yoke. Note that the angle made by the pendulum with the vertical axis $Y$ is ($\beta+\theta$).
\item Hull frame ($\mathbf{H}$): $\mathbf{Y}$ is rotated along its $z$-axis $z_y$ by angle $\psi$ to obtain $\mathbf{H}$ as shown in Figure \ref{fig:fig3d_new}. $\psi$ represents the forward spin of the bot responsible for moving it forward or backwards. This angle characterises the forward rolling of the robot.
\end{itemize}
\subsection{Rotation Matrices}
The rotation matrix ${^A}R_B$ converts the velocities, angular velocities and other vectors from frame $B$ to frame $A$. Rotation matrices $R_X(\alpha)$, $R_Y(\alpha)$ and $R_Z(\alpha)$ are used to rotate vectors by an angle $\alpha$ about the x-, y-, or z-axes. The rotation matrices relating frames $\mathbf{Y}$, $\mathbf{P}$ and $\mathbf{H}$ to $\mathbf{G}$ can be written as, ${{^G}R_Y} = {R_Y(\phi)} {R_X(\theta)}$, ${{^G}R_P} = {R_Y(\phi)} {R_X(\theta)} {R_X(\beta)}$ and ${{^G}R_H} = {R_Y(\phi)} {R_X(\theta)} {R_Z(\psi)}$.
\subsection{Angular Velocities}
Vector $\vec{{^Y}\omega_X}$ denotes the angular velocity of frame X represented in Y frame. The angular velocities obtained by differentiating the rotation matrices relating that frame with the global frame are:
\begin{equation}
\vec{{^Y}\omega_Y} = \begin{bmatrix}
\dot{\theta} && \dot{\phi}\,cos(\theta) && -\dot{\phi}\,sin(\theta)
\end{bmatrix}^\intercal{}
\end{equation}
\begin{equation}
\vec{{^P}\omega_P} = \begin{bmatrix}
\dot{\beta}+\dot{\theta} &&
\dot{\phi}\,\cos(\beta+\theta) &&
-\dot{\phi}\,\sin(\beta+\theta)
\end{bmatrix}^\intercal{}
\end{equation}
\begin{equation}
\vec{{^H}\omega_H} = \begin{bmatrix}
\dot{\theta}\,cos(\psi)\,+\,\dot{\phi}\,cos(\theta)\,sin(\psi) \\
\dot{\phi}\,cos(\psi)\,cos(\theta)\,-\,\dot{\theta}\,sin(\psi) \\
\dot{\psi}\,-\,\dot{\phi}\,sin(\theta)
\end{bmatrix}
\end{equation}
\subsection{Linear Velocities}
Position vector of a point P in coordinate frame $B(oxyz)$ expressed in the global frame $G(OXYZ)$ is denoted by $\vec{{}^{G}_{B}r_{P}}$. The position vector corresponding to the center of mass of the yoke and the hull are
\begin{equation}
\vec{{}^{G}_{G}r_{Y_c}} = \vec{{}^{G}_{G}r_{H_c}} =
\begin{bmatrix}
X & R_s & Z
\end{bmatrix}^\intercal{}
\label{rHull}
\end{equation}
where $X$ and $Z$ are the co-ordinates of the hull centre along the $x$ axis and $z$ axis of the global frame respectively and $R_s$ is the radius of the sphere. Note that the origins of frames $\mathbf{Y, P, H}$ coincide at the center of the bot. The position vector of Pendulum's centre of mass in global frame expressed in global frame is given by
\begin{equation}
\vec{{}^{G}_{G}r_{P_c}} = \vec{{}^{G}_{P}r_{P_c}} + \vec{{}^{G}_{G}r_{P_o}}
\label{rPendulum}
\end{equation}
\begin{equation}
\vec{{}^{G}_{P}r_{P_c}} = {^G}R_P \vec{{}^{P}_{P}r_{P_c}}
\text{ and }\vec{{}^{P}_{P}r_{P_c}} =
\begin{bmatrix}
0 & -R_p & 0
\end{bmatrix}^\intercal{}
\end{equation}
\begin{equation}
\vec{{}^{G}_{G}r_{P_o}} = \begin{bmatrix}
X & R_s & Z
\end{bmatrix}^\intercal{}
\end{equation}
where $P_c$ stands for pendulum's centre of mass, $P_o$ stands for origin of pendulum frame and $R_p$ is the distance between $P_c$ and $P_o$.
Differentiating the position vectors, we obtain the velocities of the center of mass of the hull, yoke and the pendulum as
\begin{equation}
\vec{{^G}v_{H_c}} = \vec{{^G}v_{Y_c}} =
\begin{bmatrix}
\dot{X} & 0 & \dot{Z}
\end{bmatrix}^\intercal{}
\label{linVelHull}
\end{equation}
\begin{equation}
\vec{{^G}v_{P_c}} =
\begin{bmatrix}
\dot{X} \\ 0 \\ \dot{Z}
\end{bmatrix} + \vec{{^G}\omega_P} \times {^G}R_P \begin{bmatrix}
0 \\ -R_p \\ 0
\end{bmatrix}
\end{equation}
\subsection{Energies}
The Kinetic Energy of the system can now be written as
\begin{multline}
K = {\frac{1}{2}}(m_H \|\vec{{^G}v_{H_c}}\|^2 +
m_Y \|\vec{{^G}v_{Y_c}}\|^2 +
m_P \|\vec{{^G}v_{P_c}}\|^2) + \\
{\frac{1}{2}}(\vec{{^H}\omega_H}^\intercal {^H}I_H \vec{{^H}\omega_H} +
\vec{{^Y}\omega_Y}^\intercal {^Y}I_Y \vec{{^Y}\omega_Y} +
\vec{{^P}\omega_P}^\intercal {^P}I_P \vec{{^P}\omega_P})
\end{multline}
where $m_X$ denoted the mass of body X and ${^X}I_X$ denotes the mass moment of inertia matrix of body X calculated in its own frame of reference and can be approximated as ${^H}I_H = {\frac{2}{3}}*diag(m_H R_s^2, m_H R_s^2, m_H R_s^2)$, ${^Y}I_Y = {\frac{1}{4}} * diag(m_Y R_s^2, 2 m_Y R_s^2, m_Y R_s^2)$ and ${^P}I_P = {\frac{1}{3}} * diag(m_P R_p^2, 0,m_P R_p^2)$.
The Potential Energy of the system is given by
\begin{equation}
V = m_P g (\vec{{}^{G}_{P}r_{P_c}}\cdot{[0\text{ }1\text{ }0]}^\intercal) +
m_Y g (\vec{{}^{G}_{Y}r_{Y_c}}\cdot{[0\text{ }1\text{ }0]}^\intercal)
\end{equation}
where the datum point for potential energy is chosen as the geometric centre of the sphere.
\subsection{Non-holonomic Constraints}
The constraint of rolling without slipping is non-holonomic and is given by
\begin{equation}
\vec{{^G}v_H} = \vec{{^G}\omega_H} \times \vec{{^G}r_H}
\end{equation}
where $\vec{{^G}r_H} = \begin{bmatrix} 0 & R_s & 0
\end{bmatrix}^\intercal$ denotes the vector (written in global frame) between the stationary point on the hull which is in contact with the ground and the sphere centre. The above equation can be simplified further as
\begin{equation}
\dot{X} = R_s (\dot{\theta} sin(\phi) - \dot{\psi}cos(\phi)cos(\theta))
\label{xdot_eq}
\end{equation}
\begin{equation}
\dot{Z} = R_s (\dot{\theta} cos(\phi) + \dot{\psi}sin(\phi)cos(\theta))
\label{zdot_eq}
\end{equation}
Note that equations \ref{xdot_eq} and \ref{zdot_eq} cannot be substituted directly in equation \ref{linVelHull} and solved in Lagrangian because an important assumption while deriving the Lagragian equation is that the generalized co-ordinates are independent of each other \citep{ambloch}. Therefore the constraint equations are written in the form
\begin{equation}
\vec{a_1}^\intercal\cdot\dot{\vec{q}} = 0 \text{ , } \vec{a_2}^\intercal\cdot\dot{\vec{q}} = 0
\end{equation}
where
\begin{equation}
\dot{\vec{q}} = \begin{bmatrix}
\dot{X} & \dot{Z} & \dot{\phi} & \dot{\theta} & \dot{\psi} & \dot{\beta}
\end{bmatrix}^\intercal \\
\end{equation}
is the vector of time derivative of generalized co-ordinates and $\vec{a_1}$ and $\vec{a_2}$ are
\begin{equation}
\vec{a_1} = \begin{bmatrix}
1 & 0 & 0 & -R_s sin(\phi) & R_s cos(\phi)cos(\theta) & 0
\end{bmatrix}^\intercal \\
\label{q_a1_vec}
\end{equation}
\begin{equation}
\vec{a_2} = \begin{bmatrix}
0 & 1 & 0 & -R_s cos(\phi) & -R_s sin(\phi)cos(\theta) & 0
\end{bmatrix}^\intercal \\
\label{q_a2_vec}
\end{equation}
\subsection{Lagrange D'Alembert equations}
The Lagrangian of the system is given by $L = K - V$.
The Lagrange D'Alembert equations for non-holonomic systems are given by
\begin{equation}
\frac{d}{dt}(\frac{\partial L}{\partial \dot{q}}) - \frac{\partial L}{\partial q} = Q + \lambda_1 a_1 + \lambda_2 a_2
\end{equation}
where Q is the generalized forces, $\lambda_1$ and $\lambda_2$ are Lagrangian parameters and $a_1$ and $a_2$ are vectors are given by equations \ref{q_a1_vec}, \ref{q_a2_vec}. As $T_s$ is the torque applied for forward motion and $T_p$ is the torque applied on the pendulum, Q is given by $[0\text{ }0\text{ }0\text{ }0\text{ }T_s\text{ }T_p]^\intercal$. Rewriting the equation for each generalized co-ordinate,
\begin{equation}
\frac{d}{dt}(\frac{\partial L}{\partial \dot{X}}) = \lambda_1
\label{eqX}
\end{equation}
\begin{equation}
\frac{d}{dt}(\frac{\partial L}{\partial \dot{Z}}) = \lambda_2
\label{eqZ}
\end{equation}
\begin{equation}
\frac{d}{dt}(\frac{\partial L}{\partial \dot{\phi}}) - \frac{\partial L}{\partial \phi} = 0
\label{eqPhi}
\end{equation}
\begin{equation}
\frac{d}{dt}(\frac{\partial L}{\partial \dot{\theta}}) - \frac{\partial L}{\partial \theta} = -R_s sin(\phi)\lambda_1 -R_s cos(\phi)\lambda_2
\label{eqTheta}
\end{equation}
\begin{multline}
\frac{d}{dt}(\frac{\partial L}{\partial \dot{\psi}}) - \frac{\partial L}{\partial \psi} = R_s cos(\phi)cos(\theta)\lambda_1 - \\ R_s sin(\phi)cos(\theta) \lambda_2 + T_s
\label{eqPsi}
\end{multline}
\begin{equation}
\frac{d}{dt}(\frac{\partial L}{\partial \dot{\beta}}) - \frac{\partial L}{\partial \beta} = T_p
\label{eqBeta}
\end{equation}
Substituting equations \ref{eqX}-\ref{eqZ} into \ref{eqTheta}-\ref{eqPsi} along with the equations \ref{xdot_eq} and \ref{zdot_eq} gives the dynamic model.
Using chain rule of differentiation, the equations can be written in a consolidated form as
\begin{equation}
M(\mathbf{x},t)\dot{\mathbf{x}} = f(\mathbf{x},t)
\label{dynamics_final}
\end{equation}
where $\mathbf{x}$ is the state vector given by
\begin{equation}
{\mathbf{x}} =
\begin{bmatrix}
\phi & \theta & \psi & X & Z & \dot{\phi} & \dot{\theta} & \dot{\psi} & \dot{X} & \dot{Z} \textbf{ }\textbf{ } \beta \textbf{ }\textbf{ } \dot{\beta}
\end{bmatrix}^\intercal
\label{state_vector}
\end{equation}
The dynamics of spherical robot given by equation \ref{dynamics_final} are simulated using the in-built MATLAB solver ODE15s.
\section{Analysis of Wobbling and Precession}
In this section, we analyze the system response exhibited by the bot for a turning maneuver using the dynamic model given by equation \ref{dynamics_final}. The bot is controlled via forward rolling torque $T_s$ and pendulum torque $T_p$ to command a specific forward velocity $\dot{\psi}$ and pendulum angle $\beta$ with respect to yoke respectively.
We design a pendulum controller using $T_p$ consisting of a feed-forward term to counter-act gravity torque and a feed-back term comprising a PD controller with high gains to quickly settle $\beta$ to a desired value. We also design a speed controller using $T_s$ having a feedback term with a proportional controller with high gain so that the bot starts moving with a desired forward velocity $\dot{\psi}$. High gains in both controllers help to keep the focus on the steady-state response instead of the transients leading to it.
We present a case where the bot starts moving in a straight line. Then the pendulum is displaced to an angle that introduces wobbling, as shown in figure \ref{fig:TurningWobble} in the system while moving the bot in an arc. We finally bring back the pendulum to its initial configuration. This case represents a turning motion of the bot as shown in figure \ref{fig:TurningTrajectory}.
Based on the steady-state analysis discussed in the literature, we can expect the bot to go straight, turn in a circle and finally resume moving in a straight line when it executes a turning motion introduced above. In reality, the spherical bot wobbles along its way as observed in the experiments\footnote{The videos at \url{https://tiny.cc/SphericalBotMotion} show the turning maneuver of a spherical bot in reality vs. 3D simulations generated using the dynamic model presented in this paper}.
The system response presented in figure \ref{fig:TurningStateResponse} demonstrates that our model
qualitatively captures
wobbling and precession exhibited by spherical robots. Our mathematical model characterizes wobbling (fig \ref{fig:TurningWobble}) by the Euler angle $\theta$ about the yoke's x-axis and precession (fig \ref{fig:TurningHeading}) by the Euler angle $\phi$ about the global y-axis. Wobbling rate (fig \ref{fig:TurningWobbleRate}) and precession rates (fig \ref{fig:TurningWobbleRate}) are thus given by $\dot\theta$ and $\dot\phi$ respectively.
In the following subsections, we separately analyze the initial linear, circular, and final linear motions that constitute the bot's turning maneuver. In this setup, a fixed value of $\beta$ implies that the pendulum doesn't oscillate relative to the yoke. Note that the angle made by the pendulum with the vertical ($\theta+\beta$) determines the gravitational torque $\vec{\tau}$ experienced by the bot that acts along the yoke's x-axis. This torque is one of the main drivers behind the precession of the bot. A 3D visualization\footnote{See footnote 1} of the simulation demonstrate the pose of the spherical robot using generalised coordinates $\phi$, $\theta$, $\psi$, $\beta$, $X$ and $Z$.
\begin{figure}
\centering
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/trajTurn.eps}
\caption{\centering{Bot's trajectory}}
\label{fig:TurningTrajectory}
\end{subfigure}\hfil
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/betaTurn.eps}
\caption{\centering{Pendulum angle vs. time}}
\label{fig:TurningBeta}
\end{subfigure}
\medskip
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/thetaTurn.eps}
\caption{\centering{Wobbling vs. time }}
\label{fig:TurningWobble}
\end{subfigure}\hfil
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/thetadTurn.eps}
\caption{\centering{Wobbling rate vs. time}}
\label{fig:TurningWobbleRate}
\end{subfigure}
\medskip
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/phiTurn.eps}
\caption{\centering{Precession vs. time}}
\label{fig:TurningHeading}
\end{subfigure}\hfil
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/phidTurn.eps}
\caption{\centering{Precession rate vs. time}}
\label{fig:TurningHeadingRate}
\end{subfigure}
\medskip
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/pendTurn.eps}
\caption{\centering{Pendulum angle wrt vertical vs. time}}
\label{fig:TurningPendulum}
\end{subfigure}\hfil
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.85\linewidth, trim = 1cm 0 1cm 0]{Figures/psidTurn.eps}
\caption{\centering{Forward rolling angular speed vs. time}}
\label{fig:TurningPsid}
\end{subfigure}
\caption{Analysis for Turning motion}
\label{fig:TurningStateResponse}
\end{figure}
\subsection{Initial straight line motion: t < 5 sec}
The bot is disturbed from its initial equilibrium configuration by imparting a forward velocity $\dot\psi$ of 1 rad/s using $T_s$. The bot quickly attains that value, as shown in figure \ref{fig:TurningPsid}. The bot continues to move at that speed in a straight line for 5 seconds, as shown in figure \ref{fig:TurningTrajectory}. $\dot{X}_0$ also stays non-zero due to pure rolling constraints. The bot does not wobble or precesses in this period, as shown in figures \ref{fig:TurningWobble} and \ref{fig:TurningHeading}. The pendulum also stays aligned with the vertical (ref. fig \ref{fig:TurningPendulum}.), and the pendulum stays perpendicular to the yoke as $\beta$ remains zero initially (ref. fig \ref{fig:TurningBeta}).
\subsection{Circular motion: 5 sec < t < 20 sec}
The bot's pendulum angle relative to the yoke $\beta$ is changed to 15$^{\circ}$ after moving in a straight line for 5 seconds by varying $T_p$. $\beta$ is further kept fixed for some duration as shown in (ref. fig \ref{fig:TurningBeta}). Bot's forward velocity $\dot\psi$ is left unchanged (ref. fig \ref{fig:TurningPsid}). As a result, the robot starts tracing a circular trajectory as shown in fig \ref{fig:TurningTrajectory}.
\subsubsection{Wobbling}
Figure \ref{fig:TurningWobble} shows the evolution of $\theta$ with time. The bot experiences wobbling because the pendulum motor disturbs the bot's stable configuration by changing the pendulum angle. To satisfy the pure rolling constraints, the oscillations in $\theta$ give rise to lateral displacements perpendicular to the heading direction of the bot when it is moving in a circle.
\subsubsection{Precession}
Figure \ref{fig:TurningHeading} shows the evolution of $\phi$ with time. It keeps increasing with time and also has an oscillatory component. The torque $\vec{\tau}$ generated due to the pendulum's weight contributes to both wobbling and precession. At high forward speed $\dot\psi$, precession dominates over wobbling as $\vec{\tau}$ is approximately perpendicular to the angular momentum $\vec{\mathbf{L}}$ that acts along the z-axis of the yoke and exists due to non-zero $\dot\psi$. At low forward speed $\dot\psi$, wobbling dominates over precession as $\vec{\tau}$ contributes to the change in $\theta$. In this case, bot's $\vec{\mathbf{L}}$ turns out to be either approximately zero or is parallel to yoke's x-axis if $\dot\theta$ is non-zero. In the simulation performed here, a moderate forward speed $\dot\psi$ (ref. fig \ref{fig:TurningPsid}) was simulated to observe a combination of both behaviors.
The linearly increasing behavior of $\phi$ can be attributed to the non-zero average value of ($\beta+\theta$) as seen in figure \ref{fig:TurningPendulum}. This ensures that the mean position of the angle made by the pendulum with the vertical is non-zero. Due to this, $\vec{\tau}$ stays approximately unchanged. As a result, the bot precesses in an arc of approximately constant radius.
The oscillations in $\phi$ occur due to the sinusoidal nature of ($\beta+\theta$) about a non-zero mean position, as seen in figure \ref{fig:TurningPendulum}. This implies that the pendulum-yoke-hull assembly oscillates about its mean position giving rise to a minor fluctuation in $\vec{\tau}$. This leads to variation in the bot's precession.
\subsection{Final straight line motion: t > 20 sec}
The bot's pendulum is finally made perpendicular to the yoke by changing $\beta$ back to 0$^{\circ}$ after by varying $T_p$. $\beta$'s position is maintained for the rest of the simulation time (ref. fig \ref{fig:TurningBeta}). Bot's forward velocity $\dot\psi$ is left unchanged (ref. fig \ref{fig:TurningPsid}). As a result, the robot moves in a straight line tangentially to the circular arc it executed, as shown in fig \ref{fig:TurningTrajectory}.
\subsubsection{Wobbling}
Figure \ref{fig:TurningWobble} shows the evolution of $\theta$ with time. Even though the pendulum becomes perpendicular to the yoke, the bot oscillates as it has already been displaced from its equilibrium configuration. Note that the wobbling amplitude is greater than when it was moving in a circle. To satisfy the pure rolling constraints, the wobbling nature of the bot also induced a lateral motion perpendicular to its heading direction, as observed in Figure \ref{fig:TurningTrajectory}.
\subsubsection{Precession}
Figure \ref{fig:TurningHeading} shows the evolution of $\phi$ with time. The net change in $\phi$ over time averages out to be zero because of its oscillatory behavior. In this simulation, the angular momentum $\vec{\mathbf{L}}$ is dominated by $\dot\psi_0$ in comparison to $\dot\theta_0$ by an order of 10 (refer figure \ref{fig:TurningPsid} and \ref{fig:TurningWobbleRate}). Hence, the gravitational torque $\vec{\tau}$ due to the pendulum is approximately perpendicular to $\vec{\mathbf{L}}$. This results in the precession of the bot. But, as the value of ($\beta+\theta$) averages out to 0 over time, as seen in Figure \ref{fig:TurningPendulum}, the torque vector keeps flipping its direction resulting in a net zero change in precession over time.
\section{Conclusion and future work}
This paper investigates the option of building novel applications combining UAVs and spherical robots. In this regard, we present a dynamic model of pendulum actuated spherical bot that accounts for the coupling between forward and sideways motion.
Our model successfully captures the oscillations of internal assembly.
Two important classes of oscillations, i.e., wobbling and precession, were defined, analyzed, and explained in this context of spherical bots. The visualization tool (simulation set-up) for spherical robot motion further aids in qualitatively comparing the system response with actual robot behavior. Modeling these oscillations is essential to open up the scope to design a controller in the future for a stabilized yoke to mount sensors on the spherical robot.
Controller for heading correction, trajectory correction, wobbling control, etc., can facilitate the uses of the spherical robot. E.g., $\phi$ and $\theta$ as output would give a heading and a wobbling controller, respectively.
|
\section{INTRODUCTION}
\label{sec:introduction}
Robotics research has long drawn inspiration from nature. Researchers have examined how animals move, communicate, and interact with their environments in order to build robots capable of performing similar tasks. Leonardo da Vinci, for instance, created the Codex on the Flight of Birds, which provided details and guidelines for achieving mechanical flight by studying the biology of birds and aerodynamics extensively. Biomimicry, the idea of designing and building technology inspired by nature, has resulted in the development of efficient robots that can adapt in response to their environment.
Our inspiration stems from the flow of a fluid around a rock offering a glimpse into how robots can navigate around obstacles in their environments. By studying the way in which fluid flows around a rock, we can gain insight into effective paths that a robot could take to achieve collision avoidance. For example, the principles of fluid dynamics can be applied to the design of a robot's movement, allowing it to move smoothly and efficiently around obstacles.
\subsection{Related Work}
Multi-Agent Systems (MASs) have been deployed in a plethora of robotics applications such as search and rescue missions \cite{murphy2016disaster, drew2021multi}, forest robotics \cite{oliveira2021advances, yuan2019learning}, surveillance \cite{acevedo2014decentralized}, and environmental monitoring \cite{dunbabin2012robots} etc, due to their significant advantages
when compared to a single agent. Such MAS must be equipped with robust algorithms that can safely navigate around both static and dynamic obstacles in order for all agents to successfully complete the cooperative job. Various collision-free path planning works have been previously published such as collision cone \cite{sunkara2019collision}, artificial potential fields \cite{koditschek1992exact}, navigation functions \cite{tanner2012multiagent}, and sampling based methods \cite{ichter2018learning}.\\
Artificial Potential Fields (APF)\cite{khatib1986real} is a simple and mathematically elegant technique originally proposed for manipulators and mobile robots in an operational space. Combining a positive potential around goal location and a negative potential around obstacles, this method guides the robot toward its goal by following the gradients of potential field while steering away from obstacles. A well-known issue of such an approach is getting trapped in local minima and in a real-world dynamic environment it has been shown that APF is inefficient \cite{montiel2015path}. Several APF variants have since been proposed for robot path planning applications \cite{kim2011path, chen2009evolutionary, goerzen2010survey, bing2011route}. \\
Control Barrier Functions (CBFs) have emerged as a potential mathematical tool for safety assurances \cite{ames2016control, jankovic2018robust}. They are a set of rules used to ensure that a robot's actions do not violate certain safety constraints, such as avoiding collisions with obstacles or other objects in the environment. Using system dynamics, CBFs can be used to define a admissible region in the robot's workspace, and the robot's control inputs are then calculated to ensure that the robot's state remains within this region at all times. CBFs for a safe behavior in multi-agent robotics was studied previously\cite{borrmann2015control} and a decentralized supervisory controller based on CBFs has been presented \cite{chen2020guaranteed}. A combination of CBFs with Control Lyapunov Functions (CLFs) via quadratic programming was studied for cruise control applications \cite{ames2014control}. Other applications include model predictive control \cite{wills2004barrier}, and bipedal robotics walking \cite{hsu2015control}.
We have recently developed a physics-based automation to safely plan large-scale coordination of multi-agent systems in the presence of disturbances and unforeseen failure(s)
\cite{rastgoftar2019physics, rastgoftar2020fault, UPPALURU2022107960}. The proposed automation consists of two operation mode that include Homogeneous Deformation Mode (HDM) and Failure Resilient Mode (FRM), where: HDM activates when all agents are healthy; FRM commands when there exists at least one noncooperative/faulty agent. We applied the principles for continuum mechanics to formalize transitions between the HDM and FRM. This multi-agent coordination approach was experimentally evaluated in Ref. \cite{romano2022quadrotor}.
\subsection{Contributions}
This work proposes a novel approach for safe and resilient coordination of multiple agent teams moving collectively in a shared motion environment. By drawing inspiration from the available ideal fluid flow models, each team considers its agents as cooperative and treat them as particles of a ideal fluid flow field whereas the other teams' agents are considered as singular points of the fluid flow field. To safely wrap the noncooperative agents and assure inter-agent collision avoidance, the cooperative agents slide along the streamlines of an ideal fluid flow field. This proposed approach will be experimentally validated using crazyflie quadcopters in an indoor flight space. Compared to the existing literature and the authors' previous work this paper offers the following key contributions:
\begin{itemize}
\item While the experimental evaluation of fluid flow navigation previously tested in the presence of a single failed quadcopter in Ref. \cite{romano2022quadrotor}, we model and experimentally validate fluid flow navigation of a multi-agent system in the presence of multiple non-concurrent failures.
\item We model and experimentally validate the fluid flow navigation of multiple agent teams simultaneously coordinating in a shared motion space.
\item We provide algorithmic approaches for the fluid flow navigation of multiple agent teams under (i) Stationary Non-Concurrent Failures (SNCF), (ii) Time-Varying Non-Cooperative (TVNC), and Time-Varying Cooperative (TVC) scenarios.
\end{itemize}
\subsection{Organization}
The remainder of the paper is organized as follows: The
detailed description of our proposed methods are presented in Section \ref{sec:methodology}. Section \ref{sec:ExperimentsAndDiscussion} describes the experimental setup and provides the results of the experiments. We finally conclude the paper in Section \ref{sec:conclusion} with thoughts about future direction.
\section{METHODOLOGY}
\label{sec:methodology}
We consider a MAS
defined by set $\mathcal{V}=\left\{1,\cdots,N\right\}$ and cluster it into $m$ separate groups, where group identification numbers are defined by set $\mathcal{M}=\left\{1,\cdots,m\right\}$. Let $\mathcal{V}_l$ define agents belonging to cluster $l\in \mathcal{M}$, then, $\bar{\mathcal{V}}_l=\mathcal{V}\setminus \mathcal{V}_l$ defines agents that do not belong to cluster $l\in \mathcal{M}$. While $\mathcal{V}$ is time-invariant, the number of agents of $\mathcal{V}_l(t)$ can change with time which in turn implies that $\mathcal{V}_l$ may either lose or absorb agents at any time $t$.
Assuming $\mathcal{V}$'s agents move in a ($2$-dimensional) plane, position of agent $i\in \mathcal{V}$ is denoted by complex variable $\mathbf{z}_i=x_i+\mathbf{j}y_i$. To safely plan a collision-free coordination of $\mathcal{V}_l$, in the presence of $\bar{\mathcal{V}}_l$'s agents, we treat ${\mathcal{V}}_l$'s agents as a finite number of particles in a time-varying ideal fluid flow field. Therefore, $\mathcal{V}_l$ considers $\bar{\mathcal{V}}_l$'s agents as singularity points in the $x-y$ plane and exclude them by using vertical cylinders obtained by defining complex function
\begin{equation}\label{MainTransformation}
\resizebox{0.99\hsize}{!}{%
$
\begin{split}
\mathbf{f}\left(\mathbf{z}_i\mathrm{e}^{-\mathbf{j}\theta_l(t)},t\right)&=\phi_l(x_i,y_i,\theta_l(t),t)+\mathbf{j}\psi_l(x_i,y_i,\theta_l(t),t)\\
&=\left(1-\beta_l\right)\mathbf{z}_i\mathrm{e}^{-\mathbf{j}\theta_l}+\beta_l\sum_{h\in \bar{\mathcal{V}}_l}\left(\left(\mathbf{z}_i\mathrm{e}^{-\mathbf{j}\theta_l}-{\mathbf{z}}_h(t)\right)+\dfrac{\Delta_h^2}{\mathbf{z}_i\mathrm{e}^{-\mathbf{j}\theta_l}-{\mathbf{z}}_h(t)}\right),
\end{split}
$
}
\end{equation}
for every $l\in \mathcal{M}$ and $i\in \mathcal{V}_l$, where $\beta_l\in \left\{0,1\right\}$ is a binary variable, $\theta_l(t)$ assigns the bulk motion direction of cluster $l\in \mathcal{M}$, and $\Delta_h\in \mathbb{R}_+$ is chosen such that the size of agent $h\in \bar{\mathcal{V}}_l$ is properly incorporated. Also, $\phi_l(x_i,y_i,\theta_l(t),t)$ and $\psi_l(x_i,y_i,\theta_l(t),t)$ are called \textit{potential function} and \textit{stream function}, respectively. We note that stream and potential function satisfy the Cauchy-Riemann conditions which in turn implies that the level curves $\phi_l$-constant and $\psi_l$-constant are perpendicular at intersection points in the $x-y$ plane.
For the sake of simplicity, we use $\phi_{il}(t)$ and $\psi_{il}(t)$ to denote $\phi_{il}(t)=\phi_l(x_i,y_i,\theta_l,t)$ and $\psi_{il}(t)=\psi_l(x_i,y_i,\theta_l,t)$ in the continuation of this paper.
Equation \eqref{MainTransformation} establishes a nonsingular transformation between $\mathbf{z}_i=x_i+\mathbf{j}y_i$ and $\phi_{il}+\mathbf{j}\psi_{il}$ for every $i\in \mathcal{V}_l$ and $l\in \mathcal{M}$. Therefore, $(x_i,y_i)$ can be uniquely obtained based on $\left(\phi_{il}(t),\psi_{il}(t)\right)$ at any time $t$ by
\begin{subequations}
\begin{equation}\label{xvshs}
x_i=g_1\left(\phi_{il},\psi_{il},\theta_l,t\right).
\end{equation}
\begin{equation}\label{yvshs}
y_i=g_2\left(\phi_{il},\psi_{il},\theta_l,t\right).
\end{equation}
\end{subequations}
Note that two different agents $i\in \mathcal{V}_l$ and $j\in \mathcal{V}_l$ do not collide if: (i) they are properly distanced in the reference configuration; and (ii) their desired trajectories can be stably tracked by the on-board tracking (position) controller. Furthermore, every agent $i\in \mathcal{V}_l$ can avoid inter-agent collision and hitting $\bar{\mathcal{V}}_l$'s agents when it slides along level curve $\psi_l\left(x_i,y_i,\theta_l(t),t\right)=\bar{\psi}_{i,l}(t)$ constant \cite{UPPALURU2022107960}. Therefore, the tangent vector to the desired sliding path of agent $i\in \mathcal{V}_l$ is obtained by
\begin{equation}
\hat{\mathbf{T}}_i(t)=\begin{bmatrix}
\dfrac{\partial \psi_l\left(x_i,y_i,\theta_l,t\right)}{\partial y}&- \dfrac{\partial \psi_l\left(x_i,y_i,\theta_l,t\right)}{\partial x}
\end{bmatrix}
^T,
\end{equation}
for every $i\in \mathcal{V}_l$ and $l\in \mathcal{M}$ (See Fig. \ref{connectionschematic}).
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{motionstrategy.png}
\caption{Schematic of the desired path of agent $i\in \mathcal{V}_l$}
\label{connectionschematic}
\end{figure}
The desired velocity of agent $i\in \mathcal{V}_l$ is given by
\begin{equation}
\mathbf{V}_i=v_l\hat{\mathbf{T}}_i,\qquad \forall i\in \mathcal{V}_l,~l\in \mathcal{M},
\end{equation}
and we use the Algorithm \ref{alg1} to update $\mathbf{z}_i$ at any time $t$.
\begin{table}[]
\caption{Properties of the investigated fluid-flow navigation problems.}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Scenarios & $\theta_l~(l\in \mathcal{M})$ & $\mathcal{V}_l~(l\in \mathcal{M})$ & $\beta_l~(l\in \mathcal{M})$&$m$\\
\hline
SNCF & Time-Invariant & Time-Varying & $\beta_l=1$ & $m=2$\\
TVNC & Time-Varying & Time-Invariant & $\beta_l=1$ & $m=2$\\
TVC & Time-Varying & Time-Invariant & $\beta_l\in \left\{0,1\right\}$ & $m>1$\\
\hline
\end{tabular}
\label{table:Properties}
\end{table}
\begin{algorithm}
\caption{Position Update Algorithm for Every Cluster $l\in \mathcal{M}$ under Fluid-Flow Navigation Strategy.}\label{alg1}
\begin{algorithmic}[1]
\State \textit{Get:} Time increment $\Delta t$, $\Delta_h$ and $\mathbf{z}_h$ for every $h\in \bar{\mathcal{V}}_l$, $\beta_l\in \left\{0,1\right\}$, $\theta_l$ sliding speed $v_l$, and current position $\mathbf{z}_i$ of every agent $i\in \mathcal{V}_l$.
\State \textit{Obtain:} Next position $\mathbf{z}'_i=x'_i+\mathbf{j}y'_i$.
\For{\texttt{ $i\in \mathcal{V}_l$}}
\State Compute current $\phi_{il}\left(t\right)$ and $\psi_{il}(t)$ using Eq. \eqref{MainTransformation}.
\State Compute next potential $\phi'_{il}$: $\phi_{il}=\phi_{il}+v_l\Delta t$.
\State Compute next stream $\psi'_{il}$: $\psi_{il}=\psi_{il}$.
\State Compute next $x'_i$: $ x'_i=g_1\left(\phi'_{il},\psi'_{il},\theta_l,t\right)$.
\State Compute next $y'_i$: $ y'_i=g_2\left(\phi'_{il},\psi'_{il},\theta_l,t\right)$.
\EndFor
\end{algorithmic}
\end{algorithm}
Given above problem setting and foundations, this paper implements and experimentally evaluates collision-free fluid flow navigation of multiple groups of agents under (i) Stationary Non-Concurrent Failures (SNCF), (ii) Time-Varying Non-Cooperative (TVNC), and Time-Varying Cooperative (TVC) scenarios, with the properties listed in Table \ref{table:Properties}.
\subsection{SNCF Fluid Flow Navigation}
\label{subsec:SNCF}
For the SNCF fluid flow navigation model, set $\mathcal{V}$ is divided into $\mathcal{V}_1$ and $\mathcal{V}_2=\bar{\mathcal{V}}_1$, where $\mathcal{V}_1(t)$ and $\mathcal{V}_2(t)$ are disjoint subsets of $\mathcal{V}$ defining the ``healthy'' and ``faulty'' agents, respectively, at time $t$.
\begin{definition}
We define $t_{\mathrm{fail}}\geq t_0$ as the \textbf{most recent time} when the status of failure of the agents has changed.
\end{definition}
For the SNCF coordination, we make the following assumptions:
\begin{assumption}\label{assum1}
Angle $\theta_1(t)$ is constant at any time $t\geq t_0$, where $t_0$ is the start time of the SNCF coordination.
\end{assumption}
\begin{assumption}\label{assum2}
Every faulty agent $h\in \mathcal{V}_2$ remains inside a stationary vertical cylinder after its failure is detected.
\end{assumption}
For the SNCF, we define potential and stream functions only for the healthy agents. Thus, potential function $\phi_1$, stream function $\psi_1$, $\beta_1$, and $\theta_1$ are denoted by $\phi$, $\psi$, $\beta$, and $\theta$, respectively, and substituted in Eq. \eqref{MainTransformation} to compute the potential and stream functions under SNCF. By imposing Assumptions \ref{assum1} and \ref{assum2}, potential function $\phi(x,y,t_{\mathrm{fail}})$ and stream $\psi(x,y,t_{\mathrm{fail}})$ are piece-wise time-invariant and remains spatially-varying at any time $t\geq t_{\mathrm{fail}}$ until the status of agents' failures change. We apply Algorithm \ref{alg2} to safely plan coordination of healthy agents under the SNCF strategy.
\begin{algorithm}
\caption{Algorithm for SNCF Fluid-Flow Navigation.}\label{alg2}
\begin{algorithmic}[1]
\State \textit{Get:} Initial time $t_0$, number of time steps denoted by $n$, time increment $\Delta t$, healthy agent set $\mathcal{V}_1(t_0)$, faulty agent set ${\mathcal{V}}_2(t_0)$,
time increment $\Delta t$, $\Delta_h=\Delta$ for every $h\in {\mathcal{V}}_2$, $\theta_l$, $v_l$, initial position $\mathbf{z}_i(t_0)$ of every healthy agent $i\in \mathcal{V}_1$, and initial position $\mathbf{z}_h(t_0)$ of every faulty agent $h\in \mathcal{V}_2$
\State \textit{Set:} $\beta=1$, $k=1$, $t_h=t_0$.
\While{$k < n$}
\If{$\mathcal{V}_2\left(t_k\right)\neq \mathcal{V}_2\left(t_{k-1}\right)$}
\State $t_{\mathrm{fail}}\leftarrow t_k$
\State Update $\phi\left(x,y,t_{\mathrm{fail}}\right)$.
\State Update $\psi\left(x,y,t_{\mathrm{fail}}\right)$.
\EndIf
\State $\phi\left(x,y,t_k\right)\leftarrow \phi\left(x,y,t_{\mathrm{fail}}\right)$.
\State $\psi\left(x,y,t_k\right)\leftarrow \psi\left(x,y,t_{\mathrm{fail}}\right)$.
\State Obtain $\mathbf{z}_i(t_{k+1})$ for every $i\in \mathcal{V}_1(t_k)$ by Algorithm \ref{alg1}.
\State Return $\mathbf{z}_i(t_{k+1})$.
\State $\mathbf{z}_i(t_{k})\leftarrow \mathbf{z}_i(t_{k+1})$ for every $i\in \mathcal{V}_1(t_k)$.
\State $k\leftarrow k+1$.
\EndWhile
\end{algorithmic}
\end{algorithm}
\subsection{TVNC Fluid Flow Navigation}
\label{subsec:TVNC}
For the TVNC fluid flow navigation model, set $\mathcal{V}$ is divided into time-invariant subsets $\mathcal{V}_1$ and $\mathcal{V}_2=\bar{\mathcal{V}}_1$, where $\mathcal{V}_1$ and $\mathcal{V}_2$ define ``cooperative'' and ``non-cooperative'' agents, respectively. The noncooperative agents have a predefined trajectories in the motion space whereas the cooperative agents uses the fluid flow navigation model, presented in Algorithm \ref{alg1}, to safely update their positions and reach their target positions.
Similar the SNCF coordination model, we denote potential function $\phi_1$, stream function $\psi_1$, $\beta_1$, and $\theta_1$ by $\phi$, $\psi$, $\beta$, and $\theta$, respectively, and substitute them in Eq. \eqref{MainTransformation} to compute the potential and stream functions. Let $\mathbf{z}_{i,f}=x_{i,f}+\mathbf{j}y_{i,f}$ be the known target position of cooperative agent $i\in \mathcal{V}_1$, then, then angle $\theta$, assigning the bulk motion direction of $\mathcal{V}_1$ is obtained by
\begin{equation}\label{thetasecond}
\theta(t)=\theta_l=\arg \left(\sum_{i\in \mathcal{V}_1}\left(\mathbf{z}_{i,f}-\mathbf{z}_i(t)\right)\right),\qquad \forall t\geq t_0,~l\in \mathcal{M}
\end{equation}
where $t_0$ is the initial time. We use Algorithm \ref{alg3} to safely plan coordination of every agent $i\in \mathcal{V}$ in a shared motion space.
\begin{algorithm}
\caption{Algorithm for TVNC Fluid-Flow Navigation.}\label{alg3}
\begin{algorithmic}[1]
\State \textit{Get:} Initial time $t_0$, time increment $\Delta t$, $\epsilon>0$, $\Delta_h$ for every $h\in \bar{\mathcal{V}}_l$, $\theta_l$, $v_l$, initial position $\mathbf{z}_i(t_0)$ of every cooperative agent $i\in \mathcal{V}_1$.
\State \textit{Set:} $\beta=1$, $k=0$.
\While{$\sum_{i\in \mathcal{V}_1}\left|\mathbf{z}_i(t_k)-\mathbf{z}_{i,f}\right|> \left|\mathcal{V}_1\right|\epsilon$ for every $i\in \mathcal{V}_1$}
\State Get $\mathbf{z}_h\left(t_k\right)$ for every $h\in \mathcal{V}_2$.
\State Compute $\theta(t_k)$ by Eq. \eqref{thetasecond}.
\State Update $\phi\left(x,y,\theta(t_k),t_k\right)$ and $\psi\left(x,y,\theta(t_k),t_k\right)$.
\State Obtain $\mathbf{z}_i(t_{k+1})$ for every $i\in \mathcal{V}_1$ by Algorithm \ref{alg1}.
\State Return $\mathbf{z}_i(t_{k+1})$.
\State $\mathbf{z}_i(t_{k})\leftarrow \mathbf{z}_i(t_{k+1})$ for every $i\in \mathcal{V}_1$.
\State $k\leftarrow k+1$.
\EndWhile
\end{algorithmic}
\end{algorithm}
\subsection{TVC Fluid Flow Navigation}
\label{subsec:TVC}
For the TVC Fluid flow navigation, we enable every agent $i\in \mathcal{V}_l$ to check if there is a possibility of colliding with an agent $h\in \bar{\mathcal{V}}_l$ within the next $n_\tau$ time steps. To this end, we define virtual box $\mathcal{B}_i(t)\subset \mathbb{C}$ for every agent $i\in \mathcal{V}_l$, with side lengths $2\delta$ and $v_ln_\tau
\Delta t$, to check possibility of collision with an agent $j\in \bar{\mathcal{V}}_l$ within the next $n_\tau \Delta t$ seconds. To formally specify collision avoidance condition, we define condition $\zeta$ as follows:
\begin{equation}\label{motionspacecontainment}
\displaystyle\bigvee_{l\in \mathcal{M}}\displaystyle\bigvee_{i\in \mathcal{V}_l}\bigvee_{j\in \bar{\mathcal{V}}_l}\left(\mathbf{z}_j\in \mathcal{B}_i\right),
\tag{$\zeta$}
\end{equation}
where ``$\bigvee$'' is used to specify ``at least one''.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{collisionavoidance.png}
\caption{The virtual box $\mathcal{B}_i$ with side lengths $2\delta$ and $v_ln_\tau\Delta t$ used by agent $i\in \mathcal{V}_l$ to estimate the possibility of colliding an agent $j\in \bar{\mathcal{V}}_l$.}
\label{connectionschematic}
\end{figure}
Note that $\zeta$ is satisfied, if there exists at least one agent $j\in \bar{\mathcal{V}}_l$ that is inside one of the safety boxes of $\mathcal{V}_l$'s agents.
Therefore, $\beta_l$ is specified as follows:
\begin{subequations}
\begin{equation}
\zeta \implies \bigwedge_{l\in \mathcal{M}}\left(\beta_l=1\right),
\end{equation}
\begin{equation}
\neg \zeta \implies \bigwedge_{l\in \mathcal{M}}\left(\beta_l=0\right),
\end{equation}
\end{subequations}
where ``$\bigwedge$'' is used to specify ``include all''; ``$\implies$'' means ``implies that''; and ``$\neg$'' is the ``negation'' symbol.
For the TVC, we use Algorithm 4 to safely plan coordination of every agent $i\in \mathcal{V}$ in a shared motion space.
\begin{algorithm}
\caption{Algorithm for TVC Fluid-Flow Navigation.}\label{alg4}
\begin{algorithmic}[1]
\State \textit{Get:} Initial time $t_0$, number of sample times denoted by $n$, time increment $\Delta t$, $\Delta_h$ for every $h\in \bar{\mathcal{V}}_l$, $v_l$, $\delta$, $n_\tau$, initial position $\mathbf{z}_i(t_0)$ of every agent $i\in \mathcal{V}_l$ and every cluster $l\in\mathcal{M}$.
\State \textit{Set:} $k=0$.
\State \textit{Set:} $\beta_l(t_0)=0$, $\cdots$, $\beta_l(t_n)=0$ for every $l\in \mathcal{M}$.
\For{\texttt{ $k\in \left\{0,\cdots,n\right\}$}}
\For{\texttt{ $l\in \mathcal{M}$}}
\If{$\zeta$~ is satisfied}
\State $\beta_l\left(t_k\right)=1$.
\State Get $\mathbf{z}_h\left(t_k\right)$ for every $h\in \bar{\mathcal{V}}_l$.
\EndIf
\State Compute $\theta_l(t_k)$ by Eq. \eqref{thetasecond}.
\State Update $\phi_l\left(x,y,\theta(t_k),t_k\right)$ and $\psi_l\left(x,y,\theta(t_k),t_k\right)$.
\State Obtain $\mathbf{z}_i(t_{k+1})$ for every $i\in \mathcal{V}_l$ by Algorithm \ref{alg1}.
\State Return $\mathbf{z}_i(t_{k+1})$.
\State $\mathbf{z}_i(t_{k})\leftarrow \mathbf{z}_i(t_{k+1})$ for every $i\in \mathcal{V}_l$.
\EndFor
\State $k\leftarrow k+1$.
\EndFor
\end{algorithmic}
\end{algorithm}
\section{EXPERIMENTS AND DISCUSSION}
\label{sec:ExperimentsAndDiscussion}
We experimentally evaluate the performance of the proposed algorithms and validate the results on a group of tiny quadcopters. Tthe multi-media of our experiments is available at YouTube (Please see https://www.youtube.com/watch?v=5jAmiUvzsCU).
\subsection{Experiment Setup}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{system.png}
\caption{An overview of the experiment setup}
\label{fig:experiment_setup}
\end{figure}
This section mainly provides a high level overview of the environment and hardware setup which consists of $4$ major components shown in Fig. \ref{fig:experiment_setup}:
\begin{itemize}
\item Motion Capture System (MCS)
\item Ground Control Station (GCS)
\item Crazyradio PA
\item Crazyflie $2.1$
\end{itemize}
The MCS captures the position of the each crazyflie in $3$-dimensional space and sends the information to the GCS through an Ethernet cable at \SI{100}{Hz}. The GCS is an Intel i$7$ $11$-th gen desktop, with $16$ GB of RAM running Ubuntu $20.04$ and ROS Noetic. GCS is also installed with the Crazyswarm\footnote{\url{https://crazyswarm.readthedocs.io/en/latest/}} \cite{preiss2017crazyswarm} ROS stack built by USC-ACT Lab and acts as a centralized planner for the system. The GCS uses the information from MCS to compute the desired states for each crazyflie and then transmits the data to the onboard controller through Crazyradio PA (Fig. \ref{fig:devices}(b)). The Crazyflie $2.1$ (CF) (Fig. \ref{fig:devices}(a)) is equipped with two microcontroller units (MCUs) capable of inertial state estimation, control tasks, radio and power management and a single motion capture marker. Each Crazyflie can fly for up to $7$ minutes on a single charge of its $240$mAh LiPo battery. The onboard controller calculates the necessary motor values (PWM signals) and send them to the onboard motors. We conducted the flight tests at the University of Arizona's Scalable Move and Resilient Transversability (SMART) lab's indoor flying area with a volume of \SI{5}{m} $\times$ \SI{5}{m} $\times$ \SI{2}{m} equipped with $8$ VICON motion capture cameras.
\begin{figure}
\centering
\subfigure[]{\includegraphics[height=4cm, width=0.48\linewidth]{cf_lights_axis.png}}
\subfigure[]{\includegraphics[height=4cm, width=0.48\linewidth]{radio.jpg}}
\caption{(a) Crazyflie $2.1$ (b) Crazyradio PA}
\label{fig:devices}
\end{figure}
\subsection{Experimental Results}
In this section, we provide a variety of situations considered while experimentally evaluating our proposed algorithms. We assume that all crazyflies are flying at the altitude of \SI{1}{m}.
\subsubsection{Stationary Non-Concurrent Failures (SNCF)}
For this experiment, we follow the approach presented in Section \ref{subsec:SNCF} where $\Delta_h = $ \SI{0.4}{m} is the chosen value as shown in Eq. \eqref{MainTransformation}. The CFs used in the experiment are uniquely identified using the set $\mathcal{V}=\{1,\cdots,6\}$. As indicated in Table \ref{table:Properties}, we are dealing $m=2$ groups of crazyflies. Therefore, at time $t_0$, $\mathcal{V}$ is divided into $\mathcal{V}_1 = \{1,\cdots,6\}$ and $\mathcal{V}_2 = \emptyset$. Until the first failure, the CFs all move together as shown in Fig. \ref{fig:first_failure} represented as solid lines. At $t_{\mathrm{fail}} =$ \SI{2}{s}, CF$4$, chosen randomly, is subjected to failure and is wrapped by a green cylinder representing the unsafe zone as shown in Fig. \ref{fig:first_failure}. At this instant, $\mathcal{V}_1 = \{1,2,3,5,6\}$ and $\mathcal{V}_2 = \mathcal{V} \setminus \mathcal{V}_1 = \{4\}$. The desired paths for all CFs belonging to $ \mathcal{V}_1$ are computed based on algorithm \ref{alg2}. As shown in Fig. \ref{fig:second_failure}, we observe that healthy crazyflies are safely avoiding the unsafe-zone. We deploy another failure, CF$5$, in the system at $t_{\mathrm{fail}} =$ \SI{12}{s} as shown by the shaded cyan circle in Fig. \ref{fig:second_failure}. The sets $\mathcal{V}_1$ and $\mathcal{V}_2$ are updated: $\mathcal{V}_1 = \{1,2,3,6\}$ and $\mathcal{V}_2 = \{4,5\}$. The desired paths for the healthy CFs are computed until $ \mathcal{V}_1$'s CFs have completely passed the unsafe-zones as shown in Fig. \ref{fig:time_invariant_failures}.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{time_invariant_multiple_failure_t0.png}
\caption{Location of healthy crazyflies at the time of first failure. The green circle corresponds to the unsafe-zone of CF$4$.}
\label{fig:first_failure}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{time_invariant_second_failure.png}
\caption{Location of healthy crazyflies at the time of second failure. The dashed lines represent the desired paths whereas the solid lines represent the actual paths. The cyan circle corresponds to the unsafe-zone of CF$5$.}
\label{fig:second_failure}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{time_invariant_multiple_failure.png}
\caption{Desired vs Actual paths tracked by crazyflies to avoid unsafe-zones.}
\label{fig:time_invariant_failures}
\end{figure}
\subsubsection{Time-Varying Non-Cooperative}
In this experiment, we evaluate the performance of the algorithm \ref{alg3} proposed in Section \ref{subsec:TVNC} using $m=2$ groups of crazyflies. Specifically, at time $t_0$, the set $\mathcal{V}=\{1,\cdots,6\}$ is divided into time-invariant subsets $\mathcal{V}_1=\{1,2,3\}$ and $\mathcal{V}_2=\{4,5,6\}$. The aim of agents in $\mathcal{V}_2$, known as non-cooperative agents, is to reach their goal locations quickly. Therefore, trajectories of $\mathcal{V}_2$'s agents are predefined, as indicated in Fig. \ref{fig:time_varying_six_cf} (See the green, cyan, and black paths). However, the agents belonging to $\mathcal{V}_1$, termed as cooperative agents, use the fluid-flow navigation function to safely plan paths in the shared motion space. As shown in Fig. \ref{fig:time_varying_six_cf}, cooperative CFs $1$, $2$, and $3$ reach their target locations by following the red, blue, and pink paths.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{time_varying_six_cf.png}
\caption{Desired (dashed line) vs actual (solid line) paths undertaken by crazyflies. We can see $\mathcal{V}_1$ taking advantage of the fluid flow recovery algorithm in order to avoid any collision with non-cooperative crazyflies in set $\mathcal{V}_2$}
\label{fig:time_varying_six_cf}
\end{figure}
\subsubsection{Time-Varying Cooperative}
In this experiment, we define the set $\mathcal{V}=\{1,\cdots,6\}$ to uniquely identify all CFs. CFs are divided into two groups identified by $\mathcal{V}_1 = \{1,2,3\}$ and $\mathcal{V}_2 = \{4,5,6\}$. For this experiment, we choose $v_l = $ \SI{0.3}{m/s}, $\delta = 0.15$, and $n_{\tau} = 3$. We have experimentally evaluated the scenario when $\theta$ is varying in time but constant for each individual group of agents. This approach ensures that each agent in a group move parallel to other agents in the group. Figure \ref{fig:game_theory_six_cf_alternate} plots the result of our experiment.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{game_theory_six_cf_alternate.png}
\caption{Desired (dashed line) vs Actual (solid line) paths tracked by crazyflies using Algorithm \ref{alg4} when $N = 6$.}
\label{fig:game_theory_six_cf_alternate}
\end{figure}
\section{CONCLUSION}
\label{sec:conclusion}
In this work, we proposed recovery algorithms based on ideal fluid-flow for collision-free coordination between multiple groups of agents. Our algorithms were able to handle different scenarios including stationary non-concurrent failures, time-varying non-cooperative failures, and time-varying cooperative failures. Experimental results using teams of crazyflies displayed the advantage of our proposed algorithms in handling different situations. Future work on this direction include incorporating reinforcement learning techniques.
\section{Acknowledgement}
The authors gratefully thank Mr. Mohammad Ghufran for his help with performing the flight experiments.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
Recently, self-supervised learning \cite{Chen2020, He2020} has shown significant results in feature learning. Self-supervised learning is a promising approach to fixing the fundamental problem in training machine learning models where labeled data is scarce or expensive. For self-supervised learning, one of the most important techniques is data augmentation. In the case of image classification, self-supervised learning involves giving the model a set of unlabeled images and asking it to predict some properties of the images \cite{moon2022tailoring, Gidaris2018}. By predicting the images, the model can learn valuable features for classification.
Using data augmentation techniques in image recognition enhances the model's ability to generalize by teaching insensitive features to spatial changes. This can be achieved by applying various modifications, such as geometric transformations (cropping, flipping, rotating) and photometric adjustments (brightness, contrast, and color). Recent methods have demonstrated the best performance, balancing complexity with accuracy and robustness. In addition, various methods for feature learning have specific to certain locations, which can also be used to transfer knowledge to tasks related to localization, like object detection or image labeling. These features allow the model to understand what and where to focus on making correct predictions.
The rotation has been attempted to be used for these augmentations, but relying solely on rotation as a self-supervised transformation can limit the ability of the model to learn rich features from the data. In this paper, we propose to use additional localizable augmentations, such as flip and shuffle channels, to provide a more diverse set of self-supervised signals. These augmentations allow the model to learn more complex features relevant to classification.
To ensure that the model focuses on the most relevant augmentations for classification, we use a Mixture of Expert (MoE), which gates the effects of each augmentation on the loss. This allows the model to dynamically adjust the importance of each augmentation, allowing it to focus on the most useful transformations for classification.
By utilizing these additional augmentations and the MoE method, our proposed approach can improve performance on image classification benchmarks, allowing the model to learn more complex features relevant to classification. Furthermore, our approach can improve the performance of other computer vision tasks, such as object detection and image annotation, as the learned features are transferable and localizable. Additionally, our approach can be used to improve the performance of models trained with scarce or expensive labeled data. Overall, our proposed method of using additional localizable augmentations and the MoE method to adjust the weighting of their importance can effectively improve self-supervised representation learning and computer vision tasks.
\section{Related Work}
\label{sec:format}
\textbf{Self-Supervised Learning:} Self-supervised learning has gained significant interest in recent years. It aims to learn general characteristics by solving tasks explicitly created for this purpose, known as pretext tasks. Depending on the number of examples used for these tasks, self-supervised learning can be divided into relation-based and transformation-based. Relation-based approaches focus on increasing the similarity between a sample and its transformed positive counterparts, and some also treat other samples as negative examples. Notable techniques in this category include memory bank \cite{chen2021exploring} and in-batch \cite{ye2019unsupervised} sampling. On the other hand, some methods use positive pairs with Siamese networks \cite{grill2020bootstrap} or add a relationship module \cite{patacchiola2020self} to perform the self-supervised task.
Transformation-based self-supervision is another popular approach, which involves generating new classes with data augmentation, predicting relative positions of patches, solving puzzles, or predicting rotations. One unique method in this category is LoRot \cite{moon2022tailoring}, designed for a different objective: to aid supervised learning.
Recently, there have been attempts to transfer the benefits of self-supervised learning to supervised learning. SupCLR \cite{khosla2020supervised} adapted the relation-based self-supervised framework to utilize labeled data since class labels clearly define both positive and negative examples. Additionally, self-label augmentation (SLA) \cite{gao2022decoupled} expanded the label space by combining the supervised class labels with data transformation labels, as using auxiliary pretext tasks can decrease performance. On the other hand, LoRot is a self-supervised method that can be directly applied to improve supervised learning.
\textbf{Mixture of Experts:} Sparsely-gated MoE \cite{shazeer2017} is the first model to show major improvements in model capacity, training time, or model quality upon activation. Switch Transformer \cite{fedus2021switch} simplifies activation by selecting only the best expert for each token using softmax in the hidden state and exhibits better scalability than previous works. All previous work required ancillary losses to encourage balance. This provision for loss must be carefully considered not to overwhelm the original loss. However, the auxiliary loss does not guarantee balance, and a hard power factor must be applied. Therefore, many tokens may not be affected by the MoE layer. Hard MoE \cite{gross2017hard} with a single decoding layer can be trained effectively in large-scale hashtag prediction tasks. Base Layers \cite{pmlr-v139-lewis21a} construct a linear assignment to maximize the token-expert relationship while ensuring that each expert receives an equal number of tokens. Hash classes design hashing techniques on input tokens. Unlike previous works, our method is a learned one that enables heterogeneous MoE and effectively improves downstream fine-tuning performance. We use MoE as a way to control the effect of each self-supervised transformation on the loss function. Thus, we can ensure that the model focuses on the most relevant augmentations for classification. Our method is similar to \cite{yudistira2017gated} where the gates are attached to the last classification layer to fuse the two-stream CNN backbone on video classification. This approach differs from previous self-supervised methods, which typically rely on a single task or set of tasks and do not have a way to adjust the importance of different augmentations.
\section{Methodology}
\label{sec:pagestyle}
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{Image/Diagram_Transformasi.jpg}
\caption{Diagram of Transformation}
\label{fig:diagram_transformation}
\end{figure}
We first discussed each transformation we used as data augmentation in Sec. \ref{ssec:transformation}. Then, we used LDAM-DRW to train the imbalanced dataset in Sec. \ref{ssec:ldam-drw}. Finally, we proposed Gated Self-Supervised Learning based on a Mixture of Expert approaches, allowing the model to focus on relevant transformations.
\subsection{Transformation}
\label{ssec:transformation}
The first instances of the usefulness of data augmentation were shown using simple modifications like flipping images horizontally, changing the color space, and applying random croppings (Figure \ref{fig:diagram_transformation}). These techniques address issues related to image recognition tasks, such as insensitivity to spatial changes. This section will cover various augmentations based on geometric transformations and other image-processing methods. The augmentations discussed are notable for their simplicity of implementation. Understanding these basic transformations will serve as a foundation for exploring more advanced data augmentation techniques.
Our proposed method uses a combination of data augmentation, LDAM-DRW, and a Mixture of Expert (MoE) approaches to improve image classification performance on imbalanced datasets.
We utilized the different transformations as data augmentations, including rotation (using Lorot-E \cite{moon2022tailoring}), flip, and shuffle channels. These augmentations are chosen as they are easy to implement yet are effective in encoding invariances and challenges present in image recognition tasks. Then, LDAM-DRW is used to train imbalanced datasets, which improves the performance and accuracy of the model by balancing the distribution of the classes. Finally, the MoE is attached to gating every self-supervised task. This allows the model to learn the importance of each transformation and dynamically adjust the weighting of each one for classification. Additionally, the gating network is built with a fully-connected layer with softmax activation to output a class for each transformation.
\subsubsection{Rotation}
\label{sssec:rotation}
We use Lorot-E\cite{moon2022tailoring} as the transformation for the rotation transformation. The image will be divided into four quadrants (2x2 grid), then randomly selected and rotated from {0, 90, 180, 270}. This transformation will create 16 classes. The rotation degree parameter heavily determines the degree of rotation augmentations.
\subsubsection{Flip}
\label{sssec:flip}
Flipping is one of the easiest to implement and has proven useful on datasets. This task will randomly flip the selected quadrant of the image along the x-axis. The number of classes resulting from this transformation is two classes.
\subsubsection{Shuffle Channel}
\label{sssec:sc}
This transformation is used to shuffle the arrangement of the RGB channel of the selected quadrant of the image. The number of classes resulting from this transformation is six classes (3P3 permutation).
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{Image/Diagram_paper.jpg}
\caption{Diagram of Mixture of Expert}
\label{fig:diagram_paper}
\end{figure}
\subsection{LDAM-DRW}
\label{ssec:ldam-drw}
\begin{table*}\centering
\caption{Imbalanced classification accuracy (\%) in CIFAR-10/100}
\begin{tabular}{l|lll|lll}
\hline
\label{tab:results}
Imbalance Ratio & 0.01 & 0.02 & 0.05 & 0.01 & 0.02 & 0.05 \\ \hline
LDAM-DRW & 77.03 & 80.94 & 85.46 & 42.04 & 46.15 & 53.25 \\
+SSP & 77.83 & 82.13 & - & 43.43 & 47.11 & - \\
+SLA-SD & 80.24 & - & - & 45.53 & - & - \\
+LoRot-E & 81.82 & 84.41 & 86.67 & 46.48 & 50.05 & 54.66 \\ \hline
+MoE(LoRot-E+Flip) & 81.65 & 83.93 & 86.64 & 46.48 & 49.96 & 54.63 \\
+MoE(LoRot-E+ShuffleChannel) & \textbf{81.91} & 84.33 & \textbf{86.91} & \textbf{46.53} & 49.95 & \textbf{54.85} \\
+MoE(LoRot-E+Flip+ShuffleChannel) & 81.67 & \textbf{84.65} & 86.35 & \textbf{47.37} & \textbf{50.57} & \textbf{54.75} \\ \hline
\end{tabular}
\end{table*}
LDAM-DRW (Label-Distribution-Aware Margin - Deferred Re-Weighting) is a combination of two techniques designed to improve the performance of machine learning models in situations where the training data is heavily imbalanced among different classes, and the evaluation criteria require good generalization to the less common classes \cite{Cao2019}. In this experiment, we use the LDAM-DRW method with Gated Self-Supervision Method, which we propose to improve the performance and accuracy of the model.
\subsection{Mixture of Experts}
Mixture-of-Experts (MoE) is a type of deep learning architecture that combines multiple models, referred to as experts, to divide a complex task into simpler sub-problems that can each be addressed by an individual expert \cite{Chen2022}. In this work, we use MoE to gating every self-supervision task, which is used to learn the importance of each transformation used in self-supervised learning as illustrated in Figure \ref{fig:diagram_paper}. Each transformation has its linear head to output the class of the transformation. The weight gate of each transformation is also learned using an MoE, which allows the model to adjust the importance of each transformation for classification dynamically. We use a fully-connected layer with softmax activation function for the gating network as in eq. \ref{eq:softmax}:
\begin{equation}
\label{eq:softmax}
G = softmax(W^{T}X + b)
\end{equation}
Where $G$, $X$, $W$, and $B$ refer to the gate, baseline output, weight gating network, and bias gating network, respectively. Specifically, we gates every loss from each self-supervision task $L$ and sum all the gating loss. Then, we combine the loss of the classifier (supervised) $L_{C}$ and the loss of the self-supervision as follows in eq. \ref{eq:gating}:
\begin{equation}
\label{eq:gating}
L_{tot} = L_{C}+ \lambda \sum_{n=1}^{t}G_{n}^{T}L_{n}
\end{equation}
where $t$ is the number of the self-supervision task and $\lambda$ is SSL ratio.
\section{Experimental Setup}
\label{sec:experimentalsetup}
In the imbalanced task experiment, we tested the gated self-supervised learning method using the Google Colab Pro environment. The GPU that we use for this experiment is Nvidia T4 GPU. For a fair comparison, we use the same backbone and baseline as previous research to train the model for the experiment. We use the Resnet-32 architecture for the backbone of the network and LDAM-DRW \cite{Cao2019} as the baseline and follows the baseline settings. We set the batch size to 128 and the epochs for training the model to 300. For the learning rate, we set the initial value to 0.1, which is dropped by a factor of 0.01 at the 160-$th$ epoch and 180-$th$ epoch. The optimizer we use in this experiment is Stochastic Gradient Descent with the momentum of 0,9 and weight decay $2\times10^{-4}$. For the SSL ratio, we set all of the experiments to 0.1.
We also train the model using our proposed method in the Tiny-Imagenet dataset \cite{le2015tiny} using a single GPU Nvidia Quadro RTX 8000. For a fair comparison, we use the same backbone and setups in each method that will be tested. To train the model, we set the batch size to 256 and the epochs for training the model to 300. Resnet18 will be used as the backbone, with Stochastic Gradient Descent as the optimizer. We set the momentum to 0,9 and weight decay to $2\times10^{-4}$. For the learning rate, we set the initial value to 0.1, decaying by a factor of 0.1 every 75 epochs. Finally, for the SSL ratio, we set all of the experiments to 0.1.
\begin{table}[]
\caption{Additional Experiments on Tiny-Imagenet}
\begin{tabular}{l|c}
\hline
\label{tab:results_tiny-imagenet}
& Val accuracy \\ \hline
ResNet 18 & 46.68 \\
+LoRot-E & 48.65 \\ \hline
+LoRot-E+Flip+ShuffleChannel & 48.86 \\ \hline
+Moe(LoRot-E+Flip) & 47.86 \\
+MoE(LoRot-E+ShuffleChannel) & 48.52 \\
+MoE(LoRot-E+Flip+ShuffleChannel) & \textbf{48.99} \\ \hline
\end{tabular}
\end{table}
\section{Experiments}
\label{sec:experiments}
For the experiment, we use LDAM-DRW \cite{Cao2019} to do the Imbalanced Classification and Cifar dataset to design the imbalanced scenario. We create three combinations of gated self-supervised learning for the proposed method, such as Lorot-E+flip, Lorot-E+ShuffleChannel, and Lorot-E+FLip+ShuffleChannel. Moreover, we also vary the imbalance scenario of Cifar 10 and Cifar 100 datasets. We apply the imbalance ratio to 0.01, 0.02, and 0.05. To compare our proposed method with other self-supervision techniques, we report the result of LDAM-DRW\cite{Cao2019}, SSP\cite{yang2020rethinking}, SLA+SD \cite{lee2020self} and, Lorot-E\cite{moon2022tailoring}.
The results of the imbalanced classification are shown in Table \ref{tab:results}. As we can see, all of the combination tasks for MoE have a clear complementary effect and improve the accuracy of the LDAM-DRW, SSP, and SLA+SD method, with a gain of up to +4.87\%. However, several combinations of MoE improve the Lorot-E method, such as Lorot-E + flip and Lorot-E + flip + shuffle channel. However, the MoE combination of Lorot+Flip can not improve the Lorot-E model in two imbalanced datasets, such as Cifar 10 and Cifar 100, thus reducing the model's accuracy. Meanwhile, the other combinations of Lorot-E + ShuffleChannel, and Lorot-E+flip+ShuffleChannel improve the accuracy of the Lorot-E model in the imbalanced Cifar 10 ad Cifar 100. Specifically, the Lorot-E+flip+Shuffle-Channel combination successfully improves the model's accuracy in all imbalanced scenarios in Cifar 100 dataset. Therefore, Lorot-E+flip+ShuffleChannel is the best combination to improve the model's accuracy in the imbalanced task.
In order to further evaluate the effectiveness of our proposed method, we tested our method on the Tiny-Imagenet dataset. We found that it consistently outperformed other methods, as shown in Table \ref{tab:results_tiny-imagenet}. This was particularly impressive given the large number of classes in the Tiny-Imagenet dataset, as our method improved even on this challenging dataset. Overall, our results demonstrate the versatility and effectiveness of our method, as it was able to achieve better performance on all datasets tested, so our method has the potential to be widely applicable and useful for a variety of image classification tasks.
This work considers each of these transformations as a non-linear one. Therefore, the gated self-supervised learning method can tackle the problem of determining which transformation is important for the model to learn. However, to make further improvements, we need to find and select the combination self-supervision task of MoE that will best improve the model's accuracy.
\section{Conclusion}
\label{sec:majhead}
Self-supervised learning has gained increasing attention in recent years as a way to train deep learning models using large amounts of unlabeled data. The main idea behind self-supervised learning is to use the inherent structure in the data to create a supervised learning problem that can be used to train a model. In this work, we proposed a self-supervised learning method that adds additional transformations, such as flip and shuffle channels, to the past method, rotation transformation. Our results show that adding these additional transformations helps to increase the accuracy of the model, especially when using larger datasets such as CIFAR 100. Furthermore, compared to other methods, our proposed method gave better accuracy in all experiments when using CIFAR 100. On CIFAR 10, our method also gave better accuracy in certain experiments. Finally, on Tiny-Imagenet, we achieve better results using a self-supervised learning method on top of the model.
Overall, our results suggest that adding more transformations to self-supervised learning, in combination with the gating method, can help improve the model's performance. This highlights the potential of self-supervised learning as a powerful tool for training deep learning models in various settings.
\vfill\pagebreak
\bibliographystyle{IEEEbib}
\section{Introduction}
\label{sec:intro}
Recently, self-supervised learning \cite{Chen2020, He2020} has shown significant results in feature learning. Self-supervised learning is a promising approach to fixing the fundamental problem in training machine learning models where labeled data is scarce or expensive. For self-supervised learning, one of the most important techniques is data augmentation. In the case of image classification, self-supervised learning involves giving the model a set of unlabeled images and asking it to predict some properties of the images \cite{moon2022tailoring, Gidaris2018}. By predicting the images, the model can learn valuable features for classification.
Using data augmentation techniques in image recognition enhances the model's ability to generalize by teaching insensitive features to spatial changes. This can be achieved by applying various modifications, such as geometric transformations (cropping, flipping, rotating) and photometric adjustments (brightness, contrast, and color). Recent methods have demonstrated the best performance, balancing complexity with accuracy and robustness. In addition, various methods for feature learning have specific to certain locations, which can also be used to transfer knowledge to tasks related to localization, like object detection or image labeling. These features allow the model to understand what and where to focus on making correct predictions.
The rotation has been attempted to be used for these augmentations, but relying solely on rotation as a self-supervised transformation can limit the ability of the model to learn rich features from the data. In this paper, we propose to use additional localizable augmentations, such as flip and shuffle channels, to provide a more diverse set of self-supervised signals. These augmentations allow the model to learn more complex features relevant to classification.
To ensure that the model focuses on the most relevant augmentations for classification, we use a Mixture of Expert (MoE), which gates the effects of each augmentation on the loss. This allows the model to dynamically adjust the importance of each augmentation, allowing it to focus on the most useful transformations for classification.
By utilizing these additional augmentations and the MoE method, our proposed approach can improve performance on image classification benchmarks, allowing the model to learn more complex features relevant to classification. Furthermore, our approach can improve the performance of other computer vision tasks, such as object detection and image annotation, as the learned features are transferable and localizable. Additionally, our approach can be used to improve the performance of models trained with scarce or expensive labeled data. Overall, our proposed method of using additional localizable augmentations and the MoE method to adjust the weighting of their importance can effectively improve self-supervised representation learning and computer vision tasks.
\section{Related Work}
\label{sec:format}
\textbf{Self-Supervised Learning:} Self-supervised learning has gained significant interest in recent years. It aims to learn general characteristics by solving tasks explicitly created for this purpose, known as pretext tasks. Depending on the number of examples used for these tasks, self-supervised learning can be divided into relation-based and transformation-based. Relation-based approaches focus on increasing the similarity between a sample and its transformed positive counterparts, and some also treat other samples as negative examples. Notable techniques in this category include memory bank \cite{chen2021exploring} and in-batch \cite{ye2019unsupervised} sampling. On the other hand, some methods use positive pairs with Siamese networks \cite{grill2020bootstrap} or add a relationship module \cite{patacchiola2020self} to perform the self-supervised task.
Transformation-based self-supervision is another popular approach, which involves generating new classes with data augmentation, predicting relative positions of patches, solving puzzles, or predicting rotations. One unique method in this category is LoRot \cite{moon2022tailoring}, designed for a different objective: to aid supervised learning.
Recently, there have been attempts to transfer the benefits of self-supervised learning to supervised learning. SupCLR \cite{khosla2020supervised} adapted the relation-based self-supervised framework to utilize labeled data since class labels clearly define both positive and negative examples. Additionally, self-label augmentation (SLA) \cite{gao2022decoupled} expanded the label space by combining the supervised class labels with data transformation labels, as using auxiliary pretext tasks can decrease performance. On the other hand, LoRot is a self-supervised method that can be directly applied to improve supervised learning.
\textbf{Mixture of Experts:} Sparsely-gated MoE \cite{shazeer2017} is the first model to show major improvements in model capacity, training time, or model quality upon activation. Switch Transformer \cite{fedus2021switch} simplifies activation by selecting only the best expert for each token using softmax in the hidden state and exhibits better scalability than previous works. All previous work required ancillary losses to encourage balance. This provision for loss must be carefully considered not to overwhelm the original loss. However, the auxiliary loss does not guarantee balance, and a hard power factor must be applied. Therefore, many tokens may not be affected by the MoE layer. Hard MoE \cite{gross2017hard} with a single decoding layer can be trained effectively in large-scale hashtag prediction tasks. Base Layers \cite{pmlr-v139-lewis21a} construct a linear assignment to maximize the token-expert relationship while ensuring that each expert receives an equal number of tokens. Hash classes design hashing techniques on input tokens. Unlike previous works, our method is a learned one that enables heterogeneous MoE and effectively improves downstream fine-tuning performance. We use MoE as a way to control the effect of each self-supervised transformation on the loss function. Thus, we can ensure that the model focuses on the most relevant augmentations for classification. Our method is similar to \cite{yudistira2017gated} where the gates are attached to the last classification layer to fuse the two-stream CNN backbone on video classification. This approach differs from previous self-supervised methods, which typically rely on a single task or set of tasks and do not have a way to adjust the importance of different augmentations.
\section{Methodology}
\label{sec:pagestyle}
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{Image/Diagram_Transformasi.jpg}
\caption{Diagram of Transformation}
\label{fig:diagram_transformation}
\end{figure}
We first discussed each transformation we used as data augmentation in Sec. \ref{ssec:transformation}. Then, we used LDAM-DRW to train the imbalanced dataset in Sec. \ref{ssec:ldam-drw}. Finally, we proposed Gated Self-Supervised Learning based on a Mixture of Expert approaches, allowing the model to focus on relevant transformations.
\subsection{Transformation}
\label{ssec:transformation}
The first instances of the usefulness of data augmentation were shown using simple modifications like flipping images horizontally, changing the color space, and applying random croppings (Figure \ref{fig:diagram_transformation}). These techniques address issues related to image recognition tasks, such as insensitivity to spatial changes. This section will cover various augmentations based on geometric transformations and other image-processing methods. The augmentations discussed are notable for their simplicity of implementation. Understanding these basic transformations will serve as a foundation for exploring more advanced data augmentation techniques.
Our proposed method uses a combination of data augmentation, LDAM-DRW, and a Mixture of Expert (MoE) approaches to improve image classification performance on imbalanced datasets.
We utilized the different transformations as data augmentations, including rotation (using Lorot-E \cite{moon2022tailoring}), flip, and shuffle channels. These augmentations are chosen as they are easy to implement yet are effective in encoding invariances and challenges present in image recognition tasks. Then, LDAM-DRW is used to train imbalanced datasets, which improves the performance and accuracy of the model by balancing the distribution of the classes. Finally, the MoE is attached to gating every self-supervised task. This allows the model to learn the importance of each transformation and dynamically adjust the weighting of each one for classification. Additionally, the gating network is built with a fully-connected layer with softmax activation to output a class for each transformation.
\subsubsection{Rotation}
\label{sssec:rotation}
We use Lorot-E\cite{moon2022tailoring} as the transformation for the rotation transformation. The image will be divided into four quadrants (2x2 grid), then randomly selected and rotated from {0, 90, 180, 270}. This transformation will create 16 classes. The rotation degree parameter heavily determines the degree of rotation augmentations.
\subsubsection{Flip}
\label{sssec:flip}
Flipping is one of the easiest to implement and has proven useful on datasets. This task will randomly flip the selected quadrant of the image along the x-axis. The number of classes resulting from this transformation is two classes.
\subsubsection{Shuffle Channel}
\label{sssec:sc}
This transformation is used to shuffle the arrangement of the RGB channel of the selected quadrant of the image. The number of classes resulting from this transformation is six classes (3P3 permutation).
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{Image/Diagram_paper.jpg}
\caption{Diagram of Mixture of Expert}
\label{fig:diagram_paper}
\end{figure}
\subsection{LDAM-DRW}
\label{ssec:ldam-drw}
\begin{table*}\centering
\caption{Imbalanced classification accuracy (\%) in CIFAR-10/100}
\begin{tabular}{l|lll|lll}
\hline
\label{tab:results}
Imbalance Ratio & 0.01 & 0.02 & 0.05 & 0.01 & 0.02 & 0.05 \\ \hline
LDAM-DRW & 77.03 & 80.94 & 85.46 & 42.04 & 46.15 & 53.25 \\
+SSP & 77.83 & 82.13 & - & 43.43 & 47.11 & - \\
+SLA-SD & 80.24 & - & - & 45.53 & - & - \\
+LoRot-E & 81.82 & 84.41 & 86.67 & 46.48 & 50.05 & 54.66 \\ \hline
+MoE(LoRot-E+Flip) & 81.65 & 83.93 & 86.64 & 46.48 & 49.96 & 54.63 \\
+MoE(LoRot-E+ShuffleChannel) & \textbf{81.91} & 84.33 & \textbf{86.91} & \textbf{46.53} & 49.95 & \textbf{54.85} \\
+MoE(LoRot-E+Flip+ShuffleChannel) & 81.67 & \textbf{84.65} & 86.35 & \textbf{47.37} & \textbf{50.57} & \textbf{54.75} \\ \hline
\end{tabular}
\end{table*}
LDAM-DRW (Label-Distribution-Aware Margin - Deferred Re-Weighting) is a combination of two techniques designed to improve the performance of machine learning models in situations where the training data is heavily imbalanced among different classes, and the evaluation criteria require good generalization to the less common classes \cite{Cao2019}. In this experiment, we use the LDAM-DRW method with Gated Self-Supervision Method, which we propose to improve the performance and accuracy of the model.
\subsection{Mixture of Experts}
Mixture-of-Experts (MoE) is a type of deep learning architecture that combines multiple models, referred to as experts, to divide a complex task into simpler sub-problems that can each be addressed by an individual expert \cite{Chen2022}. In this work, we use MoE to gating every self-supervision task, which is used to learn the importance of each transformation used in self-supervised learning as illustrated in Figure \ref{fig:diagram_paper}. Each transformation has its linear head to output the class of the transformation. The weight gate of each transformation is also learned using an MoE, which allows the model to adjust the importance of each transformation for classification dynamically. We use a fully-connected layer with softmax activation function for the gating network as in eq. \ref{eq:softmax}:
\begin{equation}
\label{eq:softmax}
G = softmax(W^{T}X + b)
\end{equation}
Where $G$, $X$, $W$, and $B$ refer to the gate, baseline output, weight gating network, and bias gating network, respectively. Specifically, we gates every loss from each self-supervision task $L$ and sum all the gating loss. Then, we combine the loss of the classifier (supervised) $L_{C}$ and the loss of the self-supervision as follows in eq. \ref{eq:gating}:
\begin{equation}
\label{eq:gating}
L_{tot} = L_{C}+ \lambda \sum_{n=1}^{t}G_{n}^{T}L_{n}
\end{equation}
where $t$ is the number of the self-supervision task and $\lambda$ is SSL ratio.
\section{Experimental Setup}
\label{sec:experimentalsetup}
In the imbalanced task experiment, we tested the gated self-supervised learning method using the Google Colab Pro environment. The GPU that we use for this experiment is Nvidia T4 GPU. For a fair comparison, we use the same backbone and baseline as previous research to train the model for the experiment. We use the Resnet-32 architecture for the backbone of the network and LDAM-DRW \cite{Cao2019} as the baseline and follows the baseline settings. We set the batch size to 128 and the epochs for training the model to 300. For the learning rate, we set the initial value to 0.1, which is dropped by a factor of 0.01 at the 160-$th$ epoch and 180-$th$ epoch. The optimizer we use in this experiment is Stochastic Gradient Descent with the momentum of 0,9 and weight decay $2\times10^{-4}$. For the SSL ratio, we set all of the experiments to 0.1.
We also train the model using our proposed method in the Tiny-Imagenet dataset \cite{le2015tiny} using a single GPU Nvidia Quadro RTX 8000. For a fair comparison, we use the same backbone and setups in each method that will be tested. To train the model, we set the batch size to 256 and the epochs for training the model to 300. Resnet18 will be used as the backbone, with Stochastic Gradient Descent as the optimizer. We set the momentum to 0,9 and weight decay to $2\times10^{-4}$. For the learning rate, we set the initial value to 0.1, decaying by a factor of 0.1 every 75 epochs. Finally, for the SSL ratio, we set all of the experiments to 0.1.
\begin{table}[]
\caption{Additional Experiments on Tiny-Imagenet}
\begin{tabular}{l|c}
\hline
\label{tab:results_tiny-imagenet}
& Val accuracy \\ \hline
ResNet 18 & 46.68 \\
+LoRot-E & 48.65 \\ \hline
+LoRot-E+Flip+ShuffleChannel & 48.86 \\ \hline
+Moe(LoRot-E+Flip) & 47.86 \\
+MoE(LoRot-E+ShuffleChannel) & 48.52 \\
+MoE(LoRot-E+Flip+ShuffleChannel) & \textbf{48.99} \\ \hline
\end{tabular}
\end{table}
\section{Experiments}
\label{sec:experiments}
For the experiment, we use LDAM-DRW \cite{Cao2019} to do the Imbalanced Classification and Cifar dataset to design the imbalanced scenario. We create three combinations of gated self-supervised learning for the proposed method, such as Lorot-E+flip, Lorot-E+ShuffleChannel, and Lorot-E+FLip+ShuffleChannel. Moreover, we also vary the imbalance scenario of Cifar 10 and Cifar 100 datasets. We apply the imbalance ratio to 0.01, 0.02, and 0.05. To compare our proposed method with other self-supervision techniques, we report the result of LDAM-DRW\cite{Cao2019}, SSP\cite{yang2020rethinking}, SLA+SD \cite{lee2020self} and, Lorot-E\cite{moon2022tailoring}.
The results of the imbalanced classification are shown in Table \ref{tab:results}. As we can see, all of the combination tasks for MoE have a clear complementary effect and improve the accuracy of the LDAM-DRW, SSP, and SLA+SD method, with a gain of up to +4.87\%. However, several combinations of MoE improve the Lorot-E method, such as Lorot-E + flip and Lorot-E + flip + shuffle channel. However, the MoE combination of Lorot+Flip can not improve the Lorot-E model in two imbalanced datasets, such as Cifar 10 and Cifar 100, thus reducing the model's accuracy. Meanwhile, the other combinations of Lorot-E + ShuffleChannel, and Lorot-E+flip+ShuffleChannel improve the accuracy of the Lorot-E model in the imbalanced Cifar 10 ad Cifar 100. Specifically, the Lorot-E+flip+Shuffle-Channel combination successfully improves the model's accuracy in all imbalanced scenarios in Cifar 100 dataset. Therefore, Lorot-E+flip+ShuffleChannel is the best combination to improve the model's accuracy in the imbalanced task.
In order to further evaluate the effectiveness of our proposed method, we tested our method on the Tiny-Imagenet dataset. We found that it consistently outperformed other methods, as shown in Table \ref{tab:results_tiny-imagenet}. This was particularly impressive given the large number of classes in the Tiny-Imagenet dataset, as our method improved even on this challenging dataset. Overall, our results demonstrate the versatility and effectiveness of our method, as it was able to achieve better performance on all datasets tested, so our method has the potential to be widely applicable and useful for a variety of image classification tasks.
This work considers each of these transformations as a non-linear one. Therefore, the gated self-supervised learning method can tackle the problem of determining which transformation is important for the model to learn. However, to make further improvements, we need to find and select the combination self-supervision task of MoE that will best improve the model's accuracy.
\section{Conclusion}
\label{sec:majhead}
Self-supervised learning has gained increasing attention in recent years as a way to train deep learning models using large amounts of unlabeled data. The main idea behind self-supervised learning is to use the inherent structure in the data to create a supervised learning problem that can be used to train a model. In this work, we proposed a self-supervised learning method that adds additional transformations, such as flip and shuffle channels, to the past method, rotation transformation. Our results show that adding these additional transformations helps to increase the accuracy of the model, especially when using larger datasets such as CIFAR 100. Furthermore, compared to other methods, our proposed method gave better accuracy in all experiments when using CIFAR 100. On CIFAR 10, our method also gave better accuracy in certain experiments. Finally, on Tiny-Imagenet, we achieve better results using a self-supervised learning method on top of the model.
Overall, our results suggest that adding more transformations to self-supervised learning, in combination with the gating method, can help improve the model's performance. This highlights the potential of self-supervised learning as a powerful tool for training deep learning models in various settings.
\vfill\pagebreak
\bibliographystyle{IEEEbib}
|
\section{Introduction }
As is well known, a non-negative function $\omega$ on $\mathbb{R}^n$ is an $A_p$ weight with $p>1,$ if there exists a constant $C$ for all cubes $Q$ such that
$$
\bigg( \frac{1}{|Q|} \int_Q \omega \, d\mu \bigg) \bigg( \frac{1}{|Q|} \int_Q \omega^{-\frac{1}{p-1}} \, d\mu \bigg)^{p-1} \leq C.
$$
Muckenhoupt \cite{MR293384} observed that the weight has an open property $A_p=\cup_{1<q<p}A_q,$ and shortly after Muckenhoupt \cite{MR350297} defined the $A^M_\infty$ weight as follows: there exist $0<\varepsilon,~\delta<1$ such that for all $E\subseteq Q$ it holds that
$$
|E| < \delta |Q| \Rightarrow \omega(E) < \varepsilon \omega(Q).
$$
Then he showed $A^M_{\infty}=\cup_{p>1}A_p.$ Independently,
Coifman and Fefferman \cite{MR358205} introduced an $A^{CF}_\infty$ weight and proved $A^{CF}_{\infty}=\cup_{p>1}A_p,$ where the $A^{CF}_\infty$ weight $\omega$ is defined as follows: there exist $C,~\delta>0$ such that for all $E\subseteq Q$
$$
\frac{\omega(E)}{\omega(Q)} \leq C \bigg( \frac{|E|}{|Q|} \bigg)^\delta.
$$
Later, a condition $A_{\infty}^{exp}$ defined by a limit of the $A_p$ weight as $p\uparrow\infty$ was studied almost simultaneously in \cite{MR727244} and \cite[p.405]{MR807149} and $\cup_{p>1}A_p=A^{exp}_{\infty}.$
As we have seen, $A^{M}_\infty,$ $A^{CF}_\infty$ and $A_{\infty}^{exp}$ are equivalent. These are geometric characterizations of $\cup_{p>1}A_p$ and systematically studied in Grafakos \cite[Theorem 7.3.3]{MR3243734}. Very recently, Duoandikoetxea, Mart\'{\i}n-Reyes and Ombrosi \cite{MR3473651} compared and discussed different characterizations of $\cup_{p>1}A_p$ in the setting of general bases. Indeed, they studied many other characterizations which are not geometric. Here we list four characterizations for cubes mentioned in \cite{MR3473651}:
\begin{enumerate}
\item [($A^{*}_{\infty}$)]There exists $C > 0$ such that
$$\int_Q M (\omega \chi_Q) dx \leq C \omega(Q),$$
where $\omega(Q) := \int_Q \omega d\mu$ and $M(\cdot)$ is the Hardy-Littlewood maximal operator(see \cite{MR481968}, \cite{MR883661}). Hyt\"{o}nen and P\'{e}rez \cite{MR3092729} used the weight $A^{*}_{\infty}$ to improve estimates of the bounds in the weighted inequalities.
\item [($A^{log}_{\infty}$)] There exists $C > 0$ such that
$$\int_Q \omega \log^+ \frac{\omega}{\omega_Q}dx \leq C \omega(Q),$$
where $\omega_Q:=\omega(Q)/|Q|$(see \cite{MR481968}).
\item [($A^{med}_{\infty}$)] There exists $C > 0$ such that
$$\omega_Q \leq C m(\omega; Q)$$
where the median of $\omega$ in $Q$ is a number $m(\omega;Q)$ such that $|\{x \in Q : \omega(x) < m(\omega;Q)\}| \leq |Q| / 2$ and $|\{x \in Q : \omega(x) > m(\omega;Q)\}| \leq |Q| / 2$ (see \cite{MR529683}). Using the median, Lerner \cite{MR2721744}
obtain a decomposition of an arbitrary measurable function in terms of local mean oscillations.
\item [($A^{\lambda}_{\infty}$)] There exist $C, \beta > 0$ such that
$$w\big(\{ x \in Q : \omega(x) > \lambda \}\big) \leq C \lambda \, \big|\{x \in Q : \omega(x) > \beta \lambda \}\big|,$$
where $\lambda>\omega_Q$. This kind of characterization appeared independently in \cite{MR402038} and \cite{MR358205}.
\end{enumerate}
Although $A^{*}_{\infty},$ $A^{log}_{\infty},$ $A^{med}_{\infty}$ and $A^{\lambda}_{\infty}$ are not geometric, they are equivalent to $A^{M}_\infty,$ $A^{CF}_\infty$ and $A_{\infty}^{exp}$ for cubes. In
the context of general bases,
the relations between them are more complicated(see \cite{MR3473651} and \cite{MR4446233} for more information).
In this paper, we study $A_\infty$ weights in martingale spaces. Izumisawa and Kazamaki first introduced the
$A_p$ weight for martingales. Under some additional
conditions, they obtained the open property $A_p=\cup_{1<q<p}A_q.$ In martingale setting, it is well known that this property is false in general, because Bonami and L\'{e}pingle \cite{MR544802} showed that for any $p>1,$ there exists a weight $\omega\in A_p,$ but $\omega\notin A_{p-\varepsilon}$ for all $\varepsilon>0$(see also \cite[p.241]{MR1224450}).
Under some additional restrictions, the $A_p$ weight was extensively studied by Dol\'{e}ans-Dade and Meyer \cite{MR544804}. Motivated by the work
of \cite[Theorem 7.3.3]{MR3243734}, \cite{MR3473651} and \cite{MR4446233}, we study
several characterizations of $A_{\infty}$ weights in the setting of martingales. Our first result is the following Theorem \ref{Thm:equa}, which partially depends on a regularity condition $S$(see Definition \ref{regular_p}).
\begin{theorem}\label{Thm:equa}Let $\omega$ be a weight. If $\omega\in S,$ then the following are
equivalent.
\begin{enumerate}[\rm (1)]
\item \label{Thm:equa_Ap}There exist
$C,~p>1$ such that for all $n\in \mathbb{N}$ we have
\begin{equation}\label{Ap}
\mathbb{E}(\omega|\mathcal {F}_n)\mathbb{E}(\omega^{-\frac{1}{p-1}}|\mathcal
{F}_n)^{p-1}\leq C,
\end{equation}
which is denoted by $\omega\in\bigcup\limits_{p>1}A_p.$
\item \label{Thm:equa_A_exp_infty}There exists a positive
constant $C$ such that for all $n\in
\mathbb{N}$ we have
\begin{equation}\label{A_exp_infty}
\mathbb{E}(\omega|\mathcal {F}_n)\leq C\exp \mathbb{E}(\log\omega|\mathcal {F}_n),
\end{equation}
which is denoted by $\omega\in A^{exp}_{\infty}.$
\item \label{Thm:equa_wAinfty}There exist $0<\gamma,~\delta<1$ such that for all $n\in
\mathbb{N}$ we have
\begin{equation}\label{wAinfty}
\mathbb{E}(\chi_{\{\omega\leq\gamma\omega_n\}}|\mathcal
{F}_n)\leq\delta<1,
\end{equation}
which is denoted by $\omega\in A^{con}_{\infty}.$
\item \label{Thm:equa_R}There exist $0<\alpha,~\beta<1$ such that for all $n\in
\mathbb{N}$ and $A\in\mathcal {F}$ we have
\begin{equation}
\label{R}\mathbb{E}(\chi_A|\mathcal {F}_n)\leq\alpha<1\Rightarrow
\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)\leq\beta<1,\end{equation}
which is denoted by $\omega \in A_{\infty}^{M}.$
\item \label{Thm:equa_RH} There exist $C,~q>1$ such that for all $n\in \mathbb{N}$ we have
\begin{equation}\label{RH}
\mathbb{E}(\omega^q|\mathcal {F}_n)\leq
C\mathbb{E}(\omega|\mathcal {F}_n)^q,\end{equation}
which is the reverse H\"{o}lder condition and denoted by $\omega\in \bigcup\limits_{q>1}RH_{q}$.
\item \label{Thm:equa_RR}There exist $0<\varepsilon'<1$ and $C>1$ such that
for all $n\in \mathbb{N}$ we have
\begin{equation}
\label{RR}\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)\leq C\mathbb{E}(\chi_A|\mathcal
{F}_n)^{\varepsilon'}.
\end{equation}
which is denoted by $\omega\in A^{CF}_{\infty}.$
\item \label{Thm:equa_rR} There exist $0<\alpha,~\beta<1$ such that for all $n\in
\mathbb{N}$ and $A\in\mathcal {F}$ we have
\begin{equation}
\label{rR}\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)\leq\alpha<1\Rightarrow
\mathbb{E}(\chi_A|\mathcal {F}_n)\leq\beta<1,\end{equation}
which is denoted by $\omega \in \hat{A}_{\infty}^{M}.$
\end{enumerate}
\end{theorem}
We remark that $\omega\in S$ is used only in $\ref{Thm:equa_R}\xRightarrow{S}\ref{Thm:equa_RH}$ and
$\ref{Thm:equa_rR}\xRightarrow{S}\ref{Thm:equa_Ap}$ in Theorem \ref{Thm:equa}.
It is natural to discuss what happens without the condition $S.$ Using weights modulo
conditional expectations $\omega/\omega_n$ instead of weights $\omega,$ we study several characterizations of $A_\infty$ weights.
First we have Theorems \ref{thm:Reverse-test} and \ref{thm:Cond}.
\begin{theorem}\label{thm:Reverse-test}The following statements are equivalent.
\begin{enumerate}
\item \label{Reverse-test1}$\omega\in \bigcup\limits_{q>1}RH_{q}$.
\item \label{Reverse-test2}$\omega\in A^{CF}_{\infty}.$
\end{enumerate}
\end{theorem}
\begin{theorem}\label{thm:Cond}The following statements are equivalent.
\begin{enumerate}
\item \label{thm:Cond1}$\omega\in A^{con}_{\infty}.$
\item \label{thm:Cond2}$\omega \in A_{\infty}^{M}.$
\end{enumerate}
\end{theorem}
Using a kind of reverse H\"{o}lder condition which appeared in Str\"{o}mberg and Wheeden \cite{MR766221}, we give a characterization of $\omega\in A^{exp}_{\infty},$ which is Theorem \ref{thm:exp-s}.
\begin{theorem}\label{thm:exp-s}The following statements are equivalent.
\begin{enumerate}
\item \label{thm:exp-s2} $\omega\in A^{exp}_{\infty}.$
\item \label{thm:exp-s1}There exists $C>1$ such that for every $s\in(0,1)$ we have
\begin{equation}\label{thm:eq-exp-s}
\mathbb{E}(\omega|\mathcal {F}_n)\leq C\mathbb{E}(\omega^s|\mathcal {F}_n)^{\frac{1}{s}},\end{equation}
which is denoted by $\omega\in A^{SW}_{\infty}.$
\end{enumerate}
\end{theorem}
Modifying $A_{\infty}^{\lambda}$ and $A_{\infty}^{med}$ in
\cite{MR402038} and \cite{MR529683}, respectively, we have
one-way implications. Indeed,
introducing the quotient $\omega/\omega_n$ into $A_{\infty}^{\lambda}$ of \cite{MR402038}, we obtain Theorem \ref{thm:imp1}.
\begin{theorem}\label{thm:imp1}Let $\omega$ be a weight. We have the sequence of implications
$\eqref{thm:lev}\Rightarrow\eqref{thm:rev} \Rightarrow\eqref{thm:log}\Rightarrow\eqref{thm:imp_Cond2}$
for the following statements.
\begin{enumerate}
\item \label{thm:lev}There exist $0<\beta<1$ and $C>1$ such that for all $n\in \mathbb{N}$ and $\lambda>1$ we have
\begin{equation}\label{thm:equ_lev}
\mathbb{E}_\omega(\chi_{\{\frac{\omega}{\omega_n}>\lambda\}}|\mathcal {F}_n)\leq C\lambda\mathbb{E}(\chi_{\{\frac{\omega}{\omega_n}>\beta\lambda\}}|\mathcal {F}_n),\end{equation}
which is denoted by $\omega\in A_{\infty}^{\lambda}.$
\item \label{thm:rev}
$\omega\in\bigcup\limits_{q>1}RH_q.$
\item \label{thm:log}There exists $C>1$ such that for all $n\in \mathbb{N}$ we have
\begin{equation}\label{thm:equ_log}
\mathbb{E}_{\omega}(\log^+\frac{\omega}{\omega_n}|\mathcal {F}_n)\leq C,
\end{equation}
which is denoted by $\omega\in A_{\infty}^{log}.$
\item \label{thm:imp_Cond2}$\omega \in A_{\infty}^{M}.$
\end{enumerate}
\end{theorem}
As for $A_{\infty}^{med}$ in \cite{MR529683}, we replace the median $m(\omega;Q)$ by the median function $m(\omega,n)$(see Definition \ref{media_f}), which is the key observation in Theorem \ref{thm:imp2}.
\begin{theorem}\label{thm:imp2}Let $\omega$ be a weight. We have the sequence of implications
$\eqref{thm:imp2_Ap}\Rightarrow\eqref{thm:imp2_Log} \Rightarrow\eqref{thm:imp2_Mid}\Rightarrow\eqref{thm:imp2_Dou}$
for the following statements.
\begin{enumerate}
\item \label{thm:imp2_Ap}$\omega\in\bigcup\limits_{p>1} A_p.$
\item \label{thm:imp2_Log}$\omega\in A^{exp}_{\infty}.$
\item \label{thm:imp2_Mid} There exists $C>1$ such that for all $n\in \mathbb{N}$ we have
\begin{equation}\label{thm:imp2_EMid}
\omega_n\leq Cm(\omega,n),
\end{equation}
which is denoted by $\omega\in A_{\infty}^{med}.$
\item \label{thm:imp2_Dou}$\omega \in A_{\infty}^{M}.$
\end{enumerate}
\end{theorem}
Now we give Theorem \ref{thm:Wi} which is related to $A^*_{\infty}$ in \cite{MR481968} and \cite{MR883661}.
The main ingredient of Theorem \ref{thm:Wi} is the conditional expectation of tailed maximal operators(see Definition \ref{tailed_o}). It is worth observing that \eqref{tailed_mo} equals $\mathbb{E}(M^*_n(\omega/\omega_n)|\mathcal {F}_n)\leq C.$
The tailed maximal operators first appeared in \cite{MR1301765} and were used to proved two-weight inequalities for martingales under some additional assumption.
In view of Theorem \ref{thm:Wi}, we have $A^{exp}_{\infty}\subseteq A^*_{\infty}.$
\begin{theorem}\label{thm:Wi} Given the following statements.
\begin{enumerate}
\item \label{thm:Wi_Ap}$\omega\in \bigcup\limits_{p>1}A_p.$
\item \label{thm:Wi_rev}
$\omega\in\bigcup\limits_{q>1}RH_q.$
\item \label{thm:Wi_log}$\omega\in A^{log}_{\infty}.$
\item \label{Thm:Wi_exp}$\omega\in A^{exp}_{\infty}.$
\end{enumerate}
Then each of these statements implies $\omega\in A^*_{\infty},$ i.e., for all $n\in\mathbb{N},$
\begin{equation}\label{tailed_mo}
\mathbb{E}(M^*_n(\omega)|\mathcal {F}_n)\leq C\omega_n.
\end{equation}
\end{theorem}
The paper is organized as follows. Some preliminaries are contained in Sect. \ref{preli}.
In Sect. \ref{regular} we prove Theorem \ref{Thm:equa} for regular weights. Sect. \ref{without} is devoted to
theorems without additional assumptions.
\section{Preliminaries}\label{preli}
Let $(\Omega,\mathcal {F},\mu)$ be a complete probability space and
$(\mathcal {F}_n)_{n\geq0}$ an increasing sequence of
sub$\hbox{-}\sigma\hbox{-}$fields of $\mathcal{F}$ with
$\mathcal{F}=\bigvee_{n\geq0}\mathcal{F}_n.$
The conditional expectation with respect
to $(\Omega,\mathcal{F},\mu,\mathcal{F}_n)$ is denoted by $\mathbb{E}(\cdot|\mathcal{F}_n).$
In this paper, for $p\geq1,$ a martingale
$f=(f_n)_{n\geq0}\in L^p(\omega)$ is meant as $f_n=E(f|\mathcal
{F}_n),~f\in L^p(\omega).$ A weight
$\omega$ is a random variable with $\omega>0$ and $
E(\omega)<\infty.$ Without loss of generality, we may assume
$E(\omega)=1$ since otherwise we can replace $\omega$ by $\omega/E(\omega).$
\begin{definition}\label{tailed_o}The Doob maximal operator $M$ and the tailed maximal
operator $M_n^*$ for martingale $f=(f_n)$ are defined by
\begin{equation*}Mf=\sup\limits_{n\geq 0}|f_n|\hbox{ and }M_n^*f=\sup\limits_{m\geq n}|f_n|,
\end{equation*}
respectively.
\end{definition}
\begin{definition}\label{regular_p}The weight $\omega$ is said to satisfy an regularity condition $S$ , if there exists $C>1$
such that for all $n\in \mathbb{N}$ we have
$$\frac{1}{C}\omega_{n-1}\leq\omega_n\leq C\omega_{n-1},$$
which is denoted by $\omega\in S.$
\end{definition}
Let $\omega$ be a weight and $d\hat{\mu}=\omega
d\mu.$ We denote the conditional expectation
with respect
to $(\Omega,\mathcal{F},\hat{\mu},\mathcal{F}_n)$ by $\hat{\mathbb{E}}(\cdot|\mathcal{F}_n)$ or $\mathbb{E}_{\omega}(\cdot|\mathcal{F}_n).$ It follows that
$$\mathbb{E}_{\omega}(\cdot|\mathcal{F}_n)=\mathbb{E}(\cdot\omega|\mathcal{F}_n)/\omega_n=\mathbb{E}(\cdot\frac{\omega}{\omega_n}|\mathcal{F}_n).$$ If $A\in \mathcal{F}$,
we denote $\int_A\omega d\mu $ by $|A|_\omega$ and $\int_Ad\mu$ by
$|A|$, respectively.
For
$(\Omega,\mathcal {F},\mu)$ and $(\mathcal {F}_n)_{n\geq0},$ the
family of all stopping times is denoted by $\mathcal {T}.$
\begin{definition}\label{media_f}The median function of $\omega$ relative to $\mathcal {F}_n$ is defined as a $\mathcal {F}_n$ measurable function $m(\omega;n)$ such that $\mathbb{E}(\chi_{\{\omega> m(\omega,n)\}}|\mathcal {F}_n)\leq 1/ 2$
and $\mathbb{E}(\chi_{\{\omega< m(\omega,n)\}}|\mathcal {F}_n)\leq 1/ 2.$
\end{definition}
We denote the set of non-negative integers by $\mathbb{N}$ and all integers
by $\mathbb{Z},$ respectively. Throughout the paper letter $C$ always denotes a positive constant which may be different in each occurrence.
\section{Equivalent Characterizations of Regular $A_{\infty}$ Weights}\label{regular}
The following Lemma \ref{key_lemma_sta} will be used in the proof of Theorem \ref{Thm:equa}.
\begin{lemma}\label{key_lemma_sta}Let $v$ be a positive measurable function
and let $0<s_0<+\infty$. If $v\in L^{s_0},$ then
\begin{equation}\label{key_lemma}
\mathbb{E}(v^s|\mathcal {F}_n)^{\frac{1}{s}}\downarrow\exp \mathbb{E}(\log v|\mathcal {F}_n),~\text{~as~}s\downarrow0^+,
\end{equation}\end{lemma}
\begin{proof}[Proof of Lemma \ref{key_lemma_sta}] Because of $v\in L^{s_0},$ we have $v\in L^{s}$ with $0<s<s_0.$
H\"{o}lder's inequality for the conditional expectation(\cite[p.3]{MR1224450}) gives
$\mathbb{E}(v^s|\mathcal {F}_n)^{\frac{1}{s}}\leq\mathbb{E}(v^t|\mathcal {F}_n)^{\frac{1}{t}}$ with $0<s<t<s_0.$
Following Jensen's inequality for the conditional expectation(\cite[p.5]{MR1224450}), we have
$$
\exp \mathbb{E}(\log v^s|\mathcal {F}_n)\leq\mathbb{E}(v^s|\mathcal {F}_n),
$$
which implies \begin{equation}\label{eq_left}\exp \mathbb{E}(\log v|\mathcal {F}_n)\leq\mathbb{E}(v^s|\mathcal {F}_n)^{\frac{1}{s}}.
\end{equation}
Because of $x\leq \exp(x-1)$ for $x>0,$ then
$$\mathbb{E}(v^s|\mathcal {F}_n)\leq\exp \big(\mathbb{E}(v^s|\mathcal {F}_n)-1\big).$$
It follows that
\begin{equation}\label{eq_right}\mathbb{E}(v^s|\mathcal {F}_n)^{\frac{1}{s}}\leq\exp (\frac{\mathbb{E}(v^s|\mathcal {F}_n)-1}{s})=\exp \mathbb{E}(\frac{v^s-1}{s}|\mathcal {F}_n).
\end{equation}
Let $f(x)=\frac{x^s-1}{s}-\frac{x^t-1}{t}$ with $s>t>0$ and $x>0.$ Then $f(1)=0$ is the minimum value of $f$ on $(0,+\infty).$
It follows that for all $x>0$ we have $\frac{x^s-1}{s}\downarrow\log x,$ as $s\downarrow0^+.$ Using Monotone Convergence Theorem for the conditional expectation(\cite[p.5]{MR1224450}), we obtain that
\begin{eqnarray*}
\mathbb{E}\big(\frac{v^{s_0}-1}{s_0}-\frac{v^s-1}{s}|\mathcal {F}_n\big)&\uparrow&\mathbb{E}\big(\frac{v^{s_0}-1}{s_0}-\log v|\mathcal {F}_n\big), ~\text{~as~}s\downarrow0^+.
\end{eqnarray*}
Thus \begin{equation}\label{eq_further}
\lim\limits_{s\rightarrow0^+}\mathbb{E}(\frac{\omega^s-1}{s}|\mathcal {F}_n)=\mathbb{E}(\log\omega|\mathcal {F}_n).
\end{equation}
Combining \eqref{eq_left}, \eqref{eq_right} and\eqref{eq_further}, we deduce
$$\lim\limits_{s\rightarrow0^+}\mathbb{E}(\omega^s|\mathcal {F}_n)^{\frac{1}{s}}=\exp \mathbb{E}(\log\omega|\mathcal {F}_n).$$
This completes the proof of \eqref{key_lemma}
\end{proof}
To prove Theorem \ref{Thm:equa}, we use $\omega\in S$ only in $\ref{Thm:equa_R}\xRightarrow{S}\ref{Thm:equa_RH}$ and
$\ref{Thm:equa_rR}\xRightarrow{S}\ref{Thm:equa_Ap}.$
\begin{proof}[Proof of Theorem \ref{Thm:equa}] We shall follow the scheme:
\begin{center}
\begin{figure}[H]
\begin{tikzpicture}[scale=0.75]
\draw (1,3) node {\ref{Thm:equa_Ap}} -- (1,3);
\draw (4,3) node {\ref{Thm:equa_A_exp_infty}} -- (4,3);
\draw (7,3) node {\ref{Thm:equa_wAinfty}} -- (7,3);
\draw (10,3) node {\ref{Thm:equa_R}} -- (10,3);
\draw (2.5,1) node {\ref{Thm:equa_rR}} -- (2.5,1);
\draw (5.5,1) node {\ref{Thm:equa_RR}} -- (5.5,1);
\draw (8.5,1) node {\ref{Thm:equa_RH}} -- (8.5,1);
\draw (1.4,1.8) node {S} -- (1.4,1.8);
\draw (9.6,1.8) node {S} -- (9.6,1.8);
\draw[-implies,double equal sign distance, line width=0.25mm] (1.5,3) -- (3.5,3);
\draw[-implies,double equal sign distance, line width=0.25mm] (4.5,3) -- (6.5,3);
\draw[-implies,double equal sign distance, line width=0.25mm] (7.5,3) -- (9.5,3);
\draw[-implies,double equal sign distance, line width=0.25mm] (9.8,2.6) -- (8.7,1.4);
\draw[-implies,double equal sign distance, line width=0.25mm] (8,1) -- (6,1);
\draw[-implies,double equal sign distance, line width=0.25mm] (5,1) -- (3,1);
\draw[-implies,double equal sign distance, line width=0.25mm] (2.3,1.4) -- (1.2,2.6);
\draw (0,0) node {} -- (0,0);
\draw (0,3) node {} -- (0,3);
\end{tikzpicture}
\end{figure}
\end{center}
$\ref{Thm:equa_Ap}\Rightarrow\ref{Thm:equa_A_exp_infty}.$ Let $\omega\in A_p.$ Then for all $q\geq p,$ we have $\omega\in A_q.$ In view of Lemma \ref{key_lemma_sta} with $s=\frac{1}{q-1}$ and $v=\frac{1}{\omega},$ we obtain that
$$
\mathbb{E}(\omega^{-\frac{1}{q-1}}|\mathcal {F}_n)^{{q-1}}\downarrow\exp \mathbb{E}(\log \frac{1}{\omega}|\mathcal {F}_n),~\text{~as~}q\uparrow+\infty.
$$
Thus
$$\mathbb{E}(\omega|\mathcal {F}_n)\exp \mathbb{E}(\log\frac{1}{\omega}|\mathcal
{F}_n)\leq C,$$ which implies $\mathbb{E}(\omega|\mathcal {F}_n)\leq C\exp \mathbb{E}(\log\omega|\mathcal
{F}_n).$
$\ref{Thm:equa_A_exp_infty}\Rightarrow\ref{Thm:equa_wAinfty}.$
Fix $n\in \mathbb{N}.$
Letting $v_n=\exp
\mathbb{E}(\log\omega|\mathcal {F}_n),$ we have that
\begin{eqnarray*}
1=\frac{1}{v_n}\exp\mathbb{E}(\log\omega|\mathcal {F}_n)
=\exp\mathbb{E}(\log\frac{\omega}{v_n}|\mathcal {F}_n).
\end{eqnarray*}
It follows that
\begin{equation}\label{equa0}\mathbb{E}(\log\frac{\omega}{v_n}|\mathcal {F}_n)=0.
\end{equation}
Using \eqref{A_exp_infty}, we obtain that
\begin{equation}\label{eqbound}\frac{1}{v_n}\mathbb{E}(\omega|\mathcal {F}_n)
\leq\frac{C}{v_n}\exp
\mathbb{E}(\log\omega|\mathcal {F}_n)
=C.\end{equation}
For some $\gamma>0$ to
be chosen later, we observe that
\begin{eqnarray*}
\{\omega\leq\gamma\omega_n\}
&=&\{\frac{\omega}{v_n}\leq
\frac{\gamma}{v_n}\mathbb{E}(\omega|\mathcal {F}_n)\}\\
&\subseteq&\{\frac{\omega}{v_n}\leq\gamma C\}\\
&\subseteq&\{\log(1+\frac{v_n}{\omega})\geq\log(1+\frac{1}{\gamma
C})\}.\end{eqnarray*}
Thus
\begin{eqnarray*}
\mathbb{E}(\chi_{\{\omega\leq\gamma\omega_n\}}|\mathcal{F}_n)
&\leq&\mathbb{E}(\chi_{\{\log(1+\frac{v_n}{\omega})\geq\log(1+\frac{1}{\gamma C})\}}|\mathcal{F}_n)\\
&\leq&\frac{1}{\log(1+\frac{1}{\gamma
C})}\mathbb{E}\big(\log(1+\frac{v_n}{\omega})|\mathcal
{F}_n\big)\\
&=&\frac{1}{\log(1+\frac{1}{\gamma
C})}\Big(\mathbb{E}\big(\log(1+\frac{\omega}{v_n})|\mathcal
{F}_n\big)-\mathbb{E}\big(\log\frac{\omega}{v_n}|\mathcal
{F}_n\big)\Big)\\
&=&\frac{1}{\log(1+\frac{1}{\gamma
C})}\mathbb{E}\big(\log(1+\frac{\omega}{v_n})|\mathcal
{F}_n\big),\end{eqnarray*}
where we have used \eqref{equa0}.
It follows from \eqref{eqbound} that
\begin{eqnarray*}
\frac{1}{\log(1+\frac{1}{\gamma
C})}\mathbb{E}\big(\log(1+\frac{\omega}{v_n})|\mathcal
{F}_n\big)
&\leq&\frac{1}{\log(1+\frac{1}{\gamma
C})}\mathbb{E}(\frac{\omega}{v_n}|\mathcal{F}_n)\\
&\leq&\frac{C}{\log(1+\frac{1}{\gamma C})}.
\end{eqnarray*}
Since
$\lim\limits_{\gamma\rightarrow0}\frac{C}{\log(1+\frac{1}{\gamma
C})}=0,$ we deduce that \eqref{wAinfty} holds.
$\ref{Thm:equa_wAinfty}\Rightarrow\ref{Thm:equa_R}.$ Fix $n\in \mathbb{N}.$ Suppose that $A\in \mathcal {F}$
with $\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)>\beta$ for some $\beta$ to be
chosen later. Denote $A_1^c=A^c\cap\{\omega>\gamma\omega_n\}$ and
$A_2^c=A^c\cap\{\omega\leq\gamma\omega_n\}.$ Then we have
$$\mathbb{E}(\chi_{A^c}|\mathcal {F}_n)= \mathbb{E}(\chi_{A_1^c}|\mathcal {F}_n)+\mathbb{E}(\chi_{A_2^c}|\mathcal {F}_n)
\leq \frac{1}{\gamma\omega_n}\mathbb{E}(\omega\chi_{A_1^c}|\mathcal
{F}_n)+\delta.$$ Since $\mathbb{E}(\omega\chi_{A^c}|\mathcal
{F}_n)=\mathbb{E}_\omega(\chi_{A^c}|\mathcal {F}_n)\omega_n,$ it follows that
$$\mathbb{E}(\chi_{A^c}|\mathcal
{F}_n)\leq\frac{1}{\gamma}\mathbb{E}_\omega(\chi_{A^c}|\mathcal
{F}_n)+\delta<\frac{1-\beta}{\gamma}+\delta.$$ Because of
$\lim\limits_{\beta\rightarrow1}(\frac{1-\beta}{\gamma}+\delta)=\delta<1,$ it is possible to choose $\beta\in(0,1)$ such that
$\alpha=1-\frac{1-\beta}{\gamma}-\delta\in(0,1).$ Therefore, we
obtain that $$\mathbb{E}(\chi_{A^c}|\mathcal {F}_n)<1-\alpha,$$ which implies $\mathbb{E}(\chi_A|\mathcal {F}_n)>\alpha.$
Thus \ref{Thm:equa_R} is valid.
$\ref{Thm:equa_R}\xRightarrow{S}\ref{Thm:equa_RH}.$ Fix $n\in \mathbb{N}.$ Let
$\tilde{\omega}=:\frac{\omega}{\omega_n}.$ For $m\in \mathbb{N},$ denote
$\tilde{\mathcal {F}}_m=:\mathcal {F}_{n+m}.$ Since $\omega\in
S,$ we obtain a constant $C$ such that
$$\frac{1}{C}\mathbb{E}(\tilde{\omega}|\tilde{\mathcal {F}}_m)\leq
\mathbb{E}(\tilde{\omega}|\tilde{\mathcal {F}}_{m+1})\leq C
\mathbb{E}(\tilde{\omega}|\tilde{\mathcal {F}}_m),~\forall~m\in \mathbb{N}.$$ For
$k\in \mathbb{N}$, define
$\tilde{\tau}_k:=\inf\{m:\mathbb{E}(\tilde{\omega}|\tilde{\mathcal
{F}}_m)>\frac{1}{2}2^{kL}\},$ where $L\geq1$ is a large integer to
be chosen momentarily. Trivially, $\tilde{\tau}_0\equiv0$ and
$\tilde{\tau}_k\geq1,~\forall~k\geq1.$ For $k,~m\geq1,$ we have
\begin{eqnarray*}
\mathbb{E}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k=m\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})
&\leq&\mathbb{E}\big(\frac{E(\tilde{\omega}|\mathcal
{F}_{\tilde{\tau}_{k+1}})}{\frac{1}{2}2^{(k+1)L}}\chi_{\{\tilde{\tau}_k=m\}}|\tilde{\mathcal{F}}_m\big)\\
&=&2\mathbb{E}\big(\frac{E(\tilde{\omega}|\mathcal
{F}_{\tilde{\tau}_{k+1}})}{2^{(k+1)L}}|\tilde{\mathcal{F}}_{\tilde{\tau}_k}\big)\chi_{\{\tilde{\tau}_k=m\}}\\
&=&\frac{2}{2^{(k+1)L}}\mathbb{E}\big(\mathbb{E}(\tilde{\omega}|\tilde{\mathcal
{F}}_{\tilde{\tau}_{k+1}})|\tilde{\mathcal{F}}_{\tilde{\tau}_k}\big)\chi_{\{\tilde{\tau}_k=m\}}.
\end{eqnarray*} It is clear that $\tilde{\tau}_k\leq\tilde{\tau}_{k+1},$ then
\begin{eqnarray*}
\mathbb{E}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k=m\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})
&\leq&\frac{2}{2^{(k+1)L}}\mathbb{E}(\tilde{\omega}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})\chi_{\{\tilde{\tau}_k=m\}}\\
&=&\frac{2}{2^{(k+1)L}}\mathbb{E}(\tilde{\omega}|\tilde{\mathcal
{F}}_m)\chi_{\{\tilde{\tau}_k=m\}}\\
&\leq&\frac{2C}{2^{(k+1)L}}\mathbb{E}(\tilde{\omega}|\tilde{\mathcal
{F}}_{m-1})\chi_{\{\tilde{\tau}_k=m\}}\\
&\leq&\frac{2^{kL}C}{2^{(k+1)L}}\chi_{\{\tilde{\tau}_k=m\}}.\end{eqnarray*}
Here
we choose $L$ so large that $\frac{2^{kL}C}{2^{(k+1)L}}\leq \alpha,$
then$$
\mathbb{E}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k=m\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})
\leq\alpha\chi_{\{\tilde{\tau}_k=m\}}.$$ Moreover,
\begin{eqnarray*}
\mathbb{E}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}\cap{\{\tilde{\tau}_k=m\}}}|\mathcal
{F}_{n+m})
&=&\mathbb{E}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k=m\}}|\mathcal
{F}_{n+m})\\
&=&\mathbb{E}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k=m\}}|\tilde{\mathcal
{F}}_m)\\
&=&\mathbb{E}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k=m\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})\\
&\leq&\alpha\chi_{\{\tilde{\tau}_k=m\}}\leq\alpha.\end{eqnarray*}
Combining with
$\mathbb{E}_{\tilde{\omega}}(\cdot|\mathcal
{F}_{n+m})=\mathbb{E}_\omega(\cdot|\mathcal
{F}_{n+m})$ and\eqref{R}, we have $$
\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}\cap{\{\tilde{\tau}_k=m\}}}|\mathcal
{F}_{n+m})\leq\beta.$$
Thus
\begin{eqnarray*}
\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{{\{\tilde{\tau}_k=m\}}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})
&=&\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}\cap{\{\tilde{\tau}_k=m\}}}|\mathcal
{F}_{n+m})\chi_{\{\tilde{\tau}_k=m\}}\\
&\leq&\beta\chi_{\{\tilde{\tau}_k=m\}}.\end{eqnarray*} Consequently,
\begin{eqnarray*}
\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})
&=&\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k<\infty\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})\\
&=&\sum\limits_{m=1}\limits^{\infty}\mathbb{E}_{\tilde{\omega}}
(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}\chi_{\{\tilde{\tau}_k=m\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})\\
&=&\sum\limits_{m=1}\limits^{\infty}\beta\chi_{\{\tilde{\tau}_k=m\}}\leq\beta\chi_{\{\tilde{\tau}_k<\infty\}}.\end{eqnarray*}
It follows that \begin{eqnarray*}
\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}|\mathcal
{F}_n)
&=&\mathbb{E}_{\tilde{\omega}}(\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_{k+1}<\infty\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_k})|\tilde{\mathcal
{F}}_{\tilde{\tau}_0})\\
&\leq&\beta
\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_k<\infty\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_0})\\
&\leq&\beta^k
\mathbb{E}_{\tilde{\omega}}(\chi_{\{\tilde{\tau}_1<\infty\}}|\tilde{\mathcal
{F}}_{\tilde{\tau}_0})\\
&\leq&\beta^k,\end{eqnarray*} which is also valid for $k=0.$ For some positive
number $\varepsilon$ to be determined later, we have \begin{eqnarray*}
\mathbb{E}\big((\frac{\omega}{\omega_n})^{1+\varepsilon}|\mathcal
{F}_n\big)
&=&\mathbb{E}(\tilde{\omega}^{1+\varepsilon}|\mathcal
{F}_n)\\
&=&\mathbb{E}(\tilde{\omega}^\varepsilon\tilde{\omega}|\mathcal
{F}_n)\\
&\leq&\mathbb{E}\big(M^*_n(\tilde{\omega})^\varepsilon\tilde{\omega}|\mathcal
{F}_n\big),\end{eqnarray*} where $M^*_n(\cdot)=\sup\limits_{m\geq0}\mathbb{E}(\cdot|\mathcal
{F}_{m+n})=\sup\limits_{m\geq0}\mathbb{E}(\cdot|\tilde{\mathcal
{F}}_{m}).$
Because $\bigcap\limits_{k\in \mathbb{N}}\{\tilde{\tau}_k=\infty\}={\O},$
$\{\tilde{\tau}_0<\infty\}=\Omega$ and $\mathbb{E}_{\tilde{\omega}}\big(
\chi_{\{\tilde{\tau}_0<\infty\}}|\mathcal
{F}_n\big)=1,$
we obtain that \begin{eqnarray*}
\mathbb{E}\big((\frac{\omega}{\omega_n})^{1+\varepsilon}|\mathcal
{F}_n\big)&\leq&\sum\limits_{k=0}\limits^{\infty}
\mathbb{E}\big((M^*_n(\tilde{\omega})^\varepsilon\tilde{\omega}
\chi_{\{\tilde{\tau}_k<\infty\}\cap\{\tilde{\tau}_{k+1}=\infty\}}|\mathcal
{F}_n\big)\\
&\leq&\frac{1}{2^\varepsilon}\sum\limits_{k=0}\limits^{\infty}
2^{(k+1)\varepsilon L}\mathbb{E}\big(\tilde{\omega}
\chi_{\{\tilde{\tau}_k<\infty\}\cap\{\tilde{\tau}_{k+1}=\infty\}}|\mathcal
{F}_n\big)\\
&=&\frac{1}{2^\varepsilon}\sum\limits_{k=0}\limits^{\infty}
2^{(k+1)\varepsilon L}\mathbb{E}_{\tilde{\omega}}\big(
\chi_{\{\tilde{\tau}_k<\infty\}\cap\{\tilde{\tau}_{k+1}=\infty\}}|\mathcal
{F}_n\big)\\
&\leq&\frac{1}{2^\varepsilon}\sum\limits_{k=0}\limits^{\infty}
2^{(k+1)\varepsilon L}\mathbb{E}_{\tilde{\omega}}\big(
\chi_{\{\tilde{\tau}_k<\infty\}}|\mathcal
{F}_n\big)\\
&\leq&\frac{1}{2^\varepsilon}\sum\limits_{k=0}\limits^{\infty}
2^{(k+1)\varepsilon L}\beta^{k-1}\\
&=&\frac{2^{\varepsilon
(L-1)}}{\beta}\sum\limits_{k=0}\limits^{\infty}
(2^{\varepsilon L}\beta)^k.\end{eqnarray*}
Choosing an $\varepsilon$ small enough, we have
$\sum\limits_{k=0}\limits^{\infty}
(2^{\varepsilon L}\beta)^k<\infty.$ Thus, \eqref{RH} is
valid with $C=\frac{2^{\varepsilon
(L-1)}}{\beta}\sum\limits_{k=0}\limits^{\infty}
(2^{\varepsilon L}\beta)^k$ and $q=1+\varepsilon.$
$\ref{Thm:equa_RH}\Rightarrow\ref{Thm:equa_RR}.$ For $A\in\mathcal {F},$ it is clear that
$\mathbb{E}_\omega(\chi_{A}|\mathcal
{F}_n)\omega_n=E(\omega\chi_{A}|\mathcal {F}_n).$ Applying
H\"{o}lder's inequality, we obtain that
$$\mathbb{E}_\omega(\chi_{A}|\mathcal {F}_n)\omega_n
\leq \mathbb{E}(\omega^{1+\varepsilon}|\mathcal
{F}_n)^{\frac{1}{1+\varepsilon}}\mathbb{E}(\chi_{A}|\mathcal
{F}_n)^{\frac{\varepsilon}{1+\varepsilon}}.$$ It follows from
\eqref{RH} that $$\mathbb{E}_\omega(\chi_{A}|\mathcal {F}_n)\leq
C^{\frac{1}{1+\varepsilon}}\mathbb{E}(\chi_{A}|\mathcal
{F}_n)^{\frac{\varepsilon}{1+\varepsilon}},$$ which implies
\eqref{RR} with
$\varepsilon'=\frac{\varepsilon}{1+\varepsilon}.$
$\ref{Thm:equa_RR}\Rightarrow\ref{Thm:equa_rR}.$ For $\varepsilon'$ and $C$ in the assumption \ref{Thm:equa_RR}, we fix $\alpha'$ small enough such that
$\beta'=C\alpha'^{\varepsilon'}<1.$ For $B\in\mathcal {F}$ with
$\mathbb{E}(\chi_B|\mathcal {F}_n)\leq\alpha',$ it follows from \eqref{RR}
that$$\mathbb{E}_\omega(\chi_B|\mathcal {F}_n)\leq\beta'.$$ Thus for all $A\in \mathcal{F}$ we have
$$\mathbb{E}(\chi_A|\mathcal {F}_n)>1-\alpha'\Rightarrow
\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)>1-\beta'.$$
Let $\alpha=1-\beta'$ and $\beta=1-\alpha'.$ Then
$$\mathbb{E}(\chi_A|\mathcal {F}_n)>\beta\Rightarrow
\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)>\alpha,$$
which is equivalent to \ref{Thm:equa_rR}.
$\ref{Thm:equa_rR}\xRightarrow{S}\ref{Thm:equa_Ap}.$ Let $\omega_1=\frac{1}{\omega}.$ Then
$\omega_1$ is a weight relative to $\omega d\mu.$ Recalling the definition of $\hat{\mathbb{E}}(\cdot|\mathcal {F}_n),$ we deduce that
$\mathbb{E}(\cdot|\mathcal {F}_n)=\hat{\mathbb{E}}_{\omega_1}(\cdot|\mathcal
{F}_n),~\forall ~n\in \mathbb{N}.$ It follows from \eqref{rR} that
$$\hat{\mathbb{E}}(\chi_A|\mathcal {F}_n)\leq\alpha<1\Rightarrow
\hat{\mathbb{E}}_{\omega_1}(\chi_A|\mathcal {F}_n)\leq\beta<1,~\forall~A\in
\mathcal{F}.$$ Combining with $\omega_1\in S_{(\omega d\mu)},$ similar to
$\ref{Thm:equa_R}\xRightarrow{S}\ref{Thm:equa_RH},$ we have $\varepsilon$ and $C$ such that
$$\hat{\mathbb{E}}(\omega_1^{1+\varepsilon}|\mathcal {F}_n)\leq
C\hat{\mathbb{E}}(\omega_1|\mathcal {F}_n)^{1+\varepsilon},~\forall~n\in
\mathbb{N},$$ that is, $$\omega_n\mathbb{E}(\omega^{-\varepsilon}|\mathcal
{F}_n)^{\frac{1}{\varepsilon}}\leq
C^{\frac{1}{\varepsilon}},~\forall~n\in \mathbb{N}.$$ Thus \eqref{Ap} is
valid with
$\varepsilon=\frac{1}{p-1}.$
\end{proof}
\section{$A_{\infty}$ Weights without Additional Assumptions}\label{without}
In this section, we study relations between different characterizations of $A_{\infty}$ weights without additional assumptions. These relations are showed in Figure \ref{figure}.
\begin{center}
\begin{figure}[H]
\begin{tikzpicture}[scale=0.85]
\draw (1,5) node {$\boxed{A^{CF}_{\infty}}$} -- (1,5);
\draw (4,-1) node {$\boxed{A^{con}_{\infty}}$} -- (4,-1);
\draw (4,3) node {$\boxed{A^{log}_\infty}$} -- (4,3);
\draw (4,5) node {$\boxed{\cup_{p>1}RH_p}$} -- (4,5);
\draw (4,7) node {$\boxed{A^{\lambda}_{\infty}}$} -- (4,7);
\draw (7,-1) node {$\boxed{A^{M}_\infty}$} -- (7,-1);
\draw (7,1) node {$\boxed{A^{*}_\infty}$} -- (7,1);
\draw (10,3) node {$\boxed{A^{med}_{\infty}}$} -- (10,3);
\draw (10,5) node {$\boxed{A_{\infty}^{exp}}$} -- (10,5);
\draw (10,7) node {$\boxed{\cup_{p>1}A_p}$} -- (10,7);
\draw (13,5) node {$\boxed{A_{\infty}^{SW}}$} -- (13,5);
\draw[implies-implies,double equal sign distance, line width=0.25mm] (4.7,-1) -- (6.4,-1);
\draw[implies-implies,double equal sign distance, line width=0.25mm] (1.7,5) -- (2.8,5);
\draw[implies-implies,double equal sign distance, line width=0.25mm] (10.7,5) -- (12.3,5);
\draw[-implies,double equal sign distance, line width=0.25mm] (4,4.5) -- (4,3.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (4,6.5) -- (4,5.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (4,2.5) -- (6.9,-0.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (4.7,2.5) -- (6.5,1.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (5.2,4.5) -- (6.9,1.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (8.95,6.5) -- (7.1,1.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (9.3,4.5)-- (7.5,1.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (10,4.5) -- (10,3.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (10,6.5) -- (10,5.5);
\draw[-implies,double equal sign distance, line width=0.25mm] (10,2.5) -- (7.1,-0.5);
\draw (0,0) node {} -- (0,0);
\draw (0,7.5) node {} -- (0,7.5);
\end{tikzpicture}
\caption{Relations without additional assumptions}
\label{figure}
\end{figure}
\end{center}
We first prove equivalent characterizations in Theorems \ref{thm:Reverse-test}, \ref{thm:Cond} and \ref{thm:exp-s}.
\begin{proof}[Proof of Theorem \ref{thm:Reverse-test}]
\eqref{Reverse-test1}$\Rightarrow$\eqref{Reverse-test2} This is $\ref{Thm:equa_RH}\Rightarrow\ref{Thm:equa_RR}$ in the proof of Theorem \ref{Thm:equa}. It holds without additional assumptions $\omega\in S.$
\eqref{Reverse-test2}$\Rightarrow$ \eqref{Reverse-test1} Let $B\in \mathcal {F}_n$ and let $E_{\lambda}=\{\frac{\omega}{\omega_n}>\lambda\}.$ We have
\begin{eqnarray*}
\lambda\mu(B\cap E_{\lambda})
&\leq&\int_{B\cap E_{\lambda}}\frac{\omega}{\omega_n}d\lambda\\
&=&\int_{B}\frac{\omega\chi_{E_{\lambda}}}{\omega_n}d\mu\\
&=&\int_{B}\frac{\mathbb{E}(\omega\chi_{E_{\lambda}}|\mathcal
{F}_n)}{\omega_n}d\mu\\
&=&\int_{B}\mathbb{E}_{\omega}(\chi_{E_{\lambda}}|\mathcal
{F}_n)d\mu.
\end{eqnarray*}
It follows from \eqref{Reverse-test2} and H\"{o}lder's inequality that
\begin{eqnarray*}
\lambda\mu(B\cap E_{\lambda})
&\leq&C\int_{B}\mathbb{E}(\chi_{E_{\lambda}}|\mathcal{F}_n)^{\varepsilon'}d\mu\\
&\leq&C\big(\int_{B}\mathbb{E}(\chi_{E_{\lambda}}|\mathcal{F}_n)d\mu\big)^{\varepsilon'}\mu(B)^{1-\varepsilon'}\\
&=&C\mu(B\cap E_{\lambda})^{\varepsilon'}\mu(B)^{1-\varepsilon'}.
\end{eqnarray*}
Thus, we have $\lambda^{\frac{1}{1-\varepsilon'}}\mu(B\cap E_{\lambda})\leq C^{\frac{1}{1-\varepsilon'}}\mu(B).$
Let $1<q<\frac{1}{1-\varepsilon'}.$ We obtain that
\begin{eqnarray*}
\int_{B}(\frac{\omega}{\omega_n})^q\mu
&=&q\int_{0}^{+\infty}\lambda^{q-1}\mu(B\cap E_{\lambda})d\lambda\\
&=&q\int_{0}^{1}\lambda^{q-1}\mu(B\cap E_{\lambda})d\lambda
+q\int_{1}^{+\infty}\lambda^{q-1}\mu(B\cap E_{\lambda})d\lambda\\
&\leq&q(\int_{0}^{1}\lambda^{q-1}d\lambda)\mu(B)
+qC^{\frac{1}{1-\varepsilon'}}(\int_{1}^{+\infty}\lambda^{q-1-\frac{1}{1-\varepsilon'}}d\lambda)\mu(B)\\
&=&\mu(B)+\frac{C^{\frac{1}{1-\varepsilon'}}q}{\frac{1}{1-\varepsilon'}-q}\mu(B)\\
&=&(1+\frac{C^{\frac{1}{1-\varepsilon'}}q}{\frac{1}{1-\varepsilon'}-q})\mu(B).
\end{eqnarray*}
Since $B$ is arbitrary, we obtain
$$\mathbb{E}(\omega^q|\mathcal {F}_n)(\frac{1}{\omega_n})^q
\leq 1+\frac{C^{\frac{1}{1-\varepsilon'}}q}{\frac{1}{1-\varepsilon'}-q}.$$
Then $\mathbb{E}(\omega^q|\mathcal {F}_n)\leq (1+\frac{C^{\frac{1}{1-\varepsilon'}}q}{\frac{1}{1-\varepsilon'}-q})(\omega_n)^q.$
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:Cond}]
Because \eqref{thm:Cond1}$\Rightarrow$\eqref{thm:Cond2} is $\ref{Thm:equa_wAinfty}\Rightarrow\ref{Thm:equa_R}$ in the proof of Theorem \ref{Thm:equa} without further
assumptions $\omega\in S,$
it suffices to prove \eqref{thm:Cond2}$\Rightarrow$
\eqref{thm:Cond1}. Let $0<\gamma <1-\beta.$ Setting $E=\{\omega\leq\gamma\omega_n\},$ we have
\begin{eqnarray*}
\mathbb{E}_\omega(\chi_{E}|\mathcal {F}_n)&=&\frac{\mathbb{E}(\omega\chi_{E}|\mathcal {F}_n)}{\omega_n}\\
&\leq&\frac{\mathbb{E}(\gamma\omega_n\chi_{E}|\mathcal {F}_n)}{\omega_n}\\
&\leq&\gamma\mathbb{E}(\chi_{E}|\mathcal {F}_n)\\
&\leq&\gamma<1-\beta.
\end{eqnarray*}
It follows that $\mathbb{E}_\omega(\chi_{E^c}|\mathcal {F}_n)>\beta.$ In view of \eqref{thm:Cond2}, we obtain that $$\mathbb{E}(\chi_{E^c}|\mathcal {F}_n)>\alpha,$$
which implies $\mathbb{E}(\chi_{E}|\mathcal {F}_n)\leq1-\alpha.$
Thus $\eqref{thm:Cond1}$ is valid with $1-\alpha<\delta<1.$
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:exp-s}]
In view of Lemma \ref{key_lemma_sta}, we have that
\begin{equation*}
\mathbb{E}(\omega^s|\mathcal {F}_n)^{\frac{1}{s}}\downarrow\exp \mathbb{E}(\log\omega|\mathcal {F}_n),~\text{~as~}s\downarrow0^+,
\end{equation*}
which establishes the
equivalence between \eqref{thm:exp-s2} and \eqref{thm:exp-s1}.
\end{proof}
In the rest of this section, we prove Theorems \ref{thm:imp1}, \ref{thm:imp2} and \ref{thm:Wi}.
These are one-way implications.
\begin{proof}[Proof of Theorem \ref{thm:imp1}]
$\eqref{thm:lev}\Rightarrow\eqref{thm:rev}$
Let $B\in \mathcal {F}_n.$ We have
\begin{eqnarray*}
\int_B(\frac{\omega}{\omega_n})^{1+\delta}d\mu
&=&\int_B(\frac{\omega}{\omega_n})^{\delta}\frac{\omega}{\omega_n}d\mu\\
&=&\delta\int_0^{+\infty}\lambda^{\delta-1}
\frac{\omega}{\omega_n}(B\cap\{\frac{\omega}{\omega_n}>\lambda\})d\lambda\\
&=&\delta\int_0^{1}\lambda^{\delta-1}
\frac{\omega}{\omega_n}(B\cap\{\frac{\omega}{\omega_n}>\lambda\})d\lambda\\
&&+\delta\int_1^{+\infty}\lambda^{\delta-1}
\frac{\omega}{\omega_n}(B\cap\{\frac{\omega}{\omega_n}>\lambda\})d\lambda.
\end{eqnarray*}
It follows that
\begin{eqnarray*}
\delta\int_0^{1}\lambda^{\delta-1}
\frac{\omega}{\omega_n}(B\cap\{\frac{\omega}{\omega_n}>\lambda\})d\lambda
&\leq&\frac{\omega}{\omega_n}(B)\delta\int_0^{1}\lambda^{\delta-1}
d\lambda\\
&=&\mu(B),
\end{eqnarray*}
where we have used $\frac{\omega}{\omega_n}(B)
=\int_B\frac{\omega}{\omega_n}d\mu
=\int_B\frac{\omega_n}{\omega_n}d\mu=\mu(B).$
Using \eqref{thm:equ_lev}, we obtain the following estimate
\begin{eqnarray*}
&~&\delta\int_{1}^{+\infty}\lambda^{\delta-1}
\frac{\omega}{\omega_n}(B\cap\{\frac{\omega}{\omega_n}>\lambda\})d\lambda\\
&\leq&C\delta\int_1^{+\infty}\lambda^{\delta}\mu(B\cap\{\frac{\omega}{\omega_n}>\beta\lambda\})
d\lambda\\
&=&\frac{C\delta}{\beta^{1+\delta}}\int_\beta^{+\infty}\lambda^{\delta}(B\cap\{\frac{\omega}{\omega_n}>\lambda\})
d\lambda\\
&\leq&\frac{C\delta}{(1+\delta)\beta^{1+\delta}}\int_B(\frac{\omega}{\omega_n})^{1+\delta}d\mu.
\end{eqnarray*}
Because of $\lim\limits_{\delta\rightarrow0}\frac{C\delta}{(1+\delta)\beta^{1+\delta}}=0,$
we can choose $\delta$ such that $\frac{C\delta}{(1+\delta)\beta^{1+\delta}}<\frac{1}{2}.$ Then we have
$$\int_B(\frac{\omega}{\omega_n})^{1+\delta}d\mu\leq2\mu(B).$$
Since $B\in\mathcal {F}_n$ is arbitrary, it follows that $$\mathbb{E}(\omega^{1+\delta}|\mathcal {F}_n)\leq2(\omega_n)^{1+\delta}.$$
$\eqref{thm:rev} \Rightarrow\eqref{thm:log}$
Let $E_k=\{2^k<\frac{\omega}{\omega_n}\leq2^{k+1}\}$ for $k\in \mathbb{N}.$ In view of \eqref{thm:rev}, we have
\begin{eqnarray*}
2^{kp}\mathbb{E}(\chi_{E_k}|\mathcal {F}_n)&\leq&\mathbb{E}\big((\chi_{E_k}\frac{\omega}{\omega_n})^{p}|\mathcal {F}_n\big)\\&\leq&\mathbb{E}\big((\frac{\omega}{\omega_n})^{p}|\mathcal {F}_n\big)\\&\leq& C,
\end{eqnarray*}
which implies $\mathbb{E}(\chi_{E_k}|\mathcal {F}_n)\leq C2^{-kp}.$
It follows that
\begin{eqnarray*}
\mathbb{E}_{\omega}(\log^+\frac{\omega}{\omega_n}|\mathcal {F}_n)
&=&\mathbb{E}_{\omega}(\sum\limits_{k=0}^{+\infty}\chi_{E_k}\log^+\frac{\omega}{\omega_n}|\mathcal {F}_n)\\
&=&\sum\limits_{k=0}^{+\infty}\mathbb{E}_{\omega}(\chi_{E_k}\log^+\frac{\omega}{\omega_n}|\mathcal {F}_n)\\
&=&\sum\limits_{k=0}^{+\infty}\mathbb{E}(\chi_{E_k}\frac{\omega}{\omega_n}\log^+\frac{\omega}{\omega_n}|\mathcal {F}_n)\\
&\leq&\sum\limits_{k=0}^{+\infty}2^{k+1} (k+1)\mathbb{E}(\chi_{E_k}|\mathcal {F}_n)\\
&\leq&C\sum\limits_{k=0}^{+\infty}(k+1)2^{k+1}2^{-kq},
\end{eqnarray*}
where the series $\sum\limits_{k=0}^{+\infty}(k+1)2^{k+1}2^{-kq}$ is convergent. Then we have \eqref{thm:equ_log}.
$\eqref{thm:log}\Rightarrow\eqref{thm:imp_Cond2}$
Let $\mathbb{E}(\chi_A|\mathcal {F}_n)\leq\alpha<1.$ Recall that $ab\leq a\log a-a+e^b$ where $a>1$ and $b\geq0.$
Then
\begin{eqnarray*}
\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)
&=& \mathbb{E}_\omega(\chi_{A\cap\{\frac{\omega}{\omega_n}\leq1\}}|\mathcal {F}_n)+ \mathbb{E}_\omega(\chi_{A\cap\{\frac{\omega}{\omega_n}>1\}}|\mathcal {F}_n)\\
&=& \mathbb{E}(\frac{\omega}{\omega_n}\chi_{A\cap\{\frac{\omega}{\omega_n}\leq1\}}|\mathcal {F}_n)+ \mathbb{E}(\frac{\omega}{\omega_n}\chi_{A\cap\{\frac{\omega}{\omega_n}>1\}}|\mathcal {F}_n)\\
&\leq&
\mathbb{E}(\chi_A|\mathcal {F}_n)+\frac{1}{b+1}\mathbb{E}(\frac{\omega}{\omega_n}\log^+\frac{\omega}{\omega_n}+e^b\chi_A|\mathcal {F}_n)\\
&=&
\mathbb{E}(\chi_A|\mathcal {F}_n)+\frac{1}{b+1}\mathbb{E}(\frac{\omega}{\omega_n}\log^+\frac{\omega}{\omega_n}|\mathcal{F}_n)+\frac{e^b}{b+1}\mathbb{E}(\chi_A|\mathcal {F}_n)\\
&\leq&\alpha(1+\frac{e^b}{b+1})+\frac{C}{b+1}.
\end{eqnarray*}
Setting $b=2C-1,$ we can pick an $\alpha$ small enough that
$\alpha(1+\frac{e^b}{b+1})\leq\frac{1}{4}$ because of $\lim\limits_{\alpha\rightarrow0}\alpha(1+\frac{e^b}{b+1})=0.$
Thus $\mathbb{E}_\omega(\chi_A|\mathcal {F}_n)\leq\frac{3}{4}.$
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:imp2}] It suffices to prove $\eqref{thm:imp2_Log} \Rightarrow\eqref{thm:imp2_Mid}\Rightarrow\eqref{thm:imp2_Dou},$ because $\eqref{thm:imp2_Ap}\Rightarrow\eqref{thm:imp2_Log}$ is the one $\ref{Thm:equa_Ap}\Rightarrow\ref{Thm:equa_A_exp_infty}$ in Theorem \ref{Thm:equa}.
$\eqref{thm:imp2_Log} \Rightarrow\eqref{thm:imp2_Mid}$
Let $E=\{\omega>m(\omega,n)\}.$ Using H\"{o}lder's inequality for the conditional expectation,
we have
\begin{eqnarray*}
\mathbb{E}(\omega^s\chi_E|\mathcal {F}_n)
&\leq&\mathbb{E}(\omega|\mathcal {F}_n)^s\mathbb{E}(\chi_E|\mathcal {F}_n)^{1-s}\\
&\leq&2^{s-1}C^s\mathbb{E}(\omega^s|\mathcal {F}_n),
\end{eqnarray*}
where we have used Theorem \ref{thm:exp-s}.
It follows that $\mathbb{E}(\omega^s\chi_E|\mathcal {F}_n)
\leq\frac{3}{4}\mathbb{E}(\omega^s|\mathcal {F}_n)$ provided $2^{s-1}C^s<\frac{3}{4}.$
Then $\mathbb{E}(\omega^s\chi_{E^c}|\mathcal {F}_n)
\geq\frac{1}{4}\mathbb{E}(\omega^s|\mathcal {F}_n).$ Thus
\begin{eqnarray*}
\frac{1}{4}\mathbb{E}(\omega|\mathcal {F}_n)^s &\leq&\frac{1}{4}C^s\mathbb{E}(\omega^s|\mathcal {F}_n)\\ &\leq&C^s\mathbb{E}(\omega^s\chi_{E^c}|\mathcal {F}_n)\\
&\leq&C^s\mathbb{E}(m(\omega,n)^s\chi_{E^c}|\mathcal {F}_n)\\
&\leq&C^s\mathbb{E}(m(\omega,n)^s|\mathcal {F}_n)\\
&=&C^sm(\omega,n)^s,
\end{eqnarray*}
which implies $\mathbb{E}(\omega|\mathcal {F}_n)\leq4^{\frac{1}{s}}Cm(\omega,n).$
$\eqref{thm:imp2_Mid}\Rightarrow\eqref{thm:imp2_Dou}$ Let $\alpha<\frac{1}{4}$ and $\mathbb{E}(\chi_E|\mathcal {F}_n)<\frac{1}{4}.$ We claim that $$\mathbb{E}(\chi_{E^c\cap\{\omega\geq m(\omega,n)\}}|\mathcal {F}_n)\geq\frac{1}{4}.$$
Indeed, we have
\begin{eqnarray*}\mathbb{E}(\chi_{E\cup\{\omega< m(\omega,n)\}}|\mathcal {F}_n)
&\leq&\mathbb{E}(\chi_E|\mathcal {F}_n)+\mathbb{E}(\chi_{\{\omega< m(\omega,n)\}}|\mathcal {F}_n)\\
&<&\frac{1}{4}+\frac{1}{2}=\frac{3}{4}.\end{eqnarray*}
This proves that $$\mathbb{E}(\chi_{E^c\cap\{\omega\geq m(\omega,n)\}}|\mathcal {F}_n)\geq\frac{1}{4}.$$
It follows that
\begin{eqnarray*}
\frac{1}{4}\omega_n
&\leq& \frac{C}{4}m(\omega,n)\\
&\leq& Cm(\omega,n)\mathbb{E}(\chi_{E^c\cap\{\omega\geq m(\omega,n)\}}|\mathcal {F}_n)\\
&=& C\mathbb{E}(m(\omega,n)\chi_{E^c\cap\{\omega\geq m(\omega,n)\}}|\mathcal {F}_n)\\
&\leq& C\mathbb{E}(\omega\chi_{E^c\cap\{\omega\geq m(\omega,n)\}}|\mathcal {F}_n)\\
&\leq& C\mathbb{E}(\omega\chi_{E^c}|\mathcal {F}_n).
\end{eqnarray*}
Then we have $\omega_n\leq 4C\mathbb{E}(\omega\chi_{E^c}|\mathcal {F}_n)$ which implies $1\leq 4C\mathbb{E}_{\omega}(\chi_{E^c}|\mathcal {F}_n).$ Thus $\mathbb{E}_{\omega}(\chi_{E}|\mathcal {F}_n)<\beta$ with $\beta=1-\frac{1}{4C}.$
\end{proof}
Before we prove Theorem \ref{thm:Wi}, we make a couple of observations on the tailed maximal operator which are Lemmas
\ref{doob_con} and \ref{doob_imp}. Lemma \ref{doob_con} is
the conditional version of Doob's inequality which appeared in \cite[p.189]{MR1301765}.
\begin{lemma}\label{doob_con} Let $p>1.$ There exists $C>1$ such that for all $n\in \mathbb{N}$ we have
$$\mathbb{E}(M^*_n(f)^{p}|\mathcal {F}_n)\leq C\mathbb{E}(f^{p}|\mathcal {F}_n).$$
\end{lemma}
Lemma \ref{doob_imp} shows that the tailed operator has the following local property.
\begin{lemma}\label{doob_imp} Let $f\in L^1.$ For all $\lambda>0$ and $n\in\mathbb{N}$ we have
$$\mu(B\cap\{M^*_n(f)>\lambda\}\leq\frac{2}{\lambda}\int_{B\cap\{|f|>\lambda/2\}}|f|d\mu,$$
where $B\in \mathcal{F}_n.$
\end{lemma}
\begin{proof} [Proof of Lemma \ref{doob_imp}] Because of $B\in \mathcal{F}_n,$ we have
\begin{eqnarray*}
&~&\mu(B\cap\{M^*_n(f)>\lambda\}\\&=&\mu(\{M^*_n(f\chi_{B})>\lambda\})\\
&\leq&\mu(\{M^*_n(f\chi_{B\cap\{|f|>\lambda/2\}})>\lambda/2\})+\mu(\{M^*_n(f\chi_{B\cap\{|f|\leq\lambda/2\}})>\lambda/2\})\\
&=&\mu(\{M^*_n(f\chi_{B\cap\{|f|>\lambda/2\}})>\lambda/2\}).
\end{eqnarray*}
For $m\in \mathbb{N},$ denote
$\tilde{\mathcal {F}}_m=:\mathcal {F}_{n+m}$ and $\tilde{f}_m=:\mathbb{E}(f\chi_{B\cap\{|f|>\lambda/2\}}|\mathcal {F}_{n+m}).$
It follows that $(\tilde{f}_m)_{m\geq0}$ is a martingale with respect
to $(\Omega,\mathcal{F},\mu,(\tilde{\mathcal{F}}_m)_{m\geq0}),$ which leads to
$\tilde{M}(\cdot)=M^*_n(\cdot).$ Using the weak $(1,1)$ type inequality for $\tilde{M}(\cdot),$ we have
\begin{eqnarray*}
\mu(\{M^*_n(f\chi_{B\cap\{|f|>\lambda/2\}})>\lambda/2\})
&\leq&\frac{2}{\lambda}\int_{B\cap\{|f|>\lambda/2\}}|f|d\mu,
\end{eqnarray*}
which completes the proof.
\end{proof}
Using these Lemmas, we give the proof of Theorem \ref{thm:Wi}.
\begin{proof}[Proof of Theorem \ref{thm:Wi}] In each case, we show that there exists $C>1$ such that
$$\int_{B}M^*_n(\omega)d\mu\leq C\int_{B}\omega d\mu,$$
for all $B\in \mathcal {F}_n.$ Since B is arbitrary, this proves $\mathbb{E}(M^*_n(\omega)|\mathcal {F}_n)\leq C\omega_n.$
\eqref{thm:Wi_Ap}
Because $\omega\in A_p,$ we have $\omega^{1-p^\prime}\in A_{p^\prime}$ with $1/p+1/p'=1.$ Recall that $M^*_n$ is bounded on $L^{p^\prime}(\omega^{1-p^\prime})$(\cite[Theorem 6.6.3]{MR1224450}). Then
\begin{eqnarray*}
\int_{B}M^*_n(\omega)d\mu
&=&\int_{B}M^*_n(\omega)\omega^{-\frac{1}{p}}\omega^{\frac{1}{p}}d\mu\\
&\leq&(\int_{B}M^*_n(\omega)^{p^\prime}\omega^{-\frac{1}{p-1}}d\mu)^{\frac{1}{p^{\prime}}}
(\int_{B}\omega d\mu)^{\frac{1}{p}}\\
&\leq&C(\int_{B}\omega^{p^\prime}\omega^{-\frac{1}{p-1}}d\mu)^{\frac{1}{p^{\prime}}}
(\int_{B}\omega d\mu)^{\frac{1}{p}}\\
&=&C(\int_{B}\omega d\mu)^{\frac{1}{p^{\prime}}}
(\int_{B}\omega d\mu)^{\frac{1}{p}}\\
&\leq&C\int_{B}\omega d\mu.
\end{eqnarray*}
\eqref{thm:Wi_rev} By Jensen's inequality for the
conditional expectation, we have
\begin{eqnarray*}
\int_{B}M^*_n(\omega)d\mu
&=&\int_{B}\mathbb{E}(M^*_n(\omega)|\mathcal {F}_n)d\mu\\
&\leq&\int_{B}\mathbb{E}(M^*_n(\omega)^{p}|\mathcal {F}_n)^{\frac{1}{p}}d\mu.
\end{eqnarray*}
Because of the conditional version of Doob's inequality(Lemma \ref{doob_con}) and $\omega\in\bigcup\limits_{1<q<\infty} RH_q,$ we obtain that
\begin{eqnarray*}
\int_{B}\mathbb{E}(M^*_n(\omega)^{p}|\mathcal {F}_n)^{\frac{1}{p}}d\mu
&\leq&C\int_{B}\mathbb{E}(\omega^{p}|\mathcal {F}_n)^{\frac{1}{p}}d\mu\\
&\leq&C\int_{B}\omega d\mu.
\end{eqnarray*}
Thus
$$\int_{B}M^*_n(\omega)d\mu\leq C\int_{B}\omega d\mu.$$
\eqref{thm:Wi_log} In view of Lemma \ref{doob_imp}, we have the following estimate
\begin{eqnarray*}
\int_{B}M^*_n(\frac{\omega}{\omega_n})d\mu
&=&\int_0^{+\infty}\mu(B\cap\{M^*_n(\frac{\omega}{\omega_n})>\lambda\}
)d\lambda\\
&\leq&\int_0^{2}\mu(B)d\lambda+\int_2^{+\infty}\mu(B\cap\{M^*_n(\frac{\omega}{\omega_n})>\lambda\}
)d\lambda\\
&\leq&2\mu(B)+\int_2^{+\infty}\frac{2}{\lambda}\int_{B\cap\{ \frac{\omega}{\omega_n}>\frac{\lambda}{2}\}}\frac{\omega}{\omega_n}d\mu d\lambda\\
&=&2\mu(B)+2\int_1^{+\infty}\frac{1}{\lambda}\int_{B\cap\{\frac{\omega}{\omega_n}>\lambda\}}\frac{\omega}{\omega_n}d\mu d\lambda\\
&=&2\mu(B)+2\int_B\frac{\omega}{\omega_n}\log^+\frac{\omega}{\omega_n}d\mu\\
&=&2\mu(B)+2\int_B\mathbb{E}_{\omega}(\log^+\frac{\omega}{\omega_n}|\mathcal {F}_n)d\mu.
\end{eqnarray*}
It follows from \eqref{thm:Wi_log} that
$$\int_{B}M^*_n(\frac{\omega}{\omega_n})d\mu\leq C\mu(B).$$
Thus $\mathbb{E}\big(M^*_n(\frac{\omega}{\omega_n})|\mathcal {F}_n\big)\leq C,$ which implies
$\mathbb{E}\big(M^*_n(\omega)|\mathcal {F}_n\big)\leq C\omega_n.$
\eqref{Thm:Wi_exp} In view of Theorem \ref{thm:exp-s},
there exists $C>1$ such that for every $s\in(0,1)$ we have
$$
\mathbb{E}(\omega|\mathcal {F}_n)\leq C\mathbb{E}(\omega^s|\mathcal {F}_n)^{\frac{1}{s}}.
$$
Then $M^*_n(\omega)\leq C\cdot M^*_n(\omega^s)^{\frac{1}{s}}.$ Let $p=:\frac{1}{s}.$ Using Doob's inequality,
we have
\begin{eqnarray*}
\int_{B}M^*_n(\omega)d\mu
&\leq&C\int_{B}M^*_n(\omega^s)^{\frac{1}{s}}d\mu\\
&\leq&C(p')^p\int_{B}\omega d\mu,
\end{eqnarray*}
where $\frac{1}{p}+\frac{1}{p'}=1.$
Following from $(p')^p\downarrow e$ as $s\downarrow0,$ we obtain
$$\int_{B}M^*_n(\omega)d\mu
\leq C e \int_{B}\omega d\mu.$$
\end{proof}
\bibliographystyle{alpha,amsplain}
\begin{bibdiv}
\begin{biblist}
\bib{MR544802}{incollection}{
author={Bonami, A.},
author={L\'{e}pingle, D.},
title={Fonction maximale et variation quadratique des martingales en
pr\'{e}sence d'un poids},
date={1979},
booktitle={S\'{e}minaire de {P}robabilit\'{e}s, {XIII} ({U}niv.
{S}trasbourg, {S}trasbourg, 1977/78)},
series={Lecture Notes in Math.},
volume={721},
publisher={Springer, Berlin},
pages={294\ndash 306},
review={\MR{544802}},
}
\bib{MR1301765}{article}{
author={Chang, Xiang-Qian},
title={Some {S}awyer type inequalities for martingales},
date={1994},
ISSN={0039-3223},
journal={Studia Math.},
volume={111},
number={2},
pages={187\ndash 194},
url={https://doi.org/10.4064/sm-111-2-187-194},
review={\MR{1301765}},
}
\bib{MR358205}{article}{
author={Coifman, R.~R.},
author={Fefferman, C.},
title={Weighted norm inequalities for maximal functions and singular
integrals},
date={1974},
ISSN={0039-3223},
journal={Studia Math.},
volume={51},
pages={241\ndash 250},
url={https://doi.org/10.4064/sm-51-3-241-250},
review={\MR{358205}},
}
\bib{MR544804}{incollection}{
author={Dol\'{e}ans-Dade, C.},
author={Meyer, P.-A.},
title={In\'{e}galit\'{e}s de normes avec poids},
date={1979},
booktitle={S\'{e}minaire de {P}robabilit\'{e}s, {XIII} ({U}niv.
{S}trasbourg, {S}trasbourg, 1977/78)},
series={Lecture Notes in Math.},
volume={721},
publisher={Springer, Berlin},
pages={313\ndash 331},
review={\MR{544804}},
}
\bib{MR3473651}{article}{
author={Duoandikoetxea, Javier},
author={Mart\'{\i}n-Reyes, Francisco~J.},
author={Ombrosi, Sheldy},
title={On the {$A_\infty$} conditions for general bases},
date={2016},
ISSN={0025-5874},
journal={Math. Z.},
volume={282},
number={3-4},
pages={955\ndash 972},
url={https://doi.org/10.1007/s00209-015-1572-y},
review={\MR{3473651}},
}
\bib{MR481968}{article}{
author={Fujii, Nobuhiko},
title={Weighted bounded mean oscillation and singular integrals},
date={1977/78},
ISSN={0025-5513},
journal={Math. Japon.},
volume={22},
number={5},
pages={529\ndash 534},
review={\MR{481968}},
}
\bib{MR807149}{book}{
author={Garc\'{\i}a-Cuerva, Jos\'{e}},
author={Rubio~de Francia, Jos\'{e}~L.},
title={Weighted norm inequalities and related topics},
series={North-Holland Mathematics Studies},
publisher={North-Holland Publishing Co., Amsterdam},
date={1985},
volume={116},
ISBN={0-444-87804-1},
note={Notas de Matem\'{a}tica [Mathematical Notes], 104},
review={\MR{807149}},
}
\bib{MR402038}{article}{
author={Gehring, F.~W.},
title={The {$L^{p}$}-integrability of the partial derivatives of a
quasiconformal mapping},
date={1973},
ISSN={0001-5962},
journal={Acta Math.},
volume={130},
pages={265\ndash 277},
url={https://doi.org/10.1007/BF02392268},
review={\MR{402038}},
}
\bib{MR3243734}{book}{
author={Grafakos, Loukas},
title={Classical {F}ourier analysis},
edition={Third},
series={Graduate Texts in Mathematics},
publisher={Springer, New York},
date={2014},
volume={249},
ISBN={978-1-4939-1193-6; 978-1-4939-1194-3},
url={https://doi.org/10.1007/978-1-4939-1194-3},
review={\MR{3243734}},
}
\bib{MR727244}{article}{
author={Hru\v{s}\v{c}ev, Sergei~V.},
title={A description of weights satisfying the {$A_{\infty }$} condition
of {M}uckenhoupt},
date={1984},
ISSN={0002-9939},
journal={Proc. Amer. Math. Soc.},
volume={90},
number={2},
pages={253\ndash 257},
url={https://doi.org/10.2307/2045350},
review={\MR{727244}},
}
\bib{MR3092729}{article}{
author={Hyt\"{o}nen, Tuomas},
author={P\'{e}rez, Carlos},
title={Sharp weighted bounds involving {$A_\infty$}},
date={2013},
ISSN={2157-5045},
journal={Anal. PDE},
volume={6},
number={4},
pages={777\ndash 818},
url={https://doi.org/10.2140/apde.2013.6.777},
review={\MR{3092729}},
}
\bib{MR4446233}{article}{
author={Kosz, Dariusz},
title={{$A_\infty$} condition for general bases revisited: complete
classification of definitions},
date={2022},
ISSN={0002-9939},
journal={Proc. Amer. Math. Soc.},
volume={150},
number={9},
pages={3831\ndash 3839},
url={https://doi.org/10.1090/proc/16014},
review={\MR{4446233}},
}
\bib{MR2721744}{article}{
author={Lerner, Andrei~K.},
title={A pointwise estimate for the local sharp maximal function with
applications to singular integrals},
date={2010},
ISSN={0024-6093},
journal={Bull. Lond. Math. Soc.},
volume={42},
number={5},
pages={843\ndash 856},
url={https://doi.org/10.1112/blms/bdq042},
review={\MR{2721744}},
}
\bib{MR1224450}{book}{
author={Long, Rui~Lin},
title={Martingale spaces and inequalities},
publisher={Peking University Press, Beijing; Friedr. Vieweg \& Sohn,
Braunschweig},
date={1993},
ISBN={7-301-02069-4},
url={https://doi.org/10.1007/978-3-322-99266-6},
review={\MR{1224450}},
}
\bib{MR293384}{article}{
author={Muckenhoupt, Benjamin},
title={Weighted norm inequalities for the {H}ardy maximal function},
date={1972},
ISSN={0002-9947},
journal={Trans. Amer. Math. Soc.},
volume={165},
pages={207\ndash 226},
url={https://doi.org/10.2307/1995882},
review={\MR{293384}},
}
\bib{MR350297}{article}{
author={Muckenhoupt, Benjamin},
title={The equivalence of two conditions for weight functions},
date={1973/74},
ISSN={0039-3223},
journal={Studia Math.},
volume={49},
pages={101\ndash 106},
url={https://doi.org/10.4064/sm-49-2-101-106},
review={\MR{350297}},
}
\bib{MR529683}{article}{
author={Str\"{o}mberg, Jan-Olov},
title={Bounded mean oscillation with {O}rlicz norms and duality of
{H}ardy spaces},
date={1979},
ISSN={0022-2518},
journal={Indiana Univ. Math. J.},
volume={28},
number={3},
pages={511\ndash 544},
url={https://doi.org/10.1512/iumj.1979.28.28037},
review={\MR{529683}},
}
\bib{MR766221}{article}{
author={Str\"{o}mberg, Jan-Olov},
author={Wheeden, Richard~L.},
title={Fractional integrals on weighted {$H^p$} and {$L^p$} spaces},
date={1985},
ISSN={0002-9947},
journal={Trans. Amer. Math. Soc.},
volume={287},
number={1},
pages={293\ndash 321},
url={https://doi.org/10.2307/2000412},
review={\MR{766221}},
}
\bib{MR883661}{article}{
author={Wilson, J.~Michael},
title={Weighted inequalities for the dyadic square function without
dyadic {$A_\infty$}},
date={1987},
ISSN={0012-7094},
journal={Duke Math. J.},
volume={55},
number={1},
pages={19\ndash 50},
url={https://doi.org/10.1215/S0012-7094-87-05502-5},
review={\MR{883661}},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section*{1 Introduction}
\subsection*{\textbf{1.1 Background and related work}}
Recently, with the evolution from the fifth and beyond fifth generation mobile communication (5G/beyond 5G) \cite{b1,b0004} to the sixth generation mobile communication (6G)\cite{b0001}, some novel technologies, such as mobile edge computing (MEC) \cite{b2,b3}, satellite communication \cite{b4} and spatial multiplexing \cite{b5}, have been proposed to promote the development of wireless sensor network (WSN) \cite{b6,b0002}, IoT \cite{b7} and other novel applications \cite{b01,b02,b0003}, so as to realize the interconnection of all things.
IoV as a special application of IoT attracts extensive attention in terms of road safety, smart transportation and information service \cite{b8}. However, the demand of computing resources and battery capacity becomes more and more urgent due to a large number of data traffic of these novel applications in IoV \cite{b10}. Hence, mobile edge computing is proposed to mitigate the computing burden of electric vehicles \cite{b12}. Meanwhile, the spectrum efficiency is further improved by MIMO and FD technologies \cite{b13}. Besides, the energy consumption of data processing will increase significantly due to the large number of data traffic. Therefore, SWIPT is supposed to be a meaningful method which provides sufficient energy for electric vehicles to complete their intensive computing tasks.
The last decade, mobile cloud computing (MCC) \cite{b14} as a promising method proposed by researcheres to deal with the limited computing resources of vehicles \cite{b15}. However, the network edge meets uncovered service even though massive resources are provided by cloud to afford computing power for vehicles \cite{b16}. In addition, long latency resulting from the long propagation between vehicles and remote access unit are significant \cite{b18}. In consideration of the problems above, MEC technology is proposed to enhance the computing power at the edge of the network \cite{b19,b20,b21}.
Hence, the author in \cite{b22} investigated the cooperation between MCC and MEC in order to improve the quality of service for vehicles at the network edge. Both \cite{b21} and \cite{b23} aimed at maximizing the system performance and minimizing the time latency assisted by MEC.
Taking account of the energy limitation of electric vehicles, SWIPT \cite{b35} is deployed to extend the battery lifetime of vehicles. Comparing with other practical methods, such as wireless power transfer (WPT) \cite{b002} and energy harvesting (EH) \cite{b32}, SWIPT is used for information decoding and energy harvesting by time-shifting and power splitting, respectively \cite{b2}. Besides, MIMO and FD technologies are able to achieve further improvements of spectral efficiency and low latency \cite{b37,b38}.
Although considerable research has been devoted to improve performance of vehicular network with MEC-assisted technologies, such as a cooperative computing task and redundant data scheme are established to reduce the redundancy of data and allocate the computing tasks through MEC server to abtain the minimal system cost \cite{b40}. The author in \cite{b41} developed a software defined network enabled heterogeneous vehicular network to improve the scalability of the network as well as provide high reliability and low-latency communication. K. Zhang, et al, \cite{b42} proposed a task offloading scheme by jointly consider the data transmission and server selection with Q-learning technology. Based on a multi-objective algorithm, an adaptive strategy of offloading is proposed to optimize the resource allocation in \cite{b43}. The author in \cite{b0005} investigated the secrecy energy efficiency in a distributed massive MIMO Systems. In \cite{b001}, an alternative optimization problem is also raised to maximize the transmission rate in a 5G-based IoT system by jointly optimizing the node powers and allocation coefficient.
Rather less attention has been paid to improve energy efficiency of electric vehicles in IoV considering computing tasks, time latency and spectrum efficiency.
\subsection*{\textbf{1.2 Contributions}}
In this paper, we propose an IoV system assisted by MEC which is used for cross-layer offloading to provide low latency and abundant computation resources. Meanwhile, electric vehicles are able to replenish battery capacity by SWIPT technology and the spectrum efficiency is further improved by MIMO and FD technologies. On the premise of computation requirement, time latency and energy constraint of electric vehicles, we maximize the average energy efficiency of electric vehicles. Since the problem is nonconvex, we decouple it into two subproblems at first. According to the previous research \cite{b44}, a closed-form solution of the central processing unit (CPU) frequency is obtained by solving the first sub-problem. Likewise, in order to tackle with the second sub-problem, three sub-problems are divided from the second sub-problem. Finally, an alternate interior-point iterative scheme is proposed to address these subproblems. Numerical simulations are executed to demonstrate the superior performance of our scheme comparing with the benchmark schemes. Our main contributions in this paper can be summarized as follows:
\begin{itemize}
\item We come up an IoV system by using the MEC, the FD and the SWIPT technologies comprehensively, so as to achieve low-latency, abundant computation resources and lower energy consumption by jointly optimizing the CPU frequency, power transfer of electric vehicles, the uplink rate and the offloading tasks.
\item Under the same computation task, energy consumption and time latency constraints, the proposed scheme can yield high energy efficiency, lower time transmission and is able to tackle more computing tasks comparing with the benchmark schemes.
\item MIMO and FD technologies deployed at anchor node are used to futher improve the spectrum efficiency and the time latency. Meanwhile, we analysis the energy efficiency fairness of the electric vehicles as well.
\end{itemize}
The remainder of this paper is organized as follows. In section 2, we illustrate the system model. A mathematical problem is formulated and solved in section 3. Section 4 provides numerical results and discussions. Finally, section 5 concludes this paper.
\section*{2 Methods}
\subsection*{2.1 System model}
As shown in Fig. \ref{fig1}, we consider an IoV system assisted by MEC, in which a MEC server is deployed at anchor node (AN) and the electric vehicle as a cognitive node connected with AN. Hence, computing tasks will be offloaded partially from electric vehicle nodes (eVNs) to MEC server. Moreover, we assume that the AN works in full-duplex mode which can transmit and receive signals using the same time-frequency resources so as to utilize SWIPT technology to fullfill energy of eVNs as well as receive the offloading computing tasks from the other eVNs at the same time. What's more, each eVN is equipped with a power splitting receiver for ID and EH with the energy conversion ratio $\beta$ ($\beta \in (0,1)$).
In Fig. \ref{fig2}, each vehicle node $VN_i(i \in \{1,...,K\})$ offloads partial computing tasks $m_i(0 \leq m_i \leq M_i)$ to the AN within the time duration $\tau_i^{\text u}$, where $M_i$ denotes the total computation tasks of $VN_i$ and $K$ is the number of vehicle nodes. Simultaneously, AN transmits the computation tasks $m_j^{\text r}(j \in \{1,...,K\})$ executed by MEC server to $VN_j$ within the time duration $\tau_j^{\text d}$, where $m_j^{\text r} = \alpha m_j (0 \leq \alpha \leq 1)$, and $\alpha$ is the ratio of the completed computation tasks from the MEC server. Moreover, the computing time consumption of the MEC server is ignored resulting from the powerful computing ability of it comparing with the vehicle nodes \cite{b45}. Furthermore, the local computing tasks $m_i^{\text {lo}}$, which equals $M_i-m_i$, is completed during the time duration $\tau_i^{\text {lo}}$. Besides, the uplink time consumption of each VN is assumed to be more than the local computing time consumption of that, which is expressed as $0 \leq \tau_i^{\text {lo}} \leq \tau_i$. More details are given as follows.
Notations: The channel fading is modeled as Rayleigh distribution in this paper. $\bm{h}_i^{\text u} \in \mathbb{C c}^{N\times1}$ and $\bm{h}_j^{\text d} \in \mathbb{C c}^{N\times1}$ satisfy uniform distribution$\sim U(0.5,1)$ which present the uplink and downlink channel, respectively, where $N$ is the antenna numbers of AN, and the superscript $\text u$ and $\text d$ refer to uplink and downlink transmission, respectively. Meanwhile, $\bm{H}_{\text {an}}$ represents the self-interference channel of AN. $d_i $ and $\bm{d}_j \in \mathbb{C c}^{N\times1}$ represent the transmitted signal from vehicle node and AN respectively, where $|\bm{d}_j|=1$. Moreover, $p_i$ and $p_j$ denote the transmitting power of VN and AN, respectively. Besides, the additive white Gaussian noise (AWGN) of AN is denoted as $\bm{n}_{\text an}$ with covariance matrix $\sigma_{\text an}^2 \bm{I}_M$. Similarly, $n_j$ and $n_{\text ps}$ are the AWGN of $VN_j$ and PS receiver with variance $\sigma_j^2$ and $\sigma_{\text ps}^2$, respectively. Finally, the bandwidth of the system is expressed as $B$.
\subsubsection*{2.1.1 Computation task processing}
In uplink transmission, partial computing tasks are uploaded from vehicle nodes to the AN in turns. Specifically, $VN_i$ transmits $d_i$ with transmitting power $p_i$ to the AN within the time slot $\tau_i^{\text u}$. Simultaneously, the AN transmits $\bm{d}_j$ with transmitting power $p_j$ to VN. Therefore, we can obtain the received signal at AN,
\begin{equation}
\begin{aligned}
\bm{x}_i^{u} = \sqrt {{p_i}} {\bm{h}}_i^{\text u}{d_i} + {{\bm{H}}_{\text an}}(\sqrt {{ p_j}} {\bm{d}_j}) + \bm{n}_{\text {an}}
\end{aligned}\label{eq_xi}
\end{equation}
where the first iterm represents the desired signal from $VN_i$, while the self-interference at the AN is expressed as the second iterm, and the rest one is the AWGN of AN. Meanwhile, $p_i$ is the transmitting power of vehicle node which satisfies $p_{\text {min}} \leq p_i \leq p_{\text {max}}$, where $p_{\text {min}}$ and $p_{\text {max}}$ are the minimal and maximal transmitting power of each vehicle node.
According to (\ref{eq_xi}), the received signal of VN is expressed as equation (\ref{eq_Pi}), while equation (\ref{eq_SINRi}) represents the signal to interference and noise ratio (SINR)
\begin{equation}
\begin{aligned}
P_i&=\text{Tr}\{ \bm{x}_i(\bm{x}_i)^H \}\\
&=p_i \text{Tr}\{ \bm{h}_i^{\text u}(\bm{h}_i^{\text u})^H\}+p_j \text{Tr}\{ \bm{H}_{\text {an}}(\bm{H}_{\text {an}})^H\}+\delta_{\text {an}}^2
\end{aligned}\label{eq_Pi}
\end{equation}
\begin{equation}
\begin{aligned}
\gamma_i =\frac{p_i \text{Tr}\{ \bm{h}_i^{\text u}(\bm{h}_i^{\text u})^H\}}{p_j \text{Tr}\{ \bm{H}_{\text {an}}(\bm{H}_{\text {an}})^H\}+\delta_{\text {an}}^2}
\end{aligned}\label{eq_SINRi}
\end{equation}
where $(.)^H$, $\text{Tr}\{.\}$ represent the conjugate and trace of matrix, respectively. According to the equation above, we can obtain the uplink rate which is written as
\begin{equation}
\begin{aligned}
r_i ={\text {B$log_2$}}(1+\gamma_i)
\end{aligned}\label{eq_ri}
\end{equation}
As we know, if the self-interference can be eliminated completely by self-interference cancelation (SIC), the maximal SINR can be expressed as
\begin{equation}
\begin{aligned}
\gamma_i^{\text{max}} =\frac{p_i \text{Tr}\{ \bm{h}_i^{\text u}(\bm{h}_i^{\text u})^H\}}{\delta_{\text {an}}^2}
\end{aligned}\label{eq_SIRimax}
\end{equation}
Hence, we can obtain the following ideal uplink rate
\begin{equation}
\begin{aligned}
r_i^{\text{max}} ={\text B} log_2 (1+\gamma_i^{\text{max}})
\end{aligned}\label{eq_rimax}
\end{equation}
However, limited by SIC device, self interference cannot be completely eliminated. Therefore, the uplink rate should not exceed the ideal maximal uplink rate, i.e.,
\begin{equation}
\begin{aligned}
r_i \leq r_i^{\text{max}}
\end{aligned}\label{eq_ricom}
\end{equation}
Based on the analysis above, each time slot $\tau_i^{\text u}$ of uplink transmission and the energy consumption of $VN_i$ can be formulated as follow,
\begin{equation}
\begin{aligned}
\tau_i =\frac{m_i}{r_i}
\end{aligned}\label{eq_ti}
\end{equation}
\begin{equation}
\begin{aligned}
e^{\text{vn}}_i=p_i\frac{ m_i}{r_i}
\end{aligned}\label{eq_evni}
\end{equation}
In consideration of the theory in \cite{b45}, the local computing time duration and the consumption of energy in each vehicle node is written as
\begin{equation}
\begin{aligned}
\tau_i^{\text {lo}}=\sum_{i=1}^{C(M_i-m_i)} \frac{1}{f_i^{\text t}}
\end{aligned}\label{eq_tlo}
\end{equation}
\begin{equation}
\begin{aligned}
e_i^{\text {lo}}=\sum_{n=1}^{C(M_i-m_i)} {\kappa (f_i^{\text t})^2}
\end{aligned}\label{eq_eloi}
\end{equation}
where the required CPU cycle is represented as $C$ which for computing one bit of data. And, for ${\text t}$-th CPU cycles, the CPU frequency of $VN_i$ is $f_i^{\text t}$. Here, $f_i^{\text t}$ should no more than the maximal CPU frequency $f_i^{\text {max}}$. And $\kappa$ denotes the efficient capacitance coefficient in view of the vehicles' chip architecture.
\subsubsection*{2.1.2 Energy harvesting}
Similarly, in the downlink transmission, $VN_j$ receives signal from the AN including the desired RF signal and the AWGN at $VN_j$, which are expressed as
\begin{equation}
\begin{aligned}
x_j=\sqrt{p_j}(\bm{h}_j^{\text d})^T \bm{d}_j+n_j
\end{aligned}\label{eq_xj}
\end{equation}
In (\ref{eq_xj}), $(.)^T$ is the transposition operation. Here, the co-channel interference between vehicle nodes are not considered, since the VNs upload the computing tasks to the anchor node in turns. Likewise, we obtain the signal power received at $VN_i$ in (\ref{eq_Pj}) and equation (\ref{eq_SNRj}) is the downlink signal to noise ratio (SNR).
\begin{equation}
\begin{aligned}
P_j={ p_j} \text{Tr}\{(\bm{h}_j^{\text d})^T (\bm{h}_j^{\text d})^{TH}\} +\delta_j^2
\end{aligned}\label{eq_Pj}
\end{equation}
\begin{equation}
\begin{aligned}
\gamma_j =\frac{\beta p_j \text{Tr}\{ (\bm{h}_j^{\text d})^T(\bm{h}_j^{\text d})^{TH}\}}{\delta_j^2}
\end{aligned}\label{eq_SNRj}
\end{equation}
Hence, the downlink rate and the time slot can be derived as
\begin{equation}
\begin{aligned}
r_j =Blog(1+\gamma_j)
\end{aligned}\label{eq_rj}
\end{equation}
\begin{equation}
\begin{aligned}
\tau_j =\frac{\alpha m_j}{r_j}
\end{aligned}\label{eq_tj}
\end{equation}
Based on the coefficient $\beta$ of the PS receiver for EH, the energy of EH at $VN_i$ is shown as
\begin{equation}
\begin{aligned}
e_j^{\text{eh}}= \beta {p_j} \text{Tr}\{(\bm{h}_j^{\text d})^T (\bm{h}_j^{\text d})^{TH}\} \tau_j^{\text d}
\end{aligned}\label{eq_Ehi}
\end{equation}
\subsubsection*{2.1.3 Problem formulation}
The computing time of the MEC server is ignored, since the computing ability of it is more powerful than that of vehicle's \cite{b45}.
And, we assume that ${\text T}$ is the overall duration. Hence, the relationship between the ${\text T}$ and the total transmission time in uplink and downlink of vehicles as shown in formula (\ref{eq_sumt_i}) and (\ref{eq_sumt_j}), respectively.
\begin{equation}
\begin{aligned}
0 \leq \sum_{i=1}^K \tau_i^{\text u} \leq {\text T}
\end{aligned}\label{eq_sumt_i}
\end{equation}
\begin{equation}
\begin{aligned}
0 \leq \sum_{j=1}^K \tau_j^{\text d} \leq {\text T}
\end{aligned}\label{eq_sumt_j}
\end{equation}
Moreover, since vehicle node is not able to transmit and receive signal simultaneously, there should be no overlap between uplink and downlink transmission time of a same vehicle node, i.e.,
\begin{equation}
\begin{aligned}
\sum_{i=1}^K \tau_i^{\text u}+ \tau_K^{\text d} \leq {\text T}
\end{aligned}\label{eq_t_i_t_j_1}
\end{equation}
\begin{equation}
\begin{aligned}
\sum_{i=1}^{K_{tmp}} \tau_i^{\text u} \leq {\text T}-\sum_{j=K_{tmp}}^{K} \tau_j^{\text d}
\end{aligned}\label{eq_t_i_t_j_2}
\end{equation}
where $K_{tmp} \in \{1,...,K\}$.
Meanwhile, the downlink transmission slot is less than the uplink transmission slot, so as to gurantee the computing tasks offloading from the next vehicle node can be processed at the AN without time latency and queuing. i.e.,
\begin{equation}
\begin{aligned}
\tau_j^{\text d} \leq \tau_i^{\text u}, j = i-1
\end{aligned}\label{eq_tcom}
\end{equation}
Besides, we assume that each vehicle node has enough energy to complete the computing due to the sufficient energy supply from the AN. Hence, the energy harvested from the AN should no more than the total transmitting power from the AN. Hence, the energy consumption of each vehicle should satisfy,
\begin{equation}
\begin{aligned}
e_i^{\text{vn}} + e_i^{\text {lo}} \leq e_i^{\text{eh}}
\end{aligned}\label{eq_ecom}
\end{equation}
\begin{equation}
\begin{aligned}
e_i^{\text{eh}} \leq e_i^{\text{an}}
\end{aligned}\label{eq_ecom}
\end{equation}
where $e_i^{\text{an}}=p_j \tau_j^{\text d}$. According to the derivations above, we formulate the following problem ($P1$) which is aim to maximizing the average enery efficiency of the total vehicle nodes.
\begin{equation}\label{eq:P1}
\begin{array}{cl}
(P1) & \mathop {\max}\limits_{{\bm{f}^{\text t}, \bm{p}, \bm{m},\bm{r}}} \frac{1}{K}\{\sum_{i=1}^K \frac{r_i}{e_i^{\text{vn}}+e_i^{\text {lo}}}\}\\
& s.t.
\left \{
\begin{array}{l}
C1: e_i^{\text{vn}} + e_i^{\text {lo}} \leq e_i^{\text{eh}}\\
C2: e_i^{\text{eh}} \leq e_i^{\text{an}}\\
C3: 0 \leq f_i^{\text t} \leq f_i^{\text{max}}\\
C4: p_{\text {min}} \leq p_i \leq p_{\text {max}}\\
C5: 0 \leq m_i \leq M_i\\
C6: 0 \leq r_i \leq r_i^{\text {max}}\\
C7: 0 \leq \tau_i^{\text {lo}} \leq \tau_i^{\text u}\\
C8: \tau_j^{\text d} \leq \tau_i^{\text u}, j = i-1\\
C9: 0 \leq \sum_{i=1}^K \tau_i^{\text u} \leq {\text T}\\
C10: 0 \leq \sum_{j=1}^K \tau_j^{\text d} \leq {\text T}\\
C11: \sum_{i=1}^{K_{tmp}} \tau_i^{\text u} \leq {\text T}-\sum_{j=K_{tmp}}^{K} \tau_j^{\text d}
\end{array} \right.
\end{array}
\end{equation}
where $\bm{f}^{\text t}=[f_1^{\text t},...,f_K^{\text t}]^T$, $\bm{p}=[p_1,...,p_K]^T$, $\bm{m}=[m_1,...,m_K]^T$, $\bm{r}=[r_1,...,r_K]^T$ represent the CPU frequency, the uplink transmitting power, the offloading computation tasks and the uplink rate of all vehicle nodes, respectively. Moreover, the $f_i^{\text{max}}$ is the maximum CPU frequency.
\subsection*{2.2 Problem solution}
Since the cost function and the constraints in problem ($P1$) are nonlinear and nonconvex, two subproblems are derived from the problem ($P1$) first.
For the first subproblem, a closed-form expression of $f^{\text t}$ with fixed variables $\bm{p}, \bm{m}, \bm{r}$ is obtained.
Combining the first subproblem with ($P1$), we derive the second subproblem. For the second subproblem, we utilize an alternate interior-point iterative scheme after decompose it into three subproblems. Finally, the suboptimal varibles $<\bm{f}^{\text t}, \bm{p}, \bm{m}, \bm{r}>$ are abtained.
\subsubsection*{2.2.1 Local computing optimization}
Inspired by \cite{b45}, the optimal CPU frequency should satisfy
\begin{equation}
\begin{aligned}
f_i^1=f_i^2=...=f_i^{C(M_i-m_i)}=\frac{C(M_i-m_i)}{\tau_i^{\text {lo}}}=\overline f_i
\end{aligned}\label{eq_fn}
\end{equation}
where the average CPU frequency is denoted as $\overline f_i$.
Based on the derivation above, the original problem ($P1$) is converted into the problem ($P2$) as following.
\begin{equation}\label{eq:P2}
\begin{array}{cl}
(P2) & \mathop {\min}\limits_{{\bm{\overline f}}} \frac{1}{K}\{\sum_{i=1}^K e_i^{\text {lo}}\}\\
& s.t.
\left \{
\begin{array}{l}
C1: e_i^{\text{vn}} + e_i^{\text {lo}} \leq e_i^{\text{eh}}\\
C2: 0 \leq \tau_i^{\text {lo}} \leq \tau_i^{\text u}\\
\end{array} \right.
\end{array}
\end{equation}
To futher simplify the problem ($P2$), we obtain the following problem ($P3$)
\begin{equation}\label{eq:P3}
\begin{array}{cl}
(P3) & \mathop {\min}\limits_{{\bm{\overline f}}} \frac{1}{K}\{\sum_{i=1}^K {\kappa C(M_i-m_i)(\overline{f}_i)^2}\}\\
& s.t.
\left \{
\begin{array}{l}
C1: e_i^{\text{vn}} + {\kappa C(M_i-m_i)(\overline{f}_i)^2} \leq e_i^{\text{eh}}\\
C2: 0 \leq \frac{C(M_i-m_i)}{\overline{f}_i} \leq \tau_i^{\text u}\\
\end{array} \right.
\end{array}
\end{equation}
where C2 indicates that $\frac{C(M_i-m_i)}{\tau_i^{\text u}} $ is the lower limit of $\overline{f}_i$ as well as the optimal CPU frequency $f_i^{\text{opt}}$.
\subsubsection*{2.2.2 Communication optimization}
We obtan the rewritted problem ($P4$) by substituting the optimal CPU frequency $f_i^{\text{opt}}$ into ($P1$).
\begin{equation}\label{eq:P4}
\begin{array}{cl}
(P4) & \mathop {\max}\limits_{{\bm{p}, \bm{m},\bm{r}}} \frac{1}{K}\{\sum_{i=1}^K \frac{r_i}{e_i^{\text{vn}}+e_i^{\text {lo}}}\}\\
& s.t.
\left \{
\begin{array}{l}
C1: e_i^{\text{vn}} + e_i^{\text {lo}} \leq e_i^{\text{eh}}\\
C2: e_i^{\text{eh}} \leq e_i^{\text{an}}\\
C3: p_{\text {min}} \leq p_i \leq p_{\text {max}}\\
C4: 0 \leq m_i \leq M_i\\
C5: 0 \leq r_i \leq r_i^{\text{max}}\\
C6: \tau_j^{\text d} \leq \tau_i^{\text u}, j = i-1\\
C7: 0 \leq \sum_{i=1}^K \tau_i^{\text u} \leq {\text T}\\
C8: 0 \leq \sum_{j=1}^K \tau_j^{\text d} \leq {\text T}\\
C9: \sum_{i=1}^{K_{tmp}} \tau_i^{\text u} \leq {\text T}-\sum_{j=K_{tmp}}^{K} \tau_j^{\text d}
\end{array} \right.
\end{array}
\end{equation}
Since the problem ($P4$) is still a NP-hard problem, we continue to decompose it into three subproblems, which are denoted as ($P4.1$) $\sim$ ($P4.3$), respectively.
\begin{equation}\label{eq:P4.1}
\begin{array}{cl}
(P4.1) & \mathop {\max}\limits_{{\bm{p}}} \frac{1}{K}\{\sum_{i=1}^K \frac{r_i}{e_i^{\text{vn}}+e_i^{\text {lo}}}\}\\
& s.t.
\left \{
\begin{array}{l}
C1: e_i^{\text{vn}} + e_i^{\text {lo}} \leq e_i^{\text{eh}}\\
C2: e_i^{\text{eh}} \leq e_i^{\text{an}}\\
C3: p_{\text {min}} \leq p_i \leq p_{\text {max}}\\
C4: \tau_j^{\text d} \leq \tau_i^{\text u}, j = i-1\\
C5: 0 \leq \sum_{i=1}^K \tau_i^{\text u} \leq {\text T}\\
C6: 0 \leq \sum_{j=1}^K \tau_j^{\text d} \leq {\text T}\\
C7: \sum_{i=1}^{K_{tmp}} \tau_i^{\text u} \leq {\text T}-\sum_{j=K_{tmp}}^{K} \tau_j^{\text d}
\end{array} \right.
\end{array}
\end{equation}
The problem ($P4.1$) is only related to $\bm{p}$, which can be tackled by classical interior-point algorithm.
\begin{equation}\label{eq:P4.2}
\begin{array}{cl}
(P4.2) & \mathop {\max}\limits_{{\bm{m}}} \frac{1}{K}\{\sum_{i=1}^K \frac{r_i}{e_i^{\text{vn}}+e_i^{\text {lo}}}\}\\
& s.t.
\left \{
\begin{array}{l}
C1: e_i^{\text{vn}} + e_i^{\text {lo}} \leq e_i^{\text{eh}}\\
C2: e_i^{\text{eh}} \leq e_i^{\text{an}}\\
C3: 0 \leq m_i \leq M_i\\
C4: \tau_j^{\text d} \leq \tau_i^{\text u}, j = i-1\\
C5: 0 \leq \sum_{i=1}^K \tau_i^{\text u} \leq {\text T}\\
C6: 0 \leq \sum_{j=1}^K \tau_j^{\text d} \leq {\text T}\\
C7: \sum_{i=1}^{K_{tmp}} \tau_i^{\text u} \leq {\text T}-\sum_{j=K_{tmp}}^{K} \tau_j^{\text d}
\end{array} \right.
\end{array}
\end{equation}
As same as the problem ($P4.1$), interior-point algorithm is used to address the problem ($P4.2$) to achieve the optimal $\bm{m}$.
\begin{equation}\label{eq:P4.3}
\begin{array}{cl}
(P4.3) & \mathop {\max}\limits_{{\bm{r}}} \frac{1}{K}\{\sum_{i=1}^K \frac{r_i}{e_i^{\text{vn}}+e_i^{\text {lo}}}\}\\
& s.t.
\left \{
\begin{array}{l}
C1: e_i^{\text{vn}} + e_i^{\text {lo}} \leq e_i^{\text{eh}}\\
C2: e_i^{\text{eh}} \leq e_i^{\text{an}}\\
C3: 0 \leq r_i \leq r_i^{\text{max}}\\
C4: \tau_j^{\text d} \leq \tau_i^{\text u}, j = i-1\\
C5: 0 \leq \sum_{i=1}^K \tau_i^{\text u} \leq {\text T}\\
C6: 0 \leq \sum_{j=1}^K \tau_j^{\text d} \leq {\text T}\\
C7: \sum_{i=1}^{K_{tmp}} \tau_i^{\text u} \leq {\text T}-\sum_{j=K_{tmp}}^{K} \tau_j^{\text d}
\end{array} \right.
\end{array}
\end{equation}
Similarly, the optimal variable $\bm{r}$ is obtained by solving the problem ($P4.3$) using the interior-point algorithm.
Each subproblem of ($P4.1$) $\sim$ ($P4.3$) is regarded as an approximate convex problem of the corresponding variable, while the other variable are fixed. In order to solve the problem ($P4$), each subproblem is addressed by classical interior-point algorithm. After that, the whole problem is solved by alternate iterative sheme, which is denoted as \textbf{Algorithm1}.
\begin{algorithm}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\caption{Alternate interior-point iterative scheme (AIIS)}
\label{}
\begin{algorithmic}[1]
\REQUIRE$\bm{h}_i^{\text u}, \bm{h}_j^{\text d}, \bm{H}_{\text {an}}, \bm{p}_{(0)}, \bm{m}_{(0)}, \bm{r}_{(0)}, \bm{d}_j, d_i, K, N, B, {\text T}, C, \kappa, \alpha, \beta, \delta_j^2, \delta_{\text ps}^2, \delta_{\text {an}}^2, p_{\text {min}}, p_{\text {max}}, M_i, p_{j}$,
where $\{i,j\}\in \{1,...,K\}$, $\bm{p}_{(0)}$, $\bm{m}_{(0)}$ and $\bm{r}_{(0)}$ represent the initial offloading computation tasks, the uplink transmission power and rate of the vehicle nodes, respectively.
\STATE According to ($P3$), $f_i^{\text{opt}} = \frac{C(M_i-m_i)}{\tau_i^{\text u}}$, $\tau_i^{\text {lo}}=\tau_i^{\text u}$, and $e_i^{\text {lo}} =\frac{ \kappa C^3 (M_i-m_i)^3}{(\tau_i^{\text u})^2}$.
\STATE Turn ($P3$) into ($P4$) through substituting $f_i^{opt}=\frac{C(M_i-m_i)}{\tau_i^{\text u}}$, $\tau_i^{\text {lo}}=\tau_i^{\text u}$ and $e_i^{\text {lo}} =\frac{ \kappa C^3 (M_i-m_i)^3}{(\tau_i^{\text u})^2}$ into ($P3$).
\STATE Let $\bm{p}_{ini}=\bm{p}_{(0)}, \bm{m}_{ini}=\bm{m}_{(0)}, \bm{r}_{ini}=\bm{r}_{(0)}$, and iteration number $k=1$ be the initial value.
\REPEAT
\STATE Deal with ($P4.1$) with fixed $\bm{m}_{(k-1)}, \bm{r}_{(k-1)}$, $\bm{p}_{ini}=\bm{p}_{(k-1)}$, to abtain $\bm{p}_{(k)}^{(\text{opt})}$;
\STATE Deal with ($P4.2$) with fixed $\bm{p}_{(k-1)}, \bm{r}_{(k-1)}$, $\bm{m}_{ini}=\bm{m}_{(k-1)}$, to abtain $\bm{m}_{(k)}^{(\text{opt})}$;
\STATE Deal with ($P4.3$) with fixed $\bm{p}_{(k-1)}, \bm{m}_{(k-1)}$, $\bm{r}_{ini}=\bm{r}_{(k-1)}$, to abtain $\bm{r}_{(k)}^{(\text{opt})}$;
\STATE Let $k=k+1$;
\UNTIL the objective function in ($P4$) converges.
\ENSURE $ \bm{f}^{(\text{opt})}, \bm{p}^{(\text{opt})}, \bm{m}^{(\text{opt})}, \bm{r}^{(\text{opt})}$
\end{algorithmic}
\end{algorithm}
\section*{4 Results and discussion}
Based on numerical simulations, we dicuss the performance of the proposed scheme in this papaer comparing with the two benchmark schems. Three offloading strategies are denoted as below and the simulation parameters are listed in \textbf{Table 1}.
\begin{itemize}
\item FVS: all variables are fixed.
\item FOS: computing tasks completely offload to the AN which indicates the local computing task is zero.
\item AIIS: computing tasks are offloaded arbitrarily based on the maximum average energy efficiency of total vehicle nodes.
\end{itemize}
In Fig. \ref{fig3}, we find that the average energy efficiency of the total vehicles decrease as computing tasks increase. However, the proposed AIIS scheme still has the maximum average energy efficiency comparing with the benchmark schemes.
Fig. \ref{fig4} investigates the influence of antenna numbers on system performance. Here, we assume that the number of vehicles equals 10 and the antenna numbers are \{6,8,10\}. As shown in figure, with the number of antennas increasing, the average energy efficiency for each schemes is improved. However, the average energy efficiency of the proposed scheme is higher than that of the comparing schemes with the same antenna numbers.
Fig. \ref{fig5} studies the system performance with different number of vehicle nodes from 1 to 10, where we can find that the number of vehicles make no difference to the average energy efficiency for each schemes due to the total energy efficiency of all vehicles are averaged. Nerverthless, the proposed scheme is superior to the comparing scheme under the same number of vehicles.
Fig. \ref{fig6} studies the effect of the downlink transmitting power from the AN. Generally speaking, additional downlink transmitting power means that vehicle node may have more energy to complete the computation tasks. However, since each vehicle node tends to maximize its uplink rate as well as minimize its uplink energy consumption so as to maximize the average energy efficiency. Hence, the redundant downlink transmitting power seems to have no effect on the average energy efficiency.
In other words, we may achieve lower energy consumption with higher energy efficiency. Besides, the system performance of the proposed scheme is better than that of the other two schemes as well.
Under the same coditions, Fig. \ref{fig7} shows that the transmission time of three schemes increase with the number of computation tasks increasing. However, the proposed scheme has the lower transmission time compared to the other schemes.
To futher study the fairness of each vehicles, we obtain the Fig. \ref{fig8} and Fig. \ref{fig9}. For the former, variance of energy efficiency of all vehicles are given to prove that the proposed scheme AIIS has the best energy efficiency fairness between the same vehicle numbers compared to the other schemes, which indicates that the qulity of service of different vehicles can be guranteed.
For the latter, the proposed scheme has the lower variance of transmission time consumptions between all vehicles with the same computation tasks comparing with the benchmark schemes, which indicates that the time-latency fairness between the different vehicles of AIIS is superior to that of the comparing schemes.
\section*{5 Conclusion}
In this paper, we propose an IoV system assisted by MEC server which is deployed at anchor node. Electric vehicle as a cognitive node uploads their intensive computing tasks to the AN as well as harvests energy from the RF signal transmitted by the AN with SWIPT technology, so as to alleviate the heavy computing tasks, reduce the time latency and compensate the limited battery capacity of vehicle node. Besides, the spectral efficiency is further improved by MIMO and FD technologies. Finally, an alternate interior-point iterative scheme (AIIS) is proposed to deal with a non-convex problem which is aim to maximize the average energy efficiency of vehicles by jointly optimize the computing and communication resources. Simulation results verify that the proposed AIIS scheme outperforms the other two comparison schemes. Furthermore, the service of future IoV system may benefit from the proposed scheme.
|
\section{Introduction}
Motivated by advances in the learning community~\cite{ref:amos2017input,ref:kolter2019learning}, we discuss in this note the ramifications of assuming (control) Lyapunov functions---as pioneered by Artstein~\cite{ref:artstein1983stabilization} and Sontag~\cite{ref:sontag1989universal}---to be convex. A secondary contribution of this note is to address---in arguably the most simple setting---continuation questions as posed by Conley~\cite{ref:conley1978isolated} and later Kvalheim~\cite{ref:kvalheim2022obstructions}. Overall, this note is in the spirit of the work by Zabczyk~\cite{ref:zabczyk1989}, Reineck~\cite{ref:reineck1991continuation},~Sepulchre \& Aeyels~\cite{ref:sepulchre1996homogeneous}, Coron~\cite[Chapter~11]{ref:coron2007control}, Byrnes~\cite{ref:byrnes2008brockett} and Cieliebak \& Eliashberg~\cite[Chapter~9]{ref:cieliebak2012stein}.
We start by introducing Lyapunov functions for the dynamical control systems at hand. Then, in Section~\ref{sec:topo} we highlight topological properties of levelsets of Lyapunov functions. These observations are the motivation for Section~\ref{sec:cvx} where we infer continuation results by considering several notions of convexity. This note is concluded in Section~\ref{sec:conclusion}.
\paragraph*{Notation} Let $r\in \mathbb{N}\cup \{\infty\}$, then, $C^r(U;V)$ denotes the set of $C^r$-smooth functions from some set $U$ to $V$. The inner product on $\mathbb{R}^n$ is denoted by $\langle \cdot , \cdot \rangle$ and $\mathbb{S}^{n-1}=\{x\in \mathbb{R}^n:\|x\|_2=1\}$ is the embedded unit sphere in $\mathbb{R}^n$. The Lie derivative of a smooth function $h$ over some open set $U\subseteq \mathbb{R}^n$ with respect to a smooth vector field $X$ over the same set $U$ is denoted by $L_X h$ and is defined pointwise by $L_Xh(p):=\langle \nabla h(p), X(p)\rangle$ for any $p\in U$~\cite[Proposition~12.32]{Lee2}. By $\mathrm{cl}(W)$ we denote the closure of a set $W$ and by $\mathrm{int}(W)$ we denote its interior. The identity map $x\mapsto x$ on a space $\mathbb{R}^n$ is denoted by $\mathrm{id}_{\mathbb{R}^n}$ and tangent spaces of appropriately defined sets $M$ are denoted by $T_pM$, for $p\in M$, with $TM$ denoting the corresponding tangent bundle~\cite[p.~65]{Lee2}.
\subsection{Dynamical control systems}
We study \textit{\textbf{dynamical systems}} over $\mathbb{R}^n$ of the form
\begin{equation}
\label{equ:dyn:sys}
\frac{\mathrm{d}}{\mathrm{d}t}x(t)=F(x(t)) : \begin{cases}
F:\mathbb{R}^n\to T\mathbb{R}^n\\
\pi\circ F = \mathrm{id}_{\mathbb{R}^n},
\end{cases}
\end{equation}
where $F$ is $C^r$-smooth with $r\geq 0$, $\pi:T\mathbb{R}^n\to \mathbb{R}^n$ defined by $(x,v)\mapsto \pi(x,v)=x$ is the canonical projection and for any $x\in \mathbb{R}^n$ we have with some abuse of notation $F(x)\in T_x\mathbb{R}^n$. Evidently, $T\mathbb{R}^n\simeq \mathbb{R}^n\times \mathbb{R}^n$, but~\eqref{equ:dyn:sys} is useful to keep in mind when comparing objects to assess if generalizations beyond $\mathbb{R}^n$ are possible. When $r\geq 1$, \textit{integral curves} of~\eqref{equ:dyn:sys} are differentiable curves $t\mapsto \xi(t)\in \mathbb{R}^n$ such that $\dot{\xi}(t)=F(\xi(t))$ for all $t\in \mathrm{dom}(\xi)$, which is non-empty by the assumption on $r$. We will not go into further regularity conditions and always assume for simplicity that $r=0$ and that the vector field is \textit{complete},~\textit{i.e.}, a global flow (see below) is induced, such that we are allowed to make global statements\footnote{We remark that \textit{completeness} is the important property here as we will appeal to a global flow, \textit{smoothness} of $F$ (going beyond $C^0$), on the other hand, is rarely exploited. The only reason to potentially keep smoothness is that one can naturally relax completeness and make some local statements. Without completeness, global statements can break down, consider $\dot{x}=x^2$. However, as the emphasis of this note is on \textit{global asymptotic stability}, examples of that form are somewhat obsolete.}, for further information we point the reader to~\cite{ref:sontag2013mathematical,ref:hale2009ordinary}.
Going beyond \textit{descriptions}, when aiming to \textit{prescribe} the dynamics of a system we consider (time-invariant) \textit{\textbf{dynamical control systems}} over $\mathbb{R}^n\times \mathbb{R}^m$ of the form
\begin{equation}
\label{equ:control:sys}
\frac{\mathrm{d}}{\mathrm{d}t}x(t) = f(x(t),u) : f(x,u)\in T_x\mathbb{R}^n\simeq \mathbb{R}^n\,\forall (x,u)\in \mathbb{R}^n\times \mathbb{R}^m ,
\end{equation}
where $x$ and $u$ denote the state and input, respectively.
Again, with some abuse of notation, we will assume that $f\in C^{0}(\mathbb{R}^n\times \mathbb{R}^m;\mathbb{R}^n)$, but again omit integrability discussions.
Input functions are of the form $t\mapsto \mu(t)\in \mathbb{R}^m$,~\textit{e.g.}, a state feedback is of the form $t\mapsto\mu(x(t))$.
Note, we use $\mu$ instead of $u$ to differentiate between the function and the point. A subclass of~\eqref{equ:control:sys} of interest are the so-called \textit{control affine} systems of the form
\begin{equation}
\label{equ:control:sys:affine}
\frac{\mathrm{d}}{\mathrm{d}t}x(t) = f(x(t)) + \textstyle\sum^m_{i=1}g_i(x(t))u_i,
\end{equation}
where $u_i$ is the $i^{\mathrm{th}}$ element of $u\in \mathbb{R}^m$ and again $f,g_i\in C^{0}(\mathbb{R}^n;\mathbb{R}^n)$ for $i=1,\dots,m$~\cite{nonlin}. Based on the control system at hand, one might say more about the space of allowable inputs $t\mapsto \mu(t)$,~\textit{e.g.}, one might consider \textit{absolutely integrable} ($L^1_{\mathrm{loc}}$) or \textit{essentially bounded} ($L^{\infty}_{\mathrm{loc}}$) function spaces~\cite[Appendix~C]{ref:sontag2013mathematical}.
\subsection{Stability}
Let $F$ parametrize a dynamical system of the form~\eqref{equ:dyn:sys}. By our standing completeness and smoothness assumptions, $F$ will give rise to a continuous \textit{flow}\footnote{Flows satisfy: (1) the \textit{identity} $\varphi^0=\mathrm{id}_{\mathbb{R}^n}$; and (2) \textit{group} property $\varphi^{s+t}=\varphi^s\circ\varphi^t$ $\forall s,t\in \mathbb{R}$.} $\varphi:\mathbb{R}\times \mathbb{R}^n\to \mathbb{R}^n$, with its evaluation denoted by $\varphi^t(x_0):=\varphi(t,x_0)$, which is understood to describe a solution to~\eqref{equ:dyn:sys} at time $t$ starting at time $0$ from $x_0$. A point $x^{\star}\in \mathbb{R}^n$ is an \textit{equilibrium point} of $F$ when $F(x^{\star})=0$, \textit{w.l.o.g.} we set $x^{\star}=0$. Then, $0$ is said to be \textit{\textbf{globally asymptotically stable}} (GAS) (with respect to $F$) if
\begin{enumerate}[(s-i)]
\item $0$ is \textit{Lyapunov stable}, that is, for any open neighbourhood $U_{\varepsilon}\ni 0$ there is an open set $U_{\delta}\subseteq U_{\varepsilon}$ such that a solution (with respect to $F$) starting in $U_{\delta}$ stays in $U_{\varepsilon}$;
\item $0$ is \textit{globally attractive}, that is, $\lim_{t\to+\infty}\varphi^t(x_0)=0$ for all $x_0\in \mathbb{R}^n$.
\end{enumerate}
We will not further digress into solutions and stability and refer to~\cite{ref:sontag2013mathematical}. In general it is not straightforward to capture if $0$ is GAS or not. A fruitful tool that \textit{does} allow for conclusions of this form has been devised by Lyapunov in the late 1800s~\cite{ref:liapunov1892general}. A function $V\in C^{\infty}(\mathbb{R}^n;\mathbb{R}_{\geq 0})$ is said to be a (smooth, strict and proper) \textbf{\textit{Lyapunov function}} (with respect to $F$ and $0$) when
\begin{enumerate}[(V-i)]
\item \label{prop:i:V}$V(x)>0$ for all $x\in \mathbb{R}^n\setminus\{0\}$ and $V(0)=0$;
\item \label{prop:ii:V}$\langle \nabla V(x),F(x) \rangle <0$ for all $x\in \mathbb{R}^n\setminus\{0\}$;
\item \label{prop:iii:V}and $V$ is \textit{radially unbounded}, that is, $V(x)\to +\infty$ for $\|x\|\to+\infty$.
\end{enumerate}
Property~(V-\ref{prop:iii:V}) implies sublevelset compactness and is sometimes referred to as \textit{weak coercivity}. Now, based on work by Massera, Kurzweil and others~\cite{ref:kurzweil1963inversion,ref:fathi2019smoothing}, we will exploit the celebrated theorem stating that $0$ is GAS if and only if there is a (corresponding) smooth Lyapunov function~\cite[Theorem~2.4]{ref:bacciotti2005liapunov}. Note, we dropped the adjective ``\textit{strict and proper}'' as we exclusively look at Lyapunov functions of that form. See also that, given that $V$ satisfies Property~(V-\ref{prop:i:V}), then $\langle \nabla V(x),F(x)\rangle \leq -V(x)$ implies Property~(V-\ref{prop:ii:V}).
For further references on Lyapunov stability theory we point the reader to~\cite{ref:bhatia1970stability,ref:sontag2013mathematical,ref:bacciotti2005liapunov}.
Now, given a control system~\eqref{equ:control:sys}, when it comes to the task of \textit{globally asymptotically stabilizing} $0$ (we will exclusively focus on stabilization by means of state feedback\footnote{Considering more general input functions,~\textit{e.g.}, of the form $t\mapsto \mu(t,x(t))$, integral curves of the corresponding closed-loop system are generally understood to be absolutely continuous curves $\xi:I\to \mathbb{R}^n$ such that the differential relation $\dot{\xi}(t)=F(\xi(t),\mu(t,\xi(t)))=:F'(t,\xi(t))$ holds for almost all $t\in I$, in the sense of Lebesgue. This requires rethinking some concepts,~\textit{e.g.}, global asymptotic stability and what a closed-loop \textit{vector field} really is.}), the Lyapunov function paradigm can be adjusted. Given our stabilization goal, we seek a function $t\mapsto \mu(x(t))$ such that under $f(x,\mu(x))=:F(x)$ the origin is GAS. Then, analogously to the definition of a Lyapunov function, one can define \textit{\textbf{control Lyapunov functions}} (CLFs), yet, Property~(V-\ref{prop:ii:V}) is now replaced by asking that for any $x\in \mathbb{R}^n\setminus\{0\}$ the following holds
\begin{equation}
\label{equ:CLF:cond}
\inf_{u\in \mathbb{R}^m}\langle \nabla V(x), f(x,u) \rangle < 0.
\end{equation}
It is not evident that a choice of input function based on~\eqref{equ:CLF:cond} can result in a continuous---let alone smooth---feedback. The next section elaborates on this problem.
\subsection{On control Lyapunov functions}
Consider a dynamical control affine system with scalar input of the form
\begin{equation}
\label{equ:basic:control:sys}
\frac{\mathrm{d}}{\mathrm{d}t}x(t) = f(x(t))+g(x(t))u,
\end{equation}
Then, for $V$ to be a smooth CLF for~\eqref{equ:basic:control:sys}, we must have that for any $x\in \mathbb{R}^n\setminus\{0\}$ there exists a $u\in \mathbb{R}$ such that $L_f V(x) + u L_g V(x) < 0$. However, the existence of a \textit{smooth} \textit{control}-Lyapunov function is topologically strong in the sense that it generally implies (see below) that an asymptotically stabilizing \textit{continuous} feedback exists~\cite[Chapter~5]{ref:sontag2013mathematical}. Indeed, the controller attributed to Sontag is
\begin{equation}
\label{equ:Sontag}
\mu_s(x) := \begin{cases}
- \dfrac{L_f V(x)+\sqrt{(L_fV(x))^2+(L_gV(x))^4}}{L_gV(x)} \quad & \text{if } L_gV(x)\neq 0\\
0 \quad & \text{otherwise},
\end{cases}
\end{equation}
\textit{e.g.}, see~\cite[p.~249]{ref:sontag2013mathematical}. Although~\eqref{equ:Sontag} appears singular, $\mu_s(x)$ can be shown to be continuous under the following condition; we speak of the \textit{{small control property}} when for all $\varepsilon>0$ there is a $\delta>0$ such that if $x\in \mathbb{R}^n\setminus \{0\}$ satisfies $\|x\|<\delta$, then, there is a $u$ such that $\|u\|<\varepsilon$ and $L_fV(x)+L_gV(x) u<0$~\cite[p.~247]{ref:sontag1989universal}. As such, the existence of a \textit{smooth} CLF is strictly stronger than being (globally) asymptotically controllable\footnote{See for example~\cite[Section~2]{ref:rifford2002semiconcave} and references therein for more on this notion.},~\textit{e.g.}, continuous feedback can be easily obstructed for globally controllable systems that even admit smooth CLFs\footnote{A well-known example attributed to Ledyaev \& Sontag is of the form $\dot{x}_1=u_2u_3$, $\dot{x}_2=u_1u_3$, $\dot{x}_3=u_1u_2$~\textit{cf.}~\cite{ref:ledyaev1999lyapunov}.}. To add, the small control property does not always hold and it is well-known that CLF-based-controllers can be singular, and ever since their inception so-called ``\textit{desingularization techniques}'' emerged~\cite[Section~12.5.1]{ref:coron2007control}. For instance, under structural assumptions a backstepping approach to handle CLF singularities is studied in~\cite{ref:li1997maximizing} and a PDE reformulation to avoid singularities is presented in~\cite{ref:yamashita2000global}.
Nevertheless, in case the dynamical control system is affine in the input $u$ and $u$ is constrained to a compact convex set, then the \textit{existence} of a $C^{\infty}$ CLF is equivalent to the existence of a $C^0$ (on $\mathbb{R}^n\setminus\{0\}$) stabilizing feedback~\cite{ref:artstein1983stabilization}. Indeed, the work by Sontag aimed at making the \textit{construction} of such a feedback transparent. Further relaxing regularity of a CLF, it can be shown that the existence of a so-called ``\textit{proximal CLF}'' is equivalent to asymptotic controllability. These proximal CLFs are $C^r$-smooth with $r\in [0,1)$,~\textit{e.g.}, see~\cite{ref:clarke2010discontinuous} for more on non-smooth CLFs. Better yet, it can be shown that global asymptotic controllability implies the existence of a---possibly discontinuous---feedback~\cite{clarke1997asymptotic}. Even more, Rifford showed that when the control system is globally asymptotically controllable, a---possibly nonsmooth---semiconcave\footnote{A continuous function $f$ is said to be \textit{semiconcave} when there is a $C>0$ such that $x\mapsto f(x)-C\|x\|_2^2$ is concave.} CLF always exists. Exploiting this structure, for control affine systems, Rifford could extend Sontag's formula~\eqref{equ:Sontag} to this setting~\cite[Theorem~2.7]{ref:rifford2002semiconcave} and \textit{get} again an explicit feedback.
The existence of a smooth CLF is not only topologically strong, it implies there exists a \textit{robustly} stabilizing feedback~\cite{ref:ledyaev1999lyapunov}.
\subsection{On learning-based stabilization}
\label{sec:stab:project}
Neural networks are becoming increasingly populair in the context of controller synthesis~\cite{ref:jin2020neural,ref:gaby2021lyapunov,ref:mukherjee2022neural,ref:zhang2022neural}. A principled approach, however, that guarantees some form of stability is largely lacking. Progress has been made when it comes to handling side-information~\cite{ref:ahmadi2020learn}, obtaining statistical stability guarantees \cite{ref:boffi2021learn}, in the context of input-state stability under CTRNN modelling assumptions~\cite{ref:yang2022input}, in the context of input-output stability by exploiting the Hamilton-Jacobi inequality~\cite{ref:okamoto2022learning}, by exploiting contraction theory~\cite{ref:rezazadeh2022learning} and by exploiting Koopman operator theory~\cite{ref:zinage2022neural}, to name a few.
As these methods are data-driven, errors inevitably slip in and great care must be taken when one aims to mimic CLF-based controllers,~\textit{i.e.}, if $L_gV(x)=0\implies L_fV(x)<0$ holds for the estimated system, does it hold for the real system and what happens if it does not? In particular, recall~\eqref{equ:Sontag}. Moreover, in this setting the underlying dynamical control system is frequently unknown and a function class for $V$ needs to be chosen \textit{a priori}, what does this choice imply? These questions motivate this note.
We also point out that these methods continue a long history of research on computational methods for Lyapunov functions,~\textit{e.g.}, see~\cite{ref:giesl2015review} for a review.
\section{Topological perspective on levelsets and singularities}
\label{sec:topo}
We start by detailing (recalling) how levelsets of smooth Lyapunov functions, with respect to points, look like topologically. This result has some ramifications and provides for motivation in the next section.
For simplicity, we momentarily focus on~\eqref{equ:basic:control:sys}.
Section~\ref{sec:stab:project} discussed why one might be interested in studying terms of the form $L_gV(x)^{-1}=\langle \nabla V(x), g(x) \rangle^{-1}$~\textit{cf.}~\eqref{equ:Sontag}. In this section we show that for practical purposes, the properties of $V$ frequently obstruct this term to be well-behaved. Indeed, singularities are studied and shown to be unavoidable when $g(x):=g$ for some $g\in \mathbb{R}^n$.
To start, consider a $C^{0}$ dynamical system of the form~\eqref{equ:dyn:sys} on $\mathbb{R}^n$, with $n\geq 2$, and assume that $0\in \mathbb{R}^n$ is globally asymptotically stable (and hence isolated). This implies that there is a (strict) $C^{\infty}$ Lyapunov function $V:\mathbb{R}^n\to \mathbb{R}_{\geq 0}$. In particular, this implies that $V$ is also a Lyapunov function for the $C^{\infty}$ auxiliary system
\begin{equation}
\label{equ:aux}
\frac{\mathrm{d}}{\mathrm{d}t}z(t) = - \nabla V(z(t)).
\end{equation}
Hence, $0\in \mathbb{R}^n$ is also GAS under~\eqref{equ:aux}. By a classical topological result largely\footnote{Earlier comments can be found in~\cite{ref:bobylev1974deformation}.} due to Krasnosel'ski\u{\i} \& Zabre\u{\i}ko~\cite[Section~52]{ref:krasnosel1984geometrical} this directly implies that the corresponding \textit{vector field index} (with respect $0$) satisfies
\begin{equation*}
\mathrm{ind}_0(-\nabla V)=(-1)^n\neq 0.
\end{equation*}
As the vector field index is the (oriented) degree of the ``\textit{normalizing}'' map $v:\partial U \to \mathbb{S}^{n-1}$ for any open neighbourhood $U$ of $0$ containing no other equilibrium points in its closure~\cite[Section~6]{ref:milnor65},~\cite[Chapter~3]{ref:guillemin2010differential}, this can only be true if
\begin{equation*}
v:\partial U\ni z \mapsto \frac{-\nabla V(z)}{\| \nabla V(z)\|_2}
\end{equation*}
is surjective. As $U$ is arbitrary, it follows that the (normalized) gradient of $V$ along any non-trivial levelset hits any vector in $\mathbb{S}^{n-1}$. Differently put, fix any $g\in \mathbb{S}^{n-1}$ then, for any $c>0$ there is always a $z\in V^{-1}(c)=: V_c\subset \mathbb{R}^n$ such that $\langle \nabla V(z), g \rangle =0$. Indeed, this is why we assumed $n\geq 2$, otherwise the claim is not true~\textit{cf.}~\cite[p.~121]{ref:sontag1989universal}. Summarizing, we have shown the following---which is attributed to Wilson~\cite{ref:wilson1967structure} and Byrnes~\cite[Theorem~4.1]{ref:byrnes2008brockett}.
\begin{proposition}[Levelsets of smooth Lyapunov functions (Wilson, Byrnes)]
\label{prop:Lyapunov:levelset}
Let $n\geq 2$ and fix some $g\in \mathbb{R}^n\setminus \{0\}$. Then, for any levelset $V_{c}$, with $c>0$, of any $C^{\infty}$-smooth Lyapunov function $V:\mathbb{R}^n\to \mathbb{R}_{\geq 0}$, asserting $0\in \mathbb{R}^n$ to be GAS under some dynamical system~\eqref{equ:dyn:sys}, there is a $x\in V_c$ such that $\langle \nabla V(x),g \rangle =0$.
\end{proposition}
Note, Proposition~\ref{prop:Lyapunov:levelset} implicitly assumes that $0\in \mathbb{R}^n$ is the only equilibrium point as we assume the origin is \textit{globally} asymptotically stable. If desired, one can adapt the statement and work with the domain of attraction. Also note that the discussion above detailed that the normalized vector $\nabla V(x)$ will hit \textit{any} vector in $\mathbb{S}^{n-1}$, our focus on $\langle \nabla V(x),g \rangle $ being equal to $0$ at some point is purely application-driven. Moreover, we see that the set of points that render the inner product zero is of codimension $1$.
Indeed, Proposition~\ref{prop:Lyapunov:levelset} is itself classical as this result can also be understood more intuitively by directly appealing to work by Wilson. Namely, due to the work by Wilson and later Perelman we know that the levelsets of (strict and proper) $C^{\infty}$ Lyapunov functions $V:\mathbb{R}^n\to \mathbb{R}_{\geq 0}$ are homeomorphic to $\mathbb{S}^{n-1}$~\cite{ref:wilson1967structure,ref:stillwell2012poincare}. Although we might assume that these levelsets $V_c$ and $\mathbb{S}^{n-1}$ come equipped with a smooth structure, this does not immediately imply the manifolds are diffeomorphic,~\textit{e.g.}, consider Milnor's \textit{exotic spheres}~\cite{ref:Milnor7}. Nevertheless, one expects that the gradient of $V$ along $V_c$ hits any direction as seen as a vector in $\mathbb{S}^{n-1}$, as indeed succinctly shown above. Visualizations can be found in~\cite{ref:sontag1999stability} and further comments of this nature are collected by Byrnes in~\cite{ref:byrnes2008brockett}, in particular, the diffeomorphism question is addressed.
The ramifications for smooth CLFs are immediate as one observes that the argument with respect to the auxiliary system~\eqref{equ:aux} extends \textit{mutatis mutandis}.
This note was motivated by renewed interest in CLFs from the neural network community. The following example highlights some work that arguably would benefit from Proposition~\ref{prop:Lyapunov:levelset}.
\begin{example}[(Almost) Singular CLF-based controllers]
In~\cite[Section~IV]{ref:kashima2022learning} the authors consider a dynamical control system of the form $\dot{x}=f(x)+gu$ with $f\in C^{\infty}(\mathbb{R}^2;\mathbb{R}^2)$, $g\in \mathbb{R}^2$ and $u\in \mathbb{R}$. Their to-be-learned control Lyapunov function is of the form $V(x)=\sigma_k(\gamma(x)-\gamma(0))+\varepsilon \|x\|^2$ for $\sigma_i:\mathbb{R}\to \mathbb{R}_{\geq 0}$ $C^{\infty}$-smooth ReLU functions and $\gamma:\mathbb{R}^n\to \mathbb{R}$ being an input-convex neural network. Hence, $V\in C^{\infty}(\mathbb{R}^n;\mathbb{R}_{\geq 0})$. Indeed, the authors report that the learned CLF leads to large control values (under a Sontag-type controller~\eqref{equ:Sontag}), they do not detail why. The above discussion provides a topological viewpoint.
\end{example}
One can also interpret Proposition~\ref{prop:Lyapunov:levelset} through the lens of feedback linearization.
Consider some input-output system $\Sigma$ of the form
\begin{equation}
\label{equ:IO}
\Sigma: \left\{ \begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}{x(t)} &= f(x(t))+g u\\
y(t) &= h(x(t))
\end{aligned}\right.
\end{equation}
for the output function $h$ given by the CLF $V$ (with respect to $f$ and $g$). Let the desired output be $y_d\equiv 0$ such that $e(t)=y(t)-y_d(t)=y(t)$. Hence, $\dot{e}=\dot{V}$. Now the standard (relative degree $1$) feedback linearizing controller for~\eqref{equ:IO} is of the form $u=(L_gV)^{-1}(v-L_fV)$ with $v$ denoting the new auxiliary input~\cite{ref:isidori1985nonlinear,nonlin}. Indeed, under the choice
\begin{equation*}
v=-\sqrt{(L_fV)^2+(L_gV)^4}
\end{equation*}
one recovers Sontag's controller~\eqref{equ:Sontag}. Now Proposition~\ref{prop:Lyapunov:levelset} tells us that the \textit{decoupling} term $(L_gV)^{-1}$ must be singular in any sufficintly small neighbourhood of $0$,~\textit{i.e.}, the relative degree assumption fails to hold.
\begin{remark}[Generalizations]
To go beyond input vector fields of the form $g(x):= g\in \mathbb{R}^n$ we look at two scenarios.
\begin{enumerate}[(g-i)]
\item (Dependency on $x$): Introduce the function class
\begin{equation*}
\mathscr{G}_n:=\{g\in C^{0}(\mathbb{R}^n;\mathbb{R}^n):g(x)=g_1+g_2(x),\,g_1\in \mathbb{R}^n\setminus\{0\},\,\lim_{x\to 0}g_2(x)=0\}.
\end{equation*}
Indeed, for any $g\in \mathscr{G}_{n}$, with $n>1$, it follows that for sufficiently small $c>0$ there is a $x\in V_c$ such that $\langle \nabla V(x),g(x)\rangle=0$. The reason being that since $g\in \mathscr{G}_{n}$ there are always $x_1,x_2\in V_c$ such that $\langle \nabla V(x_1),g(x_1)\rangle<0$ while $\langle \nabla V(x_2),g(x_2)\rangle>0$. Then the claim follows from standing regularity assumptions and the intermediate value theorem.
\item (Multidimensional input): Assume that $u\in \mathbb{R}^m$ with $1<m<n$ and let the dynamical control system be of the form $\dot{x}=f(x)+\sum^m_{i=1}g_i u_i$ (dependence on $x$ can be generalized as above). Then, as $\mathrm{span}\{g_1,\dots,g_m\}\neq \mathbb{R}^n$ there is a nonzero $v\in \mathrm{span}\{g_1,\dots,g_m\}^{\perp}$.
\end{enumerate}
\end{remark}
Exploiting the remark from above, we recover a slightly weaker version of a well-known result~\textit{cf.}~\cite[Proposition~6.1.4]{ref:bloch}.
\begin{corollary}[Obstruction for nonholonomic systems]
Assume that $u\in \mathbb{R}^m$ with $1<m<n$ and let the dynamical control system be of the form $\dot{x}=\sum^m_{i=1}g_i u_i$, then, there is no smooth CLF with respect to $0\in \mathbb{R}^n$.
\end{corollary}
\begin{proof}
Indeed this result follows for example Brockett's condition~\cite{ref:brockett1983asymptotic}. However, from Proposition~\ref{prop:Lyapunov:levelset} we know there is a point $x'\in \mathbb{R}^n\setminus \{0\}$ such that $\nabla V(x')\perp \mathrm{span}\{g^1,\dots,g^m\}$. This implies that $L_fV(x')<0$ must hold for $V$ to be a CLF. As $f:=0$, this is impossible and no smooth CLF can exist.
\end{proof}
In a similar vein, using $\mathscr{G}_{n}$, one recovers (locally) a slightly weaker version of the highly influential obstruction to continuous asymptotic stabilization of Brockett's nonholonomic integrator~\textit{e.g.}, see~\cite[Example~5.9.16]{ref:sontag2013mathematical}.
\section{On convexity and continuation}
\label{sec:cvx}
The previous section illustrated why levelsets of Lyapunov functions are topological spheres. As such, this motivates the hope that all those Lyapunov functions can be transformed---in some sense---to the canonical Lyapunov function $V(x)=\tfrac{1}{2}\langle x,x\rangle$. We start this study of transformations by looking at \textit{convex} Lyapunov functions, as this class is particularly simple to handle. Better yet, by exploiting this structure, it follows that any convex Lyapunov function also asserts stability of the ``\textit{canonical}'' inward pointing vector field on $\mathbb{R}^n$ indeed, which we will denote with some abuse of notation by the map $-\mathrm{id}_{\mathbb{R}^n}$. Exactly this observation will be formalized and further studied below.
Convexity in the context of Lyapunov stability theory has been an active research area.
For example, convexity in linear optimal control~\cite{AREbook}, convexity in the dual density formulation due to Rantzer~\cite{ref:prajna2004nonlinear} and convexity of the set of Lyapunov functions due to Moulay~\cite{ref:moulay2010some}. We are, however, interested in understanding convexity of Lyapunov functions themselves. It is known that simple asymptotically stable dynamical systems do not always admit polynomial Lyapunov functions. For instance
\begin{equation}
\label{equ:ahmadi}
\frac{\mathrm{d}}{\mathrm{d}t}\begin{pmatrix}
x_1(t)\\ x_2(t)
\end{pmatrix} = \begin{pmatrix}
-x_1(t)+x_1(t)x_2(t)\\ -x_2(t)
\end{pmatrix}
\end{equation}
does not admit a polynomial Lyapunov function~\cite{ref:ahmadi2011globally}, but one can show that $V(x)=\log(1+x_1^2)+x_2^2$ checks out as a Lyapunov function asserting $0\in \mathbb{R}^2$ is GAS. Indeed, $V$ is smooth, yet \textit{not} convex. We will come back to this several times below.
Similar obstructions can be found for analytic or rational Lyapunov functions~\cite{ref:bacciotti2005liapunov,ref:ahmadi2018globally}.
The (computational) assumption to look for \textit{convex} Lyapunov functions is a populair one in the learning community,~\textit{e.g.}, propelled by~\cite{ref:amos2017input,ref:kolter2019learning}. However, this assumption evidently restricts the problem class that can be handled. The ramifications of assuming Lyapunov functions to be convex are understood in the context of linear systems, linear differential inclusions even~\cite{ref:goebel2006conjugate}, but not in the $C^0$ nonlinear setting. Indeed, in our setting, for $n>1$, one can construct vector field examples $\dot{x}=F(x)$ over $\mathbb{R}^n$ such that $0$ is globally asymptotically stable, $F$ is smooth, yet no smooth convex Lyapunov function exists. To see why, for the sake of contradiction, one can exploit that by convexity we must have
\begin{equation*}
\langle \nabla V(x), x\rangle \geq 0 \quad \forall x\in \mathbb{R}^n
\end{equation*}
and due to the stability assumption we have
\begin{equation*}
\langle \nabla V(x),F(x)\rangle <0 \quad \forall x\in \mathbb{R}^n\setminus\{0\}
\end{equation*}
such that the function $V$ must satisfy $\langle \nabla V(x), F(x)-x \rangle < 0$ for all $x\in \mathbb{R}^n\setminus \{0\}$. Hence, if there is a non-zero fixed point\footnote{Note, here we heavily exploit the underlying vector space structure to be able to compare $x$ and $F(x)$.} of $F$, we contradict the existence of such a $V$. For a visual example, see Figure~\ref{fig:ncvxLyap} for phase portraits illustrating dynamical systems with such a fixed point that obstruct the existence of a \textit{smooth convex} Lyapunov function. As one will be able to infer from the results below, $F$ cannot point (radially) outward. Indeed, it is known that for \textit{homogeneous} Lyapunov functions this can also not be true~\cite[Proposition~1]{ref:sepulchre1996homogeneous}.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.4]{F1.png}
\caption{$\alpha=-0.2$, $\beta=1/100$, $\gamma=10$ and $p=(0.3,0.3)$}
\label{fig:F1}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.4]{F2.png}
\caption{$\alpha=-0.1$, $\beta=1/100$, $\gamma=10$ and $p=(0.5,0.5)$}
\label{fig:F2}
\end{subfigure}
\caption{Example~\ref{ex:nec:ncvx}: integral curves of a smooth dynamical systems that obstruct the existence of a \textit{convex} Lyapunov function. The figures are made with Python.}
\label{fig:ncvxLyap}
\end{figure}
We will now formalize this observation.
\begin{theorem}[Convex Lyapunov functions]
\label{thm:cvx:Lyap}
Let $F\in C^{0}(\mathbb{R}^n;\mathbb{R}^n)$ give rise to $\dot{x}=F(x)$ with $0\in \mathbb{R}^n$ globally asymptotically stable (GAS) under $F$. Then, if there is a convex $C^{\infty}$ Lyapunov function asserting $0$ is GAS, the vector field $F$ is straight-line homotopic to $-\mathrm{id}_{\mathbb{R}^n}$ such that $0$ is GAS throughout the homotopy.
\end{theorem}
\begin{proof}
By construction there is a $C^{\infty}$ Lyapunov function $V$ such that $\langle \nabla V(x),F(x) \rangle <0$ for all $x\in \mathbb{R}^n\setminus\{0\}$. By the convexity of $V$ we also know that
\begin{equation}
\label{equ:V:cvx}
V(y)\geq V(x) + \langle \nabla V(x),y-x \rangle,\quad \forall y,x\in \mathbb{R}^n.
\end{equation}
In particular,~\eqref{equ:V:cvx} must hold for $y=0$, which yields $\langle V(x),-x\rangle \leq- V(x)$, that is, $V$ is also a Lyapunov function for $\dot{x}=-x$.
Alternatively, construct the \textit{(Legendre-Fenchel) conjugate} of $V$, that is
\begin{equation}
\label{equ:conj}
V^*(y):=\sup\{\langle y,x \rangle -V(x):x\in \mathbb{R}^n\},
\end{equation}
where $V^*(y)\geq 0\,\forall y\in \mathbb{R}^n$ follows from selecting $x=0$. Now, by convexity and smoothness of $V$ we have that $V^*(\nabla V(x))+V(x) = \langle \nabla V(x),x\rangle$~\cite[Proposition~11.3]{ref:rockafellar2009variational}.
Exploiting~\eqref{equ:conj} and $V(x)>0\,\forall x\in \mathbb{R}^n\setminus\{0\}$, this implies that $\langle V(x),-x\rangle < 0\, \forall x\in \mathbb{R}^n\setminus\{0\}$.
Regardless of the approach, we find that $0$ is also GAS under $sF(x)-(1-s)x$ for all $s\in [0,1]$ since for any such $s$
\begin{equation*}
\langle \nabla V(x), sF(x)-(1-s)x\rangle <0 \quad \forall x\in \mathbb{R}^n\setminus \{0\}.
\end{equation*}
Simultaneously, this parameterization in $s$ provides the desired homotopy.
\end{proof}
To illustrate the homotopy resulting from Theorem~\ref{thm:cvx:Lyap}, two vector fields $F_1$ and $F_2$ on $\mathbb{R}^n$ such that $0$ is GAS---asserted via possibly different smooth convex Lyapunov functions---are homotopic through a continuous map $H:[0,1]\times \mathbb{R}^n\to \mathbb{R}^n$ of the form
\begin{equation*}
H(s,x) = \begin{cases}
-2sx +(1-2s)F_1(x) \quad & s\in [0,\tfrac{1}{2}]\\
-(2-2s)x+(2s-1)F_2(x) \quad & s\in (\tfrac{1}{2},1].
\end{cases}
\end{equation*}
A variety of known topological conditions capture the existence of a (local) homotopy (in far more general settings), but not that \textit{along} the homotopy stability is preserved~\textit{cf.}~\cite{ref:kvalheim2022obstructions}.
Similar statements can be made about \textit{control} Lyapunov functions.
\begin{corollary}[Convex control Lyapunov functions]
\label{cor:cvx:control:Lyap}
Let $f\in C^{0}(\mathbb{R}^n\times \mathbb{R}^m;\mathbb{R}^n)$ give rise to the control system $\dot{x}=f(x,u)$. If there is a convex control Lyapunov function (CLF) $V\in C^{\infty}(\mathbb{R}^n;\mathbb{R}_{\geq 0})$ for this control system with respect to $0$, then, $V$ is a CLF for any control system on the straight-line homotopy between $f$ and the map $(x,u)\mapsto -x$.
\end{corollary}
\begin{proof}
The proof is identical to that of Theorem~\ref{thm:cvx:Lyap}, yet, now we start from $V$ satisfying
\begin{equation*}
\forall x\in \mathbb{R}^n\setminus\{0\}\,\exists u\in \mathbb{R}^m\,:\,\langle V(x), f(x,u)\rangle < 0.
\end{equation*}
and again exploit convexity of $V$ to conclude.
\end{proof}
In conclusion, we see from Theorem~\ref{thm:cvx:Lyap} that a necessary condition for $\dot{x}=F(x)$ to admit a smooth convex Lyapunov function, asserting $0$ is GAS, is that
\begin{subequations}
\label{equ:nec}
\begin{equation}
\label{equ:cvx:L:nec}
F(x)\neq \lambda x \quad \forall \lambda\in \mathbb{R}_{\geq 0},\, \forall x\in \mathbb{R}^n\setminus\{0\}.
\end{equation}
Differently put, if $V$ is a Lyapunov function for $\dot{x}=F(x)$, it is also a Lyapunov function for $\dot{x}=F(x)-\lambda x$, with $\lambda\geq 0$. The next example shows we can find families of dynamical systems that do not obey Condition~\eqref{equ:cvx:L:nec}.
\begin{example}[Necessarily nonconvex]
\label{ex:nec:ncvx}
The systems as shown in Figure~\ref{fig:ncvxLyap} can be made explicit. Consider a $C^{\infty}$ dynamical system of the form~\eqref{equ:dyn:sys} on $\mathbb{R}^2$ as given by
\begin{equation}
\label{equ:nlin:ncvx}
\frac{\mathrm{d}}{\mathrm{d}t}\begin{pmatrix}
x_1(t)\\ x_2(t)
\end{pmatrix} = \begin{pmatrix}
\alpha & 1\\
-1 & \alpha
\end{pmatrix}\begin{pmatrix}
x_1(t)\\ x_2(t)
\end{pmatrix} + \gamma \mathrm{exp}(-\beta \|x(t)-p\|_2^2 )\begin{pmatrix}
1\\ 1
\end{pmatrix}-\gamma \mathrm{exp}(-\beta \|p\|_2^2 )\begin{pmatrix}
1 \\ 1
\end{pmatrix}
\end{equation}
where $x=(x_1,x_2)\in \mathbb{R}^2$, and for instance $\alpha=-0.2$, $\beta=1/100$, $\gamma=10$ and $p=(0.3,0.3)\in \mathbb{R}^2$ correspond to $F_1$ in Figure~\ref{fig:F1}. Indeed, as eigenvectors of the linear part are never of the form $(1,1)\in \mathbb{R}^2$, only $0\in \mathbb{R}^2$ is an equilibrium point of this dynamical system. Note,~\eqref{equ:nlin:ncvx} can also be understood as a stabilizable linear system under a (bounded) nonlinear feedback. We can look at another example when we change $\alpha$ and $p$ to $\alpha=-0.1$ and $p=(0.5,0.5)\in \mathbb{R}^2$, then, regarding our necessary condition for convexity, we find, for instance, that $x=(0.51,0.45)\in \mathbb{R}^2$ and $\lambda=0.0629$ provide for a (numerical) invalidation of~\eqref{equ:cvx:L:nec} indeed. This setting corresponds to $F_2$ as shown in Figure~\ref{fig:F2}.
\end{example}
\begin{remark}[A nonconvex conic structure]
Let $0\in \mathbb{R}^n$ be GAS under $\dot{x}=F(x)$, then $0$ is also GAS under $\dot{x}=\theta F(x)$ for any $\theta>0$,~\textit{e.g.}, consider $\langle \nabla V(x), F(x)\rangle $ and $\langle \nabla V(x), \theta F(x)\rangle$ for some Lyapunov function $V$ with respect to $F$. Hence, by Theorem~\ref{thm:cvx:Lyap}, all $\theta F$ are straight-line homotopic to $-\mathrm{id}_{\mathbb{R}^n}$. However, convexity breaks down as already the set of Hurwitz stable matrices is nonconvex,~\textit{e.g.},
\begin{equation*}
s\begin{pmatrix}
-1 & 10\\ 0 & -1
\end{pmatrix} + (1-s)\begin{pmatrix}
-1 & 0\\ 10 & -1
\end{pmatrix}
\end{equation*}
becomes unstable (not all eigenvalues lie in $\mathbb{C}_{\Re<0}$) for $s=\tfrac{1}{2}$.
\end{remark}
Similarly, from Corollary~\ref{cor:cvx:control:Lyap} we see that the control system $\dot{x}=f(x,u)$ admits a smooth convex CLF only when
\begin{equation}
\label{equ:cvx:CLF:nec}
\forall x\in \mathbb{R}^n\setminus \{0\}\,\exists u \in \mathbb{R}^m\,:\, f(x,u)\neq \lambda x \quad \forall \lambda\in \mathbb{R}_{\geq 0}.
\end{equation}
\end{subequations}
Indeed, one can replace $\lambda\in \mathbb{R}_{\geq 0}$ in~\eqref{equ:nec} by $\lambda(x)\in \mathbb{R}_{\geq 0}$, for example, $\lambda\in C^{0}(\mathbb{R}^n;\mathbb{R}_{\geq 0})$~\textit{cf.}~\cite{ref:sepulchre1996homogeneous}.
\begin{example}[Linear dynamical systems]
\label{ex:lin:sys}
Consider the linear dynamical system $\dot{x}=Ax$ for some matrix $A\in \mathbb{R}^{n\times n}$. Theorem~\ref{thm:cvx:Lyap} implies that for a convex Lyapunov function to exist (asserting $0$ is GAS) the expression $sAx-(1-s)x$ cannot vanish for some $s\in[0,1]$ and $x\in \mathbb{R}^n\setminus 0$. Reformulating, we get~\eqref{equ:cvx:L:nec},~\textit{i.e.}, $Ax=\lambda x$ cannot have a solution for some $\lambda\geq 0$ and $x\in \mathbb{R}^n\setminus \{0\}$. However, this is precisely stating that $A$ cannot have an unstable eigenvalue of the form $\lambda\in \mathbb{R}_{\geq 0}$. Indeed, for globally asymptotically stable linear systems a convex (quadratic) Lyapunov function of the form $V(x)=\tfrac{1}{2}\langle Px, x \rangle$ always exists~\cite[Theorem~18]{ref:sontag2013mathematical}.
\end{example}
\begin{remark}[On sufficiency]
\label{rem:suff}
Example~\ref{ex:nec:ncvx} showed that for dynamical systems of the form~\eqref{equ:nlin:ncvx} no convex Lyapunov function can exist. Going back to~\eqref{equ:ahmadi}, the provided Lyapunov function is nonconvex. Concurrently, one can check that~\eqref{equ:cvx:L:nec} holds, so a convex Lyapunov function is not ruled out. We come back to this below. As discussed above in Example~\ref{ex:lin:sys}, in the linear case Condition~\ref{equ:cvx:L:nec} is, however, necessary and sufficient for a convex Lyapunov function to exist.
\end{remark}
To elaborate on Example~\ref{ex:lin:sys}, for controllable linear systems,~\textit{e.g.}, of the form $\dot{x}=Ax+Bu$, one can always parameterize a quadratic Lyapunov function for the LQ optimally controlled closed-loop system by the positive definite solution to the corresponding Ricatti equation (for any appropriate cost)~\cite[Theorem~42, Exercise~8.5.4]{ref:sontag2013mathematical}.
\begin{example}[Linear dynamical control systems and Hautus' test]
\label{ex:lin:control:sys}
A celebrated condition largely attributed to Hautus (and Belevitch and Popov) states that a linear dynamical control system of the form $\dot{x}=Ax+Bu$, is stabilizable when
\begin{equation}
\label{equ:Hautus}
\mathrm{rank}\left( \begin{pmatrix}A-\lambda I_n & B \end{pmatrix} \right)=n \quad \forall \lambda \in \sigma(A)\cap \mathbb{C}_{\Re\geq 0},
\end{equation}
where $\sigma(A)$ denotes the spectrum of $A$. See for instance~\cite[Chapter~3]{ref:trentelman2012control}.
Now, elementary algebraic arguments show that Hautus' condition~\eqref{equ:Hautus} implies that~\eqref{equ:cvx:CLF:nec} holds, as it should for linear control systems.
\end{example}
Using the above, one can readily verify that for example
\begin{equation}
\label{ex:sys:1}
\frac{\mathrm{d}}{\mathrm{d}t}
\begin{pmatrix}
{x_1}(t)\\ {x_2}(t)
\end{pmatrix} = \begin{pmatrix}
x_1(t) u\\ x_1(t)x_2(t) u
\end{pmatrix}
\end{equation}
does not admit a smooth convex CLF. Indeed, for~\eqref{ex:sys:1}, controllability is lost at $(0,x_2)\in \mathbb{R}^2$.
Example~\ref{ex:lin:sys} and Example~\ref{ex:lin:control:sys} show that conditions~\eqref{equ:cvx:L:nec} and~\eqref{equ:cvx:CLF:nec} are to some extent generalizations of known conditions for linear systems, yet, lifted to nonlinear systems under convexity assumptions. These conditions are, however, weak.
A stronger set of conditions one can derive from Theorem~\ref{thm:cvx:Lyap} is of the form: $\dot{x}=f(x,u)$ admits a smooth convex CLF only if $\dot{x}=f(x,u)-\lambda(x)x$ does, for any $\lambda\in C^{0}(\mathbb{R}^n;\mathbb{R}_{\geq 0})$.
We are not the first to observe something of this form, \textit{e.g.},~Sepulchre \& Aeyels~ \cite[Section~4.1]{ref:sepulchre1996homogeneous} look at homogeneous CLFs and recover a similar condition.
\subsection{On compact convex sets}
We briefly show that without too much effort the results extend from $0\in \mathbb{R}^n$ being GAS under some dynamical system parametrized by $F\in C^{0}(\mathbb{R}^n;\mathbb{R}^n)$ to a compact convex set $A\subseteq \mathbb{R}^n$ being GAS\footnote{For more on the generalization of stability notions from points to sets we point the reader to~\cite{ref:hurley1982attractors} for a topological treatment.} under $F$. As $0$ and $A$ are (straight-line) homotopy equivalent this is perhaps not surprising. Define the projection operator by
\begin{equation*}
\Pi_A(x):=\argmin_{y\in A}\|x-y\|_2.
\end{equation*}
We have the following.
\begin{corollary}[Convex Lyapunov functions for convex compact sets]
\label{cor:cvx:Lyap:A}
Let $F\in C^{0}(\mathbb{R}^n;\mathbb{R}^n)$ give rise to $\dot{x}=F(x)$ with a compact convex set $A\subseteq \mathbb{R}^n$ being globally asymptotically stable (GAS) under $F$. Then, if there is a convex $C^{\infty}$ Lyapunov function asserting $A$ is GAS, the vector field $F$ is straight-line homotopic to $\Pi_A-\mathrm{id}_{\mathbb{R}^n}$ on $\mathbb{R}^n\setminus A$ such that $A$ is GAS throughout the homotopy.
\end{corollary}
\begin{proof}
The Lyapunov function is such that $V(x)=0\iff x\in A$, hence for the convexity condition $V(y)\geq V(x)+\langle \nabla V(x),y-x \rangle$ we pick $y=\Pi_A(x)$ such that for all $x\in \mathbb{R}^n\setminus A$ we have $\langle \nabla V(x),\Pi_A(x)-x\rangle <0$. We can conclude.
\end{proof}
Some comments are in place, we do not need $F(A)=0$, $A$ merely needs to be invariant\footnote{Let $\varphi$ be the flow corresponding to $F$, then $A$ is said to be invariant (under $\varphi$) when $\varphi(\mathbb{R},A)=A$.}. This is why we cannot say anything about the homotopy on $A$ itself. Moreover, settings like these easily obstruct $V\in C^{\omega}(\mathbb{R}^n;\mathbb{R}_{\geq 0})$ (real-analyticity), not to contradict real-analytic function theory (bump functions cannot be analytic). Also, when $A$ is not convex, $\Pi_A$ is potentially set-valued, obstructing our vector field construction and perhaps $\Pi_A-\mathrm{id}_{\mathbb{R}^n}$ is not the expected ``\textit{canonical}'' inward vector field (due to the non-scaled offset $-x$). At last, we point out that although $V$ is smooth, this does not imply that $\partial A$ must be a smooth manifold. For instance, consider $V(x_1,x_2)=(x_1-x_2)^2(x_1+x_2)^2$ (although here, $V_0$ is clearly not convex).
We direct the reader to~\cite{ref:fathi2019smoothing} and references therein for more on general (converse) Lyapunov theory with respect to sets.
\subsection{Geodesic convexity}
To go beyond convexity, we follow~\cite[Chapter~3]{ref:udriste2013convex},~\cite[Chapter~11]{ref:boumal2020introduction} and show how the situation is hardly different in the context of \textit{geodesic} convexity. We will be brief, for the details on geodesic convexity we point the reader to the references above and for background information on Riemannian geometry we suggest~\cite{Lee3}.
Let $(\mathbb{R}^n,g)$ be a $C^{\infty}$ \textit{Riemannian manifold} for some Riemannian metric $g$. One can think of $g$ as inducing a change of coordinates via the inner product $\langle \cdot,\cdot \rangle_g$, in particular, this metric has an effect on gradients,~that is, the (Riemmanian) gradient of a differentiable function $f:\mathbb{R}^n\to \mathbb{R}$, with respect to $g$, satisfies $Df(x)[v]=\langle \mathrm{grad}\,f(x), v\rangle_g$ for any $(x,v)\in T\mathbb{R}^n$, with $Df(x)[v]$ being the directional derivative in the direction $v\in T_x\mathbb{R}^n$. For example, let $g$ be parametrized by a symmetric positive definite matrix $P$, that is, $\langle v,w \rangle_g:=\langle Pv,w\rangle$ for any $v,w\in T_x\mathbb{R}^n$ and $x\in \mathbb{R}^n$, then, $\mathrm{grad}\,f(x) = P^{-1}\nabla f(x)$. Indeed, for a practical application of this in $\mathbb{R}^n$, we point the reader to a discussion of Newton's method as used in second-order optimization~\cite[Section~9.5]{ref:Boyd_04}. The metric $g$ also has ramifications for ``\textit{straight lines}'', a $C^1$ curve $[0,1]\ni s\mapsto \gamma(s)$ is a \textit{geodesic}, with respect to $g$, when it is an \textit{extremal} of the energy functional $E(\gamma):=\tfrac{1}{2}\int_{[0,1]} \langle \dot{\gamma}(\tau),\dot{\gamma}(\tau)\rangle_g \mathrm{d}\tau$. This implies geodesics are \textit{locally} minimizing length and in that sense they generalize straight lines. As this statement is local, geodesics are by no means always unique. Then, a subset $U\subseteq \mathbb{R}^n$ is called \textit{geodesically convex} ($g$-convex) when for all points $x,y\in U$ there is an \textit{unique}\footnote{See the discussion in~\cite[Section~11.3]{ref:boumal2020introduction} on various slightly different definitions of \textit{geodesic convexity} and their implications.} geodesic $\gamma:[0,1]\to \mathbb{R}^n$ (with respect to $g$) connecting $x$ to $y$ such that $\gamma([0,1])\subseteq U$. A function $f:U\subseteq \mathbb{R}^n\to \mathbb{R}$, over some $g$-convex domain $U$, is said to be \textbf{\textit{geodesically convex}} (\textbf{\textit{$g$-convex}}) when
\begin{equation}
\label{equ:g:cvx:geod}
f(\gamma(t)) \leq tf(x)+(1-t)f(y)\quad \forall t\in [0,1]
\end{equation}
for $\gamma:[0,1]\to \mathbb{R}^n$ a geodesic, with $\gamma([0,1])\subseteq U$ connecting the point $x$ to $y$. Indeed,~\eqref{equ:g:cvx:geod} generalizes the standard $C^0$ definition of convexity. A $C^1$ condition is now given by
\begin{equation*}
f(\mathrm{Exp}_x(tv))\geq f(x) + t\langle \mathrm{grad}\,f(x),v\rangle_g\quad \forall t\in [0,1],
\end{equation*}
where $v\in T_x\mathbb{R}^n$, $\mathrm{Exp}_x$ is the (Riemannian) \textit{exponential map} at $x\in U\subseteq \mathbb{R}^n$ and $\mathrm{grad}\,f(x)$ is the \textit{Riemannian gradient} of $f$. Here, the exponential map is defined, locally, by $\mathrm{Exp}_x(v)=\gamma(1)$ for $\gamma$ the unique geodesic with $\gamma(0)=x$ and $\dot{\gamma}(0)=v$.
Similarly, for a $C^2$ condition, a function $f$ is $g$-convex when the \textit{Riemannian Hessian} satisfies $\mathrm{Hess}\,f(x)\succeq 0$ for all $x\in U\subseteq \mathbb{R}^n$. The interest in $g$-convex functions stems from the fact that local minima are again global minima, as with standard convex functions.
We are now equipped to generalize Theorem~\ref{thm:cvx:Lyap}.
\begin{theorem}[Geodesically convex Lyapunov functions]
\label{thm:geod:cvx:Lyap}
Let $(\mathbb{R}^n,g)$ be a \textit{Riemannian manifold} and let $U\subseteq \mathbb{R}^n$ be open and $g$-convex. Let $F\in C^{0}(U;\mathbb{R}^n)$ give rise to $\dot{x}=F(x)$ with $0\in U$ globally asymptotically stable (GAS) (on $U$) under $F$. Then, if there is a g-convex $C^{\infty}$ Lyapunov function asserting $0$ is GAS, the vector field $F$ is straight-line homotopic to $\mathrm{Exp}^{-1}(0)$ such that $0$ is GAS throughout the homotopy.
\end{theorem}
In Theorem~\ref{thm:geod:cvx:Lyap}, $\mathrm{Exp}^{-1}(0)$ should be understood as the map being defined by $x\mapsto \mathrm{Exp}_x^{-1}(0)\in T_x\mathbb{R}^n$.
\begin{proof}
By assumption, there is a $C^{\infty}$ Lyapunov function $V$ such that $\langle \nabla V(x),F(x) \rangle <0$ for all $x\in U\setminus\{0\}$. By the $g$-convexity of $V$ we also know that for any $(x,v)\in TU$
\begin{equation*}
V(\mathrm{Exp}_x(tv))\geq V(x)+t\langle \mathrm{grad}\,V(x),v\rangle_g= V(x) + t\langle \nabla V(x),v\rangle \quad \forall t\in [0,1],
\end{equation*}
where we removed the dependency on the metric $g$ by identifying both inner products with the directional derivative $DV(x)[v]$. We consider $t=1$ and pick $v:=\mathrm{Exp}_x^{-1}(0)$. This map is always well-defined since our manifold $(\mathbb{R}^n,g)$ is geodesically complete by construction and all geodesics are unique, implying that the exponential map $\mathrm{Exp}_x:T_x\mathbb{R}^n\to \mathbb{R}^n$ must be a global diffeomorphism. Now we proceed exactly as in the proof of Theorem~\ref{thm:cvx:Lyap} and conclude.
\end{proof}
Indeed, we recover Theorem~\ref{thm:cvx:Lyap} for the identity metric on $\mathbb{R}^n$ and $U=\mathbb{R}^n$. In particular, in that case we can define our Riemannian exponential map as $\mathrm{Exp}_x(v)=x+v$ for (sufficiently small) $v\in T_xU$. Hence, the tangent vector $v$ such that $\mathrm{Exp}_x(v)=0$ is simply $-x$ (now seen as a tangent vector),~\textit{i.e.}, $\mathrm{Exp}_x^{-1}(0)=-x$. Recall, formally speaking, $-\mathrm{id}_{\mathbb{R}^n}$ should be understood as $x\mapsto (x,-x)\in TU$ while ignoring the first component of the image. With this in mind we can again understand $\mathrm{Exp}^{-1}(0)$ as the canonical ``\textit{inward}'' vector field, yet now on a subset of $(\mathbb{R}^n,g)$.
Generalizing to compact manifolds and so forth (beyond contractible sets) is somewhat nonsensical as no smooth function function with a single critical point exists on those spaces. This restriction comes from the demand that our geodesics are unique, obstructing nontrivial topologies. See~\cite[Chapter~4]{ref:udriste2013convex} for more pointers.
A similar generaliztion can be achieved through the lens of \textit{contraction analysis}~\cite{ref:lohmiller1998contraction}. See in particular~\cite{ref:wensing2020beyond} for a relation between $g$-convexity and contraction metrics.
We end this section by returning to Remark~\ref{rem:suff}, the Lyapunov function with respect to~\eqref{equ:ahmadi} is nonconvex, yet the dynamical system satisfies the necessary condition~\eqref{equ:cvx:L:nec}. Indeed, the function \textit{is} $g$-convex\footnote{In fact, the function $x\mapsto \log(1+x^2)$ is also semiconcave.} (under quadrant-wise exponential geodesics), and the necessary condition effectively extends (as inferred from Theorem~\ref{thm:geod:cvx:Lyap}).
\subsection{Continuation}
The existence of a mere homotopy is not immediately informative. Often, only when the homotopy itself satisfies certain properties, one can draw nontrivial conclusions.
In our case the homotopies as detailed in Theorem~\ref{thm:cvx:Lyap}, Corollary~\ref{cor:cvx:control:Lyap}, Corollary~\ref{cor:cvx:Lyap:A} and Theorem~\ref{thm:geod:cvx:Lyap} all preserve qualitative properties of the underlying dynamical system. More formally, this construction provides a \textit{continuation} in the sense of Conley, albeit from a different perspective. Again, we are decidedly brief, but we point the reader to~\cite{ref:conley1978isolated,ref:mischaikow2002conley} for more details on Conley index theory and suggest~\cite{Hatcher} as a reference on algebraic topology.
Recall that under our standing assumptions a dynamical system of the form~\eqref{equ:dyn:sys} gives rise to a global flow $\varphi:\mathbb{R}\times \mathbb{R}^n\to \mathbb{R}^n$. Let $S\subset \mathbb{R}^n$ be an \textit{isolated invariant set} (with respect to $\varphi)$, that is,
\begin{equation*}
S = \mathrm{Inv}(M,\varphi):=\{x\in M:\varphi(\mathbb{R},x)\subseteq M\}\subseteq \mathrm{int}(M)
\end{equation*}
for some compact set $M\subset \mathbb{R}^n$. Note that not every invariant set is isolated,~\textit{e.g.} consider an equilibrium point of the \textit{center}-type. Then, a pair of compact sets $(N,L)\subset \mathbb{R}^n\times \mathbb{R}^n$ is an \textit{index pair} for $S$ when
\begin{enumerate}[({I}-i)]
\item $S=\mathrm{Inv}(\mathrm{cl}(N\setminus L),\varphi)$ and $N\setminus L$ is a neighbourhood of $S$;
\item $L$ is positively invariant in $N$;
\item $L$ is an exit set for $N$ (a trajectory that leaves $N$, must leave through $L$).
\end{enumerate}
Now, the (homotopy) \textit{Conley index} of $S$ is the homotopy type of the pointed (quotient) space $(N/L,[L])$,~\textit{e.g.}, for $N=\mathbb{B}^n$, $L=\partial \mathbb{B}^n=\mathbb{S}^{n-1}$, we have that $N/L\simeq \mathbb{S}^n$ such that $(N/L,[L])$ is the pointed $n$-sphere. As this object is hard to computationally work with, let $H^k(A,B;\mathbb{Z})$ denote the $k^{\mathrm{th}}$ singular cohomology group of $A$ relative to $B\subseteq A$, then, the \textit{homological} \textbf{\textit{Conley index}} defined as $\mathrm{CH}^k(S,\varphi):=H^k(N/L,[L];\mathbb{Z})$ is of larger interest,~\textit{e.g.}, as computational tools \textit{are} available~\cite{ref:kaczynski2004computational}. Going back to our setting, assume for that moment that $0\in \mathbb{R}^n$ is a GAS hyperbolic fixed point of the flow $\varphi$. By hyperbolicity (local linearity), we can pick $N=\varepsilon\mathbb{B}^n$ (a sufficiently small closed ball in $\mathbb{R}^n$) and $L=\emptyset$. Now see that
\begin{equation*}
\mathrm{CH}^k(0,\varphi)=H^k(\varepsilon\mathbb{B}^n/\emptyset,[\emptyset];\mathbb{Z})\simeq H^k(\varepsilon\mathbb{B}^n;\mathbb{Z}) \simeq \begin{cases}
\mathbb{Z}\quad & \text{if }k=0\\
0\quad & \text{otherwise}
\end{cases}
\end{equation*}
since $\varepsilon\mathbb{B}^n$ is homotopic to a point. If $0$ is not hyperbolic, pick $N$ to be a sublevelset of a smooth Lyapunov function that asserts $0$ is GAS, this set is compact by Property~(V-\ref{prop:iii:V}). Indeed, constructions like these provide for topological obstructions~\cite{ref:moulay2011conley}.
Now, if some $N$ can be chosen to be an isolating neighbourhood \textit{throughout} a homotopy, then the Conley index is preserved along that homotopy~\cite[Theorem~1.10]{ref:mrozek1994shape}. Simply put, we speak in this case of a \textit{\textbf{continuation}} between the dynamical system at the beginning and the end of the homotopy. A question asked by Conley concerns the opposite~\cite[p.~83]{ref:conley1978isolated}, to what extent do equivalent Conley indices relate to the existence of such a continuation. See also the discussion in~\cite{ref:mrozek2000conley,ref:kvalheim2022obstructions}. Indeed, we see that if there is a homotopy through flows $[0,1]\ni\lambda\mapsto \varphi_{\lambda}$ such that $0$ is GAS along the homotopy, then $\mathrm{CH}^k(0,\varphi_0)\simeq \mathrm{CH}^k(0,\varphi_1)$.
For the other direction, based on the above we have the following. One can extend the statement to compact convex sets or $g$-convexity if desired.
\begin{corollary}[On continuation and convex Lyapunov functions]
Let $0\in \mathbb{R}^n$ be GAS under two dynamical systems of the form~\eqref{equ:dyn:sys} parametrized by $F_0$ and $F_1$, giving rise to the flows $\varphi_0$ and $\varphi_1$. Assume that $0$ being GAS is asserted by---possibly different---smooth convex Lyapunov functions $V_0$ and $V_1$. Then $0$ (with respect to $\varphi_0$) and $0$ (with respect to $\varphi_1$) are related by continuation where $N$ can be chosen to be of the form $N=N_0\cup N_1$ (based on sublevelsets of $V_0$ and $V_1$).
\end{corollary}
A further study of this observation is the topic of future work.
To return to similarities pointed out in the introduction, the work by Reineck~\cite{ref:reineck1991continuation} and the proof of \cite[Theorem~11.4]{ref:coron2007control} provide the homotopy (preserving the Conley index) between $F$ and the (a) negative gradient flow $-\nabla V$. However, how to link---if at all--- multiple dynamical systems is unclear. The book by Cieliebak \& Eliashberg does contain results in this direction, yet under $C^k$-nearness assumptions~\cite[Chapter~9]{ref:cieliebak2012stein}, not in general.
Then, this work alludes to convexity being a simple structural ingredient to actually link several dynamical systems together via some canonical dynamical system.
\section{Conclusion}
\label{sec:conclusion}
To return to Section~\ref{sec:stab:project}, the space of dynamical systems over $\mathbb{R}^n$ with $0$ being GAS subject to the existence of a convex Lyapunov function is path-connected and in that sense practically convenient.
To elaborate, the necessary conditions as set forth in Section~\ref{sec:cvx} allow one to check if a problem at hand is at least not infeasible,~\textit{e.g.}, when searching for $g$-convex (control) Lyapunov functions. In the context of policy optimization, when starting with a pre-stabilized system (a controller that at least ``\textit{works}''), one likes to be sure that the pre-stabilized system can be safely transformed to the optimally controlled system throughout the optimization/learning process. Indeed, this desire boils down to the initial and final system being homotopic through qualitatively equivalent systems. There is work in this area in the context of linear optimal control~\cite{ref:bu2019lqr,ref:ECCJongeneelKuhn21}, but further extensions are largely lacking.
This work focused on the complete(d) $C^{0}$ setting with the emphasis on Euclidean space, future work aims at studying dynamical control systems under weaker regularity assumptions in more general spaces.
Also, this work focused on the exploitation of a $g$-convex structure, however, more general structures have been proposed and studied,~\textit{e.g.}, a compositional structure~\cite{ref:Grune2021}. It seems worthwhile to study more structural assumptions along the lines of this note and previous work by Aeyels \& Sepulchre~\cite{ref:sepulchre1996homogeneous}.
\paragraph*{Acknowledgements} The authors are grateful to Matthew Kvalheim and Matteo Tacchi for their feedback.
\pagestyle{basicstyle}
\addcontentsline{toc}{section}{Bibliography}
\subsection*{Bibliography}
\printbibliography[heading=none]
\end{document}
|
\section{Introduction}\label{secintorf}
As identified by Shannon and Weaver \cite{6773024}, communications can be categorized into three levels:
i) transmission of symbols;
ii) transmission of semantic information behind transmitted symbols;
iii) effectiveness of semantic information interaction.
In 1950s, Shannon focused on the first level and derived a rigorous mathematical theory of communications based on probabilistic models, yielding the famous Shannon capacity limit \cite{thomas2006elements}.
In the past decades, wireless communication systems based on the Shannon information theory have evolved from the first generation (1G) to the fifth generation (5G). Various advanced technologies, including massive multiple-input multiple-output (MIMO) \cite{8354789}, mmWave communications \cite{8443598}, and deep learning (DL) based communications \cite{8663966}, continuously promote the system capacity to Shannon capacity limit.
In particular, DL has been recognized as a promising candidate for beyond-5G (B5G) communications due to its remarkable
capability of learning intricate inter-relationships hidden in massive data.
Great progresses have been achieved by DL in physical layer communications such as channel estimation \cite{9288911,8353153,8795533,9175003}, data detection \cite{8052521,9791486,9018199},
channel feedback \cite{8482358,guo2019convolutional,9136588}, and beamforming
\cite{9279228,liu2021deep,9112250}, etc.
By either modifying or replacing conventional communication modules,
these studies exploit deep neural networks (DNNs) to improve system accuracies or reduce computational complexities.
Despite of continuously emerging DL based communication applications,
the Shannon capacity limit is still the insurmountable upper bound.
Driven by the ambitious goals of the sixth generation (6G) and empowered by the development of artificial intelligence, a higher level communication paradigm, i.e., semantic communications, has been proposed very recently, which no longer focuses on accurately recovering the transmitted symbols but concerns on precisely recovering the meaning behind the transmitted symbols \cite{9679803,9530497}. The term ``semantic information'' refers to the information/meaning associated with the source signals.
Semantic information exists in various signal modalities like texts, speeches, images, videos, and even random variables etc \cite{Kalfaee,MalikMalik,8100143,8844998}.
Studies on semantic communications in the area of computer vision (CV) and natural language processing (NLP) have attracted great attentions.
The paradigm of semantic communications is to extract semantic information, transmit semantic information, and ensure semantic information being correctly interpreted by the receiver.
Several recent works \cite{9398576,9632815,9450827,9685667} have provided concrete examples of semantic text \cite{9398576,9632815}, speech \cite{9450827}, and image \cite{9685667} transmissions.
In \cite{9398576}, the authors propose a DL based semantic communication architecture
for text transmission, where the Transformer structure \cite{vaswani2017attention} is adopted in the semantic encoder/decoder.
In the training process, the semantic accuracy and the system capacity are jointly
optimized, while in the testing process the word-error-rate and the peak signal-to-noise ratio (PSNR) are used to measure the accuracy of the source information recovery.
The corresponding simulations demonstrate the superiority of the semantic architecture over traditional approaches.
Following the semantic communication architecture in \cite{9398576}, the work \cite{9632815} introduces an adaptive circulation mechanism into the Transformer structure \cite{vaswani2017attention}.
With the adaptive Transformer,
the semantic communication system becomes more flexible to transmit sentences with different semantic
information and exhibits better robustness over channel conditions.
In \cite{9685667}, the authors propose a DL based image semantic coding algorithm under a rate-perception-distortion optimization framework, which can achieve the state-of-the-art visually pleasing reconstruction and semantic preserving performance in the extreme low bit rate case.
{In fact, all the existing semantics communication systems only utilize the semantics of source signals to enhance the information representation power of the transmission bits, which can be more precisely referred to as source-oriented semantic communications (SOSCs).} The SOSC systems aim to promote the achieve rate reaching the Shannon capacity limit, which is consistent with the goal of enhanced mobile broad band (eMBB).
Meanwhile, how to realize ultra-reliable and low-latency communications (URLLCs) is also an essential topic for the implementation of an efficient practical system. Since mmWave bands have abundant spectrum resources and support wider subcarrier spacing, mmWave communication systems can better meet low latency requirements in URLLCs.
However, mmWave communication system is inherently unreliable due to its sensitivity to blockage and high penetration loss. To explore mmWave systems in URLLCs, the acquisitions of channel, interference, and blockage information are critical, especially for decisions like the beam switching or the base station (BS) hand-off.
Intrinsically, channels are determined by the wireless propagation environment, while the propagation environment can be captured by the images/visions from the cameras in cars or streets.
Hence, one may extract the channel semantic information from environment images, named as environment semantics or channel semantics, to help channel related downstream tasks.
By exploiting environment semantics, channel related downstream tasks could be accomplished without conventional channel estimation, which can significantly save the time and the computation cost related to the pilot training or the channel feedback, and even predict burst channel interference to improve the communication reliability.
Actually, some works have already proposed to use environment images to help channel related downstream tasks, e.g., beam selection \cite{weihua20192d}, channel covariance matrix estimation \cite{9523557}, and blockage prediction \cite{9512383}. These vision based works demonstrate that channel information could be effectively attained from environment images to build communication connections without pilot overheads.
However, original image usages for channel related tasks require expensive storage space, high computational cost, and huge energy consumption, which is unacceptable for the base stations (BSs) or users.
Moreover, environment image usages may involve complex social issues such as user privacy, public safety, and management policies, especially when the third party cameras (e.g., the ubiquitous surveillance cameras) are employed to provide the environment images.
Compared with directly utilizing environment images, operating with environment semantics naturally protects privacy. This is because that only the class and the layout information of objects are preserved in environment semantics, and the surface characteristics of objects are eliminated.
Furthermore, environment semantics is associated with various objects, including key scatterers of propagation channels and objects that are irrelevant to propagation channels. By retaining channel relevant semantics and mitigating channel irrelevant semantics, the system overheads such as storage space and computational cost could be also reduced.
The environment semantics aided communication system focuses on exploring channel semantics hidden in environment images, and therefore can be regarded as channel-oriented semantic communications (COSCs)
In this paper, we propose a framework for COSCs,
where environment semantics is extracted from environment images, selectively encoded based on its channel-relevance, and then fused to make decisions for specific channel related downstream tasks.
To demonstrate the effectiveness of the proposed framework, we develop an environment semantics aided network architecture for mmWave beam prediction and blockage prediction as a case study. The environment semantics aided network architecture is composed of an environment semantics extraction network, a feature selection (FS) algorithm, a task-oriented encoder network, and a decision network.
With images taken from street cameras, the environment semantics is obtained, selected, encoded, and then transmitted to BS.
After fusing the encoded channel semantics and identification information of the target user, BS can obtain the predicted beam index and blockage state by the decision network.
The environment semantics aided beam prediction can realize extremely low-latency without pilot training or the costly beam scans, while the environment semantics aided blockage prediction can support ultra-reliable communication links. Therefore, the proposed environment semantics aided framework offers extraordinary application values for URLLCs.
Simulation results based on the autonomous driving and the ray-tracing channel simulators demonstrate the effectiveness of the proposed environment semantics aided communication framework, which further validates its great application potential in URLLCs.
The rest of this paper is organized as follows.
The frameworks of COSCs and SOSCs are proposed in Section~\ref{secsena}.
As a case study, the environment semantics aided network architecture for mmWave beam prediction and blockage prediction is developed in Section~\ref{seccase}. Numerical simulations are provided in Section \ref{secsimu}. The main conclusions are
given in Section \ref{secconcul}.
\emph{Notation:}
The bold and lowercase letters denote vectors;
the notation $(\cdot)^T$ denotes the transpose of a matrix or a vector;
the notations ${\mathds{C}^{m\times n}}$ and ${\mathds{R}^{m\times n}}$ represent the $m\times n$ complex and the $m\times n$ real vector space, respectively;
the notation $\left\| {\bm x} \right\|_2$ denotes the $L_2$ norm of $\bm x$;
the notation $| {\bm x} |$ denotes the $L_1$ norm of $\bm x$ while $| X|$ denotes the size of the set $X$.
\section{Source Semantics and Channel Semantics}\label{secsena}
In this section, we will introduce
SOSCs,
and then propose a framework of environment semantics aided communication systems as a representative instance of
COSCs.
A communications system working with both SOSCs and COSCs will be named as a generalized semantics communications systems. We will then heuristically analyze the source semantics and channel semantics.
\subsection{Existing Semantic Communication Systems: SOSCs}
Fig.~\ref{figsemantic} illustrates the frameworks of a classical (the upper part) and an existing semantic (the bottom part) communication systems.
In a classical communication system, signals with various modalities are first compressed by a source encoder to save transmission cost. Then, the designed redundancies are added by a channel encoder to combat physical channel
noises in transmission links, and the received signals experience the inverse process in the receiver side.
Unlike classical communications whose goal is to ensure each transmitted bit being correctly received, the
SOSC systems \cite{9398576,9632815,9450827,9685667} aim to guarantee the meaning behind the transmitted bits, i.e., the semantic information of the transmitted signals, being correctly understood by the receiver.
Details about the modules of the existing semantic communication systems are given in the following.
\noindent \textbf{Semantic encoder} extracts semantic features from the source signals and compresses them to save the transmission cost. Next, the compressed semantic features go through a channel
encoder and are transmitted over physical channels.
\noindent \textbf{Semantic decoder} interprets the received semantic features and recover the original source signals, i.e., conducting the inverse process of the semantic encoder.
\noindent \textbf{Background knowledge base (BKB)} is the prerequisite of semantic communication systems, which can be independently established or interactively shared by the transmitter and the receiver. The semantic encoder extracts semantic features based on the background knowledge at the transmitter, while the semantic decoder interprets semantic features and restores the original source signals based on the background knowledge at the receiver.
The BKBs can learn from previous signal data and periodically update with the gradual progress of communications.
Naturally, BKBs are different in various contexts. For example, high-frequency words in business and medical scenarios are different, and therefore the corresponding BKBs contain different semantic corpora. Moreover, BKBs should have different formats for signals in various modalities, e.g., text, image, and speech, etc.
Besides, BKBs at the transmitter and the receiver might be different due to cognitive bias or belated sharing.
\noindent \textbf{Semantic noise} comes from semantic ambiguity of received signals or mismatches between the BKBs at the transmitter and the receiver, resulting in semantic errors in the interpretation and recovery processes. Note that the physical channel noise in transmission environments might lead to symbol errors of the received signals, which can also cause semantic errors.
\begin{figure*}[!t
\centering
\includegraphics[width=160mm]{semantic.eps}
\caption{Frameworks of a classical (the upper part), an existing semantic (the bottom part), and an environment semantics aided (the middle part) communication systems. The environment semantics aided and the existing semantic communication systems are the instances of COSCs and SOSCs, respectively. }
\label{figsemantic}
\end{figure*}
\subsection{Environment Semantics Aided Communication System: COSCs}
The framework of the environment semantics aided communication system is displayed in the middle part of Fig.~\ref{figsemantic}, where semantic information is extracted from environment images for channel related downstream tasks, e.g., beam prediction, blockage prediction, and codebook designs, etc.
We refer the semantic information extracted from environment images as environment semantics or the channel semantics, which naturally protects user privacy by eliminating the surface characteristics of objects and preserving only the class and the layout information of objects.
Since the environment semantics aided communication system exploits the channel semantics hidden in the environment images to assist channel related downstream tasks, it can be regarded as an instance of COSCs.
To illustrate the idea, suppose the goal of environment semantics aided communication systems is to predict the parameter $Y$ from a set of multimodal observations/samples, including environment images \cite{9277535}. Here $Y$ can be the beam index/vector in beam prediction tasks, the blockage flag in blockage prediction tasks, the codebook in codebook design tasks, or the channel covariance matrix in channel covariance matrix estimation tasks, etc.
We refer features extracted from other relevant modalities as \emph{auxiliary features}, which can help to identify target users or provide extra system information.
Denote the universal set of the extracted features as $X_{\textrm{uni}}$.
By utilizing FS methods, the most efficient feature set
could be selected from $X_{\textrm{uni}}$.
Then, the task-oriented encoder maps $X$ into $\tilde{X}$ to save system overheads (e.g., storage space, transmission delay, and energy consumption, etc).
The compressed semantic feature set $\tilde{X}$ is sent to the fusion module for further decisions in channel related downstream tasks.
\subsection{Analyses of Source Semantics and Channel Semantics}
Signals and channels are the two major components of communication systems.
Existing semantic communication systems \cite{9398576,9632815,9450827,9685667} are SOSCs that focus on the semantics of the source signals to enhance the representation capacity of transmitted bits. The proposed environment semantics aided communication system belongs to COSCs that extract the channels semantics from multimodal data like environment images to assist the accomplishment of channel related downstream tasks without pilot training.
In other words, SOSCs improve the representation capacity of transmitted bits by utilizing the source signal semantics, i.e., pursuing the goal of eMBB.
In contrast, COSCs improve the transmission efficiency by exploiting the channel semantics hidden in multimodal data, i.e., targeted at the requirements of URLLCs. In the environment semantics aided communication system, channel semantics refers to the shape and classes of channel relevant scatterers, which can be obtained from the environment images.
Besides, the proposed environment semantics aided communication system
is a concrete realization of task-oriented communications, where the module interactions are carried out to accomplish a specific task while the information not strictly relevant to the task would be mitigated. \cite{strinati20216g}.
\section{Case Study of Channel Semantics: MmWave Beam Prediction and Blockage Prediction }\label{seccase}
In this section, we develop an environment semantics aided network architecture for mmWave beam prediction and blockage prediction to meet requirements of URLLCs as
a case study.
\subsection{System Model}
Consider a massive MIMO system, where a BS is equipped with $N_t\!\gg\! 1$ antennas in the form of uniform linear array (ULA) and serves multiple single-antenna users.
The received frequency domain signal of a user on the $k$-th subcarrier is
\begin{equation}\label{equreceive}
r[k]=\bm h ^{T}[k] \bm w s[k] + \varepsilon [k],
\end{equation}
where $r[k]\in \mathds{C}^{1\times 1}$ is the received signal, $ s[k]\in \mathds{C}^{1\times 1}$ is the transmit signal, $\bm w \in \mathds{C}^{N_t\times 1} $ is the transmit beamforming vector, and
$ \varepsilon[k]\!\!\sim\!\! \mathcal{CN}\left(0, \sigma_{\varepsilon}^{2} \right) $ is the additive white Gaussian noise.
Moreover, $\bm h [k]\in \mathds{C}^{M\times 1}$ is the downlink channel that can be written as \cite{alkhateeb2019deepmimo}
\begin{equation}\label{equchannel}
\bm h [k] = \sum\nolimits_{l} {\alpha _{ l}{e^{ - j2\pi f_{\textrm{D},k}\tau_{ l} + j{\phi _{ l}}}} \bm a\left( {\theta_{\mathrm{az}}^{ l}, \theta_{\mathrm{el}}^{ l}} \right)},
\end{equation}
where $f_{\textrm{D},k}$ is the frequency of the $k$-th downlink subcarrier, while $\alpha _{ l}$, $\phi _{ l}$, and $\tau _{ l}$ are the attenuation, phase shift, and delay
of the $l$-th path, respectively. In addition, $\bm a\left( {\theta_{\mathrm{az}}^{ l}, \theta_{\mathrm{el}}^{ l}} \right)$ is the array manifold vector\footnote{When other types of antenna arrays are adopted, the array manifold vector should be changed accordingly. Note that the proposed network architecture is not limited by specific antenna array shape, and therefore is applicable for array with arbitrary geometry.} defined as
\begin{eqnarray}\label{equavec}
\bm a\left( {\theta_{\mathrm{az}}^{ l}, \theta_{\mathrm{el}}^{ l}} \right){\kern -8pt}&={\kern -8pt}&\left[1, e^{j\varpi \sin \left(\theta_{\mathrm{el}}^{ l}\right) \cos \left(\theta_{\mathrm{az}}^{ l}\right)}, \ldots\right. \nonumber \\
& &{\kern 5pt}\left.\ldots, e^{j \varpi\left(M-1\right) \sin \left(\theta_{\mathrm{el}}^{ l}\right) \cos \left(\theta_{\mathrm{az}}^{ l}\right)}\right]^{T},
\end{eqnarray}
where $\varpi ={{2\pi d f_{\textrm{D},k}}}/{c }$, $d$ is the antenna spacing, $c$ is the speed of light, and $\{\theta_{\mathrm{az}}^{ q}, \theta_{\mathrm{el}}^{ l}\}$ is the \{azimuth, elevation\} angle of arrival.
Assume the beamforming vector $\bm w $ is selected from the beam codebook $\mathcal{W}$, where $M_{\textrm{bm}}=|\mathcal{W}|$ is the codebook size.
The achievable transmission rate corresponding to the beam vector ${\bm w}$ can be written as
\begin{eqnarray}\label{equarate}
\textrm{Rate}_{\bm w}= \frac{1}{K} \sum_{k=1}^{K} \log _{2}\left(1+\frac{P_{k}}{\sigma^{2}}\left|\bm h ^{T}[k] \bm w\right|^{2}\right),
\end{eqnarray}
where $P_{k}=E[\left|s [k]\right|^{2}]$ is the power of the transmitted signal\footnote{To simplify the notation, we drop the sub-carrier index $k$
in the rest of the paper, e.g., replacing $\bm r[k], \bm h_{u}(f_{\textrm{D},k})$, and $\bm s[k]$
with $\bm r, \bm h(f_{\textrm{D}})$, and $\bm s$, respectively.}.
The optimal beamforming vector can be obtained by maximizing the transmission rate, i.e.,
\begin{eqnarray}\label{equaraterhh}
\bm{w}^{\textrm{opt}}=\underset{\bm w \in \mathcal{W}}{\arg \max }\ \textrm{Rate}_{\bm w}.
\end{eqnarray}
In fact, the beamforming vector $\bm w $ is designed to concentrate the transmitter power on a narrow beam to compensate for the high penetration loss of mmWaves. To obtain the optimal beamforming vector $\bm{w}^{\textrm{opt}}$, conventional algorithms need to estimate accurate channel information by extensive downlink pilot training, thus leading to unacceptable time delay especially for massive MIMO systems. Besides, exhaustively searching on the a large codebook also greatly increases the time delay, which makes it impractical to apply conventional algorithms to URLLCs.
In the meantime, based on field measurements \cite{9689054}, the channel power is generally concentrated in the first few paths, especially the line-of-sight (LOS) path.
Therefore, LOS link blockages caused by obstacles would significantly reduce the achievable transmission rate, and even lead to link disconnections. To ensure the link reliability of mmWave communications, the prediction of future link blockages is critical to beam/BS switching decisions.
\begin{figure*}[!t
\centering
\includegraphics[width=160mm]{model.eps}
\caption{The overall diagram of the environment semantics aided communication system. Here two cameras are schematically installed in the deployment scenario. In fact, the number and positions of the cameras can be set according to practical conditions and demands, which will not affect the algorithm design. }
\label{fignetmodel}
\end{figure*}
\subsection{Problem Formulation}
As indicated in Eq.~\eqref{equchannel}, millimeter massive MIMO channels have sparse structures associated with the parameters such as attenuations, phase shifts, and delays, etc. Naturally, these channel
parameters can be regarded as one type of semantic information. However, from an intrinsical perspective, wireless channels are determined by the propagation paths or more straightforward by the key scatters in propagation environment, e.g., vehicles or roads. Therefore, there exists channel related semantic information in corresponding environment images that could be extracted to improve the channel related tasks, e.g., the beam prediction, blockage prediction, hybrid precoding, and codebook design, etc.
Suppose that cameras, i.e., the sensing devices installed on both sides of streets, can capture image data flows to assist BS in channel related tasks. As shown in Fig.~\ref{fignetmodel},
different cameras can provide image data from different perspectives, thus avoiding the target user being blocked by other vehicles nearby.
Since images from cameras are municipal infrastructure data, which involve complex social issues such as user privacy, public safety, and management policies, etc, they should not be directly sent to the fusion center for communication services.
By extracting environment semantics from images, the surface characteristics of buildings, vehicles, or roads, etc, are eliminated, while only the class and the layout information of objects are preserved. Therefore, environment semantics is naturally more private than original images, and facilitates potential new privacy transmission protocols.
Meanwhile, user's identification information, such as locations, historical beam indices, or other predetermined information that could be used to distinguish users, should also be sent to the fusion center in a secure and private way\footnote{If restricted, the feature extraction could be independently implemented by terminal devices without communications with subsequent communication tasks. More details about how to strictly protect user privacy involve specific device transmission protocols and encryption methods, which are possibly left as future work. }.
We refer the features extracted from users' identification information and environment semantics from cameras as auxiliary and semantic features, respectively.
Raw multimodal data from communication or sensing devices are first transformed into auxiliary or semantic features, and then selectively encoded by the task-oriented encoder.
After fusing the encoded features from the target user and the cameras on streets, optimal decisions can be obtained by the designed network at BS.
\begin{figure*}[!t
\centering
\includegraphics[width=165mm]{network.eps}
\caption{Diagram of the environment semantics aided network architecture for beam prediction and blockage prediction.}
\label{fignetwork}
\end{figure*}
\subsection{Environment Semantics Aided Network Architecture for Beam Prediction and Blockage Prediction}
Fig.~\ref{fignetwork} displays the diagram of the environment semantics aided network architecture for beam prediction and blockage prediction, consisting of the environment semantics extraction network, the FS algorithm, the task-oriented encoder, and the decision network.
Specifically, the environment semantics extraction network is deployed independently at each camera to extract the environment semantics from the corresponding environment images.
The output of the environment semantics extraction network is the segmentation map, i.e., the pixel-wise category labels of the environmental semantic concepts (e.g., ``road'', ``vehicle'', ``trafficlight'', and ``sky'', etc).
After multiplying the segmentation map by the corresponding zero-masks, one can obtain maps of various environmental semantic concepts, as shown in the lower left quarter of Fig.~\ref{fignetwork}.
We refer a map of each environmental semantic concept as an instance of one semantic feature.
The universal feature set $X_{\textrm{uni}}\buildrel \Delta \over =\{X_{\textrm{uni},v}| i=1,\cdots,V_{\textrm{uni}} \} $ contains both the semantic features from environment images and the auxiliary features from users' identification information, where $V_{\textrm{uni}}=|X_{\textrm{uni}}|$ denotes the size of $X_{\textrm{uni}}$.
Note that many types of user's identification information, such as locations, historical beam indexes, or other predetermined information, could all be added to $X_{\textrm{uni}}$ for further selection. In this work, we adopt the location as an example.
Then, we utilize the FS algorithm to select the optimal feature sets, i.e., \[X_{\textrm{bm}}\buildrel \Delta \over =\{X_{\textrm{bm},v}| i=1,\cdots,V_{\textrm{bm}},X_{\textrm{bm},v}\in X_{\textrm{uni}} \} \] and \[X_{\textrm{bl}}\buildrel \Delta \over =\{X_{\textrm{bl},v}| i=1,\cdots,V_{\textrm{bl}},X_{\textrm{bl},v} \in X_{\textrm{uni}} \},\] from $X_{\textrm{uni}}$ for beam prediction and blockage prediction, where ${V_{\textrm{bm}}}$ and ${V_{\textrm{bl}}}$ are the sizes of $X_{\textrm{bm}}$ and $X_{\textrm{bl}}$, respectively.
Since the user's identification information is necessary to distinguish the target user from other possible users,
the user location would definitely be retained by the FS algorithm as the auxiliary feature for both beam prediction and blockage prediction\footnote{When more types of users' identification information are involved, which type of users' identification information is retained depends on the FS algorithm.}. For clarify, we denote the location as the feature $X_{\textrm{bm},1} \buildrel \Delta \over = X_{\textrm{bl},1}$. Hence, $\{X_{\textrm{bm},v}\}_{v=2}^{V_{\textrm{bm}}}$ and $\{X_{\textrm{bl},v}\}_{v=2}^{V_{\textrm{bl}}}$ are the semantic features selected from the environment semantics for beam prediction and blockage prediction, respectively.
The task-oriented encoder is composed of the auxiliary, the semantic-beam, and the semantic-blockage modules.
The decision network includes the blockage-prediction and the beam-prediction modules.
With $X_{\textrm{bm}}$ as the input, the three modules, i.e., the auxiliary, the semantic-beam, and the beam-prediction modules, are jointly trained to predict the beam.
With $X_{\textrm{bl}}$ as the input, the three modules, i.e., the auxiliary, the semantic-blockage, and the blockage-prediction modules, are jointly trained to predict the blockage.
Since pilot training and costly beam scans are unnecessary, the environment semantics aided network architecture is expected to realize low-latency beam prediction and blockage prediction to ensure efficient and reliable mmWave communication links, so as to fulling the goal of URLLCs.
In the following, we will present the environment semantics extraction network, followed by the structures of above-mentioned modules.
Then, the FS algorithm for beam prediction and blockage prediction are presented respectively.
Lastly, the overall training steps of the environment semantics aided network architecture for beam prediction and blockage prediction are provided.
\begin{figure*}[!t
\centering
\includegraphics[width=150mm]{seg.eps}
\caption{Illustration of the environment semantics extraction. A green cuboid represents a feature map output by a basic residue block. The number next to the cuboid is the marked index of the feature map but does not necessarily correspond to the actual number of blocks. Cuboids in other colors represent feature maps of different size in the pyramid pooling module.
Notice that cuboids with smaller side areas typically are thicker in the horizontal direction. This is because that when images are down sampled, we tend to increase the filter number to avoid information loss, which is also a common trick in computer vision. }
\label{fignetseg}
\end{figure*}
\subsubsection{The environment semantics extraction network}
In semantic coded transmission work \cite{9685667}, a semantic-perceptual loss defined by the semantic segmentation network PSPNet \cite{8100143} is adopted for image compression and reconstruction, which is experimentally demonstrated to yield better reconstruction performance than common pixel-level losses at the same transmission bit rate.
The work \cite{9685667} also validates that the reconstructed images suffering less semantic information losses could achieve better performance in downstream tasks like object detection.
Inspired by this, we utilize PSPNet to extract environment semantics from environment images, as shown in Fig.~\ref{fignetseg}. Inheriting from the well-known ResNet \cite{7780459}, PSPNet is composed of a serial of basic residue blocks, and each residue block is sequentially stacked by several convolution, batch normalization (BatchNorm), and ReLu layers.
We refer the outputs of hidden layers as the feature maps, and mark them with serial numbers in Fig.~\ref{fignetseg} for illustration convenience.
The main difference of PSPNet compared with ResNet is the pyramid pooling module, where the $3$-rd feature map first
goes through three pooling layers with different down-sampling rates in parallel.
Then, the three down-sampled feature maps independently and successively go through a convolution and a BatchNorm layers, obtaining three feature maps with different pyramid scales, i.e., the \{4,5,6\}-th feature maps.
Next, the \{4,5,6\}-th feature maps are up-sampled by bilinear interpolations such that they can be concatenated together with the $3$-rd feature map as the final pyramid pooling global feature, i.e., the $7$-th feature map.
Note that the number of residue blocks, convolutional filters, and pyramid levels can all be adjusted on demand. The principle of the pyramid pooling module is that 1) feature maps with smaller scales has larger sizes of receptive fields\footnote{The receptive field is the locations in deeper layers correspond to the locations in the input image they are path-connected to.}; 2) layers with larger receptive fields could learn more about the global features of images while layers with smaller receptive fields could learn more about the local features of images;
3) the pyramid pooling module could learn a comprehensive feature of
images by fusing feature maps with varying scales.
Denote the input image as $\bm g \in \mathds{R}^{3\times H\times W}$, where ``3'' denotes the RGB channels, $H$ is the image height, and $W$ is the image width.
The label and the output of PSPNet are $\hat{\bm c} \in \mathds{R}^{ M_{\textrm{con}}\times H\times W}$ and $\bm c \in \mathds{R}^{ H\times W}$, respectively, where $M_{\textrm{con}}$ is the number of semantic feature categories, and the $(i,j)$-th entry of $\bm c$ is the category label of the environmental semantic concept at the pixel coordinate $(i,j)$, i.e., $\bm c(i,j)=0,\cdots,M_{\textrm{con}}-1$.
The loss function of the PSPNet is given by
\begin{align}\label{equdentsemp}
f_{\{\textrm{Lpsp}\}}=- \frac{1}{HW}\sum_{i,j}\log \left(\frac{\exp (\hat{\bm c}(\bm c(i,j),i,j))}{\sum\nolimits_{m=0}^{M_{\textrm{con}}-1} \exp (\hat{\bm c}(m,i,j))}\right).
\end{align}
By using the ADAM algorithm to minimize $f_{\{\textrm{Lpsp}\}}$ until the convergence,
the segmentation map can be obtained by
\begin{equation}\label{equaccfptr}
\bm c_\textrm{map} = \underset{m=0,\cdots,M_{\textrm{con}}-1}{\arg\max} \ \hat{\bm c}(m,i,j).
\end{equation}\label{equaccpsp}
The extraction accuracy of the environment semantics is then given by
\begin{eqnarray}\label{equaccpsp}
A_{\textrm{PSP}}= \frac{\sum_{i,j}\sum\nolimits_{n=1}^{N_{\textrm{tot}}}\mathds{1}(\bm c_\textrm{map}^{(n)}(i,j) =\bm c^{(n)}(i,j))}{HWN_{\textrm{tot}}} ,
\end{eqnarray}
where the superscript ${(n)}$ denotes the $n$-th sample, $\mathds{1}$ is the indicator function, and $N_{\textrm{tot}}$ is the total number of testing samples.
The accuracy $A_{\textrm{PSP}}$ measures the accuracy of the environment semantics extraction.
After obtaining the segmentation map, i.e., $\bm c_\textrm{map}$, the environment semantics such as ``road'', ``vehicle'',
``trafficlight'', ``pedestrian'', ``sky'', and ``building'', etc, can be easily separated by multiplying zero-masks.
\subsubsection{The auxiliary module}
With the location as the input\footnote{In this example work, we can prejudge that the user location would be the input of the auxiliary-module. When more types of users' identification information are involved, the design of the auxiliary module would be more complex and depends on which type of users' identification information is chosen by the FS algorithm.}, the auxiliary module is composed of several fully connected (FC) blocks, and each FC block is sequentially stacked by an FC, a BatchNorm, and a ReLu layers.
Denote the mathematical function of the auxiliary module as $f_{\{\textrm{en,a}\}}$ for future use.
Although we adopt the same network structure of the auxiliary module for both beam prediction and blockage prediction, the specific network parameters of the auxiliary module would be trained independently for beam prediction and blockage prediction.
\subsubsection{The semantic-beam module}
With the selected semantic features $\{X_{\textrm{bm},v}\}_{v=2}^{V_{\textrm{bm}}}$ as the input, the semantic-beam module is built by stacking several convolution-blocks, and each convolution-block contains a convolution, a BatchNorm, and a Relu layers in sequence.
The output size of the semantic-beam module can be changed by adjusting the number of convolution blocks, the number of filters in the convolution layer, or the down-sampling rate of the pooling layer.
Denote the mathematical function of the semantic-beam module as $f_{\{\textrm{en,sbm}\}}$ for future use.
\subsubsection{The semantic-blockage module}
With the selected semantic features $\{X_{\textrm{bl},v}\}_{v=2}^{V_{\textrm{bl}}}$ as the input, the semantic-blockage module has similar structure with the semantic-beam module, i.e., stacked by several convolution blocks.
Denote the mathematical function of the semantic-blockage module as $f_{\{\textrm{en,sbl}\}}$ for future use.
\subsubsection{The beam-prediction module}
The input of the beam-prediction module can be given as
\begin{eqnarray}\label{equfusing}
\bm x_{\{\textrm{bm},X_{\textrm{bm}}\}}=[f_{\{\textrm{en,a}\}}(X_{\textrm{bm},1}), f_{\{\textrm{en,sbm}\}} (\{X_{\textrm{bm},v}\}_{v=2}^{V_{\textrm{bm}}})].
\end{eqnarray}
With $\bm x_{\{\textrm{bm},X_{\textrm{bm}}\}}$ as the input, the beam-prediction module is composed of several FC, dropout, and Relu layers. Define the mathematical function of the beam-prediction module as $f_{\{\textrm{dec,bm}\}}$.
Then, the output of the beam-prediction module is
\begin{eqnarray}\label{equdecf}
\bm y_{\{\textrm{bm},X_{\textrm{bm}}\}} =f_{\{\textrm{dec,bm}\}}(\bm x_{\{\textrm{bm},X_{\textrm{bm}}\}}).
\end{eqnarray}
The cross entropy loss is adopted for beam prediction, which is given by
\begin{align}\label{equdentrop}
f_{\{\textrm{Lbm},X_{\textrm{bm}}\}}=-\log \left(\frac{\exp (\bm y_{\{\textrm{bm},X_{\textrm{bm}}\}}[y_{\{\textrm{bm,lab}\}}])}{\sum_{m=1}^{M_{\textrm{bm}}} \exp (\bm y_{\{\textrm{bm},X_{\textrm{bm}}\}}[m])}\right),
\end{align}
where $y_{\{\textrm{bm,lab}\}}$ is the beam label.
The auxiliary, the semantic-beam, and the beam-prediction modules are jointly trained by applying adaptive moment estimation (ADAM) algorithm \cite{kingmaadam} to minimize $f_{\{\textrm{Lbm},X_{\textrm{bm}}\}}$ until the convergence.
Denote the Top-$G$ beam-index set of the $n$-th testing sample as $Y_{\{G,X_{\textrm{bm}}\}}^{(n)}$, which contains the indices corresponding to $G$ maximum values of the output vector $\bm y_{\{\textrm{bm},X_{\textrm{bm}}\}}^{(n)} $.
Then, the Top-$G$ beam prediction accuracy can be obtained by
\begin{eqnarray}\label{equacckdcn}
A_{\{G,X_{\textrm{bm}}\}}=\sum\nolimits_{n}^{N_{\textrm{tot}}}\mathds{1}( y_{\{\textrm{bm,lab}\}}^{(n)}\in Y_{\{G,X_{\textrm{bm}}\}}^{(n)})/N_{\textrm{tot}}.
\end{eqnarray}
Denote the Top-$G$ beam-vector set as $\mathcal{W}_{\{G,X_{\textrm{bm}}\}}^{(n)} $, where each beam vector in $\mathcal{W}_{\{G,X_{\textrm{bm}}\}}^{(n)} $ is selected from the codebook $\mathcal{W}$ according to the beam index set $Y_{\{{G},X_{\textrm{bm}}\}}^{(n)}$.
The transmission rate corresponding to ${Y_{\{{G},X_{\textrm{bm}}\}}^{(n)}}$ can be written as
\begin{eqnarray}\label{equarate}
\textrm{Rate}_{Y_{\{{G},X_{\textrm{bm}}\}}^{(n)}} = \underset{\bm w \in \mathcal{W}_{\{G,X_{\textrm{bm}}\}}^{(n)} }{\max }\ \textrm{Rate}_{\bm w}.
\end{eqnarray}
Define the transmission rate ratio (TRR) of the Top-$G$ beam as the ratio between the transmission rate corresponding to ${Y_{\{{G},X_{\textrm{bm}}\}}^{(n)}}$ and the optimal transmission rate corresponding to the beam $\bm{w}^{\textrm{opt}}$ in Eq.~\eqref{equarate}, i.e.,
\begin{eqnarray}\label{equarate}
\textrm{TRR}_{\{G,X_{\textrm{bm}}\}} =\frac{1}{N_{\textrm{tot}}} \sum\nolimits_{n}^{N_{\textrm{tot}}} \frac{\textrm{Rate}_{Y_{{\{G,X_{\textrm{bm}}\}}}^{(n)}}} {\textrm{Rate}_{\bm{w}^{\textrm{opt}}}}.
\end{eqnarray}
It should be mentioned that with the Top-$G$ beam-vector set $\mathcal{W}_{\{G,X_{\textrm{bm}}\}}^{(n)} $, BS only needs to scan $G$ times to obtain the optimal beam. In this way, the environment semantics aided network architecture offers an ideal initial beam search range, and thus significantly reduces the delay caused by beam scans.
\subsubsection{The blockage-prediction module}
The input of the blockage-prediction module can be written as
\begin{eqnarray}\label{equfusing}
\bm x_{\{\textrm{bl},X_{\textrm{bm}}\}}=[f_{\{\textrm{en,a}\}}(X_{\textrm{bl},1}), f_{\{\textrm{en,sbl}\}} (\{X_{\textrm{bl},v}\}_{v=2}^{V_{\textrm{bl}}})].
\end{eqnarray}
With $\bm x_{\{\textrm{bl},X_{\textrm{bl}}\}}$ as the input, the blockage-prediction module is composed of several FC, dropout, and Relu layers. Define the mathematical function of the blockage-prediction module as $f_{\{\textrm{dec,bl}\}}$.
Then, the output of the blockage-prediction module is
\begin{eqnarray}\label{equdecf}
y_{\{\textrm{bl},X_{\textrm{bl}}\}} =\sigma( f_{\{\textrm{dec,bl}\}}(\bm x_{\{\textrm{bl},X_{\textrm{bl}}\}})),
\end{eqnarray}
where $\sigma(z)=1/(1+e^{-z})$ is the sigmoid function.
We adopt the binary cross entropy loss for blockage prediction, which is given by
\begin{align}\label{equdblockghg}
f_{\{\textrm{Lbl},X_{\textrm{bl}}\}}&=
y_{\{\textrm{bl,lab}\}} \log y_{\{\textrm{bl},X_{\textrm{bl}}\}} \nonumber \\
&{\kern 30pt} + \left(1-y_{\{\textrm{bl,lab}\}}\right) \log \left(1- y_{\{\textrm{bl},X_{\textrm{bl}}\}} \right),
\end{align}
where $y_{\{\textrm{bl,lab}\}}$ is the blockage label.
The auxiliary, the semantic-blockage, and the blockage-prediction modules are jointly trained by ADAM algorithm to minimize $f_{\{\textrm{Lbl},X_{\textrm{bl}}\}}$ until the convergence.
In the testing stage, the blockage accuracy can be obtained by
\begin{eqnarray}\label{equaccblcok}
A_{\{\textrm{bl},X_{\textrm{bl}}\}}=\sum\nolimits_{n}^{N_{\textrm{tot}}}\mathds{1}(y_{\{\textrm{bl},X_{\textrm{bl}}\}} ^{(n)}=y_{\{\textrm{bl,lab}\}}^{(n)})/N_{\textrm{tot}}.
\end{eqnarray}
\subsubsection{FS for beam prediction}
For simplicity, we adopt the Top-1 beam prediction accuracy as the optimization objective of FS.
Based on Eq.~\eqref{equacckdcn}, the Top-1 beam prediction accuracy with the feature set $X$ can be calculated by
\begin{eqnarray}\label{equacc}
A_{\{1,X_{\textrm{bm}}\}}=\sum\nolimits_{n}^{N_{\textrm{tot}}}\mathds{1}( y_{\{1,X_{\textrm{bm}}\}}^{(n)}= y_{\{\textrm{bm,lab}\}}^{(n)})/N_{\textrm{tot}},
\end{eqnarray}
where $y_{\{1,X_{\textrm{bm}}\}}^{(n)}$ is the predicted Top-1 beam of the $n$-th sample, i.e.,
\begin{eqnarray}\label{equrep}
y_{\{1,X_{\textrm{bm}}\}}^{(n)}=\underset{m=0,\cdots,M_{\textrm{bm}}-1}{\arg \max } \ \bm y_{\{\textrm{bm},X_{\textrm{bm}}\}} ^{(n)}[m] .
\end{eqnarray}
Originated from the SFFS algorithm, the proposed FS algorithm takes the universal feature set $X_{\textrm{uni}}$ as the input and then outputs the selected subset $X_{\textrm{bm}}$, where the mapping mechanism from the feature set $X_{\textrm{bm}}$ to the Top-1 beam prediction accuracy $A_{\{1,X_{\textrm{bm}}\}}$ serves as the optimization objective.
Denote the selected subset at the $i$-th iteration as $X_{\textrm{bm}}^{(i)}$.
The history feature set $X_{\{\textrm{bm,hist}\}}$ contains the history iterations of the selected subset.
We initialize both $X_{\textrm{bm}}^{(0)}$ and $X_{\{\textrm{bm,hist}\}}$ as empty sets.
The FS algorithm can be divided into two alternate steps, i.e., {Step 1} and {Step 2}, until a termination criterion is satisfied.
{Step 1} aims to find the most significant feature $x^{+}$ among the unselected features, i.e., the feature leading to the best accuracy increase for $X_{\textrm{bm}}^{(i)}$:
\begin{equation}\label{equxxplus}
x^{+} \leftarrow \underset{x \in X_{\textrm{uni}}-X_{\textrm{bm}}^{(i)} }{\arg\max} A_{\{1,X_{\textrm{bm}}^{(i)}+x\}}.
\end{equation}
After successively updating $X_{\textrm{bm}}^{(i+1)}$, $X_{\{\textrm{bm,hist}\}}$, and $i$ by
\begin{align}\label{equxxplusv}
X_{\textrm{bm}}^{(i+1)} &\leftarrow X_{\textrm{bm}}^{(i)} + x^{+},\\
X_{\{\textrm{bm,hist}\}}& \leftarrow \{X_{\{\textrm{bm,hist}\}},X^{(i)}\},
\end{align}
and $ i \leftarrow i+1$, respectively,
we go over to {Step 2} with the goal to remove the least significant feature in $X_{\textrm{bm}}^{(i)}$.
In {Step 2}, if there exists one feature in $X_{\textrm{bm}}^{(i)}$ that has a negative contribution to the accuracy, i.e., satisfying
\begin{equation}\label{equxx2if}
\underset{x \in X_{\textrm{bm}}^{(i)} }{\max} A_{\{1,X_{\textrm{bm}}^{(i)}-x\}} >A_{\{1,X_{\textrm{bm}}^{(i)}\}} ,
\end{equation}
then we will remove the feature $ x^{-}$ that results in the maximum accuracy drop, i.e., obtaining $ x^{-}$ by
\begin{equation}\label{equmina}
x^{-} \leftarrow \underset{x \in X_{\textrm{bm}}^{(i)} }{\arg\max} A_{\{1,X_{\textrm{bm}}^{(i)}-x\}}.
\end{equation}
\begin{algorithm}[!t]
\LinesNumbered
\algsetup{linenosize=\tiny} \small
\caption{FS algorithm for beam prediction }
\label{notr}
\KwIn{ Universal feature set $X_{\textrm{uni}}\buildrel \Delta \over =\{X_{\textrm{uni},v}| i=1,\cdots,V_{\textrm{uni}} \}$}
\KwOut {Selected feature set $X_{\textrm{bm}}\buildrel \Delta \over =\{X_{\textrm{bm},v}| i=1,\cdots,V_{\textrm{bm}},X_{\textrm{bm},v} \in X_{\textrm{uni}} \}$}
Initialize feature set $X_{\textrm{bm}}^{(0)} \leftarrow \emptyset $ \\
history feature set $X_{\{\textrm{bm,hist}\}}\leftarrow \emptyset $; $i=0$\\
\textbf{Step 1: (Inclusion)} \\
\quad\ \textbf{if} {$X_{\textrm{bm}}^{(i)} \notin X_{\{\textrm{bm,hist}\}}$} \textbf{then} \label{linedff}\\
\quad\ \quad\ $x^{+} \leftarrow \underset{x \in X_{\textrm{uni}}-X_{\textrm{bm}}^{(i)} }{\arg\max} A_{\{1,X_{\textrm{bm}}^{(i)}+x\}}$; $X_{\textrm{bm}}^{(i+1)} \leftarrow X_{\textrm{bm}}^{(i)} + x^{+} $\\
\quad\ \quad\ $X_{\{\textrm{bm,hist}\}}\leftarrow \{X_{\{\textrm{bm,hist}\}},X_{\textrm{bm}}^{(i)}\} $ \\
\quad\ \quad\ $ i \leftarrow i+1$\\
\quad\ \quad\ \textbf{go to Step 2}\\
\quad\ \textbf{else}\\
\quad\ \quad\ \textbf{Output} $ X_{\textrm{bm}} \leftarrow X_{\textrm{bm}}^{(i)} $\\
\quad\ \textbf{end}\\
\textbf{Step 2: (Conditional Exclusion)} \\
\quad\ \textbf{if} {$\underset{x \in X_{\textrm{bm}}^{(i)} }{\max} A_{\{1,X_{\textrm{bm}}^{(i)}-x\}} >A_{\{1,X_{\textrm{bm}}^{(i)}\}}$} \textbf{then} \\
{
\quad\ \quad\ $x^{-} \leftarrow \underset{x \in X_{\textrm{bm}}^{(i)} }{\arg\max}A_{\{1,X_{\textrm{bm}}^{(i)}-x\}} $; $X_{\textrm{bm}}^{(i+1)} \leftarrow X_{\textrm{bm}}^{(i)} - x^{-} $ \\
\quad\ \quad\ $ i \leftarrow i+1$\\
\quad\ \quad\ \textbf{go to Step 2}\\
\quad\ \textbf{else}\\
\quad\ \quad\ \textbf{go to Step 1}\\
\quad\ \textbf{end}\\
}
\end{algorithm}
After successively updating $ X_{\textrm{bm}}^{(i+1)}$ and $i$ by
\begin{equation}\label{equxxpddlusv}
X_{\textrm{bm}}^{(i+1)} \leftarrow X_{\textrm{bm}}^{(i)} - x^{-}
\end{equation}
and $ i \leftarrow i+1$, we go back to check the condition Eq.~\eqref{equxx2if}.
If Eq.~\eqref{equxx2if} is not satisfied, then it indicates any feature in $X_{\textrm{bm}}^{(i)}$ has a positive contribution to the accuracy. In this case, we will come to {Step 1} to select the most significant feature $x^{+}$ in the feature set $X_{\textrm{uni}}-X_{\textrm{bm}}^{(i)} $.
In {Step 1}, when $X_{\textrm{bm}}^{(i)}$ has appeared in history feature set $X_{\{\textrm{bm,hist}\}}$, the iteration will enter an infinite same loop, i.e., the feature added in {Step 1} will be removed in {Step 2}, which means we cannot find a feature set that outperforms $X_{\textrm{bm}}^{(i)}$ in terms of the beam prediction accuracy. Therefore, the algorithm ends and outputs $ X_{\textrm{bm}} \leftarrow X_{\textrm{bm}}^{(i)} $.
The detailed steps of the FS algorithm for beam prediction are given in \textbf{Algorithm~\ref{notr}}.
In fact, other termination criterion could also be adopted. For instance, to limit the number of select features, i.e., $|X_{\textrm{bm}}|$ being no more than $V_{\textrm{max}}$, we could replace Line~\ref{linedff} of \textbf{Algorithm~\ref{notr}} with ``$
\textbf{if} \quad X_{\textrm{bm}}^{(i)} \notin X_{\{\textrm{bm,hist}\}}\ \textbf{and} \ |X_{\textrm{bm}}^{(i)}|< V_{\textrm{max}} \quad \textbf{then}$''.
In this way, \textbf{Algorithm~\ref{notr}} can output the optimal feature subset $ X_{\textrm{bm}} $ with specified size.
\subsubsection{FS for blockage prediction}
The FS algorithm for blockage prediction aims to select the optimal feature set $X_{\textrm{bl}}$. The detailed steps of the FS algorithm for blockage prediction
is the same with that for beam prediction except that its optimization objective is replaced by $A_{\{\textrm{bl},X_{\textrm{bl}}\}}$, as shown in \textbf{Algorithm~\ref{notrnlck}}.
\begin{algorithm}[!t]
\LinesNumbered
\algsetup{linenosize=\tiny} \small
\caption{FS algorithm for blockage prediction }
\label{notrnlck}
\KwIn{ Universal feature set $X_{\textrm{uni}}\buildrel \Delta \over =\{X_{\textrm{uni},v}| i=1,\cdots,V_{\textrm{uni}} \}$}
\KwOut {Selected feature set $X_{\textrm{bl}}\buildrel \Delta \over =\{X_{\textrm{bl},v}| i=1,\cdots,V_{\textrm{bl}},X_{\textrm{bl},v} \in X_{\textrm{uni}} \}$}
Initialize feature set $X_{\textrm{bl}}^{(0)} \leftarrow \emptyset $\\
history feature set $X_{\{\textrm{bl,hist}\}}\leftarrow \emptyset $; $i=0$\\
\textbf{Step 1: (Inclusion)} \\
\quad\ \textbf{if} {$X_{\textrm{bl}}^{(i)} \notin X_{\textrm{hist}}$} \textbf{then} \label{linedff}\\
\quad\ \quad\ $x^{+} \leftarrow \underset{x \in X_{\textrm{uni}}-X_{\textrm{bl}}^{(i)} }{\arg\max} A_{\{\textrm{bl},X_{\textrm{bl}}^{(i)}+x\}}$; $X_{\textrm{bl}}^{(i+1)} \leftarrow X_{\textrm{bl}}^{(i)} + x^{+} $ \\ \quad\ \quad\ $X_{\{\textrm{bl,hist}\}}\leftarrow \{X_{\{\textrm{bl,hist}\}},X_{\textrm{bl}}^{(i)}\} $ \\
\quad\ \quad\ $ i \leftarrow i+1$\\
\quad\ \quad\ \textbf{go to Step 2}\\
\quad\ \textbf{else}\\
\quad\ \quad\ \textbf{Output} $ X_{\textrm{bl}} \leftarrow X_{\textrm{bl}}^{(i)} $\\
\quad\ \textbf{end}\\
\textbf{Step 2: (Conditional Exclusion)} \\
\quad\ \textbf{if} {$\underset{x \in X_{\textrm{bl}}^{(i)} }{\max} A_{\{\textrm{bl},X_{\textrm{bl}}^{(i)}-x\}} >A_{\{\textrm{bl},X_{\textrm{bl}}^{(i)}\}}$} \textbf{then} \\
{
\quad\ \quad\ $x^{-} \leftarrow \underset{x \in X_{\textrm{bl}}^{(i)} }{\arg\max}A_{\{\textrm{bl},X_{\textrm{bl}}^{(i)}-x\}} $; $X_{\textrm{bl}}^{(i+1)} \leftarrow X_{\textrm{bl}}^{(i)} - x^{-} $\\
\quad\ \quad\ $ i \leftarrow i+1$\\
\quad\ \quad\ \textbf{go to Step 2}\\
\quad\ \textbf{else}\\
\quad\ \quad\ \textbf{go to Step 1}\\
\quad\ \textbf{end}\\
}
\end{algorithm}
\subsubsection{The overall training steps of the environment semantics aided
network architecture}
The overall implementation steps can be given as follows:
\begin{enumerate}[i)]
\item Train the environment semantics extraction network at each camera end, i.e, using ADAM algorithm to minimize Eq.~\eqref{equdentsemp} until the convergence. The extraction accuracy of the environment semantics is calculated by Eq.~\eqref{equaccpsp}.
\item Apply \textbf{Algorithm~\ref{notr}} to select the optimal feature set $X_{\textrm{bm}}$ from the universal feature set $X_{\textrm{uni}}$. For each iterated feature set, the auxiliary, the semantic-beam, and the beam-prediction modules would be jointly retrained by minimizing Eq.~\eqref{equdentrop} until the convergence. The Top-1 beam prediction accuracy corresponding to the iterated feature set is tested by Eq.~\eqref{equacc}.
\item Save and fix the parameters of the auxiliary, the semantic-beam, and the beam-prediction modules corresponding to the optimal feature set $X_{\textrm{bm}}$.
\item Apply \textbf{Algorithm~\ref{notrnlck}} to select the optimal feature set $X_{\textrm{bl}}$ from the universal feature set $X_{\textrm{uni}}$. For each iterated feature set, the auxiliary, the semantic-blockage, and the blockage-prediction modules would be jointly retrained by minimizing Eq.~\eqref{equdblockghg} until the convergence. The blockage accuracy corresponding to the iterated feature set is tested by Eq.~\eqref{equaccblcok}.
\item Save and fix the parameters of the auxiliary, the semantic-blockage, and the blockage-prediction modules corresponding to the optimal feature set $X_{\textrm{bl}}$. Note that the initialization, the training, and the storage of the auxiliary module in Step iv)-Step v) are all independent with those in Step ii)-Step iii).
\item The cameras only send the selected environment semantics, i.e., $X_{\textrm{bm}}\cup X_{\textrm{bl}}$, to the task-oriented encoder. Then, BS obtains the predicted Top-$G$ beam and blockage by the decision network.
\end{enumerate}
\textbf{Remark:}
Compared with the pervious vision based works \cite{weihua20192d,9523557,9512383} that directly utilize environment image data to assist channel related tasks, the environment semantics aided network architecture enjoys two main benefits:
\begin{itemize}
\item Environment semantics are extracted from environment images for subsequent data processing, which can not only protect user privacy especially in the case of using the third part cameras, but also reduce the system overheads (e.g., storage space and computational cost) brought by channel-irrelevant information.
\item Channel semantics is extracted from environment images, which reveals the key scatterers associated with channel propagation paths and further provides an interpretable insight into the joint communication and sensing systems.
\end{itemize}
\section{Simulation Results}\label{secsimu}
\subsection{Simulation Setup}
\subsubsection{Environment modeling}
The autonomous driving simulator CARLA \cite{Dosovitskiy17} is utilized to simulate the street traffic environment,
including the street landscape, vehicles, and cameras, etc.
Vehicles in the deployment scenario are randomly chosen from the three types, i.e., the car (3.71$\times$1.79$\times$1.55m$^3$), the van (5.20$\times$2.61$\times$2.47m$^3$), and the bus (11.08$\times$3.25$\times$3.33m$^3$).
Then, we utilize the traffic simulation software SUMO \cite{SUMO2018} to control the speed and
trajectory of all the vehicles.
A partial view of the street traffic environment is shown in the left of Fig.~\ref{fignetmodel}.
\subsubsection{Scene image and semantic label generation} \label{secsema}
As shown in Fig.~\ref{fignetmodel}, two cameras are set at 5m on both sides of the street, both orienting towards the street. The two cameras keep taking images as vehicles cross the street.
Fortunately, the CARLA simulator can generate the corresponding semantic label for each images taken by cameras.
There are 20 classes of environmental semantic concepts, including ``building'', ``fence'', ``pedestrian'', ``pole'', ``roadline'',
``sidewalk'', ``vegetation'', ``vehicle'', ``wall'', ``trafficsign'', ``sky'', ``ground'', ``bridge'', ``railtrack'', ``trafficlight'', ``static'', ``dynamic'', ``water'', ``terrain'', and ``unlabeled''.
In particular, the concept ``static'' refers to the elements that are immovable, e.g., fire hydrants, fixed benches, and bus stops, etc. The concept ``dynamic'' refers to the elements that are susceptible to move over time, e.g., wheelchairs, animals, and buggies, etc.
The concept ``unlabeled'' refers to the elements that have not been categorized.
By collecting the \{image, semantic label\} pairs and splitting them to the training and the testing datasets,
the semantic feature extraction network PSPNet \cite{8100143} could be trained to produce the corresponding environment semantics.
\begin{figure}[!t
\centering
\includegraphics[width=65mm]{wireless.eps}
\caption{A partial view of the synchronization simulation in the Wireless Insite simulator.
To save the cost of art designs, we utilize simple cubes to model the buildings and the vehicles in Wireless Insite and ignore the surface details of these elements. The losses in surface details have limited influences on the channels, which will not affect the reliability of simulations.}
\label{figinste}
\end{figure}
\begin{table*}[!t]\small
\centering
\caption{The network parameters for the beam prediction and blockage prediction}
\label{tabsfgder}
\begin{tabular}{|c|c|c|c|c|c| }
\hline
\footnotesize Module & \footnotesize Layer & \footnotesize Kernel & \footnotesize Stride & \footnotesize Filter/Neuron &\footnotesize Output Shape\\
\hline
\multirow{2}[2]{*}{ \makecell[c]{\footnotesize Auxiliary}} &BatchNorm & / & / & 3 & 3 \\
&(FC-BatchNorm-ReLu)$\times$2 & / & / & 256,16 & 16 \\
\hline
\multirow{4}[2]{*}{ \makecell[c]{\footnotesize Semantic-beam}}
&(Conv-BatchNorm-ReLu)$\times$2 & 3, 3 & 4, 2 & 32, 16 & 16$\times$40$\times$80 \\
& AvgPool & 3 & 2 & / & 16$\times$20$\times$40 \\
&(Conv-BatchNorm-ReLu)$\times$2$<$Residual$>$ & 3, 3 & 4, 1 & 8, 8 & 8$\times$5$\times$10 \\
&(Conv-BatchNorm-ReLu)$\times$2$<$Residual$>$ & 3, 3 & 1, 1 & 8, 8 & 8$\times$5$\times$10 \\
\hline
\multirow{3}[2]{*}{ \makecell[c]{\footnotesize Beam-prediction}} &(FC-BatchNorm-ReLu) & / & / & 512 & 512 \\
\cline{3-5}
& Dropout & \multicolumn{3}{c|}{Rate: 0.1} & 512 \\
\cline{3-5}
& FC & / & / & 64 & 64 \\
\hline
\multirow{3}[2]{*}{ \makecell[c]{\footnotesize Semantic-blockage}}
&(Conv-BatchNorm-ReLu) & 3 & 2 & 16 & 16$\times$40$\times$80 \\
& AvgPool & 3 & 2 & / & 16$\times$20$\times$40 \\
&(Conv-BatchNorm-ReLu)$\times$2$<$Residual$>$ & 3, 3 & 4, 1 & 8, 8 & 8$\times$5$\times$10 \\
\hline
\multirow{3}[2]{*}{ \makecell[c]{\footnotesize Blockage-prediction}} &(FC-BatchNorm-ReLu) & / & / & 64 & 64 \\
\cline{3-5}
& Dropout & \multicolumn{3}{c|}{Rate: 0.1} & 64 \\
\cline{3-5}
& FC & / & / & 1 & 1 \\
\hline
\end{tabular}
\end{table*}
\subsubsection{Channel, blockage label, and beam label generation}
To obtain the channels of users in vehicles, we synchronize the street traffic environment in CARLA with
the 3D ray-tracing simulator Wireless InSite \cite{alkhateeb2019deepmimo} at each moment that cameras take images, which ensures that the images taken by the two cameras exactly correspond to the channels.
Since one image may contain several vehicles or users, we also record the location of the target user, to form the uniquely identified sample pair of the two environment images, the location, and the channel.
Fig.~\ref{figinste} displays a partial view of the synchronization simulation in the Wireless Insite simulator, where
the buildings and the sizes/locations/orientations of vehicles in Wireless Insite are exactly the same as that in CARLA at each image shooting moment. The carrier frequency is 28 GHz, the number of OFDM subcarriers is 128, and the BS is equipped with 64 antennas.
After setting the transceiver antennas at the corresponding locations, the ray-tracing simulator shoots thousands of rays in all directions from the transmitter and records the strongest 20 paths that reach the receiver, obtaining the corresponding channel parameters, i.e., $\{\alpha _{ l},\phi_{ l},\tau_{ l},\theta_{\mathrm{az}}^{ l}, \theta_{\mathrm{el}}^{ l}\}$, and further obtaining the channels using Eq.~\eqref{equchannel}.
The blockage labels can be obtained according to whether there exists a LOS path among the strongest 20 paths during the future time period between two image shootings. Each time slot is 50ms. Unless otherwise specified, the blockage prediction is to predict the next future time slot, i..e., blockage state in next 50ms.
Following Eq.~\eqref{equaraterhh}, the corresponding beam labels can be obtained by exhaustively searching on the codebook. Here we adopt the discrete fourier transformation codebook as an example.
\subsubsection{Network parameters}
The network structure of the feature extraction network follows the work \cite{8100143}, except that the input and the output dimensions are adjusted according to the shapes of the images and the semantic labels generated by CARLA. Moreover, various network scales, including ResNet-18, ResNet-34, ResNet-50, and ResNet-100 \cite{7780459}, are respectively adopted and trained on the datasets in Section~\ref{secsema}.
The network parameters for the auxiliary, the semantic-beam, the semantic-blockage, the beam-prediction, and the blockage-prediction modules are given in Tab.~\ref{tabsfgder}, where
``-'' represents sequential stacking, ``$(\cdot)\times 2$'' represents that the layers insides $(\cdot)$ are sequentially stacked twice, the numbers split by ``,'' represents the parameters for adjacent convolution or FC layers, ``$<$Residual$>$'' represents the residual connection between the adjacent convolutional layers, and ``Conv'' is the abbreviation of ``convolution''.
The initial learning rate of the ADAM optimizer is 0.001, and the batch size is 128.
\begin{figure}[!t
\centering
\includegraphics[width=65 mm]{psp.eps}
\caption{The prediction accuracy of PSPNet, i.e., the accuracy of the environment semantics extraction, trained on datasets in Section~\ref{secsema}.}
\label{figpspggr}
\end{figure}
\begin{figure*}
\centering
\subfigure[Top-5 beam prediction accuracy versus the number of selected features.]{\label{figfeaff1}\includegraphics[width=75mm]{pic_num.eps}}
\subfigure[Top-5 beam prediction accuracy under various selected features, $V_{\textrm{bm}} =1$.]{
\label{figfeaff2}\includegraphics[width=75mm]{pic_acc1.eps}
}\\
\subfigure[Top-5 beam prediction accuracy under various selected features, $V_{\textrm{bm}} =2$.]{
\label{figfeaff3}\includegraphics[width=75mm]{pic_acc2.eps}
}
\subfigure[Top-5 beam prediction accuracy under various selected features, $V_{\textrm{bm}} =3$.]{\label{figfeaff4}\includegraphics[width=75mm]{pic_acc3.eps}}
\caption{The Top-5 beam prediction accuracy $A_{\{5,X_{\textrm{bm}}\}}$ under various selected feature sets.}
\label{figfeatureselect}
\end{figure*}
\subsection{Performance of the Environment Semantics Extraction}
We adopt PSPNet to extract the environment semantics from environment images, which implies that the prediction accuracy of PSPNet, i.e., $A_{\textrm{PSP}}$, is the accuracy of the environment semantics extraction.
Fig.~\ref{figpspggr} exhibits the prediction accuracies of PSPNet versus the network scales.
The accuracy of PSPNet improves as the network scale increases, and the accuracy saturates when the network scale reaches that of ResNet-50.
The work \cite{8100143} presents the accuracy of PSPNet on the ADE20K dataset \cite{ZhougdtBolei}, which is much lower than that on
the datasets generated by CARLA.
This is because that the number of semantic concepts to be classified in the ADE20K dataset is 150, which is much larger than that in our datasets, i.e., 20. Therefore, the classification task for the ADE20K dataset is harder than that for our datasets.
This indicates that the accuracy of the semantic feature extraction could be further improved if we could remove unimportant semantic features and further reduce the number of categories to be classified.
\subsection{Performance of FS for Beam Prediction and Blockage Prediction}
Fig.~\ref{figfeatureselect} displays the Top-5 beam prediction accuracy $A_{\{5,X_{\textrm{bm}}\}}$ under various feature sets selected by \textbf{Algorithm~\ref{notr}}. Fig.~\ref{figfeaff1} plots the Top-5 beam prediction accuracy versus the number of selected features $V$, where the Top-5 beam prediction accuracy is obtained through the execution of \textbf{Algorithm~\ref{notr}}.
More specifically, the Top-5 beam prediction accuracy at $V_{\textrm{bm}} =v$ can be determined by
\begin{equation*}
\underset{X_{\textrm{bm}} \in X_{\{\textrm{bm,hist}\}} ,|X_{\textrm{bm}}|=v } {\arg\max} A_{\{5,X_{\textrm{bm}}\}}.
\end{equation*}
The Top-5 beam prediction accuracy improves as $V_{\textrm{bm}} $ increases, while the accuracy gain is less than 0.01 when $V_{\textrm{bm}} $ is larger than 3, which implies that three features are efficient enough for beam prediction.
Fig.~\ref{figfeaff2}-Fig.~\ref{figfeaff4} present the Top-5 beam accuracies under various selected feature sets with $V_{\textrm{bm}} =1,2,$ and 3, respectively, where the feature names are listed in descending order along the $x$-axis according to the corresponding accuracies.
In Fig.~\ref{figfeaff2}, only one feature is utilized to predict the beam.
The location yields the highest accuracy among all the features. This is because that the location is the user's
identification information, which is essential to distinguish the target user from other possible users\footnote{The accuracy performance associated with the location also explains that location based works are popular \cite{weihua20192d,9523557,9512383}. In fact, other user's identification information, such
as historical beam indices or other predetermined information, could also be used to
distinguish the target user from other possible users. Therefore, the location based prediction is only a special case of the environment semantics aided communication system. }.
The reason that other features like the vehicle still yield higher accuracies than random predictions is due to the limited concurrent users within the same image.
In Fig.~\ref{figfeaff3}, two features are used to predict the beam, and one of them is the location.
It can be seen than the three environment semantic concepts, i.e., ``vehicle'', ``road'', and ``sidewalk'', lead to the highest beam prediction accuracy.
In Fig.~\ref{figfeaff4}, three features are utilized to predict the beam, and two of them are the location and the vehicle. The feature set \{location, vehicle, sidewalk\} can achieve the highest accuracy, which conforms to the intuition that the vehicles provide the information of users' directions and the possible obstacles, while sidewalks offer the information of street layouts.
\begin{figure*}
\centering
\subfigure[Blockage prediction accuracy versus the number of selected features.]{\label{figfeablock1}\includegraphics[width=75mm]{picblock_num.eps}}
\subfigure[Blockage prediction accuracy under various selected features, $V_{\textrm{bl}}=1$.]{
\label{figfeablock2}\includegraphics[width=75mm]{picblock_acc1.eps}
}\\
\subfigure[Blockage prediction accuracy under various selected features, $V_{\textrm{bl}}=2$.]{
\label{figfeablock3}\includegraphics[width=75mm]{picblock_acc2.eps}
}
\subfigure[Blockage prediction accuracy under various selected features, $V_{\textrm{bl}}=3$.]{\label{figfeablock4}\includegraphics[width=75mm]{picblock_acc3.eps}}
\caption{The blockage prediction accuracy $A_{\{\textrm{bl},X_{\textrm{bl}}\}}$ under various selected feature sets.}
\label{figfeablock}
\end{figure*}
Fig.~\ref{figfeablock} displays the blockage prediction accuracy $A_{\{\textrm{bl},X_{\textrm{bl}}\}}$ under various feature sets selected by \textbf{Algorithm~\ref{notrnlck}}.
Fig.~\ref{figfeablock1} plots the blockage prediction accuracy versus the number of selected features $V_{\textrm{bl}}$, where the blockage prediction accuracy at $V_{\textrm{bl}} =v$ can be obtained by
\[\underset{X_{\textrm{bl}} \in X_{\{\textrm{bl,hist}\}} ,|X_{\textrm{bl}}|=v } {\arg\max} A_{\{{\textrm{bl}},X_{\textrm{bl}}\}}.\]
It can be observed that two features are efficient enough for blockage prediction.
Fig.~\ref{figfeablock2}-Fig.~\ref{figfeablock4} present the blockage prediction accuracies under various selected feature sets with $V_{\textrm{bl}} =1,2,$ and 3, respectively, where the feature names are listed in descending order along the $x$-axis according to the corresponding accuracies.
Similar with Fig.~\ref{figfeatureselect}, the feature set \{location, vehicle, sidewalk\} can achieve the highest accuracy.
This validates that the feature set \{location, vehicle, sidewalk\} are
the most significant features for channel related tasks, and other unselected semantic features
have lower relevancy with channels.
\begin{table}[!t]
\centering
\caption{Environment semantics analyses}
\begin{tabular}{p{14em}p{8em}}
\rowcolor[rgb]{ .267, .447, .769} \textcolor[rgb]{ 1, 1, 1}{\footnotesize \textbf{Environmental semantic concept}} & \textcolor[rgb]{ 1, 1, 1}{\footnotesize \textbf{Accuracy contribution}} \\
\hline
\footnotesize Vehicles, Road, Sidewalk & \footnotesize Critical \\
\hline
\footnotesize Dynamic, Terrain & \footnotesize Moderate \\
\hline
\footnotesize{Sky, Building, Pedestrian, Roadline, Pole, Vegetation, Wall, Ground, Bridge, Railtrack, Trafficlight, Water, Fence, Static} & \footnotesize \multirow{2}[1]{*}{Negligible} \\
\hline
\end{tabular
\label{tabanattt
\end{table
\begin{table}[!t]
\centering
\caption{Beam prediction and blockage prediction accuracies with various inputs}
\begin{tabular}{ccc}
\hline
\footnotesize {\textbf{Input}} & \footnotesize {\textbf{Top-5 beam}} & \footnotesize {\textbf{Blockage}} \\
\hline
\footnotesize Original images & \footnotesize 0.910$\pm$0.091 & \footnotesize 0.971$\pm$0.088 \\
\hline
\footnotesize All environment semantics &\footnotesize 0.914$\pm$0.042& \footnotesize 0.974$\pm$0.036 \\
\hline
\footnotesize \{Vehicles, Road, Sidewalk\} &\footnotesize 0.914$\pm$0.020 & \footnotesize 0.974$\pm$0.018 \\
\hline
\end{tabular
\label{tabbeamblock
\end{table
\begin{figure*}
\centering
\subfigure[Beam accuracy versus the number of training epoches.]{\label{figtopk1}\includegraphics[width=80mm]{topk_acc.eps}}
\subfigure[TRR versus the number of training epoches.]{
\label{figtopk2}\includegraphics[width=80mm]{topk_rate.eps}
}
\caption{The beam prediction accuracy (a) and the TRR (b) of the Top-$G$ beam versus the number of training epoches.}
\label{figtopk}
\end{figure*}
\begin{table*}[!t]
\centering
\caption{Blockage prediction accuracy versus future time slots}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
\footnotesize Time slots &\footnotesize 1 &\footnotesize 6 &\footnotesize 11 & \footnotesize 16 & \footnotesize 21 & \footnotesize 26 & \footnotesize 31 & \footnotesize 36 \\
\hline
\footnotesize $A_{\{\textrm{bl},X_{\textrm{bl}}\}}$ &\footnotesize 0.9741 &\footnotesize 0.966 &\footnotesize 0.951 &\footnotesize 0.94 & \footnotesize 0.929 & \footnotesize 0.922 &\footnotesize 0.901 & \footnotesize 0.882 \\
\hline
\end{tabular}
\label{tabggggfhjkd}
\end{table*}
Based on the results of Fig.~\ref{figfeatureselect} and Fig.~\ref{figfeablock}, we divide the environmental semantic concepts into three levels of accuracy contributions, i.e., critical, moderate, and negligible, as shown in Tab.~\ref{tabanattt}.
Obviously, the environmental semantic concepts ``vehicles'', ``road'', and ``sidewalk'' are the most critical features due to their contributions to the beam prediction and blockage prediction.
Therefore, the three environmental semantic concepts ``vehicles'', ``road'', and ``sidewalk'' can be interpreted as the most significant channel semantics, and other unselected semantic features have lower relevancy with the channels.
For beam prediction, as shown in Fig.~\ref{figfeaff3} and Fig.~\ref{figfeaff4}, ``road'' is more effective than ``sidewalk'' for $V_{\textrm{bm}}=2$ but is less effective than ``sidewalk'' for $V_{\textrm{bm}}=3$.
To understand the phenomenon, one can glean some indications in the right part of Fig.~\ref{fignetseg}, where the shapes of ``road'' and ``vehicles'' are complementary, which implies that the better accuracy of ``road'' for $V_{\textrm{bm}}=2$ owns to its high dependency on ``vehicles'', while the gain of ``road'' would be weakened for $V_{\textrm{bm}}=3$ where ``vehicles'' has been selected.
When ``vehicle'' has been selected, ``sidewalk'' could offer extra beam prediction accuracy gains.
Moreover, ``dynamic'' has moderate contributions to the beam prediction accuracy. This is because that ``dynamic'' objects in the road may change the shape of ``vehicles'', and the acquisition of ``dynamic'' can modify the shape distortion, yielding noticeable contributions to the beam prediction accuracy.
For blockage prediction, as shown in Fig.~\ref{figfeablock3} and Fig.~\ref{figfeablock4}, the accuracy gain brought by the third feature sidewalk is marginal. This indicates that the environmental semantic concept ``vehicles'' is sufficient for blockage prediction, and blockage prediction is less sensitive to other environmental semantic concepts than beam prediction. This is because that the beam vector is highly related to various scatterers in environments while the blockage state is only related to the obstacle in the LOS path.
Tab.~\ref{tabbeamblock} compares the beam prediction and blockage prediction accuracies with various inputs, i.e., original images, all environment semantics, and the selected channel semantics set \{location, vehicle, sidewalk\}. The overall network structures for original images and all environment semantics are the same with those for the selected channel semantics, and the only difference is that the network input has been replaced by the original images and the segmentation maps, respectively.
It can be observed that the channel semantics set is much sparse than the environment semantics set, but is informative enough to obtain almost the same prediction accuracy, which also demonstrates that the proposed FS algorithm can significantly compress the original image flows while preserve the most informative features.
Furthermore, the standard deviations of the selected channel semantics set are smaller than those of original images and all environment semantics. This is because irrelevant semantics or fine image details expand the input feature space, making available samples sparser and more dissimilar, which results in a network more easier to overfit and generate unstable predictions.
Therefore, the proposed environment semantics aided network architecture cannot only reduce system overheads brought by channel-irrelevant information, but also improve the prediction robustness.
\subsection{Performance of Top-$G$ Beam Prediction}
Fig.~\ref{figtopk} depicts the beam prediction accuracy and the TRR of the Top-$G$ beam versus the number of training epoches, where the feature set \{location, vehicle, sidewalk\} is adopted.
It can be seen that the task-oriented encoders and decision network converge when the number of epoches is larger than 50.
As shown in Fig.~\ref{figtopk1}, the Top-1 beam prediction accuracy reaches 66\%, and the Top-5 beam prediction accuracy reaches 91\%.
As shown in Fig.~\ref{figtopk2}, the TRR of Top-1 beam reaches 88\%, and the TRR of Top-5 beam reaches 98\%.
These results demonstrate the effectiveness of the proposed environment semantics aided network architecture.
\subsection{Performance of Blockage Prediction Versus Future Time Slots}
Tab.~\ref{tabggggfhjkd} displays the blockage prediction accuracies versus future time slots, where each time slot is 50ms. For each recorded time slot point, i.e., 6, 12, $\cdots$, 36, the network would be retrained with the corresponding future blockage labels. It can be seen that the blockage prediction accuracy is higher than 90\% for future 30 time slots, i.e., 1.5s, which demonstrates the effectiveness of the proposed environment semantics aided network architecture.
\section{Conclusion} \label{secconcul}
In this paper, we propose a framework for environment semantics aided wireless communications to improve the transmission efficiency and protect the user privacy.
We also developed an environment semantics aided network architecture for mmWave beam prediction and blockage prediction as a case study. Simulation results show that only three features of the environment semantics can achieve almost the same accuracy as the whole environment semantics, which can significantly reduce system overheads and improve the transmission efficiency.
The superiority of the environment semantics aided wireless communication framework in
realizing extremely low-latency beam prediction and supporting ultra-reliable communication links
demonstrates its great application potential in URLLCs.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Edge Computing (EC) is an emerging technology deployed in mobile edge networks, which coordinates computing and memory resources in the edge to support low-latency and high-bandwidth-demanding applications for end-side mobile and Internet of Things (IoT) devices.
With the increasing concern for intelligent services and data privacy protection of devices, Federated Edge Learning (FEL) \cite{tak2020federated} is proposed to jointly train Machine Learning (ML) models with decentralized data at the edge of the network.
Unlike conventional centralized training paradigms, FEL keeps \textcolor{black}{private} data on devices and only transmits model parameters or \textcolor{black}{encrypted data information} to the edge server, which is a promising solution for privacy-preserving edge intelligence.
However, FEL faces severe challenges related to resources, personalization, network environments, etc \cite{tak2020federated,yu2021toward,lim2020federated}.
Shortage of individual device resources and imbalance of resources among devices significantly increase the difficulty of utilizing end-side resources.
Besides, personalized needs of users call for differentiated on-device models, while the uniform models trained by conventional parameter-averaging-based FEL methods cannot generalize to all devices.
Moreover, non-ideal communication channels and network topology constrain the system design of FEL.
As an ML technique that both enables knowledge transfer and model collaborative training, \textcolor{black}{Knowledge Distillation (KD) \cite{hinton2015distilling,wang2021knowledge} transfers knowledge from one ML model to another, allowing interactive learning among heterogeneous ML models to achieve constructive optimization.}
Due to flexibility and effectiveness, KD has been applied to solve numerous ML problems, such as model compression \cite{hinton2015distilling,he2019knowledge}, domain adaptation \cite{wu2021spirit,nguyen2021unsupervised}, distributed learning \cite{anil2018large,bistritz2020distributed}, etc.
Recent trends suggest the great potential to address the above challenges in the context of FEL.
Previous works integrating KD into the training process of FEL have been successful in tackling constrained device resources \cite{itahara2021distillation,jeong2018communication,sattler2021cfd,song2022federated,cheng2021fedgems,he2020group,cho2022heterogeneous}, adapting to heterogeneous devices and user requirements \cite{wu2023fedict,jin2022personalized,zhou2021source,zhang2022fedzkt,yu2022resource,qi2022fedbkd,mishra2021network}, and adapting to complex communication channels as well as network topologies \cite{ahn2019wireless,ahn2020cooperative,oh2020mix2fld,li2021decentralized,taya2022decentralized}. Therefore, a survey is urgently needed to review how KD applies to FEL.
To the best of our knowledge, \textbf{this paper is the first work to investigate the application of knowledge distillation in federated edge learning.}
Different from existing surveys \cite{tak2020federated,yu2021toward,lim2020federated,mora2022knowledge}, we take the challenges faced by FEL as the main clue, introducing existing FEL approaches based on diverse forms of KD techniques and providing guidance for both future research directions and real deployment.
Specifically, the reminders of this paper are organized as follows.
Section \ref{pre} provides preliminary knowledge of related research directions, including FEL and KD, and elaborates the reasons for concerning KD in FEL.
Section \ref{challenge} investigates FEL based on KD in addressing resource-constrained, resource-heterogeneous, personalization, non-ideal channels and decentralization challenges in mobile edge networks.
Section \ref{discuss} summarizes the limitations of existing methods, raises open problems in KD-based FEL research, and provides guidance for real deployment.
Section \ref{conclude} summarizes the whole paper.
\section{Preliminary}
\label{pre}
\subsection{Federated Edge Learning}
As a practical way to realize edge intelligence, Federated Edge Learning (FEL) implements Federated Learning (FL) systems in mobile edge networks, where massively distributed mobile and Internet of Things (IoT) devices jointly train machine learning models without sharing private data on devices \cite{tak2020federated}.
Due to diverse user behaviors, limited device capabilities and non-ideal communication environments, FEL faces more severe challenges than conventional FL: more restricted on-device computation and communication resources, more prominent cross-devices data and resource heterogeneity, more complex channel environments, etc \cite{yu2021toward,lim2020federated}.
How to address the above challenges to accommodate the system infrastructure and user characteristics of mobile edge networks is crucial for practical applications of FEL.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8 \textwidth]{
KD.pdf}
\caption{General framework of knowledge distillation.}
\label{kd}
\end{figure}
\subsection{Knowledge Distillation}
Knowledge Distillation (KD) is a machine learning technique for constructive model training through knowledge transfer \cite{wang2021knowledge,gou2021knowledge}. In common KD frameworks, the transferred model output (typically logits \cite{hinton2015distilling,peng2019correlation} or features \cite{liu2020structured,he2019knowledge,wu2021spirit2}) is referred to as knowledge.
\textcolor{black}{Fig. \ref{kd} shows the general framework of KD.}
The core of KD is that given the same input, one model (student network) simulates the adjusted outputs of another model (teacher network) to learn from the representation of the latter model, that is:
\[\mathop {\arg \min }\limits_{{W_S}} \mathop E\limits_{X\sim{\cal X}} [dist(f(X;{W_S}),\tau (f(X;{W_T})))],\]
where $W_S$ and $W_T$ are model parameters of \textcolor{black}{the student and the teacher networks}, respectively. $f(\cdot;W_*)$ is a nonlinear function determined by parameters $W_*$, which is commonly a neural network. $\tau(\cdot)$ is a knowledge distribution adjustment function for knowledge interpretation, making the transferred knowledge easier to be learned by the student network.
\textcolor{black}{
During the distillation process, the student network inputs data or features $X\sim \mathcal{X}$, where $\mathcal{X}$ is the set of input data or extracted features. The student network then accepts the knowledge (outputs) extracted by the teacher network on the same input $X$ and achieves knowledge transfer by minimizing the distance $dist(\cdot,\cdot)$ between the two networks' outputs.
}
KD can be customized to solve various machine learning problems, such as model compression via knowledge transfer from bulky models to compact models \cite{wu2021spirit,hinton2015distilling}, distributed model training via knowledge exchange between models \cite{anil2018large,bistritz2020distributed}, etc.
With elaborate system design, recent studies confirm that KD can be leveraged to solve problems in FL as well \cite{chen2022metafed,li2019fedmd,cho2021personalized,zhang2021parameterized,wu2022exploring,zhu2021data,jiang2022towards,lee2022preservation}.
\subsection{Why Concern Knowledge Distillation in Federated Edge Learning}
KD has achieved success in FL, such as realizing efficient communication \cite{itahara2021distillation,sattler2021cfd}, enabling model heterogeneity \cite{li2019fedmd,zhu2021data} and personalization \cite{zhang2021parameterized,cho2021personalized} among devices, etc.
\textcolor{black}{Recent trends suggest that KD has great potential to apply to various learning processes in FEL as an important tool for knowledge transfer or model collaborative training in diversity-constrained mobile edge networks.}
Specifically, the technical characteristics of KD meet the core demands of FEL, and the roles it can play include but not limited to compressing large-scale edge models for on-device deployment \cite{mishra2021network}, transferring local adaptive knowledge to on-device models for personalization \cite{wu2023fedict,jin2022personalized,zhou2021source}, and helping establish novel FL frameworks for enabling heterogeneous device supports \cite{qi2022fedbkd,zhang2022fedzkt,yu2022resource}.
Representative works that apply KD in FEL are summarized in Table \ref{methods}, \ref{knowledge-and-data}.
\color{black}
As shown in Table \ref{methods}, prior works are classified into four types according to the role of KD in the FEL: 1) Knowledge transfer, where KD is utilized to transfer knowledge from one ML model to another, occurring only at the end-side or edge-side.
2) Model representation exchange protocol, where on-device and edge models achieve collaborative optimization based on exchanged logits or features. 3) Component of backbone algorithm, where KD-based FEL algorithms are improved as the base algorithm.
4) Dataset distillation, where small datasets derived from on-device data are synthesized via KD.
In addition, the deployment modes of KD in FEL are classified into edge, end and edge-end collaboration, depending on the application location of KD and whether end-side and edge-side collaboration are required.
\textcolor{black}{As shown in Table \ref{knowledge-and-data}, some FEL approaches require public datasets, in which edge models and on-device models leverage the same shared public data during distillation to synchronize the knowledge they generate.}
\color{black}
In the following sections, we will detailly introduce the role played by KD in \textcolor{black}{tackling a variety of challenges} faced by FEL.
\begin{table}[t]
\caption{Comparison of representative works that apply KD in FEL \textcolor{black}{regarding challenges to tackle, roles play, and deployment modes.}}
\centering
\setlength{\tabcolsep}{5pt}
\renewcommand\arraystretch{1.12}
\begin{adjustbox}{center}
\begin{tabular}{l|c|c|c}
\hline
\multicolumn{1}{c|}{\textbf{Method}} & \textbf{Challenge to Tackle} & \textbf{Role of KD in FEL} & \textbf{Deployment Mode} \\ \hline
Fed-ET \cite{cho2022heterogeneous} & \begin{tabular}[c]{@{}c@{}}Limited\\ Computation\end{tabular} & \multirow{11}{*}{Knowledge Transfer} & \multirow{4}{*}{Edge} \\ \cline{1-2}
FedZKT \cite{zhang2022fedzkt} & \begin{tabular}[c]{@{}c@{}}Heterogeneous\\ Computation\end{tabular} & & \\ \cline{1-2}
Mix2FLD \cite{oh2020mix2fld} & Non-ideal Channel & & \\ \cline{1-2} \cline{4-4}
pFedSD \cite{jin2022personalized} & \multirow{2}{*}{Personalization} & & \multirow{5}{*}{End} \\ \cline{1-1}
STU-KD \cite{zhou2021source} & & & \\ \cline{1-2}
NRFL \cite{mishra2021network} & \begin{tabular}[c]{@{}c@{}}Heterogeneous\\ Communication\end{tabular} & & \\ \cline{1-2}
Def-KT \cite{li2021decentralized} & Decentralization & & \\ \cline{1-2} \cline{4-4}
FedKEMF \cite{yu2022resource} & \begin{tabular}[c]{@{}c@{}}Heterogeneous\\ Computation\end{tabular} & & Edge+End \\ \hline
DS-FL \cite{itahara2021distillation} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Limited\\ Communication\end{tabular}} & \multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Model Representation\\ Exchange Protocol\end{tabular}} & \multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Edge-end\\ Collaboration\end{tabular}} \\ \cline{1-1}
FD \cite{jeong2018communication} & & & \\ \cline{1-2}
FedGEMS \cite{cheng2021fedgems} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Limited\\ Computation\end{tabular}} & & \\ \cline{1-1}
FedGKT \cite{he2020group} & & & \\ \cline{1-2}
FedBKD \cite{qi2022fedbkd} & \begin{tabular}[c]{@{}c@{}}Heterogeneous\\ Computation\end{tabular} & & \\ \cline{1-2}
CMFD \cite{taya2022decentralized}
& Decentralization & & \\ \hline
FedICT \cite{wu2023fedict} & Personalization & \begin{tabular}[c]{@{}c@{}}Knowledge Transfer+\\ Model Representation\\ Exchange Protocol\end{tabular} & \begin{tabular}[c]{@{}c@{}}Edge+End+\\ Edge-end\\ Collaboration\end{tabular} \\ \hline
CFD \cite{sattler2021cfd} & \begin{tabular}[c]{@{}c@{}}Limited \\ Communication\end{tabular} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Component of \\ Backbone Algorithm\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Edge-end\\ Collaboration\end{tabular}} \\ \cline{1-2}
Ahn, et al. \cite{ahn2019wireless,ahn2020cooperative} & Non-ideal Channel & & \\ \hline
FedD3 \cite{song2022federated} & \begin{tabular}[c]{@{}c@{}}Limited\\ Communication\end{tabular} & Dataset Distillation & End \\ \hline
\end{tabular}
\end{adjustbox}
\label{methods}
\end{table}
\begin{table}[]
\caption{\textcolor{black}{Comparison of representative works that apply KD in FEL regarding knowledge types and requirements for public datasets.}}
\centering
\setlength{\tabcolsep}{10pt}
\renewcommand\arraystretch{1.12}
\begin{tabular}{l|c|c}
\hline
\multicolumn{1}{c|}{\textbf{Method}} & \textbf{Knowledge Type} & \textbf{Require Public Dataset} \\ \hline
STU-KD \cite{zhou2021source} & \multirow{15}{*}{Logits} & \multirow{12}{*}{Yes} \\ \cline{1-1}
DS-FL \cite{itahara2021distillation} & & \\ \cline{1-1}
FD \cite{jeong2018communication} & & \\ \cline{1-1}
CFD \cite{sattler2021cfd} & & \\ \cline{1-1}
FedGEMS \cite{cheng2021fedgems} & & \\ \cline{1-1}
Fed-ET \cite{cho2022heterogeneous} & & \\ \cline{1-1}
FedKEMF \cite{yu2022resource} & & \\ \cline{1-1}
FedBKD \cite{qi2022fedbkd} & & \\ \cline{1-1}
Mix2FLD \cite{oh2020mix2fld} & & \\ \cline{1-1}
Ahn, et al. \cite{ahn2019wireless,ahn2020cooperative} & & \\ \cline{1-1}
Def-KT \cite{li2021decentralized} & & \\ \cline{1-1}
CMFD \cite{taya2022decentralized} & & \\ \cline{1-1} \cline{3-3}
pFedSD \cite{jin2022personalized} & & \multirow{3}{*}{No} \\ \cline{1-1}
FedZKT \cite{zhang2022fedzkt} & & \\ \cline{1-1}
NRFL \cite{mishra2021network} & & \\ \hline
FedGKT \cite{he2020group} & \multirow{2}{*}{Features+Logits} & \multirow{2}{*}{No} \\ \cline{1-1}
FedICT \cite{wu2023fedict} & & \\ \hline
FedD3 \cite{song2022federated} & Distilled Datasets & No \\ \hline
\end{tabular}
\label{knowledge-and-data}
\end{table}
\section{Challenges in Federated Edge Learning with Solutions based on Knowledge Distillation}
\label{challenge}
\subsection{Limited Resources}
As hardware configurations and running power of devices are commonly inferior in FEL, communication and computation resources at the network termination are strictly limited.
Therefore, how to fully utilize available resources to achieve satisfactory system performance is the core issue that KD need to tackle in FEL.
To reduce the communication overhead in FEL, Itahara \cite{itahara2021distillation} proposes DS-FL, which is a distillation-based semi-supervised FL framework. DS-FL achieves efficient communication via on-device local models' outputs exchange between heterogeneous devices \textcolor{black}{and optimizes} local models based on KD with common inputs of an open dataset, thus avoiding model parameters exchange conducted by conventional FEL approaches \cite{jiang2022fedsyl,mills2021multi}.
Jeong \cite{jeong2018communication} employs label-level knowledge alignment in FEL, where the per-label mean logit vectors of each device (knowledge) are periodically uploaded to the edge server. The server then broadcasts the aggregated knowledge to all devices in turn, and the on-device models conduct KD based on the downloaded knowledge for local models' regularization.
Sattler \cite{sattler2021cfd} extends existing KD-based FEL methods \cite{jeong2018communication,sattler2021cfd} through imposing quantization and delta coding to exchanged knowledge, showing $\times 100$ communication efficiency improvement compared with prior works. On top of this, Song \cite{song2022federated} distils local datasets on devices and uploads the distilled dataset to the edge server, requiring only one-shot communication during the entire training process.
To reduce on-device computation overhead in FEL, a series of approaches \cite{he2020group,cheng2021fedgems,cho2022heterogeneous} are proposed to enable devices training much smaller models than the edge server, leveraging KD as an exchange protocol across model representations.
Specifically, He \cite{he2020group} and Cheng \cite{cheng2021fedgems} establish alternating minimization FEL frameworks to transfer knowledge from compact on-device models to the large edge model via KD, after which the on-device models are optimized based on the knowledge transferred back from the edge.
In addition, Cho \cite{cho2022heterogeneous} proposes an ensemble knowledge transfer algorithm for collaborative training across devices, where the large edge model learns the weighted consensus of the uploaded small on-device models via KD.
\subsection{Heterogeneous Resources}
\textcolor{black}{Resources among devices exist huge divergence in FEL} deriving from disparate hardware configurations, cruising ability, and network connectivity of different devices.
Resource gap among devices poses requirements that KD-based FEL approaches should support heterogeneous training demands depending on the hardware and power of devices, and tolerate disparate communication capabilities of devices.
Existing works employ heterogeneous on-device models to adapt to the computing power of heterogeneous devices' hardware, and leverage KD to transfer knowledge from the shared edge model to heterogeneous on-device models.
Specifically, Zhang \cite{zhang2022fedzkt} proposes a zero-shot knowledge transfer approach that adversarially trains a generator with uploaded on-device models. After that, pseudo-data generated by the trained generator is applied to a bi-directional KD process across the edge and the devices, aiming to integrate the knowledge from heterogeneous on-device models to the edge server and guide the on-device models to achieve more generalized training performance in turn.
Yu \cite{yu2022resource} adopts a variant of co-distillation \cite{anil2018large} at the network termination for collaboratively optimizing on-device models and the downloaded edge model, while integrating the knowledge from heterogeneous devices via ensemble KD in the edge.
In addition, Qi \cite{qi2022fedbkd} employs KD after on-device model uploading and edge model downloading, and uses a conditional variational autoencoder to complement both the required public data and private data.
It is also worth noting that previously mentioned approaches taking KD as the end-edge model representation exchange protocol \cite{he2020group,cheng2021fedgems,cho2022heterogeneous,wu2023fedict,itahara2021distillation,jeong2018communication} also support heterogeneous on-device models, which are partly applicable on heterogeneous devices as well.
Another line of research addresses the problem of unequal available bandwidth among devices in FEL, where Mishra \cite{mishra2021network} proposes a network resource-aware FL approach using KD. Precisely, devices are grouped depending on their available bandwidth resources and are set models that accommodate their bandwidth. During the training process, devices accept the initialized on-device model parameters from the edge server and transfer back the locally trained models in groups from large to small.
The edge server iteratively distributes the aggregated on-device models to their corresponding groups of devices and uses KD to compress the uploaded models for initializing model parameters of the next group of devices.
\subsection{Personalization}
Local data of different devices are generated from the daily use of their respective users.
Due to differentiated user behaviours, data among devices tend to exhibit non-independent identically distributed (non-IID), and will make it difficult for the trained edge-side global model to generalize to all devices.
Therefore, techniques for global model personalization are urgently needed to address the challenges posed by non-IID data in FEL.
In \cite{wu2023fedict}, a novel bi-directional knowledge distillation framework practical for multi-access edge computing is proposed to achieve personalized optimization of on-device models while ensuring fast convergence of the global model. Besides, Jin \cite{jin2022personalized} proposes to reserve personalized models on devices. At the end of each communication round, self-distillation is employed on devices to transfer knowledge from the personalized on-device models of the previous round, and the distilled model is retained to the next round for further personalized optimization. In addition, Zhou \cite{zhou2021source} models FEL as a federated domain adaptation problem, and leverages distillation-based source-free unsupervised domain adaptation to transfer the knowledge from the edge server in a memory-efficient manner, aiming to achieve high inference accuracy in local data environments on edge devices with acceptable training cost.
\subsection{Non-ideal Channels}
Most existing KD-based FEL approaches assume that communication links between devices and the edge server are ideal channels, which are impractical in mobile edge networks in reality.
For instance, the capacity of uplink and downlink channels are often unequal in real network environments, or the FEL system usually is deployed on fading channels, etc.
Such wireless communication constraints present new challenges to implementing KD-based FEL.
In \cite{oh2020mix2fld}, Oh addresses uplink-downlink capacity asymmetry by allowing devices to upload model outputs but download model parameters during training, and the edge model conducts optimization via KD, with uploaded mixed samples from devices as input.
Ahn \cite{ahn2019wireless} models devices' uplink as gaussian multiple-access channels, and provides digital implementations with separate source-channel coding and over-the-air computing implementations with joint source-channel coding for KD-based FEL. Further, the case in which uplink is a multiple access fading channel and downlink is a fading broadcast channel is considered in \cite{ahn2020cooperative}, and concrete implementations for KD-based FEL are given as well.
\subsection{Decentralization}
In mobile edge networks, locating powerful and reliable edge servers can be challenging due to weak infrastructure, restricted network topologies, and a lack of trust in remote servers.
Therefore, decentralization of FEL is a promising research direction in which KD provides a tool for stable convergence of decentralized FEL systems with heterogeneous device support.
Taya \cite{taya2022decentralized} proposes a decentralized FL algorithm for Internet of Everything (IoE) in multi-hop networks, where each device learns the transferred knowledge from neighboring devices via KD. In addition, Li \cite{li2021decentralized} leverages mutual knowledge transfer to allow clients to learn knowledge from each other, enabling training on heterogeneous datasets with client-drift prevented.
\section{Discussion}
\label{discuss}
\subsection{Limitations}
Although KD-based FEL has witnessed a phased success, quite a few approaches \cite{he2020group,itahara2021distillation,jeong2018communication,sattler2021cfd,oh2020mix2fld,ahn2019wireless} only achieve comparable or even worse performances than inchoate FL methods represented by FedAvg \cite{mcmahan2017communication}.
On the one hand, the amount of useful information transferred from knowledge in a single communication round is less than model parameters, and the gain in communication efficiency achieved by the above methods is at the cost of inferior on-device model performance.
On the other hand, adopting KD as model representation exchange protocols implicitly tolerates greater divergence of on-device models during training, which results in models receiving knowledge irrelevant to their own training goals \cite{cho2021personalized,zhang2021parameterized}.
Therefore, it is necessary to reconstruct the transferred knowledge so that on-device models can capture more useful information and better exploit the unique advantages of KD-based FEL with acceptable performance in mobile edge networks.
\subsection{Open Problems}
Since the related research is still at an early stage, there are many open problems in KD-based FEL.
\textcolor{black}{In terms of device connectivity,} it is impractical to keep devices online all the time, and how to cope with devices offline and dropout in KD-based FEL systems remains unsolved.
\textcolor{black}{Considering training acceleration,} how to schedule devices in KD-based FEL to achieve efficient asynchronous training has so far never been studied.
\textcolor{black}{When it comes to incentive mechanism, ways of} measuring device contributions in the KD-based FEL and creating incentives to keep devices consistently motivated to participate in FEL training is also worth attention.
\textcolor{black}{Discussing about privacy protection,} how to encode knowledge in KD-based FEL systems for counteracting inversion attacks is also worth concerns, especially for methods that upload features \cite{he2020group,wu2023fedict} or transformed data \cite{oh2020mix2fld,song2022federated} during training which are relatively easy to be attacked.
\textcolor{black}{Last but not least,} many prior works \cite{zhou2021source,itahara2021distillation,sattler2021cfd,cheng2021fedgems,cho2022heterogeneous,yu2022resource,taya2022decentralized} rely on a publicly available dataset that is inaccessible in real scenarios \cite{yu2021toward,wu2023fedict}, \textcolor{black}{thus novel KD-based FEL methods should provide normative solutions.}
\subsection{Practical Guidance for Real Deployment}
Applying KD-based FEL in practice is a multi-dimensional problem that requires ensembled techniques with performance trade-offs. We suggest that a practical FEL system with KD could include but not limited to resource-aware FEL architecture, model-agnostic representation exchange protocols without open datasets, knowledge adaptation and refinement, KD-specific knowledge compression, privacy-preserving knowledge transfer, techniques for learning from knowledge, solutions for complex communication channels in mobile edge networks, etc.
Only in this way can FEL be improved by KD, with wider applications in real-world scenarios.
\section{Conclusion}
\label{conclude}
Taking the challenges faced by Federated Edge Learning (FEL) as the main clue, this paper surveys prior works on applying Knowledge Distillation (KD) to FEL, and classifies the role of KD in FEL into four types of knowledge transfer, model representation protocol, component of backbone algorithm and dataset distillation. \textcolor{black}{In addition, the deployment modes of KD are categorized as edge-side, end-side, and edge-end collaboration.}
We also provide guidance for future research directions and real deployment of KD-based FEL.
In principle, KD can effectively help address the core challenges in FEL and be deemed a functional tool for knowledge transfer and model collaborative training.
In addition, we suggest that future research on KD-based FEL should focus on improving model accuracy, while paying attention to open problems we raised. Real deployment requires reasonable integration of existing techniques and multifaceted performance trade-offs.
~\\
\noindent
\textbf{Acknowledgment.} We thank Hui Jiang, Yuchen Zhu and Tianliu He from Institute of Computing Technology, Chinese Academy of Sciences for inspiring suggestions.
\bibliographystyle{splncs04}
|
\section{Introduction}
\label{sec1}
Speech emotion recognition (SER) is the process of automatic prediction of speaker's emotional state from his/her speech samples. A speech sample generally remains enriched with various information, such as speaker, language, emotion, context, recording environment, gender and age, intricately entangled to each other~\cite{krothapalli2013speech}. Human mind is congenitally trained to disentangle such information, however, same is not true for machines~\cite{picard2003affective}. Machines need to be specifically trained to extract cues pertaining to a particular information. Among such, extraction of emotion-specific cues for SER is still considered a challenging task. The challenge basically persists because of the differences in the manner of emotion expression across individuals \cite{el2011survey}. These differences stem from factors such as speaker's culture and background, ethnicity, speaker's mood, gender, manner of speech, etc.~\cite{el2011survey, SHAHFAHAD2021}. For automatic SER, a machine should be capable of extracting emotion-specific cues in the presence of all such variabilities. \\
SER finds application in several human-computer interaction domains such as sentiment analysis in customer service, health care systems, self-driving vehicles, auto-pilot systems, product advertisement and analysis~\cite{krothapalli2013speech, el2011survey, akccay2020speech}. One of the first seminal works in SER was aimed towards emotion information extraction using different speech cues~\cite{dellaert1996recognizing}. Various works that followed discovered that \emph{speech prosody} (pitch, intonation, energy, loudness, etc.) contain significant information for emotion discrimination~\cite{eyben2010towards,eyben2015geneva,chen2012speech}. Similarly, several other works report that \emph{spectral features} (spectral flux, centroid, mel-frequency cepstral coefficients (MFCCs), etc.) and \emph{voice-quality features} (jitter, shimmer, harmonic-to-noise ratio (HNR), etc.) of speech are also important for SER~\cite{li2007stress}. For classification, these extracted features are processed with a classifier back-end such as \emph{support vector machine} (SVM), \emph{Gaussian mixture model} (GMM), and \emph{k-nearest neighbour} (k-NN) for emotion class prediction. These approaches which employ certain signal processing algorithm for feature extraction are termed hand-crafted feature based approaches for SER. Hand-crafted approaches enjoy the advantage of being interpretable, in terms of which feature or speech characteristic is more relevant for emotions, and are computationally inexpensive. However, hand-crafted features often suffer from \emph{curse of dimensionality}, especially when \emph{brute-force} method based SER system is used~\cite{batliner2011whodunnit}. \\
Recent advancements in signal processing have introduced \emph{deep neural networks} (DNN) into the speech processing domain. DNNs have the impressive ability by which, given the required data, they automatically learn to obtain a possible solution to pattern recognition problems. This is accomplished by automatically updating the DNN parameters so as to reduce the defined loss function and approach towards the local minima.
In SER system, deep networks are either used as automatic feature extractors or as classifiers for emotion class prediction.
Recently, a new deep learning paradigm is also introduced which performs both feature extraction and emotion classification in an end-to-end fashion. Along these lines, several works use \emph{convolutional neural network} (CNN) as automatic feature extractor for SER~\cite{zhang2017speech, mao2014learning}. In contrast, other approaches use hand-crafted methods for feature extraction which are then used as features for DNN classifier input~\cite{ghosh2016representation, zhao2019speech, issa2020speech}. To obtain an end-to-end solution for SER works in~\cite{trigeorgis,tzirakis2018end,tang2018end} have used DNNs where the initial layers extract the emotion-relevant features and final layers act as classifier. In recent years, deep learning methods have been consistently shown to outperform hand-crafted feature based SER techniques. \\
In spite of their tremendous success, DNNs have major practical disadvantages. One such disadvantage is the requirement of large labelled database for proper DNN training~\cite{rolnick2017deep}. In contrast to other speech classification problems, such as speech and speaker recognition, large speech corpora are not available for evaluating SER task. Various ethical and legal issues make it difficult to collect large dataset of natural emotional voices from real-world scenario~\cite{el2011survey, akccay2020speech}. To somewhat alleviate this issue, acted emotion recordings are generally used where skilled actors enact a predefined set of emotions. However, this approach is not considered very appropriate as acted emotions are often exaggerated versions of natural emotions~\cite{el2011survey, akccay2020speech}. Another disadvantage of DNN is its complexity. Due to large trainable parameter set and high non-linear relationship between input and output, DNNs are often termed \emph{black-box} models, which are very difficult to understand/interpret~\cite{lipton2018mythos, BARREDOARRIETA202082, kimura2020new}. As the training includes optimization of all DNN parameters, it takes much larger time for DNNs to train as compared to classical statistical methods (e.g., SVM or GMM). Hence, even though very appealing, DNN models are still far-off a completely optimized SER approach. \\
The above discussion leads to the conclusion that both hand-crafted and DNN based feature extraction methods have their own set of advantages and disadvantages. In this work, we aim to exploit the advantages of both the methods for improved SER performance. Our framework is similar to the combination of hand-crafted feature, in the form of time-frequency representation, and a DNN model for further feature enrichment as used in other SER works~\cite{ghosh2016representation, zhao2019speech, issa2020speech}. However, our approach incorporates the prior (domain) knowledge of speech processing in humans, i.e., early auditory and cortex-based processing of speech \cite{sshamma}, for an improved hand-crafted feature representation. Being data-driven, DNN-based machine learning approaches suffer in performance, especially when there are constraints over the training data, e.g., limited size, ethical concerns in recording and poor quality of data \cite{muralidhar2018incorporating}, all of which are relevant for SER databases. Evidences reveal that such disadvantages can be alleviated by the use of domain knowledge \cite{muralidhar2018incorporating, rueden2021} in hand-crafted feature generation. Further, regarding speech representations, spectrogram and mel-spectrogram are considered the \emph{de-facto} standard of time-frequency representations in SER. However, they encompass only the early auditory processing of speech and lack cortical information. \\
Inspired by this fact, we first employ a hand-crafted feature extraction technique which combines an emotion relevant early auditory representation with corresponding cortex-based representation of the speech for SER. These features are then processed by a deep convolutional neural network which further extracts the emotion relevant information. Two machine learning frameworks are used at the back-end: Convolutional network with fully connected layer, and convolutional layer for embedding extraction with SVM classifier for final emotion class prediction. Such combination of multi-stage hand-crafted feature with DNN at back-end more closely follows the natural speech processing workflow in humans where the auditory system captures the signal and extracts the features, which are then transmitted to the inner regions of brain for further analysis and understanding. The achieved improvement in performance over different databases further consolidates our hypothesis of two-staged hand-crafted speech processing for SER. Figure~\ref{flow_diag} provides a general overview of the two-staged processing framework in human auditory system for SER. \\
\begin{figure}[t!]
\centering
\includegraphics[scale=0.65]{Fig1_MSF_figure_new.png}
\caption{The two-staged speech processing in the human auditory system for SER. The input speech captured at ear is converted to a form similar to time-frequency (TF) representation by the early auditory processing filters present in cochlea. This representation is then passed on to the auditory cortex in the brain for processing with cortical filters. The highlighted part in brain image identifies the auditory cortex region of the brain. The cortical filter processing leads to a modulation spectrogram based representation. This is further processed by the inner regions of the brain to finally decode emotions. Our employed deep neural network, which is loosely based on the studies of the brain and nervous system, models the inner processing of the brain to identify the emotion classes from the input modulation spectrogram feature. The $h_{\omega_n^a}(t)$ and $g_{\omega_k^m}(t)$ depict the impulse response of $n$th early auditory and $m$th cortical processing filter, respectively. The Figure shows logarithm applied TF and modulation spectrogram representation.}
\label{flow_diag}
\end{figure}
In the next section (Section~\ref{lit_review}), we describe the relevant literature and discuss the motivation and major contributions of this work. Section~\ref{cqt_section} and \ref{mod_spec_section} provides a brief introduction to the early auditory and cortex-based feature representations used in this work. Section~\ref{exp_setup} describes the experimental setup used to perform the experiments. Section~\ref{res_and_dis} describes the results obtained with the proposed feature and comparison with the standard features followed by corresponding discussion. Finally, Section~\ref{conc_sec} includes the conclusive statements of the work.
\section{Related Works and Motivation}
\label{lit_review}
In this section, we provide a brief review of works related to the frequency localisation of emotions. We then discuss some works which describe the relevance of modulation spectrogram in speech processing. This is followed by description of the motivation of our proposed feature and the major contributions of this work.
\subsection{Literature Review}
Several studies, aimed towards analysing the importance of spectral frequencies, have reported the prominence of low frequencies in SER. Authors in~\cite{bou2000comparative} report the prominence of first formant frequency (F1) for recognition of \emph{Anger} and second formant (F2) for recognition of \emph{Neutral}. Studies performed in~\cite{goudbeek2009emotion} found that high arousal emotions, e.g, \emph{Anger}, \emph{Happy}, have higher average F$1$ value and lower F$2$ value. They also found that positive valence emotions (e.g., \emph{Happy}, \emph{Pride}, \emph{Relief}) have higher average F$2$ value. Authors in~\cite{bozkurt2011formant} also report discrimination between idle and negative emotions using the temporal patterns of first two formant frequencies. In~\cite{lech2018amplitude}, authors show that non-linear frequency scales (e.g., equivalent rectangular bandwidth (ERB), mel, logarithmic) when applied for sub-band partitioning and energy computation over discrete Fourier transform based spectrogram, results in improved SER accuracy. Such studies hint toward the requirement of a non-linear frequency scale based time-frequency representation with higher emphasis on low-frequency regions of speech. \\
Regarding human sound perception, evidences in literature suggest that the process of auditory signal analysis can be modelled into two stages: (i)~\emph{Early auditory stage}, which models the incoming audio signal into a spectrogram based representation. (ii)~\emph{Cortical analysis stage}, which extracts the spectro-temporal modulation relationship among different audio cues from the auditory spectrogram \cite{sshamma,tchi}. Such modelling strategy has been found effective in the analysis of both speech and music signals~\cite{sshamma}. The spectral and temporal modulation features of speech spectrogram are also highly related to speech intelligibility, noise and reverberation effects~\cite{tchi}. In~\cite{kumar2012features}, authors report that the spectro-temporal representation of audio (non-speech) signals with positive\textbackslash negative valence is different from that of neutral sounds. They also report that spectral frequency and temporal modulation frequency can represent the valence information of sounds.
Authors in~\cite{ARNAL20152051} compared the temporal modulations of human speech with scream voice and concluded that slow temporal variations ($< 20$~Hz) contain most linguistic (both prosodic and syllabic cues) information. \\
In speech analysis, temporal modulation features are called \emph{modulation spectral features} (MSFs). Owing to their relatedness to speech intelligibility, MSFs have been extensively used in speech processing. Some works also successfully explored modulation features for speaker identification, verification and audio coding~\cite{vuuren1998importance, nsephus}. Author in~\cite{hermanskyhistory} provides a comprehensive description of the history of the use of modulation features in speech recognition. \\
Modulation features have also been explored for emotion recognition in speech. Authors in~\cite{wu2011automatic} used MSF for emotion recognition and provided a detailed explanation of its relevance for SER. In~\cite{alam2013amplitude}, authors used a smoothed nonlinear operator to obtain the amplitude modulated power spectrum of the gammatone filterbank generated spectrogram and showed improvement over standard MFCC for SER. Authors in~\cite{Zhu+2016} studied the relationship between human emotion perception and the MSFs of emotional speech and concluded on the suitability of modulation features for emotion recognition. Authors in~\cite{peng2020} used 3-D convolutions and attention-based recurrent networks to combine auditory analysis and attention mechanisms for SER. This work also explains that the temporal modulations extracted from auditory analysis contain periodicity information important for emotion recognition. In~\cite{avila}, various feature pooling, such as \emph{mean}, \emph{standard deviation}, and \emph{kurtosis}, on frame-level measure of MSF to be used for ``in-the-wild'' dimension-based SER (dimensional SER includes projection of speech onto three emotion dimensions: \emph{valence}, \emph{arousal} and \emph{dominance)}. The authors report improvement in results over frame-wise modulation spectral feature baseline for various noise and reverberated speech scenarios. Similar MSF measures when used with Bag-of-Audio-Words (BoAW) approach showed SER improvement against environmental noise in~\cite{kshirsagar2022}. In~\cite{avila2019}, authors use modulation spectral features with convolutional neural networks to discriminate between stress-based speech and neutral speech. The authors show that the modulation spectral features when used with CNN with the time frames intact (without statistics pooling over time frames of MSF) gives better performance, especially over increased number of target emotion classes. In~\cite{yeh2010spectro}, authors show that joint spectro-temporal modulation representation outperforms standard MFCC in emotion classification of noisy speech. Recently, the authors in~\cite{PENG2021261} have also used MSF over cochleagram features with a long short term memory (LSTM) based system for dimensional SER. The work explains that arousal information can be characterised by the amplitude envelope of speech signal, whereas valence information is characterised by the temporal dynamics of amplitude envelope. Since it is difficult to obtain such dynamics from low-level descriptor (LLD) features, auditory analysis based temporal modulation features can potentially represent the required temporal dynamics for SER.
\subsection{Motivation and Contributions}
The literature in SER reveals two important speech characteristics for emotion prediction: the importance of low frequencies, and the importance of temporal modulations of spectrogram. To address the importance of low-frequency information, we use constant-Q transform (CQT) based time-frequency representation for SER. CQT provides higher frequency resolution and increased time invariance at low frequencies thereby emphasizing the low-frequency regions of speech~\cite{SINGH2022103712}. This helps in better resolution of emotion salient frequency regions of speech and improved SER performance~\cite{singh2021}. CQT is also known to provide a representation with visible pitch frequency and well-separated pitch harmonics~\cite{chandra}. Because of high relevance of pitch information in emotion discrimination, this property of CQT makes it more suitable for SER over standard mel-based features. \\
To further enhance the CQT-based system while utilising the understanding of domain knowledge of human auditory-cortical physiology~\cite{tchi}, we propose to use temporal modulation of CQT spectrogram representation for SER. Specifically, we use CQT spectrogram representation for auditory analysis and extract temporal modulations of CQT by again using constant-Q filters, for cortical analysis. In this way, we obtain the temporal modulation of emotion salient low-frequency regions which are emphasized by CQT. Studies show that such use of constant-Q modulation filterbank better approximates the cortical sound processing in humans~\cite{suki2002, suki2004}. The constant-Q factor characteristic of modulation filters also lead to higher resolution at lower modulation frequencies, hence, providing an arrangement that helps in identifying any deviation from general (or \emph{Neutral}) speech modulation rate ($2$-$4$~Hz)~\cite{hermansky2011speech}. Our choice of constant-Q filters in both stages is also inspired from the study of early auditory and cortical stages of mammalian auditory cortex~\cite{tchi, zotkin2003}. We term our proposed feature as \emph{constant-Q transform based modulation spectral feature} (CQT-MSF). A 2-dimensional convolution neural network architecture (2-D CNN) is used to further refine the emotion information present in CQT-MSF feature. We compare the performance of CQT-MSF with mel-frequency spectral coefficients (MFSC) and show that the constant-Q non-linearity based auditory-cortical features outperform the mel-scale non-linearity based features. We also investigate the performance differences obtained with auditory and cortical representations taken separately. We also highlight the striking similarity of CQT-MSF with the wavelet-based time-shift and deformation invariant coefficients, known as scattering transform coefficients~\cite{anden}. Our main contributions in this work are as follows:
\begin{itemize}
\item This study proposes a new human auditory-cortical physiology based SER framework.
\item We propose a modulation feature extraction technique using constant-Q filterbank over constant-Q spectrogram and analyse its relevance from vocal emotion perspective.
\item We perform similarity analysis with another two-staged auditory-cortical feature representation: Scattering transform.
\item We also perform explainability analysis to visually inspect different regions of CQT-MSF that weigh the most in prediction of a particular emotion class.
\item The study further hints correlation between music training and emotion understanding by discussing the case of \emph{Amusia}~\cite{sydney2021}, and the possible analogy between modulation computed over CQT spectrogram and the cortex-level processing of sound in music trained individuals~\cite{dmitrieva2006ontogenetic, fuller2014musician, twaite2016examining, weijkamp2017attention, thompson2004decoding, nussbaum2021}.
\end{itemize}
\section{Early auditory processing: Constant-Q Transform (CQT)}
\label{cqt_section}
Our proposed features are based on the use of constant-Q filterbanks for both time-frequency (early auditory) and temporal modulation (cortical) based analysis of speech. In this section, we briefly discuss the CQT method of time-frequency representation. CQT uses constant \emph{quality factor} (Q-factor) bandpass filters with logarithmically spaced center frequencies~\cite{todisco2017constant}. Mathematical formulation of constant-Q transform is given by,
\begin{equation}
X^{CQT}[k,n]~=\sum_{j~=~n-\floor{N_K/2}}^{n+\floor{N_k/2}}~x(j)a_k^*(j-n+N_k/2)
\label{eq:CQT}
\end{equation}
\noindent where $k$ denotes the CQT frequency index, $\floor{.}$ denotes the rounding-off to nearest integer towards negative infinity and $a_k^*(n)$ is the complex conjugate of the CQT basis function for $k$\textsuperscript{th} CQT bin. The CQT basis, or the time-frequency \emph{atom}, is a complex time domain waveform given as,
\begin{equation}
a_k(n)~=~ \frac{1}{N_k}w\left(\frac{n}{N_k}\right) exp\left[-i2\pi n\frac{f_k}{f_s}\right]
\end{equation}
\noindent where $f_k$ is the center frequency of $a_k$, $f_s$ is the sampling frequency and $w(n)$ is the window function with length $N_k$. We use the standard \emph{Hann} window in this work for CQT computation. The center frequencies of filters in constant-Q transform are spaced by the relation $f_k~=~f_{\mathrm{min}}2^\frac{k-1}{B}$ where $f_k$ is the frequency of $k$th filterbank, $f_{\mathrm{min}}$ being the frequency of the lowest bin and $B$ the number of frequency bins used per octave of frequency. This binary logarithmic spacing leads to more frequency bins at lower frequencies, as compared to high frequencies, and hence provides higher frequency resolution at low frequencies~\cite{todisco2017constant}. In time domain, such filters can be given as truncated sinusoids (e.g., truncated with \emph{Hann} window) with different lengths~\cite{schorkhuber2010constant}, given by,
\begin{equation}
N_k~=~\frac{q f_s}{f_k(2^{\frac{1}{B}}-1)}
\label{eq:scale}
\end{equation}
\renewcommand{\thesubfigure}{\alph{subfigure}}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\hbox{\hspace{-0.7cm}\includegraphics[scale=0.23]{Fig2_1_Aud_filt.png}}
\caption{Early auditory filterbank}
\label{early_aud_filters}
\end{subfigure}%
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[scale=0.23]{Fig2_2_Mod_filt_scale2.png}
\caption{Modulation filterbank}
\label{cortical_filters}
\end{subfigure}%
\caption{Auditory and Modulation filter banks used in CQT-MSF. The modulation filters shown here have scale factor ($q$) value 2.}
\label{CQT-MSF filterbanks}
\end{figure}
\noindent where $q$ is the filter scaling factor. This scaling factor offers to change the time (and hence frequency) resolution of CQT bases without affecting $B$~\cite{schorkhuber2010constant}. When compared with mel-based features, the mel-scale is also logarithmic in nature. However, mel-scale uses a decadic logarithm scale (or natural logarithm in some implementations), because of which, the emphasis on low-frequencies is not as prominent as CQT. \\
The computation of CQT, as described in Eq.~\ref{eq:CQT}, includes convolution of \emph{atom} with every time sample of the input signal. However, the fast CQT computation algorithm~\cite{schorkhuber2010constant} introduced a \emph{hop length} parameter, which describes the number of samples the time window is shifted for next time frame CQT computation. The hop length is kept equal to integer multiples of $2^ {\mathrm{No. \ of \ octaves}}$ so that the corresponding signal frames at different frequencies do not fall out of alignment \cite{schorkhuber2010constant}. In CQT representation, the number of octaves is given by $\log_{2}{\frac{F_{\mathrm{max}}}{F_{\mathrm{min}}}}$ \cite{todisco2017constant} where $F_{\mathrm{min}}$ and $F_{\mathrm{max}}$ are the minimum and maximum frequency of operation, respectively. For CQT computation in this work, we use the \emph{LibROSA}\footnote{\url{https://librosa.github.io/}} toolkit~\cite{brian_mcfee_2021_4792298} which follows all the computational details of the fast CQT implementation mentioned above.
\section{Cortex-based processing: Modulation Spectrogram}
\label{mod_spec_section}
Modulation spectrogram shows the temporal variation pattern of the spectral components in spectrogram. According to~\cite{hermansky2011speech}, speech signal is composed of two parts, the carrier, i.e., the vocal cords excitation, and the varying modulation envelope which is the result of changes in orientation of different vocal organs over time. The low-frequencies of the modulation envelope characterise slow variations of the complete spectral structure, which is known to encode most of the phonetic information~\cite{greenberg1997, hermanskyhistory, elhilali2019modulation}. Let $S(t, ~\omega)$ be the speech spectrogram. The temporal evolution of a frequency bin $\omega_o$ in $S(t, ~\omega)$, over time $t$, is a one-dimensional time-series. The spectral representation of this time-series $S(t, ~\omega_o)$ constitutes the modulation spectrum of frequency bin $\omega_o$ over $T$, where $T$ is the spectrogram time window (with duration equal to window length $N_k$). \\
For speech, most of the modulation energy remains concentrated around $2$-$4$~Hz range with peak at $4$~Hz~\cite{hermansky2011speech}. This makes $4$~Hz to be considered as the syllabic rate of normal (\emph{Neutral}) speech. Deviations from this rate generally result from infliction of noise or reverberation effects over speech~\cite{elhilali2019modulation, moritz2011}. It is studied in SER literature that rate of speech is higher than \emph{Neutral} class for high arousal emotions, such as, \emph{Anger}, \emph{Fear} and lower for low arousal emotions, such as \emph{Sad}, \emph{Boredom}~\cite{banse1996acoustic}.
Hence, this deviation of modulation energy peak from $4$~Hz can be used for emotion discrimination over arousal scale.
\begin{figure}[t!]
\centering
\includegraphics[scale=0.66]{Fig3_block_diagram.png}
\caption{Block diagram of proposed CQT-MSF feature extraction method. In figure, $\omega^a$ refers to acoustic frequency and $\omega^m$ refers to the modulation frequency.}
\label{block_diagram}
\end{figure}
\subsection{Constant-Q based Modulation Spectral Features (CQT-MSF)}
\label{cqmsf}
\begin{figure}[t]
\centering
\hbox{\hspace{-0.4cm}\includegraphics[scale = 0.35]{Fig4_Analysis_scale2.png}}
\caption{Modulation spectral features (averaged over time) for different emotions in EmoDB. The `\textbf{M F}' refers to the modulation frequency channels and `\textbf{A F}' refers to the acoustic frequency channels. The modulation filters used in this analysis has filter scale ($q$) value 2.}
\label{cqt_msf_analysis}
\end{figure}
In this subsection, we compute modulations of CQT bins and combine them with CQT spectrogram to generate CQT-MSF. The first stage early auditory analysis (CQT spectrogram) in the CQT-MSF feature can be given as,
\begin{equation}
Y(t, ~\omega^a) = s(t) * h_{\omega_n^a}(t); \ \omega_0^a \leq \omega_n^a < \omega_C^a,
\end{equation}
\noindent where, $s(t)$ is the input speech signal, $h_{\omega_n^a}(t)$ is the impulse response of $n$th constant quality factor auditory filter with $\omega_n^a$ center frequency, $C$ is the number of auditory filters and $Y(t, ~\omega^a)$ is the corresponding time-frequency representation. Fig.~\ref{early_aud_filters} shows the frequency response of different $h_{\omega_n^a}(t)$ used. For envelope extraction, modulus operation is applied over $Y(t, ~\omega^a)$, i.e., $|Y(t, ~\omega^a)|$. The resulting representation provides the temporal trajectories of different frequency bins in $Y(t, ~\omega^a)$. \\
For cortical analysis, the $|Y(t, ~\omega^a)|$ is passed through a modulation filterbank. The modulation spectrogram computed over time-frequency representation $|Y(t, ~\omega^a)|$, is given as,
\begin{equation}
Y(t, \omega^a, \omega^m) = |Y(t, ~\omega_n^a)| * g_{\omega_k^m}(t); \ \omega_0^m \leq \omega_k^m < \omega_M^m, ~\omega_0^a \leq \omega_n^a <\omega_C^a,
\label{eq2}
\end{equation}
\noindent where, $g_{\omega_k^m}(t)$ is the impulse response of $k$th modulation filter with $\omega_k^m$ center frequency and $M$ is the total number of modulation filters. Similar to the output of the first stage, we use the modulus of computed modulation spectrum coefficients computed over all frequency bins of CQT spectrogram, i.e., $|Y(t, \omega^a, \omega^m)|$~\cite{hermansky2011speech}. Fig.~\ref{block_diagram} shows the block diagram of CQT-MSF feature extraction. The complete MSF includes concatenation of temporal modulations, computed using every modulation filter ($g_{\omega_0^m} \leq g_{\omega_k^m} < g_{\omega_M^m}$), of all frequency bins in the time-frequency representation, i.e., for $\omega_0^a \leq \omega_n^a < \omega_C^a$ bins where $C = 24$ auditory channels in our experiments. Regarding the properties of MSF, study performed in~\cite{tchi} report distinction between three different temporal modulation rates: slow, intermediate, and fast. The slow modulation rate is shown to roughly correspond to the syllable or speaking rate. Whereas the intermediate modulation rate appearing because of interharmonic interaction is shown to reflect the fundamental frequency of the signal. This shows the importance of temporal modulation for pitch representation, and hence, SER. Temporal modulation extracted by MSF represent tempo~\cite{DING2017181}, pitch, and timber~\cite{zotkin2003}, all of which are related to emotion information in speech. \\
Fig.~\ref{cqt_msf_analysis} shows the time-averaged CQT-MSF coefficients for utterances of different emotion classes of the EmoDB database. The MF and AF refer to the modulation and auditory frequency channels, respectively. In terms of modulation frequency, the highest peak in \emph{Neutral} emotion is observed at $4$~Hz modulation frequency with another peak around $0.5$~Hz. Compared to \emph{Neutral} class, low arousal emotions (\emph{Boredom} and \emph{Sad}) also have energies extending towards $0$-$4$~Hz modulation frequency range. High arousal emotions (\emph{Anger}, \emph{Fear}) have peak around $4$-$8$~Hz modulation frequency range. In contrast, \emph{Happy} also has a peak at $4$~Hz similar to \emph{Neutral}. Similarly, \emph{Disgust} also shows a peak at $4$~Hz followed by another peak at $2$~Hz. From AF perspective, \emph{Anger} emotion shows peak at high frequencies (high AF), whereas, in \emph{Sad} low auditory frequencies are more dominant. For remaining emotions, auditory energy distribution extends almost similarly over mid-auditory frequencies. This analysis shows the higher emotion discrimination potential of combined MF and AF channels as compared to only AF channel-based representation. \\
\begin{figure}[t!]
\centering
\hbox{\hspace{-1cm}{\includegraphics[scale = 0.4]{Fig5_Modspec_F_ratio_Q=2.png}}}
\caption{F-ratio values of different auditory frequency (AF) and modulation frequency (MF) bins of modulation spectrum features computed over EmoDB database. The F-ratio is calculated over time-averaged modulation spectrum features (MSF) between \emph{Neutral} and every other emotion class. The modulation filters used in this analysis have filter scale ($q$) value 2.}
\label{f_ratio_modspec}
\end{figure}
To further analyze the discriminative potential of modulation spectrum features, we perform F-ratio analysis between the time-averaged modulation features of various emotion classes and the \emph{Neutral} class of the EmoDB database~\cite{LU2008, dipjyoti2017}. Fig.~\ref{f_ratio_modspec} shows the $3$-D projection of F-ratio values in AF-MF plane. Different auditory and modulation bins show varying discriminative characteristics for different emotions. For every emotion class, F-ratio peaks are observed at low MF bins showing their potential for emotion discrimination. Similarly, low AF also shows high F-ratio for every class except \emph{Sad}. High arousal emotions (\emph{Anger}, \emph{Happy}, \emph{Fear, etc.}) in general show greater F-ratios at high MF bins as compared to low arousal emotions (\emph{Sad}, \emph{Boredom}). The highest F-ratio value with respect to \emph{Neutral} is observed for \emph{Anger} and lowest for \emph{Boredom} emotion class. For \emph{Anger} and \emph{Happy} classes, low AF bins exhibit higher discriminative characteristic. \emph{Disgust} shows F-ratio peaks over wide range of AF and corresponding MF bins. In \emph{Fear}, F-ratio peaks are observed at low and high AF values with gradual slope towards increasing MF bins. The presence of moderately higher F-ratio values at MF bins is a result of increase in speaking rate in high arousal emotions.
\emph{Boredom} class has lowest F-ratio values, mostly focused at low AF and low MF bins, whereas, \emph{Sad} shows higher discrimination w.r.t. \emph{Neutral} over low and high AF and MF bins. Lower F-ratio values for \emph{Boredom} also indicate its similarity in characteristics with \emph{Neutral} class. The F-ratio analysis again shows the higher discrimination potential of joint AF and MF bins, with respect to \emph{Neutral} emotion.
\subsection{Comparison between CQT-MSF and Scattering Transform}
\label{comp_scat_msf}
Our proposed CQT-MSF feature, combined with standalone CQT, has striking similarity with \emph{scattering transform} feature representation of $1$-D signals. Authors in \cite{anden} compute scattering coefficients of $1$-D signals and show their characteristic invariance against temporal shifts and deformations. The features (or coefficients) are computed by convolving the signal with a set of predefined filter kernels. The feature extraction process includes the following steps: 1) Scalogram computation by passing the signal through a bank of wavelet filters. 2) Passing the obtained time-series of frequency bins in scalogram through another set of wavelet filterbank to obtain modulation spectrogram. 3) Introduce stability to deformations by low-pass filtering the signal, scalogram and modulation spectrogram coefficients. The scattering transform coefficients are mathematically described as,
\begin{equation}
S_{J_2} x(t) = U_2 x*\phi_{2^J}(t) = \int U_2 x(u)\phi_{2^J}(t-u)du
\end{equation}
\noindent where,
\begin{equation}
U_2 x = U[\lambda_2]U[\lambda_1]x = ||x*\psi_{\lambda 1}|*\psi_{\lambda 2}|.
\end{equation}
Here, $x$ is the $1$-D signal, $\phi_{2^J}(t)$ defines the averaging low-pass filter with scale $2^J$, $\psi_{\lambda_N}$ describes the $N$th layer complex \emph{Morlet} wavelet filterbank (layer $1$ are scalogram coefficients and layer $2$ constitutes modulation coefficients) and operator `$*$' is the convolution operator. Scattering coefficients are found useful in various speech and audio processing domains, e.g., speech recognition \cite{anden}, speaker identification \cite{ghezaiel:hal-03086433}, urban and environmental sound classification \cite{bauge2013}, etc. In \cite{singh2021deep}, scattering coefficients also showed improvement in SER performance over mel-frequency cepstral coefficients (MFCCs).\\
\renewcommand{\thesubfigure}{\alph{subfigure}}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\hbox{\hspace{-0.7cm}\includegraphics[scale=0.6]{Fig6_1_stft.png}}\caption{STFT}
\end{subfigure}%
\begin{subfigure}[t]{0.5\textwidth}
\includegraphics[scale=0.6]{Fig6_2_cq.png}\caption{CQT}
\end{subfigure}%
\vspace{0.75cm}
\begin{subfigure}[t]{0.6\textwidth}
\includegraphics[scale=0.6]{Fig6_3_lin_cq.png}
\caption{CQT with linear frequency scale}
\end{subfigure}
\caption{Visual description of deformation stability of STFT and CQT. a), Signal $x(t)$ (left) and its deformed version $x`(t)$ (right). b), CQT representation of the same signal and its deformed version. c) CQT of the original and deformed signal projected on linear frequency scale. Figure taken with permission from~\cite{SINGH2022103712}.}
\label{deform_stable}
\end{figure}
In our proposed CQT-MSF feature, the CQT time-frequency representation is similar to the first layer scalogram coefficients computed by scattering transform. Similarly, the MSF computed over CQT is similar to the modulation spectrogram computed over scalogram in second layer of scattering transform. Also, CQT follows the same constant-Q non-linearity as followed by the filterbanks in both first and second layers of the scattering transform. However, the averaging performed in scattering coefficients to obtain invariance to time-shift and deformations is absent in CQT-MSF. The design parameter (e.g., bandwidth) of the low-pass filter which performs this averaging in scattering transform is manually selected depending upon the input signal characteristics. To address the absence of time-shift invariance, we employ a $2$-D convolutional neural network over the computed CQT-MSF. The employed CNN architecture includes multiple layers with different filter scale values in different layers. According to \cite{wiatowski}, convolutional neural networks inherently exhibit invariance in vertical direction (direction of network depth) which mainly appears due to feature pooling. Hence, a generic CNN architecture with pooling layers can learn to apply the required averaging and obtain the required time-shift invariance characteristic.\\
Regarding sensitivity to deformation, authors in \cite{wiatowski} and \cite{philipp} prove that convolutional feature extractors provide inherent but limited deformation stability. The extent of this stability depends upon the deformation sensitivity of the input signal. Signals which are slowly varying or band-limited are more deformation insensitive than signal with sudden changes or discontinuities \cite{philipp}. As the CQT also provides a non-uniform filterbank representation as provided by mel-filters, similar stability to temporal deformation can be assumed in CQT filterbank as well. Fig.~\ref{deform_stable} shows the deformation stability of STFT, CQT and CQT with linear frequency scale for a signal $x(t)$ deformed by a factor $\epsilon t$ (i.e, $x'(t) = x(t-\epsilon t) = x((1-\epsilon)t)$) \cite{anden}. The upward shift of spectral response in STFT appears due to the `$\epsilon t$' term, leading to instability to deformation imposed by $\epsilon t$. However, in CQT, spectral responses of deformed signal do not show any major frequency shift. Instead, the deformed signal well-overlaps with the original signal because of higher filter bandwidth at higher frequencies. This shows that CQT is indeed deformation stable as compared to STFT. The linear frequency CQT plot is given for comparison of STFT and CQT over linear frequency scale, hence confirming the deformation stability of CQT in both linear and non-linear frequency scales. \\
Therefore, convolutional neural network layers can be used to inherently provide the required time-shift and deformation invariance for better emotion-rich representation at its output. We hence write the feature extracted by convolution layers of our employed DNN model as,
\begin{equation}
S = F(||x*\psi_{\lambda1}|*\psi_{\lambda2}|)
\end{equation}
\noindent where, $F(.)$ is the function estimated by $2$-D convolution layers, and $\psi_{\lambda1}$ and $\psi_{\lambda2}$ corresponds to the filterbanks $h_{\omega_n^a}(t)$ and $g_{\omega_k^m}(t)$ used in the CQT-MSF generation (Fig.~\ref{CQT-MSF filterbanks}). Another point of dissimilarity is the difference between the basis functions used in the filterbank of scattering transform and CQT-MSF. The former uses \emph{Morlet} wavelets, whereas, the latter employs sinusoids multiplied with \emph{Hann} window function. The ripples observed in the frequency response of the filters in Fig.~\ref{CQT-MSF filterbanks} is because of the small spectral leakage in the \emph{Hann} window.
\section{Experimental Setup}
\label{exp_setup}
\subsection{Database Description}
\label{data_used}
For analysis of CQT-MSF and its comparison with mel-scale features, we perform experiments with two different speech corpora. We use Berlin EmoDB and RAVDESS datasets which are most widely used and publicly available.
\subsubsection{Berlin Emotion Database (EmoDB)}
Berlin Emotion Database \cite{burkhardt2005database} contains acted emotional speech recordings of $10$ professional artists ($5$ female and $5$ male). The actors speak ten emotionally neutral and phonetically rich sentences in German language. Seven different emotion categories are used in the database: \textit{Anger, Happy, Fear, Sad, Boredom, Disgust,} and \textit{Neutral}. To evaluate the authenticity of recordings, listening test was performed by $20$ subjects. A total of $800$ utterances were recorded but only $535$, having more than $80\%$ recognition rate and $60\%$ naturalness, were finally selected. Our choice of this database is explained by its diligent recording setup, popularity in SER domain \cite{eyben2015geneva, zhang2017speech, wu2011automatic, mao2014learning, bitouk2010class, wang2015speech, deb2018multiscale, Ntalampiras2012} and its free availability.
\subsubsection{Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS)}
The RAVDESS database \cite{livingstone2018ryerson} contains acted utterances of $12$ male and $12$ female artists speaking English language. A total of $7536$ clips were recorded in three different modalities, namely audio-only, video-only, and audio-video, out of which the audio-only modality contains $1440$ spoken utterances from all speakers. The database includes eight different emotion categories (\textit{Happy, Anger, Sad, Neutral, Disgust, Calm, Surprised,} and \textit{Fear}) with two intensity levels, strong and normal. Recorded clips were evaluated by $319$ subjects out of which $247$ tested the validity and $72$ evaluated test-retest reliability of recordings. An average of $60\%$ accuracy was obtained in validity test over recordings of all emotions. Up-to-date design and inclusion of an extensive emotion set with varying intensities make this an important database for SER.\\
To further increase the diversity in training data, five-fold data augmentation following the x-vector \emph{Kaldi}\footnote{\url{https://github.com/kaldi-asr/kaldi/tree/master/egs/voxceleb/v2}}~recipe is used \cite{snyder2018x}. The augmented data involves adding additive and reverberation noises over clean speech samples. The RAVDESS database is downsampled to $16$~kHz before data augmentation and feature extraction.
\subsection{Parameter settings for feature extraction}
We compare the performance of proposed CQT-MSF features with baseline CQT and MFSC. The different parameter values used for CQT and MFSC are based on our preliminary comparison of the two methods~\cite{singh2021}. For CQT computation, we select the minimum frequency value $F_{\mathrm{min}}$ to be $32.7$~Hz and $F_{\mathrm{max}}$ equal to the Nyquist frequency. This provides a total of eight octaves over complete frequency range. Every frequency octave contains three bins which provides a total of $24$ frequency bins over complete frequency range ($F_{\mathrm{min}}$ to $F_{\mathrm{max}}$). The obtained CQT representation corresponds to $24$-channel early auditory stage feature extraction. Another important parameter in CQT computation is hop length which is the number of samples the observation window advances between successive frame-shifts. We keep the hop length value fixed at $64$. For cortical analysis, the above-mentioned CQT representation is passed through another set of constant-Q filterbank, referred to as modulation filterbank, as described in Section~\ref{mod_spec_section}. We use an 8-channel modulation filterbank with center frequencies ranging from $0.5$ to $64$~Hz covering a total of eight frequency octaves.\\
For a fair comparison with CQT and CQT-MSF, similar modification in parameter values of MFSC is applied. We use $24$-filter bank for MFSC computation over STFT with $512$ frequency bins. The frame size is fixed to $320$ samples with hop length of $64$ samples over $16$~kHz sampling frequency. We also compute the modulation spectrum coefficients by using MFSC time-frequency representation for comparison with CQT-MSF features. We refer to these as MFSC-MSF features. We use the same \emph{LibROSA} toolkit for MFSC feature generation.
\subsection{Evaluation Methodology}
\label{eval_metho}
Unlike other speech processing tasks, such as automatic speech recognition (ASR), automatic speaker verification (ASV), the SER lacks standardization of evaluation methodology for performance benchmarking on publicly available datasets. This leads to a wide variation in results across different works reported in the literature. Some examples of differences in evaluation protocol are the use of some selected emotions from the databases, the choice of performance evaluation metric, the selection of cross-validation strategy, the differences in the selected emotion classes, etc. Because of these reasons, meaningful comparison of obtained results with those reported in the literature becomes inaccurate, if not impossible, in SER research. \\
We adopt a leave-one-speaker-out (LOSO) cross-validation strategy for evaluation and benchmarking. The databases are divided into train/validation/test groups with every group containing disjoint speakers. The test and validation group contain utterances from one speaker and the remaining speakers are kept for training. This keeps the total number of train/validation/test sets equal to the number of speakers in every database. The final performance over a database is reported by averaging the performance metrics obtained for every train/validation/test group. For SER, speaker-dependent testing is known to fair better than speaker-independent testing \cite{schuller2005speaker}. However, speaker-independent sets of the database eliminate the chances of the trained classifier being biased towards a set of speakers and also simulates the real-world scenario in a better way. Although LOSO cross-validation is computationally expensive, due to small database sizes in SER, the increase in complexity can be safely ignored. Also, keeping a single speaker for testing allows more training data to be available which is essential with small databases. \\
\begin{table}[t!]
\centering
\caption{The parameters of CNN architecture for SER. The number of $2$D-Conv layers and the kernel sizes are inspired from the x-vector TDNN architecture~\cite{snyder2018x}. Maxpooling applied after every $2$D-Conv layer provides time and frequency invariant feature representations.}
\begin{adjustbox}{width=0.575\columnwidth,center}
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{cccc}
\hline
\textbf{Layer} & \begin{tabular}[c]{@{}c@{}} \textbf{No. of} \\ \textbf{Filters} \end{tabular} & \begin{tabular}[c]{@{}c@{}} \textbf{Height} \\ \textbf{(Frequency)}\end{tabular} & \begin{tabular}[c]{@{}c@{}} \textbf{Length} \\ \textbf{(Time)}\end{tabular} \\ \hline
\hline
2-D Conv & 128 & 5 & 5 \\
Maxpool & - & 2 & 1 \\
2-D Conv & 128 & 3 & 3 \\
Maxpool & - & 2 & 1 \\
2-D Conv & 128 & 3 & 3 \\
Maxpool & - & 2 & 1 \\
2-D Conv & 128 & 1 & 1 \\
Maxpool & - & 2 & 1 \\
\begin{tabular}[c]{@{}c@{}} Global Average \\ Pool (GAP) \end{tabular} & - & - & - \\
Fully Connected & 64 & - & - \\
Softmax & \#Classes & - & - \\
\hline
\end{tabular}
\end{adjustbox}
\label{cnn_param_tab}
\end{table}
\subsection{Classifier Description}
\label{classifier_desc_sec}
In this work, we use two different machine learning frameworks to evaluate performances of studied features: (1)~convolutional neural network with fully connected layer for emotion classification (termed henceforth as DNN). (2)~Convolutional layers to extract emotion embeddings and SVM to classify embeddings into emotion classes (termed as DNN-SVM). Our selection of these is inspired from the success of embeddings based networks~\cite{zhang2017speech, snyder2018x, dawalatabad21_interspeech} and fully DNN-based frameworks~\cite{peng2020} in speech processing. Performance evaluation over these also enables us to compare the SER efficiency of the two DNN frameworks. \\
The DNN-SVM framework comprises of two parts: embedding extraction from convolutional layers of the trained model used in the DNN framework, and final classification using SVM. The SVM is trained and tested over the embeddings extracted from the trained DNN model.
We extract embeddings at the output of global average pooling (GAP) layer, placed after the final convolutional layer. These embeddings are processed with an SVM back-end for performance evaluation. In SVM model, we empirically select the value of regularization parameter $C$ and the expanse/width of the \emph{radial basis function} kernel (parameter $\gamma$) to $1$ and $0.001$ respectively~\cite{libsvm}. \\
In the DNN framework, to train and validate the model, the speech utterances from every database are chunked into segments of $100$ frame length with $50\%$ overlap across consecutive frames. With $64$ samples hop and $16$~kHz sampling rate, this corresponds to $400$~ms speech duration. Our choice of $400$~ms is based on the reports in SER literature which explain that segment length greater than $250$~ms contains required information for emotion prediction~\cite{zhang2017speech}. However, for testing, complete utterances are used to test model performance. This is done as the labels provided to emotion speech recordings are over complete utterances and not over segments. This approach also leads to increase in available data samples for training. Similarly, in DNN-SVM framework, the train embeddings are generated over segments of speech utterances with $100$ frame length ($400$~ms), whereas, test embeddings are generated from complete utterances. \\
Table~\ref{cnn_param_tab} describes the DNN architecture employed in this work. We use cross entropy optimizer with learning rate value of $0.001$ with $64$ batch size and dropout value of $0.3$ applied over only the fully connected (FC) layer. The model is trained for $50$ epochs and the version with the best performance on validation set is used for testing. \\
\begin{figure}[t]
\begin{subfigure}[t]{0.55\textwidth}
\hspace{-0.25cm}
\includegraphics[scale=0.45]{Fig7_1_CQT_MSF_plot_W.png}
\caption{CQT-MSF of \emph{Anger} emotion}
\end{subfigure}%
\begin{subfigure}[t]{0.55\textwidth}
\hspace{-0.25cm}
\includegraphics[scale=0.45]{Fig7_2_CQT_MSF_plot_T.png}
\caption{CQT-MSF of \emph{Sad} emotion}
\label{cqt_msf_example_figure}
\end{subfigure}%
\caption{Logarithm of feature fusion of CQT and modulation spectral features, i.e., CQT-MSF, extracted from CQT over utterances taken from EmoDB database. The first $24$ bins on y-axis correspond to the CQT spectrogram. Bins that follow include stacking of $8$ modulations bins corresponding to every auditory bin. The total number of bins then becomes, $24$ auditory bins $+$ $24$ auditory bins $\times~8$ modulation bins for every auditory bin$~=~24~$auditory bins$~+~192~$MSF bins$~=~216$ total bins on y-axis).}
\label{CQT-MSF-plot}
\end{figure}
As the convolution layers accept input in $2$-D (time-frequency) form, we use two different strategies to combine CQT/MFSC time-frequency representation with their MSF features. In the first method, we directly concatenate the CQT/MFSC with its corresponding MSF features over the frequency axis. This leads to a representation with time frames in x-axis and early auditory frequency bins, followed by modulation frequency bins corresponding to every auditory bin, placed in succession over y-axis. Fig.~\ref{CQT-MSF-plot} shows the $2$-D representation obtained with concatenation of CQT/MFSC with the corresponding MSF.
In second approach, to better combine the information from time-frequency and modulation features, we use an embedding fusion based DNN architecture. The architecture consists of two parallel but similar branches of convolutional and GAP layer, followed by a common FC and softmax layer. For both feature fusion and embedding fusion, the embeddings are extracted from the GAP layer of DNN model.
\subsection{Evaluation Metrics}
For performance evaluation, we use accuracy and UAR metrics. We chose these metrics owing to their popularity in SER and also for better comparison of results with the literature. Accuracy is defined as the ratio between the number of correctly classified utterances to the total number of utterances in test set. According to \cite{rosenberg2012classifying}, the UAR metric is given as:
\begin{equation}
\mathrm{UAR} = \frac{1}{K} \sum_{i=1}^{K} \frac{A_{ii}}{\sum_{j=1}^{K} A_{ij}}
\end{equation}
\noindent here, $A$ is called the contingency matrix, $A_{ij}$ refers to the number of samples in class $i$ classified as class $j$, and $K$ is the total number of classes. As accuracy is considered \emph{unintuitive} for databases with uneven samples across different classes, we use UAR to measure the validation set performance of the DNN model and to select the best performing model over the set of epochs.
\section{Results \& Discussion}
\label{res_and_dis}
\subsection{Performance Comparison of Different Features}
\renewcommand{\thesubfigure}{\alph{subfigure}}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.55\textwidth}
\hbox{\hspace{-0.4cm}\includegraphics[scale=0.23]{Fig8_1_Mod_filt_scale1.png}}\caption{Filter scale $q~=~1$}
\end{subfigure}%
\begin{subfigure}[t]{0.5\textwidth}
\hbox{\hspace{-0.4cm}\includegraphics[scale=0.23]{Fig2_2_Mod_filt_scale2.png}}\caption{Filter scale $q~=~2$}
\label{Mod_filts_2}
\end{subfigure}%
\caption{Modulation filter banks for different values of filter scale factor $q$. Filters with $q~=~1$ have same center frequencies but higher bandwidth than filters with $q~=~2$.}
\label{Mod_filts}
\end{figure}
\begin{table}[t]
\centering
\caption{Comparison between different filter scale ($q$) values of modulation filters for experiments performed with feature-fused CQT-MSF over EmoDB database. Given values are in percentages.}
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{ccccccc}
\hline
\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}} Classification \\ Framework\end{tabular}}} & \multicolumn{2}{c}{\textbf{$\boldsymbol{q~=~1}$}} & \multicolumn{2}{c}{\textbf{$\boldsymbol{q~=~2}$}} & \multicolumn{2}{c}{\textbf{$\boldsymbol{q~=~3}$}} \\
& \multicolumn{1}{c}{\textbf{Accuracy}} & \textbf{UAR} & \multicolumn{1}{c}{\textbf{Accuracy}} & \textbf{UAR} & \multicolumn{1}{c}{\textbf{Accuracy}} & \textbf{UAR} \\ \hline
\hline
\begin{tabular}[c]{@{}c@{}} DNN\end{tabular} & \multicolumn{1}{c}{76.97} & 68.25 & \multicolumn{1}{c}{70.79} & 64.93 & \multicolumn{1}{c}{69.77} & 64.27 \\
\begin{tabular}[c]{@{}c@{}} DNN-SVM\end{tabular} & \multicolumn{1}{c}{79.86} & 76.17 & \multicolumn{1}{c}{79.50} & 77.00 & \multicolumn{1}{c}{74.28} & 70.91 \\ \hline
\end{tabular}
\label{Qvalue_analysis}
\end{table}
\begin{table}[t!]
\centering
\renewcommand{\arraystretch}{1.4}
\caption{Performance comparison of combined early auditory and cortical features over EmoDB database. Filter scale ($q$) value of $1$ is used in modulation filterbank for MSF computation. Given values are in percentages.}
\begin{tabular}{ccccc}
\hline
\multirow{2}{*}{\textbf{Features}} & \multicolumn{2}{c}{\textbf{DNN}} & \multicolumn{2}{c}{\textbf{DNN-SVM}} \\
& \multicolumn{1}{c}{\textbf{Accuracy}} & \multicolumn{1}{c}{\textbf{UAR}} & \multicolumn{1}{c}{\textbf{Accuracy}} & \textbf{UAR} \\ \hline
\hline
\begin{tabular}[c]{@{}c@{}}CQT-MSF \\ (Feature Fusion)\end{tabular} & \multicolumn{1}{c}{76.97} & 68.25 & \multicolumn{1}{c}{79.86} & 76.17 \\
\begin{tabular}[c]{@{}c@{}}CQT-MSF\\ (Embedding Fusion)\end{tabular} & \multicolumn{1}{c}{77.99} & 71.74 & \multicolumn{1}{c}{79.32} & 75.48 \\
\begin{tabular}[c]{@{}c@{}}MFSC-MSF\\ (Feature Fusion)\end{tabular} & \multicolumn{1}{c}{61.45} & 55.23 & \multicolumn{1}{c}{69.02} & 64.45 \\
\begin{tabular}[c]{@{}c@{}}MFSC-MSF\\ (Embedding Fusion)\end{tabular} & \multicolumn{1}{c}{60.80} & 57.92 & \multicolumn{1}{c}{67.74} & 64.89 \\ \hline
\end{tabular}
\label{msf_analysis}
\end{table}
\begin{table}[t!]
\centering
\renewcommand{\arraystretch}{1.4}
\caption{Performance comparison of early auditory and cortical features taken separately over EmoDB database. Filter scale ($q$) value of $1$ is used in modulation filterbank for MSF computation. Given values are in percentages.}
\begin{tabular}{ccccc}
\hline
\multirow{2}{*}{\textbf{Features}} & \multicolumn{2}{c}{\textbf{DNN}} & \multicolumn{2}{c}{\textbf{DNN-SVM}} \\
& \multicolumn{1}{c}{\textbf{Accuracy}} & \multicolumn{1}{c}{\textbf{UAR}} & \multicolumn{1}{c}{\textbf{Accuracy}} & \textbf{UAR} \\ \hline
\hline
\begin{tabular}[c]{@{}c@{}}MSF \\ (Computed over CQT)\end{tabular} & \multicolumn{1}{c}{72.67} & 66.32 & \multicolumn{1}{c}{78.73} & 76.33 \\
\begin{tabular}[c]{@{}c@{}}MSF \\ (Computed over MFSC)\end{tabular} & \multicolumn{1}{c}{59.35} & 53.44 & \multicolumn{1}{c}{65.24} & 60.54 \\
CQT & \multicolumn{1}{c}{71.77} & 64.91 & \multicolumn{1}{c}{76.74} & 73.21 \\
MFSC & \multicolumn{1}{c}{61.62} & 57.32 & \multicolumn{1}{c}{66.18} & 62.58 \\ \hline
\end{tabular}
\label{feat_analysis}
\end{table}
\begin{table}[h!]
\caption{Feature performance comparison over different databases with DNN-SVM framework. Given values are in percentages}
\centering
\label{data_comp}
\renewcommand{\arraystretch}{1.4}
\begin{adjustbox}{width=\columnwidth, center}
\begin{tabular}{@{}ccccccccc@{}}
\hline
\multirow{3}{*}{\textbf{Database}} & \multicolumn{2}{c}{\textbf{MFSC}} & \multicolumn{2}{c}{\textbf{CQT}} & \multicolumn{2}{c}{\begin{tabular}[c]{@{}c@{}} \textbf{CQT-MSF} \\ \textbf{(Feature Fusion)}\end{tabular}} & \multicolumn{2}{c}{\begin{tabular}[c]{@{}c@{}} \textbf{Chance level} \\ \textbf{performance}\end{tabular}}\\
& \textbf{Accuracy} & \textbf{UAR} & \textbf{Accuracy} & \textbf{UAR} & \textbf{Accuracy} & \textbf{UAR} & \multicolumn{2}{c}{\textbf{Accuracy}} \\ \hline
\hline
EmoDB & 66.01 & 61.75 & 76.74 & 73.21 & 79.86 & 76.17 & \multicolumn{2}{c}{14.28}\\
RAVDESS & 36.94 & 36.16 & 48.68 & 44.64 & 52.24 & 48.83 & \multicolumn{2}{c}{12.5}\\ \hline
\end{tabular}
\end{adjustbox}
\end{table}
First, we experimentally optimize the constant-Q filters used for CQT-MSF computation by varying the value of filter scaling factor ($q$).
Fig.~\ref{Mod_filts} shows the frequency response of modulation filter banks with $q$~=~$1$ and $2$. Since $q$ affects the time resolution of filters as given in Eq.~\ref{eq:scale}, filters with $q$~=~$1$ are wider (have higher bandwidth) which leads to clipping of the filter frequency response at low frequencies. This leads to inclusion of zero-frequency (or DC) components as well. Also, because of greater overlap between filters, the generated filter outputs have higher redundancy. Increased redundancy helps convolutional layers to better extract required emotion correlation among modulation frequency bins. With $q$~=~$2$, the filter responses remain limited inside the frequency range providing a less redundant filterbank structure as shown in Fig.~\ref{Mod_filts_2}. Table~\ref{Qvalue_analysis} reports the difference in results obtained for modulation features computed with different values of scaling factor $q$. Filterbank structure with $q$~=~$1$ outperforms the arrangement with $q$~=~$2$ and $3$. Hence, we select modulation filters with $q$~=~$1$ for further experiments. We perform optimization and detailed experimentation of $q$ over only EmoDB database due to its small size and similar trends with other databases.\\
\begin{figure}
\centering
\begin{subfigure}[h]{0.5\textwidth}
\hbox{\hspace{0cm}\includegraphics[scale=0.2]{Fig9_1_cqtmsf_conf_mats.png}
\end{subfigure}%
\begin{subfigure}[h]{0.5\textwidth}
\hbox{\hspace{0cm}\includegraphics[scale=0.2]{Fig9_2_cqt_conf_mats.png}
\end{subfigure}
\begin{subfigure}[h]{0.5\textwidth}
\hbox{\hspace{0cm}\includegraphics[scale=0.2]{Fig9_3_mfscmsf_conf_mats.png}}
\end{subfigure}%
\begin{subfigure}[h]{0.5\textwidth}
\hbox{\hspace{0cm}\includegraphics[scale=0.2]{Fig9_4_mfsc_conf_mats.png}}
\end{subfigure}
\caption{Confusion matrices for CQT-MSF (feature fusion), MFSC-MSF (feature fusion), CQT, and MFSC features. CQT-MSF shows best comparative performance in classifying different emotion classes of EmoDB database. Matrices are computed over DNN-SVM framework owing to its greater performance.}
\label{conf_mats}
\end{figure}
Table~\ref{msf_analysis} shows the performance of time-frequency representations combined with their corresponding modulation features for different classification frameworks over EmoDB database. The DNN-SVM classification framework outperforms DNN framework for every feature and over both performance metrics. This observation is counter-intuitive as the activations extracted from the same convolutional layer in DNN-SVM and DNN framework, end up performing better with SVM at back-end but not with fully-connected layer at the back-end. Regarding the two different feature fusion types, there is no specific pattern, in terms of performance improvement, among the classification frameworks. With CQT-MSF, feature fusion outperforms embedding fusion over DNN framework, whereas, embedding fusion outperforms feature fusion for DNN-SVM framework. For MFSC-MSF, the trend is opposite to that of CQT-MSF fusion results. However, CQT-MSF in both fusion styles performs better than MFSC-MSF. \\
To further analyse the contribution of early auditory and cortical features taken separately over SER, we perform experiments to analyse the performance of standalone MSF extracted over CQT/MFSC (without any type of fusion). From the results in Table~\ref{feat_analysis}, cortical features (standalone MSF over CQT) outperforms standalone CQT. However, the same is not true for mel-scale based features. The MSF computed over MFSC shows poor performance as compared to standalone MFSC. This questions the usability of temporal trajectories of the mel-scale time-frequency representation for emotion classification. The inferiority of MFSC against CQT is also indicated by direct comparison between CQT and MFSC. CQT outperforms MFSC in both DNN and DNN-SVM classification frameworks. Among different CQT feature types, MSF computed over CQT assumes very similar performance in contrast to both fusion types of CQT-MSF. This phenomenon describes the higher emotion relevance of the temporal modulations of the low-frequency regions emphasised by CQT. \\
Fig.~\ref{conf_mats} shows the confusion matrices for different features over EmoDB database. We choose only the feature-fused CQT-MSF and MFSC-MSF for comparison, owing to their improved performance as compared to embedding fusion in DNN-SVM framework. Even though CQT-MSF is comparatively better at classifying different emotion classes, some instances of \emph{Happy} and \emph{Disgust} are confused with \emph{Anger}, and that of \emph{Fear} are confused with \emph{Happy}. The highest misclassification is in \emph{Happy-Anger} emotion pair which have similar arousal but opposite valence characteristic.
This observation is found to be consistent among various SER works~\cite{zhang2017speech, peng2020, deb2019} and can also be related to the similar F-ratio characteristics of \emph{Anger} and \emph{Happy} in Fig.~\ref{f_ratio_modspec}. Among low-arousal classes, some confusion in \emph{Sad-Boredom} is also visible in CQT-MSF. As prosody features are less effective in valence discrimination~\cite{iker2010}, the confusion among pairs with similar arousal but opposite valence characteristics can be attributed to higher emphasis over pitch in constant-Q scale based spectral representation. However, modulations computed across constant-Q scale reduce this confusion, which is evident from the comparison between standalone CQT and CQT-MSF. In standalone CQT, confusion among both low and high arousal emotions is higher and appears among multiple emotion classes (e.g., \emph{Disgust} and \emph{Fear} with \emph{Happy}). In MFSC and MFSC-MSF, as compared to CQT-based features, the misclassification is more prominent across multiple classes. Unable to emphasise speech prosody, the emotion classification ability of MFSC over arousal scale is inferior to CQT-based features (e.g., increased confusion of \emph{Boredom} with \emph{Neutral}, and \emph{Fear} with \emph{Sad}). Also, modulations of MFSC are computed with more focus on high and mid speech frequencies and less focus on prosody at low frequencies, further deteriorating the performance. \\
At the end, we compare the SER performance of CQT-MSF feature with the scattering transform coefficients. As scattering network is also a deep convolutional network, its comparison with our proposed CQT-MSF with DNN-SVM classification framework shows the superiority of automatic feature extraction and required time and frequency-shift invariance learning for SER. The scattering coefficients are computed using the similar train/test strategy. The training speech utterances are chunked to $400$~ms segments with $50$\% overlapping, whereas testing utterances are used as is. Following our experiments in~\cite{singh2021deep}, the Q-factor value (Q) for first layer coefficients is chosen as Q~=~$5$. As the training segment size is fixed to $400$~ms ($6400$ samples at $16$~kHz), the maximal wavelet length or averaging scale ($T$) is kept $4096$ samples for both training and validation. However, since we use complete utterances in test, the duration $N$ for testing is empirically fixed to $51000$ samples ($3.18$ seconds at $16$~kHz). Longer utterances are chopped to contain only $51000$ samples, whereas shorter frames are zero-padded. We obtain \boldsymbol{$72.67\%$} accuracy and \boldsymbol{$69.8\%$} UAR with scattering coefficients outperforming MFSC, and indicating the requirement of time and deformation stability in SER. However, the performance is inferior to that with the proposed CQT-MSF (especially with feature-fusion). The superiority of CQT-MSF shows that deep networks learn to provide better time and deformation stability, as described in Section~\ref{comp_scat_msf}, while extracting the emotion relevant information from multiple convolution layers. Although scattering transform also involves convolutions and averaging for stability, it is performed using fixed kernels which are not automatically learned/optimized to improve performance.\\
Table~\ref{data_comp} shows the results obtained with different features over EmoDB and RAVDESS databases. The proposed CQT-MSF feature outperforms other features over RAVDESS database as well. This explains the suitability of CQT-MSF features or two stage auditory analysis for SER over different databases. Compared to EmoDB, the relative performance improvement with CQT-MSF, CQT over MFSC is higher in RAVDESS database. \\
\subsection{Comparison with Related Works}
In this subsection, we compare our obtained results with related works in SER. Among SER literature, different strategies are used to evaluate system performances, for example, use of different databases, number of emotion classes used in the databases, different evaluation methodologies, etc. These differences make direct comparison of SER works difficult. The evaluation methodology, in terms of, cross-validation scheme, train/test split, speaker dependent/independent testing, etc. are found to differ substantially in SER literature. Lack of reproducible research in SER domain also leads to uncertainty, leading to difficulty in comparison. Hence, a comparison made with other relevant SER literature can not be considered accurate. \\
Due to the above mentioned issues, to justify our obtained results, we implement different studies from the literature by using our proposed CQT-MSF feature and experimental framework. Table~\ref{comp_res_table} shows the list of selected works and the corresponding performances obtained. Section~\ref{eval_metho} and~\ref{classifier_desc_sec} of the manuscript describe the experimental framework employed in the studies listed in the table. Our choice of selected works is based on the use of modulation spectrogram related features, and use of advanced neural network architectures (e.g., contextual long short-term memory (LSTM), multi-head attention, ResNet architecture, multi-time-scale kernel, etc.). Below we briefly describe the details of selected works.
\begin{table}[t]
\caption{Performance comparison with selected works. \textbf{Boldface} values show the best obtained results.}
\centering
\hspace*{-1.75cm}
\begin{footnotesize}
\begin{tabular}{ccccccc}
\hline
\multirow{3}{*}{\textbf{References}} & \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}\textbf{Brief Description} \\ \textbf{(Feature \& Classifier)} \end{tabular}} & \multicolumn{4}{c}{\begin{tabular}[c]{@{}c@{}}\textbf{Performance (in \%)}\end{tabular}} \\
& & \multicolumn{2}{c}{\begin{tabular}[c]{@{}c@{}}\textbf{EmoDB} \end{tabular}} & \multicolumn{2}{c}{\begin{tabular}[c]{@{}c@{}}\textbf{RAVDESS} \end{tabular}} \\
& & \textbf{Acc.} & \textbf{UAR} & \textbf{Acc.} & \textbf{UAR} \\ \hline \hline \\
\citeauthor{avila}~(\citeyear{avila}) & \begin{tabular}[c]{@{}c@{}}Different modulation \\ spectral measures. \\ DNN classifier.\end{tabular} & 55.40 & 46.83 & 37.77 & 29.10 \\ \\
\citeauthor{lightsernet}~(\citeyear{lightsernet}) & \begin{tabular}[c]{@{}c@{}} MFCC. \\ Parallel path fully \\ convolutional network. \end{tabular} & 73.37 & 67.21 & 48.68 & 44.51 \\ \\
\citeauthor{LIU20221}~(\citeyear{LIU20221}) & \begin{tabular}[c]{@{}c@{}} Interspeech 2009 feature set. \\ bc-LSTM~+~Multi-head attention. \end{tabular} & 53.45 & 47.49 & 24.58 & 21.79 \\ \\
\citeauthor{emonet}~(\citeyear{emonet}) & \begin{tabular}[c]{@{}c@{}} Mel-spectrogram. \\Convolutional ResNet. \end{tabular} & 76.34 & 71.07 & 52.56 & 49.46 \\ \\
\citeauthor{guizzo2020multi}~(\citeyear{guizzo2020multi}) & \begin{tabular}[c]{@{}c@{}} Spectrogram. \\ Multi-time-scale kernel \\ based CNN. \end{tabular} & 63.01 & 58.03 & 40.27 & 36.46 \\ \\
~\citeauthor{PARRAGALLEGO2022103286}~(\citeyear{PARRAGALLEGO2022103286}) & \begin{tabular}[c]{@{}c@{}}x-vectors, i-vectors, INTERSPEECH \\ 2010 feature set, articulation, \\ phonation, and prosody features. \\ SVM classifier. \end{tabular} & 50.21 & 42.74 & 46.59 & 43.68 \\ \\
This work & \begin{tabular}[c]{@{}c@{}} GMT-MSF feature.\\ DNN-SVM framework. \end{tabular} & 77.05 & 74.35 & \textbf{54.05} & \textbf{51.16} \\ \\
This work (Proposed method) & \begin{tabular}[c]{@{}c@{}} CQT-MSF feature. \\ DNN-SVM framework. \end{tabular} & \textbf{79.86} & \textbf{76.17} & 52.24 & 48.83 \\ \\ \hline
\end{tabular}
\label{comp_res_table}
\end{footnotesize}
\end{table}
\begin{itemize}
\item Study performed by \citeauthor{avila}~(\citeyear{avila}) proposes feature pooling techniques over different measures of modulation spectral features for dimensional emotion recognition. For performance comparison, we compute the same modulation spectral measures over CQT representation, unlike \citeauthor{avila} which uses GMT representation, and show the results on our experimental framework. Feature-pooling schemes reported in the study are skipped to maintain similarity in comparison as our framework does not include handcrafted pooling operations.
\item \citeauthor{lightsernet}~(\citeyear{lightsernet}) use mel-frequency cepstral coefficients (MFCCs) over a fully convolutional neural network architecture with parallel paths of different kernels sizes followed by stacked convolutional layers for SER with impressive reported SER performance. We use the GitHub implementation\footnote{\url{https://github.com/AryaAftab/LIGHT-SERNET}} of the state-of-the-art architecture with our experimental framework and CQT-MSF feature for performance comparison.
\item We also select the state-of-the-art study performed by~\citeauthor{LIU20221}~(\citeyear{LIU20221}) for multi-modal emotion recognition in our work. As our primary focus is emotion recognition from speech, we use only the bidirectional-contextualised LSTM (bc-LSTM) with multi-head attention block used for speech modality in~\cite{LIU20221} with our experimental framework and databases. As emotion information spreads temporally across utterances, temporal pattern extraction architectures like LSTM, attention, are selected to compare with handcrafted temporal modulation feature, i.e., CQT-MSF.
\item Study reported by~\citeauthor{emonet}~(\citeyear{emonet}) employs an adapter ResNet architecture for multi-corpora SER scenario. As our study does not include mixing different corporas, we select the reported ResNet architecture without the adapter module but with CQT-MSF feature for performance comparison. We use the open-source GitHub implementation\footnote{\url{https://github.com/EIHW/EmoNet}} of the model.
\item Study performed by~\citeauthor{guizzo2020multi}~(\citeyear{guizzo2020multi}) uses a multi-time-scale convolutional kernel based front-end which employs multiple temporally resampled versions of the original convolution kernel and parallelly perform convolution with every version. We utilize this method by employing multi-time-scale convolution layer as front-end over our DNN framework (mentioned in Section~\ref{classifier_desc_sec}). We use the available open-source GitHub implementation\footnote{\url{https://github.com/ericguizzo/multi_time_scale}} of multi-time-scale convolution layer. Similar to CQT-MSF, the multi-time-scale kernels also focus on the extraction of speech temporal patterns for emotion recognition.
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{Fig10_filt_comp.png}
\caption{Comparison between different non-linear time-frequency representation scales with filterbank center frequencies (shown by dots). The zoomed-in portion shows the difference between the non-linearities for emotion-salient low frequency filter bins. For better visibility of differences among the scales, we plot every scale with $96$ frequency bins.}
\label{filt_comp}
\end{figure}
\item We select the feature set based study by \citeauthor{PARRAGALLEGO2022103286}~(\citeyear{PARRAGALLEGO2022103286}) for notably high performance reported on the RAVDESS database. The work includes combination of x-vector, i-vector, Interspeech $2010$ paralinguistic (IS10) feature set, and articulation, phonation and prosody features extracted from \emph{Disvoice}\footnote{\url{https://github.com/jcvasquezc/DisVoice}} framework for emotion classification with SVM as classifier. As the study report that the combination of only x-vector, IS10, and Disvoice framework provide the best performance, we use the same shortened feature set with SVM classifier, but on LOSO cross-validation framework. To maintain similarity, we use complete utterances (no segmentation) for both training and testing in this particular implementation. Note that study does not integrate our CQT-MSF and is rather based on the original implementation in the paper.
\item As several modulation feature based SER works use gammatone-scale to generate a time-frequency representation over which modulations are computed~\cite{alam2013amplitude, avila, wu2011automatic}, we compare the performance of CQT-MSF with gammatone-scale based modulation spectral features (GMT-MSF) on the DNN-SVM framework mentioned in Section~\ref{classifier_desc_sec}. The obtained performance is reported in Table~\ref{comp_res_table} along with other employed comparison techniques. Our implementation of gammatone-spectrogram includes gammatone filter design using python~\emph{Spafe}~\cite{ayoub_malek_2022_6824667} toolkit, followed by application of the filters over the signal spectrogram. The GMT-MSF is computed over the designed gammatone time-frequency representation following the same steps used to compute CQT-MSF (Fig.~\ref{block_diagram}).
\end{itemize}
\begin{figure}[t]
\centering
\begin{minipage}{0.5\textwidth}
\hbox{\hspace{-2cm}\includegraphics[scale=0.48]{Fig11_1_emodb_ESD.png}}
\end{minipage}%
\begin{minipage}{0.5\textwidth}
\hbox{\hspace{0cm}\includegraphics[scale=0.48]{Fig11_2_ravdess_ESD.png}}
\end{minipage}
\caption{Average energy spectral density (ESD) computed over all utterances of EmoDB and RAVDESS. The spectrogram (Spec.), MFSC, CQT, and GMT ESD corresponds to the energy of corresponding coefficients average across all utterances. For better visibility of differences among the features, we plot every ESD with $96$ frequency bins.}
\label{energy_curves}
\end{figure}
From Table~\ref{comp_res_table} we observe that CQT-MSF outperforms GMT-MSF on EmoDB database but performs relatively poor for RAVDESS database. This observation is due to the difference in non-linearity between constant-Q and gammatone scale. Previous studies on speech recognition~\cite{paliwalsfcc}, speaker recognition~\cite{SARANGI2020102795}, and SER~\cite{SINGH2022103712} also show how different non-linearity during frequency wrapping affect recognition performances. Inspired by those studies, we compare the three concerned non-linear scales used in our experiments: mel, constant-Q, and gammatone in Fig.~\ref{filt_comp} along with the filterbank center frequencies. For better visibility of differences in filter placement, we use $96$ frequency bins for every scale in this analysis. The figure shows that constant-Q scale provides highest low-frequency emphasis (because of binary logarithm) followed by gammatone and mel-scale. Considering the relevance of the low-frequency information for SER~\cite{singh2021, SINGH2022103712}, the underperformance of constant-Q scale is an interesting observation. \\
Previous studies indicate that the dataset-dependent non-linearity scale is more appropriate than a general-purpose scale and providing importance to higher energy region helps in performance optimization~\cite{paliwalsfcc, dipjyoti2017}. Hence, to further investigate the performance gap, we compare the energy spectral density (ESD) averaged across all utterances for both the databases. Fig.~\ref{energy_curves} shows the respective averaged energy density plots. The energy densities are computed by time averaging the squared feature coefficients, followed by averaging across all utterances. Fig.~\ref{energy_curves} shows that when compared to EmoDB, the energy density in RAVDESS is more shifted towards higher frequency regions. The constant-Q scale provides greater emphasis at low-frequency bins (refer Fig.~\ref{filt_comp}) but due to high non-linearity (binary logarithm), the resolution reduces drastically as we move towards higher frequencies. Hence, for RAVDESS, the resolution on the frequencies with the larger ESD value is lower in CQT compared to the resolution provided by gammatone filterbank placed in the ERB scale. Thus, for RAVDESS, GMT-MSF better captures the signal energy information and leads to better performance when compared to CQT-MSF. \\
From Table~\ref{comp_res_table} we also observe that the employed EmoNet-based ResNet architecture also shows competitive performance on RAVDESS database when compared to CQT-MSF feature. Larger model size with comparatively larger database leads to this observation.
\subsection{Visual Inspection of Learned Features}
In spite of the performance gain achieved from the proposed CQT-MSF feature, the complexity of the model (Table~\ref{cnn_param_tab}) in terms of network parameters makes it very difficult to understand which information in input helps the network to recognize the pattern. To obtain a general insight into the operation of deep networks, several works use gradient generated at the ultimate layer with respect to the network input to generate a saliency map. This map shows corresponding input regions which weigh the most in generating the output probability scores~\cite{simonyan2014deep, springenberg2015striving, selvaraju2019gradcam}. We use one such analysis, called the \emph{gradient-based class activation mapping} (Grad-CAM), to obtain insight into the working of the model employed~\cite{selvaraju2019gradcam}. \\
\renewcommand{\thesubfigure}{\alph{subfigure}}
\begin{figure}[hp!]
\begin{subfigure}[t]{\textwidth}
\hbox{\hspace{1cm}\includegraphics[height=6cm, width=11cm]{Fig12_1_03a05Wb.wav.png}}
\caption{Constant-Q MSF with Grad-CAM, pitch and first three formants of \emph{Anger} utterance}
\end{subfigure}
\begin{subfigure}[t]{\textwidth}
\hbox{\hspace{1cm}\includegraphics[height=6cm, width=11cm]{Fig12_2_03a05Nd.wav.png}}
\caption{Constant-Q MSF with Grad-CAM, pitch and first three formants of \emph{Neutral} utterance}
\end{subfigure}
\begin{subfigure}[t]{\textwidth}
\hbox{\hspace{1cm}\includegraphics[height=6cm, width=11cm]{Fig12_3_03a05Tc.wav.png}}
\caption{Constant-Q MSF with Grad-CAM, pitch and first three formants of \emph{Sad} utterance}
\label{pitch_gradcam}
\end{subfigure}
\caption{Grad-CAM output of three different emotion utterances of EmoDB database. To analyse the significance of various frequency regions, pitch and first three formants are also shown along with the Grad-CAM output. For MSF plot, the y-axis labels on the left describe the auditory frequencies and ticks on the right describe the modulation bins corresponding to every auditory frequency bin. The title of every plot describes the EmoDB file name used for analysis.}
\label{gradcam}
\end{figure}
Grad-CAM uses the class-wise gradient generated at network output with respect to the activations of the final convolution layer to generate a \emph{heatmap} showing the importance of different regions of the input. The steps included in Grad-CAM heatmap generation are:
\begin{enumerate}
\item Compute the gradient of network output score (before softmax non-linearity) with respect to the activations of final convolution layer, i.e., $\frac{\partial y_c}{\partial A_k}$, where $y_c$ is the score of the $c$th class and $A_k$ is the 3-dimensional (height, width and channel dimension) class activations from the final convolution layer.
\item Average the computed gradients over length and width dimensions (global average pooling) to obtain a single vector representation of gradients. Mathematically, $\alpha_k~=~\frac{1}{N} \sum \limits_{\mathrm{length}} \sum \limits_{\mathrm{width}} \frac{\partial y_c}{\partial A_k}$.
\item Multiply the computed gradient vector with the final convolution layer activations and average the result over the number of filters in convolution layer, i.e., $map~=~\frac{1}{N} \sum \limits_k(\alpha_k A_k) $.
\item Apply \emph{ReLU} activation over computed class activation maps in the previous step, $L^{c}_{\mathrm{Grad}-\mathrm{CAM}}~=~ReLU(map)$.
\item Upsample the computed $2$-D heatmap over length and width axes to make its shape similar to the input image. The upsampled heatmap shows the importance of various regions of input image which led to the final class prediction.
\end{enumerate}
Fig.~\ref{gradcam} shows the Grad-CAM output of the CQT-MSF feature of \emph{Anger}, \emph{Neutral}, and \emph{Sad} utterances of the same speaker (speaker $03$) and context (a$05$) from EmoDB database. For analysis of Grad-CAM output, we also plot the pitch and first three formant frequencies (F$1$, F$2$ \& F$3$) of utterances to compare the frequency regions which are most focused upon by the network to predict emotion classes. We observe that pitch and the first two formants (F$1$ \& F$2$) are important for the emotion class prediction. Lower pitch harmonics are apparently more significant for \emph{Sad} emotion as shown in Fig.~\ref{pitch_gradcam}. Another important observation for \emph{Sad} is the presence of high Grad-CAM score at the silence and unvoiced regions (where pitch frequency is $0$~Hz) of utterance. This shows that the silence between spoken phonemes is also important for emotion recognition. In \emph{Anger} and \emph{Neutral} emotions, formants F$1$ \& F$2$ are more prominent than pitch. Both emotions are identified mostly in the voiced region of utterances. Interestingly, the Grad-CAM response of \emph{Sad} also shows some focus on high frequencies (near formant F3) over complete utterance as compared to \emph{Neutral} emotion class. Observations made from the Grad-CAM analysis indicate the importance of low frequencies for emotion recognition, especially for low-arousal emotions like \emph{Sad}, further justifying the use of CQT time-frequency representation for SER. Also, MSF representation provides the Conv2D classifier with different modulation rates of different speech characteristics, such as pitch harmonics, formants etc. This helps the classifier to emphasize the emotion-wise differences appearing across different modulation rates, for different speech characteristics (pitch, formant, etc.), hence improving the performance.
\subsection{Discussion}
Our performed experiments show higher relevance of CQT-based features, as compared to mel-scale features, for SER. We summarise and interpret the results obtained from the experiments in the following points:
\begin{itemize}
\item \textbf{The two-staged auditory processing based features, i.e., early auditory with cortical analysis based features improve emotion classification performance}. This also justifies the combination of human auditory analysis (domain knowledge) with neural networks for the betterment of SER. However, such improvement is observed over temporal modulations extracted from CQT representation only. Temporal modulations of MFSC show opposite effect and degrade the performance.
\item \textbf{The improvement observed with CQT-MSF (both fusion types) and standalone CQT over MFSC features show the relevance of increased low-frequency resolution in CQT for SER}. As low-frequency resolution in mel-scale is not as high, it does not provide enough emphasis over the emotion relevant low-frequencies to capture the information required for emotion discrimination. Hence, its modulation spectrum coefficients also end up with less emotion relevant parts of speech, e.g., irrelevant high frequency regions, leading to reduction in performance.
\item \textbf{DNN framework lags behind in performance as compared to DNN-SVM classification framework with RBF kernel function}. This observation is consistent across every employed feature. The advantage in performance of DNN-SVM framework has also been mentioned in SER \cite{zhang2017speech} and speaker recognition \cite{snyder2018x} works.
\item \textbf{The confusion matrices of different features show a general misclassification trend in \emph{Happy-Anger} and \emph{Fear-Happy} emotion pairs}. This confusion mainly appears due to very similar arousal characteristics of features. However, \emph{Happy-Anger} pair are placed very distant in valence plane. This is due to higher focus on speech prosody in constant-Q representation, and higher sensitivity of speech prosody over arousal characteristics~\cite{iker2010}. Although, inclusion of modulations of constant-Q representation reduces the confusion among emotions with opposite valence characteristics.
\item \textbf{Both CQT-MSF and standalone CQT also outperforms scattering transform coefficients}. Scattering transforms apply averaging over features with empirically defined averaging scale to obtain a translation invariant representation. The convolutional neural network, when used with constant-Q features as input, automatically learns this requires invariance with cross-entropy objective function. Hence, the joint effect of CQT/CQT-MSF and automatic translation invariance makes our frameworks superior. Although, scattering transform manages to outperform mel-scale features, again because of the constant-Q filter banks and time shift and deformation stability in scattering coefficients.
\item \textbf{CQT-MSF feature outperforms GMT-MSF on EmoDB but underperforms on RAVDESS database.} Comparative analysis shows a slight high-frequency shift in average energy spectral density of the RAVDESS database compared to EmoDB database. The difference in non-linearities of the gammatone and CQT scales result in gammatone-spectrogram better capturing energies with a slight high-frequency shift. This explains the observed anomaly in the performance with RAVDESS database.
\end{itemize}
Studies in psychology report that individuals with music expertise are better capable of perceiving emotions from speech \cite{lima2011speaking, good2017benefits, fuller2014musician, dmitrieva2006ontogenetic, twaite2016examining, weijkamp2017attention, thompson2004decoding, nussbaum2021}. This finding falls in line with our experiments. As CQT was originally invented for music analysis, its better suitability for SER can be considered as a mathematical evidence, supporting the findings in psychology domain. Another justification towards increased SER suitability of CQT, can be proposed by analysing studies performed over \emph{amusia} in \cite{sydney2021, nussbaum2021}. Amusia is a medical condition in which individuals have limited capability to perceive or resolve pitch. The study in \cite{sydney2021} reports that emotion recognition ability of amusic individuals is below par with that of normal individuals, which is attributed to their limited ability to resolve pitch or low-frequencies of speech. Amusics then utilize the high-frequency content of speech to decipher emotions but are not as efficient as healthy individuals. Therefore, an amusic brain can be assumed to represent speech as a time-frequency representation with low resolution at low-frequencies and comparatively higher resolution at high frequencies. In a contrastive manner, a representation with high low-frequency resolution should improve SER ability, which is what we observe in our experiments with CQT-based features. Also, modulation coefficients computed over CQT is analogous to cortex-level analysis performed over music trained brain. Study performed in~\cite{nussbaum2021} reports that such cortical analysis has further beneficial effects over SER ability.
\section{Conclusion}
\label{conc_sec}
This paper proposed the use of constant-Q transform based modulation spectral features for SER. The proposed feature employs the knowledge of two-staged sound processing in humans as domain knowledge and is tested over two different deep network based classification frameworks. We show that the proposed feature outperforms standard mel-frequency based feature and scattering transform coefficients. From the performed experiments, we conclude the following:
\begin{enumerate}
\item A representation with increased low-frequency resolution is a better contender for SER. Similar conclusion is endorsed in psychology based studies as well.
\item The combination of a time-frequency representation with higher low-frequency resolution, and its temporal modulation (two-staged representation), efficiently represent the emotion contents in speech.
\item Mel-scale based feature and its temporal modulations are not very significant from speech emotion information perspective.
\item Similar, to mel-scale features, CQT and its temporal modulation representation are also time deformation invariant. With CQT-MSF as input, the CNN can learn the required invariance to time-frequency shifts, leading to a representation stable to shifts and deformations.
\item The DNN-SVM framework provides better SER performance as compared to the standard DNN framework.
\item Grad-CAM based analysis performed over MSF reestablishes the importance of pitch and formant frequencies for SER. It also describes the importance of different modulation rates of pitch and formants, apart from their crude values, for SER.
\end{enumerate}
Although the proposed feature performs better than standard features, the performance is still not optimum for practical real-world deployment of the feature. Also, the performance varies by a large margin when the feature is used over different databases. Even though found more efficient than mel-scale, constant-Q scale is effective for utterances with larger average energy in low-frequency regions. This opens up the opportunity to explore database-dependent non-linear scale for SER. In future, we would also like to experiment with joint spectral and temporal modulation feature and analyse its suitability for SER. To combine the domain knowledge with self-learning, a deep network based architecture for self-learned modulation feature extraction can also be explored.
|
\section{Introduction} \label{intro}
CG Dra has remained an enigmatic object for more than half a century. Initially found on the photographic plates taken by K. Loechel in 1964 with the 1.34-m Schmidt telescope of the Karl Schwarzschild Observatory, \citet{1966AN....289....1H, 1966AN....289..139H} lists CG Dra under preliminary designation S 9370 as U Gem-like, ``most likely belonging to the CN Ori group.'' U Gem stars are dwarf novae (DNe)---a particular type of cataclysmic variable (CV) consisting of a white dwarf (WD) primary star accreting matter from a red dwarf secondary. Comprehensive literature reviews of DNe are available from \citet{1995cvs..book.....W} and \citet{2001cvs..book.....H}. U Gem DNe are generally divided into four subtypes: SS Cyg, SU UMa, Z Cam and WZ Sge. While the latter two are completely inapplicable classifications in the case of CG Dra due to the lack of standstills and its long orbital period, CG Dra is usually referred to as a SS Cyg star \citep{1997A&A...325..601B, 10.1093/pasj/56.sp1.S1}. This type of DN is characterized by periodic outbursts, $\Delta V \sim$ 2--6 mag, in some cases with plateau, outburst rise time of a few d and a slightly longer decline, outburst interval of 10 d to a few yr, and orbital period, $P_{orb} > 3$ h. It is now widely accepted that DN outbursts occur due to the thermal limit cycle instability in the primary's accretion disk (AD), in particular due to the pile-up of material flowing from the secondary at a rate exceeding the rate of accretion in quiescence \citep{1974PASJ...26..429O, 1984AcA....34..161S}. At some point the disk becomes hot and ionized, this boosts the system's luminosity, increases disk viscosity, which then causes the orbiting material to spread, due to the exchange of angular momentum, partially inwards, falling towards the WD primary. The increased accretion rate quickly drains the AD and the system is returned to the cool, quiescent state.
The other subtype of DN, SU UMa, is characterized by the addition of periodic superoutbursts, typically brighter by $\sim$ 2 mag and occurring about $3\times$ less frequently than normal outbursts, and a shorter $P_{orb}<3$ h. All SU UMa-type DNe exhibit superhumps---periodic modulations appearing near superoutburst maximum with a period a few percent longer than $P_{orb}$. While normal SU UMa outbursts are believed to be caused by the same mechanism as in SS Cyg stars, multiple models exist explaining superoutbursts and superhumps: the thermal-tidal instability (TTI) model by \citet{1989PASJ...41.1005O, 1996PASP..108...39O}, the enhanced mass transfer (EMT) model currently supported by \citet{1991AcA....41..269S, 2004AcA....54..221S, 2017AcA....67..273S}, and the pure thermal instability (PTI) model by \citet{2010ApJ...725.1393C, 2012ApJ...747..117C}.
Overviews of these models and description of associated problems are available from \citet{2013PASJ...65...50O} and \citet{2017AcA....67..273S}. In brief, the TTI model requires eccentric AD to explain superoutbursts. This condition occurs due to the 3:1 resonance in DNe with the mass ratio of components below a critical value, $q \equiv M_2/M_1 \lesssim q_{crit}$, where $q_{crit}\approx 0.3$, coinciding with the DN period gap between $P_{orb} \approx 2 \to 3$ h. \citet{2020AcA....70..313S} proposes $q_{crit} = 0.22$, which corresponds to the gap at $\approx 2.6$ h. SU UMa DNe showing superoutbursts, hence, should not be found above the period gap. The EMT model explains superoutbursts as the result of variable hot spot brightness during enhanced mass transfer episodes; and the PTI model simulations suggest that thermal instability alone used to explain normal and long SS Cyg outbursts is sufficient to explain superoutbursts of SU UMa stars. At the current moment there is no consensus on which model is true. The TTI model has an issue of producing superhumps of excessive amplitude in simulations, and it is unable to explain them appearing in DNe above the period gap.
CG Dra has entered BAA Variable Star Section (VSS) observing campaigns since 2001 as a poorly characterized DN. \citet{2007JBAA..117...22S} have identified its outburst period of $\approx$ 11 d and noted the bi-modality of outbursts. In the time-resolved photometry of \citet{2008JBAA..118..343S}, $P_{orb}=0.18864 \pm 0.00004$ d, noting the system's short and shallow grazing eclipses, consistent with its high inclination. The 14 eclipses recorded in both quiescent and outburst states appeared symmetrical, with both eclipse duration and depth independent of the state. The flickering appeared to be continuing throughout eclipses, suggesting that the system's inner AD is not occulted. In the earlier radial velocity measurements, \citet{1997A&A...325..601B} have found two principle power spectrum peaks corresponding to $P_{orb}=0.1893 \pm 0.0006$ and $0.2343 \pm 0.0021$ (d).
In this work we present new photometric data, collected during a BAA VSS observing campaign in 2022. As a result, the number of observed eclipses has increased significantly. Most of the data were obtained by the authors on the 0.43-m \textit{A1} telescope of Alnitak Remote Observatories in Nerpio, Spain, and additional observations come from BAA observers acknowledged below. The long-term light curve, eclipse profile comparison in various states of the system, its outburst cycle and the analysis of outburst decline rates are provided and discussed in \S \ref{photometry}. CG Dra eclipse ephemerides, orbital period and spectrograms of frequencies in the $P_{orb}$ frequency domain are revealed in \S \ref{periods}. Results and other challenges present with this system are discussed in \S \ref{discussion} and this work is concluded in \S \ref{conclusion}.
\section{Photometry} \label{photometry}
\subsection{Light Curve and Eclipse Profiles}
\begin{figure*}
\centering
\includegraphics[width=7in]{Figure_1.png}
\caption{The light curve of CG Dra obtained throughout the 2022 BAA VSS observing campaign. Filters are color-coded. The light curve spans JD 2459709.39673 to 2459904.43000 (195 d), containing \nobs observations. BAA observers who have contributed to the light curve above are D. G. Buczynski, D. Shepherd. F. Tabacco, G. Poyner, I. L. Walton, M. Mobberley, M. Usatov, N. D. James, P. Bouchier and R. Sargent.}
\label{f:lightcurve}
\end{figure*}
CG Dra light curve obtained throughout the 2022 BAA VSS observing campaign is shown in figure \ref{f:lightcurve}. The data collected covers JDs \jdbegin $\to$ \jdend spanning \dataspan d. A total of \nobs observations were made, with \neclipses eclipses recorded. The 0.43 m \textit{A1} telescope used is a Corrected Dall-Kirkham (CDK) optical design equipped with the latest generation back-illuminated Sony IMX455 CMOS 24 $\times$ 36 mm sensor with 3.76 $\mu$m square pixels. All CG Dra observations were taken in the 2 × 2 binned mode with 0.53”/px scale that is suitable for typical 1-2” FWHM seeing conditions available at the site. Photometric reduction was done using \textsc{MetroPSF} Python code\footnote{The original \textsc{MetroPSF} source code is available at https://github.com/blackhaz/MetroPSF}, modified for batch processing. \textsc{SExtractor} routines \citep{1996A&AS..117..393B} were used for flux determination. \textsc{MetroPSF} performed automatic blind astrometric calibration via local copy of the Astrometry.net service \citep{2010AJ....139.1782L}, requested comparison photometry data from the AAVSO Photometric All-Sky Survey (APASS DR9) catalog \citep{2016yCat.2336....0H}, matched stellar sources and performed differential photometry of CG Dra via linear regression fits to a weighted V-band ensemble \citep{2010JAVSO..38..202P}. The ensemble was limited to known constant stars with AAVSO Unique Identifiers (AUIDs), within $\pm 3$ mag of CG Dra. All known variable stars in the AAVSO VSX database were excluded from the ensemble. The selection of stars was fixed throughout each night whenever possible. Typically 5-6 comparison stars were used per night, and one check star not part of the ensemble for control. Photometry with uncertainty exceeding $\pm 0.1$ mag was discarded, and good observations were submitted manually to BAA VSS and AAVSO databases after visual control.
Most of observations were made without filters (CV mode) due to faint magnitudes in quiescence and short cadence requirement. Only a short period of observations of the first bright outburst was taken with the Johnson V filter. All exposures were of 30 s duration.
CG Dra is quiescent near 17 mag, bursting to $\approx 15.75$ mag. Visually two outburst types can be identified---normal and bright, the latter being $\approx 0.25$ mag brighter. We have captured \boutbursts bright outbursts, and the outburst cycle is analyzed in \S \ref{outburstcycle}.
\begin{figure}
\centering
\includegraphics[width=3.4in]{profile_depth_5.png}
\caption{Phase plot of CG Dra normalized average eclipse profiles, in different system states. From top to bottom: bright outbursts, normal outbursts, rising and fading, and quiescence. Ordinates are in the [0, 1] range. Black points represent individual observations of \neclipses eclipses. The bottom pane compares averaged eclipse profiles.}
\label{f:eclipseprofiles}
\end{figure}
Averaged and normalized profiles of eclipses are shown in figure \ref{f:eclipseprofiles}. All eclipses were divided into bright outburst, outburst, rising and falling, and quiescent states manually. The system was considered in outburst state if $V \lesssim 16.25$ and in bright outburst state if $V \lesssim 15.75$. Eclipses recorded on the decline from these states were considered fading. Quiescent state assumed once the system reached $V \approx 16.50$, and rising once $V \lesssim 16.50$ or if the linear slope of a nightly light curve showed significant negative linear trend. Ephemerides found in \S \ref{periods} were used to calculate predicted eclipse minima. Each eclipse light curve was truncated to the length of orbital period, centered on its calculated minimum, and smoothed using Gaussian filter with $\sigma = 0.2$. Here, the filter was applied in time domain on unevenly sampled data, and the standard deviation of the Gaussian distribution was found empirically to remove light curve flickering and other noise. Approximate eclipse boundaries were found around the dimmest magnitude point of the truncated smoothed light curve using a simple condition that intensity of a smoothed light curve can only increase with distance from the eclipse minimum towards each boundary. The light curve was truncated further to found boundaries, and a Gaussian function was fitted to the remaining data. The minimum of the fit was used to record the observed eclipse minimum. Eclipses were folded using $P_{orb}$ found in \S \ref{periods} and the observed minimum epoch. For each state of the system, folded light curves were combined into unevenly sampled data sets which were averaged using Gaussian filter with $\sigma = 5$, found empirically to produce a smoothly varying average eclipse profile per state. This provides a case of weighted averaging whereby data sets with fewer data points, i.e. higher inter-sampling time difference, receive a higher degree of smoothing.
CG Dra exhibits a variety of eclipse profiles in line with canonical DN outburst models. Quiescent states are dominated by asymmetric eclipses with a lower egress flux, compared to ingress, and an orbital hump of $\Delta V \sim 0.1$ mag, extending throughout orbital phases $\phi \approx -0.4 \to 0.2$, representing the bright spot where the flow of material from the secondary impacts the primary's AD. Occasionally orbital hump amplitudes exceed 0.2 mag. The eclipse ingress appears to occur at $\phi \approx -0.07$ when the secondary begins occulting the AD and the bright spot. As noted by \citet{2008JBAA..118..343S}, the short duration of eclipses suggests that CG Dra's inclination is close to critical, hence its eclipses are grazing---occulting only the bright spot and, partially, the AD. The presence of flickering, typically of $\Delta V \sim 0.05$ mag, throughout the eclipses indicates that its likely source, the inner part of the AD, is not occulted. Quite often the flickering reaches $\approx 0.1$ and, sometimes, $0.2$ mag amplitudes. The egress occurs at $\phi \approx 0.08$.
The rising and fading states occurring before and after normal outbursts are dominated by eclipses with low-amplitude ($\Delta V \lesssim 0.1$ mag) orbital hump and symmetric profiles, with a slightly gentler egress slope. The subsidence of orbital hump indicates that the hot and ionized AD contributes a higher proportion of the total system's light in these states, dominating their profiles. Symmetric low-amplitude hump eclipses are the most common throughout non-quiescent states. In bright outburst state the egress flux is higher than on the ingress. This is explained by the increasing presence of post-egress ``flares'' during this state, which could be attributed to flickering.
CG Dra demonstrates a wide variety of eclipse profiles throughout all of its states. The new data shows that CG Dra ingress and egress phases vary from quiescence to bright outbursts, getting slightly wider as the system progresses from quiescence to the bright outburst state. This suggests that the size of the AD varies with the state of the system. Eclipse depth also increases in this sequence. Both effects are more pronounced at the point of egress. Although only \boeclipses eclipses have been recorded in the bright outburst state, their profiles are of particular interest, as they represent the system in its extreme state. These eclipses are shown in figure \ref{f:brightoutbursts}. Data cadence is 30 s for all figures mentioned below. Common in this state are post-egress ``flares'' occurring immediately after the egress---see panes A, B and D for example, and broad post-egress humps, shown in panes C and D. Post-egress humps and ``flares'' occasionally appear not only during the bright outburst state---for example see pane C in figure \ref{f:pehs} for those in fading state after a normal outburst, and, also, panes B, E and F showing broad post-egress humps during normal outbursts. High-amplitude ``flares'' can be explained by the enhanced accretion flow and, hence, strong flickering in this state. However, the nature of post-egress humps at $\phi \approx 0.1 \to 0.5$ is uncertain. In contrast to eclipses with post-egress humps are highly asymmetric eclipses with egress magnitudes only slightly above the preceding eclipse minima---see examples in figure \ref{f:ha}. Four such eclipses were observed in the fading state of CG Dra, all on the final return to the quiescent state.
\begin{figure}
\centering
\includegraphics[width=3.4in]{brightoutbursts.png}
\caption{CG Dra light curves showing eclipses in its bright outburst state, with post-egress ``flares'' (panes A, B and D) and broad humps (panes C and D).}
\label{f:brightoutbursts}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=3.4in]{pehs_orb.png}
\caption{Occasionally CG Dra light curves show post-egress ``flares'' and broad post-egress humps. Phase plot, panes A and D: bright outburst, pane C: fading after normal outburst, panes B, E and F: normal outburst. }
\label{f:pehs}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=3.4in]{ha.png}
\caption{Highly asymmetric CG Dra eclipse profiles observed during fading to quiescent state, with egress magnitudes close or near the eclipse minima.}
\label{f:ha}
\end{figure}
\subsection{Outburst Cycle and Decline Rate} \label{outburstcycle}
\begin{figure}
\centering
\includegraphics[width=3.4in]{outburst_spectrogram.png}
\caption{Two-dimensional AoV spectrogram of frequencies in the domain of the CG Dra outburst cycle. Normal outbursts occur every $\sim$ 10 d.}
\label{f:os}
\end{figure}
CG Dra follows a quasiperiodic outburst cycle. A two-dimensional spectrogram of frequencies in the domain of the outburst cycle is shown in figure \ref{f:os}. The method to compute the spectrogram is similar to that described by \citet{2013PASJ...65...50O}, except for a few changes. Non-detrended normalized data is used as the input, with spectrogram step equal to $1/300$ and the window of $1/6$ of the whole dataset, without smoothing. An orthogonal multi-harmonic analysis of variance (AoV) algorithm was used to find periods \citep{1996ApJ...460L.107S} implemented in the \textsc{P4J} Python module \citep{2018ApJS..236...12H}.An AoV periodogram was computed for each step with resolution $R=1000$, and the number of harmonics, $n_h = 1$. Some discontinuities are visible along the time axis of the spectrogram which are attributed to the uneven sampling of the data.
\citet{2001IBVS.5124....1K}, observing two CG Dra outbursts, noted that CG Dra has an unusual outburst decline rate for its orbital period---0.14 and 0.31 mag d$^{-1}$. DNe are known to follow a well-defined decay time--$P_{orb}$ relation \citep{1975JBAA...86...30B, 1995cvs..book.....W}:
\begin{equation}
\tau_d = 0.53 P_{orb}^{0.84} \textup{(h) d mag}^{-1}.
\end{equation}
For CG Dra, applying $P_{orb}$ found in \S \ref{periods}, the expected decay time scale $\tau_d = 1.88$ d mag$^{-1}$, corresponding to 0.53 mag d$^{-1}$ expected decline rate. The derivative of CG Dra magnitude resampled, smoothed and interpolated to 1-hour bins indicates decline rates peak at $\sim 0.36$ and average at $\approx 0.25$ (mag d$^{-1}$), approximately half way to quiescence. Rise rates are about twice as fast. This confirms that CG Dra deviates from the expected relation for DNe.
\section{Ephemerides} \label{periods}
\subsection{Orbital Period and the O--C Chart}
In order to find the orbital period, the long-term CG Dra light curve was detrended using a locally weighted polynomial regression algorithm \citep{doi:10.1080/01621459.1979.10481038} implemented via \textsc{LOWESS} Python package by A. Lee. The bandwidth parameter $b=0.005$ was used which produces a smoothed light curve while preserving signals in the orbital period domain. The smoothed curve was subtracted from the original data to produce detrended light curve. The \textsc{P4J} AoV algorithm implementation was used to find the orbital period from the detrended data, with $R=10^4$ and $n_h=8$. The most prominent peak was refined with $R=10^5$. To estimate period error, as no interfering frequencies were detected and the level of the orbital period is quite strong, $S/N=76$, we adopt a simple procedure described by \citet{1991MNRAS.253..198S}, applicable to strong signals. The mean noise power level, $N^2$, was defined as the median of the spectrum in the vicinity of the frequency $f$ of the most prominent spectral line of power $p$, with boundaries set to $f\pm10\%$. The width of the spectral line at the $p-N^2$ level represents the $1\sigma$ confidence interval of the spectral line. The $P_{orb}$ found is 0.188640 $\pm$ 0.000007 d. This corresponds to $4^\textup{h} 31^\textup{m} 38^\textup{s} \pm 1^\textup{s}$.
The $P_{orb}$ value was tested by constructing the observed minus calculated (O--C) chart, shown in figure \ref{f:oc}, comparing observed eclipse minima calculated as described in \S \ref{photometry} with predicted values. Eclipses with poor Gaussian fits were removed. No significant period change from test ephemeris was detected.
\begin{figure}
\centering
\includegraphics[width=3.4in]{oc.png}
\caption{Observed minus calculated chart for CG Dra $P_{orb}=0.188640$ $\pm$ 0.000007 d. Fitted linear trend line is shown in blue.}
\label{f:oc}
\end{figure}
To determine the epoch of minimum, the method of \citet{1956BAN....12..327K} was used, implemented in the \textsc{Peranso} software. High-quality light curve was selected of an eclipse in bright outburst state of CG Dra, with suppressed flickering and symmetric eclipse profile dominated by AD light. The following ephemeris was obtained:
\begin{equation}
\textup{JD}_{min} = 2459788.508796(694) + E \times 0.188640(7).
\end{equation}
\subsection{Spectrogram in the Orbital Period Domain}
To search for potential superhump signatures, typically expected within a few percent of $P_{orb}$, we build a spectrogram within this domain of frequencies, shown in figure \ref{f:porbspec}. LOWESS-detrended data is used as the input, with spectrogram step equal to $1/100$ and the window of $1/20$ of the whole dataset, without smoothening. AoV parameters used were $R=1000$ and $n_h=8$. The orbital period is prominent in the spectrogram, along with a spectral line at $\approx$ 0.237 d, and a weaker one at $\approx 0.159$ d. These additional periodicities are likely associated with most data being taken at night at the same location, as they correspond to the true orbital frequency $f_{orb}=5.30 \pm 1$ (c d$^{-1})$. In other words, these lines likely represent an additional artificial diurnal cycle on both sides of the true frequency and are the aliases of the orbital frequency.
As an additional test we constructed a phase-dispersion minimization (PDM) \citep{1978ApJ...224..953S} spectrogram of the same detrended data. The PDM algorithm used was implemented in the \textsc{Astrobase} Python package \citep{wbhatti_astrobase}. A spectrogram window of 1/20 and step of 1/40, of the whole data set were used, with $R=100$. Again, only the orbital period is prominent, with hints of artificial aliases mentioned above. No other significant periods are detected in the orbital period domain via both AoV and PDM methods.
\begin{figure}
\centering
\includegraphics[width=3.4in]{porbspec.png}
\caption{Two-dimensional AoV spectrogram of frequencies in the domain of the CG Dra orbital period, $P_{orb}=0.18864$ d, prominent in this chart.}
\label{f:porbspec}
\end{figure}
\section{Discussion} \label{discussion}
CG Dra appears to exhibit two distinct types of outbursts, which we can call normal and bright outbursts. Bright outbursts resemble SU UMa-type superoutbursts, however, CG Dra does not show any superhump signatures and is located above the period gap, which makes its unambiguous classification difficult. No SU UMa-type stars have been found that show superoutbursts without superhumps. If bright outbursts are the equivalent of SU UMa superoutbursts then our observations might support models that do not specifically require eccentric ADs, or suggest that CG Dra may possibly be an intermediary U Gem subtype between SS Cyg and SU UMa.
Assuming CG Dra bright outbursts correspond to SU UMa superoutbursts, the lack of superhumps provides a challenge for the TTI model that requires superhumps to be present due to the eccentric AD. If superoutbursts appear due to the tidal dissipation in the eccentric disk then this also creates a problem, as CG Dra is located well above the period gap where these effects are not supposed to happen. Assuming CG Dra is related to the SS Cyg subtype, what then explains the appearance of bright outbursts? Long outbursts can actually occur in SS Cyg DNe---for example, the 45-day long outburst of U Gem in 1985, somewhat brighter than the normal outbursts of this star. \citet{2004AcA....54..433S} have, in fact, detected 0.3 mag amplitude superhumps in the 1985 data, ``appearing not later than 2--3 days after reaching maximum, and disappearing 4 days before final decline.'' This poses a challenge to the TTI model that is unable to explain superoutbursts and superhumps in systems with $P_{orb}$ above the period gap (U Gem $P_{orb} = 4.25$ h.)
\citet{2000A&A...353..244H} estimate that the mass accreted during the 1985 U Gem outburst was higher than the mass of the whole AD in quiescence. A likely trigger for such an outburst would have been the increased mass transfer rate from the secondary, probably caused by the irradiation of the secondary, supporting the EMT superoutburst model. However, this long outburst was an isolated episode, unlike the periodic bright outbursts we observe in CG Dra.
SS Cyg itself exhibits bi-modal outburst distribution, with long outbursts lasting $>12$ d, and short, lasting $< 12$ d \citep{2007PASP..119.1361P}. These are generally sequenced LS (long-short) and, less commonly, LLS, LSSS or LLSS. The amplitude of long and short outbursts tend be the same, while there are also occasional anomalous short outbursts of smaller amplitude. To compare this with CG Dra, its normal and bright outbursts differ significantly in amplitude and, with two bright outburst cycles observed, 4--5 normal outbursts are seen between the bright ones.
There is an example of a DN which exhibits properties of different subtypes at the same time---NY Serpentis that shows three distinctive type of outbursts: normal, wide outbursts without superhumps, and superoutbursts with superhumps \citep{2014PASJ...66..111P}. NY Ser could be an intermediary DN subtype---a case potentially applicable to CG Dra.
In addition to the problems presented above, CG Dra provides another challenge: unexpected spectral type of the secondary star. Although this work is focused on photometry, it is worth mentioning this problem, at least in brief. The first spectrophotometric observations of CG Dra were done by \citet{1990AGAb....4...24S, 1992A&A...266..225S} using the 2.2-m telescope at Calar Alto Observatory, and then later on 4.2-m William Herschel Telescope at La Palma by \citet{1997MNRAS.287..271S}, who identified the secondary's spectral type within the range of K5--7. \citet{1997A&A...325..601B}, observing with the 3.5-m telescope at Calar Alto, identified it as K5 $\pm$ 2. Assuming that the canonical CV mass-period relationship holds for CG Dra, the secondary mass can be estimated as
\begin{equation}
M_2 = 0.065 P_{orb}^{5/4} (\textup{h})
\end{equation}
for $1.3 \leq P_{orb} (\textup{h}) \leq 9$ \citep{1984ApJS...54..443P}. For CG Dra, $M_2 \approx 0.43 M_{\odot}$. A main sequence star of this mass is expected to be of M3--M4 spectral type \citep{2018MNRAS.479.5491E}. A K5 secondary would be significantly overweight, requiring longer $P_{orb} \sim 7$ h.
Yet another spectroscopic issue with CG Dra is that its emission and absorption spectrum lines move in phase \citep{1997A&A...325..601B}. In CVs, the absorption component is generally attributed to the secondary star, and emission---to the hotter AD, thus an opposite is expected. No reasonable explanation of this observation exists.
\section{Conclusion} \label{conclusion}
We have presented the most extensive photometric data set available on CG Dra to date. The orbital period measured was $P_{orb}=4^\textup{h} 31^\textup{m} 38^\textup{s} \pm 1^\textup{s}$. Throughout different states of the system a variety of eclipse profiles was observed, consistent with standard DN models with the exception of post-egress ``flares'' and broad humps. The presence of two types of quasi-periodic outbursts---normal and bright---was evident in the data. No superhump signatures were found in AoV and PDM spectrograms at frequencies in the domain of $P_{orb}$. Assuming the bright outbursts of CG Dra correspond to SU UMa-type superoutbursts, this creates an issue of explaining, within the TTI model, how superoutbursts appear without superhumps in a DN that is significantly above the period gap, where SU UMa-type stars should not be found. For both normal and bright outbursts, we confirm that the decline rates observed are too slow for known DN decay time--$P_{orb}$ relations.
Interpreting CG Dra within canonical DN classification and models is problematic. Do post-egress ``flares'' during bright outburst and, occasionally and to a lesser extent, in normal outburst state, represent EMT episodes that lead to superoutbursts? What triggers bright outbursts in this system and are post-egress humps related to them? Further observations are required to understand the physical nature of this system. We hope that observations presented herein will be helpful for further research. The photometry obtained is openly available via BAA and AAVSO variable star databases.
\section*{Acknowledgements}
This research has made use of National Aeronautics and Space Administration (NASA) Astrophysics Data System Bibliographic Services and the SIMBAD database, operated at CDS, Strasbourg, France. The authors acknowledge with thanks the variable star observations from the AAVSO International Database and BAA Variable Star Section contributed by observers worldwide and used in this research.
|
\section{Introduction}
\label{sec:introduction}
Metamorphism is an integral transform recently introduced to treat partial differential equations~\cite{Kisil21c}. In particular, metamorphism allows one to reduce the order of a differential equations: e.g. a second order differential equation can be transformed to a first order admitting a straightforward solution and transparent geometrical structure~\cites{AlmalkiKisil18a,AlmalkiKisil19a}. Basic properties of the metamorphism can be verified by direct calculations---the path which was intentionally chosen to reduce the amount of prerequisites in the introductory paper~\cite{Kisil21c}.
Yet, a genuine origin of metamorphism is a covariant transform related to the Schr\"o\-din\-ger--Jacobi group~\cites{Folland89,Berndt07a} as was already presented in the Jupyter notebooks~\cite{Kisil21b} with respective symbolic computations.
This paper systematically utilises the group theory and covariant transform technique to reinstall the metamorphism transform from a scratch. Furthermore, some sister integral transforms are appearing as well. The paper can be seen as a readable narrative to a Jupyter notebook~\cite{Kisil21b}, which will be frequently referred here to replace some boring calculations. Our main result is a characterisation of the metamorphism image space in Thm.~\ref{th:metamorphism-image-characterisation}.
We made this paper as accessible as possible. Its reading does not require an advanced knowledge of group representations and the theory of covariant transform. We provide most of required information with further references to more detailed presentations if needed.
In Sect.~\ref{se:groups} we introduce several groups: the Heisenberg, $\SL$, affine, Schr\"o\-din\-ger, and finally our main object---the group SSR. Essential relations between those groups are presented as well. We describe some (not all) induced representations of the group SSR in Sect.~\ref{se:induced-representations}. The corresponding covariant transform and its properties are described in Sect.~\ref{se:covariant-transform}. Finally, we connect a selection of a fiducial vector with the properties of the image space of covariant transform in Sect.~\ref{sec:image-space-covar-trans}. In particular, the metamorphism is defined as the covariant transform with a remarkable fiducial vector---the Gaussian. Covariant transforms with some other mentioned fiducial vectors are still awaiting their investigation.
\section{Heisenberg, $\SL$, affine, Schr\"o\-din\-ger and SSR groups}
\label{se:groups}
We start from a brief account of groups involved in the consideration. An element of the one-dimensional Heisenberg group \(\mathbb{H}\)~\cites{Folland89,Kisil10a,Kisil19b} will be denoted by \(( s, x, y)\in \Space{R}{3}\). The group law on \(\mathbb{H}\) is defined as follows:
\begin{displaymath}
(s, x, y)\cdot (s ',\, x ',\, y' ) =( s+ s'+ \half{\omega}( x, y; x', y'),\, x+\, x', \,y+ \, y'),
\end{displaymath}
where
\begin{equation}
\label{eq:symplectic-form}
{\omega} ( x, y; x', y')= xy'- x'y
\end{equation}
is the symplectic form~\amscite{Arnold91}*{\textsection{}41} on \(\Space{R}{2}\).
The identity element in \(\Space{H}{}\) is \(( 0, 0, 0)\), and the inverse of \(( s, x, y)\) is \(( -s, -x, -y)\).
There is an alternative form of \(\Space{H}{}\) called the polarised Heisenberg group \(\Space[p]{H}{}\) with the group law~\citelist{\amscite{Folland89}*{\textsection{}1.2} \cite{AlameerKisil21a}}
\begin{displaymath}
(s,\, x, \,y)\cdot (s ',\, x ',\, y' )= (\, s+ s' + x y',\, x+ x ',\, y+ y ').
\end{displaymath}
and the group isomorphism \(\theta : \Space{H}{}\rightarrow \Space[p]{H}{}\) given by
\begin{displaymath}
\Theta: ( s, x, y)\rightarrow ( s + \half xy,\, x,\, y).
\end{displaymath}
The special linear group \(\SL\) is the group of \(2\times 2\) matrices with real entries and the unit determinant~\cites{Lang85,Kisil12a}. The group law on \(\SL\) coincides with the matrix multiplication.
A matrix \(A\in\SL\) acts on vectors in \(\Space{R}{2}\) by a symplectomorphism, i.e. an
automorphisms of the symplectic form \(\omega\)~\eqref{eq:symplectic-form}:
\begin{displaymath}
\omega(A(x,y);A(x',y'))=\omega(x,y;x',y').
\end{displaymath}
Therefore, the transformation \(\theta_{A}: \Space{H}{}\rightarrow \Space{H}{}\)
\begin{displaymath}
\theta_{A}: ( s, x, y)\rightarrow ( s, A( x, y))
\end{displaymath}
is an automorphism of \(\Space{H}{}\)~\amscite{Folland89}*{\textsection{}1.2}.
The corresponding polarised automorphism \(\theta_{A}^{p} = \Theta \circ \theta_A \circ \Theta^{-1}: \Space[p]{H}{}\rightarrow \Space[p]{H}{}\) is
\begin{displaymath}
\theta_{A}^{P}( s, x, y)
\left( s + \half ( ac x^{2} +2bc xy + bd y^{2}), ax + by, cx+dy\right),
\end{displaymath}
where \(A=\begin{pmatrix} a&b\\c&d\end{pmatrix}\).
Upper-triangular matrices in \(\SL\) with positive diagonal entries form a subgroup \(\Space{A}{}\). We parameterise it by pairs \((b, r) \in \Space[+]{R}{2}\) with \(b\in\Space{R}{}\) and \(r>0\) as follows:
\begin{equation}
\label{eq:affine-group}
\begin{pmatrix}1&b\\0&1\end{pmatrix}\,
\begin{pmatrix}r&0\\0&1\slash r\end{pmatrix}\,
=\begin{pmatrix}r&b\slash r \\0&1\slash r \end{pmatrix}.
\end{equation}
The subgroup is isomorphic to the affine group of the real line also known as the \(ax+b\) group~\cite{Kisil12d}.
For a group acting by automorphism on another group we can define their semi-direct product. The model case is the affine group itself, where dilations act as automorphisms of shifts. Formally, let \(G\) and \(H\) be two groups and assume \(\theta : H\rightarrow Aut(G)\), where \(\theta_h\) is an automorphism of \(G\) corresponding to \(h\in H\). The semi-direct product of \(G\) by \(H\) denoted by \(G\rtimes H\) is the Cartesian product of \(G\times H\) with the group law
\begin{equation}
\label{eq:semiderect-law}
( g_{1}, h_{1})\,\cdot ( g_{2}, h_{2})= ( g_{1} \theta_{h_{1}}(g_{2}), h_{1} h_{2}),
\end{equation}
where \(( g_{1}, h_{1}),\, ( g_{2}, h_{2})\in G \times H\).
The semidirect product of the Heisenberg group and \(\SL\) is called Schr\"o\-din\-ger group \(\mathbb{S}\), which is the group of symmetries of the Schr\"o\-din\-ger equation~\cites{Niederer72a,KalninsMiller74a} and parabolic equations~\cite{Wolf76a} with applications in optics~\cites{ATorre08a,ATorre10a}. In the context of number theory it is also known as the Jacobi group~\cite{Berndt07a}.
Our main object here is the group \(\Space{G}{} \coloneqq \Space{H}{} \rtimes \Space{A}{}\), which is the semi-direct product of the Heisenberg group \(\Space[p]{H}{}\) and the affine group \(\Space{A}{}\)~\eqref{eq:affine-group} acting by symplectic automorphism of \(\Space[p]{H}{}\). Thus, \(\Space{G}{}\) is a subgroup of the Schr\"o\-din\-ger group. It can be also called shear-squeeze-rotation (SSR) group~\cite{Kisil21b} by three types of transformations of Gaussian coherent states. A subgroup of \(\Space{G}{}\) without squeeze (i.e. \(r=1\) in~\eqref{eq:affine-group}) is called the shear group and it was used in a similar context in~\cites{AlmalkiKisil18a,AlmalkiKisil19a}. This nilpotent step 3 group is also known as the Engel group~\cite{Chatzakou22a}.
Let \(( s, x, y, b, r)\in \Space{G}{}\)
where \( ( s, x, y) \in \Space[p]{H}{}\) and \(( b, r) \in \mathbb{A}\).
Explicitly the group law~\eqref{eq:semiderect-law} on \(\Space{G}{}\) is~\cite{Kisil21b}
\begin{align*}
( s, x, y, b, r) \cdot ( s', x', y', b', r')
&=( s+ s'+ x {r}^{-1} y' - \half\, b \,{({r}^{-1} {y'})}^{2},\\
& x + r x'- b {r}^{-1} y',\, y + {r}^{-1} y', \, b + b' {r}^{2},\, r r').
\end{align*}
There is a convenient matrix realisation of \(\Space{G}{}\)~\cite{Kisil21b}
\begin{displaymath}
( s, x, y, b, r)=
\begin{pmatrix}
1&-yr&({x+ b y})/{r}&2s-yx\\
0&r&-b/r&x\\
0&0&1/r&y\\
0&0&0&1
\end{pmatrix}.
\end{displaymath}
The corresponding solvable Lie algebra \(\algebra{g}\) has a basis \(\{ S, X, Y, B, R\}\), with the following non-vanishing commutators:
\begin{equation}
\label{eq:lie-algebra-commutators}
[ X, Y]= S,\quad [ X, R]= -X,\quad [ Y, R]= Y, \quad
[ Y, B]= X,\quad [ R, B]= 2B.
\end{equation}
Clearly, the group \(\Space{G}{}\) is a not a commutative Lie group.
\section{Induced representations the group $\Space{G}{}$}
\label{se:induced-representations}
In this section we construct several induced representations of the group \(\Space{G}{}\), which are required for our study. First, we recall the general scheme of induced representations. For simplicity, only inductions from characters of subgroups are considered and it is sufficient for our present purposes. For further details and applications of induced representations see~\cites{AliAntGaz14a,Folland16a,Mensky76,Mackey70a,Kisil09e}.
\subsection{Induced representation from a subgroup character}
Let \(G\) be a group and \(H\) be a subgroup of \(G\). The space \(X= G \slash H\) of the left cosets \(gH\) of the subgroup \(H\) is given by the equivalence relation: \( g\sim g'\) if there exists \( h\in H\) such that \(g = g' h\). We define the natural projection \( \mathbf{p}: G \rightarrow X\) such that \( \mathbf{p}(g)= g H\).
Let us fix a section \(\mathbf{s}: X\rightarrow G\) such that \(\mathbf{p} \circ \mathbf{s}= I\), where \(I\) is the identity map on \(X\).
An associated map \(\mathbf{r}: G\rightarrow H\) by
\begin{equation}
\label{eq:map-r-formula}
\mathbf{r}( g)= {\mathbf{s}( \mathbf{p}( g))}^{-1}\cdot g.
\end{equation}
provides the unique decomposition of the form~\amscite{Kirillov76}*{\textsection{}13.2}
\begin{displaymath}
g = \mathbf{s} ( \mathbf{p}(g)) \cdot \mathbf{r}(g), \qquad \text{ for any } g \in G.
\end{displaymath}
Thus, \(X\) is a left homogeneous space with the \(G\) action as follows:
\begin {equation}
\label{eq:G-space-action}
g^{-1}: x\rightarrow g^{-1} \cdot x = \mathbf{p}\, ( g^{-1} * \mathbf{s}(x)),
\end{equation}
where \( * \) is the multiplication of \(G\) and \(\cdot\) is the action of \(G\) on \(X\) from the left.
Suppose \(\chi : H\rightarrow \Space{T}{}\) be a character of the subgroup \(H\).
Let \(\FSpace[\chi]{L}{2}( G)\) be a Hilbert space of functions on \(G\) with a \(G\)-invariant inner product and the \(H\)-covariance property
\cite{Kisil17a},
\begin{equation}
\label{eq:covariance-property}
F( g h) = \Bar{\chi} (h) \, F (g), \qquad \text{ for all } g\in G, \ h \in H.
\end{equation}
The space \(\FSpace[\chi]{L}{2}( G)\) is invariant under the left regular representation by \(G\)-shifts
\begin{equation}
\label{eq:left-regular-action}
\Lambda (g): F(g')\rightarrow F( g^{-1} g'), \quad \text{ where } g , g' \in G.
\end{equation}
The restriction of \( \Lambda\) to the space \(\FSpace[\chi]{L}{2}( G)\) is called the induced representation from the character \(\chi\).
An equivalent form of the induced representation can be constructed as follows~\cites{Kirillov76,Kisil17a}. We define a lifting \( \oper{L}^{\chi} : \FSpace{L}{2}(X) \rightarrow \FSpace[\chi]{L}{2}(G) \) as the map
\begin{equation}
\label{eq:lifting}
[\oper{L}^{\chi} f] ( g)= \overline{\chi}( \mathbf{r}( g)) \,f (\mathbf{p}( g)).
\end{equation}
The pulling \(\mathcal{P}: \FSpace{L}{2} ^{\chi}(G)\rightarrow \FSpace{L}{2}( X)\) given by
\begin{equation}
\label{eq:pulling}
[\mathcal{P} F] (x)= F (\mathbf{s} ( x)).
\end{equation}
Clearly \(\mathcal{P} \circ \oper{L}^{\chi} = I\) on \(\FSpace{L}{2}(X)\).
From~\eqref{eq:lifting},~\eqref{eq:pulling}, the induced representation \( \rho: \FSpace{L}{2}( X) \rightarrow \FSpace{L}{2}( X)\) is defined by the formula:
\begin{displaymath}
\rho_{\chi}( g) = \mathcal{P} \circ \Lambda ( g) \circ \oper{L}^{\chi},
\end{displaymath}
where \(\Lambda ( g)\) is the left regular representation~\eqref{eq:left-regular-action}.
The representation \(\rho_{\chi}\) explicitly is
\begin{equation}
\label{eq:induced-rep-homogen}
[\rho_{\chi}(g)](x)= \bar{\chi}( \mathbf{r}( g^{-1} \,\mathbf{s} (x))) \, f( g^{-1}\cdot x),
\end{equation}
where \(g\in G\) and \(x\in X\) and \(g^{-1}\cdot x\) is defined by~\eqref{eq:G-space-action}. For a \(G\)-invariant measure \(\mu\) on \(X\) the representation~\eqref{eq:induced-rep-homogen} is unitary on the space \(\FSpace{L}{2}(X,\mu)\)
\subsection{Derived representations}
In this subsection \(G\) is a Lie group with the corresponding Lie algebra \(\algebra{g}\).
Let \(\rho\) be a representation of \(G\) in a Hilbert space \(\mathcal{H}\), the derived representation of \(X\in \algebra{g}\) denoted as \(\rmd\rho^{X}\) is given by
\begin{equation}
\label{eq:derived-representation}
\ifspringer
\rmd{\rho}^{X} \phi = \frac{\rmd\ }{\rmd t}\rho(\exp(t X)) \phi \mid_{t=0},
\else
\rmd{\rho}^{X} \phi = \left. \frac{\rmd\ }{\rmd t}\rho(\exp(t X)) \phi \right|_{t=0},
\fi
\end{equation}
where the vector \(\phi \in \mathcal{H}\) is such that the vector-function \(g \rightarrow \rho (g) \phi\) is infinitely-differentiable for any \(g \in G\). These vectors are called smooth and constitute a linear subspace, denoted \(\mathcal{D}^{\infty}\), of \(\mathcal{H}\) which is dense in \(\mathcal{H}\). It is easy to show that \(\mathcal{D}^{\infty}\) is invariant under \(\rho (g)\)~\amscite{Lang85}*{\textsection{}6.1}.
If \(\mathcal{H}\) is \(\FSpace{L}{2}(\Space{R}{n})\) then the space \(D^{\infty}\) contains the Schwartz space, which is a dense subspace of \(\FSpace{L}{2}(\Space{R}{n})\).
Also, we define the Lie derivative \(\mathcal{L}^{X}\) for \(X \in \algebra{g}\) as the derived right regular representation~\amscite{Lang85}*{\textsection{}6.1}, that is
\begin{equation}
\label{eq:lie-derivative}
\ifspringer
[\mathcal{L}^{X} F](g)= \frac{d\ }{dt} F ( g\, \exp(t X))\mid_{t=0},
\else
[\mathcal{L}^{X} F](g)= \left. \frac{d\ }{dt} F ( g\, \exp(t X))\right|_{t=0},
\fi
\end{equation}
for any differentiable function \(F\) on \(G\).
\subsection{Left regular representation of group $\Space{G}{}$}
The left and right invariant Haar measures of the group \(\Space{G}{}\) are given by
\begin{displaymath} \rmd_l( s, x, y, b, r) = \rmd s\, \rmd x \, \rmd y\, \rmd b \, \frac{\rmd r}{{r}^{3}},\end{displaymath}
\begin{displaymath} \rmd_r( s, x, y, b, r) = \rmd s\, \rmd x \, \rmd y \, \rmd b\, \frac{\rmd r}{r}.\end{displaymath}
Thus, the group \(\Space{G}{} \) is non-unimodular with the modular function \(\Delta (s, x, y, b, r)=\textstyle \frac{1}{r^{2}}\).
We extend the action~\eqref{eq:left-regular-action} of \(\Space{G}{}\) on itself by left shifts to the left regular unitary representation on the linear space of functions \(\FSpace{L}{2}(\Space{G}{},d_l)\):
\begin{equation}
\label{G-left-regular}
\begin{split}
\lefteqn{[ \Lambda ( s, x, y, b, r) F]( s', x', y', b', r') =F( s'- s + x (y' - y) - \half b {( y' - y)}^{2},}&\qquad{}\\
&\qquad \qquad \qquad \qquad \qquad \frac{1}{r} (x' -x) + \frac{b}{r} (y'- y),\, r (y'- y), \, \frac{1}{r^{2}} ( b'- b),\, \frac{r'}{r}),
\end{split}
\end{equation}
where \(( s, x, y, b, r)\), \(( s', x', y', b', r') \in \Space{G}{}\).
This representation is reducible, i.e. there are \(\Lambda\)-invariant proper subspaces in \(\FSpace{L}{2}(\Space{G}{},d_l)\). In particular, many types of induced representations of \(\Space{G}{}\) are realised as restrictions of the left regular representations~\eqref{G-left-regular} to some subspaces with a covariance property~\eqref{eq:covariance-property}.
We describe here two of them---called the quasi-regular type representation and the Schr\"o\-din\-ger type representation---together with equivalent forms on the respective homogeneous spaces.
\subsection{Quasi-regular representation of the group $\Space{G}{}$}
Let
\begin{displaymath}
Z=\{( s, 0, 0, 0, 1), s\in \Space{R}{}\}
\end{displaymath}
be the centre of the group \(\Space{G}{}\). The space of left cosets \(X= \Space{G}{}\slash Z\) can be parametrised by
\begin{displaymath}
\Space[+]{R}{4} = \{ (x, y, b, r) \in \Space{R}{4}: \ r > 0\}.
\end{displaymath}
Consider the natural projection and the section maps
\begin{align}
\nonumber
\mathbf{p}( s, x, y, b, r)&\rightarrow ( x, y, b, r),\\
\label{eq:map-s-center}
\mathbf{s}( x, y, b, r) &\rightarrow( 0, x, y, b, r).
\end{align}
We calculate the respective map \( \mathbf{r}\)~\eqref{eq:map-r-formula} as follows
\begin{displaymath}
\begin{split}
\mathbf{r} ( s, x, y, b, r)&= \mathbf{s}( \mathbf{p}( s, x, y, b, r))^{-1} ( s, x, y, b, r)\\
&=( s, 0, 0, 0, 1).
\end{split}
\end{displaymath}
Let \(\chi_{\hbar} : Z\rightarrow \Space{T}{}\) be an unitary character of \(Z\):
\begin{displaymath} \chi_{\hbar} ( s, 0, 0, 0, 1) = \rme^{ 2 \pi \rmi \hbar s},\end{displaymath}
defined by a parameter \( \hbar \in \Space{R}{}\). In quantum mechanical framework \(\hbar\) is naturally associated to the Planck constant~\cites{Folland89,Kisil02e,Kisil09e,Kisil17a}.
The corresponding induced representation \(\tilde{\rho} : \FSpace{L}{2}(\Space[+]{R}{4})\rightarrow \FSpace{L}{2}(\Space[+]{R}{4})\) is~\cite{Kisil21b}
\begin{equation}
\label{eq:quasi-regular}
\begin{split}
\lefteqn{[\tilde{\rho} ( s, x, y, b, r)f]( x', y', b', r')=
\rme^{ 2 \pi \rmi \hbar( s + x ( y' - y) - b {( y' - y)}^{2}/2)}}&\qquad\\
& \qquad \qquad \qquad \qquad {} \times f( \textstyle\frac{1}{r}( x'- x) +\textstyle \frac{b}{r} ( y'- y), r ( y'- y), \textstyle\frac{1}{r^{2}}(b' - b ), \textstyle\frac{r'}{r}).
\end{split}
\end{equation}
It is called the quasi-regular type representation on \(\FSpace{L}{2}(\Space[+]{R}{4})\). One can check that \(\tilde{\rho}\) is unitary and we will discuss its reducibility below.
\subsection{Schr\"o\-din\-ger type representation of the group $\Space{G}{}$}
Let
\begin{displaymath}
H_{1}=\{( s, x, 0, b, r),\, s,\, x, b \in \Space{R}{}, r \in \mathbb{R_{+}}\}
\end{displaymath}
be a subgroup of \(\Space{G}{}\), which is a semidirect product of a maximal abelian subgroup of \(\Space{H}{}\) and the affine group \(\Space{A}{}\). The space of the left cosets \( \Space{G}{}\slash H_{1}\) is parameterized by \(\Space{R}{}\).
We define the natural projection \(\mathbf{p}: \Space{G}{} \rightarrow \Space{R}{}\) and a section map \(\mathbf{s}: \Space{R}{} \rightarrow \Space{G}{} \) by
\begin{align*}
\mathbf{p}( s, x, y, b, r) &= y, \\
\mathbf{s}( y) & = ( 0, 0, y, 0, 1).
\end{align*}
The respective map \( \mathbf{r}\)~\eqref{eq:map-r-formula} is
\begin{displaymath}
\begin{split}
\mathbf{r}( s, x, y, b, r)&=\mathbf{s}( \mathbf{p}( s, x, y, b, r))^{-1} ( s, x, y, b, r)\\
&=( s, x, 0, b, r) .
\end{split}
\end{displaymath}
Let \({\chi}_{\hbar \lambda} : H_{1} \rightarrow \Space{T}{}\) be a character \(H_{1}\)
\begin{displaymath}
\chi_{ \hbar \lambda} ( s, x, 0, b, r) = \rme^{ 2\pi \rmi \hbar s} \, r^{ \lambda +\frac{1}{2}},
\end{displaymath}
where \(\hbar\in \Space{R}{}\), \(\lambda \in i{\Space{R}{}}\). For simplicity, we will consider here the case of \(\lambda =0\) only. The induced representation on \( \FSpace{L}{2}(\Space{R}{})\) is~\cite{Kisil21b}
\begin{equation}
\label{eq:schrodinger-type}
[\rho ( s, x, y, b, r)f]( u)=\sqrt{r} \, \rme^{ 2\pi \rmi \hbar ( s + x( u - y) - b {( u - y)}^{2}/2)} \, f( r \,(u - y)).
\end{equation}
\begin{rem}
The structure of this representation can be illuminated through its restrictions to the following subgroups:
\begin{itemize}
\item The affine group \(\mathbb{A}\), i.e. the substitution \(s=x=y=0\). The restriction is the co-adjoint representation of the affine group~\citelist{\amscite{Folland16a}*{\textsection{}6.7.1} \cite{Kisil12d}}:
\begin{displaymath}
[\rho ( 0, 0, 0, b, r)f]( u)= \sqrt{r}\, \rme^{ \pi \rmi \hbar \, b\, { u}^{2} } \, f( r\, u).
\end{displaymath}
Through the Fourier transform it is unitary equivalent to the quasi-regular representation of the affine group, which is the keystone of the wavelet theory and numerous results in complex and harmonic analysis~\cite{Kisil12d}.
\item The Heisenberg group, that is \(r=1\) and \(b=0\). The restriction is the celebrated Schr\"o\-din\-ger representation~\cites{Folland89,Kisil17a}:
\begin{displaymath}
[\rho ( s, x, y, 0, 1)f]( u)=\rme^{ 2\pi \rmi \hbar ( s + x( u - y)) } \, f(u - y),
\end{displaymath}
which plays the crucial r\^ole in quantum theory.
\item The third subgroup is the Gabor group with \(b=0\). The representation is
\begin{displaymath}
[\rho ( s, x, y, r, 0)f]( u)= \rme^{ 2\pi \rmi \hbar ( s + x( u - y))} r^{ \frac{1}{2} }\, f( r \,(u - y)).
\end{displaymath}
It is involved in Gabor analysis and Fourier--Bros--Iagolnitzer (FBI) transform~\amscite{Folland89}*{\textsection{}3.3}.
\item Finally, the shear group corresponding to \(r=1\). The restriction is
\begin{displaymath}
[\rho ( s, x, y, 1, b)f]( u)= \rme^{ 2\pi \rmi \hbar ( s + x( u - y) - b {( u - y)}^{2}/2)} f(u - y),
\end{displaymath}
It was employed in~\cites{AlmalkiKisil18a,AlmalkiKisil19a} to reduce certain quantum Hamiltonians to first-order differential operators.
\end{itemize}
\end{rem}
In view of the mentioned connections, we call representation~\eqref{eq:schrodinger-type} as Schr\"o\-din\-ger type representation. It is irreducible since its restriction to the Heisenberg group coincides with the irreducible Schr\"o\-din\-ger representation~\cites{Folland89, Kisil17a}.
The derived representation~\eqref{eq:derived-representation} of the Schr\"o\-din\-ger type representation~\eqref{eq:schrodinger-type} is
\begin{align}
\nonumber
\rmd{\rho}^{X}&= 2 \pi \rmi \hbar u I, &
\rmd{\rho}^{B}&= - \pi \rmi \hbar u^{2} I,\\
\label{eq:schrodinger-derived}
\rmd{\rho}^{Y}&= -\frac{\rmd\ }{\rmd u},&
\rmd{\rho}^{R}&= \half I+u\, \frac{\rmd\ }{\rmd u},\\
\nonumber
\rmd{\rho}^{S}&= 2 \pi \rmi \hbar I.
\end{align}
It is easy to check that the above sets of operators
\eqref{eq:schrodinger-derived} represents commutators~\eqref{eq:lie-algebra-commutators} of the Lie algebra \(\algebra{g}\) of the group \(\Space{G}{}\).
\section{Covariant transform}
\label{se:covariant-transform}
The covariant transform plays a significant r\^ole in various fields of mathematics and its applications~\cites{Perelomov86, Berezin86, AliAntGaz14a, Folland89, Kisil11c, Kisil17a, Kisil12d}. We present here some fundamental properties of the covariant transform which have implications for the metamorphism transform.
\subsection{Induced covariant transform}
Let \(G\) be a group and let \(\rho\) be a unitary irreducible representation of the group \(G\) in a Hilbert space \(\mathcal{H}\). For a fixed unit vector \(\phi \in \mathcal{H}\), called here a fiducial vector (aka vacuum vector, ground state, mother wavelet, etc.), the covariant transform \(\oper{W}_{\phi}: \mathcal{H}\rightarrow L(G)\)
\ifspringer
is~\cite[\textsection{}8.1]{AliAntGaz14a}, \cite{Berezin86,Perelomov86}
\else
is~\citelist{\cite{AliAntGaz14a}*{\textsection{}8.1} \cite{Berezin86} \cite{Perelomov86}}
\fi
\begin{equation}
\label{eq:covariant-transform}
[ {\oper{W}_{\phi}}f](g)= \langle f, \rho(g) \phi \rangle, \qquad
\text{ where } f\in \mathcal{H} \text{ and } g\in G.
\end{equation}
The main property of~\eqref{eq:covariant-transform} is that \({\oper{W}_{\phi}}\) intertwines the representation \(\rho\) on \(\mathcal{H}\) and the left regular action \(\Lambda\)~\eqref{eq:left-regular-action} on \(G\):
\begin{equation}
\label{eq:covariant-intertwining}
\oper{W}_{\phi} \circ \rho( g) = \Lambda (g) \circ \oper{W}_{\phi},\quad \text{ for all } g\in G.
\end{equation}
A representation \(\rho\) is square-integrable if for some \(\phi \in \mathcal{H}\), the map \(\oper{W}_{\phi}: \mathcal{H} \rightarrow \FSpace{L}{2}(G, dg)\) is unitary for a left Haar measure \(dg\) on \(G\). Some representations are not square-integrable, but can still be treated by the following modification of covariant transform by Perelomov~\cite{Perelomov86}. Let \(H\) be a closed subgroup of the group \(G\) and the corresponding homogeneous space is \(X= G\slash H\). Let for some character \(\chi\) of \(H\) a fiducial vector \(\phi \in \mathcal{H}\) is a joint eigenvector
\begin{equation}
\label{eq:eigenvector-subgroup}
\rho (h) \, \phi = \chi (h) \phi, \qquad \text{for all } h \in H.
\end{equation}
Then, the respective covariant transform satisfies the covariant property, cf.~\eqref{eq:covariance-property}:
\begin{displaymath}
[\oper{W}_{\phi} f](g h
= \overline{\chi}(h) [\oper{W}_{\phi} f] (g).
\end{displaymath}
Thus, the image space of \(\oper{W}_{\phi}\) belongs to the induced representation by the character \(\chi\) of the subgroup \(H\). This prompts to adopt the covariant transform to the space of function on the homogeneous space \(X=G/H\).
To this end, let us fix a section \(\mathbf{s}: X \rightarrow G\) and a fiducial vector \(\phi\in \mathcal{H}\) satisfying~\eqref{eq:eigenvector-subgroup}. The induced covariant transform from the Hilbert space \(\mathcal{H}\) to a space of functions \(\FSpace{L}{\phi}(X)\) is
\begin{displaymath}
[ {\oper{W}_{\phi}}f](x)= \langle f, \rho(s(x)) \phi \rangle,\quad
\text{ where } x\in X.
\end{displaymath}
Then, the induced covariant transform intertwines \(\rho\) and \(\tilde{\rho}\)---an induced representation from the character \(\chi\) of the subgroup \(H\), cf.~\eqref{eq:covariant-intertwining}:
\begin{equation}
\label{eq:covariant-intertwining-induced}
W^{\rho}_{\phi}\circ \rho(g) = \tilde{\rho}(g) \circ W^{\rho}_{\phi}, \qquad \text{ for all } g\in G.
\end{equation}
In particular, the image space \(\FSpace{L}{\phi}( G\slash H)\) of the induced covariant transform is invariant under \(\tilde{\rho}\). Induced covariant transforms for the Heisenberg group~\cite{Kisil09e} and the affine group~\cite{Kisil12d} are the most familiar examples.
\subsection{Induced covariant transform of the group $\Space{G}{}$}
On the same way as above, we can calculate the induced covariant transform of \(\Space{G}{}\).
Consider the subgroup \(Z\) of \(\Space{G}{}\), which is \(Z= \{( s, 0, 0, 0, 1), s\in \Space{R}{}\}\). For the Schr\"o\-din\-ger type representation~\eqref{eq:schrodinger-type}, any function \(\phi \in \FSpace{L}{2}(\Space{R}{})\) satisfies the eigenvector condition \(\rho( s, 0, 0, 0, 1) \phi= \rme^{2 \pi \rmi \hbar s} \phi\) with the character \(\chi ( s, 0, 0, 0, 1)= \rme^{-2 \pi \rmi \hbar s}\), cf.~\eqref{eq:eigenvector-subgroup}. Thus, the respective homogeneous space is \(\Space{G}{}\slash Z \simeq \Space[+]{R}{4}\) and we take the above section \(\mathbf{s}: \Space{G}{}\slash Z \rightarrow \Space{G}{}: \ \mathbf{s}( x, y, b, r)= ( 0, x, y, b, r)\)~\eqref{eq:map-s-center}. Then, the induced covariant transform is
\begin{equation}
\label{eq:covariant-tr-schrodinger-type}
\begin{split}
[ \oper{W}_{\phi} f]( x, y, b, r)&= \langle f, \rho(\mathbf{s} ( x, y, b, r)) \phi \rangle \\
&= \langle f,\rho( 0,x, y, b, r) \phi \rangle\\
&= \int_{\Space{R}{}}{f (u)\, \overline{\rho( 0, x, y, b, r) \, \phi ( u)}} \, \rmd u\\
&= \int_{\Space{R}{}}{ f ( u) \,\rme^{- 2 \pi \rmi \hbar (x ( u -y) - b (u - y)^{2}/2)} \,r^{ \frac{1}{2}} \, \overline\phi (r( u- y))} \, \rmd u\\
&=\sqrt{r} \int_{\Space{R}{}}{ f ( u)\, \rme^{-2 \pi \rmi \hbar (x ( u -y) - b (u - y)^{2}/2)}\, \overline\phi ( r( u- y))} \, \rmd u.
\end{split}
\end{equation}
From~\eqref{eq:covariant-intertwining-induced}, \(\oper{W}_{\phi}\) intertwines the Schr\"o\-din\-ger type representation~\eqref{eq:schrodinger-type} with quasi-regular~\eqref{eq:quasi-regular}.
The last integral in~\eqref{eq:covariant-tr-schrodinger-type} is a composition of five unitary operators \(\FSpace{L}{2}(\Space{R}{2})\rightarrow \FSpace{L}{2}(\Space{R}{2}) \) applied to a function \(F(y,u)=f(y)\overline {\phi}(u)\) in the space \(\FSpace{L}{2}(\Space{R}{})\otimes \FSpace{L}{2}(\Space{R}{}) \simeq \FSpace{L}{2}(\Space{R}{2})\):
\begin{enumerate}
\item The unitary operator \(R: \FSpace{L}{2}(\Space{R}{2}) \rightarrow \FSpace{L}{2}(\Space{R}{2})\) based on the dilation
\begin{displaymath}
R : F( y, u)\rightarrow \sqrt{r}\, F( y, ru), \qquad \text{ where } r > 0.
\end{displaymath}
\item The change of variables \(T: \FSpace{L}{2}(\Space{R}{2}) \rightarrow \FSpace{L}{2}(\Space{R}{2})\)
\begin{displaymath}
T: F( y, u)\rightarrow F(u, u-y).
\end{displaymath}
\item The operator of multiplication by an unimodular function \(\psi_{b}( x, y)= \rme^{\pi \rmi \hbar b (u -y)^{2}}\),
\begin{displaymath}
M_{b}: F( y,u) \rightarrow \rme^{\pi \rmi \hbar b (u -y)^{2}} \,F ( y, u), \qquad
\text{ where } b\in \Space{R}{}.
\end{displaymath}
\item The partial Fourier transform \(u\rightarrow x\) in the second variable
\begin{displaymath}
[\mathcal{F}_{2} F]( y, x)= \int_{\Space{R}{}}{ F( y, u)\, \rme^{- 2 \pi \rmi \hbar x u}\, \rmd u}.
\end{displaymath}
\item The multiplication \(M\) by the unimodular function \(\rme^{2 \pi \rmi \hbar x y}\).
\end{enumerate}
Thus, we can write \(\oper{W}_{\phi}\) as
\begin{equation}
\label{eq:covarinat-as-composition}
[\oper{W}_{\phi}f]( x, y, b, r) = [\left(M \circ \mathcal{F}_{2} \circ M_{b} \circ T \circ R\right)\,F ]( x, y),
\end{equation}
and obtain
\begin{prop}
\label{pr:unitarity-G-covariant}
For a fixed \(r_{0}\in\Space[+]{R}{}\) and \(b_{0}\in \Space{R}{}\), the map \(f \otimes \overline{\phi}\rightarrow [\oper{W}_{\phi} f](\cdot, \cdot, b_0, r_0)\) is a unitary operator from \(\FSpace{L}{2}(\Space{R}{})\otimes \FSpace{L}{2}(\Space{R}{})\) onto \( \FSpace{L}{2}(\Space{R}{2})\).
\end{prop}
Also, the induced covariant transform preserves the Schwartz space, that is, if \(f, \phi \in \mathcal{S}(\Space{R}{})\) then \(\oper{W}_{\phi}f ( \cdot, \cdot, b_0, r_0)\in \mathcal{S}(\Space{R}{2})\). This is because the \(\mathcal{S}(\Space{R}{2})\) is invariant under the all five above components of \(\oper{W}_{\phi}f\) in~\eqref{eq:covarinat-as-composition}.
Note, that the induced covariant transform~\eqref{eq:covariant-tr-schrodinger-type} does not define a square-integrable function on \(\Space{G}{}\slash Z\sim \Space[+]{R}{4}\). To discuss unitarity we need to introduce a suitable inner product. In general we can start from a probability measure \(\mu\) on \(\Space[+]{R}{2}\), that is \(\int_{\Space[+]{R}{2}} \,\rmd \mu =1\). Then we define the inner product
\begin{equation}
\label{scalar-product-mu}
\langle f, g \rangle_{\mu} = \int_{\Space[+]{R}{4}}{f ( x, y, b, r) \, \overline{ g( x, y, b, r)}\, \frac{\hbar \, \rmd x \, \rmd y \, \rmd \mu(b,r)}{\sqrt{2 r}},}
\end{equation}
for \(f, g \in \FSpace{L}{\phi}(\Space[+]{R}{4})\). The factor \(\hbar\) in the measure \(\textstyle\frac{\hbar \, \rmd x \, \rmd y}{\sqrt{2 r_{0}}}\) makes it dimensionless, see discussion of this in~\cites{Kisil02e,AlmalkiKisil18a}.
Important particular cases of probability measures parametrised by \((b_0, r_0)\in \Space[+]{R}{2}\) are
\begin{equation}
\label{eq:dirac-measure}
\rmd \mu_{(b_0,r_0)}(b,r) = \delta(b-b_0)\, \delta(r-r_0) \,\rmd b \,\rmd r \, ,
\end{equation}
where \(\delta(t)\) is the Dirac delta. The respective inner products becomes:
\begin{equation}
\label{scalar-product-b0-r0}
\langle f, g \rangle_{(b_0, r_0)} = \int_{\Space{R}{2}}f ( x, y, b_0, r_0) \, \overline{ g( x, y, b_0, r_0)}\, \frac{\hbar \, \rmd x \, \rmd y}{\sqrt{2 r_{0}}}\,.
\end{equation}
From now on we consider \(\FSpace{L}{\phi}(\Space[+]{R}{4})\) as a Hilbert space with the inner product~\eqref{scalar-product-mu} or specifically \eqref{scalar-product-b0-r0}. The respective norms are denoted by \(\| \cdot \|_{\mu}\) and \(\| \cdot \|_{(b_0, r_0)}\).
Using the above inner product, we can derive from Proposition~\ref{pr:unitarity-G-covariant} the following orthogonality relation:
\begin{cor}
Let \(f, g, \phi, \psi \in \FSpace{L}{2}(\Space{R}{})\), then
\begin{displaymath}
\langle \oper{W}_{\phi} f, \oper{W}_{\psi} g\rangle_{\mu}= \langle f, g \rangle \, \overline{\langle \phi, \psi \rangle },
\end{displaymath}
for any probability measure \(\mu\), in particular~\eqref{eq:dirac-measure} with fixed \((b_0, r_0)\in \Space[+]{R}{2}\).
\end{cor}
\begin{cor}
Let \(\phi \in \FSpace{L}{2}( \Space{R}{})\) have a unite norm. Then, the induced covariant transform \(\oper{W}_{\phi}\) is an isometry from \(\FSpace{L}{2}(\Space{R}{})\) to
\(\FSpace{L}{\phi}( \Space[+]{R}{4})\) and its inverse is given by the adjoint operator---contravariant transform:
\begin{equation}
\label{eq:contravariant-transform-mu}
f(u) = \int_{\Space[+]{R}{4}}F ( x, y, b, r) \, [\rho( \mathbf{s}(x, y, b, r)) \phi](u)\, \frac{\hbar \, \rmd x \, \rmd y \, \rmd \mu(b,r)}{\sqrt{2 r}},
\end{equation}
where \(F \in \FSpace{L}{\phi}( \Space[+]{R}{4})\). In particular:
\begin{equation}
\label{eq:contravariant-transform}
f(u) = \int_{\Space{R}{2}} F ( x, y, b_0, r_0) \, [\rho( \mathbf{s}(x, y, b_0, r_0)) \phi](u) \, \frac{\hbar \, \rmd x \, \rmd y}{\sqrt{2 r_{0}}}.
\end{equation}
\end{cor}
\begin{proof}
For \(f\in \FSpace{L}{2}(\Space{R}{})\), we have
\begin{displaymath}
\| f\|_{\FSpace{L}{2}(\Space{R}{})} = \| f \otimes \overline{\phi} \|_{\FSpace{L}{2}(\Space{R}{2})}= \| \oper{W}_{\phi}f\|_{\mu},
\end{displaymath}
which follows from isometry \( \FSpace{L}{2}(\Space{R}{2}) \rightarrow \FSpace{L}{2}(\Space{R}{2})\) in Prop.~\ref{pr:unitarity-G-covariant}. Then verification of formulae~\eqref{eq:contravariant-transform-mu}--\eqref{eq:contravariant-transform} is a technical exercise.
\end{proof}
A reader may note that~\eqref{eq:contravariant-transform} with \(\phi(u) = \sqrt[4]{2} \rme^{-\pi \hbar\, u^{2}} \) is essentially the inverse Fock--Segal--Bargmann transform.
\section{Image spaces of the covariant transforms}
\label{sec:image-space-covar-trans}
Clearly, not every function on \(\Space[+]{R}{4}\) is a covariant transform~\eqref{eq:covariant-tr-schrodinger-type} of a function from \(\FSpace{L}{2}(\Space{R}{})\). In this section we discuss the image space of the covariant transform.
\subsection{Right shifts and covariant transform}
Let \(R(g)\) be the right regular representation of the group \(G\), which acts on the functions defined in the group \(G\) as follows:
\begin{displaymath}
R(g): f( g')\rightarrow f( g' \, g), \qquad \text{ where } g\in G.
\end{displaymath}
In contrast to the intertwining property of the covariant transform for the left regular representation~\eqref{eq:covariant-intertwining}, the right shift satisfies the relation
\begin{align*}
R(g) [\oper{W}_{\phi}] (g') & =[\oper{W}_{\phi}]( g' \,g) \\
& = \langle f, \rho( g'\,g) \phi\rangle \\
& =\langle f ,\rho(g') \rho(g) \phi \rangle \\
& =[\oper{W}_{\rho (g)\,\phi} f](g').
\end{align*}
That is, the covariant transform intertwines the right shift with the action of \(\rho\) on the fiducial vector \(\phi\). Therefore, we obtain the following result, which plays an important r\^ole in exploring the nature of the image space of the covariant transform.
\begin{cor}
\textup{\cite{Kisil10c}}
\label{co:analyticity-fiducial}
Let \(G\) be a Lie group with a Lie algebra \(\algebra{g}\) and \(\rho\) be a representation of \(G\) in a Hilbert space \(\mathcal{H}\). Let a fiducial vector \(\phi\) be a null-solution, \(A\phi=0\), for the operator \(A= \sum_{j} a_{j}{d\rho}^{X_{j}}\), where \({d\rho}^{X_{j}}\) are the derived representation of some \(X_{j}\in \algebra{g}\) and \(a_{j}\) are constants. Then, for any \(f\in \mathcal{H}\) the wavelet transform \([W_{\phi}f](g)= \scalar{f}{\rho(g)\phi}\) satisfies
\begin{displaymath}
D(W_{\phi}f)=0, \quad \text{ where } \quad D= \sum_{j} \overline{a}_{j} \mathcal{L}^{X_{j}}.
\end{displaymath}
Here \(\mathcal{L}^{X_{j}}\) are the left invariant fields (Lie derivatives)~\eqref{eq:lie-derivative} on \(G\) corresponding to \(X_{j}\).
\end{cor}
Illustrative examples are the classical spaces of analytical functions: the Fock--Segal--Bargmann space and the Hardy space, see~\cites{Kisil11c,Kisil13c} for details.
\begin{rem}
\label{rem:analyticity-polynom}
It is straightforward to extend the result of Cor.~\ref{co:analyticity-fiducial} from a linear combination of elements in the Lie algebra \(\algebra{g}\) to an arbitrary polynomial from the enveloping algebra of \(\algebra{g}\) or even to more general functions/distributions, cf.~\amscite{Kisil11c}*{Cor.~5.8}.
\end{rem}
\subsection{Characterisation of the image space for the group $\Space{G}{}$}
The above Cor.~\ref{co:analyticity-fiducial} can be used to construct covariant transforms with desired properties through purposely selected fiducial vectors. We are illustrating this for the group \(\Space{G}{}\). First, we need to compute the Lie derivatives~\eqref{eq:lie-derivative} reduced to the representation space of the quasi-regular representation~\eqref{eq:quasi-regular}, see~\cite{Kisil21b}:
\begin{align}
\nonumber
\mathcal{L}^{X}&= r \partial_{x}, &
\mathcal{L}^{B}&= r^{2}\,\partial_{b},\\
\label{eq:lie-derivatives}
\mathcal{L}^{Y} &=\textstyle\frac{1}{r}( -2\pi \rmi \hbar x I- b\,\partial_{x} + \partial_{y}), &
\mathcal{L}^{R} &= r\, \partial_{r} ,\\
\nonumber
\mathcal{L}^{S}&= -2 \pi \rmi \hbar I.
\end{align}
One can check that those Lie derivatives make a representation of the Lie algebra of the group \(\Space{G}{}\)~\cite{Kisil21b}.
Now we are looking for a covariant transform \(\oper{W}_\phi: \FSpace{L}{2}(\Space{R}{}) \rightarrow \FSpace{L}{2}(\Space[+]{R}{4})\) with the image space annihilated by a generic linear combination of Lie derivatives~\eqref{eq:lie-derivatives}. To this end the fiducial vector \(\phi\) shall be a null solution of the following differential operator composed from the derived Schr\"o\-din\-ger type representation~\eqref{eq:schrodinger-derived}
\begin{equation}
\label{eq:generic-derivative}
\begin{split}
\lefteqn{ {\rmd\rho}^{\rmi E_s S + E_{x}X+ \rmi E_y Y+ \rmi E_{b} B+ E_{r} R}}\qquad&\\% &= E_s\, {d\rho}^{S} + E_{x}\, {d\rho}^{X} + \rmi E_y\, {d\rho}^{Y} + \rmi E_{b}\, {d\rho}^{B}+ E_{r}\,{d\rho}^{R}\\
&= (E_ru-\rmi E_y) \frac{\rmd\ }{\rmd u} +(\pi \hbar(E_b u^{2}+{2 \rmi } E_x u - 2 E_s)+\half E_{r} )I.
\end{split}
\end{equation}
where \(E_s\), \(E_x\), \(E_y\), \(E_b\) and \(E_r\) are arbitrary real coefficients.
This equation has two different solutions depending on a value of \(E_r\). If \(E_r=0\) (which requires \(E_y\neq 0\) for non-trivial operator~\eqref{eq:generic-derivative}) then a generic solution of~\eqref{eq:generic-derivative} is~\cite{Kisil21b}
\providecommand{\genphizero}{\phi_0}
\begin{equation}
\label{eq:fiducial-Er=0}
\genphizero (u)= C\,\exp\!\left(\pi \hbar\left( 2\rmi \frac{E_s}{E_y} u+ \frac{E_x}{E_y} u^2 - \rmi \frac{E_b}{3E_y} u^3 \right)\right)\,,
\end{equation}
where \(E_x< 0\) for square integrability of \(\genphizero\) and the constant \(C\) is determined from the normalisation condition \(\|\genphizero\|_2 = 1\). We have here a sort of Airy beam~\cite{BerryBalazs79a}, which was employed in~\cite{AlmalkiKisil19a} in the context of the share group, i.e. the absence of \({d\rho}^R\) in~\eqref{eq:generic-derivative}.
For \(E_r\neq 0\) we find the generic solution in the form~\cite{Kisil21b}:
\providecommand{\genphi}{\phi_1}
\begin{equation}
\label{eq:fiducial-Er-not-0}
\begin{split}
\genphi (u)&= C \, {(E_{r} u - \rmi E_y)^{- \frac{ 1}{2}+2
{ \pi \hbar E_s}/{E_r} -
{ \pi \hbar E_y( 2E_{x} E_{r}+ E_{b} E_y)}/{E_{r}^{3}}}}\\
&\qquad \times {\exp\left(-\pi \hbar\left(\frac{\rmi( 2E_{x}E_{r}+ E_{b} E_y)} {E_{r}^{2}} u +\frac{ E_{b}}{2 E_{r}} {u}^{2}\right)\right)}
.
\end{split}
\end{equation}
Again, for \(\genphi\in \FSpace{L}{2}(\Space{R}{})\) we need \(\frac{\hbar E_{b}}{E_{r}}>0\) and a proper normalising constant \(C\).
A detailed study of all arising covariant transforms is still awaiting further work. Here we concentrate on some special aspects which appear in this extended group setting for the most traditional fiducial vector---the Gaussian. First, we note that it steams from both solutions~\eqref{eq:fiducial-Er=0} and~\eqref{eq:fiducial-Er-not-0}:
\begin{itemize}
\item For \(E_r=0\) letting \(E_s=E_b=0\), \(E_x=-1\) and \(E_y=1\) with \(C=\sqrt[4]{2}\) in \(\genphizero\)~\eqref{eq:fiducial-Er=0} produces
\begin{equation}
\label{eq:gaussin-first}
\phi(u) = \sqrt[4]{2} \rme^{-\pi \hbar\, u^{2}} \quad \text{ with the identity } \quad {\rmd \rho}^{ -X+ \rmi Y}\, \phi = 0,
\end{equation}
i.e. \(\phi\) is annihilated by the Heisenberg group part of \(\Space{G}{}\).
\item For \(E_r=1\) substitution of \(E_s= \frac{1}{4 \pi \hbar }\), \(E_{x}=E_y=0\) and \(E_{b}=2\) with \(C=\sqrt[4]{2}\) into the vacuum \(\genphi\)~\(\genphi\)~\eqref{eq:fiducial-Er-not-0} again produces
\begin{equation}
\label{eq:gaussian-second}
\phi( u)= \sqrt[4]{2} \rme^{-\pi \hbar\, u^{2}} \quad \text{ with the identity } \quad {\rmd \rho}^{\rmi/(4\pi \hbar) S+ 2\rmi B+ R}\, \phi = 0,
\end{equation}
i.e. \(\phi\) is also annihilated by the affine group part of \(\Space{G}{}\).
\end{itemize}
Let us introduce the covariant transform \(\oper{W}_\phi: \FSpace{L}{2}(\Space{R}{}) \rightarrow \FSpace{L}{2}(\Space[+]{R}{4})\)~\eqref{eq:covariant-tr-schrodinger-type} with the fiducial vector \(\phi\)~\eqref{eq:gaussin-first}--\eqref{eq:gaussian-second}:
\begin{equation}
\label{eq:metamorphism}
\begin{split}
\lefteqn{
[\oper{W}_\phi
f](x,y,b,r)
=\sqrt{r} \int_{\Space{R}{}}{ f ( u)\, \rme^{-2 \pi \rmi \hbar (x ( u -y) - b (u - y)^{2}/2)}\, \overline\phi ( r( u- y))} \, \rmd u} \qquad &\\
&=\sqrt[4]{2r^2} \int_{\Space{R}{}} f ( u)\, \rme^{-2 \pi \rmi \hbar (x ( u -y) - b (u - y)^{2}/2)}\, \rme^{ -\pi \hbar r^2( u- y)^2} \, \rmd u\\
&= \sqrt[4]{2r^2}
\int_{\Space{R}{}} f(u)\,
\exp\left( -\pi \myhbar \left(
( r^{2}-\rmi b)(u-y)^2
+2 \rmi (u-y) x
\right)\right)
\,\rmd
\,.
\end{split}
\end{equation}
It was introduced in~\cites{Kisil21c,Kisil21b} and called metamorphism. We are also using the notation \( \metamorph{ f} \coloneqq \oper{W}_\phi f\) from~\cite{Kisil21c} which now can be explained as the double covariant transform for the Heisenberg and the affine groups simultaneously.
The image space \(\FSpace{L}{\phi}(\Space[+]{R}{4})\) of the metamorphism is a subspace of square-integrable functions on \(\FSpace{L}{2}(\Space[+]{R}{4}, \|\cdot\|_{\mu})\), see~\eqref{scalar-product-mu}.
\begin{rem}
Another feature of the Gaussian as a fiducial vector is that an extension of the group \(\Space{G}{}\) to the full Schr\"o\-din\-ger group does not add a value. Indeed, the Iwasawa decomposition \(\SL=ANK\)~\citelist{\amscite{Lang85}*{\textsection{}III.1} \amscite{Kisil12a}*{\textsection{}1.1}} represents \(\SL\) as the product of the affine subgroup \(AN\) and the compact subgroup \(K\). Yet, the Gaussian is invariant under the action of the phase-space rotations produced by \(K\). Thus, we get the same set of coherent states from the actions of the group \(\Space{G}{}\) and the Schr\"o\-din\-ger group.
\end{rem}
From the annihilation property~\eqref{eq:gaussin-first} by the derived representation \({d\rho}^{ -X+ \rmi Y}\) and Cor.~\ref{co:analyticity-fiducial} we conclude that \(\mathcal{L}^{-X - \rmi Y}\, \metamorph{ f} =0\) for any \(f\). Using~\eqref{eq:lie-derivatives} we find~\cite{Kisil21b}:
\begin{equation}
\label{eq:Cauchy-first}
\begin{split}
\oper{C}_1 &= -\mathcal{L}^{X} - \rmi\mathcal{L}^{Y}\\
&= \frac{1}{r}\left({(r^{2} - \rmi b)}\, \partial_x+ \rmi \, \partial_y+2 x \myhbar \pi\, I\right).
\end{split}
\end{equation}
The operator \(\oper{C}_1\) is called the first Cauchy--Riemann type operator. Similarly from~\eqref{eq:gaussin-first} we conclude that \(\oper{C}_2 \metamorph{ f} =0\) for the second Cauchy--Riemann type operator~\cite{Kisil21b}:
\begin{equation}
\begin{split}
\label{eq:Cauchy-second}
\oper{C}_2 & = -\frac{\rmi}{4 \pi \hbar }\mathcal{L}^{S} -2 \rmi \mathcal{L}^{B} + \mathcal{L}^{R} \\
&= 2 r^{2}\, \partial_b +\rmi r\, \partial_r - \half \rmi\, I\,.
\end{split}
\end{equation}
It is convenient to view operators \(\oper{C}_1\) and \(\oper{C}_2\) as the Cauchy--Riemann operators for the following complexified variables:
\begin{equation}
\label{eq:complex-variables}
w=b + \rmi r^2 \quad \text{ and } \quad z=x+(b + \rmi r^2)y = x+wy \,.
\end{equation}
\begin{rem}
As was pointed out in~\cite{Kisil13c}, the analyticity conditions~\eqref{eq:Cauchy-first}--\eqref{eq:Cauchy-second} are consequences of minimal uncertainty properties of the fiducial vector. The first condition~\eqref{eq:Cauchy-first} follows from the celebrated Heisenberg--Kennard uncertainty relation~\cites{Folland89,Kisil13c}
\begin{displaymath}
\Delta_\phi(M) \cdot \Delta_\phi(D) \geq \frac{\myh}{2}
\end{displaymath}
for the coordinate \(M=\rmd{\rho}^{\rmi X}\) and momentum \(D=\rmd{\rho}^{\rmi Y}\) observables in the Schr\"o\-din\-ger representation~\eqref{eq:schrodinger-derived}. The second condition~\eqref{eq:Cauchy-second} is due to the similar minimal joint uncertainty of the Gaussian state for the Euler operator \({\rmd \rho}^{1/(4\pi \hbar) S-\rmi R}=-\rmi u \, \frac{\rmd\ }{\rmd u}\) and the quadratic potential \({\rmd \rho}^{\rmi B}= \pi \hbar u^{2} I\).
\end{rem}
Besides the two operators \(\oper{C}_1\) and \(\oper{C}_2\) which are based on the special properties~\eqref{eq:gaussin-first}--\eqref{eq:gaussian-second} of the Gaussian we can note a couple of polynomial identities in the Schr\"o\-din\-ger type representation of the Lie algebra \(\algebra{g}\). Indeed, using~\eqref{eq:schrodinger-derived} one can check:
\begin{equation}
\label{eq:second-order-rep-zero}
\left(\rmd{\rho}^{X}\right)^2+2\, \rmd {\rho}^{S}\, \rmd {\rho}^{B} =0, \quad \text{ and } \quad
\rmd{\rho}^{X}\, \rmd {\rho}^{Y}+d{\rho}^{Y}\, \rmd {\rho}^{X} + 2\, \rmd {\rho}^{S}\, \rmd {\rho}^{R} =0.
\end{equation}
These relations express the affine subalgebra generators \(B\) and \(R\) through the Heisenberg ones \(X\) and \(Y\). That is related to so-called quadratic algebra concept~\amscite{Gazeau09a}*{\textsection{}2.2.4}. Because operators in~\eqref{eq:second-order-rep-zero} annihilate any function, including the fiducial vector of the metamorphism, Rem.~\ref{rem:analyticity-polynom} implies that the image space \(\FSpace{L}{\phi}(\Space[+]{R}{4})\) is annihilated by by the second-order differential operators~\cite{Kisil21b}:
\begin{align}
\label{eq:strctural-1}
\oper{S}_1 &=\left(\mathcal{L}^{X}\right)^2+2\,\mathcal{L}^{S}\,\mathcal{L}^{B} = r^{2}( 4 \pi \rmi \myhbar \, \partial_{b} - \partial_{xx}^{2}) \,;
\intertext{and}
\label{eq:strctural-2}
\begin{split}
\oper{S}_2 &= \mathcal{L}^{X}\,\mathcal{L}^{Y}+\mathcal{L}^{Y}\,\mathcal{L}^{X} + 2\,\mathcal{L}^{S}\,\mathcal{L}^{R}\\
& = -4 \pi \rmi r \myhbar \, \partial_{r}- 2b \, \partial_{xx}^{2}+2\, \partial_{xy}^{2} - 4 \pi \rmi x \myhbar \, \partial_{x}-2 \pi \rmi \myhbar\, I\,.
\end{split}
\end{align}
The identities \(\oper{S}_1\metamorph{ f} = 0 \) and \(\oper{S}_2\metamorph{ f} = 0 \) are called the first and second structural conditions.
Their presence is a notable difference between covariant transforms on the group \(\Space{G}{}\) and the Heisenberg group.
Of course, the list of annihilators is not exhausting and the above conditions are not independent. If \( \oper{S}_1 F =0\) for a function \(F\) satisfying both the Cauchy--Riemann-type operators~\eqref{eq:Cauchy-first}--\eqref{eq:Cauchy-second} then the function \(F\) has to be in the kernel of \(\oper{S}_2\)~\eqref{eq:strctural-2} as well, see~\cite{Kisil21b}.
It was shown~\cite{Kisil21c} that a generic solution of two differential operators~\eqref{eq:Cauchy-first}--\eqref{eq:Cauchy-second} is:
\begin{equation}
\label{eq:first-order-generic-complex}
[\oper{G}f_2](z,w) \coloneqq
\sqrt{r}\, \rme^{-\pi i \myhbar x^2/w}\, f_2 (z, w)\,,
\end{equation}
where \(f_2\) is a holomorphic function of two complex variables \(z\) and \(w\)~\eqref{eq:complex-variables}. Additionally, the structural condition \(\oper{S}_1\)~\eqref{eq:strctural-1} applied to \(\oper{G}f_2\)~\eqref{eq:first-order-generic-complex} produces a parabolic equation for \(f_2\):
\begin{equation}
\label{eq:strctural-generic}
4\pi i h w\partial_wf_2(z,w)- w\partial_{zz}^2 f_2(z,w)+4\pi i h z\partial_zf_2(z,w) +2\pi i h f_2(z,w) =0\,.
\end{equation}
which is equivalent through a change of variables~\amscite{PolyaninNazaikinskii16a}*{3.8.3.4}:
\begin{displaymath}
(z,w,f_2)\rightarrow \left (\frac{z}{ w}, \frac{1}{w}, \frac{1}{\sqrt{w}} f_2\right) .
\end{displaymath}
to the free particle Schr\"o\-din\-ger equation.
The above discussion allows us to characterise the image space of the metamorphism:
\begin{thm}
\label{th:metamorphism-image-characterisation}
A function \(F(x,y,b,r)\) on \(\Space[+]{R}{4}\) is the metamorphism~\eqref{eq:metamorphism} of a function \(f\in \FSpace{L}{2}(\Space{R}{})\) if and if only \(F\) satisfies to the following conditions:
\begin{enumerate}
\item
\(F(x,y,b,r)\) is annihilated by operators \(\oper{C}_1\)~\eqref{eq:Cauchy-first}, \(\oper{C}_2\)~\eqref{eq:Cauchy-second} and \(\oper{S}_1\)~\eqref{eq:strctural-1}.
\item
\(F(\cdot,\cdot, b_0,r_0)\) is square-integrable in the sense of the inner product \(\langle \cdot, \cdot \rangle_{(b_0, r_0)}\)~\eqref{scalar-product-b0-r0} for some \((b_0,r_0)\in\Space[+]{R}{2}\).
\end{enumerate}
\end{thm}
\begin{proof}
The necessity of both conditions was discussed above. For sufficiency, let a function \(G\) is annihilated by \(\oper{C}_1\), \(\oper{C}_2\) and \(\oper{S}_1\). If \(G(x,y, b_0,r_0) = 0\) for all \((x,y) \in \Space{R}{2}\) then the initial value problem for the parabolic equation~\eqref{eq:strctural-generic} implies \(G \equiv 0\) on \(\Space[+]{R}{4}\).
Now, based on the square-integrability of \(F\) we use contravariant transform expression~\eqref{eq:contravariant-transform} to construct a function \(f \in \FSpace{L}{2}(\Space{R}{})\).
Then \(G\coloneqq F - \metamorph{f}\) is annihilated by operators \(\oper{C}_1\), \(\oper{C}_2\), \(\oper{S}_1\) and \(G(x,y, b_0, r_0) = 0\) for all \((x,y) \in \Space{R}{2}\). Thus \(G\equiv 0\) (as explained above) and therefore \(F = \metamorph{f}\) on \(\Space[+]{R}{4}\).
\end{proof}
Although the Gaussian and the metamorphism based on it are genuinely remarkable in many respects, other covariant transforms~\eqref{eq:covariant-tr-schrodinger-type} with fiducial vectors~\eqref{eq:fiducial-Er=0} and~\eqref{eq:fiducial-Er-not-0} desrve further attention as well.
\medskip
\section*{Acknowledgments}
\label{sec:acknowledgments}
Authors are grateful to Prof.~Alexey Bolsinov for useful comments and suggestions on this work. The first named author was sponsored by the Albaha University (SA).
\small
|
\section{Introduction}
High-precision experiments with polarized or unpolarized electron beams \cite{Gr20,Au18} require an accurate knowledge of additional multiple photon processes which modify the Coulombic scattering
cross section.
To these belong, besides dipersion, the vacuum polarization and the vertex correction, renormalized by the self-energy and made infrared
finite by the soft bremsstrahlung.
For vacuum polarization it is well-known that the addition of the Uehling potential to the Coulombic target field $V_T$, which arises from the nuclear charge distribution, provides a nonperturbative consideration of this quantum electrodynamical (QED) effect \cite{Ue35,SM88}.
It is the first nonvanishing term in the decomposition of the vacuum loop in powers of $V_T$ \cite{Sh00}.
Indeed, if the Uehling potential were treated to first order \cite{Jaku21}, the respective cross-section modification would agree with Tsai's result \cite{T60,MT00} from the first-order Born approximation.
However, the deviations from this Born prediction are, even for the $^{12}$C nucleus, formidable
in the vicinity of a diffractive cross-section minimum \cite{Jaku21b}.
The relation between the first-order Born amplitude and the underlying potential was recently applied in the context of the contribution to the beam-normal spin asymmetry,
also known as Sherman function \cite{Mo64}, which results from dispersion.
In their method, Koshchii et al \cite{Ko21} constructed an absorptive potential from the respective Born amplitude,
to be included in the Dirac equation for the electronic scattering states, in order to provide a nonperturbative representation of the dispersive spin asymmetry.
In the present work this procedure is adopted for generating a potential $V_{\rm vs}$ for the vertex and self-energy (vs) correction from the respective first-order Born amplitude.
Apart from the nonperturbative treatment of the cross-section modifications induced by adding $V_{\rm vs}$ to $V_T$ in the Dirac equation,
this allows for a
consistent estimate of the respective changes in the spin asymmetry.
By considering a light ($^{12}$C) and a heavy ($^{208}$Pb) target nucleus and electrons with energies between 1 MeV and 240 MeV,
the QED corrections and their dependence on the Coulomb distortion are investigated in a large region of momentum transfers.
The paper is organized as follows. In section 2 the vs potential is derived. Results for the radiative modifications of the differential cross section and the spin asymmetry are provided in section 3 for the two target nuclei.
Concluding remarks follow (section 4).
Atomic units ($\hbar=m=e=1$) are used unless indicated otherwise.
\section{Theory}
In the Born approximation, the differential cross section for the elastic scattering of an electron into the solid angle $d\Omega_f$, which includes the radiative corrections to lowest order \cite{Lan}, is given by
$$\frac{d\sigma^{B1}}{d\Omega_f}=\frac{k_f}{k_i}\;\frac{1}{f_{\rm rec}}\;\sum_{\sigma_f}[ \;|A_{fi}^{B1}|^2$$
\begin{equation}\label{2.1}
\;+\;2\mbox{ Re}\,\{A_{fi}^{\ast B_1}\,(A_{fi}^{\rm vac} + {A}_{fi}^{\rm vs}+A_{fi}^{\rm box})\} \;
+\; \frac{d\sigma^{\rm soft}}{d\Omega_f}],
\end{equation}
where it is summed over the final spin polarization $\sigma_f$ of the electron.
$A_{fi}^{B1}$ ist the first-order Born amplitude for potential scattering in the Coulombic target field $V_T$,
and $A_{fi}^{\rm vac}$ and $A_{fi}^{\rm box}$ are the lowest-order amplitudes for vacuum polarization \cite{MT00} and dispersion \cite{Sch55,Le56,FR74}, respectively.
Recoil effects are
considered by the
pre\-factor $f_{\rm rec}^{-1}$ \cite{Jaku21b}.
Here and in the following $k_i$ and $k_f$ denote the moduli of the initial and final electron momenta $\mbox{\boldmath $k$}_i$ and $\mbox{\boldmath $k$}_f$, respectively.
The lowest-order Born amplitude for the vertex correction, after eliminating the UV divergence by renormalizing with the help of the self energy, is given by \cite{BS19,Va00}
$$A_{fi}^{\rm vs}\;=\; F_1^{\rm vs}(-q^2)\;A_{fi}^{B1},$$
$$F_1^{\rm vs}(-q^2)\;=\;\frac{1}{2\pi c}\left[ \frac{v^2+1}{4v}\left(\ln \,\frac{v+1}{v-1}\right)\left( \ln\,\frac{v^2-1}{4v^2}\right)\right.$$
\begin{equation}\label{2.2}
+\;\frac{2v^2+1}{2v} \,\ln\,\frac{v+1}{v-1}
-\,2\,+\;\frac{v^2+1}{2v}\left\{\mbox{Li }\left(\frac{v+1}{2v}\right)\right.
\end{equation}
$$ \left. \left.\;-\;\mbox{Li }\left(\frac{v-1}{2v}\right)\right\}\right]\;+\;\mbox{IR},$$
where
$q^2=(E_i-E_f)^2/c^2-\mbox{\boldmath $q$}^2$, with $\mbox{\boldmath $q$}=\mbox{\boldmath $k$}_i-\mbox{\boldmath $k$}_f$, is the squared 4-momentum transfer to the nucleus.
$E_i$ and $E_f$ are the initial, respectively final, total energies of the scattering electron.
Moreover,
$v=\sqrt{1-4c^2/q^2}$ and Li$(x)=-\int_0^x dt \frac{\ln|1-t|}{t}\;$ is the Spence function \cite{T61,Va00}.
IR denotes the infrared divergent term. There is also a tiny magnetic contribution to $A_{fi}^{\rm vs}$ \cite{BS19}, which is omitted here.
The differential cross section for the soft bremsstrahlung reads in Born approximation
\begin{equation}\label{2.3}
\frac{d\sigma^{\rm soft}}{d\Omega_f}\;=\;W_{fi}^{\rm soft}\;|A_{fi}^{B1}|^2
\end{equation}
with (correcting printing errors in \cite{MT00} and \cite{BS19})
$$ W_{fi}^{\rm soft}\,=\,-\frac{1}{\pi c}\left\{ 2 \,\ln\frac{2\omega_0}{c^2}\,+\,\frac{E_i}{k_ic}\,\ln\frac{c^2}{E_i+k_ic}\right. $$
$$\,+\,\frac{E_f}{k_fc}\,\ln \frac{c^2}{E_f+k_fc}
\,-\left[ 2\,\left(\ln \frac{2\omega_0}{c^2}\right)
\frac{v^2+1}{2v}\,\ln \frac{v+1}{v-1}\right.$$
$$+\,c^4\beta \,\frac{1-q^2/(2c^2)}{\zeta (\beta E_i-E_f)}\left( \frac{1}{4}\left( \ln \frac{E_i-k_ic}{E_i+k_ic}\right)^2 \!\!
-\,\frac{1}{4}\left( \ln \frac{E_f-k_fc}{E_f+k_fc}\right)^2 \right.$$
\begin{equation}\label{2.4}
\,+\,\mbox{Li}(1\,-\,\beta \,\frac{E_i-k_ic}{\zeta})
-\,\mbox{Li}(1\,-\,\frac{E_f-k_fc}{\zeta})
\end{equation}
$$ \left.\left. \left.\,+\,\mbox{Li}(1\,-\,\beta\,\frac{E_i+k_ic}{\zeta})\,-\,\mbox{Li}(1\,-\,\frac{E_f+k_fc}{\zeta})\right)\right] \right\}-\,2 \mbox{ IR},$$
introducing the cutoff frequency $\omega_0$ of the soft photons and the abbreviations
$$\beta\,=\,1\,-\,\frac{q^2}{2c^2}\,+\,\sqrt{-\frac{q^2}{c^2}\left( 1\,-\,\frac{q^2}{4c^2}\right)}$$
\begin{equation}\label{2.5}
\zeta\,=\,c^4\left[ \beta \left( 1\,-\,\frac{q^2}{2c^2}\right) -1\right]\frac{1}{\beta E_i-E_f}.
\end{equation}
The validity of (\ref{2.4}) for $W_{fi}^{\rm soft}$ is subject to the requirement that $\omega_0$ is not too small ($\frac{1}{\pi c} |\ln \frac{\omega_0}{c^2}|\ll 1$ \cite{Sh00}).
Due to mutual cancellations in (\ref{2.4}), a very large integration step number
for the Spence functions is necessary (some 50 000,
increasing with energy and angle).
For $-q^2/c^2\gtrsim 100$, the much simpler asymptotic formula for $W_{fi}^{\rm soft}$ can be used, as e.g. given in \cite{MT00} or \cite{Jaku21}.
Hard bremsstrahlung is disregarded in (\ref{2.1}), since it is assumed that the resolution $\Delta E$ of the electron detector
(which defines the upper limit of the photon frequency by $\omega_0=\Delta E$) is at most 1 MeV.
There is a simple connection between the first-order Born amplitude and the potential by which it is generated.
This is exemplified for the scattering amplitude $A_{fi}^{B1}$ which can be represented in terms of the nuclear
charge form factor $F_L(|\mbox{\boldmath $q$}|)$ \cite{BD64},
\begin{equation}\label{2.6}
A_{fi}^{B1}(\mbox{\boldmath $q$}) \,=\, -\,\frac{2\sqrt{E_iE_f}}{c^2}\;\frac{Z}{\mbox{\boldmath $q$}^2}\;\left( u_{k_f}^{(\sigma_f)+}\;u_{k_i}^{(\sigma_i)}\right)\,F_L(|\mbox{\boldmath $q$}|),
\end{equation}
where $Z$ is the nuclear charge number and $u_{k_i}^{(\sigma_i)}, u_{k_f}^{(\sigma_f)}$ are, respectively, the free 4-spinors of the initial and final electronic states to the spin polarization $\sigma_i, \sigma_f$.
In turn, the form factor is related to the Fourier transform of the target potential $V_T$,
\begin{equation}\label{2.7}
F_L(|\mbox{\boldmath $q$}|)\,=\,-\,\frac{\mbox{\boldmath $q$}^2}{4\pi Z}\int d\mbox{\boldmath $r$}\;e^{i\mbox{{\scriptsize \boldmath $q$}} \mbox{{\scriptsize \boldmath $r$}}} \,V_T(r).
\end{equation}
This provides us with the basic relation between the
potential and the first-order Born amplitude,
$$V_T(r)\;=\;\frac{1}{(2\pi)^3}\int d\mbox{\boldmath $q$}\;e^{-i\mbox{{\scriptsize \boldmath $q$}}\mbox{{\scriptsize \boldmath $r$}}}\;A_{fi}^{B1}(\mbox{\boldmath $q$})/A_0,$$
\begin{equation}\label{2.8}
A_0\;=\;\frac{\sqrt{E_iE_f}}{2\pi c^2}\;\left( u_{k_f}^{(\sigma_f)+}\;u_{k_i}^{(\sigma_i)}\right).
\end{equation}
For the construction of a nonperturbative theory, the IR contributions in (\ref{2.2}) and (\ref{2.4}) are omitted because it is known that they
cancel to all orders \cite{T61,Ye61}.
In order to derive the potential $V_{\rm vs}$ for the vertex and self-energy process, use is made of the
proportionality (\ref{2.2}) of its amplitude $A_{fi}^{\rm vs}$ to the scattering amplitude $A_{fi}^{B1}$.
Hence the application of (\ref{2.8}) yields
$$V_{\rm vs}(r)\;=\;\frac{1}{(2\pi)^3}\int d\mbox{\boldmath $q$}\;e^{-i\mbox{{\scriptsize \boldmath $q$}} \mbox{{\scriptsize \boldmath $r$}}}\;A_{fi}^{\rm vs}/A_0$$
\begin{equation}\label{2.9}
\approx\,-\,\frac{2Z}{\pi} \int_0^\infty d|\mbox{\boldmath $q$}|\;\frac{\sin(|\mbox{\boldmath $q$}|r)}{|\mbox{\boldmath $q$}|\,r}\;F_L(|\mbox{\boldmath $q$}|)\;F_1^{\rm vs}(-q^2).
\end{equation}
When performing the angular integration, the weak dependence of $F_1^{\rm vs}$ on $E_i-E_f$ (and hence on the scattering angle $\vartheta_f$) by means of recoil has been disregarded.
For a nonperturbative consideration of vacuum polarization and the vs correction, the Dirac equation
with the additional potentials is solved,
\begin{equation}\label{2.10}
\left[ -ic\mbox{\boldmath $\alpha$} \mbox{\boldmath $\nabla$} + \gamma_0 c^2 + V_T(r) +U_e(r)+V_{\rm vs}(r) \right]\psi(\mbox{\boldmath $r$})=E\,\psi(\mbox{\boldmath $r$}),
\end{equation}
where $U_e$ is the Uehling potential \cite{Kl77}
and $\mbox{\boldmath $\alpha$}, \gamma_0$ refer to Dirac matrices.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig1.pdf}
\vspace{-0.5cm}
\caption
{
Differential cross section $\frac{d\sigma_{\rm coul}}{d\Omega_f}$ as a function of scattering angle $\vartheta_f$ for electrons of 56 MeV $(----$), 150.2 MeV $(-\cdot -\cdot -$) and 238.1 MeV (-------------) colliding with $^{12}$C. Also shown are the experimental data by Reuter et al ($\blacksquare$, \cite{Reu82}) at 150.2 MeV and by Offermann et al ($\blacklozenge$, \cite{Off91}) at 238.1 MeV.
}
\end{figure}
\section{Results}
\setcounter{equation}{0}
The Coulombic target potentials of $^{12}$C and $^{208}$Pb are generated from the Fourier-Bessel representation of the respective ground-state charge densities \cite{VJ}.
The electronic scattering state $\psi$ is expanded in terms of partial waves which, together with their phase shifts, are determined with the help of
the Fortran code RADIAL of Salvat et al \cite{Sal}.
Since the two additional potentials $U_e$ and $V_{\rm vs}$ are of long range
(as compared to the nuclear radius), they require matching points between the inner and outer radial
solutions of the Dirac equation of the order of 2000 fm.
The determination of the scattering amplitude involves weighted summations of the phase shifts \cite{Lan}, which are performed with the help of a threefold convergence acceleration \cite{YRW}.
In order to minimize the difference between the nonperturbative and the Born QED results, (\ref{2.1}) is in the actual calculations modified by including the Coulomb distortion throughout, as suggested by Maximon \cite{Ma69}.
This is done by replacing the Born amplitude $A_{fi}^{B1}$ by the exact Coulomb amplitude $f_{\rm coul}$, indicated in the replacement of $A_{fi}^{\rm vac},A_{fi}^{\rm vs}, d\sigma^{\rm soft}/d\Omega_f$ by $\tilde{A}_{fi}^{\rm vac}, \tilde{A}_{fi}^{\rm vs}, d\tilde{\sigma}^{\rm soft}/d\Omega_f$,
and is leading to
$$\frac{d\sigma^{B1-C}}{d\Omega_f} \,=\,\frac{k_f}{k_i}\,\frac{1}{f_{\rm rec}} \;\sum_{\sigma_f} [\;|f_{\rm coul}|^2 $$
\begin{equation}\label{3.1}
+2\mbox{ Re}\, \{ f_{\rm coul}^\ast ( \tilde{A}_{fi}^{\rm vac}+\tilde{A}_{fi}^{\rm vs}+A_{fi}^{\rm box}) \}+\frac{d\tilde{\sigma}^{\rm soft}}{d\Omega_f}].
\end{equation}
Hence,
noting that $A_{fi}^{\rm vac}$ is, like $A_{fi}^{\rm vs}$, proportional to $A_{fi}^{B1}$ and
disregarding dispersion, the expression on the rhs of (\ref{3.1}) is proportional to the Coulombic cross section,
\begin{equation}\label{3.2}
\frac{d\sigma_{\rm coul}}{d\Omega_f}\,=\,\frac{k_f}{k_i}\,\frac{1}{f_{\rm rec}} \sum_{\sigma_f} |f_{\rm coul}|^2.
\end{equation}
The scattering amplitude $f_{\rm coul}$ is obtained from the phase-shift analysis
relating to the potential $V_T$ \cite{Lan}.
Recoil is included in the phase-shift analysis in terms of a reduced collision energy $\sqrt{(E_i-c^2)(E_f-c^2)}$, in a similar way as done for excitation \cite{MG64}.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig2a.pdf}
\vspace{-1.5cm}
\vspace{-0.5cm}
\includegraphics[width=11cm]{qedvertfig2b.pdf}
\caption
{
Cross section change $\Delta \sigma^C$ in (a) 56 MeV and (b) 150 MeV $e+^{12}$C collisions as a function of scattering angle $\vartheta_f$.
Shown are the results for vacuum polarization $(-\cdot -\cdot -$), vertex and self-energy correction $(----$) and the consideration of both (----------, thin line), as well as the additional inclusion of the soft-bremsstrahlung contribution for $\omega_0=1$ MeV (------------, thick line).
Included are the Born results $\Delta \tilde{\sigma}^{\rm vac}$ for vacuum polarization and $\Delta \tilde{\sigma}^{\rm vs}$ for the vs correction $(\cdots\cdots$).
In Fig.2b, the crosses mark the sum of the combined QED corrections and dispersion \cite{Jaku22}.
}
\end{figure}
On the other hand, the nonperturbative treatment of vacuum polarization and the vs process (leading to the scattering amplitude $f_{\rm vac+vs}$) results in the following expression for the differential cross section,
$$
\frac{d\sigma^{C}}{d\Omega_f} \,=\,\frac{k_f}{k_i} \,\frac{1}{f_{\rm rec}} \sum_{\sigma_f}\left[ \,|f_{\rm vac+vs}|^2\,+\,2\mbox{ Re } \{ f_{\rm coul}^\ast \,A_{fi}^{\rm box}\}\right. $$
\begin{equation}\label{3.3}
\left. +\; W_{fi}^{\rm soft}\;|f_{\rm vac+vs}|^2\right].
\end{equation}
In this prescription of the soft-photon cross section the fact has been accounted for that the cross section for emitting an additional soft photon during a certain scattering process
is given by the cross section for this scattering process times a factor which describes the attachment of one soft-photon line to the
respective diagram \cite{We65}.
This factorization holds as long as the scattering process is undisturbed by this photon emission.
In particular, the photon energy has to be sufficiently low ($\omega_0 \ll E_i-c^2)$ and the change $\delta |\mbox{\boldmath $q$}|/|\mbox{\boldmath $q$}|$ of momentum transfer sufficiently small.
This restricts the scattering angle by means of \cite{Lan,Lo58}
\begin{equation}\label{3.4}
\sin \frac{\vartheta_f}{2}\;\gg\;\frac{\omega_0c^4}{4\,E_i^3}.
\end{equation}
For the present cases of interest, both conditions are well satisfied. In particular, one has for $\omega_0 \lesssim 1$ MeV and
$E_i\gtrsim 50$ MeV the condition $\vartheta_f \gtrsim 1^\circ$, or for $E_i-c^2 \gtrsim 1 $ MeV and an energy resolution of at most 1\% the requirement of $\vartheta_f \gtrsim 10^\circ$.
By using the Born factor $W_{fi}^{\rm soft}$ in (\ref{3.3}) the approximation is made that this soft-photon line corresponds to a free electron, in the same spirit as in the second-order Born representation of dispersion.
The effect of the QED and dispersion processes is illustrated by considering the cross-section change, defined
with respect to the Coulombic cross section,
\begin{equation}\label{3.5}
\Delta \sigma \;=\;\frac{d\sigma/d\Omega_f}{d\sigma_{\rm coul}/d\Omega_f}\;-\,1,
\end{equation}
where in the two cross sections an additional averaging over the initial spin polarizaton $\sigma_i$ has to be made.
The cross section changes from the individual radiative processes are additive, such that
\begin{equation}\label{3.6}
\Delta \sigma^{B1-C} \,=\,\Delta \tilde{\sigma}^{\rm vac}
\,+\,\Delta \tilde{\sigma}^{\rm vs}\,+\,\Delta \sigma^{\rm box}\,+\,\Delta\tilde{\sigma}^{\rm soft},
\end{equation}
and
\begin{equation}\label{3.7}
\Delta \sigma^C\;=\;\Delta \sigma^{\rm vac+vs}\,+\,\Delta \sigma^{\rm box} \,+\,\Delta \sigma^{\rm soft},
\end{equation}
where the summands in (\ref{3.6}) and (\ref{3.7}) correspond to the contributions to $\Delta\sigma$
from the individual terms in (\ref{3.1}) and (\ref{3.3}), respectively.
It should be noted that $\Delta \sigma^{B1-C}$ --
without dispersion -- is approximately target-independent, since the Coulombic cross section drops out and recoil effects in vacuum polarization, $F_1^{\rm vs}$ and $W_{fi}^{\rm soft}$ are small.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig3a.pdf}
\vspace{-1.5cm}
\vspace{-0.5cm}
\includegraphics[width=11cm]{qedvertfig3b.pdf}
\caption
{
Cross section change $\Delta \sigma^C$ in 238.1 MeV $e+^{12}$C collisions as a function of scattering angle $\vartheta_f$.
In (a), the lines have the same meaning as in Fig.2. In (b), $\Delta \sigma^{\rm box}$ calculated with the three transiently excited states ($----$) and with the Friar-Rosen theory $(\cdots\cdots$) are shown, as well as $\Delta\sigma^{\rm box} + [\Delta\sigma^{\rm vac+vs} - (\Delta \tilde{\sigma}^{\rm vac} + \Delta \tilde{\sigma}^{\rm vs})]$ (-----------, without closure, $-\cdot -\cdot -$ with the Friar-Rosen theory for $\Delta\sigma^{\rm box}$).
Included is the relative deviation of the experimental cross section from the Coulombic result ($\bullet$, connected by lines \cite{Off91}).
}
\end{figure}
\subsection{The $^{12}$C nucleus}
The angular distribution of the Coulombic cross section is shown in Fig.1 for the collision energies 56, 150.2 and 238.1 MeV.
Whereas at the two lower energies there is a monotonous decrease with scattering angle,
a diffraction minimum exists near $100^\circ$ for the 238.1 MeV electron impact. There is good agreement with the available experimental data \cite{Reu82,Off91}, which are corrected for global QED effects.
Fig.2 displays the corresponding cross section changes (\ref{3.7}) by the QED effects and dispersion in comparison with the (Coulomb-distorted)
Born approximation (\ref{3.6}). Apart from showing the combined influence of these effects,
the vacuum polarization as well as the vs effect and their supersposition are provided separately. This is done by only retaining the respective potentials in (\ref{2.10}) or the corresponding perturbative parts in (\ref{3.1}).
For 56 MeV impact (Fig.2a), the deviations from the Born results are, as expected for this light nucleus, extremely small.
Bremsstrahlung enhances the cross-section change, dependent on the cut-off frequency $\omega_0$.
If not stated otherwise, $\omega_0=1 $ MeV is taken throughout,
which is well below the first excited state at 4.4 MeV for $^{12}$C or at 2.6 MeV for $^{208}$Pb.
The contribution $\Delta \sigma^{\rm box}$ is tiny at this energy (decreasing from $\sim -10^{-4}$ to $\sim - 10^{-3}$ with angle), and its inclusion is not visible in the graph.
At 150 MeV (Fig.2b) the influence of the QED effects on the cross section is notably smaller at the larger angles when the vs potential is considered nonperturbatively, than when predicted by the Born approximation.
Also a dispersion effect is peceptible at angles beyond $70^\circ$.
The situation is different at 238.1 MeV (Fig.3).
While the Born results for vacuum polarization or the vs effect still have a monotonous angular dependence,
the nonperturbative results mimic the presence of the
diffraction minimum in $d\sigma_{\rm coul}/d\Omega_f$ by a resonance-like structure.
The comparision with Fig.2 indicates that the global strength of the vs correction increases with $E_i$, whereas vacuum polarization remains at $1-2\%$.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig4.pdf}
\vspace{-0.5cm}
\caption
{
Differential cross section $\frac{d\sigma_{\rm coul}}{d\Omega_f}$ for electrons of 56 MeV $(-\cdot -\cdot -$), 150 MeV $(----)$ and 167 MeV (---------)\\ colliding with $^{208}$Pb as a function of scattering angle $\vartheta_f$. Included are the relative experimental data by Friedrich and Lenz ($\bullet$, \cite{Fr72}) at 167 MeV.
}
\end{figure}
The dispersion correction
$\Delta \sigma_{\rm box}$ is displayed in Fig.3b. It is estimated in the second-order Born approximation by considering
three dominant transiently excited nuclear states
of low angular momentum \cite{Jaku22}. It is seen that the resonant-like structure is also present in this correction.
For comparison, the result from the Friar-Rosen theory \cite{FR74} is included, which employs a closure approximation by setting all nuclear excitation
energies to the fixed value of 15 MeV \cite{Jaku21}.
We recall that in the experimental data by Offermann et al \cite{Off91} only a smooth background from the QED effects had been subtracted.
In order to account for the influence of the structure in the nonperturbative approach, it is isolated by forming the difference
between $\Delta\sigma^C$ and $\Delta \sigma^{B1-C}$.
This difference is added to the dispersion correction, which is also not considered in the data, and the result is included in Fig.3b for
the two theories of $\Delta \sigma^{\rm box}$. Comparison is made with the experimental cross section change, obtained from (\ref{3.5}) by identifying $d\sigma/d\Omega_f$ with the cross section measurements.
One has to keep in mind that these so generated data points depend crucially on the way
how recoil is incorporated into the Coulombic result.
For a light target like $^{12}$C at such a high impact energy even the recoil prefactor $k_f/(k_i f_{\rm rec})$ in (\ref{3.2}) reduces the cross section by 5\%,
apart from the shift in angle by the reduced collision velocity \cite{Jaku21b}.
Although, as compared to $\Delta \sigma^{\rm box}$ alone, the deviation between theory and experiment around $95^\circ$ persists,
the consideration of the oscillatory bahaviour of the QED correction improves the agreement on the wings.
We note that at this energy the Friar-Rosen theory performs still reasonably well, while at higher energies its estimates for $\Delta \sigma^{\rm box}$ are far too small \cite{Jaku22}.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig5a.pdf}
\vspace{-1.5cm}
\vspace{-0.5cm}
\includegraphics[width=11cm]{qedvertfig5b.pdf}
\caption
{
Cross section change $\Delta \sigma^C$ in (a) 56 MeV and (b) 150 MeV $e + ^{208}$Pb collisions as a function of scattering angle $\vartheta_f$.
Shown are the results for vacuum polarization ($-\cdot - \cdot -)$, vertex and self-energy correction $(----)$ and for both\\ (-----------, thin line),
as well as the additional inclusion of the soft-bremsstrahlung contribution for $\omega_0=1$ MeV (-----------, thick line). Also shown are the Born results $\Delta \tilde{\sigma}^{\rm vac}$ ($\cdots\cdots$, upper line) and $\Delta \tilde{\sigma}^{\rm vs}$ ($\cdots\cdots$, lower line).
}
\end{figure}
\subsection{The $^{208}$Pb nucleus}
Fig.4 provides the angular distribution of the Coulombic cross section at the impact energies 56 and 150 MeV, as well as for 167 MeV where experimental data are available,
which were measured relative to $^{12}$C and normalized to the $^{12}$C phase-shift theory \cite{Fr72}.
For the extended lead nucleus, diffraction oscillations are already present at 150 MeV, while having a still earlier onset (i.e. at smaller angles) at 167 MeV.
The QED changes in the differential cross section are plotted in Fig.5, again in comparison with the Born results from (\ref{3.1}).
The deviations between the two prescriptions are considerably larger than for $^{12}$C, even at 56 MeV (Fig.5a), with notable differences already at the smallest angles.
Since the Born results $\Delta\tilde{\sigma}^{\rm vac}$ and $\Delta \tilde{\sigma}^{\rm vs}$ coincide with those from Fig.2, the effect of Coulomb distortion when proceeding from $^{12}$C to $^{208}$Pb becomes obvious.
We note that the combined inclusion of $U_e$ and $V_{\rm vs}$, leading to $\Delta \sigma^{\rm vac+vs}$, differs from the sum resulting from the separate treatments, $\Delta \sigma^{\rm vac} + \Delta \sigma^{\rm vs}$.
At 56 MeV, this difference is up to 3\% (in comparison to 1\% for $^{12}$C), increasing with energy.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig6.pdf}
\vspace{-0.5cm}
\caption
{
Angular dependence of the nonperturbative cross-section change by all QED effects for 56 MeV electrons colliding with $^{208}$Pb
and cut-off frequencies $\omega_0=1$ MeV (------------), 0.1 MeV $(----)$ and 0.03 MeV $(-\cdot -\cdot -$).
Also shown is the isolated contribution from soft bremsstrahlung for $\omega_0=0.1$ MeV $(\cdots\cdots)$.
}
\end{figure}
For 150 MeV (Fig.5b), the nonperturbative QED effects show oscillations, the minima of which correspond to the minima in the respective differential cross section.
In a similar way as for $^{12}$C, the cross-section modifications are particularly large and thus easily discernable when the Coulombic cross section does not notably change with angle (which is the case in the region of a diffraction minimum).
Fig.6 provides the influence of the soft bremsstrahlung when the detector resolution is changed.
The brems\-strahlung itself is approximately constant in angle at 56 MeV,
apart from the foremost regime. It increases, however, in strength when the cut-off frequency $\omega_0$ is lowered,
according to the logarithmic dependence $(\ln 2\omega_0/c^2$) in the formula (\ref{2.4}).
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig7.pdf}
\vspace{-0.5cm}
\caption
{
Coulombic spin asymmetry $- S_{\rm coul}$ for 56 MeV electrons colliding with $^{12}$C (----------) and $^{208}$Pb $(-\cdot - \cdot -$) as a function of scattering angle $\vartheta_f$.
}
\end{figure}
\subsection{Spin asymmetry}
For perpendicularly polarized incident electrons, the Sherman function $S$ is defined as the relative cross section difference when the initial spin is flipped from up ($\uparrow$) to down ($\downarrow$),
\begin{equation}\label{3.8}
S\;=\;\frac{d\sigma/d\Omega_f(\uparrow) - d\sigma/d\Omega_f(\downarrow)}{d\sigma/d\Omega_f(\uparrow)+d\sigma/d\Omega_f(\downarrow)}.
\end{equation}
Correspondingly, the modification of the spin asymmetry by the radiative corrections, relative to the Coulombic Sherman function $S_{\rm coul}$, is calculated from
\begin{equation}\label{3.9}
d\,S\;=\;\frac{S}{S_{\rm coul}}\;-\,1.
\end{equation}
For higher collision energies when diffraction induces zeros in $S_{\rm coul}$, the definition (\ref{3.9}) is no longer meaningful. Therefore we have restricted the spin investigations to an energy of 56 MeV.
For this energy, the Coulombic spin asymmetry is displayed in Fig.7 for both targets.
A logarithmic scale is used (and hence $-S_{\rm coul}$ is shown) to demonstrate the strong increase of the spin asymmetry with scattering angle.
It should be noted that $S$ is much larger for $^{208}$Pb, since the spin asymmetry increases with nuclear charge due to the stronger relativistic effects in the
electron-nucleus encounter.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig8.pdf}
\vspace{-0.5cm}
\caption
{
Change $dS$ of the Sherman function from 56 MeV $e+^{12}$C collisions by the nonperturbative QED effects and dispersion (------------, thick line) as a function of scattering angle $\vartheta_f$.
Also shown are the separate contributions from vacuum polarization ($dS_{\rm vac}, \;-\cdot -\cdot -$), from the vs correction ($dS_{\rm vs},\;----$), from both ($dS_{\rm vac+vs}$, ------------- thin line) and from dispersion ($dS_{\rm box},\;\cdots\cdots$).
}
\end{figure}
Fig.8 depicts the change $dS$ for $^{12}$C by means of the QED effects and the dispersion correction.
Like in the case of the cross-section modifications, the effect of vacuum polarization is small, at most 1\%. The vs contribution is of opposite sign
and considerably larger in magnitude, well beyond the factor of $2$ anticipated from exact bound-state QED investigations \cite{Sh00}.
We recall that the spin-asymmetry change by the QED effects is zero in the Born approximation (\ref{2.1}) or (\ref{3.1}), since the leading-order cross section
is only multiplied by a factor which drops out in (\ref{3.8}). Moreover, as the soft-bremsstrahlung contribution contains the leading-order cross section as a factor, it does also not add to any asymmetry change
in a higher-order approach. This was already stated by Johnson et al \cite{Jo62}, who calculated the QED corrections to $S$ within the second-order Born approximation in the Coulomb field.
In this context our previous Born results for $dS_{\rm vsb}$ \cite{Jaku21} should only be considered as qualitative estimates, since Coulomb distortion was not included in the contributions from vs and from soft bremsstrahlung.
Also shown in Fig.8 is the asymmetry change from dispersion \cite{Jaku22}, which tends to large negative values for small angles.
The Sherman function with inclusion of all radiative corrections can be estimated by
\begin{equation}\label{3.10}
S_{\rm tot}\;\approx\;S_{\rm vac+vs} \,+\,d S_{\rm box} \,S_{\rm coul}\,\frac{d\sigma_{\rm coul}/d\Omega_f}{d\sigma^{\rm QED}/d\Omega_f},
\end{equation}
where $S_{\rm vac+vs}$ and $S_{\rm coul}$ are calculated from the leading-order term in (\ref{3.3}) and (\ref{3.1}), respectively, or alternatively in terms of the direct (A)
and spin-flip (B) amplitudes as obtained from the phase-shift analysis,
\begin{equation}\label{3.11}
S\;=\;\frac{2 \mbox{ Re } \{A B^\ast\}}{|A|^2 +|B|^2}.
\end{equation}
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig9.pdf}
\vspace{-0.5cm}
\caption
{
Change $dS$ of the Sherman function from 56 MeV $e+^{208}$Pb collisions by the nonperturbative QED effects ($dS_{\rm vac+vs}$, ------------) as a function of scattering angle $\vartheta_f$.
Included are the separate contributions from vacuum polarization ($dS_{\rm vac},\;-\cdot -\cdot -$) and from the vs effect ($dS_{\rm vs},\;----$).
}
\end{figure}
Furthermore, $d\sigma^{\rm QED}/d\Omega_f$ is calculated from (\ref{3.3}) by omission of dispersion, while $S_{\rm box}$ (and consequently $dS_{\rm box}$) is obtained from (\ref{3.1}) by dropping all three QED contributions.
Into the formula (\ref{3.10}) enters the assumption that the cross-section change due to dispersion is small (for a $^{12}$C target, it is below 1\% for collision energies up to 100 MeV),
such that it can be omitted in the denominator.
Consequently, the total change of asymmetry can be found from
$$dS_{\rm tot}\;=\;\frac{S_{\rm tot}}{S_{\rm coul}}\,-1$$
\begin{equation}\label{3.12}
\approx\; dS_{\rm vac+vs}\, + \,dS_{\rm box}\;\frac{1}{1+(\Delta \sigma^{\rm vac+vs}+\Delta \sigma^{\rm soft})},
\end{equation}
which is also displayed in the figure.
Actually for a low-$Z$ target like $^{12}$C the determination of the asymmetry changes suffers from large numerical instabilities,
which are partly smoothed in the figure.
In Fig.9 the spin-asymmetry change by the QED effects is displayed for a $^{208}$Pb target. Diffraction effects are already perceptible at 56 MeV, producing a zero in $S_{\rm coul}$ near $\vartheta_f=16^\circ$, which induces the strong rise of $dS$ near the smallest angles shown in the figure.
Also, as compared to the nearly constant values of $dS_{\rm vac}$ or $dS_{\rm vs}$ for a carbon target at the same energy,
strong angular variations of the QED effects take place for lead, although the Coulombic cross section shows hardly any modulations.
However, the total QED spin-asymmetry changes are smaller than the respective changes for $^{12}$C.
It is noteworthy that even slight diffraction effects cause sign changes in $dS_{\rm vac}$ and $dS_{\rm vs}$, such that there is an angular region (at 56 MeV between $120^\circ$ and $160^\circ$) where both QED modifications are of the same sign.
\begin{figure}
\vspace{-1.5cm}
\includegraphics[width=11cm]{qedvertfig10.pdf}
\vspace{-1.0cm}
\caption
{
Change $dS$ of the Sherman function for $e+^{208}$Pb collisions at $\vartheta_f=173^\circ$ by the QED effects (---------------) as a function of collision energy $E_{\rm i,kin}=E_i-c^2$.
Included are the contributions $dS_{\rm vac}$ from vacuum polarization $(-\cdot - \cdot -$) and $dS_{\rm vs}$ from the vertex and self-energy correction $(----$).
}
\end{figure}
The behaviour of the Sherman function at very low collision energies is interesting from an experimental point of view
in the context of accuracy tests of different kinds of detectors.
We provide in Fig.10 the energy dependence of the QED corrections at an angle of $173^\circ$, used in a recent precision experiment where 5 MeV electrons collided with a gold target \cite{Gr20}.
At this angle, $dS_{\rm vac}$ first increases with energy up to 0.5\%, and then decreases again beyond 30 MeV.
The vs contribution shows the opposite behaviour, with $|dS_{\rm vs}/dS_{\rm vac}|$
increasing from 1.5 at 3 MeV to 10 or more at the largest energies considered.
The similarity between the energy pattern (Fig.10) and the angular pattern (Fig.9) of the nonperturbative QED corrections indicates their basic dependence on the momentum transfer, $|\mbox{\boldmath $q$}| \approx 2 k_i \sin (\vartheta_f/2)$, into which the two variables enter as product.
\section{Conclusion}
The QED corrections to the elastic scattering cross section and to the beam-normal spin asymmetry were estimated by using a nonperturbative approach
in terms of a suitable potential for the vertex and self-energy correction, and the Uehling potential for vacuum polarization.
When investigating electron scattering from the $^{12}$C nucleus, notable deviations from the respective Born predictions for the
cross-section change were only found near and above 150 MeV impact energy, which are increasing with scattering angle.
In particular, the correction by the vs contribution, although mostly of opposite sign as compared to the effect of vacuum polarization,
is in magnitude considerably larger than the factor of two hitherto assumed from the results of exact low-energy bound-state considerations.
Like the cross-section changes by dispersion (estimated with or without the use of a closure approximation), the nonperturbative QED results show an oscillatory behaviour near the diffractive cross-section minima.
The numerical accuracy of our estimated QED cross-section changes is better than 0.5\% at 56 MeV, deteriorating to 5\% at 238 MeV for the backmost angles.
In case of the lead target, the deviations from the Born QED results are quite large, up to nearly a factor of 2 at backmost angles
even for a low energy of 56 MeV.
A diffraction pattern emerges at energies near 100 MeV, with an increasing number of structures at higher energies, in concord with the diffractive structures of the Coulombic cross section.
The numerical accuracy is higher than for $^{12}$C, below 1\% even at 150 MeV.
One has to keep in mind that the size of the total QED corrections depends strongly on the contribution of the soft bremsstrahlung, which in turn is controlled by the resolution of the electron detector.
The nonperturbative consideration of the vs effect allows also for a consistent estimate of the Sherman function. For low collision energies, its changes by the QED effects increase strongly with energy.
For lead this holds up to about 30 MeV at backward angles which are of particular interest to the experimentalists due to the large values of the spin asymmetry.
For example, at $170^\circ$ and 3.5 MeV, these QED changes amount to $dS\approx -0.5\%$, while at 5 MeV, $dS \approx -0.9\%$ for both targets.
On the other hand, at 56 MeV, they are about 5\% for $^{12}$C and somewhat less (at most 3\%) for $^{208}$Pb in the whole angular regime.
The numerical accuracy of $dS$ for carbon is unfortunately quite poor, partly due to the small absolute values of $S$ (in the forward regime), and partly due to numerical instabilities when solving the Dirac equation
(in the backward hemisphere).
It amounts up to 0.5\% at 3 MeV and 3\% at 10 MeV, but $10-15\%$ at 56 MeV.
For lead, the results are stable, with an accuracy of $\lesssim 0.25\%$ at 30 MeV and $\lesssim 1\%$ at 56 MeV.
The dispersion effects on the cross section are small, but on the Sherman function they are formidable, even for 56 MeV electron impact on $^{12}$C. They lead to a total change of $S$ by the radiative corrections up to 20\% or more at the smallest angles.
An investigation of dispersion for a lead target is in progress.
\vspace{0.5cm}
\noindent{\large\bf Acknowledgments}
I would like to thank C.Sinclair for directing my interest to his work.
|
\section{\label{sec:level1}First-level heading}
\textit{Introduction.}---The CP violation has always been playing a key role in particle physics. Measurements of CP violation effects in flavor processes are crucial to determine the Cabibbo-Kobayashi-Maskawa (CKM)~\cite{Cabibbo:1963yz,Kobayashi:1973fv} matrix, whose unitarity is a critical test of the standard model (SM). Moreover, the observed matter-antimatter asymmetry in the universe requires the CP violation as one of the criteria~\cite{Sakharov:1967dj}, but the CP violation in the SM is too small to be the only source~\cite{Bernreuther:2002uj,Canetti:2012zc}. Therefore, precision tests of CP asymmetries may open a window to physics beyond the SM.
Although the CP violation is induced by the weak interaction, its visualization in hadron decays typically requires interplay between weak and strong interactions and thus receives pollution from strong dynamics. Therefore, explorations of CP asymmetries in diverse physical observables with diverse dependence on strong dynamics would be beneficial for deciphering the mystery of the CP violation. Particularly, observables free of strong phases provide a clean environment~\cite{BESIII:2021ypr,Wang:2022fih}. We hereby propose a novel CP violation observable --- double-mixing CP asymmetry, which is induced by interferences between two mixing processes in one cascade decay. It is independent on CP asymmetry in decay and mixing, and can possibly be large even in absence of strong phases.
The double-mixing CP asymmetry exists in cascade decay chains, in which at least two oscillating neutral mesons are involved. Such a process typically starts from a primary neutral meson $M_1^0$, which decays, before or after oscillating to its antiparticle $\bar{M}_1^0$, into a secondary neutral meson $\bar{M}_2^0(M_2^0)$ and other particles. Usually, $\bar{M}_2^0(M_2^0)$ also further decays into directly detectable particles by detectors. The process happens via multiple quantum paths, which interfere with each other. One example is shown in FIG.~\ref{fig:2chain}, which allows two oscillation paths $M_1^0\to \bar{M}_2^0 \to M_2^0$ and $M_1^0\to \bar{M}_1^0 \to M_2^0$. The interference between the two paths induces the double-mixing CP asymmetry that we focus on. There also exist other interfering paths such as $M_1^0\to \bar{M}_1^0 \to M_2^0\to \bar{M}_2^0$ and $M_1^0\to \bar{M}_2^0$. If $M_2^0$ and $\bar{M}_2^0$ decay into the same final state, usually a CP eigenstate, there are interferences between the four paths, and induce very fruitful CP violation observables.
The double-mixing CP asymmetry has a dependence on two time variables, the oscillation time $t_1$ of $M_1^0$ and $t_2$ of ${M}_2^0$. Therefore, a two-dimensional time-dependent analysis on $M_1^0(t_1)\to \bar{M}_2^0(t_2)\to f$ can in principle be performed as a new measure tool. Practically, the double-mixing CP asymmetry can be very significant. For example, the new CP violation effect in $B^0_s(t_1) \to \rho^0\bar{K}^0(t_2) \to \rho^0\pi^-\ell\nu$ can reach as large as 50\%. As the new CP violation still exists in absence of strong phases, it provides the possibility of directly extracting weak phases in the CKM matrix, and can thus be sensitive to some dynamics beyond the SM. The new CP violation effects can be observed and measured at flavor experiments including BESIII~\cite{BESIII:2020nme}, Belle II~\cite{Belle-II:2018jsg}, LHCb~\cite{Cerri:2018ypt} and future lepton colliders~\cite{SuperTauCharm,Charm-TauFactory:2013cnj,CEPCStudyGroup:2018ghi,TLEPDesignStudyWorkingGroup:2013myl}, including CEPC and FCC-ee, which can also produce fruitful flavor results~\cite{Qin:2017aju,Ali:2018ifm,Cheng:2018khi,Ding:2019tqq,Qin:2020zlg,Shen:2022ffi}.
\begin{figure}[htbp]
\centering
\includegraphics[width=1.0\linewidth]{doublemixing.png}
\caption{Interference between two oscillation paths in the cascade decay. The decay products associated with $M_2$ are not displayed.}
\label{fig:2chain}
\end{figure}
In the rest of the paper, we will first present the general formulas for the double-mixing CP asymmetry in the process $M_1^0(t_1)\to \bar{M}_2^0(t_2)\to f$, and then take $B^0_s \to \rho^0 \bar{K}^0 \to \rho^0 (\pi^-\ell^+{\nu}_\ell)$ as the example to perform the numerical analysis.
{\textit{Formulae.}} ---In the following derivation, we accept the convention that the mass eigenstates $M_{H,L}$ of the neutral mesons are superpositions of their flavor eigenstates
\begin{eqnarray}
\ket{M_{H,L} } = p \ket{ M^0 } \mp q \ket{ \bar{M}^0 } \; ,
\end{eqnarray}
where $q,p$ are complex coefficients. The mass and decay width differences are defined as $\Delta m \equiv m_H - m_L$ and $\Delta \Gamma \equiv \Gamma_H-\Gamma_L$ such that the oscillation is formulated by
\begin{eqnarray}
\ket{ M^0(t) } &=& g_+(t) \ket{M^0} - {q\over p} g_-(t) \ket{\bar{M}^0} \; , \nonumber \\
\ket{ \bar{M}^0(t) } &=& g_+(t) \ket{\bar{M}^0} - {p\over q} g_-(t) \ket{{M}^0} \; , \nonumber \\
\text{with}\ g_{\pm}(t) & = & {1\over2} \left[ e^{-im_Ht-{1\over2}\Gamma_H t } \pm e^{ -im_L t-{1\over2}\Gamma_L t } \right] \; ,
\end{eqnarray}
in the case that the theory is CPT invariant.
In this letter, we will focus on the case with the decay process happening through $M_1^0\to \bar{M}_2^0 \to M_2^0$ and $M_1^0\to \bar{M}_1^0 \to M_2^0$, as shown in FIG.~\ref{fig:2chain}. A more comprehensive study including other cases is left for future~\cite{inpre}. The two-dimensional time-dependent CP asymmetry is defined as
\begin{eqnarray}
A_\mathrm{CP}(t_1,t_2) &\equiv& {|\mathcal{M}|^2(t_1,t_2) - |\mathcal{\bar{M}}|^2 (t_1,t_2) \over |\mathcal{M}|^2(t_1,t_2) + |\mathcal{\bar{M}}|^2(t_1,t_2) } \; ,
\end{eqnarray}
where the amplitude $\mathcal{M}(t_1,t_2)$ is the sum of amplitudes of the two paths $M_1^0\to \bar{M}_2^0 \to M_2^0$ and $M_1^0\to \bar{M}_1^0 \to M_2^0$, and the amplitude $\bar{\mathcal{M}}(t_1,t_2)$ is the CP conjugate of $\mathcal{M}(t_1,t_2)$. The oscillation time $t_1$ and $t_2$ can be identified at experiments by using vertex detection techniques. To give prominence to the mixing effects, we assume that there are no direct CPVs in the decay $M_1\to M_2$ or $M_2\to f$. Again, a more comprehensive analysis taking all these effects into consideration will be presented in an upcoming work~\cite{inpre}. In such a case $|\braket{\bar{M}_2^0}{M_1^0}| = |\braket{{M}_2^0}{\bar{M}_1^0}|$ and the decay amplitudes are related by $\braket{\bar{M}_2^0}{M_1^0} = \braket{{M}_2^0}{\bar{M}_1^0} e^{2i\delta}$, where $\delta$ is a pure weak phase. We further write the mixing parameters of $M_{1,2}$ as $(q/p)_{1,2} = |(q/p)_{1,2}| e^{-i\phi_{1,2}}$. Then, the time-dependent CP asymmetry is calculated to be
\begin{eqnarray}\label{eq:acp}
A_\mathrm{CP}(t_1,t_2) &=& {N(t_1,t_2)\over D(t_1,t_2)} \; ,\\
N(t_1,t_2)&=&|g_{1,+}(t_1)|^2 C_+(t_2) + |g_{1,-}(t_1)|^2 C_-(t_2) \nonumber \\
&&+ e^{-\Gamma_1t_1} \sinh{\Delta\Gamma_1t_1\over 2} \; S_h(t_2) \; , \nonumber\\
&&+ e^{-\Gamma_1t_1} \sin{(\Delta m_1t_1)}\; S_n(t_2) \; , \nonumber\\
D(t_1,t_2)&=&|g_{1,+}(t_1)|^2 C'_+(t_2) + |g_{1,-}(t_1)|^2 C'_-(t_2) \nonumber \\
&& + e^{-\Gamma_1t_1} \sinh{\Delta\Gamma_1t_1\over 2}\; S'_h(t_2) \nonumber \\
&& + e^{-\Gamma_1t_1} \sin{(\Delta m_1t_1)}\; S'_n(t_2) \; , \nonumber
\end{eqnarray}
where
\begin{eqnarray}
C_+(t_2) &=& |g_{2,-}(t_2)|^2 \; \left( \left| {p_2/ q_2} \right|^2 - \left| {q_2/ p_2} \right|^2 \right) \;, \nonumber \\
C_-(t_2) &=& |g_{2,+}(t_2)|^2 \; \left( \left| {q_1/ p_1} \right|^2 - \left| {p_1/ q_1} \right|^2 \right) \; ,
\end{eqnarray}
consisting of $M_2$- and $M_1$-mixing induced CPV, respectively. In a more complete form, they both contain the direct CP asymmetries of $M_1\to M_2$ and $M_2\to f$. The double-mixing CP asymmetries are given by
\begin{eqnarray}\label{eq:shn}
S_h(t_2) &=& {e^{-\Gamma_2t_2}\over2} [ - 2 \sin{(\Delta m_2t_2)} \sin(\phi_1+\phi_2+2\delta ) \nonumber \\
&&+ \sinh{\Delta\Gamma_2t_2\over2} \left( \left|{q_1 \over p_1}\right|\left|{p_2\over q_2}\right| - \left|{p_1\over q_1}\right| \left|{q_2\over p_2}\right| \right) \nonumber \\
&&\times \cos(\phi_1+\phi_2+2\delta ) ] \;, \nonumber \\
S_n(t_2) &=& {e^{-\Gamma_2t_2}\over2} [ 2 \sinh{\Delta\Gamma_2t_2\over2} \sin(\phi_1+\phi_2+2\delta ) \nonumber \\
&& +\sin{(\Delta m_2t_2)} \left( \left|{q_1 \over p_1}\right|\left|{p_2\over q_2}\right| - \left|{p_1\over q_1}\right| \left|{q_2\over p_2}\right| \right) \nonumber \\
&&\times \cos(\phi_1+\phi_2+2\delta ) ] \;,
\end{eqnarray}
where some doubly suppressed small quantities are neglected. The phase angle dependence on $\phi_1+\phi_2$ clearly indicates that they are induced by the double mixing interference. The two terms $S_h$ and $S_n$ have very different time dependence: $S_n$ has sine dependence on $t_1$ and hyperbolic sine dependence on $t_2$, and $S_h$ has hyperbolic sine dependence on $t_1$ and sine dependence on $t_2$. The two different types of time dependences can be combined to separate $S_h$ and $S_n$ and perform a complementary analysis. It can be observed that in absence of strong phases in decays and mixings, {\it i.e.}, $|q/p|=1$, the double-mixing CP asymmetries $S_h$ and $S_n$ still exist, and it thus provides a clean method to extract CKM phases.
The terms contributing to the denominator of \eqref{eq:acp} are given by
\begin{eqnarray}\label{eq:deno}
C'_+(t_2) &=& 2|g_{2,-}(t_2)|^2 \;, \nonumber \\
C'_-(t_2) &=& 2 |g_{2,+}(t_2)|^2 \; , \nonumber \\
S'_h(t_2) & = & {e^{-\Gamma_2t_2}} \sinh{\Delta\Gamma_2t_2\over2} \cos(\phi_1+\phi_2+2\delta ) \; , \nonumber \\
S'_n(t_2) & = & {e^{-\Gamma_2t_2}} \sin{(\Delta m_2t_2)} \cos(\phi_1+\phi_2+2\delta ) \; ,
\end{eqnarray}
with suppressed terms neglected.
For the decay chain constructed by the oscillation paths $M_1^0\to M_2^0 \to \bar{M}_2^0$ and $M_1^0\to \bar{M}_1^0 \to \bar{M}_2^0$, the results can be obtained by taking \eqref{eq:acp} and replacing $q_2/p_2$ with $p_2/q_2$, including $|q_2/p_2| \to |p_2/q_2|$ and $\phi_2\to -\phi_2$.
{\it Example.} ---
We take $B^0_s \to \rho^0 \bar{K}^0 \to \rho^0 (\pi^-\ell^+{\nu}_\ell)$ as an example. Because both $B^0_s$ and $K^0$ have large mixing effects, their interference could also be large. As the penguin operator contributions to $B^0_s \to \rho^0 \bar{K}^0$ are suppressed by Wilson coefficients~\cite{Beneke:2003zv,Ali:2007ff}, we neglect them for simplicity, {\it i.e.}, the direct CP asymmetry is neglected. The decay amplitude $\braket{\rho^0 \bar{K}^0}{B^0_s}$ and its charge conjugate $\braket{\rho^0 K^0}{\bar{B}^0_s}$ then have the identical magnitude, and the phase difference is given by $e^{i2\delta} = -{V_{ub}^*V_{ud}\over V_{ub}V_{ud}^*}$, where the minus sign is due to $CP|\rho^0 K^0\rangle = - |\rho^0 \bar{K}^0\rangle$ in a pseudoscalar meson decay. We also neglect the indirect CP asymmetries induced by $B_s$ and $K$ mixings, {\it i.e.,} taking $|q/p|=1$, and then the phases of the mixing coefficients are approximately given by
\begin{eqnarray}
e^{-i\phi_1} = {V_{tb}^*V_{ts}\over V_{tb}V_{ts}^*} \; , \qquad e^{-i\phi_2} = {V_{cd}^*V_{cs}\over V_{cb}V_{cs}^*} \; .
\end{eqnarray}
Under such assumptions, the only non-vanishing CP asymmetry is induced by the double-mixing interference, contained in $S_h$ and $S_n$ given by \eqref{eq:shn}. With the numerical inputs listed in TABLE~\ref{table:parameters}, we can calculate the numerical result for the two-dimensional time-dependent CP violation observables $A_{\rm CP}(t_1,t_2)$, and the contributions $A_{h}(t_1,t_2)$ and $A_{n}(t_1,t_2)$ from the $S_h$ and $S_n$ terms, respectively.
The result for the $A_{\rm CP}(t_1,t_2)$ dependence on $t_1$ and $t_2$ is displayed in the left panel of FIG.~\ref{fig:BsrhoK}. It can be observed that the magnitude of the peak values can be larger than 50\%. If the data sample is not large enough for a two-dimensional time dependence analysis, we can also integrate out one time dimension and get the evolution along the remaining one. Integrating $t_2$ from 0 to the Kaon average lifetime $\tau_{K}\equiv 2/(\Gamma_{K,L}+\Gamma_{K,S})$, we obtain the $t_1$ dependence of $A_{\rm CP}$ displayed in the middle panel of FIG.~\ref{fig:BsrhoK}; integrating $t_1$ from $\tau_{B_s}$ to $5\tau_{B_s}$ with $\tau_{B_s}\equiv 2/(\Gamma_{B_s,L}+\Gamma_{B_s,S})$, we obtain the $t_2$ dependence of $A_{\rm CP}$ displayed in the right panel of FIG.~\ref{fig:BsrhoK}. The contributions $A_{h}$ and $A_{n}$ are displayed by dashed and dotted curves, respectively. The time evolution along $t_2$ is dominated by $A_h$, because $A_n$ highly oscillates along the integrated time $t_1$. The time evolution along $t_1$ is dominated by $A_n$, but the $A_h$ contribution is also considerable.
\begin{table}
\caption{The input parameters and their values.}\label{table:parameters}
{
\begin{tabular}{cc}
\hline
\textbf{Parameter} & \textbf{Value} \\
\hline
$|q_1/p_1|$ & $1.0003\pm 0.0014$~\cite{HFLAV:2022pwe} \\
$\phi_1$ & $(-2.106\pm 0.135)^\circ$~\cite{ParticleDataGroup:2022pth} \\
$x_1=2\Delta m_1/(\Gamma_{B_s,L}+\Gamma_{B_s,S})$ & $27.01\pm 0.10$~\cite{ParticleDataGroup:2022pth} \\
$y_1=\Delta \Gamma_1/(\Gamma_{B_s,L}+\Gamma_{B_s,S})$ & $-0.064\pm 0.003$~\cite{ParticleDataGroup:2022pth} \\
$|q_2/p_2|$ & $0.996774\pm 0.000019$~\cite{ParticleDataGroup:2022pth} \\
$\phi_2$ & $(0.176\pm 0.001)^\circ$~\cite{ParticleDataGroup:2022pth} \\
$x_2=2\Delta m_2/(\Gamma_{K,L}+\Gamma_{K,S})$ & $0.946\pm 0.002$~\cite{ParticleDataGroup:2022pth} \\
$y_2=\Delta \Gamma_2/(\Gamma_{K,L}+\Gamma_{K,S})$ & $-0.996506\pm 0.000016$~\cite{ParticleDataGroup:2022pth} \\
$2\delta$ & $(-48.907\pm 3.094)^\circ$~\cite{ParticleDataGroup:2022pth} \\
\hline
\end{tabular}
}
\end{table}
\begin{widetext}
\
\begin{figure}[htbp]
\centering
\includegraphics[keepaspectratio,width=6.cm]{ASt1t2.png}
\hspace{0.8cm}
\includegraphics[keepaspectratio,width=4.8cm]{ASt1.png}
\hspace{0.8cm}
\includegraphics[keepaspectratio,width=4.7cm]{ASt2.png}
\caption{Time dependence of the double-mixing CP asymmetry $A_{\rm CP}$ in $B^0_s(t_1) \to \rho^0 \bar{K}^0(t_2) \to \rho^0 (\pi^-\ell^+{\nu}_\ell)$. The left panels display the two-dimensional time dependence. The middle panel and the right panel display the dependence on $t_1$ (with $t_2$ integrated from 0 to $\tau_{K}$) and $t_2$ (with $t_1$ integrated from $\tau_{B_s}$ to $5\tau_{B_s}$), respectively. }\label{fig:BsrhoK}
\end{figure}
\end{widetext}
{\it Conclusion.} --- It is found that a novel type of CP violation effect, the double-mixing CP asymmetry, exists in cascade decays involving two neutral mesons oscillating.
The double-mixing CP asymmetry depends on two time variables and thus a two-dimensional time dependence analysis can be performed. Unlike direct CP asymmetries, its existence does not require nonzero strong phases, and thus it provides opportunities to directly extract weak phases without pollution from strong dynamics, which is crucial for CKM matrix determination and new physics search.
Practically, the new CP violation effects can be very significant in some channels, which are very promising to be measured in experiments.
{\it Acknowledgement.} --- The authors are grateful to Cai-Dian L\"u and Yue-Hong Xie for useful discussions.
This work is supported by Natural Science Foundation of China under grant No. 12005068.
|
\section{Introduction}
The measurements showing that the observable Universe is in accelerated expansion \cite{Riess1998,Perlmutter1999} demand a full analysis of the implications that a {\em positive} cosmological constant $\Lambda >0$ has. One particular known result is that $\Lambda \neq 0$ imposes a finite bound
\begin{equation}\label{lim}
\Lambda A < 4\pi (1-g)
\end{equation}
on the area $A$ of marginally future-trapped\footnote{\label{foot1}All the results are applicable, {\em mutatis mutandis}, to marginally past-trapped surfaces, but for the sake of conciseness I will from the start just consider future-trapped surfaces.} closed surfaces of genus $g$ \cite{HSN,W,Simon} if these are stable in spatial directions in the sense of \cite{AMS,AMS1}---equivalently, `outer' in the sense of \cite{Hay,HSN}---and the dominant energy condition holds. The stability assumption is usually interpreted \cite{AG,HSN,SILS,Hay,PBH,BBGV} as implying that the marginally trapped surfaces (MTSs) enclose a black hole (BH) region, leading to a maximum size of BHs if $\Lambda >0$---in which case the above bound contains information only if $g=0$, topological spheres.
To understand these physical limitations, I recently \cite{Snew} considered some simple, spherically symmetric, models where the area bound is surpassed. In spherical symmetry and for fixed $\Lambda$, the area of marginally trapped round spheres depends on the total mass enclosed by the surface, and increases as the mass also does. Therefore, one may wonder what happens if a BH in formation with area very close to the bound \eqref{lim} accretes, or receives, extra mass from its surroundings. As shown with explicit models in \cite{Snew}, the dynamical horizon foliated by marginally trapped round spheres merges with a timelike membrane at a distinguished round sphere $\bar S$ with area $A_{\bar S}=4\pi/\Lambda$ forming a marginally trapped tube \cite{AG,Booth,BBGV} that changes signature precisely at $\bar S$. The entire marginally trapped tube (MTT) foliated by MTSs becomes locally a {\em holographic screen} as introduced in \cite{BE,BE1}, but globally is more general than that as it may contain null portions that are non-expanding horizons\footnote{These null portions are actually extremely important as they describe the periods where a BH or ultra-massive spacetime in formation is in equilibrium. These periods may have very large durations.}. Still, these MTTs satisfy an area law, for the area is always non-decreasing and strictly increasing outside the non-expanding horizon portions, and thus I will call them {\it generalized holographic screens} (GHS). The compact objects creating the intense gravitational field were destined to become BHs, but eventually something {\em more powerful} emerges and future infinity fully disappears: instead a universal future singularity arises. The spacetime becomes a contracting universe everywhere outside the spherical GHS. I called these spacetimes `ultra massive' because the area of MTSs goes {\it beyond} the maximal bound \eqref{lim}.
In this paper I prove that the above properties are actually universal, and {\em generically} all spacetimes in which the area bound \eqref{lim} is approached without limit by spatially stable MTSs develop GHSs with $\mathbb{R}\times \mathbb{S}^2$ topology, satisfying area laws and containing a distinguished MTS $\bar S$ that is stable in a null direction, with constant Gaussian curvature ${\cal K}=\Lambda$---and thus of maximal area $A_{\bar S}=4\pi/\Lambda$. All possible marginally trapped tubes passing through $\bar S$ change signature somewhere on $\bar S$ and are GHSs. A future singularity arises, and no BH is left over but rather an ultra-massive spacetime.\footnote{The only exception to the generic situation is given by spacetimes that are (locally) foliated by non-expanding horizons \cite{AK} including the family of near-horizon geometries that contain multiple Killing horizons with compact sections \cite{MPS,MPS1,MPS2,PLJ,LS,LSW}.}
To that end, I use the properties of the stability operator for MTSs introduced in \cite{AMS,AMS1} following pioneering work by Newman \cite{Ne}. I collect the necessary (known and new) results in section \ref{sec:L}. The main results are proven in section \ref{sec:main} for the generic situation.
The paper ends with a brief discussion in section \ref{sec:discussion}. A particularly important consequence of these results is that, at least in some cases, the merging of very big BHs may not lead to a larger BH but rather to an ultra-massive spacetime. The amount of mass required for this was estimated in \cite{Snew}, for spherically symmetric collapses, by using the present accepted value of $\Lambda \simeq 1.1 \times 10^{-52}$m$^{-2}$ based on the accelerated-expansion observations. If this estimation is approximately valid in general, at least $10^{8}$ big galaxies as massive as, say, M81 would be needed to collide and form the ultra-massive spacetime. Or equivalently, at least $10^{10}$ milky-ways. The total number of estimated galaxies in the observable Universe is about $10^{11}$, but the real number is probably larger.
But before anything, let me fix the set up and the main concepts to be used.
\subsection{Terminology and definitions}
$(M,g)$ denotes a 4-dimensional oriented Lorentzian manifold with causally oriented metric $g_{\mu\nu}$ of signature $(-,+,+,+)$ that satisfies Einstein field equations with cosmological constant $\Lambda$
\begin{equation}\label{efe}
G_{\mu\nu}+\Lambda g_{\mu\nu} = \frac{8\pi G}{c^4} T_{\mu\nu}:={\cal T}_{\mu\nu}
\end{equation}
where $G_{\mu\nu}$ is the Einstein tensor and $T_{\mu\nu}$ an energy-momentum tensor satisfying the dominant energy condition (DEC). For the purposes of this work, a closed spacelike surface $S$ is any embedded and orientable 2-dimensional connected and compact manifold with positive-definite first fundamental form $h_{AB}$\footnote{Greek indices are spacetime indices and Latin upper-case indices belong to the 2-dimensional $S$}. I denote by $\mathfrak{X}(S)$ and $\mathfrak{X}^\perp(S)$ the set of vector fields tangent and orthogonal to $S$, respectively. One can choose a couple of {\em null} vector fields in $\mathfrak{X}^\perp(S)$, denoted by $\vec \ell$ and $\vec k$, that are linearly independent all over $S$ \cite{Mars}. For convenience, I will choose them to be future pointing and normalized such that
$$
\ell_\mu k^\mu =-1.
$$
Let $\vec v,\vec w\in \mathfrak{X}(S)$, then one can split the covariant derivative $\nabla_{\vec v} \vec w$ into tangent and normal parts
$$
\nabla_{\vec v} \vec w =D_{\vec v} \vec w -\vec{K}(\vec v,\vec w)
$$
where $D$ is the intrinsic covariant derivative in $(S,h)$ and $\vec K$ is the shape tensor, also known as the second fundamental form vector \cite{O,Kri,S}. It can be decomposed into its trace and traceless parts relative to $h_{AB}$ as follows
$$
\vec{K}_{AB} = \vec\Sigma_{AB} +\frac{1}{2} \vec H h_{AB}, \hspace{1cm} h^{AB} \vec\Sigma_{AB} =0.
$$
Here $\vec\Sigma_{AB}$ is the total shear tensor \cite{CSV} and $\vec H$ is the mean curvature vector of $S$ in $M$ \cite{Kri,O,S}.
By definition, $\vec{K}(\vec v,\vec w), \vec\Sigma (\vec v, \vec w)$ and $\vec H$ are normal to $S$ and thus they can be expressed in the basis $\{\vec\ell,\vec k\}$:
$$
\vec K_{AB} =-K^\ell_{AB} \vec k-K^k_{AB} \vec\ell , \hspace{1cm} \vec\Sigma_{AB} =-\Sigma^\ell_{AB} \vec k-\Sigma^k_{AB} \vec\ell
$$
and
\begin{equation}\label{H}
\vec H=-\theta^\ell \vec k -\theta^k \vec \ell
\end{equation}
where $K^\ell_{AB} := \ell_\mu K^\mu_{AB}$ and $K^k_{AB}:=k_\mu K^\mu_{AB}$ are the null second fundamental forms, $\Sigma^\ell_{AB} := \ell_\mu \Sigma^\mu_{AB}$ and $\Sigma^k_{AB}:=k_\mu \Sigma^\mu_{AB}$ the null shear tensors, and $\theta^\ell :=\ell_\mu H^\mu$ and $\theta^k:=k_\mu H^\mu$ the null expansions of $S$ relative to $\vec\ell$ and $\vec k$ respectively.
The causal character of $\vec H$ leads to a classification into different types of surfaces, see e.g. \cite{AG,AMS,AMS1,MaSe,S,S0}. For our purposes, we just need to know the following (recall footnote \ref{foot1}): $S$ is said to be
\begin{itemize}
\item marginally `outer' trapped (MOTS) if $\vec H$ is null everywhere on $S$ consistently pointing along one of the null normals
\item marginally future trapped (MTS) if it is a MOTS with $\vec H$ future pointing all over $S$
\item weakly future trapped (WTS) if $\vec H$ is future pointing everywhere on $S$
\item future trapped (TS) if $\vec H$ is timelike and future pointing everywhere on $S$
\item (fully) untrapped if $\vec H$ is spacelike somewhere (everywhere) on $S$
\item Extremal if $\vec H$ vanishes identically on all of $S$
\end{itemize}
Equivalently, MOTSs have one of the null expansions vanishing, MTSs have in addition the other null expansion negative, WTSs and TSs have both null expansions non-positive and negative, respectively. In (fully) untrapped surfaces the expansions have opposite signs somewhere (everywhere) on $S$.
For MOTSs and MTSs the mean curvature vector points along one of the null normals ---the one with vanishing expansion---, and this is usually called the `outer' direction in the literature. But beware, `outer' has no outward meaning in principle. To fix the notation, for MOTSs and MTSs I am going to assume that $\vec k$ is the direction of the mean curvature vector, so that in this paper we are going to have
\begin{equation}\label{H}
\mbox{MOTS}: \vec H = -\theta^\ell \vec k, \hspace{1cm} \mbox{MTS}: \vec H = -\theta^\ell \vec k \, \, \, \& \, \, \, \theta^\ell <0.
\end{equation}
Notice that any MTS is, in particular, a MOTS.
A marginally (outer) trapped tube M(O)TT is a hypersurface in $M$ foliated by M(O)TSs. A spacelike MTT is usually called a dynamical horizon \cite{AK,AG} or future outer trapping horizon \cite{Hay}, while if it is timelike is called a timelike membrane \cite{AG,BBGV} or future inner trapping horizon \cite{Hay}.
\subsection{Spacetimes containing a stable marginally trapped surface}\label{subsec:possibilities}
A pioneering notion of stability of M(O)TSs was put forward by Newman \cite{Ne} and was used in \cite{Hay} to prove many relevant and useful results. An important advancement was the refinement performed in \cite{AMS,AMS1}, where for the first time an infinitesimal stability notion was found and the `stability operators' --to be defined later-- identified. See also an interesting discussion in \cite{PBH}. I am going to make extensive use of those results and the underlying ideas.
The basic notion is the {\em stability} of a MOTS along a given normal direction $\vec n\in \mathfrak{X}^\perp(S)$ with $n_\mu k^\mu \geq 0$ ---I will call these directions `external'. One can visualize such directions by imagining the local null hypersurface generated by $\vec k$, and then $\vec n$ points towards its exterior, or along it if $\vec n =\vec k$, see also figure \ref{fig:scheme} below.
\begin{defi}
A MOTS $S$ is said to be stable along an external $\vec n$ if there exists a non-negative function $f\not\equiv 0$ such that the variation of the vanishing expansion along $f\vec n$, denoted by $\delta_{f\vec n} \theta^k $, is non-negative. And it is called strictly stable if in addition $\delta_{f\vec n} \theta^k >0$ somewhere.
\end{defi}
To understand this nomenclature and the underlying conceptual scheme, note that if a MTS is embedded in a hypersurface $\Sigma$ and is strictly stable along the direction tangent to $\Sigma$ and orthogonal to $S$, then no nearby WTS extends to the `exterior' of $S$, and no nearby fully untrapped surface enters the interior of $S$, in $\Sigma$ \cite{AMS,AMS1} --recall that here exterior is defined by the direction of the null mean curvature vector. This is a kind of barrier property separating, within $\Sigma$, trapped from untrapped surfaces. Furthermore, if a MTS $S$ has this type of barrier property in $\Sigma$, then it is stable within $\Sigma$ \cite{AMS,AMS1}. It is in this sense that MTSs that are stable in spacelike directions are supposed to reside in local-in-time black-hole regions.
A deeper analysis of this stability concept will be implicit later in section \ref{sec:L}. For the present introductory purposes, let me just recall two fundamental results proven in \cite{AMS1}:
\begin{theorem}\label{th1}
Let $S_0$ be a MOTS strictly stable along a exterior-pointing normal direction $\vec n$. Choose a hypersurface $\Sigma_0$ tangent to $\vec n$ at $S_0$ such that $S_0$ is embedded in $\Sigma_0$, and a family of hypersurfaces $\{\Sigma_s\}$, $s\in[0,T]$, foliating the spacetime locally near $S_0$. Then, for some $\tau\in (0,T]$ there is a local MOTT adapted to the foliation, in the sense that each MOTS $S_s$ of the MOTT is embedded in $\Sigma_s$ for all $s\in [0,\tau)$ and strictly stable there. In addition, the MOTT thus built is never tangent to any of the $\Sigma_s$.
\end{theorem}
Obviously, the MOTT depends on the choice of reference foliation $\{\Sigma_s\}$. Notice moreover that, as long as the $S_s$ remain smooth and embedded, the `evolution' of the MOTT continues until either the MOTSs run out to infinity, or strict stability is lost. In the latter case we also have
\begin{theorem}\label{th2}
Under the same hypothesis of theorem \ref{th1}, assume that the MOTSs $\{S_s\}$ converge as $s\rightarrow \tau$ to a smooth closed MOTS $S_\tau$ that is stable, but not strictly stable, in $\Sigma_\tau$. If the function
\begin{equation}\label{W}
W:= {\cal T}_{\mu\nu}k^\mu k^\nu + \Sigma^k_{AB} \Sigma^k{}^{AB}
\end{equation}
is not identically zero on $S_\tau$, then the MOTT extends to all $s\in [0,\tau]$ and is tangent to $\Sigma_\tau$ at $S_\tau$.
\end{theorem}
Observe that the function $W$ is always non-negative (for this DEC is not needed, the mere null convergence condition is enough). This function $W$ plays an important role in what follows, as the possibility $W\equiv 0$ determines the exceptional cases.
In case that the reference foliation is composed by {\em non-timelike} hypersurfaces then the MOTT of the above theorems is spacelike everywhere near $S_0$ if $W|_{S_0} \not\equiv 0$, and is null on $S_0$ if $W$ vanishes identically there.
Therefore, every spacetime satisfying DEC and containing a MTS $S_0$ that is strictly stable in some {\em spacelike} external direction contains MTTs that are dynamical horizons around $S_0$ if $W|_{S_0} \not\equiv 0$ and must fall into one of the following categories:
\begin{enumerate}
\item All MTTs containing $S_0$ eventually become simple MOTTs--or reach an extremal surface--loosing the negativity of the non-vanishing expansion of the foliating MOTSs, or the M(O)TSs develop singularities, or in general something happens that stops the `evolution' of these MTTs.
\item evolution of (at least) some MTT continues indefinitely but the foliating MTSs have areas that never approach the bound \eqref{lim}. These include standard situations where, if there is a well defined future infinity $\mathscr{J}^+$ \cite{P00,HE,Wald,Fra}, a global event horizon encloses a BH and the MTT, and this either merges with, or asymptotically approaches, the global event horizon ---whose area is also bound by \eqref{lim}.
\item evolution of (at least) some MTT continues such that the foliating MTSs have areas that approach the bound \eqref{lim} as much as desired, but the entire MTT can never be contained in a compact set. These are marginal cases where the bound is never attained but is `reached' somehow at infinity. A particular explicit example of this behaviour is given by the toy model of figure 4 in \cite{Snew}, but note that in this example infinity has a very peculiar structure, and is reachable only by some special observers.
\item \label{pos1} evolution of (at least) some MTT continues such that the foliating MTSs have areas that approach the bound \eqref{lim} as much as desired, and the entire MTT is contained in a compact set. Several explicit examples of this case were presented in \cite{Snew}.
\item \label{pos2} there are finally the exceptional cases where the function $W\equiv 0$ for a set of MTSs that cover an open region of the spacetime and of the MTTs.
\end{enumerate}
Possibility \ref{pos1} is the one I wish to analyze in this work. This will be done in section \ref{sec:main} where I also provide a precise formulation of the type of spacetimes to be considered---ultra-massive spacetimes---in Definition \ref{ultraM}. For that, I will need some fundamental results on MTSs and MTTs that are presented in the next section. The exceptional situation \ref{pos2} will hopefully be studied in later work.
\section{The stability operator for MOTSs at work}\label{sec:L}
In this section I collect all results on MTSs and MTTs that I will need in what follows. Many of them are known and can be found in \cite{AMS,AMS1,AK,AG,BeS,BBGV,BE,BE1,Hay,JRD,Mars,Ne,PBH,SPrague,SERE,SILS,Simon}, some others are new as far as I know. The basic tool I am going to use is the {\em stability operator} for MOTSs, as introduced in \cite{AMS,AMS1}.
Let $S$ be a MOTS as explained before. The external pointing vector fields previously introduced are going to be characterized by their norms as follows, see the pictorial explanation in Figure \ref{fig:scheme}
\begin{equation}
\vec n =-\vec\ell +\frac{n_{\mu}n^{\mu}}{2}\vec k \label{n}
\end{equation}
so that the normalization with respect to the mean curvature vector is
\begin{equation}
H_\mu n^\mu = \theta^\ell \hspace{1cm} (k_\mu n^\mu=1) .\label{norm}
\end{equation}
The causal character of $\vec n$ is unrestricted. Let $\vec u\in \mathfrak{X}^\perp(S)$ be the `Hodge dual' of $\vec n$ in the normal bundle, that is, the unique vector field normal to $S$, orthogonal to $\vec n$ and with opposite norm:
\begin{equation}
\vec u = \vec\ell +\frac{n_{\mu}n^{\mu}}{2}\vec k, \hspace{1cm} u_\mu n^\mu =0, \hspace{1cm} u_\mu u^\mu =-n_\mu n^\mu .\label{u}
\end{equation}
\begin{figure}[h!]
\includegraphics[width=12cm]{PertScheme.pdf}
\caption{Scheme of the directions for a M(O)TS $S$. $\vec k$ and $\vec\ell$ are the two future null directions normal to $S$ and $\vec n \in \mathfrak{X}^\perp (S)$ is any possible direction pointing towards the shadowed zone, which is the {\em exterior} region for $S$. The choice of normalization is such that $k_\mu n^\mu =1$, so that the norm $n_\mu n^\mu$ of $\vec n$ chooses the particular direction. Of course, the direction of $\vec n$ depends on the point on $S$. Observe that the causal character of $\vec n$ is unrestricted and that $\vec k$ is the only normal direction not included in the set of vector fields $\vec n$.\label{fig:scheme}}
\end{figure}
The formula for the variation of the vanishing expansion $\theta^k=0$ reads
\begin{equation}
\delta_{f\vec n} \theta^k:=L_n f
\label{deltatheta}
\end{equation}
where $L_n$ is the {\em stability operator} for $S$ in the direction $\vec n$ explicitly given by \cite{AMS,AMS1}
\begin{equation}
L_n f =-\Delta f+2s^BD_{B}f+f\left({\cal K} -s^B s_{B}+D_{B}s^B-\frac{n^\rho n_{\rho}}{2}\Sigma^k_{AB} \Sigma^k{}^{AB}-G_{\mu\nu}k^\mu u^{\nu} \right)
\label{Ln0}
\end{equation}
where ${\cal K}$ is the Gaussian curvature on $S$, $\Delta$ its Laplacian, and $s_{B}$ the one-form on $S$ defined by
$$
\bm{s} (\vec v) =k_{\mu}\nabla_{\vec v}\ell^\mu, \hspace{1cm} \forall \vec v\in \mathfrak{X}(S) .
$$
Using the definiton of $W$ in \eqref{W}, an alternative form of $L_n$ is
\begin{equation}
L_n f =-\Delta f+2s^B D_{B}f+f\left({\cal K} -s^B s_{B}+D_{B}s^B-G_{\mu\nu}k^\mu \ell^{\nu}-\frac{n^\rho n_{\rho}}{2}\, W\right) \label{Ln}
\end{equation}
For positive $f>0$, yet another alternative way of writing these formulas reads \cite{JRD,Simon}
\begin{eqnarray}
\frac{L_n f}{f} &=& -\Delta \ln f +D_B s^B-\left(D_B \ln f -s_B\right)\left(D^B \ln f -s^B\right)+{\cal K}-\frac{n^\rho n_{\rho}}{2}\Sigma^k_{AB} \Sigma^k{}^{AB}-G_{\mu\nu}k^\mu u^{\nu}\nonumber\\
&=&-\Delta \ln f +D_B s^B-\left(D_B \ln f -s_B\right)\left(D^B \ln f -s^B\right)+{\cal K}-\frac{n^\rho n_{\rho}}{2}W-G_{\mu\nu}k^\mu \ell^{\nu} \label{Ln2}.
\end{eqnarray}
Formula (\ref{deltatheta}) is valid for all possible normal directions {\em except} for $\vec k$ (due to the normalization used). The variation in that distinguished null direction reads simply
\begin{equation}
\delta_{f\vec k} \theta^k= -f W, \hspace{5mm} \mbox{in particular}\hspace{5mm} \delta_{\vec H} \theta^k= \theta^\ell W . \label{deltak}
\end{equation}
Thus, the variation of the vanishing expansion along the corresponding {\em future} null direction is non-positive if the null convergence condition holds.
For each $\vec n$, $L_n$ is an elliptic operator on $S$, not self-adjoint in general (with respect to the $L^2$-product on $S$). It possesses a {\em real} principal eigenvalue, denoted by $\lambda_n$, and the corresponding {\em real} eigenfunction, denoted by $\phi_n$, can be chosen to be positive on all of $S$. The formal adjoint operator of $L_n$ with respect to the $L^2$-product is given by
$$
L^\dag_n f =-\Delta f-2s^B D_{B}f+f\left({\cal K} -s^B s_{B}-D_{B}s^B-\left.G_{\mu\nu}k^\mu \ell^{\nu}\right|_S-\frac{n^\rho n_{\rho}}{2}\, W\right)
$$
and has the same principal eigenvalue as $L_n$. The corresponding principal eigenfunctions are denoted by $\phi^\dag_n$, and they are real and positive on the entire $S$.
The first fundamental result that I am going to use is \cite{AMS,AMS1}
\begin{result}
A MOTS is (strictly) stable along the external direction $\vec n$ if and only if $\lambda_n$ is non-negative (positive).
\end{result}
Expression \eqref{Ln} implies that the (in)stability of $S$ is independent of the direction if $W\equiv 0$, it is in this sense that the case $W\equiv 0$ is exceptional. Therefore, it is useful to introduce the following definition.
\begin{defi}\label{Wset}
For any closed surface, ${\cal W}(S) \subset S$ is defined as the set of points in $S$ with vanishing function \eqref{W}:
$$
{\cal W}(S) := \{q\in S, \hspace{2mm} W|_q =0\}.
$$
\end{defi}
Concerning the stability along the mean curvature vector, or $\vec k$, the following holds.
\begin{result}\label{k}
For a MOTS $S$,
\begin{enumerate}
\item\label{Wnot0} if $S\setminus {\cal W}(S)\neq \emptyset$, $S$ is unstable along $\vec k$ and strictly stable along $-\vec k$.
\item\label{W=0} if $S= {\cal W}(S)$ then $S$ is stable along both $\pm\vec k$.
\end{enumerate}
\end{result}
\noindent{\em Proof.\/}\hspace{3mm} This follows immediately from \eqref{deltak}.
In fact, it is easily seen that the relation between stability operators along different directions $\vec n$ and $\vec n'$ is given simply by
\begin{equation}\label{twoL}
L_{n'} = L_n +\frac{W}{2} \left(n_\mu n^\mu -n'_\mu n'^\mu\right)
\end{equation}
making it plain that the (in)stability of $S$ is independent of $\vec n\neq \vec k$ when $W\equiv 0$. Therefore, the exceptional case with $W=0$ can be classified in three possibilities.
\begin{result}\label{excep}
Let $S$ be a MOTS with $S= {\cal W}(S)$, that is, with $W=0$ on all of $S$. Then only one of the following three cases can occur:
\begin{enumerate}
\item\label{allunsta} $S$ is unstable in all external directions and stable along $\pm\vec k$.
\item\label{allssta} $S$ is strictly stable in all external directions and stable along $\pm\vec k$
\item\label{allsta} $S$ is merely stable along all external directions including $\pm\vec k$.
\end{enumerate}
\end{result}
\noindent{\em Proof.\/}\hspace{3mm} The statement about $\pm\vec k$ is point \ref{W=0} in Result \ref{k}. From \eqref{twoL} we know that the stability operator is the same for all possible external directions (different from $\pm\vec k$) and actually equal to $L_{-\ell}$. Thus, depending on the the sign of $\lambda _{-\ell}(=\lambda_n$ for all $\vec n$) the three possibilities arise.\hfill\hbox{\vrule\vbox{\hrule\phantom{N}\hrule}\vrule}\,
These three cases arise when $S$ belongs to a non-expanding horizon (NEH), in particular for isolated horizons and Killing horizons \cite{AK}. Case \ref{allssta} is for black-hole type such horizons, while case \ref{allunsta} arises in situations such as the past cosmological horizon in de Sitter spacetime. The case \ref{allsta} is the truly exceptional possibility, and arises in regions of the spacetime foliated by NEHs, \cite{LS,LSW,PLJ} that include the multiple Killing horizons of near-horizon geometries \cite{MPS,MPS1,MPS2}.
From formula \eqref{twoL} the following relation can be derived \cite{AMS1}
$$
(\lambda_n -\lambda_{n'}) \int_S \phi^\dag_n \phi_{n'} = -\int_S \phi^\dag_n \phi_{n'} \frac{W}{2} \left(n_\mu n^\mu -n'_\mu n'^\mu\right)
$$
so that taking into account the positivity of the principal eigenfunctions one obtains the following set of results
\begin{result}\label{twon}
Assume that $S\setminus {\cal W}(S)\neq \emptyset$ on a MOTS $S$. Given any two normal (exterior pointing and normalized according to \eqref{norm}) directions $\vec n$ and $\vec n'$, the following list of results hold \cite{AMS1,JRD,Mars,SPrague,SERE}:
\begin{enumerate}
\item $\lambda_n -\lambda_{n'} +W\left(n_\mu n^\mu -n'_\mu n'^\mu\right)/2$ necessarily changes sign on $S$, or is identically zero;
\item\label{W=0} if $\vec n$ and $\vec n'$ differ exclusively within the region ${\cal W}(S)$, then $\lambda_n =\lambda_{n'}$;
\item\label{interweave} any two directions with the same principal eigenvalue $\lambda_n = \lambda_{n'}$ are such that $n_\mu n^\mu -n'_\mu n'^\mu$ either changes sign on $S$, or is non-zero only within ${\cal W}(S)$.
In the former case, $\vec n -\vec n'\propto \vec k$ necessarily changes sign, that is, causal orientation, on $S$;
\item\label{SupInf} $\lambda_{n'} + {\rm inf}_S \left[W\left(n'_\mu n'^\mu -n_\mu n^\mu\right)/2\right] \leq \lambda_n \leq \lambda_{n'} + {\rm sup}_S \left[W\left(n'_\mu n'^\mu -n_\mu n^\mu\right)/2\right]$
\item\label{l'>l} if $n'_\mu n'^\mu \leq n_\mu n^\mu $ everywhere outside ${\cal W}(S)$ and $n'_\mu n'^\mu < n_\mu n^\mu $ somewhere there, then $\lambda_{n'} > \lambda_n$;
\item\label{l'<l} if $\lambda_{n'} < \lambda_n$, then $n'_\mu n'^\mu > n_\mu n^\mu$ somewhere on $S\setminus {\cal W}(S)$.
\item if $S$ is not stable along the null direction $-\vec \ell$, then it cannot be stable along any spacelike direction.
\end{enumerate}
\end{result}
The last point implies that a necessary condition for the stability of MOTSs in spacelike external directions is its stability along the null direction $-\vec \ell$. Observe also that if a MOTS is stable along a spacelike $\vec n$, then it must be strictly stable along $-\vec\ell$.
\footnote{This also implies that the notion of {\em outer trapping surface} as defined by Hayward ---not to be confused with outer trapped surfaces!--- in \cite{Hay} is essentially equivalent to that of stability along some spacelike directions.}
Plugging $f=\phi_n$ into \eqref{Ln2} and using \eqref{efe} one gets
\begin{eqnarray}
\lambda_n &=& -\Delta \ln \phi_n +D_B s^B-\left(D_B \ln \phi_n -s_B\right)\left(D^B \ln \phi_n -s^B\right)+{\cal K}-\Lambda- \frac{n^\rho n_{\rho}}{2}\Sigma^k_{AB} \Sigma^k{}^{AB}-{\cal T}_{\mu\nu}k^\mu u^{\nu}\nonumber\\
&=&-\Delta \ln \phi_n +D_B s^B-\left(D_B \ln \phi_n -s_B\right)\left(D^B \ln \phi_n -s^B\right)+{\cal K}-\Lambda -\frac{n^\rho n_{\rho}}{2}W-{\cal T}_{\mu\nu}k^\mu \ell^{\nu}\label{lambdan}
\end{eqnarray}
so that integrating this on the compact $S$ and using the Gauss-Bonnet theorem
\begin{eqnarray}
\left(\Lambda +\lambda_n \right)A_S &=&4\pi (1-g) -\int_S\left[\left(D_B \ln \phi_n -s_B\right)\left(D^B \ln \phi_n -s^B\right)+ \frac{n^\rho n_{\rho}}{2}\Sigma^k_{AB} \Sigma^k{}^{AB}+{\cal T}_{\mu\nu}k^\mu u^{\nu}\right]\nonumber \\
&=&4\pi (1-g) -\int_S\left[\left(D_B \ln \phi_n -s_B\right)\left(D^B \ln \phi_n -s^B\right)+ \frac{n^\rho n_{\rho}}{2}W+{\cal T}_{\mu\nu}k^\mu \ell^{\nu}\right]\label{genineq}
\end{eqnarray}
where $A_S$ is the area of $S$ and $g$ its genus. Therefore, if the DEC holds and the normal direction $\vec n$ is non-timelike everywhere, the integrand on the righthand side is non-positive and one gets the bound \eqref{lim} together with some other consequences \cite{GSc,HSN,JRD,Mars,Simon,W}.
\begin{result}\label{areas}
Assume that DEC holds and $\Lambda \neq 0$
in a spacetime that satisfies the Einstein field equations \eqref{efe} and let $S$ be a MOTS with $S\setminus {\cal W}(S)\neq \emptyset$, ${\cal W}(S)$ as introduced in definition \ref{Wset}. Then
\begin{enumerate}
\item \label{ineq}For any external direction $\vec n$ that is non-timelike on $S\setminus {\cal W}(S)$, the following inequality holds
$$
\left(\Lambda +\lambda_n \right)A_S\leq 4\pi (1-g)
$$
with equality possible only if
\begin{equation}\label{zeros}
s_B =D_B \ln \phi_n, \hspace{1cm} (n_\mu n^\mu) \Sigma^k_{AB} \Sigma^k{}^{AB}= 0, \hspace{1cm} T_{\mu\nu} k^\mu u^\nu=0
\end{equation}
\item\label{ineq1} In particular, if $S$ is stable ($\lambda_n\geq 0$) along an external $\vec n$ that is non-timelike on $S\setminus {\cal W}(S)$ then
$$
\Lambda A_S\leq 4\pi (1-g)
$$
equality requiring $\lambda_n=0$ and \eqref{zeros}
\item\label{topol} $S$ cannot be stable in external directions that are non-timelike on $S\setminus {\cal W}(S)$ if $\Lambda >0$ and $g>0$.
\item\label{Alarger} If $\Lambda >0$ and the area $A_S > 4\pi /\Lambda$, then $\lambda_n <0$ necessarily along all external $\vec n$ that are non-timelike on $S\setminus {\cal W}(S)$. In particular $\lambda_{-\ell} <0$.
\item\label{Aequal} If $\Lambda >0$ and $A_S= 4\pi /\Lambda $, then $\lambda_n \leq 0$ in all external directions that are non-timelike on $S\setminus {\cal W}(S)$.
\item\label{Aequal1} If $\Lambda >0$, $A_S= 4\pi/\Lambda $ and $\lambda_n =0$ along an external $\vec n$ that is non-timelike on $S\setminus {\cal W}(S)$, then \eqref{zeros} hold and $S$ is a metric round sphere of constant curvature ${\cal K}=\Lambda$. In this case, the energy-momentum tensor must take the following form on $S$:
\begin{equation}\label{emt}
T_{\mu\nu}|_S = a^2 k_\mu k_\nu + b^2 \ell_\mu \ell_\nu +C_{\mu\nu}
\end{equation}
where $C_{\mu\nu}$ is a symmetric tensor field tangent to $S$ ($k^\mu C_{\mu\nu}=0=\ell^\mu C_{\mu\nu}$) such that
\begin{equation}\label{dec}
a^2b^2 \geq C_{\mu\rho} C_{\nu}{}^\rho v^\mu v^\nu
\end{equation}
for {\em all} unit vectors $\vec v \in \mathfrak{X}(S)$. At the points where $\vec n$ is spacelike ($n_\mu n^\mu >0$ there) one further has $b=0$ and $C_{\mu\nu}=0$. Moreover, $b=0$ and $C_{\mu\nu}=0$ on ${\cal W}(S)$.
\end{enumerate}
\end{result}
\begin{remark}
Note that the last in \eqref{zeros} always includes $T_{\mu\nu}k^\mu \ell^\nu =0$ due to DEC.
\end{remark}
\noindent{\em Proof.\/}\hspace{3mm} Statements \ref{ineq} through \ref{Aequal} are direct consequences of \eqref{genineq}.
To prove \ref{Aequal1}, first note that \eqref{zeros} hold due to statement \ref{ineq1}. Introducing \eqref{zeros} and $\lambda_n=0$ into \eqref{lambdan} one has
$$
0= -\Delta \ln\phi_n +\Delta \ln \phi_n +{\cal K} -\Lambda
$$
so that the Gaussian curvature of $S$ is constant and positive, ${\cal K}=\Lambda$ (and thus $g=0$). To deduce the form \eqref{emt} for $T_{\mu\nu}$ I repeatedly use the properties of tensors satisfying the dominant property, see e.g. \cite{BS}. As $u^\mu$ is future, $T_{\mu\nu}u^\nu$ must be past directed, and the last in \eqref{zeros} readily implies $T_{\mu\nu}u^\nu=-a^2 k_\mu$ for some $a$. Similarly, $T_{\mu\nu}k^\mu$ is past directed, so that the last in \eqref{zeros} again provides $T_{\mu\nu}k^\nu =-b^2 \ell_\mu$, and moreover $b$ must vanish at any point where $\vec u$ is timelike, that is, at any point where $\vec n $ is spacelike: $b^2 n_\mu n^\mu =0$. The structure \eqref{emt} follows. Take then any other future null direction $\vec M$ on $S$, that is
$$
\vec M = \vec v + x\vec k + y\vec\ell, \hspace{1cm} v_\mu v^\mu =1, \hspace{1cm} 2xy =1.
$$
Then, DEC states that $T_{\mu\nu} M^\nu = -a^2 y k_\mu -b^2 x \ell_\mu +C_{\mu\nu} v^\nu$ is past directed, and its non-positive norm leads to \eqref{dec}. On ${\cal W}(S)$ $T_{\mu\nu}k^\mu k^\nu =0$, and thus $b=0$ follows too. \hfill\hbox{\vrule\vbox{\hrule\phantom{N}\hrule}\vrule}\,
Let us now consider MOTTs ${\cal H}$ containing a given MOTS $S$ with $S\setminus {\cal W}(S) \neq \emptyset$. As the MOTS $S$ is embedded in ${\cal H}$, taking appropriate deformations of $S$ along the direction tangent to the MOTT and normal to $S$, say $\vec m\in \mathfrak{X}^\perp(S) \cap \mathfrak{X}(\cal H)$, will lead to another MOTS of the MOTT. Hence, along this direction $\lambda_m=0$, and the deformation leading to another MOTS is given by the vector field $\phi_m \vec m$. In other words, the possible directions of MOTTs containing a given MOTS $S$ with $S\setminus {\cal W}(S) \neq \emptyset$ are included in the set of directions with vanishing eigenvalue. Then, point \ref{interweave} in Result \ref{twon} immediately provides the next result \cite{AG,BeS,SPrague,SERE}.
\begin{result}\label{inter}
Any two MOTTs containing the same MOTS $S$ (such that $W|_S \not \equiv 0$) necessarily interweave each other near $S$ in such a way that their tangent vector fields orthogonal to the MOTS and subject to \eqref{norm}, say $\vec m$ and $\vec m'$, are such that $\vec m -\vec m' \propto \vec k$ changes causal orientation on $S$ or is only non-vanishing within ${\cal W}(S)$.
\end{result}
\subsection{Existence of (in)stability directions}
The number of directions with vanishing eigenvalue is huge \cite{SPrague,SERE}, and they contain many possible MOTTs through the same MOTS $S$ as long as $S$ is stable is some external direction $\vec n$. The existence of these many MOTTs through $S$ follows from Theorem \ref{th1} by changing the reference foliation at will. The question of whether or not a given MOTS can be unstable in {\em all} possible external directions can be easily answered in the negative if $W\not\equiv 0$ \cite{Mars,AMS1}
\begin{result}\label{always}
Assume the null convergence condition holds and let $S$ be a MOTS with $S\setminus {\cal W}(S) \neq \emptyset$. There always exist external stability directions, as well as external directions into which $S$ is strictly stable and external directions into which $S$ is unstable.
\end{result}
\noindent{\em Proof.\/}\hspace{3mm} From point \ref{SupInf} in Result \ref{twon} we know that
$$
\lambda_n \leq {\rm sup}_S \left[W(n'_\mu n'^\mu -n_\mu n^\mu)/2 \right] +\lambda_n'
$$
so that choosing $\vec n$ sufficiently close to $\vec k$, that is, choosing $n_\mu n^\mu$ large enough and as $W>0$ somewhere on $S$, the righthand side can be made negative, and thus $\lambda_n <0$ for such a choice. Similarly, using now the other inequality side in the same point \ref{SupInf} of Result \ref{twon} and choosing $n_\mu n^\mu$ negative enough one can achieve $\lambda_n >0$ for some $\vec n$. To prove that there always exist directions with $\lambda_n =0$, start with a direction $\vec n'$ of strict stability, and use Theorem \ref{th1} to construct an adapted MOTT containing $S$. As explained above the direction $\vec m$ tangent to the MOTT and orthogonal to $S$ has a vanishing variation $\delta_{\phi_m \vec m} \theta^k =0$ of the vanishing expansion and thus $\lambda_m=0$.\hfill\hbox{\vrule\vbox{\hrule\phantom{N}\hrule}\vrule}\,
Actually, many (in)stability directions can be explicitly identified in the case that $W\neq 0$ all over $S$.
\begin{result}\label{L+-}
Assume the null convergence condition holds and let $S$ be a MOTS with ${\cal W}(S) =\emptyset$. Then
\begin{enumerate}
\item\label{alwaysstab} $S$ is strictly stable along all normal directions with
$$
Wn_\mu n^\mu\leq 2\left( {\cal K}-s_B s^B + |D_B s^B| -G_{\mu\nu}k^\mu \ell^\nu\right) \hspace{5mm} \mbox{(and $<$ somewhere)}
$$
\item\label{alwaysinestab} And $S$ is unstable along all normal directions with
$$
Wn_\mu n^\mu\geq 2\left( {\cal K}-s_B s^B - |D_B s^B| -G_{\mu\nu}k^\mu \ell^\nu\right) \hspace{5mm} \mbox{(and $>$ somewhere)}
$$
\item\label{K>Lambda} If $S$ is stable along a non-timelike direction and DEC holds, then $S$ must have at least a point $q\in S$ where ${\cal K}|_q \geq \Lambda$. And if $\vec n$ is spacelike somewhere then ${\cal K}|_q >\Lambda$ is actually necessary.
\end{enumerate}
\end{result}
\noindent{\em Proof.\/}\hspace{3mm}
In \cite{SERE,SPrague} it was proven that, if $W\neq 0$, the two external directions $\vec n_\pm$ defined by
$$
W n_\pm{}_\mu n_\pm^\mu = 2\left( {\cal K}-s_B s^B \pm D_B s^B -G_{\mu\nu}k^\mu \ell^\nu\right)
$$
have vanishing principal eigenvalues $\lambda_{n_\pm} =0$, ergo these are explicit examples of stability directions. Then, Result \ref{twon} implies that every $\vec n$ with $n_\mu n^\mu \geq n_\pm{}_\mu n_\pm^\mu $ (and $>$ somewhere) will be unstable directions, while those with with $n_\mu n^\mu \leq n_\pm{}_\mu n_\pm^\mu $ (and $<$ somewhere) will be directions of strict stability, so that points \ref{alwaysstab} and \ref{alwaysinestab} follow. To prove point \ref{K>Lambda} assume on the contrary that ${\cal K} < \Lambda $. Then, using the field equations \eqref{efe} and point \ref{alwaysinestab} instability of $S$ along all directions $\vec n$ with
$$
Wn_\mu n^\mu\geq 2\left( {\cal K}-s_B s^B - |D_B s^B| -{\cal T}_{\mu\nu}k^\mu \ell^\nu -\Lambda \right)< 0
$$
would follow, in contradiction.
\hfill\hbox{\vrule\vbox{\hrule\phantom{N}\hrule}\vrule}\,
The stability operators corresponding to $\vec n_\pm$ read
$$
L_{n_+}f= -\Delta f+2s^B D_B f, \hspace{1cm} L_{n_-}f =-\Delta f +2s^BD_B f +2fD_B s^B
$$
and they have some specific interesting properties \cite{SPrague,SERE}. Notice in particular that the principal eigenfunction $\phi_{n_+}$ is a constant.
Identifying specific stability directions when ${\cal W}(S)$ is non-empty is a more difficult task.
\subsection{Area law: generalized holographic screens}
Concerning the variation of area for MOTSs foliating a MOTT, let $\epsilon_{AB}$ be the canonical area 2-form on $(S,h)$. From classical results in semi-Riemannian geometry, the variation of $\epsilon_{AB}$ along any possible direction defined by a vector field $\vec \xi$ is ruled by the mean curvature vector as follows (see e.g. \cite{Kri,O,MaSe})
$$
\delta_\xi \epsilon_{AB} = \epsilon_{AB} \left(D_B \xi^B_\parallel +H_\mu \xi^\mu \right)
$$
where $\vec\xi_\parallel$ is the part of $\vec\xi$ tangent to $S$. In particular, the total variation of area is thus
$$
\delta_\xi A_S = \int_S H_\mu \xi^\mu .
$$
Note that only the normal part of $\vec\xi$ enters in this relation. Thus, restricting to $\vec\xi\in \mathfrak{X}^\perp (S)$ and recalling \eqref{H}, the variations become
\begin{equation}
\delta_\xi \epsilon_{AB} =- \epsilon_{AB} \theta^\ell (k_\mu \xi^\mu), \hspace{1cm} \delta_\xi A_S =-\int_S \theta^\ell (k_\mu \xi^\mu).
\end{equation}
Recall that the external vector fields $\vec n$ that we are using in the stability operator have $n_\mu k^\mu =1$. This leads to a result, the area law, for MTTs \cite{Hay,AK,BE,BE1}.
\begin{result}\label{area}
The area element $\epsilon_{AB}$ and the total area of the MTSs foliating any MTT are everywhere non-decreasing along the external directions tangent to the MTT and orthogonal to the MTSs, and can only remain stationary at the regions where the MTT is tangent to the mean curvature vector of the MTSs, that is, tangent to $\vec k$.
\end{result}
This is the basic property of ``holographic screens'' as defined in \cite{BE,BE1}, see also \cite{SW}. In those references the authors required that $W\neq 0$ everywhere and that the stationary set was with empty interior. These assumptions are not necessary, and one can accept open portions where the MTT is a non-expanding (or isolated or Killing \cite{AK}) horizon tangent to $\vec k$. Explicit examples are those of Figures 2, 3 and 6 in \cite{Snew}. Thus, I will use the term ``generalized holographic screens'' for these MTTs. I would like to remark that if the MTT is tangent to the mean curvature vector --and therefore to $\vec k$-- of the foliating MTSs in an open portion, then from \eqref{deltak} it follows that $W\equiv 0$ on that portion leading to the exceptional case. Observe that the area remains constant in these portions.
Let me finally recall another important result, that provides a deeper meaning to the previous one \cite{AG,BE,BE1}.
\begin{result}
The foliation by MTSs of any MTT is unique, except in open regions where the MTT is null and tangent to the mean curvature vector of the MTSs.
\end{result}
\section{Main results: ultra massive spacetimes}\label{sec:main}
We are now ready to analyze possibility \ref{pos1} as presented in section \ref{subsec:possibilities} of the Introduction.
To that end, it is necessary to provide a more rigourous definition of what Possibility \ref{pos1} requires.
\begin{defi}[Ultra-massive spacetimes]\label{ultraM}
A spacetime will be called ultra-massive if it contains a dynamical horizon foliated by MTSs $\{S_s\}$ ($s\in [0,\tau)$) with the following properties
\begin{itemize}
\item All of them are such that the set ${\cal W}(S_s)$ is not the entire $S_s$
\item For any positive $\epsilon >0$, some of the foliating MTSs have areas larger than $4\pi/\Lambda -\epsilon$.
\end{itemize}
\end{defi}
The second condition here captures the idea that the areas of the MTSs in the dynamical horizon approach the bound \eqref{lim} indefinitely.
\begin{remark}
Observe that it would actually be sufficient to assume the existence of the MTS $S_0$ and its stability along a spacelike direction because then theorem \ref{th1} applies and, choosing the reference foliation $\{\Sigma_s\}$ composed by non-timelike hypersurfaces, the resulting MOTT will be spacelike and foliated by MTSs---as the expansion $\theta^\ell <0$ on $S_0$. Thus, the MOTT will be a dynamical horizon locally. The true assumptions in definition \ref{ultraM} are that the $\{S_s\}$ remain strictly stable along spacelike directions while at the same time their areas increase approaching the bound \eqref{lim}.
\end{remark}
\begin{theorem}
Let the spacetime be ultra-massive in the sense of definition \ref{ultraM} and assume that $T_{\mu\nu}$ satisfies the dominant energy condition and $\Lambda >0$.
Then
\begin{itemize}
\item $S_0$ belongs to a generalized holographic screen ${\cal H}$ that satisfies the area law
\item The topology of ${\cal H}$ is $\mathbb{R}\times \mathbb{S}^2$
\item There is a distinguished MTS $\bar S\in {\cal H}$ with constant Gaussian curvature ${\cal K} =\Lambda$ ---and area $4\pi /\Lambda$.
\item $T_{\mu\nu}|_{\bar S}$ takes the form \eqref{emt}.
\item All GHSs passing through $S_0$ change signature somewhere at $\bar S$. At least one GHS is null everywhere on $\bar S$.
\item If $\bar S$ is not extremal and $\bar S\setminus {\cal W}(\bar S)\neq \emptyset$, all such GHSs become (partly) timelike towards the past of $\bar S$, and continue all along towards the past as far as the foliating surfaces keep being marginally trapped (and the genericity condition $W\not\equiv 0$ holds on them)
\end{itemize}
\end{theorem}
\noindent{\em Proof.\/}\hspace{3mm}
Given the strict stability of $S_0$ along a spacelike external direction its topology must be $\mathbb{S}^2$ (point \ref{topol} in Result \ref{areas}) and a reference foliation composed by non-timelike hypersurfaces $\{\Sigma_s\}$ can be chosen adapted to the dynamical horizon of definition \ref{ultraM} according to Theorem \ref{th1}. Thus, the topology of this DH is $\mathbb{R}\times \mathbb{S}^2$.
From theorem \ref{th2}, the DH extends to $s=\tau$ as a MTT whenever $S_\tau :=\bar S$ is not extremal, and also this MTT is tangent to $\Sigma_\tau$ everywhere on $\bar S$ if $\bar S$ is just stable (but not strictly stable) there.
Let $\vec n_s$ be the non-timelike directions of strict stability for each $S_s$, $s\in[0,\tau)$, tangent to the hypersurfaces $\Sigma_s$. Point \ref{ineq} in Result \ref{areas} implies that
$$(\Lambda +\lambda_{n_s})A_{S_s}\leq 4\pi, \hspace{3mm} \forall s\in [0,\tau).
$$
where $\lambda_{n_s}$ denote the principal eigenvalue of each $S_s$ along the external direction $\vec n_s$, and thus $\lambda_{n_s} >0$ for all $s\in [0,\tau)$, so that $\Lambda A_{S_s} <4\pi $ for all $s\in [0,\tau)$.
As the areas increase monotonically with $s$ due to Result \ref{area}, it follows that necessarily in the limit
$$
A_{\bar S} = 4\pi /\Lambda
$$
as otherwise (if $A_{\bar S} < 4\pi /\Lambda$) there would exist some $\epsilon >0$ such that all the $S_s$ of the DH will have an area strictly smaller than $4\pi/\Lambda -\epsilon$, in contradiction.
Furthermore, as $\vec n_s$ are non-timelike for all $s\in [0,\tau)$, in the limit the stability direction $\vec n_\tau$ tangent to $\Sigma_\tau$ is non-timelike everywhere on $\bar S$. Point \ref{Aequal} in Result \ref{areas} ensures then that $\bar S$ cannot be strictly stable within $\Sigma_\tau$, and thus it has to be just stable so that at $\bar S$
$$\lambda_{n_\tau}=0$$
and the MTT is tangent to $\Sigma_\tau$ everywhere on $\bar S$.
Point \ref{Aequal1} in Result \ref{areas} then ensures ${\cal K}=\Lambda$ and the form \eqref{emt} of $T_{\mu\nu}$, plus \eqref{zeros}. Hence, if $\vec n_\tau$ is spacelike somewhere on $\bar S$, $\Sigma^k_{AB}=0$ and $W=0$ there. In other words, $\vec n_\tau$ is null and proportional to $-\vec \ell$ everywhere on $\bar S\setminus {\cal W}(\bar S)$ and, due to the construction, non-timelike on ${\cal W}(\bar S)$. This implies also that $\Sigma_\tau$ is null everywhere on $\bar S\setminus {\cal W}(\bar S)$ and, as the MTT is tangent to $\Sigma_\tau$ on $\bar S$, the MTT changes signature there.
Moreover, using now point \ref{W=0} of Result \ref{twon} one deduces that on $\bar S$
$$\lambda_{-\ell}=0.$$
This implies that if the reference foliation is such that $\Sigma_\tau$ is null everywhere on $\bar S$, the corresponding MTT changes signature everywhere on $\bar S$ and is null everywhere on $\bar S$.
As $\bar S$ is stable along $-\vec\ell$, then if ${\cal W}({\bar S}) \neq \bar S$ from point \ref{l'>l} in Result \ref{twon} $\bar S$ is strictly stable in all external directions that are non-spacelike on $\bar S\setminus {\cal W}(\bar S)$ and timelike somewhere there. In particular, $\bar S$ is strictly stable in all external timelike directions. Choosing an appropriate reference foliation, for instance with timelike hypersurfaces $\{\Sigma'_s\}$, with $\bar S\subset \Sigma'_0$, theorem \ref{th1} states that there is a MOTT adapted to $\{\Sigma'_s\}$ beyond $\bar S$ containing it, and this MOTT will in fact be a MTT if $\bar S$ is not extremal. Any other MTT passing through $\bar S$ will be given by other external directions $\vec n'$ on $\bar S$ with $\lambda_{n'} =0$, and thus point \ref{interweave} in Result \ref{twon} states that either $\vec n'$ is different from $-\vec \ell$ only on ${\cal W}(\bar S)$, or
$$-\vec\ell -\vec n' =-\frac{n'^\mu n'_\mu}{2} \vec k
$$
changes causal character within $\bar S \setminus {\cal W}(\bar S)$, ergo $\vec n'$ has all 3 causal characters on $\bar S \setminus {\cal W}(\bar S)$. Thus, signature change always happens at some points of $\bar S$ for all MTTs that contain it.
These MTTs going beyond $\bar S$ cannot become dynamical horizons. This follows from the area law (Result \ref{area}) because the area of the MTSs $\{S'_s\}$ embedded in each $\Sigma'_s$ will be larger than $4\pi/\Lambda $ and thus point \ref{Alarger} of Result \ref{areas} implies that $\lambda_{-\ell} <0$ for all $S'_s$, so that point \ref{l'<l} in that Result \ref{areas} states that any other external direction $\vec m'$ with $\lambda_{\vec m'}=0$ must be timelike somewhere on $S'_s\setminus {\cal W}(S'_s)$. In particular the direction tangent to any of the MTTs and orthogonal to the $S'_s$ has to be timelike somewhere on $S'_s$. This behaviour must continue along the MTTs as long as the foliating surfaces remain marginally trapped and the genericity condition holds.
\hfill\hbox{\vrule\vbox{\hrule\phantom{N}\hrule}\vrule}\,
\begin{remark}
If any one of the $S'_s$, say $S'_{s_1}$, has ${\cal W}(S'_{s_1})=\emptyset$, then (at least) some GHSs will be timelike everywhere there. To prove it, choose a {\em timelike} deformation vector $\vec n$ with $W n^\mu n_\mu =2 \lambda_{-\ell} <0$ on $S'_{s_1}$ so that, as $\lambda_{-\ell}$ is constant
$${\rm sup}_{S'_{s_1}} \left(-W n_\mu n^\mu/2\right)={\rm inf}_{S'_{s_1}} \left(-Wn_\mu n^\mu/2\right)=-\lambda_{-\ell}$$
and thus point \ref{SupInf} in Result \ref{twon} applied to $\vec n' =-\vec\ell$ provides
$$
0\leq \lambda_n \leq 0 \Longrightarrow \lambda_n =0.
$$
Thus, some MTTs passing through $S'_{s_1}$ will be tangent to $\vec n$ and/or other nearby timelike vector fields everywhere on $S'_{s_1}$.
\end{remark}
To fix ideas, let me use ${\cal H}$ for the specific GHS that is null everywhere at $\bar S$. One may wonder what the `final fate' of ${\cal H}$ may be when proceeding towards the past with increasing areas of the foliating MTSs `after' having left $\bar S$ `behind'. One possibility is that eventually the MTSs become simple MOTSs. A more interesting possibility comes about if eventually a MTS $S_f$ with ${\cal W}(S_f)=S_f$ is reached. Then, one deduces that we must be in the situation of point \ref{allunsta} in Result \ref{excep}. This follows from the fact that, as explained in the proof of Result \ref{always}, the stability directions of the previous MTSs require external directions $\vec n$ closer and closer to $-\vec k$ as $W$ approaches becoming identically zero along them. After $S_f$, ${\cal H}$ may turn back to be a local timelike membrane or keep being null. In either case the `evolution' with non-decreasing area goes indefinitely towards the past. Examples of such situations are provided by the Figures 2 and 3 in \cite{Snew}. Observe, moreover, that other null portions of ${\cal H}$ may have also existed `before' reaching $\bar S$. For instance, prior to $S_0$ there may be a portion of a NEH joining the DH containing $S_0$. In this case, and for an analogous reasoning as before, the MTSs foliating these NEH portions must belong to the category of point \ref{allssta} in Result \ref{excep}. An explicit example is the spacetime of Figure 3 in \cite{Snew}.
Standard results on geodesic incompleteness \cite{HE,P,S1,SMilestone} imply the existence of a future singularity in ultra-massive spacetimes due to the existence of the future trapped surfaces. It is enough to assume, for instance, the existence of a Cauchy hypersurface $\Sigma$ containing one of the many future-trapped surfaces such that the part of $\Sigma$ external to the trapped surface is non-compact, see also section 7 in \cite{AMMS}.
\section{Discussion}\label{sec:discussion}
The results shown above may look a little counter-intuitive at first, because the `evolution' of the GHS ${\cal H}$ indefinitely towards the past implies the existence of MTSs that are unstable in all spacelike directions at the far past (or as initial data in $\mathscr{J}^-$). However, one must keep in mind that we are dealing with spacetimes with positive $\Lambda$, and thus they may well tend to behave like de Sitter in the far past.
The implications of the existence and properties of these ultra-massive spacetimes are reminiscent of the pioneering studies performed years ago in \cite{NYM,SNKM,BHKT,NSH} both theoretically and numerically, and also using specific explicit spacetimes---the Kastor-Traschen charged multi black hole solution \cite{KT}. The words `overmassive' \cite{BHKT} and super-critical \cite{NSH} actually already appeared there. However, the maximum bound \eqref{lim} for spatially stable MTSs is never achieved in the spacetimes analyzed in those references due to the existence of electric charge (that effectively implies a smaller bound \cite{HSN,Simon}). Nevertheless, the conclusions concerning the possibility of black hole mergers are similar: it seems that black holes with areas close to the bound \eqref{lim} will not produce a new black hole if they collide but rather something else like, for instance, an ultra-massive spacetime. The same may happen if a greedy black hole accretes too much mass from its surroundings.
When the cosmological constant is positive, and if the spacetime admits a conformal completion with past null infinity $\mathscr{J}^-$, we know \cite{Frie0} that the entire future evolution of the spacetime is encoded in initial data at $\mathscr{J}^-$. Therefore, the universal properties of ultra-massive spacetimes can be characterized by some specific form of the initial data at $\mathscr{J}^-$. It would be interesting to unveil the properties of such data, which might shed some light onto the definition of mass-energy at infinity when $\Lambda >0$. In any case, this way of looking at ultra-massive spacetimes resolves any puzzles concerning the anti-intuitive behaviour of the GHS ${\cal H}$ in its timelike part `after' the bound \eqref{lim} has been reached at the distinguished MTS $\bar S$, because the true future evolution of ${\cal H}$ is already determined at $\mathscr{J}^-$ as long as ${\cal H}$ remains in the domain of dependence of $\mathscr{J}^-$.
An important remark is that the numbers provided in the Introduction and in \cite{Snew} for the total mass of ultra-massive spherically symmetric spacetimes do not take into account the expansion of the Universe, which may change the picture in several respects, see also \cite{ZCGHSW}.
It would be very interesting to identify explicit examples of ultra-massive spacetimes without spherical symmetry, for its own sake and also to gain a better intuition about their general properties and behaviours. In particular, another important question that arises is whether or not ultra-massive spacetimes may have an event horizon of an asymptotically de Sitter future infinity, or they will always lead to a universal singularity as in the examples presented in \cite{Snew}.
In this paper I have concentrated in the {\em future} case, that is to say, concerning marginally future-trapped surfaces. However, the past case could also lead to interesting conclusions. Observe that in that case the area of the MTSs along past GHS will actually increase towards the future.
\section*{Acknowledgments}
I am grateful to Marc Mars for many helpful comments and to Ken-ichi Nakao from bringing important references to my attention.
Research supported by the Basque Government grant number IT1628-22, and by Grant PID2021-123226NB-I00 funded by the Spanish MCIN/AEI/10.13039/501100011033 together with ``ERDF A way of making Europe'' . This research was carried out during a visiting professorship at Yukawa Institute for Theoretical Physics.
|
\section{Introduction}
For all real number $ \tau $ ,the gamma function $\Gamma\left({\tau} \right)$ is expressed as
\begin{equation}
\Gamma\left({\tau} \right)=\displaystyle\int_0^\infty e^{-t} \ t^{\tau-1} \, dt
\end{equation} , which is an improper integral and it satisfies the following properties:
1. $\Gamma\left({n+1} \right)=n!$, $ n \in N$
2. $\Gamma\left({n+1} \right)=n \Gamma\left({n} \right)$
for each real number $n\not\in \{0,-1,-2,-3,....\}$.\
\noindent It is obvious by previous knowledge based on sequence spaces the space of all real valued sequences are denoted by $\omega $. The spaces $c_{0} ,c $ and $l_{\infty }$ are of all null, convergent and bounded sequences respectively.\
Maddox \cite{madx1,madx2,madx3} introduced $c_{0}\left(p\right) ,c\left(p\right) $ and $l_{\infty }\left(p\right) $ for $p=\left(p_{k}\right) $ a bounded sequence of strictly positive real numbers, as:
\[\begin{array}{l} {c_{0}\left(p\right) =\left\{\zeta=\left(\zeta_{k} \right)\in \omega :{\mathop{\lim }\limits_{k\to \infty }} \left| \zeta_{k} \right|^{p_{k}} =0 \right\}}, \\
{c\left(p\right) =\left\{\zeta=\left(\zeta_{k} \right)\in \omega :{\mathop{\lim }\limits_{k\to \infty }} \left| \zeta_{k} -l\right|^{p_{k}} =0 \ for some \ l\in R \right\}}, \\
{l_{\infty }\left(p\right) =\left\{\zeta=\left(\zeta_{k} \right)\in \omega :{\mathop{\sup }\limits_{k\in N}} {\left| \zeta_{k} \right|}^{p_{k} } <\infty \right\}} \end{array}\] \
and these are complete paranormed sequence spaces with paranorm
$ g\left(x\right) = {\mathop{\sup }\limits_{k\in N}}\ {\left| x_{k} \right|}^{{p_{k} }/M } $ and where $ M=max \{1,\displaystyle \sup_{k} \ p_{k}\} $. The difference sequence spaces $X\left(\Delta \right)=\left\{x=\left(x_{k} \right):\Delta \left(x_{k} \right)\in X\right\}$ for $X=\left\{c_{0}, c, l_{\infty } \right\}$ was introduced in 1981 by K1zmaz \cite{kizm15}. Then the difference sequence spaces attracted the attention of several authors \cite{basa6,colak1,mur1} in different directions. For a proper fraction ${\tau}$, Baliarsingh \& Dutta (\cite{bali4,bali5,bali6,bali7,dutt10,dutt11}) they introduced the difference operator $\Delta ^{\left( {\tau}\right) } $ as
\begin{equation} \label{1.2}
\Delta ^{\left( {\tau}\right) } x_{k} =\sum _{i}\left(-1\right)^{i} \frac{\Gamma \left({{\tau} }+1\right)}{i!\Gamma ({{\tau} }-i+1)} x_{k-i}
\end{equation}
\begin{equation} \label{1.3}
\Delta ^{-\left( {\tau}\right) } x_{k} =\sum _{i}\left(-1\right)^{i} \frac{\Gamma \left({{-\tau} }+1\right)}{i!\Gamma ({-{\tau} }-i+1)} x_{k-i}
\end{equation}
Throughout the paper it is assumed that the series of fractional difference operators are convergent. It is also appropriate to express the difference operator and its inverse as follows:\
\[\left(\Delta ^{\left( {\tau}\right) } \right)_{nk} =\left\{\begin{array}{l} {\left(-1\right)^{n-k} \frac{\Gamma \left({{\tau} }+1\right)}{\left(n-k\right)!\Gamma ({{\tau} }-n+k+1)} \qquad if\, 0\le k\le n} \\ {0 \qquad \qquad \qquad \qquad \qquad \qquad if\, k>n} \end{array}\right. \] \
\[\left(\Delta ^{\left( {-\tau}\right) } \right)_{nk} =\left\{\begin{array}{l} {\left(-1\right)^{n-k} \frac{\Gamma \left({-{\tau} }+1\right)}{\left(n-k\right)!\Gamma ({-{\tau} }-n+k+1)} \qquad if\, 0\le k\le n} \\ {0 \qquad \qquad \qquad \qquad \qquad \qquad if\, k>n} \end{array}\right. \] \
Basar and Braha \cite{bara3} introduced Euler-Cesaro difference sequence spaces $\check{c}$ , $\check{c_{0}}$, $\check{l_\infty}$ of null, convergent and bounded sequences respectively and then Ellidokuzoglu and Demiriz \cite{demri2} introduced Euler-Riesz difference sequence spaces ${ \left[{ c_{0}} \right] _{e,r}}$ , ${ \left[ c \right] _{e,r}}$ , ${\left[ {l_\infty} \right] _{e,r}} $. Baliarsingh and Dutta \cite{bali4} introduced the fractional difference operators. Now our interest is to introduce Euler-Riesz difference sequence spaces of fractional order.\
\noindent Now we introduce the spaces $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $ , $ {\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and $ {\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} $ by using the product of the Euler mean $ E_{1} $ and Riesz mean $ R_{q} $ with fractional operator $ \Delta^{\left( \tau \right)} $ .We prove certain topological properties of these spaces along with $ \alpha- , \beta- , \gamma- $ duals.\
\section{Main Results}
\noindent Here we introduce the matrix $ {\tilde{B} \left( \Delta^{\left( \tau \right)}\right)} ={\tilde{B}^{\left( \tau \right)}} $ by the product of Euler-Riesz matrix $\tilde{B}$ \cite{demri2} and fractional ordered difference operator $ \Delta^{\left( \tau \right)}$ \cite{bali4}, and obtain its inverse, where\
\begin{multline} \label{2.1}
\left({\tilde{B}^{\left( \tau\right)}} \right)_{nk} =\left\{\begin{array}{l} {\sum _{i=k}^{n}(-1)^{i-k} } \left(\begin{array}{l} {n} \\ {i} \end{array}\right)\frac{\Gamma \left({\tau} +1\right)}{\left(i-k\right)!\Gamma \left({\tau} -i+k+1\right)} \ \frac{q_{i}}{ 2^{n} Q_{n}} , { \qquad if\, 0\le k\le n } \\ {0,\qquad \qquad \qquad \qquad \qquad \qquad\qquad \qquad \qquad \qquad if\, k>n} \end{array}\right.
\end{multline}
Equivalently we may write\
\begin{multline}\label{*}
\left({\tilde{B}^{\left( \tau\right)}} \right) =\
\left
\begin{array}{ccccccc}
{\frac{1}{2} } & 0 & 0 & 0 & \dots \\
({\frac{2 q_{1}- \tau q_{2}}{2^2 Q_{2} } }) & ({\frac{q_{2} }{2^2 Q_{2}} }) & 0 & 0 & \dots \\
({\frac{3q_{1}-3{\tau} q_{2} +\frac{{\tau}({\tau}-1)}{2!} q_{3} }{2^3 Q_{3} } }) & ({\frac{3 q_{2}- \tau q_{3}}{2^3 Q_{3} } }) & ({\frac{q_{3}}{2^3 Q_{3}}}) & 0&\dots \\
\vdots & \vdots & \vdots &\vdots &\ddots
\end{array
\right)
\end{multline}\\
By simple calculation the inverse of the matrices $\Delta^ {\left(\tau \right)} , \tilde{B}$ and $\tilde{B}^{\left( \tau\right)}$ can be obtained as given in the following lemma.
\begin{l1} \label{1.1}\cite{bali7}
\noindent The inverse of fractional difference operator $\Delta ^{\left( {\tau}\right) } $ is given as
\[\left(\Delta ^{-\left({\tau}\right)} \right)_{nk} =\left\{\begin{array}{l} {\left(-1\right)^{n-k} \frac{\Gamma \left(-{\tau} +1\right)}{\left(n-k\right)!\Gamma (-{\tau} -n+k+1)} \qquad if\, 0\le k\le n} \\ {0\qquad \qquad \qquad \qquad \qquad \qquad if\, k>n} \end{array}\right. \]
\end{l1}
\begin{l1}\label{2}\cite{demri2}
\noindent The inverse of the Euler-Riesz matrix $\tilde{B}$ is given as
\[\left(\tilde{B} \right)_{nk}^{-1} =\left\{\begin{array}{l} {(-1)^{n-k} \ \ \left(\begin{array}{l} {n} \\ {k} \end{array}\right) {\frac{Q_{k}{2^k}}{q_{n}}}, \qquad \qquad if\, 0\le k\le n} \\ {0, \qquad \qquad \qquad \qquad \qquad \qquad if\, k>n} \end{array}\right. \]
\end{l1}
\begin{t1}\label{4}
\noindent The inverse of the fractional ordered Euler-Riesz matrix
$ {\left({\tilde{B}^{\left(\tau \right)}}\right)}_{nk} $ is written as $ {\left({\tilde{B}^{\left(\tau \right)}}\right)}_{nk}^{-1} $ and given by
\[{\left({\tilde{B}^{\left(\tau \right)}}\right)}_{nk}^{-1} =\\
\left\{\begin{array}{l} { \sum _{j=k}^{n}\left(-1\right)^{n-k} \left(\begin{array}{l} {j} \\ {k} \end{array}\right) \frac{\Gamma \left({-\tau} +1\right)}{\left(n-j\right)!\Gamma \left({-\tau}-n+j+1\right)} \frac{2^k Q_{k}}{q_{j}} , \qquad \qquad if\, 0\le k\le n } \\ {0, \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad if\, k>n} \end{array}\right. \]
\end{t1}
\begin{proof}
This theorem can be proved using lemma \ref{1.1}, lemma \ref{2} and hence omitted.
\end{proof}
\noindent For a positive real number $\tau $ , we now introduce the classes of fractional ordered Euler-Riesz difference sequence spaces $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $ , $ {\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and $ {\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} $ by
\begin{tiny}
\[\begin{array}{l} {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}
{=\left\{x=\left(x_{k} \right)\in w:{\mathop{\lim }\limits_{m\to \infty }} \left| \sum _{j=0}^{m}\sum _{i=j}^{m} \left(-1\right)^{i-j} \left(\begin{array}{l} {m} \\ {i} \end{array}\right)
\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)}\frac{q_{i} x_{j}}{ 2^{m} Q_{m}}\right| ^{p_{k} }
=0 \right\}} \end{array}\]
\[\begin{array}{l} {\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} {=\left\{x=\left(x_{k} \right)\in w:{\mathop{\lim }\limits_{m\to \infty }} \left| \sum _{j=0}^{m}\sum _{i=j}^{m} \left(-1\right)^{i-j} \left(\begin{array}{l} {m} \\ {i} \end{array}\right)
\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)}\frac{q_{i} x_{j}}{ 2^{m} Q_{m}}\right| ^{p_{k} }
exists \right\}} \end{array}\]
\[\begin{array}{l} {\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} {=\left\{x=\left(x_{k} \right)\in w:{\mathop{\sup }\limits_{m}} \left| \sum _{j=0}^{m}\sum _{i=j}^{m} \left(-1\right)^{i-j} \left(\begin{array}{l} {m} \\
{i} \end{array}\right)
\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)}\frac{q_{i} x_{j}}{ 2^{m} Q_{m}}\right| ^{p_{k} }
<\infty \right\}} \end{array}\]
\end{tiny}
Above spaces can be written as :\\
$ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}
=\left(c_{0}\left( p \right) \right)_{\left(\tilde{B}^{\left( \tau\right) } \right)} , \left[ c ,\ \Delta^{\left(\tau \right)} , p \right]_{er}
{=\left(c\left( p \right) \right)_{ \left( \tilde{B}^{\left( \tau\right) } \right) }, {\left[ l_\infty ,\ \Delta^{\left(\tau \right)} ,p \right]}_{er}=\left(l_\infty\left( p \right) \right)_{\left(\tilde{B}^{\left( \tau\right) }\right)} }.$
The above sequence spaces generalize many of the known sequence spaces as particular cases which are as follows:\\
${\left( i\right)} $. For ${ \tau ={0} }$ and ${ p=\left( p_{k}\right) = e}$ , the above classes reduce to ${ \left[{ c_{0}} \right] _{e,r}}$ , ${ \left[ c \right] _{e,r}}$ , ${\left[ {l_\infty} \right] _{e,r}} $ introduced by Ellidokuzoglu and Demiriz.\cite{demri2}\\
${\left( ii\right)} $. For ${ \tau ={0} }$ and ${ p=\left( p_{k}\right) = e} , { q=\left( q_{k}\right) = e}$ , the above classes reduce to the sequence spaces $\check{c}$ , $\check{c_{0}}$, $\check{l_\infty}$ studied by Basar and Braha \cite{bara3}.\\
\noindent Now with $\tilde{B}^{\left({\tau}\right)}$ - transform of $x={\left(x_{k}\right)}$ we define the sequence $y=\left(y_{k} \right)$ as follows :
\begin{equation} \label{3.1}
\begin{array}{l} {y_{k} =\left( \tilde{B}^{\left({\tau}\right)} x \right)_{k} } {=\sum _{j=0}^{k}\sum _{i=j}^{k}\left(-1\right)^{i-j} \left(\begin{array}{l} {k} \\ {i} \end{array}\right)\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)} \frac{q_{i} x_{j}}{ 2^{k} Q_{k}} } \end{array} ,
\end{equation}
By a straightforward calculation $\left( \ref{3.1}\right)$ it can be obtained that
\begin{equation} \label{3.2}
\begin{array}{l} {x_{k} =\left(\tilde{B}^{\left({-\tau}\right)} y \right)_{k} } {=\sum _{i=0}^{k}\sum _{j=i}^{k}\left(-1\right)^{k-i} \left(\begin{array}{l} {j} \\ {i} \end{array}\right) \frac{\Gamma \left({-\tau} +1\right)}{\left(k-j\right)!\Gamma \left({-\tau}-k+j+1\right)} \frac{2^i Q_{i} y_i}{q_{j}} } \end{array} ,
\end{equation}
\begin{l1}\label{3.5}
\noindent The operator $\tilde{B}^{\left({\tau}\right)}: w \ {\rightarrow} \ w $ is linear.
\end{l1}
\begin{proof}
The proof is omitted as it can be easily obtained.
\end{proof}
Remark: $\tilde{B}^{\left({\tau}\right)} . \tilde{B}^{\left({-\tau}\right)} \cong \tilde{B}^{\left({-\tau}\right)} .\tilde{B}^{\left({\tau}\right)} \cong I $ , where I is an identity matrix.
\section{Topological structure}
\noindent This section deals with some interesting topological results of the spaces $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $, $ {\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and $ {\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} $ .
\begin{t1}\label{th3.1}
\noindent $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $ is a paranormed space with the paranorm \[ {g}_{\tilde{B}^{\left( \tau\right)}} \left( x\right){= {\mathop{\sup }\limits_{k\in N}}\left|{ \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_k\right| ^{\frac{p_{k}}{M}}}\]
\begin{equation}\label{pnorm1}
={\mathop{\sup }\limits_{n}} \left| \sum _{j=0}^{k}\sum _{i=j}^{k} \left(-1\right)^{i-j} \left(\begin{array}{l} {k} \\
{i} \end{array}\right)
\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)}\frac{q_{i}}{ 2^{k} Q_{k}} {x_{j}}\right| ^{\frac{p_k}{M}},
\end{equation}
where $ 0\ <{p_k} \leq H\ < \infty , H={\mathop{\sup }\limits_{k}}\ {p_k} , M = max(1,H) , h={\mathop{\inf }\limits_{k}}\ {p_k}.$
\end{t1}
\begin{proof}
\noindent The theorem will be proved only for $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $.\\
Assumed that $h > 0$ , then $ {g}_{\tilde{B}^{\left( \tau\right) }} \left( \theta \right)=0 $ and $ {g}_{\tilde{B}^{\left( \tau\right) }} \left( -x\right)={g}_{\tilde{B}^{\left( \tau\right) }} \left( x\right) $. To prove the linearity of $ {g}_{\tilde{B}^{\left( \tau\right) }} \left( x\right) $, we consider two sequences $x= {\left( x_k\right)} , y= {\left( y_k\right)} \in {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ and any two scalars ${{\beta_1} , {\beta_2} \in \mathbb{R} }$ .
Since $\tilde{B}^{\left({\tau}\right)}$ is linear we get ,\\
$ {g}_{\tilde{B}^{\left({\tau}\right)}} \left( {{\beta_1}x + {\beta_2}y} \right) ={\mathop{\sup }\limits_{n}} \left| \sum _{j=0}^{k}\sum _{i=j}^{k} \left(-1\right)^{i-j} \left(\begin{array}{l} {k} \\
{i} \end{array}\right)
\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)}\frac{q_{i}}{ 2^{k} Q_{k}} {\left( {{\beta_1}{x_j} + {\beta_2}{y_j}}\right) }\right| ^{\frac{p_k}{M}} \\
\leq {max\{ 1,\left| \beta_1\right| \} {\mathop{\sup }\limits_{k\in N}}\left|{ \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_k\right| ^{\frac{p_{k}}{M}} + max\{ 1,\left| \beta_2\right| \} {\mathop{\sup }\limits_{k\in N}}\left|{ \left(\left( \tilde{B}^{\left({\tau}\right)}\right)y\right)}_k\right| ^{\frac{p_{k}}{M}}} \\
{= max\{ 1,\left| \beta_1\right| \} \ {{g}_{\tilde{B}^{\left({\tau}\right)}} \left( x\right)} + max\{ 1,\left| \beta_2\right| \} \ {{g}_{\tilde{B}^{\left({\tau}\right)}} \left( y\right)} } $\\
Hence ${g}_{\tilde{B}^{\left({\tau}\right)}}$ is subadditive. i.e. ${{g}_{\tilde{B}^{\left({\tau}\right)}} \left( {x+y}\right)} \leq {{g}_{\tilde{B}^{\left({\tau}\right)}} \left( x\right)} + {{g}_{\tilde{B}^{\left({\tau}\right)}} \left( y\right)}$ , for all $x,y \in {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $. Now consider $ \{ u^n \} $ is a sequence of points in ${\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ then $ {{g}_{\tilde{B}^{\left({\tau}\right)}} \left( {u^n} - u\right)} \rightarrow {0}$ and $\left( {\lambda_n} \right) $ is a sequence of scalars such that $ \lambda_n \rightarrow \lambda $ as $ {n} \rightarrow \infty $. By using the subadditivity of ${g}_{\tilde{B}^{\left({\tau}\right)}}$ , we get
$ {g}_{\tilde{B}^{\left({\tau}\right)}} \left( {u^n} \right) \leq {g}_{\tilde{B}^{\left({\tau}\right)}} \left( u\right)+{g}_{\tilde{B}^{\left({\tau}\right)}} \left( {u^n} - u\right) $.\\
\noindent Since $ \{{g}_{\tilde{B}^{\left({\tau}\right)}} \left( {u^n}\right) \} $ is bounded , we have \\
$ {g}_{\tilde{B}^{\left({\tau}\right)}} \left( { \lambda_n u^n} - \lambda u\right) ={\mathop{\sup }\limits_{k}} \left| \sum _{j=0}^{k} \left[ \sum _{i=j}^{k} \left(-1\right)^{i-j} \left(\begin{array}{l} {k} \\
{i} \end{array}\right)
\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)}\frac{q_{i}}{ 2^{k} Q_{k}}\right] {\left( {{\lambda_n}{u_j}^n + {\lambda}{u_j}}\right) }\right| ^{\frac{p_k}{M}} \\
\leq {\left| {\lambda_n - \lambda}\right|}^{\frac{p_k}{M}} \ {g}_{\tilde{B}^{\left({\tau}\right)}} \left( {u^n} \right) + {\left| {\lambda}\right|}^{\frac{p_k}{M}} \ {g}_{\tilde{B}^{\left({\tau}\right)}} \left( {u^n} - u \right) \rightarrow 0 \ as\ {n \rightarrow \infty} $.\\
Hence it shows that the scalar multiplication of ${g}_{\tilde{B}^{\left({\tau}\right)}} \left( {x}\right)$ is continuous and ${g}_{\tilde{B}^{\left({\tau}\right)}} \left( {x}\right)$ is a paranorm of the space ${\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ and proof for other spaces can be done using similar techniques.
\end{proof}
\begin{t1}\label{th3.2}
\noindent The sequence space ${\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ is a complete linear space paranormed by ${g}_{\tilde{B}^{\left({\tau}\right)}} \left( {x} \right)$ .
\end{t1}
\begin{proof}
The proof is a routine verification and hence omitted.
\end{proof}
\begin{t1}\label{th3.3}
\noindent $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $, $ {\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and $ {\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} $
are linearly isomorphic to $c_{0} {\left(p \right)} ,\, c {\left(p \right)},\, l_{\infty } {\left(p \right)} $ where $0\ <{p_k} \leq H\ < \infty$,\ respectively.
\begin{proof}
\noindent Now define a mapping $F:{\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} \to l_{\infty} {\left(p \right)} $ by $x\to y=Fx$.
\noindent Clearly , $F$ is a linear transformation. If $Fx=\theta \ then \ x=\theta $, so $F$ is one-one. \
\noindent Let $y\in l_{\infty}{\left(p \right)} $ , define a sequence $x=\left(x_{k} \right)$ in $\left( \ref{3.1}\right)$ as
\[x_{k} =\sum _{i=0}^{k} \sum _{j=i}^{k}\left(-1\right)^{k-i} \left(\begin{array}{l} {j} \\ {i} \end{array}\right)\frac{\Gamma \left(-{\tau} +1\right)}{\left(k-j\right)!\Gamma \left(-{\tau} -k+j+1\right)} \frac{2^i Q_{i}}{q_{j}} y_{i} \]
Then \[ {g}_{\tilde{B}^{\left({\tau}\right)}} \left( x\right)
={\mathop{\sup }\limits_{n}} \left| \sum _{j=0}^{k}\sum _{i=j}^{k} \left(-1\right)^{i-j} \left(\begin{array}{l} {k} \\
{i} \end{array}\right)
\frac{\Gamma \left({\tau} +1\right)}{\left(i-j\right)!\Gamma \left({\tau} -i+j+1\right)}\frac{q_{i}}{ 2^{k} Q_{k}} {x_{j}}\right| ^{\frac{p_k}{M}} \]
\[ = \mathop{\sup }\limits_{k \in N} \left|\sum _{j=0}^{k} {\delta_{kj} \ y_{j}}\right|^{\frac{p_k}{M}} ={\mathop{\sup }\limits_{k \in N}} \left| y_{k} \right|^{\frac{p_k}{M}} \ < \infty \]
\[ where \ {\delta_{kj}} = \left\lbrace \begin{array}{l} { 1 , \qquad if\, k = j } \\ { 0 , \qquad if\, k \neq j } \end{array}\right. \]
Thus, $ x \in {\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and $F$ is a linear bijection and paranorm preserving. Hence the spaces ${\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er}$ and $l_\infty\left( p\right) $ are linearly isomorphic.
\noindent i.e. ${\left[ l_\infty , \Delta^{\left(\tau \right)} , p \right]}_{er} \cong l_\infty\left( p\right) $. The proof for other spaces can be obtained in a similar manner.
\end{proof}
\end{t1}
\section{Basis for the spaces}
\noindent In this section the Schauder basis \cite{madx3} for $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $, $ {\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} $ are constructed.
\begin{t1}\label{th4.1}
For $0\ <{p_k} \leq H\ < \infty$ , let $ \mu_{k} \left( q\right){= { \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_k} $. For $k \in N_{0}$ define $ {b^{\left( k\right)}} {\left( q\right)} = {\left\lbrace {{b_{n}}^{\left( k\right)}} {\left( q\right)}\right\rbrace}_{n \in N_{0}} $ by
\begin{multline}
{\left\lbrace {{b_{n}}^{\left( k\right)}} {\left( q\right)}\right\rbrace} =\left\{\begin{array}{l} {\sum _{j=k}^{n}(-1)^{n-k} } \left(\begin{array}{l} {j} \\ {k} \end{array}\right)\frac{\Gamma \left({-\tau} +1\right)}{\left(n-j\right)!\Gamma \left({-\tau} -n+j+1\right)} \ \frac{2^{k} Q_{k}}{q_{j} } , { \qquad if\, 0\le k \le n } \\\ {\frac{2^{n} Q_{n}}{q_{n}}, \qquad \qquad \qquad \qquad \qquad \qquad\qquad\qquad \qquad if\, k=n} \\\ \ {0,\qquad \qquad \qquad \qquad \qquad \qquad\qquad \qquad \qquad \qquad if\, k>n} \end{array}\right.
\end{multline}
\noindent$(i)$ $\left\lbrace {{b_{n}}^{\left( k\right)}} {\left( q\right)}\right\rbrace$ is a basis for ${\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and each $x\in {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and $ x $ has unique representation
\[x=\sum _{k}\mu _{k}{\left( q\right)} {{b_{n}}^{\left( k\right)}} {\left( q\right)} \] .
\noindent$(ii)$ $\left\lbrace {\left(\tilde{B}^{\left( -\tau \right) } \right){e}}, {{b_{n}}^{\left( k\right)}} {\left( q\right)}\right\rbrace$ is a basis for ${\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} $, and each $x\in {\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er} $ and $ x $ has unique representation
\[x= {le} + \sum _{k} \left( {\mu_{k}-l}\right) {b^{\left( k\right)}}, \ \ l=\lim\limits_{k \rightarrow \infty} {\mu _{k}} \ . \]
\end{t1}
\begin{proof}
\noindent $(i)$ By the definition of ${\left(\tilde{B}^{\left(\tau \right) } \right)}$ and ${{b_{n}}^{\left( k\right)}} {\left( q\right)}$ ,
\[\tilde{B}^{\left(\tau \right) }{{b_{n}}^{\left( k\right)}} {\left( q\right)}={e^{\left(k\right)}} \in c_{0} ,\]
Let $x\in {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{e,r} $, then
$ x^{\left[s \right]} =\sum _{k=0}^{s} \mu_{k}{\left( q\right)} {{b}^{\left( k\right)}}{\left( q\right)}$ for an integer $s \geq 0$.
By applying ${\tilde{B}^{\left(\tau \right) }}$ we get \ $ {\tilde{B}^{\left(\tau \right)}} x^{\left[s \right]} =\sum _{k=0}^{s} \mu _{k}{\left( q\right)} {\tilde{B}^{\left(\tau \right) }} {{b}^{\left( k\right)}}{\left( q\right)}$
\[=\sum _{k=0}^{s} {\mu_{k}{\left( q\right)}}{e^{\left(k\right)}} = { \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_k e^{\left(k\right)} \] and
\begin{multline}
{\tilde{B}^{\left(\tau \right)} \left( {x-x^{\left[s\right]}}\right)}_{r} =\left\{\begin{array}{l} {0, \qquad\qquad \qquad \qquad if\, 0\le r \le s } \\ {{\left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_k, \qquad \qquad if\, r>s} \end{array}\right. ; for \ r,s \in {N_{0}}
\end{multline}
\noindent For $\epsilon > 0$ there exist an integer $m_{o}$ s.t.
\[{\mathop{\sup }\limits_{r \geq s}}\left|{ \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_{r}\right| ^{\frac{p_{k}}{M}} < {\frac{\epsilon}{2}} \ for \ all \ s \geq {m_0} .\]
Hence \[{g}_{\tilde{B}} \left( {x-x^{\left[s \right]}} \right) ={\mathop{\sup }\limits_{r \geq s}}\left|{ \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_{r}\right| ^{\frac{p_{k}}{M}} < {\frac{\epsilon}{2}} < {\epsilon} \ , for \ all \ s\geq {m_0}. \]
\noindent Assume that $x=\sum _{k}\eta _{k} {\left( q\right) } {b}^{\left(k\right)} {\left( q\right) }$. Since the linear mapping $F$ from $ {\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er} $ to $ {c_{0} \left( p\right)}$ is continuous we have,
\[{ \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_{k}=\sum _{k}\eta _{k} {\left( q\right)} \left( {\left( \tilde{B}^{\left({\tau}\right)}\right)} {b}^{\left(k\right)} {\left( q\right)}\right)_n \]
\[ =\sum _{k}\eta _{k} {\left( q\right)} e^{\left(k\right)} = \eta _{n} {\left( q\right)} \ . \]
This contradicts to our assumption that $ { \left(\left( \tilde{B}^{\left({\tau}\right)}\right)x\right)}_{k} =\mu_{k} {\left( q\right)} $ for each $k\in {N_{0}}$.
Thus, the representation is unique.
\noindent $(ii)$ The proof as it is similar to previous one.
\end{proof}
\section{${\alpha-}, \beta-$ and $\gamma-$ duals}
Here we determine $\alpha -,\beta -$ and $\gamma - $ duals of ${\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ , ${\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er}$ and
${\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er}$.
We define \[S(X,Y)=\left\{u=\left(u\right)_{k} \in \omega :ux=\left(u_{k} x_{k} \right)\in Y,whenever\, x=\left(x_{k} \right)\in X\right\}\] as the multiplier sequence space for any two sequence spaces X and Y.
Let $\alpha -,\beta -$ and $\gamma -$duals be denoted by
\noindent $X^{\alpha } =S\left(X,l_{1} \right),\, X^{\beta } =S\left(X,cs\right),\, X^{\gamma } =S\left(X,bs\right)$ respectively.
\noindent Throughout the collection of all finite subsets of $ \mathbb{N}$ is denoted by $\kappa $ . We consider $K\in \kappa$.
\begin{l1}\cite{gros}\label{lm5}
\noindent Let $A=\left(a_{n,k} \right)$ be an infinite matrix. Then,
\begin{enumerate}
\item $A\in \left(l_{\infty}{\left(p \right) }, l{\left(q \right) } \right)$ iff \begin{equation}
{\mathop{\sup }\limits_{k\in \kappa }} \sum _{n}{\left|\sum _{k\in K}a_{nk}{B^{\frac{1}{p_k}}} \right|}^{q_n} <\infty ,\qquad for\, all\, integers\, B>1 and\, {q_n}\geq1 for\, all\, n ;
\end{equation}
\item $A\in \left(l_{\infty}{\left(p \right) }, l_{\infty}{\left(q \right) } \right)$ iff \begin{equation} \label{4.5}
{\mathop{\sup } \limits_{n\in N}} \left( \sum _{k}\left|a_{nk} \right| {B^{\frac{1}{p_k}}} \right)^{q_n} <\infty ,\qquad ;
\end{equation}
\item $A\in \left(l_{\infty}{\left(p \right) }, c{\left(q \right) } \right)$ and $q=\left({q_n} \right) $ be a bounded sequence of strictly positive real numbers iff \begin{equation} \label{4.6}
{\mathop{\sup }\limits_{n\in N}} \sum _{k}\left|a_{nk} \right| {B^{\frac{1}{p_k}}}<\infty ,\qquad for\, all\, B>1 ,
\end{equation} exists $\left({\tau_{k}} \right)\subset R $ such that \
${\mathop{\lim } \limits_{n\to \infty }}\left( \sum _{k}\left| a_{nk}-{\tau_{k}} \right|{B^{\frac{1}{p_k}}} \right)^{q_n} =0 , for\, all\, B>1$;
\item $A\in \left(l_{\infty}{\left(p \right) }, c_{0}{\left(q \right) } \right)$ iff\[ {\mathop{\lim } \limits_{n\to \infty }}\left( \sum _{k}\left| a_{nk} \right|{B^{\frac{1}{p_k}}} \right)^{q_n} =0 , \qquad for\, all\, B>1 , \] .
\end{enumerate}
\end{l1}
\begin{l1}\cite{gros}\label{lm6}
\noindent Let $A=\left(a_{nk} \right)$ be an infinite matrix. Then,
\begin{enumerate}
\item $A\in \left(c_{0}{\left(p \right) }, l_{\infty}{\left(q \right) } \right)$ iff \begin{equation} \label{4.7}
{\mathop{\sup } \limits_{n\in N}} \left( \sum _{k}\left|a_{nk} \right| {B^{\frac{-1}{p_k}}} \right)^{q_n} <\infty ,\qquad for\, all\, B>1 ;
\end{equation}
\item $A\in \left(c_{0}{\left(p \right) }, c{\left(q \right) } \right)$ iff \begin{equation} \label{4.8}
{\mathop{\sup }\limits_{n\in N}} \sum _{k}\left|a_{nk} \right| {B^{\frac{-1}{p_k}}}<\infty ,\qquad for\, all\, B>1 ,
\end{equation}
\begin{equation} \label{4.9} exists \left({\tau_{k}} \right)\subset R \, such\, that\,
{\mathop{\sup }\limits_{n\in N}}\sum _{k}\left| a_{nk}-{\tau_{k}} \right|{M^{\frac{-1}{p_k}}}{B^{\frac{-1}{p_k}}} < \infty ,\\
\end{equation} for all integers $M,B>1$;
\begin{equation} \label{5.1} exists \left({\tau_{k}} \right)\subset R \, such\, that\,
{\mathop{\lim } \limits_{n\to \infty }} \sum _{k}\left| a_{nk}-{\tau_{k}} \right|^{q_n} =0 , for\, all\, k \in N;
\end{equation}
\item $A\in \left(c_{0}{\left(p \right) }, c_{0}{\left(q \right) } \right)$ iff
\begin{equation} \label{5.2} exists \left({\tau_{k}} \right)\subset R \, such\, that\,
{\mathop{\sup }\limits_{n\in N}}\sum _{k}\left| a_{nk} \right|{M^{\frac{-1}{p_k}}}{B^{\frac{-1}{p_k}}} < \infty ,\\
\end{equation} for all integers $M,B>1$;
\begin{equation} \label{5.3} exists \left({\tau_{k}} \right)\subset R \, such\, that\,
{\mathop{\lim } \limits_{n\to \infty }} \sum _{k}\left| a_{nk} \right|^{q_n} =0 , for\, all\, k \in N.
\end{equation}
\end{enumerate}
\end{l1}
\begin{l1}\cite{gros}\label{lm4}
\noindent Let $A=\left(a_{nk} \right)$ be an infinite matrix. Then,
\begin{enumerate}
\item $A\in \left(c{\left(p \right) }, l_{\infty}{\left(q \right) } \right)$ iff (16) holds and \begin{equation} \label{4.2}
{\mathop{\sup } \limits_{n\in N}}\left| \sum _{k}a_{nk} \right|^{q_n} <\infty ;
\end{equation}
\item $A\in \left(c{\left(p \right) }, c{\left(q \right) } \right)$ iff (17), (18), (19) hold:
\begin{equation} \label{5.4} exists \left({\tau} \right)\subset R \, such\, that\,
{\mathop{\lim } \limits_{n\to \infty }} \left| \sum _{k} a_{nk}-{\tau} \right|^{q_n} =0 ;
\end{equation}
\item $A\in \left(c{\left(p \right) }, c_{0}{\left(q \right) } \right)$ iff (20), (21) hold and
\begin{equation} \label{5.5}
{\mathop{\lim } \limits_{n\to \infty }} \left| \sum _{k} a_{nk} \right|^{q_n} =0 .
\end{equation}
\end{enumerate}
\end{l1}
\begin{t1}
\noindent Let $0\ <{p_k} \leq H\ < \infty$ The ${\alpha-}, \beta-$ and $\gamma-$ duals of ${\left[ c_{0} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ , ${\left[ c , \Delta^{\left(\tau \right)} , p \right]}_{er}$ and
${\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ are the following sets
\begin{tiny}
\[D_{1}^{\left( \tau\right)}{\left( p\right) } =\\
{\bigcap_{M>1}}\left\{a=\left(a_{k} \right)\in \omega :{\mathop{\sup }\limits_{k\in \tau }} \sum _{n}\left|\sum _{k\in K}\left[{\sum _{j=k}^{n-1}(-1)^{n-k} } \left(\begin{array}{l} {j} \\ {k} \end{array}\right)\frac{\Gamma \left({-\tau} +1\right)}{\left(n-j\right)!\Gamma \left({-\tau} -n+j+1\right)} \ \frac{2^{k} Q_{k}}{q_{j} }{a_k} \right] \right|M^{\frac{1}{p_k}} <\infty \right\},\]
\[D_{2}^{\left( \tau\right)}{\left( p\right) } =
{\bigcap_{M>1}}\left\{a=\left(a_{k} \right)\in \omega : \sum _{k}\left|\tilde{B}^{\left(\tau \right) }{\left(\frac{a_{k}}{q_{k}}\right)}{Q_k} \right|M^{\frac{1}{p_k}} <\infty and \left(\frac{a_{k}}{q_{k}}{Q_k}M^{\frac{1}{p_k}} \right)\in {c_0}\right\} ,\]
\[D_{3}^{\left( \tau\right)}{\left( p\right) } =
{\bigcap_{M>1}}\left\{a=\left(a_{k} \right)\in \omega :\sum _{k}\left|\tilde{B}^{\left(\tau \right)}{\left(\frac{a_{k}}{q_{k}}\right)}{Q_k} \right|M^{\frac{1}{p_k}} <\infty and \left\lbrace\tilde{B}^{\left(\tau \right)}{\left(\frac{a_{k}}{q_{k}}\right)}{Q_k} \right\rbrace \in {l_ {\infty}} \right\},\]
\[D_{4}^{\left( \tau\right)}{\left( p\right) } =\\
{\bigcup_{M>1}}\left\{a=\left(a_{k} \right)\in \omega :{\mathop{\sup }\limits_{k\in \tau }} \sum _{n}\left|\sum _{k\in K}\left[{\sum _{j=k}^{n-1}(-1)^{n-k} } \left(\begin{array}{l} {j} \\ {k} \end{array}\right)\frac{\Gamma \left({-\tau} +1\right)}{\left(n-j\right)!\Gamma \left({-\tau} -n+j+1\right)} \ \frac{2^{k} Q_{k}}{q_{j} }{a_k} \right] \right|M^{\frac{-1}{p_k}} <\infty \right\},\]
\[D_{5}^{\left( \tau\right)}{\left( p\right) } =\\
{\bigcup_{M>1}}\left\{a=\left(a_{k} \right)\in \omega : \sum _{n}\left|\sum _{k}\left[{\sum _{j=k}^{n-1}(-1)^{n-k} } \left(\begin{array}{l} {j} \\ {k} \end{array}\right)\frac{\Gamma \left({-\tau} +1\right)}{\left(n-j\right)!\Gamma \left({-\tau} -n+j+1\right)} \ \frac{2^{k} Q_{k}}{q_{j} }{a_k} \right] \right| <\infty \right\},\]
\[D_{6}^{\left( \tau\right)}{\left( p\right) } =
{\bigcap_{M>1}}\left\{a=\left(a_{k} \right)\in \omega :\sum _{k}\left|\tilde{B}^{\left(\tau \right)}{\left(\frac{a_{k}}{q_{k}}\right)}{Q_k} \right|M^{\frac{-1}{p_k}} <\infty \right\},\]
\end{tiny}
where\begin{equation}
\tilde{B}^{\left(\tau \right)}{\left(\frac{a_{k}}{q_{k}}\right)} ={\frac{2^{k} a_{k}}{q_{k}}+{\sum _{i=k+1}^{n}(-1)^{n-k}{a_i} } {\sum _{j=k}^{i}} \left(\begin{array}{l} {i} \\ {j} \end{array}\right)\frac{\Gamma \left({-\tau} +1\right)}{\left(j-i\right)!\Gamma \left({-\tau} -j+i+1\right)} \ \frac{2^{i}}{q_{j} }}
\end{equation}
\end{t1}
\begin{proof}
\noindent We prove it for ${\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ . Considering $x=\left(x_{k} \right)$ as in \ref{3.2}, let $a=\left(a_{k} \right)\in \omega $ define
\[\begin{array}{l} {a_{n} x_{n} =\sum _{i=0}^{k} \sum _{j=i}^{k}\left(-1\right)^{k-i} \left(\begin{array}{l} {j} \\ {i} \end{array}\right)\frac{\Gamma \left(-{\tau} +1\right)}{\left(k-j\right)!\Gamma \left(-{\tau} -k+j+1\right)} \frac{2^i Q_{i}}{q_{j}}{a_n} y_{i} } \\ {\, \, \, \, \, \, \, \, \, \, =\left(Uy \right)_{n} ,\qquad {\rm for}\, n\in N} \end{array},\]
where matrix $U =\left(u_{nk} \right)$ is defined as
\begin{multline}
{u_{nk}}=\left\{\begin{array}{l} {\sum _{j=k}^{n}(-1)^{n-k} } \left(\begin{array}{l} {j} \\ {k} \end{array}\right)\frac{\Gamma \left({-\tau} +1\right)}{\left(n-j\right)!\Gamma \left({-\tau} -n+j+1\right)} \ \frac{2^{k} Q_{k}}{q_{j}{a_n} } , { \qquad if\, 0\le k \le n } \\\ {\frac{2^{n} Q_{n}}{q_{n}}, \qquad \qquad \qquad \qquad \qquad \qquad\qquad\qquad \qquad if\, k=n} \\\ \ {0,\qquad \qquad \qquad \qquad \qquad \qquad\qquad \qquad \qquad \qquad if\, k>n} \end{array}\right.
\end{multline}
Therefore we conclude that $ax=\left(d_{n} x_{n} \right)\in l_{1} $ when $x=\left(x_{k} \right) \in {\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ iff $Uy \in l_{1}$ as $y=\left({y_k} \right) \in {l_{\infty}}{\left( p\right)} $ .\\
By lemma $\left( \ref{lm5}\right) $ we conclude that $\left\lbrace {\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er} \right\rbrace ^{\alpha}=D_{1}^{\left( \tau\right)}{\left( p\right) } $.
\noindent Now,
\[\sum _{i=0}^{n}{a_{k} x_{k} =\sum _{k=0}^{n} {a_k}\left[ \sum _{i=0}^{k} \sum _{j=i}^{k}\left(-1\right)^{k-i} \left(\begin{array}{l} {j} \\ {i} \end{array}\right)\frac{\Gamma \left(-{\tau} +1\right)}{\left(k-j\right)!\Gamma \left(-{\tau} -k+j+1\right)} \frac{2^i Q_{i}}{q_{j}} y_{i} \right] }\]
\[ =\sum _{k=0}^{n} {y_k}{Q_k} \left[ {\frac{2^{k} a_{k}}{q_{k}}+{\sum _{i=k+1}^{n}(-1)^{n-k}{a_i} } {\sum _{j=k}^{i}} \left(\begin{array}{l} {i} \\ {j} \end{array}\right)\frac{\Gamma \left({-\tau} +1\right)}{\left(j-i\right)!\Gamma \left({-\tau} -j+i+1\right)} \ \frac{2^{i}}{q_{j} }} \right] \]
\[ =\sum _{k=0}^{n} {y_k}{Q_k} \tilde{B}^{\left(\tau \right)}{\left(\frac{a_{k}}{q_{k}}\right)} = \left(Vy \right)_{n},\]
where matrix $V =\left(v_{nk} \right)$ is defined as
\begin{multline}
{v_{nk}}=\left\{\begin{array}{l} {\tilde{B}^{\left(\tau \right)}{\left(\frac{a_{k}}{q_{k}}\right)}{Q_k}}, {\qquad \qquad \qquad \qquad if\, 0\le k \le n} \\{\frac{2^{n} Q_{n}}{q_{n}}, \qquad \qquad \qquad \qquad \qquad \qquad if\, k=n} \\ {0,\qquad \qquad \qquad \qquad \qquad \qquad \qquad if\, k>n} \end{array}\right.
\end{multline}
Therefore we deduce that $ax=\left(d_{n} x_{n} \right)\in cs $ when $x=\left(x_{k} \right) \in {\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er}$ iff $Vy \in c$ as $y=\left({y_k} \right) \in {l_{\infty}}{\left( p\right)} $ .\\
By using lemma $\left( \ref{lm5}\right) $ with $ q=q_{n}=1 $ we conclude that $\left\lbrace {\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er} \right\rbrace ^{\beta}=D_{2}^{\left( \tau\right)}{\left( p\right) } $.
Similarly by using lemma $\left( \ref{lm5}\right) $ with $ q=q_{n}=1 $ for all n, we conclude that $\left\lbrace {\left[ l_{\infty} , \Delta^{\left(\tau \right)} , p \right]}_{er} \right\rbrace ^{\gamma}=D_{3}^{\left( \tau\right)}{\left( p\right) } $.Hence the theorem proved and the duals of other spaces can be obtained in a similar manner using lemma $\left( \ref{lm6}\right) $ and lemma $\left( \ref{lm4}\right) $.
\end{proof}
|
\section{Introduction}
\begin{table}[htb!]
\begin{tabular}{ll}
\toprule
Glossary & \\
\midrule
PV & Photovoltaic \\
ANN & Artificial neural network \\
BP & Back propagation \\
CNN & Convolutional neural network \\
LSTM & Long short term memory \\
TFT & Temporal fusion transformer \\
GRN & Gated residual network \\
GLU & Gated linear unit \\
VSN & Variable selection network \\
NWP & Numerical weather predictions \\
GHI & Global horizontal irradiance \\
DHI & Diffuse horizontal irradiance \\
LOESS & Locally Weighted Scatterplot Smoothing \\
MSTL & Multiple seasonal-trend decomposition using LOESS \\
DKASC & Desert knowledge Australia solar Centre \\
\bottomrule
\end{tabular}
\end{table}
Energy demand is increasing dramatically as global urbanization progresses.
Conventional energy sources such as coal, oil and natural gas have limited reserves, and emit greenhouse gases when converted into electricity, exacerbating the greenhouse effect \cite{iea_GGE}.
To improve public health, more and more countries are using various policies to support clean energy.
Clean energy is renewable and produces less pollution \cite{epa_2022}, which generally includes solar, wind and hydro energy.
Compared to conventional energy sources, inexhaustible clean energy exists in all countries and regions.
Therefore, the development of clean energy has become a global trend.
Solar energy is a kind of clean energy with very low maintenance costs.
Distributed and modular system make solar energy less prone to large-scale failure.
With the steady progress of solar power generation technology, many countries and regions have installed a large number of photovoltaic (PV) facilities, the global PV power generation continues to grow.
The total cumulative installed PV capacity reached at least 942 GW \cite{iea_2022} at the end of 2021.
The development of solar energy can optimize the energy structure to build a low-carbon and efficient energy system.
PV power generation is highly correlated with the solar radiation and affected by meteorological factors such as floating dust and cloud cover.
Since the solar irradiation is strongest at noon during the day, the corresponding PV power generation also reaches its peak.
At this point, other generation methods need to reduce their output as planned to balance the grid demand.
When PV power generation deviates from the expected value, the grid needs to redeploy power from other suppliers.
This situation requires additional power generation to be temporarily supplemented by other methods, creating a challenge for grid integration.
On the other hand, the marginal tariff is five times the day-ahead tariff in many electricity markets, excessive inaccuracies can result in penalties, such as offsetting income if the forecast is exceeded by 10\% \cite{MATHIESEN2013357}.
PV producers need to provide the most accurate valuation of future power generation to achieve the highest possible economic efficiency.
Therefore, accurately predicted PV generation is of great importance for grid integration.
To describe the day-ahead PV forecasting problem accurately, we classify and summarize the research on PV forecasting in recent years.
\subsection{Classification by forecasting process}
Some studies use meteorological data to calculate solar irradiance and then predict future PV power output based on PV power generation model, which are indirect prediction methods \cite{ANTONANZAS201678}.
The PV prediction proposed by Lorenz et al. \cite{https://doi.org/10.1002/pip.1033} uses a model to obtain site-specific irradiance, and then predicts the hourly PV power output for the next two days based on the relationship between irradiance and historical PV power.
Pelland et al.\cite{https://doi.org/10.1002/pip.1180} used spatial averaging and Kalman filter method to improve the accuracy of solar irradiance prediction by reducing the instability due to cloud variability.
The PV prediction method is then used to combine solar irradiance, temperature of the PV system surface and historical PV power data to derive the future PV power generation in next 48 hours.
Since solar irradiance plays an important role in PV power forecasting and therefore precise prediction of solar irradiance is a must.
Here the authors \cite{ALONSOMONTESINOS2015387} propose a short-term and medium-term solar irradiance prediction method for solar beam, diffuse and global irradiance in sunny, partly cloudy and overcast conditions.
Lai et al.\cite{LAI2021114941} proposed a deep time series clustering method that divides global horizontal irradiance (GHI) time series data into multiple clusters and then builds a feature attention deep forecasting neural network for each cluster to predict the hour-ahead GHI in the future.
The direct prediction method directly predicts the PV power output.
Generally, PV data include meteorological data, historical PV data and PV system parameters \cite{DAS2018912}.
The authors \cite{https://doi.org/10.1002/eej.20755} compared two methods, direct prediction and indirect prediction, and summarized the use scenarios of the two prediction methods.
In the primary stage of PV system operation, the data volume is small, the indirect prediction method should be used until the PV system collects enough data.
After that, using the direct prediction method, the accuracy of the output power prediction of PV power plants can be guaranteed to the maximum extent.
\subsection{Classification by algorithm}
Models for PV power prediction can be generally classified into four categories: PV performance models, classical statistical models, artificial neural network models, and hybrid models.
PV performance models, also known as physical models, derives PV power generation mainly from the solar irradiance. So no historical data is required, but detailed parameters of the PV plant need to be obtained to achieve accurate prediction \cite{MAYER2022112772}.
In most cases , the future solar irradiance is obtained from numerical weather predictions (NWP)\cite{LIMA2016807}.
Many PV performance models use additional data to optimize the results, such as meteorological data like temperature, humidity and wind direction, and extra attributes like PV plant capacity and installation angle, leading to increased complexity of the modeling process.
However, the most important meteorological data, the solar irradiance, still requires NWP to be operative.
Making PV performance models require the assistance of nearby weather stations.
The statistical models use historical data to predict PV power generation.
Although artificial neural network (ANN) models are statistical models, their data-fit ability is much higher than other classical statistical models.
Therefore, we split statistical models into classical statistical models and ANN models.
Classical statistical models include linear and nonlinear regression models, such as linear regression, auto-regressors, and support vector regression.
They are simple and have a small number of parameters, but are less capable of fitting complex curves.
Song et al. \cite{7796490} proposed a PCA-SVM model to improve the accuracy of PV prediction, where PCA was used to extract the main features of the data and used as the input to the SVM model.
ANN models include back propagation networks (BP), convolutional neural networks (CNN), recurrent neural networks (RNN), and long-short memory networks (LSTM).
They have complicated structures, a large number of parameters, so they are tend to be more capable of fitting complex curves than classical models.
In the past, the parameters number was the limiting factor for model training and applying.
But now, with GPU, TPU and other computational acceleration methods, the model parameters number is no longer a problem.
Therefore, the ANN models are always used to predict PV power generation.
The authors \cite{ALMONACID2014389} propose an ANN-based method that uses two dynamic models to predict the hour ahead GHI and air temperature, and then integrates the configuration parameters of the PV system to calculate the power output of PV plant.
One study proposes a PV prediction method based on ANN and ELM models, and verifies prediction affect by the training data size, input variables, and variable order\cite{7387113}.
AlShafeey et al. \cite{ALSHAFEEY20217601} evaluated two different PV prediction modeling techniques, ANN and multiple regression, for predicting PV output power in the next 24 h.
The experimental results indicated that ANN had higher coefficient of determination values and lower MAE, MSE, and RMSE values, and the ANN model performed better than the multiple regression model.
The authors \cite{GAO2019115838} use LSTM to solve the problem of PV power output under different weather conditions, and use discrete grey model to the prediction results of PV power output under undesirable weather conditions such as cloudy and rainy days.
The results show that the RMSE of the LSTM-based prediction method can reach 4.62\% under the ideal sunny days.
The authors \cite{9621717} propose two extended CNN based frameworks named multi-headed CNN method and multi-channel CNN method, for predicting PV output power one day-ahead and two days-ahead.
Experiments show that the proposed two frameworks have better prediction results compared with CNN models.
The authors \cite{munawar2020framework} evaluate the combination of short-term solar power forecasting, and the result show that the combination of feature selection using PCA method and model using XGBoost works better.
Santos et al. \cite{en15145232} used temporal fusion transformer (TFT) to predict PV data for two regions, Germany and Australia, and achieved RMSE of 6.1\% on the desert knowledge Australia solar centre (DKASC) \cite{dkasc-2008,} dataset.
A hybrid model is a combination of several models by voting, bagging, and bosting, etc.
The combination may contain several different algorithmic models.
The data they used depend on the models contained inside.
That is, if two submodels that each require two different kinds of data, the hybrid model would require four kinds of data.
One study \cite{en6020733} proposed a hybrid model containing NWP and statistical models, using statistical models to predict future GHI, together with historical data to predict PV generation.
The statistical model requires high stability of PV power and meteorological data, and the large deviation of training data will lead to serious errors \cite{AASIM2019758}.
There are studies in which multiple ANN models are ensembled to achieve improved accuracy.
Agga et al \cite{AGGA2022107908} proposed a CNN-LSTM model, which uses a CNN to extract features from meteorological data and an LSTM model to fit the output, and fusing the advantages of both models to improve the accuracy of PV prediction.
The authors \cite{AGGA2021101} propose a hybrid model of CNN-LSTM and ConvLSTM, and concluded that the accuracy of CNN-LSTM and ConvLSTM is higher than the LSTM model.
Zhen et al \cite{9054985} proposed a hybrid model containing three models, CNN, LSTM and ANN, and used it for PV power prediction.
The paper eliminates some irrelevant noise by building a mapping model of the sky image and GHI, so that the prediction effect can be improved.
Then the prediction results of the hybrid model are compared with CNN, LSTM and ANN models. The results show that the hybrid model has better performance and maintains stability under different weather conditions.
One review of PV forecasting \cite{ANTONANZAS201678} found that more than half of pappers using machine learning models for modeling, which include ANN, LSTM and CNN.
On the other hand, there are points that good solar prediction methods should involve as much physics as possible \cite{yang2022concise,yang2019guideline}.
Wolff et al. \cite{WOLFF2016197} compared the prediction performance of support vector regression (SVR) with different inputs and compared the SVR and physical modeling methods, which were similar for a single PV system, but the physical model was better for regional PV prediction.
The authors \cite{RAMADHAN20211006} compare the accuracy of physical and machine learning models of PV output power and conclude that machine learning models generally outperform physical models if the parameters are chosen appropriately.
The authors also mention that even though machine learning models have better accuracy, the advantages of physical models with fewer parameters and no training should be taken into account when choosing the optimal model.
Mayer et al. \cite{MAYER2022112772} compared the predictions of physical models, machine learning models, and hybrid models in the conversion of solar irradiance to PV power, and demonstrated that hybrid models containing the most physically-calculated predictors performed best.
Hybrid models have a wide range of applications in predicting the output power of PV generators, and the physical model require lot of input data and detailed requirements, leading to complex modeling \cite{GANDOMAN2018793}.
Therefore, physical models can be used when predicted meteorological data and PV plant data are available.
When historical PV and meteorological data are available, statistical models can be used.
\subsection{Classification by forecast horizons}
Forecast horizon is the time span to forecast future PV power generation \cite{DAS2018912}.
Muhammad et al. \cite{RAZA20151352} published in 2015 classified PV forecasts into three categories, the time horizon from 1 h to 1 week indicates short-term forecasts, from 1 month to 1 year indicates medium-term forecasts, and the time horizon from 1 year to 10 years indicates long-term forecasts.
Another paper published by Muhammad et al \cite{RAZA2016125} in 2016 classified PV forecasts into four categories, adding ultra-short term forecasts to the original classification, which refers to the prediction of future PV power generation over a time span of 1 min to several min in the future.
Different prediction ranges have different implications for the power scheduling (generation, transmission and distribution) of PV plants.
Long-term and medium-term PV forecasts are good for scheduling, but most grid operators only need to make dispatch decisions one day in advance, so short-term PV forecasts directly affect the power consumption plans of PV power stations and are important for the stable operation of power stations \cite{CSEREKLYEI2019358}.
\subsection{PV forecast features}
PV forecast data can generally be categorized into endogenous features and exogenous features.
The endogenous features include PV generation and the corresponding lagged and logarithmic values, etc.
Exogenous features include solar irradiance, temperature, humidity, wind direction, and pv plant properties.
Early PV prediction methods were limited by computing power so the data amount is small.
The used features generally include only solar irradiance and historical power generation \cite{PURI1978409, BACHER20091772}.
With the development of computers, the variety of parameters for prediction models has increased to include more meteorological features and PV plant properties.
Meteorological features may include global horizontal irradiance (GHI), diffuse horizontal irradiance (DHI), solar zenith angle, temperature, wind speed, humidity, rainfall etc.
Mellit et al. \cite{MELLIT2010807} used ANN models for grid-connected PV plant prediction, using meteorological features such as temperature, wind speed and cloud density.
Li et al. \cite{LIN202256} observed that weather type had a significant impact on the fluctuations of PV power: high PV production volume and smooth PV production curve on sunny days; Frequent and dramatic fluctuations curve on cloudy and rainy days, with PV production decreased.
They classify the weather conditions based on their similarity, and use meteorological data such as solar irradiance and wind direction as additional features.
PV plant properties include azimuth and tilt angle, location, module type, power ratings, efficiency, area, shading effects, tracking effects, manufacturer and aging etc.
These additional features are important for accurate prediction of PV power \cite{8353805}.
Akhter et al. \cite{AKHTER2022118185} propose a hybrid deep learning model for hour ahead PV prediction, where three different PV systems are based on polycrystalline silicon, monocrystalline silicon and thin film technologies.
Based on the availability of PV system parameters such as PV cell temperature and power, Ayompe et al \cite{AYOMPE20104086} compared different models for PV power prediction accuracy and found that the prediction was good using a combination of PV system parameters.
\subsection{Day ahead PV forecasting}
In summary, the short-term forecasts are useful for day-ahead electricity market guidance and power deployment, and most electricity market prices vary with days, making day-ahead PV forecasts a key research direction.
Among them, ANN-based models are most widely used.
In addition, the data diversity, data resolution and forecast length are also key parts for studies.
Therefore, we surveyed the ANN-based day-ahead PV forecasting studies in recent years, as shown in Table 1.
\begin{table}[htb!]
\caption{Day-ahead PV forecasting in recent years}
\resizebox{\textwidth}{!}{%
\begin{tabular}{lllllll}
\toprule
Author &
Year &
Model &
Input Features &
Resolution &
Accuracy &
Data set \\
\midrule
\begin{tabular}[c]{@{}l@{}}Santos et al \cite{en15145232}* \end{tabular} &
2022 &
TFT &
\begin{tabular}[c]{@{}l@{}}PV power, Solar Irradiance, \\ Temperature, Humidity,\\ Solar Zenith Angle, \\ sine/cosine of month\end{tabular} &
1 h &
\begin{tabular}[c]{@{}l@{}}NRMSE 6.4\%\\ NMAE 3.3\%\end{tabular} &
\href{https://dkasolarcentre.com.au/download?location=alice-springs}{DKASC} \\
\begin{tabular}[c]{@{}l@{}}Qu \cite{QU2021120996}*\end{tabular} &
2021 &
CNN-LSTM &
\begin{tabular}[c]{@{}l@{}}PV power,Solar Irradiance, \\ Temperature, Humidity, \\ Daily rainfall, Wind Direction\end{tabular} &
5-min &
\begin{tabular}[c]{@{}l@{}}NRMSE 6.34\% \\ NMAE 4.20\%\end{tabular} &
\href{https://dkasolarcentre.com.au/download?location=alice-springs}{DKASC} \\
\begin{tabular}[c]{@{}l@{}}Gu \cite{GU2021117291} \end{tabular} &
2021 &
WOA-LSSVM &
\begin{tabular}[c]{@{}l@{}}PV power, Solar Irradiance,\\ Wind Speed,Temperature, \\ Humidity\end{tabular} &
10-min &
\begin{tabular}[c]{@{}l@{}}NRMSE 2.55\% \\ NMAE 2.00\%\end{tabular} &
Private \\
\begin{tabular}[c]{@{}l@{}}Wang \cite{WANG2020112766}\end{tabular} &
2020 &
LSTM-RNN &
PV power &
15-min &
\begin{tabular}[c]{@{}l@{}} NRMSE 6.29\% \\ NMAE 2.78\%\end{tabular} &
\href{https://www.esrl.noaa.gov}{ESRL} \\
\begin{tabular}[c]{@{}l@{}}Miraftabzadeh \cite{9203481}\end{tabular} &
2020 &
LSTM architecture &
PV power, Temperature, Time &
10-min &
NRMSE 3.35\% &
Private \\
\begin{tabular}[c]{@{}l@{}}Theocharides \cite{THEOCHARIDES2020115023} \end{tabular} &
2020 &
ANN &
\begin{tabular}[c]{@{}l@{}}PV power, SolarIrradiance, \\ Temperature, Humidity,\\ Wind Speed and Direction,\\ Solar azimuth and elevation angles\end{tabular} &
1 h &
NRMSE 6.11\% &
Private \\
\bottomrule
\end{tabular}%
}
\end{table}
According to the papers we surveyed in recent years, a significant portion of the researchers used their own private PV power generation data, which is difficult to reproduce and compare.
To compare with our proposed method, two recent studies using the DKASC are collected.
On the other hand, different studies have different predicted power ratings and units, which cannot be compared using metrics such as MAE and RMSE.
Therefore, we used NMAE and NRMSE to normalize the statistical metrics in the above studies from 0 to 100\%.
According to our research, recent day-ahead PV forecasting studies mainly include generation data decomposition, additional meteorological features and pv plant properties, improvement and integration of ANN-based models.
Due to the instability and variability of PV power generation, further research can be conducted on the above-mentioned short-term PV prediction studies.
\begin{itemize}
\item Using signal analysis methods to decompose historical PV data.
Due to the variability of the natural environment, there is periodicity and randomness in the variation of solar energy, resulting in high fluctuation of PV power sequence data.
To achieve high prediction accuracy, many studies using signal analysis methods (e.g. VMD, EMD and EEMD, etc.) to decompose the PV generation series into several subsequences.
However, it is difficult to make out the association between the subsequences and the original series.
The time series decomposition algorithm has the same capability as the above methods, and the results have similar periodicity and volatility as the original series.
\item For PV power generation forecasting, there is no feature categorization based on time series characteristics, and no corresponding model.
PV generation forecasting is an application problem and therefore the features have different specific properties.
For example, the solar angle changes regularly, the orientation of PV plant is fixed, and there are fluctuations when the solar irradiation changes.
Based on our survey, no studies have categorized data based on PV generation forecasting scenarios, and lack of corresponding forecasting models.
\item Recently, there are few studies on day-ahead PV power forecasting.
In contrast, many pappers predict the power point by point, and complete the whole day prediction by iterations.
The day-ahead PV power forecasting requires predicting multiple power points at once.
The day-ahead PV forecasting is more practical than the iterative one, and more difficult to give accurate predictions.
\item There is no comparative analysis under different PV forecasting scenarios.
Depending on the actual forecasting conditions, there are differences in corresponding methods.
For PV site containing multiple plants, total power prediction can be performed either by summing the individual predictions plant by plant, or by directly predicting the total site power.
Depending on the weather station nearby, the data at the forecast time may vary. May include meteorological data such as hourly irradiance and temperature, or only the weather type, or even nothing.
The papers we found did not discuss these cases in detail.
\end{itemize}
To address the above issue, we propose an MSTL and TFT based day-ahead PV power forecasting method.
The main contributions of this study are as follows.
\begin{itemize}
\item The MSTL, a algorithm for time series decomposition, was adopted to process the historical PV power generation data.
For PV prediction problems, common decomposition methods are signal analysis algorithms such as VMD, EMD and EEMD.
MSTL is a decomposition algorithm designed for time series problems, which could yield subsequences that reflect the periodicity of PV power series.
Results show that MSTL outperforms raw data and other decomposition methods when only historical data is used.
\item The data are tagged according to their type and time series properties, which is required by TFT.
We tag the data based on their type and whether it changes over time.
For example, real features and categorical features, time-varying features and static features, etc.
Finally, the tagged data are fed to the TFT model for prediction.
\item We surveyed day-ahead PV forecasting studies in recent years and normalized their results.
The results show that our method has significantly improvement over other studies on the DKASC dataset.
\item For the power generation forecast of PV site, we perform a comparative analysis of possible scenarios.
The results show that the TFT model works better than other common models.
Using MSTL decomposition in cloudy and rainy days, and raw data in sunny days are better than other methods in corresponding weather.
For PV site forecasting, the summation of each PV plant gives better results than the direct forecast.
For NAME and NRMSE on the test set, using meteorological data at prediction have lower values than using only historical data.
\end{itemize}
The remaining parts are structured as follows: in Chapter 2, we present the data preparation, models and evaluation metrics in the method.
The data preparation part includes data description, feature analysis and decomposition algorithms used for PV power sequences.
The model part introduces adopted forecasting model (i.e. LSTM, RNN-LSTM, CNN-LSTM, XGB Regressor and TFT).
The evaluation metrics section gives the formulas for NRMSE and NMAE.
Chapter 3 compares the prediction results of TFT and other models, and the result shows that TFT outperforms than other models.
Next, we perform a comparative analysis of the TFT model for three cases at PV sites, and offer suggestions for practical PV power prediction by MSTL-TFT based on the conclusions.
\section{Materials and Methods}
\subsection{Data preparation}
The data for this study was collected from the DKASC, the largest PV power station in the Southern Hemisphere.
DKASC is located in Alice Spring in central Australia, which is a tropical desert climate, with sunshine all year round and little rain, which has great solar energy resources.
Alice Springs has long summers and short winters.
Between late April and late November, the weather here is normally sunny.
Especially in August, the highest percentage of sunny days compared to other months, reaching more than ninety percent, almost no rainfall.
The remaining five months all have cloudy days.
The month with the highest percentage of cloudy weather compared to other months is February, with an average of half of the time being cloudy or overcast and less rainfall.
Moreover, Alice Springs has different hours of sunlight throughout the year.
The shortest sunshine day is June 21, with about 10 hours, and the longest sunshine day is December 22, with 13.5 hours.
\begin{table}[hbt!]
\caption{PV plants details}
\begin{tabular}{llllll}
\toprule
PV Plant & Manufacturer & Array Rating & PV Technology & Array Structure & Install Date \\
\midrule
PV-01 & BPSolar & 5 & poly-Si & Fixed & 2008 \\
PV-02 & Kaneka & 5.2 & Amorphoussilicon & Fixed & 2008 \\
PV-03 & Solibro & 5.8 & CIGS & Fixed & 2017 \\
PV-04 & SunPower & 5 & mono-Si & Fixed & 2009 \\
PV-05 & BPSolar & 5.1 & poly-Si & Fixed & 2008 \\
PV-06 & BPSolar & 5.3 & mono-Si & Fixed & 2008 \\
PV-07 & Trina & 5.4 & mono-Si & Fixed & 2009 \\
PV-08 & Kyocera & 2 & poly-Si & Fixed & 2008 \\
PV-09 & BPSolar & 6.3 & poly-Si & Fixed & 2008 \\
PV-10 & SunPower & 5 & mono-Si & Fixed & 2011 \\
PV-11 & Sungrid & 5 & mono-Si & Fixed & 2010 \\
PV-12 & Sungrid & 4.9 & poly-Si & Fixed & 2010 \\
PV-13 & EvergreenSolar & 16.8 & poly-Si & Fixed & 2010 \\
\bottomrule
\end{tabular}
\end{table}
We selected the PV power generation data from 13 PV plants in DKASC for this study, and their parameters and ID are listed in Table 2.
The data are collected from 2018 to 2020, with 0.72\% of entire records missing, 0.1\% of individual cell missing, and a few apparent outliers.
We use maximum or minimum to replace the out-of-bounds values, and interpolate the missing values using their adjacent time points.
For detecting outliers, skewness works well.
For example, PV-01 had almost zero solar irradiance and temperature on 2018-03-17, while the PV generation values behaved normally.
These outliers are replaced with the average of the last few identical weather conditions at the same time.
There are many records here as the interval of the original data is 5 minutes.
To reduce data density, we resampled the data and obtain the hourly generated power.
\subsubsection{Adopted features}
In this study, the data contains timestamp, power production, PV plant status and meteorological features.
The interval between each record is one hour.
PV plant status includes technology, array structure and manufacturer, etc.
Meteorological data includes global horizontal irradiation, temperature and rainfall, etc.
In addition, solar angle is important for PV prediction, so we obtained zenith angle and azimuth angle data from Global Monitoring Laboratory (GML) \cite{ESRL-SGC} for the area where DKASC is located.
In addition, derived data including lagged values of historical generation, sine and cosine values of months, seasons and weather types.
Generally, weather could be discerned by the daily solar irradiance.
According to Eq. (1), the index $k_d$ can be calculated from the daily DHI and GHI, and then the weather type can be derived from it.
Zang et al. \cite{ZANG2020105790} use $k_d$ ranges to derive weather type in the DKASC dataset, including sunny, cloudy, and rainy.
We have followed their ranges and listed them in Table 3.
Where $d$ denotes the date and $t$ denotes the hour.
\begin{equation}
k_d = \frac{\sum_t{DHI_t}}{\sum_t{GHI_t}}
\end{equation}
\begin{table}[htb!]
\caption{Weather types and their $k_d$ ranges}
\centering
\begin{tabular}{lll}
\toprule
Weather index & Weather type & k range \\
\midrule
1 & Sunny & 0$\leq k_d \leq$0.15 \\
2 & Partially Cloudy & 0.15$\le k_d \leq$0.45 \\
3 & Overcast/Rainy & 0.45$\le k_d \leq$1 \\
\bottomrule
\end{tabular}%
\end{table}
\begin{table}[htb!]
\caption{Adopted features in this study}
\begin{tabular}{llll}
\toprule
Group & Name & From & Tag \\
\midrule
\multirow{4}{*}{\begin{tabular}[c]{@{}l@{}}Endogenous \\ Features\end{tabular}} &
PV Energy &
PV Site &
Time Varying Unknown Reals \\ \cline{2-4}
&
PV Energy Lags &
Derived &
Time Varying Known Reals \\ \cline{2-4}
&
\begin{tabular}[c]{@{}l@{}}Sine of Month,\\ Cosine of Month\end{tabular} &
Derived &
Time Varying Known Reals \\ \cline{2-4}
&
Season &
Derived &
Time Varying Known Categorical \\\cline{1-4}
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Meteorology \\ Features\end{tabular}} &
\begin{tabular}[c]{@{}l@{}}Global Horizontal Irradiation, \\ Diffuse Horizontal Irradiation, \\ Temperature, Daily Rainfall, \\ Humidity\end{tabular} &
PV Site &
\begin{tabular}[c]{@{}l@{}}Time Varying Unknown Reals\\ Time Varying Known Reals\end{tabular} \\ \cline{2-4}
&
Weather &
Derived &
\begin{tabular}[c]{@{}l@{}}Time Varying Unknown Categorical\\ Time Varying known Categorical\end{tabular} \\ \cline{2-4}
&
\begin{tabular}[c]{@{}l@{}}Solar Zenith Angle,\\ Solar Azimuth Angle\end{tabular} &
GML &
Time Varying Known Reals \\ \cline{1-4}
\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}PV Plant \\ Properties\end{tabular}} &
\begin{tabular}[c]{@{}l@{}}Manufacturer, PV Technology, \\ Array Structure\end{tabular} &
PV Site &
Static Categorical \\ \cline{2-4}
&
Array Rating, Install Date &
PV Site &
Static Reals\\
\bottomrule
\end{tabular}
\end{table}
We divide the features mentioned in this paper into three groups: endogenous features, meteorological features, and PV plant properties.
Time-varying tags and static tags are added to the features depending on whether the features change in the time series or not.
Add known and unknown tags depending on whether the features are knowable in the prediction time period or not.
All features, groups and tags used in this paper are shown in Table 4. They can help to clearly explain the features performance in the time series.
For example, the PV power output is time varying unknown reals, indicating that it is a real variable that varies with time and is unknowable at prediction.
Such labels are required in many Seq2Seq models that combine attention mechanisms, such as the TFT model.
Depending on the data type of the feature, we performed different regularization.
For the category features, we denote them as their respective one-hot codes.
For the real features, we process them using the min-max normalization according to their corresponding maximum and minimum values.
The data in from 2018.1.1 to 2019.12.31 are chosen as the training and validation sets, and the 2020 data as the test set.
The ratio of training set to validation set is 3:1, and the validation set is stochastic selected in proportion from every month in two years.
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{fig/Correlation-Scatters-All.png}
\caption{Relation between PV energy production and meteorological features (normalized)}
\label{fig:Correlation-Scatters-All}
\end{figure}
Fig. 1 represents the correlation between PV power generation and meteorological features such as GHI, highlighting the differences between weathers.
The sunny, cloudy and rainy samples are red, green and blue.
In the first two subplots which representing solar irradiance, black dots indicate outlier points.
From the figure we see that GHI and DHI have a clear correlation with PV power and a large difference along with weather.
Especially for the DHI, it is obvious that the data points boundaries are clearer between different weather.
\subsubsection{PV Energy decomposition}
In signal analysis, complex signal sequences can be decomposed to a number of regular subsequences.
Predicting these subsequences is tended to be easier than predicting the original sequences.
So many studies use signal analysis decomposition methods to reduce the difficulty of analysis and prediction.
Common decomposition algorithms for PV power prediction include EMD, EEMD, VMD and their variants.
In general, the result of the decomposition algorithm includes subsequences and residual, where subsequences represent the derived components, and residual represents the offset between the original value and the sum of all components.
To clearly compare the results of different decomposition methods, we decompose the PV power sequence by VMD, EEMD and VMD-EEMD, which are common in PV power prediction, and the result are shown in Fig. 2 and Fig. 3.
The data used for the plots are taken from 2018, contains all weather types, five days each.
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{fig/eemd_and_vmd.png}
\caption{Result of EEMD and VMD decomposition}
\label{fig:vmd}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{fig/vmd-eemd2.png}
\caption{Result of VMD-EEMD decomposition}
\label{fig:vmd-eemd}
\end{figure}
From the figures we can see that the PV generation sequence is most regular on sunny days, followed by cloudy days, and is volatile when it rains.
The results of the decomposition method showed similar trends in the three weathers.
The frequency of VMD results was moderate, showing periodicity on sunny and cloudy days.
Some subsequence of EEMD and VMD-EEMD have high frequency.
Increasing the number of subsequences in EEMD (some studies even reach 10 or more) can reduce the frequency.
However, an excessive number of subsequences may cause some of them to flatten out, and increase the training time for models.
Although the above signal analysis decomposition algorithms are common for PV power forecasting, all of them can hardly reflect the patterns and periodicity of PV power generation over time.
To tackle this problem, many time decomposition methods have been proposed, such as Prophet \cite{navratil2019decomposition}, STR \cite{dokumentov2015str} and MSTL.
There are two kinds of time series decomposition, which are additive decomposition and multiplicative decomposition \cite{lu2014analysis}.
The additive decomposition divides the time series into the sum of trend, seasonality, and residuals.
The multiplicative decomposition divides the time series into the product of them.
They can be denoted as Eq. (2) and Eq. (3).
Where $T_t$ refers to the trend, which indicates the trend of the time series over a longer period.
$S_t$ refers to the seasonality, which represents the regular fluctuation of the time series in quarterly, monthly, daily, hourly or even shorter periods.
$R_t$ refers to the residuals, which represent the stochastic fluctuation of the time series, with no obvious trend compared to the two components mentioned above.
\begin{equation}
y_t^1 = \hat{T}_t + \hat{S}_t + \hat{R}_t
\end{equation}
\begin{equation}
y_t^2 = \hat{T}_t \times \hat{S}_t \times \hat{R}_t
\end{equation}
Methods such as STL \cite{cleveland1990stl} and X-13-ARIMA-SEATS \cite{doi:10.1080/07350015.1984.10509398} can perform single seasonal decomposition.
However, many time series are affected by several periodic factors, which may not be effective for single seasonal decomposition method.
The author \cite{bandara2021mstl} demonstrate that MSTL has higher accuracy and it is able to model seasonality over time, which requires the lowest computational cost compared to other seasonal decomposition methods.
MSTL is an extension algorithm based on STL, which can be seen as an iterative use of STL to decompose the multiple seasonal components of a time series.
The MSTL formula is shown in Eq. (4), where n refers to the number of seasonal components (e.g. quarter, month, day, etc.).
\begin{equation}
y_t = \hat{S}_t^1 + \hat{S}_t^2 + \cdots +\hat{S}_t^n + \hat{S}_t + \hat{R}_t
\end{equation}
We use the MSTL to decompose the PV power generation sequence, and the result is shown in Fig. 4.
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{fig/mstl2.png}
\caption{Result of MSTL decomposition}
\label{fig:mstl}
\end{figure}
It can be seen that the difference of residuals between MSTL and the other three decomposition algorithms is obvious.
The distribution pattern of MSTL residuals is closer to human expectations, with small total values and little fluctuations in good weather conditions, which is the opposite situation in poor weather conditions.
In addition, the frequency of the MSTL subsequence is the same as the original sequence, while all other decomposition methods have higher frequencies than the original ones.
\subsection{Forecasting models}
\subsubsection{LSTM, RNN-LSTM, CNN-LSTM}
We selected four models commonly used in PV power prediction, which are BP, LSTM, RNN-LSTM, and CNN-LSTM.
In this subsection, we will explain the concepts of these models.
BP model is a kind of classical ANN that refers to a multi-layer feedforward network using the BP algorithm.
It has shown significant improvement in image classification and sequence regression problems compared to other models of the same period.
As a result BP has performed well in early PV forecasts.
For sequence forecasting, RNN have a significant advantage over other ANNs due to their structure of preserving state in sequences.
RNN model passes the information from the previous time step to the next time step by recurring neurons connections.
It effectively associates historical information to current input, and is suitable for processing serial data such as time series, text utterances and speech.
With the layers of the network increases rapidly, the error will keep passing forward during the back propagation, resulting in vanishing gradient.
Due to this problem, RNN models are unable to learn long-term dependencies in sequences.
LSTM can solve the gradient vanishing by adding additional state units and gate control units on top of RNN, which can perform better in medium and long-term sequence prediction tasks.
The RNN-LSTM model is an LSTM layer stacked on bottom of an RNN layer, where the RNN receives the input and the result of the LSTM is used as the model output.
CNN is feedforward neural network commonly used to process data with grid topology.
Time series data and images can be considered as grids of different dimensions.
CNNs are generally composed of convolutional layers, pooling layers and fully connected layers.
The first two layers are used to extract and downscale features, and the last layers could map features to the model output.
CNN can integrate high level features and low level features, which can handle the nonlinear relationship between weather data and PV power sequence.
In CNN-LSTM, the CNN extracts the features and sends them to the LSTM model, and then uses the LSTM model to predict the power output.
\subsubsection{XGB Regressor}
XGBoost is an implementation of boosting in ensemble learning, which forms a strong learner by integrating many tree models together.
The algorithm grows trees and adds trees by splitting features.
The tree used is generally CART tree \cite{lewis2000introduction}, including CART classification tree and CART regression tree.
CART is a binary tree that completes classification or regression by splitting, uses the Gini coefficient to select the optimal features, and prevents overfitting by pruning.
As an implementation of GBDT, XGBoost optimizes the goal by integrating trees, which has good performance in time series prediction problems \cite{10.1145/2939672.2939785}.
According to the tree type XGBoost can also be divided into classifier and regressor.
In the PV power forecasting, the suitable one is XGB Regressor.
\subsubsection{Temporal Fusion Transformer}
TFT uses the attention mechanism in time series forecasting problems, which allows multi-horizon time series forecasting as well as explanatory characterization of time series features.
This is due to the multi-headed attention mechanism in TFT, which can be evaluated for a time horizon or a number of variables.
Similar to the LSTM that uses gates to control the data flow, TFT uses gated residual network (GRN) and gated linear unit (GLU) to filter non-essential information and avoid undesired non-linear processes.
GLU borrows the gating mechanism of LSTM and replaces the activation function with the glu function.
The results are yielded by performing a sigmoid operation on the fully connected network output.
GRN implements the gate select structure by adding the GLU module, using elu activation function and skip connection.
Variable selection network (VSN) is the key structure of TFT, which derives the weights of input features by GRN.
Then the GRN uses soft-max function to sum up the features as the output.
And it combines a self-attentive module that uses a multi-headed attention mechanism to integrate the selected features for output, which is used to detect the important features at each step.
TFT applies different data flow operations to static features, time-varying known features, and time-varying unknown features.
Time-varying features are integrated through VSN, LSTM and gating units, and input to the multi-headed attention mechanism network after GRN.
Static features are integrated through VSN and static encoders, and the final results are also fed into the multi-headed attention mechanism network.
PV forecasting, as a type of time series forecasting, is essentially a regression problem.
Common objective functions for solving regression problems are linear objective functions such as mean squared error.
The variation of PV output power with solar irradiance is non-linear or even non-single trend, as can be seen from Fig. 1.
At this point, the linear objective functions might show large deviations.
Therefore, TFT uses the quantile loss, which calculates the difference between the predicted and target values over a certain range.
The equation for quantile loss is as follows.
\begin{equation}
QuantileLoss = \sum_{q \in Q} max(q(y_i - \hat{y}_i),(q-1)(y_i - \hat{y}_i))
\end{equation}
Which $y_i$ represent real power value of the i-th point,$\hat{y}_i$ represent corresponding predicted value, $Q$ is quantile values, which is a set like \{0.1,0.25,0.5,0.75.0.9\}.
\subsection{Evaluation Metrics}
In PV forecasting studies, the units of RMSE, MAE can be various (e.g. kW, kWH, etc.), depending on the forecasting goals.
To visually evaluate our results and to compare with other studies, we used normalized root mean square error (NRMSE) and normalized mean absolute error (NMAE).
They are in percentages and defined as follows.
\begin{equation}
NRMSE = \sqrt{\frac{1}{Ty_{max}^2} \sum_{i=1}^T(y_i - \hat{y}_i)^2}
\end{equation}
\begin{equation}
NMAE = \frac{1}{Ty_{max}}\sum_{i=1}^T{|y_i - \hat{y}_i|}
\end{equation}
where $T$ is prediction length, $y_i$ represent real power value of the i-th point, $\hat{y}_i$ represent corresponding predicted value, and $y_{max}$ is the maximum of real power values.
\section{Result and discussion}
First, PV site generally contains many PV plants, and their total electrical energy constitutes the output of the site.
For the total output, it can be directly predicted or by the sum of PV plants.
Second, PV sites obtain weather data from their own weather systems or from nearby weather stations.
Some sites can obtain future weather data through weather station, while other sites may only have historical weather data.
Third, the decomposition methods, VMD, EEMD, EEMD-VMD and MSTL, may have different influence on models.
Therefore, we conducted detailed analysis for three cases mentioned above.
This paper studies day-ahead PV forecasting with a one-day forecast horizon and a three-day input horizon for each forecast.
The data resolution is 1 hour and the lagged value of PV power generation is 24 hour.
The 13 PV plants used are named PV-01, PV-02 to PV-13.
For PV site forecasting, the direct forecast is named by Site-Sum, the summation of each PV plant is named by Site-Indiv.
When training, all models are deployed with an early stopping strategy to prevent possible underfitting or overfitting due to constant training epochs.
This chapter is structured as follows.
First, we analyze the performance of BP, LSTM, RNN-LSTM, CNN-LSTM, XGB Regressor and TFT in all cases, shows that TFT is optimal among them.
Next, three cases are discussed: PV site forecasting, meteorological data availability and PV sequence decomposition.
Finally, we give some suggestions for the MSTL-TFT model in practical power prediction through the result from this study.
\subsection{Models comparison}
In this subsection, we will compare the prediction loss of BP, LSTM, RNN-LSTM, CNN-LSTM, XGB Regressor and TFT.
First, we experimented with the output power of the entire PV site and the results are shown in Table 5 and Table 6.
\begin{table}[htb!]
\caption{Models scores when meteorological data is unavailable}
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}lcccccccccc@{}}
\toprule
& \multicolumn{5}{c}{BP} & \multicolumn{5}{c}{LSTM} \\ \midrule
Metric & Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\
NMAE & 7.44\% & 6.66\% & 9.30\% & 7.84\% & \textbf{4.97\%} & \textbf{5.12\%} & 5.54\% & 5.53\% & 5.57\% & 5.59\% \\
NRMSE & 10.15\% & 9.20\% & 11.63\% & 10.67\% & \textbf{7.77\%} & 10.38\% & \textbf{8.83\%} & 10.33\% & 10.37\% & 8.92\% \\ \midrule
& \multicolumn{5}{c}{CNN\_LSTM} & \multicolumn{5}{c}{RNN\_LSTM} \\ \midrule
Metric & Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\
NMAE & \textbf{3.96\%} & 6.02\% & 5.53\% & 5.47\% & 5.03\% & 5.20\% & \textbf{4.88\%} & 5.52\% & 5.50\% & 5.41\% \\
NRMSE & \textbf{8.49\%} & 8.95\% & 10.19\% & 10.08\% & 8.70\% & 10.45\% & \textbf{8.14\%} & 10.50\% & 10.73\% & 8.98\% \\ \midrule
& \multicolumn{5}{c}{XGBRegressor} & \multicolumn{5}{c}{TFT} \\ \midrule
Metric & Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\
NMAE & 3.62\% & 3.95\% & 5.55\% & 6.72\% & \textbf{3.37\%} & 2.82\% & 2.90\% & 5.47\% & 6.19\% & {\color[HTML]{FE0000} \textbf{2.64\%}} \\
NRMSE & 7.69\% & 6.87\% & 8.62\% & 9.84\% & \textbf{6.21\%} & 7.68\% & 6.25\% & 8.83\% & 9.85\% & {\color[HTML]{FE0000} \textbf{5.63\%}} \\ \bottomrule
\end{tabular}%
}
\end{table}
\begin{table}[hbt!]
\caption{Models scores when meteorological data is available}
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}lcccccccccc@{}}
\toprule
& \multicolumn{5}{c}{BP} & \multicolumn{5}{c}{LSTM} \\ \midrule
Metric & Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\
NMAE & 6.15\% & \textbf{4.74\%} & 8.24\% & 8.98\% & 5.22\% & 4.82\% & \textbf{3.90\%} & 4.94\% & 4.79\% & 4.38\% \\
NRMSE & 8.16\% & \textbf{6.51\%} & 9.99\% & 10.93\% & 7.41\% & 8.14\% & \textbf{5.40\%} & 7.29\% & 7.07\% & 6.86\% \\ \midrule
& \multicolumn{5}{c}{CNN\_LSTM} & \multicolumn{5}{c}{RNN\_LSTM} \\ \midrule
Metric & Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\
NMAE & 4.82\% & \textbf{3.59\%} & 5.82\% & 6.90\% & 3.93\% & \textbf{3.65\%} & 5.43\% & 3.80\% & 6.02\% & 4.78\% \\
NRMSE & 7.42\% & \textbf{5.04\%} & 8.65\% & 10.27\% & 6.09\% & \textbf{5.92\%} & 7.77\% & 6.71\% & 11.13\% & 7.25\% \\ \midrule
& \multicolumn{5}{c}{XGBRegressor} & \multicolumn{5}{c}{TFT} \\ \midrule
Metric & Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\
NMAE & \textbf{2.40\%} & 3.41\% & 4.49\% & 5.71\% & 2.88\% & {\color[HTML]{FE0000} \textbf{1.04\%}} & 1.59\% & 3.72\% & 4.84\% & 2.04\% \\
NRMSE & \textbf{4.82\%} & 5.49\% & 6.33\% & 7.77\% & 5.27\% & {\color[HTML]{FE0000} \textbf{2.61\%}} & 2.86\% & 4.97\% & 6.30\% & 3.83\% \\ \bottomrule
\end{tabular}%
}
\end{table}
For all models, the scores in Table 6 is better than the scores in Table 5.
Suggesting that the result is better when meteorological data is available.
As can be seen, TFT is the best of them, XGB is the second.
BP, LSTM, RNN-LSTM, and CNN-LSTM have similar NMAE and NRMSE, which has a significant gap from TFT and XGB.
Therefore, for the next evaluation on all PV Plants, we compare only the TFT model, and the results are shown in Table 7 and Table 8.
\begin{table}[hbt!]
\caption{Loss of TFT when meteorological data is unavailable}
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}lcccccccccc@{}}
\toprule
ID & \multicolumn{5}{c}{NMAE} & \multicolumn{5}{c}{NRMSE} \\ \midrule
& Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\ \cline{2-11}
PV-01 & 2.74\% & 2.87\% & 5.15\% & 6.30\% & \textbf{2.57\%} & 7.76\% & 6.36\% & 8.89\% & 9.89\% & \textbf{5.78\%} \\
PV-02 & 2.67\% & 2.97\% & 5.30\% & 6.12\% & \textbf{2.52\%} & 8.20\% & \textbf{6.19\%} & 9.18\% & 10.20\% & 6.53\% \\
PV-03 & \textbf{2.79\%} & 3.03\% & 5.38\% & 6.38\% & 2.95\% & 8.21\% & 6.68\% & 9.46\% & 10.56\% & \textbf{6.02\%} \\
PV-04 & 2.84\% & 3.20\% & 5.53\% & 6.72\% & \textbf{2.76\%} & 8.30\% & \textbf{6.80\%} & 9.38\% & 10.25\% & 6.91\% \\
PV-05 & 3.12\% & 3.10\% & 5.47\% & 6.47\% & \textbf{2.81\%} & 8.52\% & 7.07\% & 9.79\% & 10.04\% & \textbf{6.93\%} \\
PV-06 & 3.33\% & 3.23\% & 5.65\% & 6.34\% & \textbf{2.95\%} & 8.01\% & \textbf{6.50\%} & 8.77\% & 9.89\% & 6.64\% \\
PV-07 & 3.07\% & 2.98\% & 5.47\% & 6.45\% & \textbf{2.92\%} & 9.79\% & 9.20\% & 10.73\% & 13.79\% & \textbf{7.35\%} \\
PV-08 & 3.02\% & 3.30\% & 6.17\% & 9.19\% & \textbf{2.50\%} & 8.70\% & \textbf{6.37\%} & 9.30\% & 10.35\% & 6.87\% \\
PV-09 & 3.13\% & 3.07\% & 5.59\% & 6.57\% & \textbf{3.03\%} & 8.44\% & 6.64\% & 9.27\% & 10.24\% & \textbf{6.12\%} \\
PV-10 & 3.01\% & 3.15\% & 5.52\% & 6.43\% & \textbf{2.93\%} & 8.32\% & \textbf{5.95\%} & 9.02\% & 10.21\% & 5.99\% \\
PV-11 & 2.92\% & 2.89\% & 5.32\% & 6.36\% & \textbf{2.64\%} & 8.92\% & 6.04\% & 8.76\% & 9.66\% & \textbf{5.74\%} \\
PV-12 & 3.37\% & 3.08\% & 5.26\% & 6.04\% & \textbf{2.86\%} & 8.16\% & 6.19\% & 9.09\% & 9.89\% & \textbf{5.82\%} \\
PV-13 & 2.85\% & 2.89\% & 5.38\% & 6.18\% & \textbf{2.62\%} & 7.62\% & 5.82\% & 7.48\% & 8.02\% & \textbf{5.50\%} \\
Site-Indiv & 2.59\% & 2.56\% & 3.34\% & 3.81\% & \textbf{2.48\%} & 7.68\% & 6.25\% & 8.83\% & 9.85\% & \textbf{5.63\%} \\
Site-Sum & 2.82\% & 2.90\% & 5.47\% & 6.19\% & \textbf{2.64\%} & 8.14\% & 6.36\% & 8.78\% & 10.05\% & \textbf{5.72\%} \\ \bottomrule
\end{tabular}%
}
\end{table}
\begin{table}[hbt!]
\caption{Loss of TFT when meteorological data is available}
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}lcccccccccc@{}}
\toprule
ID & \multicolumn{5}{c}{NMAE} & \multicolumn{5}{c}{NRMSE} \\ \midrule
& Raw & VMD & EEMD & VMD-EEMD & MSTL & Raw & VMD & EEMD & VMD-EEMD & MSTL \\ \cline{2-11}
PV-01 & \textbf{0.94\%} & 1.31\% & 3.89\% & 4.97\% & 1.88\% & \textbf{2.56\%} & 3.34\% & 5.17\% & 6.86\% & 3.67\% \\
PV-02 & \textbf{1.22\%} & 1.47\% & 3.95\% & 4.68\% & 2.15\% & \textbf{2.78\%} & 4.42\% & 5.67\% & 6.40\% & 4.13\% \\
PV-03 & \textbf{1.29\%} & 1.31\% & 3.85\% & 4.82\% & 2.16\% & \textbf{2.50\%} & 3.97\% & 5.40\% & 6.92\% & 4.03\% \\
PV-04 & \textbf{0.94\%} & 1.31\% & 3.91\% & 5.27\% & 2.02\% & \textbf{2.65\%} & 3.53\% & 5.72\% & 7.31\% & 3.91\% \\
PV-05 & \textbf{1.19\%} & 1.39\% & 4.00\% & 4.93\% & 2.09\% & \textbf{2.49\%} & 4.10\% & 5.48\% & 6.76\% & 4.02\% \\
PV-06 & \textbf{1.01\%} & 1.49\% & 3.90\% & 4.97\% & 2.21\% & \textbf{3.00\%} & 3.87\% & 5.55\% & 6.91\% & 4.38\% \\
PV-07 & \textbf{0.96\%} & 1.42\% & 3.87\% & 5.09\% & 2.12\% & \textbf{2.86\%} & 3.45\% & 5.45\% & 6.82\% & 4.09\% \\
PV-08 & \textbf{1.14\%} & 1.58\% & 4.48\% & 7.33\% & 2.11\% & \textbf{3.67\%} & 5.15\% & 6.47\% & 9.77\% & 4.61\% \\
PV-09 & \textbf{1.14\%} & 1.53\% & 3.81\% & 4.77\% & 1.95\% & \textbf{2.92\%} & 3.68\% & 5.42\% & 6.59\% & 3.98\% \\
PV-10 & \textbf{1.20\%} & 1.43\% & 3.82\% & 4.83\% & 2.23\% & \textbf{2.62\%} & 4.08\% & 5.02\% & 6.57\% & 4.17\% \\
PV-11 & \textbf{1.11\%} & 1.40\% & 3.73\% & 4.92\% & 2.30\% & \textbf{2.55\%} & 3.16\% & 5.15\% & 6.61\% & 4.18\% \\
PV-12 & \textbf{1.35\%} & 1.72\% & 3.79\% & 4.65\% & 2.11\% & \textbf{3.04\%} & 3.39\% & 5.08\% & 6.35\% & 3.98\% \\
PV-13 & \textbf{1.11\%} & 1.41\% & 3.85\% & 4.77\% & 2.13\% & \textbf{2.52\%} & 3.24\% & 5.14\% & 6.43\% & 4.14\% \\
Site-Indiv & \textbf{1.04\%} & 1.01\% & 1.86\% & 2.40\% & 1.78\% & \textbf{2.16\%} & 3.12\% & 3.14\% & 3.93\% & 3.52\% \\
Site-Sum & \textbf{1.04\%} & 1.59\% & 3.72\% & 4.84\% & 2.04\% & \textbf{2.61\%} & 2.86\% & 4.97\% & 6.30\% & 3.83\% \\ \bottomrule
\end{tabular}%
}
\end{table}
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{fig/whcihTFT-Norm.png}
\caption{Scores of TFT when meteorological data is unavailable}
\label{fig:whcihTFT-Norm}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{fig/whcihTFT-NWP.png}
\caption{Scores of TFT when meteorological data is available}
\label{fig:whcihTFT-NWP}
\end{figure}
In TFT, when the weather data is unknown the MSTL get most minimum values, when the weather data is known more the raw data get most minimum values.
To show the difference between the two cases above, we have plotted the results for PV site and all PV plants as shown in Fig. 5 and Fig. 6.
The dashes in the graph indicate the decomposition algorithm (raw data or MSTL), and the scatter indicates the scores of PV plants and PV site.
When meteorological data is unavailable, the MSTL curve is lower than the raw data curve in NRMSE.
In the NMAE, the curve of MSTL is lower than that of raw data, except for PV-03.
When meteorological data is available, the MSTL curve is lower than the raw data curve in both NAME and NRMSE.
As can be seen, MSTL outperform than other methods when meteorological data unavailable, which have improvement compared to raw data.
We therefore used MSTL for PV forecasting when meteorological data was unavailable.
Similarly, we use raw data when meteorological data is available.
Later we will analyze the performance of TFT in three cases.
\subsection{Case study on TFT}
\subsubsection{Case 1: Individual PV plant and total output}
\begin{figure}[hbt!]
\centering
\includegraphics[width=\textwidth]{fig/IndivAndSum2.png}
\caption{Scores of MSTL-TFT when meteorological data is unavailable}
\label{fig:IndivAndSum-Norm}
\end{figure}
In the PV plants forecasting, their point distribution is widely dispersed, and their ranking order is not consistent.
When predicting the PV site output, we found that the loss for Site-Indiv is lower than Site-Sum, which indicates that the summation of each PV plant gives better results than the direct forecast.
And Site-Indiv's predictions are also better than any other individual PV plant when meteorological unavailable.
Therefore, we conclude that for PV site forecasting, plant-by-plant prediction could give better result than direct forecast, although it is time consuming for model training.
For the individual PV plant forecasts, their score points are scattered and does not show a consistent pattern in both situations.
\subsubsection{Case 2: Availability of future meteorological data in the prediction}
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{fig/If_NWP_Bar.png}
\caption{Individual PV plant and total output}
\label{fig:Norm-NWP}
\end{figure}
According to the analysis in section 3.1, MSTL outperforms than other decomposition methods when the future meteorological data is unavailable, for raw data is when the meteorological data is available.
In order to compare the differences of TFT models under different meteorological data conditions, we compare their NMAE and NRMSE respectively.
The results are shown in Fig. 8.
For PV site and all PV plants, we can see that the NMAE and NRMSE values when meteorological data are available are nearly one-third of the values when meteorological data are not available, which is a significant decrease.
Therefore, we conclude that when meteorological data are available for the forecast time, the effect is significantly improved over when meteorological data are not available.
\subsubsection{Case 3: PV decomposition on different weather}
To further analyze the impact of data decomposition methods on forecasts, we analyzed the decomposition methods according in different weathers.
The data contains three types of weather: sunny, cloudy and rainy.
In the test set, sunny days accounted for 55.40\%, cloudy days accounted for 28.98\% and rainy days accounted for 15.62
\%.
We counted the losses of each decomposition method under different weather as shown in the following table:
\begin{table}[hbt!]
\caption{Total output when future meteorological is unavailable}
\resizebox{0.8\textwidth}{!}{%
\begin{tabular}{@{}lllllll@{}}
\toprule
& \multicolumn{3}{l}{NMAE} & \multicolumn{3}{l}{NRMSE} \\ \midrule
& Sunny & Cloudy & Rainy & Sunny & Cloudy & Rainy \\ \cline{1-7}
Raw & 1.45\% & 3.18\% & 9.37\% & 4.09\% & 7.72\% & 17.80\% \\
VMD & 1.61\% & 3.12\% & 8.51\% & 2.43\% & 5.72\% & 13.05\% \\
EEMD & 4.11\% & 5.23\% & 12.06\% & 5.28\% & 7.65\% & 19.02\% \\
MSTL & 1.75\% & 2.87\% & 6.15\% & 3.29\% & 5.55\% & 11.39\% \\
EEMD\_VMD & 4.79\% & 6.00\% & 13.15\% & 6.14\% & 8.77\% & 20.88\% \\ \bottomrule
\end{tabular}%
}
\end{table}
\begin{table}[hbt!]
\caption{Total output when future meteorological is available}
\resizebox{0.8\textwidth}{!}{%
\begin{tabular}{@{}lllllll@{}}
\toprule
& \multicolumn{3}{l}{NMAE} & \multicolumn{3}{l}{NRMSE} \\ \midrule
& Sunny & Cloudy & Rainy & Sunny & Cloudy & Rainy \\ \cline{1-7}
Raw & 0.93\% & 1.09\% & 1.34\% & 1.78\% & 2.56\% & 4.48\% \\
VMD & 1.05\% & 1.67\% & 3.29\% & 1.58\% & 2.96\% & 5.22\% \\
EEMD & 3.58\% & 3.64\% & 3.99\% & 4.54\% & 4.89\% & 6.08\% \\
MSTL & 1.61\% & 2.19\% & 3.35\% & 2.81\% & 4.06\% & 6.02\% \\
EEMD\_VMD & 4.67\% & 4.79\% & 5.39\% & 5.97\% & 6.24\% & 7.39\% \\ \bottomrule
\end{tabular}%
}
\end{table}
It can be seen from Table (9) and Table (10) that sunny days are better than cloudy days, and cloudy days are better than rainy days.
This may be due to the regularity of PV generation data on sunny days, which are strongly correlated with time and solar irradiance, and rarely change abruptly.
On sunny days, the lowest value of NMAE is obtained at raw data, and the lowest value of NRMSE is obtained at VMD.
On cloudy and rainy days, when future meteorological is unavailable, the lowest values of NMAE and NRMSE are obtained at MSTL, when future meteorological is available, the lowest values of NMAE and NRMSE are obtained at raw data
In sunny days, we conclude that the NMAE values are better when using raw data and NRMSE is better when using VMD.
In cloudy and rainy days, if future meteorological data is unavailable, then use the MSTL method.
If future meteorological data is available, then use the raw data.
\section{Conclusion}
For the open-sourced PV power generation dataset DKASC, this paper conducts experiments on four data decomposition methods, VMD, EEMD, EEMD-VMD and MSTL, and five models, TFT, BP, LSTM, RNN-LSTM, CNN-LSTM and XGB Regressor.
The results show that the TFT model performs the best among the models.
When the meteorological data in the forecast time is unavailable, the best result of NMAE is 2.48\% and NRMSE is 5.5\%, which are achieved by MSTL-TFT.
When the meteorological data is available, the best result of NMAE is 1.04\% and NRMSE is 2.61\%, which are achieved by Raw-TFT.
These results are better than any of the other studies we have surveyed on day-ahead DKASC PV forecasting.
In addition, we analyze the effects of some study cases on TFT PV prediction, which are summarized as follows.
\begin{itemize}
\item For power forecasting for PV site that contain a lot of PV plants, the summation of each PV plant gives better results than the direct forecast.
\item For the data decomposition method of the PV generation series, the optimal method depends on the weather conditions and the availability of meteorological data.
When meteorological data is available, it is better to use raw data.
When meteorological data is not available, then the optimal method is selected based on weather type.
For sunny days, using raw data could get lower NMAE values, and using VMD could get lower NRMSE values.
For cloudy and rainy days, MSTL makes the best prediction.
\end{itemize}
Based on the results of this paper, we make the following recommendations for the day-ahead PV generation forecast.
\begin{itemize}
\item In the PV generation forecast, the TFT performs better than other state of the art models such as CNN-LSTM and XGB Regressor.
\item For PV site forecasting by TFT, the summation of each PV plant gives better results than the direct forecast.
\item The meteorological data for the forecast time is very important and has a significant improvement compared to no meteorological data.
If meteorological data is available, then forecasting should use raw data for forecasting.
\item Very often, it is not possible to obtain detailed meteorological data for the forecast time, but only the weather type.
At this point, the forecast can be based on the next day's weather type.
When the weather is sunny, the raw data should be used.
When the weather is cloudy and rainy, the MSTL method should be used.
\end{itemize}
\newpage
\bigskip
\paragraph*{Acknowledgments.}
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.
\paragraph*{Data Availability Statement.}
Data from DKASC opened data. Code will be opened after receivion.
\printbibliography
\end{document} |
\section{Introduction and statement of results}
This paper is concerned with the fourth Painlev\'e (PIV) equation
\begin{equation}\label{PIV}
\frac{\mathrm{d}^2q}{\mathrm{d}x^2}
=\frac{1}{2q}\left(\frac{\mathrm{d}q}{\mathrm{d}x}\right)^{2}
+\frac{3}{2}q^{3}+4xq^{2}+2(x^{2}-2\alpha)q,
\end{equation}
where
the parameter
$\alpha\in\mathbb{R}$. We are interested in the real solutions of \eqref{PIV}
satisfying the asymptotic condition
\begin{equation}\label{Boundcondit}
q(x)\rightarrow 0\quad \mathrm{as}\quad x\rightarrow +\infty.
\end{equation}
A result due to Bassom \emph{et al.} \cite{BCHM} shows that any real solution of the PIV equation \eqref{PIV}, fulfilling the boundary condition \eqref{Boundcondit}, possesses the following asymptotic behavior
\begin{equation}\label{qasy}
q(x)\sim \kappa D_{\alpha-\frac{1}{2}}^{2}(\sqrt{2}x)\quad \mathrm{as}\quad x\to+\infty
\end{equation}
for some constant $\kappa\in\mathbb{R}$. Here, $D_{\mu}(x)$ denotes the parabolic cylinder function,
which is the solution of Weber's equation
(cf. \cite[Chapter 12]{NIST})
\begin{align*}
\frac{\mathrm{d}^2D_{\mu}(x)}{\mathrm{d}x^2}=\left(\frac{1}{4}x^2
-\mu-\frac{1}{2}\right)D_{\mu}(x),
\end{align*}
uniquely characterized by the asymptotic property
\begin{equation}\label{eq:DAsy}
D_{\mu}(x)=x^{\mu}e^{-\frac{1}{4}x^2}\left(1+O\left(x^{-2}\right)\right),\quad \mathrm{as}\quad x\to+\infty.
\end{equation}
Conversely,
for any constant $\kappa\in\mathbb{R}$, there exists a unique solution to the PIV equation \eqref{PIV} asymptotic to $\kappa D_{\alpha-\frac{1}{2}}^{2}(\sqrt{2}x) $
as $x\to +\infty$.
For convenience, we denote such solutions by $q(x)=q(x;\alpha,\kappa)$.
Concerning the asymptotic behaviors of $q(x;\alpha,\kappa)$ as $x\rightarrow-\infty$, there has been the following conjecture.
\subsection*{Conjecture (Clarkson-McLeod \cite{CM})}
There exists a constant $\kappa^*>0$ such that:
\begin{description}
\item{(a)} When $0<\kappa<\kappa^{*}$ and as $x\rightarrow-\infty$, we have
\begin{equation}\label{qAsy1}
q(x;\alpha,\kappa)\sim c_n\,2^{\alpha-\frac{1}{2}} x^{2 \alpha-1} e^{-x^{2}}
\end{equation}
if $\alpha+\frac{1}{2}=n+1\in \mathbb{N}$ with $\mathbb{N}$ being the set of all positive integers; and
\begin{equation}\label{qAsy2}
q(x;\alpha,\kappa)\sim -\frac{2 x}{3}+\frac{4d_1}{\sqrt{3}} \sin \left(\frac{x^{2}}{\sqrt{3}}-\frac{4 d_1^{2}}{\sqrt{3}}\ln(\sqrt{2}|x|)+d_2+O\left(x^{-2}\right)\right)+O\left(x^{-1}\right)
\end{equation}
if $\alpha-\frac{1}{2}\notin \mathbb{Z}$, where the constants $c_n$, $d_1$, $d_2$ are dependent on $\kappa$.
\item{(b)} When $\kappa=\kappa^{*}$, $q(x;\alpha,\kappa)$ is asymptotic to $-2x$ as $x\rightarrow-\infty$.
\item{(c)} When $\kappa>\kappa^{*}$, $q(x;\alpha,\kappa)$ has a pole at some point $x$ on the real axis.
\end{description}
The solutions $q(x;\alpha,\kappa)$ in this class are now called the \emph{Clarkson-McLeod solutions} of the PIV equation \eqref{PIV}.
In the case $\alpha+\frac{1}{2}\in\mathbb{N}$ of part (a),
the asymptotic formula \eqref{qAsy1} has been proven in \cite{BCHM,CM}. Therein, the values of $c_n$ and $\kappa^{*}$ were explicitly evaluated by
\begin{equation}\label{kappastarinteger}
c_n=\frac{\kappa}{1-\sqrt{\pi}\,n!\,\kappa},\qquad
\kappa^*=\frac{1}{\sqrt{\pi}\,n!}.
\end{equation}
While $\alpha-\frac{1}{2}\notin \mathbb{Z}$, the value of $\kappa^*$ in \cite{CM} was conjectured to be
\begin{equation}\label{kappastar}
\kappa^*=\frac{1}{\sqrt{\pi}\, \Gamma\left(\alpha+\frac{1}{2}\right)}.
\end{equation}
For $\alpha+\frac{1}{2}\in\mathbb{N}$, each solution $q(x)=q(x;\alpha,\kappa)$ of the PIV equation \eqref{PIV} can be explicitly expressed in terms of the classical special functions. The behavior of the solution as $x\to-\infty$ can then be determined in a straightforward manner. For example, when $\alpha=\frac{1}{2}$, the exact solution is given by (see \cite[Equation (4.13)]{BCHM})
\begin{equation}\label{specialsolu}
q\left(x; {1}/{2},\kappa\right)=\frac{2\kappa\,e^{-x^2}}{2-\kappa\sqrt{\pi}\,\mathrm{erfc}(x)},
\quad x\in \mathbb{R},\end{equation}
where $\mathrm{erfc}(x)=\frac{2}{\sqrt{\pi}}\int^{\infty}_{x}e^{-t^2}dt$ is the complementary error function; cf. \cite[Chapter 7]{NIST}.
It is readily verified that $\kappa^*=1/\sqrt{\pi}$. The approximation \eqref{qAsy1} is valid since $\mathrm{erfc}(x)$ is strictly monotone decreasing on $\mathbb{R}$ with range $(0, 2)$. Similarly, it holds that $q\left(x; {1}/{2},1/\sqrt\pi \right)\sim -2x$ as $x\to-\infty$. For $\kappa\sqrt \pi >1$, the solution has a real pole which is the zero of the denominator in
\eqref{specialsolu}.
For general $\alpha+\frac{1}{2}\in\mathbb{N}$, the solutions and their asymptotic approximations as given in \eqref{qAsy1} can be obtained by using \eqref{specialsolu} and the B\"{a}cklund transformation; see \cite[Equation (5.9)]{BCHM}, or \cite[Equation (3.24)]{BCH}.
The solutions $q(x; \alpha, \kappa)$ with positive half-integers $\alpha$
find prominent applications in the theory of orthogonal polynomials with the discontinuous Hermite weight \cite{Chen}, and in random matrix theory \cite{FW,TW}.
It is well known that for the Gaussian unitary ensemble of $n\times n$ Hermitian matrices,
the probability of having no eigenvalues in the interval $(x,+\infty)$ can be expressed in terms of the Fredholm determinant
\begin{equation}
\det\left(\mathbf{I}-K_{n,x}\right).
\end{equation}
Here, $K_{n,x}$ is the integrable operator acting on $L^2(0,+\infty)$ with the classical Hermite kernel
\begin{equation}\label{eq:OPkernel}
K_{n,x}(\lambda,\mu)=e^{-\frac{(x+\lambda)^2+(\mu+x)^2}{2}} \gamma_{n-1}^{2}\frac{\pi_n(\lambda+x)\pi_{n-1}(\mu+x)-\pi_{n-1}(\lambda+x)\pi_{n}(\mu+x)}{\lambda-\mu}, \end{equation}
where $\pi_n(\lambda)$ is the $n$-th monic Hermite polynomial
determined by the orthogonal relation
\begin{equation}\label{eq:OP} \int_{\mathbb{R}}\pi_{n}(x)\pi_{m}(x)e^{-x^2}dx=\gamma_{n}^{-2}\delta_{n,m},
\end{equation}
with the normalization constant
\begin{equation}\label{eq:gamman}\gamma_{n-1}^{2}=\frac{2^{n-1}}{\sqrt{\pi}\,\Gamma(n)};
\end{equation}
see \cite[Table 18.3.1]{NIST}.
It is found by Tracy and Widom in \cite{TW} that
\begin{equation}\label{eq:detH}
\frac{d}{dx}\ln \det(\mathbf{I}-K_{n,x})=\sigma_{n}(x)
\end{equation}
where $\sigma_{n}(x)$ is the unique solution of the $\sigma$-form of the PIV equation \cite{JM} with the parameter $\nu=n$
\begin{equation}\label{eq:sPIV}
\left(\sigma_{\nu}^{\prime \prime}\right)^{2}+4\left(\sigma_{\nu}^{\prime}\right)^{2}\left(\sigma_{\nu}^{\prime}+2 \nu\right)-4\left(x \sigma_{\nu}^{\prime}-\sigma_{\nu}\right)^{2}=0,
\end{equation}
characterized by the boundary condition
\begin{equation}\label{eq:sPIVAsy}
\sigma_{\nu}(x)\sim \frac{2^{\nu-1} x^{2 \nu-2} e^{-x^{2}}}{\sqrt{\pi}(\nu-1) !},\quad x\to+\infty.
\end{equation}
It is worth mentioning that $\sigma_{\nu}(x)$ is closely related to the Hamiltonian for the PIV equation \eqref{PIV}; see Remark \ref{rem:Hq} below and \cite{JM}.
For general $\alpha-\frac{1}{2}\notin \mathbb{Z}$,
the asymptotic formula \eqref{qAsy2} in part (a) was proved by Abdullayev \cite{Ab} using the integral equation method. The connection formulas, that is, the explicit expressions of the parameters $d_1$ and $d_2$ in \eqref{qAsy2}, in terms of the parameter $\kappa$, were derived later by Its and Kapaev \cite{IK}, and by Wong and Zhang \cite{WZ}, using respectively the isomonodromy method and the uniform asymptotic approach.
Moreover, according to the numerical investigations performed
in \cite{BCH1}, $q(x;\alpha,\kappa)$ might blow up at finite $x$ if $\alpha<-1/2$. While $\alpha>-1/2$, the same numerical results allow us to expect the absence of the real poles of $q(x;\alpha,\kappa)$.
The asymptotic behavior of $q(x;\alpha,\kappa)$ and the connection formulas in part (c) were recently derived by the current authors in \cite{XXZ}.
While, to the best of our knowledge, the asymptotic result in part (b) of the Clarkson-McLeod conjecture has not been confirmed.
It is also desirable to know whether there exists a determinantal representation of the $\sigma$-form of the Clarkson-McLeod solutions similar to \eqref{eq:detH}, for general $\alpha$ not being a half-integer.
In the present paper, we derive the asymptotic approximations and the connection formulas for the Clarkson-McLeod solutions as $x\to-\infty$ by using the Deift-Zhou nonlinear steepest descent method \cite{Deft,DZ,DZ1}. Particularly, we prove part (b) and revisit part (a), (c) of the Clarkson-McLeod conjecture.
We also show that the $\sigma$-form of the Clarkson-McLeod solutions of the PIV equation \eqref{PIV} with general parameter $\alpha$ can be represented by the Fredholm determinant of an integrable operator whose kernel is expressed in terms of the classical parabolic cylinder functions, thus generalizing the result \eqref{eq:detH} of Tracy and Widom.
Furthermore, the asymptotics of the Hamiltonian of the Clarkson-McLeod solutions and the evaluations of total integrals of the Clarkson-McLeod solutions are also obtained.
\subsection{Statement of results}
\subsubsection*{Asymptotics of the Clarkson-McLeod solutions}
Our first result is the following complete description of the asymptotic behaviors of the Clarkson-McLeod solutions to the PIV equation \eqref{PIV} when the parameter $\alpha\in\mathbb{R}$ with $\alpha+\frac{1}{2}\notin\mathbb{N}$.
\begin{thm}\label{thm1}
Let $\kappa\neq 0$ be a given real number and $\kappa^*$ be the constant defined by \eqref{kappastar}. For any $\alpha\in\mathbb{R}$ with $\alpha+\frac{1}{2}\notin\mathbb{N}$, there exists a unique real solution $q(x;\alpha,\kappa)$ to the PIV equation \eqref{PIV} satisfying the following asymptotic behavior
\begin{equation}\label{qAsy}
q(x;\alpha,\kappa)=\kappa\, 2^{\alpha-\frac{1}{2}}x^{2\alpha-1}e^{-x^2}\left(1+O\left(x^{-2}\right)\right)\quad \mathrm{as}\quad x\to+\infty.
\end{equation}
For $\alpha\in\mathbb{R}$ with $\alpha-\frac{1}{2}\notin\mathbb{Z}$, the solution $q(x;\alpha,\kappa)$ possesses the following asymptotic behaviors as $x\rightarrow-\infty$.
\begin{description}
\item{(1)} If $\kappa(\kappa-\kappa^*)<0$, then
\begin{equation}\label{q1}
q(x;\alpha,\kappa)=-\frac{2}{3}x+\frac{2\sqrt{6}\,b_1}{3}\sin \left(\frac{x^{2}}{\sqrt{3}}-\frac{b_1^{2}}{\sqrt{3}}\ln(2\sqrt{3} x^{2})+\psi_1\right)+O\left(\frac{1}{x}\right).
\end{equation}
\item{(2)} If $\kappa=\kappa^*$, then
\begin{equation}\label{q2}
q(x;\alpha,\kappa)=-2x+O\left(\frac{1}{x}\right).
\end{equation}
\item{(3)} If $\kappa(\kappa-\kappa^*)>0$, then
\begin{equation}\label{q3}
q(x;\alpha,\kappa)=-\frac{2}{3}x+\frac{2x}{2\cos\left(\frac{x^{2}}{\sqrt{3}}
-\frac{b_2}{\sqrt{3}}\ln \left(2 \sqrt{3} x^{2}\right)+\psi_2\right)+1}+O\left(\frac{1}{x}\right).
\end{equation}
\end{description}
The error term in the asymptotic expansion \eqref{q3} is uniform for $x$ bounded away from the singularities appearing on the right-hand side of the asymptotic expansion. Moreover, the corresponding connection formulas are respectively given by
\begin{equation}\label{connectionformula1}
\left\{
\begin{aligned}
b_1^{2}&=-\frac{\sqrt{3}}{2\pi} \ln(1-\left|\rho\right|^{2}),\quad b_1\geq0, \\
\psi_1&=-\frac{ \pi}{4}-\frac{2 \pi}{3}\alpha-\arg \Gamma\left(-\frac{b_1^{2}}{\sqrt{3}}i \right)-\arg \rho,
\end{aligned}\right.
\end{equation}
and
\begin{equation}\label{connectionformula2}
\left\{
\begin{aligned}
b_2&=-\frac{\sqrt{3}}{2\pi} \ln(|\rho|^{2}-1),\\
\psi_2&=-\frac{2\pi}{3}\alpha
-\arg\Gamma\left(-\frac{b_2}{\sqrt{3}}i+\frac{1}{2}\right)-\arg \rho,
\end{aligned}\right.
\end{equation}
where
\begin{equation}\label{rhoandkappa}
\rho=1-\frac{2\pi^{\frac{3}{2}}}{e^{\pi i\alpha}\Gamma\left(\frac{1}{2}-\alpha\right)}\kappa.
\end{equation}
When $\frac{1}{2}-\alpha\in \mathbb{N}$, it holds $\kappa^*=0$, and $q(x;\alpha,\kappa)$ possesses the asymptotic behavior \eqref{q3}
for $\kappa\in \mathbb{R}\setminus\{0\}$
as $x\rightarrow-\infty$.
\end{thm}
\begin{rem}\label{rem-1.2}
For all $\alpha\in\mathbb{R}$, the remaining case $\kappa=0$ corresponds to the trivial solution $q(x;\alpha,0)=0$. As aforementioned,
the asymptotic formula \eqref{q1} and connection formulas \eqref{connectionformula1} have been derived in \cite{IK,WZ}. In this paper, we give alternative proofs of these formulas by performing asymptotic analysis of the Riemann-Hilbert (RH, for short) problem for PIV equation. Furthermore, we accomplish the case $b_1=0$ in \eqref{q1} and \eqref{connectionformula1} which is not covered in \cite[Equation (1.10)]{IK} and \cite[Equation (1.5)]{WZ}. We also show that \eqref{q3} is true for negative half-integer $\alpha$, a case not considered in our previous work \cite{XXZ}.
We also provide a novel proof of \eqref{q2}. Theorem \ref{thm1}, along with
the case with positive half-integer $\alpha$ solved in terms of the special functions in \cite{BCHM,CM}, fully confirms the conjecture of Clarkson and McLeod. Minor extension allows $\kappa^*$ to be negative or zero.
\end{rem}\vskip .3cm
The asymptotic analysis of the RH problem for the PIV equation enables us to derive simultaneously the asymptotics of the Hamiltonian associated with the Clarkson-McLeod solutions and the total integrals of Clarkson-McLeod solutions along the real axis, which are of independent interests.
\begin{thm}\label{thm2}
Under the conditions in Theorem \ref{thm1}, for any $\alpha\in\mathbb{R}$ with $\alpha+\frac{1}{2}\notin\mathbb{N}$, the Hamiltonian $\mathcal{H}(x;\alpha,\kappa)$ of the Clarkson-McLeod solutions, defined by \eqref{eq:Hq} below,
has the following asymptotic behavior as $x\rightarrow +\infty$:
\begin{equation}\label{Hasymp+infty}
\mathcal{H}(x;\alpha,\kappa)=-\kappa\,2^{\alpha-\frac{1}{2}}
x^{2\alpha-1}e^{-x^2}\left(1+O\left(x^{-2}\right)\right).
\end{equation}
While for $\alpha\in\mathbb{R}$ with $\alpha-\frac{1}{2}\notin\mathbb{Z}$, $\mathcal{H}(x;\alpha,\kappa)$ satisfies the following asymptotic behaviors as $x\rightarrow-\infty$.
\begin{description}
\item{(1)} If $\kappa(\kappa-\kappa^*)<0$, then
\begin{equation}\label{H1}
\mathcal{H}(x;\alpha,\kappa)=-\frac{8x^3}{27}+\frac{4}{3}
\left(\alpha+b_1^2\right)x
-\frac{2\sqrt{2}\,b_1}{3}\cos
\left(\frac{x^{2}}{\sqrt{3}}-\frac{b_1^{2}}{\sqrt{3}}\ln (2 \sqrt{3}x^{2} )+\psi_1\right)+O\left(\frac{1}{x}\right),
\end{equation}
\item{(2)} If $\kappa=\kappa^*$, then
\begin{equation}\label{H2}
\mathcal{H}(x;\alpha,\kappa)=4\alpha x+O\left(\frac{1}{x}\right),
\end{equation}
\item{(3)} If $\kappa(\kappa-\kappa^*)>0$, then
\begin{equation}\label{H3}
\mathcal{H}(x;\alpha,\kappa)=-\frac{8x^3}{27}+\frac{4}{3}(\alpha+b_2)x
+\frac{\frac{4}{\sqrt{3}}x\sin\left(\frac{x^{2}}{\sqrt{3}}-\frac{b_2}{\sqrt{3}}\ln \left(2 \sqrt{3} x^{2}\right)+\psi_2\right)}
{2\cos\left(\frac{x^{2}}{\sqrt{3}}-\frac{b_2}{\sqrt{3}}\ln \left(2 \sqrt{3} x^{2}\right)+\psi_2\right)+1}+O\left(\frac{1}{x}\right).
\end{equation}
\end{description}
The parameters $b_1$, $\psi_1$ and $b_2$, $\psi_2$ are the same as in \eqref{connectionformula1} and \eqref{connectionformula2}, respectively. The error term in the asymptotic expansion \eqref{H3} is uniform for $x$ bounded away from the singularities appearing on the right-hand side of the asymptotic expansion.
Moreover, when $\frac{1}{2}-\alpha\in \mathbb{N}$, for any non-vanishing
$\kappa\in \mathbb{R}$, $\mathcal{H}(x;\alpha,\kappa)$ possesses the same asymptotic behavior \eqref{H3} as $x\rightarrow-\infty$.
\end{thm}
Our next result is the evaluation of the total integrals of the Clarkson-McLeod solutions $q(x;\alpha,\kappa)$. Similar results for the integrals of the Painlev\'e II transcendents have been derived in \cite{BBD,BBDI,DXZ,Kok,Mil}.
\begin{thm}\label{thm3}
Let $\mathrm{P.V.}$ denote the Cauchy principal value. For $\alpha\in\mathbb{R}$ with $\alpha-\frac{1}{2}\notin\mathbb{Z}$, we evaluate the total integrals of $q(x;\alpha,\kappa)$ as follows.
\begin{description}
\item{(1)} If $\kappa(\kappa-\kappa^{*})<0$, take $c < 0 < d$ such that all real poles of $q(x;\alpha,\kappa)$ lie in the interval $(c,d)$,
\begin{align}\label{integral-1}
&\exp\Bigg\{\int^{c}_{-\infty}\left(q(t;\alpha,\kappa)+\frac{2t}{3}
-\frac{2\alpha}{t}\right)dt+\mathrm{P.V.}\int_{c}^{d}q(t;\alpha,\kappa)\,dt
+\int^{+\infty}_{d}q(t;\alpha,\kappa)\,dt\Bigg\}\nonumber\\
&\qquad =\frac{(-1)^{N_{+}-N_{-}}\sqrt{\pi}e^{\frac{c^2}{3}}|c|^{-2\alpha}3^{2\alpha}(1-\rho)e^{\pi i\alpha}}{2^{\frac{1}{2}-\alpha}\Gamma(\frac{1}{2}-\alpha)(1-|\rho|^2)^{\frac{2}{3}}},
\end{align}
where $\rho$ is related to $\kappa$ by \eqref{rhoandkappa} and $N_{\pm}$ denote the numbers of real poles
of $q(x;\alpha,\kappa)$ in the interval $(c,d)$ of residues $\pm1$, respectively.
\item{(2)} If $\kappa=\kappa^{*}$, take $c< 0 < d$ such that all real poles of $q(x;\alpha,\kappa)$ lie in the interval $(c,d)$,
\begin{align}\label{integral-2}
&\exp\Bigg\{\int^{c}_{-\infty}\left(q(t;\alpha,\kappa)+2t
+\frac{2\alpha}{t}\right)dt
+\mathrm{P.V.}\int_{c}^{d}q(t;\alpha,\kappa)\,dt
+\int^{+\infty}_{d}q(t;\alpha,\kappa)\,dt\Bigg\}\nonumber\\
&\qquad =\frac{(-1)^{N_{+}-N_{-}} \sqrt{\pi}e^{-c^2}
|c|^{-2\alpha}2^{\frac{1}{2}+\alpha}}{\Gamma(\frac{1}{2}+\alpha)},
\end{align}
where $N_{\pm}$ denote the numbers of real poles
of $q(x;\alpha,\kappa)$ of residues $\pm1$ in the interval $(c,d)$.
\end{description}
\end{thm}
\begin{rem}
According to the numerical analysis performed in \cite{BCH1}, it is expected that $q(x;\alpha,\kappa)$ is pole free on the real axis when
$0<\kappa\leq\kappa^*$. However, to the best of our knowledge, there is no rigorous proof of the numerical evidence.
The Cauchy principal values in \eqref{integral-1} and \eqref{integral-2} may be removed if one can prove that $q(x;\alpha,\kappa)$
is pole free on the real axis in these cases.
\end{rem}
\subsubsection*{Determinantal representation of the Clarkson-McLeod solutions}
Let $\gamma K_{\nu,x}$ be the integrable operator acting on $L^2(0,+\infty)$ with the parabolic cylinder kernel
\begin{equation}\label{eq:PCKernel}
\gamma K_{\nu, x}(\lambda,\mu)=\gamma \frac{D_{\nu}(\sqrt{2}(\lambda+x))D_{\nu-1}(\sqrt{2}(\mu+x))
-D_{\nu-1}(\sqrt{2}(\lambda+x))D_{\nu}(\sqrt{2}(\mu+x))}{\lambda-\mu},
\end{equation}
where $D_{\nu}$ is the parabolic cylinder function with the parameter $\nu \in \mathbb{R}$ and $\gamma$ is a real parameter. Then, the Fredholm determinant of the operator is related to the $\sigma$-form of the Clarkson-McLeod solutions as stated in the following theorem.
\begin{thm}\label{thm:IntRep}
Suppose that $\nu\in\mathbb{R}$, $\gamma\in \mathbb{R}$ and $\gamma K_{\nu,x}$ is the integrable operator acting on $L^2(0,+\infty)$ with the kernel \eqref{eq:PCKernel}. We have
\begin{equation}\label{eq:IntRep}
\frac{d}{dx} \ln \det(\mathbf{I}-\gamma K_{\nu, x})=\sigma_{\nu}(x;\gamma),\end{equation}
where $\sigma_{\nu}(x;\gamma)$ is the unique solution of the $\sigma$-form \eqref{eq:sPIV} of the PIV equation,
determined by the boundary condition near positive infinity
\begin{equation}\label{eq:SigmaAsy}
\sigma_{\nu}(x;\gamma)\sim \sqrt{2}\gamma D_{\nu-1}^2(\sqrt{2}x).
\end{equation}
\end{thm}\vskip .5cm
\begin{rem}\label{rem:Hq}
It is seen from \cite[Equations (C.34)-(C.36)]{JM} and \eqref{eq:Hq} below that the $\sigma$-form of the PIV equation is related to $q(x;\alpha,\kappa)$ and the Hamiltonian $\mathcal{H}(x;\alpha,\kappa)$ by
\begin{equation}\label{eq:sigmaH}
\sigma_{\alpha+\frac{1}{2}}(x;\gamma)=\frac{1}{2}\left(q(x;\alpha,\kappa)-\mathcal{H}(x;\alpha,\kappa)\right),
\end{equation}
and conversely
\begin{equation}\label{eq:qsigma}q(x;\alpha,\kappa)= -\frac{\sigma''_{\alpha+\frac{1}{2}}(x;\gamma)+2x\sigma_{\alpha+\frac{1}{2}}'(x;\gamma)-2\sigma_{\alpha+\frac{1}{2}}(x;\gamma)}{2\sigma_{\alpha+\frac{1}{2}}'(x;\gamma)+4\alpha+2}.\end{equation}
It follows from the approximation \eqref{eq:SigmaAsy} and the relation \eqref{eq:qsigma} that $q(x;\alpha,\kappa)$ actually satisfies the boundary condition \eqref{qAsy},
and the parameters $\kappa$ and $\gamma$ are related by
\begin{equation}\label{eq:sigmaKappa}
\kappa=\sqrt{2}\gamma.
\end{equation}
In view of \eqref{eq:sigmaH}, the asymptotic behavior near negative infinity of the solution of the $\sigma$-form \eqref{eq:sPIV} of the PIV equation subject to the boundary condition \eqref{eq:SigmaAsy} can be obtained by using Theorems \ref{thm1} and \ref{thm2}.
\end{rem}
\begin{rem}
For $\nu=n\in\mathbb{N}$, the parabolic cylinder function is reduced to the Hermite polynomial \cite[(12.7.2)]{NIST}
\begin{equation}\label{eq:Hermite}
D_{n}(\sqrt{2}x)=e^{-\frac{1}{2}x^2} 2^{\frac{n}{2}} \pi_{n}(x),
\end{equation}
where $\pi_{n}(x)$ is the monic Hermite polynomial defined through
\eqref{eq:OP}. Therefore, for given $\nu\in\mathbb{N}$, the kernel $\gamma^{*} K_{\nu, x}(\lambda,\mu)$ with
\begin{equation}\label{eq:gammaSta} \gamma^{*}=\frac{1}{\sqrt{2\pi}\Gamma(\nu)}\end{equation}
is reduced to the classical Hermite kernel defined in \eqref{eq:OPkernel}.
Applying Theorem \ref{thm:IntRep}, we recover \eqref{eq:detH}-\eqref{eq:sPIVAsy}, which was obtained first by Tracy and Widom.
\end{rem}
The rest of the present paper is arranged as follows. In Section \ref{sec:RHPforPIV}, we state the RH problem for the PIV equation \eqref{PIV} and express the solutions of the PIV equation and the associated Hamiltonian in terms of the solution to this RH problem. Subsequently, in Sections \ref{Asymptotic-infty1}-\ref{Asymptotic-infty3}, we apply the Deift-Zhou nonlinear steepest descent method to the mentioned RH problem as $x\to-\infty$,
for the parameter $\kappa$ respectively in three different regimes. Using the asymptotic analysis of the RH problems we performed, Theorems \ref{thm1}-\ref{thm3} are then proved in Section \ref{proof1}. The final Section \ref{sec:proof of det} is devoted to the proof of Theorem \ref{thm:IntRep}. For the convenience of the reader, we collect in the Appendix three local parametrix models used in the asymptotic analysis of the RH problems.
\section{Riemann-Hilbert problem for the Painlev\'e IV equation}\label{sec:RHPforPIV}
In this section, we review the RH problem for the PIV equation \eqref{PIV}, a detailed description can be found in \cite[Section 2]{IK} and \cite[Chapter 5.1]{FIKN}.
Denote $\Sigma=\bigcup^{8}_{k=1}\gamma_{k}$, where $\gamma_{k}=\{\xi\in \mathbb{C}\mid\arg\xi=k\pi/4\}$; see Figure \ref{PIVj}. Then, the $2\times2$ matrix-valued function $\Psi(\xi,x)$ solves the following RH problem.
\subsection*{RH problem for $\Psi(\xi,x)$}
\begin{description}
\item{(1)} $\Psi(\xi,x)$ is analytic for all $\xi\in\mathbb{C}\setminus \Sigma$.
\item{(2)} $\Psi(\xi,x)$ satisfies the jump relations
$$
\Psi_{+}(\xi,x)=\Psi_{-}(\xi,x)\left\{
\begin{aligned}
&S_{k},\quad &\xi&\in\gamma_{k},\ k=1,\cdots,7,\\
&S_{8}e^{-2\pi i\alpha\sigma_{3}},\quad &\xi&\in\gamma_{8},
\end{aligned}
\right.
$$
where the Stokes matrices
\begin{equation}\label{eq:Sk}
S_{2i-1}=
\begin{pmatrix}
1 & s_{2i-1}\\
0 & 1
\end{pmatrix},\quad
S_{2i}=
\begin{pmatrix}
1 & 0\\
s_{2i} & 1
\end{pmatrix}, \quad \ i=1,2,3,4.
\end{equation}
The Stokes multipliers $s_k$ satisfy the following restrictions
\begin{equation}\label{srela1}
s_{k+4}=-s_{k}e^{(-1)^{k}2\pi i\alpha},\quad k=0,1,2,3,4,
\end{equation}
and
\begin{equation}\label{srela2}
[(1+s_{3}s_{4})(1+s_{1}s_{2})+s_{1}s_{4}]e^{-i\pi\alpha}-(1+s_{2}s_{3})e^{i\pi \alpha}=-2i\sin(\pi\alpha).
\end{equation}
\item{(3)} As $\xi\to\infty$, $\Psi(\xi,x)$ satisfies the following asymptotic condition
\begin{equation}\label{Asyatinfty}
\Psi(\xi,x)=\Psi^{(\infty)}(\xi,x)
e^{\Theta(\xi,x)\sigma_{3}},
\end{equation}
where
\begin{equation}\label{Psiinfty}
\Psi^{(\infty)}(\xi,x)=\mathbf{I}+\frac{\Psi_{1}}{\xi}+\frac{\Psi_{2}}{\xi^2}
+O\left(\frac{1}{\xi^{3}}\right),\quad \Theta(\xi,x)=\frac{1}{8}\xi^{4}+\frac{x}{2}\xi^{2}+\alpha\ln\xi,
\end{equation}
with the branch of $\ln\xi$ chosen so that $\arg\xi\in(0,2\pi)$.
\item{(4)} As $\xi\to0$, $\Psi(\xi,x)$ has the asymptotic behavior of the form
\begin{equation}\label{Asyatzero}
\Psi(\xi,x)=\left\{\begin{aligned}
&\Psi^{(0)}(\xi,x)\xi^{\alpha\sigma_3}E, && \mathrm{for}\quad \alpha-\frac{1}{2}\notin \mathbb{Z},\\
&\Psi^{(0)}(\xi,x)\xi^{\alpha\sigma_3}\begin{pmatrix} 1 & \frac{s_0}{\pi i}\ln\xi \\ 0 & 1\end{pmatrix}E,&& \mathrm{for}\quad \alpha+\frac{1}{2}\in \mathbb{N},\\
&\Psi^{(0)}(\xi,x)\xi^{\alpha\sigma_3}\begin{pmatrix} 1 & 0 \\ -\frac{s_0}{\pi i}\ln\xi & 1\end{pmatrix}E,&& \mathrm{for}\quad \frac{1}{2}-\alpha\in \mathbb{N},
\end{aligned}\right.
\end{equation}
where $\Psi^{(0)}(\xi,x)$ is analytic in the neighborhood of $\xi=0$.
The functions $\xi^{\alpha}$ and $\ln\xi$ take principal values. The connection matrix $E$ is given by
\begin{equation}\label{connectionmatrix}
E=E_0S_{0 }\cdots S_{k-1},\quad\xi\in\Omega_k,\quad k=1,\cdots,8,
\end{equation}
where $S_0:=\mathbf{I}$ and the regions $\Omega_k=\left\{\xi\in\mathbb{C}\mid\arg\xi\in(\frac{(k-1)\pi}{4},\frac{k\pi}{4})\right\}$ are depicted in Figure \ref{PIVj}.
Moreover, the connection matrix $E_0$ takes the form
\begin{equation}\label{E0}
E_0=\left\{\begin{aligned}
&\begin{pmatrix} 1 & 0 \\ \frac{s_0e^{2\pi i\alpha}}{e^{2\pi i\alpha}+1} & 1
\end{pmatrix}, && \mathrm{for}\quad \alpha-\frac{1}{2}\notin\mathbb{Z},\\
&\begin{pmatrix} p_1 & -1 \\ 1 & 0
\end{pmatrix}, && \mathrm{for}\quad \alpha+\frac{1}{2}\in\mathbb{N},\\
&\begin{pmatrix} 1 & 0 \\ p_2 & 1
\end{pmatrix}, && \mathrm{for}\quad \frac{1}{2}-\alpha\in\mathbb{N},
\end{aligned}\right.
\end{equation}
where $p_1$, $p_2$ are two arbitrary nonzero constants.
\end{description}
In the RH formulation, we denote the Pauli matrices by $\sigma_k$, $k=1,2,3$
\begin{equation}\label{Pauli}
\sigma_1=\begin{pmatrix}0 & 1\\ 1 & 0\end{pmatrix},\quad \sigma_2=\begin{pmatrix}0 & -i\\ i & 0\end{pmatrix},\quad \sigma_3=\begin{pmatrix}1 & 0\\ 0 & -1\end{pmatrix}.
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=7cm,height=7cm]{PIVjump}\\
\caption{The jump contour $\Sigma$ and the regions $\Omega_k$}\label{PIVj}
\end{figure}
A significant fact is that
the PIV transcendents and the associated Hamiltonian can be expressed in terms of the solution of the RH problem for $\Psi(\xi,x)$.
\begin{pro}\label{eq:PIVRHP}
The solution of the PIV equation \eqref{PIV} and the associated Hamiltonian are related to the solution to above RH problem for $\Psi(\xi,x)$ by
\begin{equation}\label{qsolu1}
q(x)=(\Psi_{1})_{12}(\Psi_{1})_{21},
\end{equation}
and
\begin{equation}\label{tau}
\mathcal{H}(x)
=\left(\Psi_2\right)_{11}-\left(\Psi_2\right)_{22},
\end{equation}
where $\Psi_1$ and $\Psi_2$ are given in \eqref{Psiinfty}.
Let $\Psi^{(0)}(\xi,x)$ be defined as \eqref{Asyatzero}, we have
\begin{equation}\label{Fequation}
F(x):=\Psi^{(0)}(0,x)=\exp\bigg[\int^x q(x)dx\; \sigma_3\bigg].
\end{equation}
\end{pro}
\begin{proof}
The relations \eqref{qsolu1} and \eqref{Fequation} have been derived in \cite[Equations (2.11) and (2.40)]{IK}.
According to Jimbo, Miwa and Ueno \cite{JMU}, it follows from \eqref{Asyatinfty} and \eqref{Psiinfty} that
\begin{equation}\label{tau-Res}
\mathcal{H}(x)=-\mathop{\mathrm{Res}}\limits_{z=\infty}
\mathrm{Tr}\left(\Psi^{(\infty)}(\xi,x)^{-1}
\frac{\mathrm{d}\Psi^{(\infty)}(\xi,x)}{\mathrm{d}\xi}
\frac{\mathrm{d}\Theta(\xi,x)}{\mathrm{d}x}\sigma_3\right)
=\left(\Psi_2\right)_{11}-\left(\Psi_2\right)_{22}.
\end{equation}
\end{proof}
We mention that using the Lax pair for $\Psi(\xi,x)$ \cite[Equations (2.1)-(2.2)]{IK}, the Hamiltonian $\mathcal{H}(x)$ can be expressed in terms of $q$
as follows
\begin{equation}\label{eq:Hq}
\mathcal{H}(x)=\frac{q^3}{4}+xq^2+(x^2-2\alpha)q-\frac{(q')^2}{4q}.
\end{equation}
Let $p=\frac{q'}{q}$, then
\begin{equation}\label{eq:H}
H(x):=-2\mathcal{H}(x)=-\frac{q^3}{2}-2xq^2-2(x^2-2\alpha)q+\frac{1}{2}p^2q.
\end{equation}
The PIV equation \eqref{PIV} can be derived by eliminating $p$ from the Hamilton equations
\begin{equation}\label{eq:Hequ}
\frac{dq}{dx}=\frac{\partial H}{\partial p},\quad \frac{dp}{dx}=-\frac{\partial H}{\partial q}.
\end{equation}
Applying \eqref{eq:Hq}-\eqref{eq:Hequ}, we also find the relation
\begin{equation}\label{eq:dH}
\frac{d}{dx}\mathcal{H}(x)=\frac{\partial }{\partial x}\mathcal{H}(x)=q^2+2xq.
\end{equation}
It was observed by Its and Kapaev \cite[Equation (2.23)]{IK} that, for any real solution of the PIV equation \eqref{PIV}, the Stokes multipliers satisfy the conditions
\begin{equation}\label{assum1}
\overline{s}_0=s_0,\quad \overline{s}_1=-e^{2\pi i\alpha}s_{3}.
\end{equation}
When $\alpha\in\mathbb{R}$ with $\alpha-\frac{1}{2}\notin\mathbb{Z}$, Its and Kapaev proved that for solutions of the PIV equation \eqref{PIV} determined by the asymptotic behavior \eqref{qAsy} with real parameter $\kappa$, the Stokes multipliers fulfill the following conditions (see \cite[Theorem 3.1]{IK})
\begin{equation}\label{assum2}
s_2=0,\quad s_1+s_3=0,\quad s_*\neq1, \quad (1-s_*)e^{\pi i\alpha}\in\mathbb{R},
\end{equation}
where $s_*$ is constituted by the Stokes multipliers $s_0$, $s_1$ through
\begin{equation}\label{sstar}
s_*=s_0s_1+1.
\end{equation}
The connection formula between
the Stokes multipliers
and the parameter $\kappa$ in the asymptotic behavior \eqref{qAsy}
was also derived therein
\begin{equation}\label{kapparep}
\kappa=-\frac{(s_*-1)e^{\pi i\alpha}}{2\pi^{\frac{3}{2}}}\Gamma
\left(\frac{1}{2}-\alpha\right).
\end{equation}
Actually, the relation \eqref{kapparep} is also true for $\frac{1}{2}-\alpha\in \mathbb{N}$.
For the case $\alpha-\frac{1}{2}\notin\mathbb{Z}$, combining \eqref{kappastar} with \eqref{kapparep}, we find the following correspondence between the conditions on the parameter $\kappa$ and the Stokes multiplier $s_*$ as shown in Table
\ref{table}.
\begin{table}[H]
\renewcommand\arraystretch{1}
\centering
\begin{tabular}{|c|c|}
\hline
$\kappa$ & $s_*$ \\
\hline
$\kappa(\kappa-\kappa^*)<0$ & $|s_*|<1$ \\
\hline
$\kappa=\kappa^*$ & $|s_*|=1,\ s_*\neq1$ \\
\hline
$\kappa(\kappa-\kappa^*)>0$ & $|s_*|>1$ \\
\hline
\end{tabular}
\caption{The correspondence between $\kappa$ and $s_*$}\label{table}
\end{table}
We point out that the exceptional case $s_*=1$ in the table is equivalent to $\kappa=0$ and the solution determined by \eqref{qAsy} is trivial; see Remark \ref{rem-1.2}. While, for the case $\frac{1}{2}-\alpha\in\mathbb{N}$, it follows from \eqref{assum1}, \eqref{sstar} and \eqref{kapparep} that we always have
$$|s_*|^2=1+\frac{4\pi^3\kappa^2}{\Gamma(\frac{1}{2}-\alpha)^2}>1.$$
For $\alpha+\frac{1}{2}\in\mathbb{N}$, as mentioned in the introduction, the solution of the PIV equation \eqref{PIV} determined by \eqref{qAsy} can be expressed in terms of the complementary error function.
These special solutions correspond to the specified values of Stokes multipliers (see \cite{Kap98})
\begin{equation}\label{SpecialsoluStokes}
s_0=s_2=0,\quad s_1+s_3=0.
\end{equation}
\section{RH analysis as $x\to-\infty$ with $0\leq |s_*|<1$}\label{Asymptotic-infty1}
In this section, we start by carrying out the Deift-Zhou nonlinear steepest descent analysis of the RH problem for $\Psi$ as $x\to-\infty$ under the condition $0< |s_*|<1$. Then, the reduced case $|s_*|=0$ is considered at the end of this section.
Assume now that $x<0$. We start with the first transformation
\begin{equation}\label{rescaling}
\Phi(z)=(-x)^{-\frac{\alpha}{2}\sigma_3}
\Psi\left((-x)^{\frac{1}{2}}z,x\right).
\end{equation}
Immediately, $\Phi(z)$ solves the following RH problem.
\subsection*{RH problem for $\Phi(z)$}
\begin{description}
\item{(1)} $\Phi(z)$ is analytic for all $z\in\mathbb{C}\setminus\widetilde\Sigma$, where $\widetilde\Sigma=\Sigma\setminus(\gamma_2\cup\gamma_6)$; see Figure \ref{PIVj}.
\item{(2)} $\Phi(z)$ satisfies the same jump conditions as $\Psi(\xi,x)$ on $\widetilde\Sigma$.
\item{(3)} As $z\to\infty$,
\begin{equation}\label{Asyatinfty1}
\Phi(z)=\left(\mathbf{I}+\frac{\Phi_{1}}{z}+\frac{\Phi_{2}}{z^2}+O(z^{-3})\right)
z^{\alpha\sigma_3}e^{x^2(\frac{1}{8}z^{4}-\frac{1}{2}z^{2})\sigma_{3}},
\end{equation}
where the branch of $z^{\alpha}$ is chosen such that $\arg z\in(0,2\pi)$.
\item{(4)} $\Phi(z)$ has the same asymptotic behaviors as $\Psi(\xi,x)$ as $z\to0$; see \eqref{Asyatzero} and \eqref{connectionmatrix}.
\end{description}
Simultaneously, by \eqref{qsolu1}, \eqref{tau} and \eqref{rescaling}, we get
\begin{equation}\label{solu2}
q(x;\alpha,\kappa)=-x(\Phi_{1})_{12}(\Phi_{1})_{21},
\end{equation}
and
\begin{equation}\label{tau1}
\mathcal{H}(x;\alpha,\kappa)=-x\left[(\Phi_2)_{11}-(\Phi_2)_{22}\right],
\end{equation}
where $\Phi_1=\Phi_1(x)$ and $\Phi_2=\Phi_2(x)$ are the coefficients in \eqref{Asyatinfty1}.
Introduce the following $g$-function
\begin{equation}\label{g-function}
g(z)=\frac{1}{8}z\left(z^{2}-\frac{8}{3}\right)^{\frac{3}{2}},
\end{equation}
where $\arg\left(z\pm\textstyle\sqrt{ {8}/{3}}\right)\in(-\pi,\pi)$. By straightforward computation,
\begin{equation}\label{gatinfty}
g(z)=\frac{1}{8}z^4-\frac{1}{2}z^2+\frac{1}{3}+\frac{4}{27}z^{-2}+O(z^{-4})
\end{equation}
as $z\rightarrow\infty$. Moreover, $g(z)$ has four saddle points
\begin{equation*}
z_{1,\pm}=\pm\sqrt{\frac{2}{3}},\quad z_{2,\pm}=\pm\sqrt{\frac{8}{3}}.
\end{equation*}
We then make the second transformation
\begin{equation}\label{U(z)}
\mathbf{U}(z)=e^{\frac{x^2}{3}\sigma_{3}}\Phi(z)e^{-x^2g(z)\sigma_{3}}.
\end{equation}
It is direct to check that $\mathbf{U}(z)$ satisfies the following RH problem.
\subsection*{RH problem for $\mathbf{U}(z)$}
\begin{description}
\item{(1)} $\mathbf{U}(z)$ is analytic for $z\in\mathbb{C}\setminus\widetilde\Sigma$.
\item{(2)} We have the jump relations $\mathbf{U}_{+}(z)=\mathbf{U}_{-}(z)J_\mathbf{U}(z)$, where
$$J_\mathbf{U}(z)=\left\{
\begin{aligned}
&\begin{pmatrix}1 & s_ke^{2x^2g(z)}\\ 0 & 1 \end{pmatrix},\quad &z&\in\gamma_{k},\ k=1,3,5,7,\\
&\begin{pmatrix}e^{x^2(g_-(z)-g_+(z))} & 0\\ -s_0e^{2\pi i\alpha}e^{-x^2(g_+(z)+g_-(z))} & e^{x^2(g_+(z)-g_-(z))} \end{pmatrix},\quad &z&\in\gamma_{4},\\
&\begin{pmatrix}e^{-2\pi i\alpha}e^{x^2(g_-(z)-g_+(z))} & 0\\ s_0e^{2\pi i\alpha}e^{-x^2(g_+(z)+g_-(z))} & e^{2\pi i\alpha}e^{x^2(g_+(z)-g_-(z))} \end{pmatrix},\quad &z&\in\gamma_{8}.
\end{aligned}
\right.
$$
\item{(3)} $\mathbf{U}(z)$ satisfies the asymptotic condition
$$\mathbf{U}(z)=\left(\mathbf{I}+O(z^{-1})\right)z^{\alpha\sigma_3},\quad \mathrm{as}\quad z\rightarrow\infty,$$
where $\arg z\in(0,2\pi)$.
\item{(4)} $\mathbf{U}(z)$ satisfies the following asymptotic behavior as $z\to0$
\begin{equation}\label{Uatzero}
\mathbf{U}(z)=\mathbf{U}_{0}(z)z^{\alpha\sigma_3}E_0S_1e^{-x^2g(z)\sigma_3},\qquad \arg z\in(\pi/4, {\pi}/{2}),
\end{equation}
where $\mathbf{U}_{0}(z)$ is analytic in the neighborhood of $z=0$. The asymptotic behaviors of $\mathbf{U}(z)$ in other regions are determined by \eqref{Uatzero} and the jump relations satisfied by $\mathbf{U}(z)$.
\end{description}
\begin{figure}[t]
\centering
\includegraphics[width=13cm,height=8cm]{ASC}\\
\caption{The anti-Stokes curves of the exponent $g(z)$}\label{ASC}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=13cm,height=7cm]{Trans1}\\
\caption{The first deformation of the jump contour} \label{Deforma1}
\end{figure}
\subsection{Deformations of the jump contour}\label{sec:RHT}
Next, we transform the RH problem for $\mathbf{U}(z)$ to a RH problem formulated on the anti-Stokes curves of $g(z)$, as illustrated in Figure \ref{ASC}.
To this aim, we first notice that the RH problem for $\mathbf{U}(z)$ is equivalent to the one posed on the contour shown in Figure \ref{Deforma1}, where we use the notations $\widetilde{S}_k$ to denote the
analytically extended jump matrices $J_{\mathbf{U}}(z)$.
Secondly, since the jump matrices on $(z_{2,-},z_{2,+})$ are now oscillating for large $|x|$, we should deform the segment $(z_{2,-},z_{2,+})$ to the anti-Stokes curves of $g(z)$. Thus, we introduce the third transformation $\mathbf{U}\rightarrow \mathbf{T}$. This transformation is based on the following factorizations of the jump matrices on $(z_{2,-},z_{2,+})$.
\begin{align}\nonumber
\widetilde{S}_{4}&=\begin{pmatrix}e^{x^2(g_-(z)-g_+(z))} & 0\\ -s_0e^{2\pi i\alpha} & e^{x^2(g_+(z)-g_-(z))}\end{pmatrix}\\ \nonumber
&=\begin{pmatrix}1 & -s_0^{-1}e^{-2\pi i\alpha}e^{2x^2g_-(z)} \\ 0 & 1\end{pmatrix}
\begin{pmatrix}0 & s_0^{-1}e^{-2\pi i\alpha} \\ -s_0e^{2\pi i\alpha} & 0\end{pmatrix}
\begin{pmatrix}1 & -s^{-1}_0e^{-2\pi i\alpha}e^{2x^2g_+(z)} \\ 0 & 1\end{pmatrix}\\
&=:\widetilde{S}_{U_1}\widetilde{S}_{P_-}\widetilde{S}_{U_2}, \label{decom1}
\end{align}
\begin{align}\nonumber
(\widetilde{S}_3\widetilde{S}_4\widetilde{S}_5)^{-1}&=\begin{pmatrix}s_*
e^{x^2(g_-(z)-g_+(z))} & s_1(e^{-2\pi i\alpha}+s_*)\\ s_0e^{2\pi i\alpha} & \overline{s}_*e^{x^2(g_+(z)-g_-(z))}\end{pmatrix}\\ \nonumber
&=\begin{pmatrix}1 & 0 \\ \frac{\overline{s}_*e^{-2x^2g_-(z)}}{s_1(e^{-2\pi i\alpha}+s_*)} & 1\end{pmatrix}
\begin{pmatrix}0 & (|s_*|^2-1)s_0^{-1}e^{-2\pi i\alpha}\\ s_0e^{2\pi i\alpha}(1-|s_*|^2)^{-1} & 0\end{pmatrix}
\\ \nonumber
&~~~~\times \begin{pmatrix}1 & 0 \\ \frac{s_*e^{-2x^2g_+(z)}}{s_1(e^{-2\pi i\alpha}+s_*)} & 1\end{pmatrix}\\ \label{decom3}
&=:\widetilde{S}_{L_1}\widetilde{S}_P\widetilde{S}_{L_2},
\end{align}
\begin{align} \nonumber
\widetilde{S}_7\widetilde{S}_8\widetilde{S}_1&=\begin{pmatrix}\overline{s}_*
e^{-2\pi i\alpha}e^{x^2(g_-(z)-g_+(z))} & s_1(e^{-2\pi i\alpha}+s_*) \\
s_0e^{2\pi i\alpha} & s_*e^{2\pi i\alpha}e^{x^2(g_+(z)-g_-(z))}\end{pmatrix}\\ \nonumber
&=\begin{pmatrix}1 & 0 \\ \frac{s_*e^{2\pi i\alpha}e^{-2x^2g_-(z)}}{s_1(e^{-2\pi i\alpha}+s_*)} & 1\end{pmatrix}
\begin{pmatrix}0 & (|s_*|^2-1)s_0^{-1}e^{-2\pi i\alpha}\\ s_0e^{2\pi i\alpha}(1-|s_*|^2)^{-1} & 0\end{pmatrix}
\\ \nonumber
&~~~~\times
\begin{pmatrix}1 & 0 \\ \frac{\overline{s}_*e^{-2\pi i\alpha}e^{-2x^2g_+(z)}}{s_1(e^{-2\pi i\alpha}+s_*)} & 1\end{pmatrix}\\ \label{decom4}
&=:\widetilde{S}_{L_3}\widetilde{S}_P\widetilde{S}_{L_4},
\end{align}
and
\begin{align} \nonumber
\widetilde{S}_{8}&=\begin{pmatrix}e^{-2\pi i\alpha}e^{x^2(g_-(z)-g_+(z))} & 0\\ s_0e^{2\pi i\alpha} & e^{2\pi i\alpha}e^{x^2(g_+(z)-g_-(z))} \end{pmatrix}\\ \nonumber
&=\begin{pmatrix}1 & s_0^{-1}e^{-4\pi i\alpha}e^{2x^2g_-(z)} \\ 0 & 1\end{pmatrix}
\begin{pmatrix}0 & -s_0^{-1}e^{-2\pi i\alpha} \\ s_0e^{2\pi i\alpha} & 0\end{pmatrix}
\begin{pmatrix}1 & s^{-1}_0e^{2x^2g_+(z)} \\ 0 & 1\end{pmatrix}\\
&=:\widetilde{S}_{U_3}\widetilde{S}_{P_+}\widetilde{S}_{U_4}.\label{decom2}
\end{align}
In the above factorizations, we have utilized the property
$$g_+(z)+g_-(z)=0, \quad \mathrm{for} \quad z\in (z_{2,-},z_{2,+}), $$
and the complex conjugate relation
\begin{equation}\label{sstarbar}
\overline{s}_*=s_0s_1e^{2\alpha\pi i}+1.
\end{equation}
Now,
we define the third transformation $\mathbf{U}\mapsto \mathbf{T}$ as
\begin{equation}\label{T(z)}
\mathbf{T}(z)=\left\{\begin{aligned}
&\mathbf{U}(z), &\mathrm{for}\ &z\ \mathrm{outside\ the\ four\ lens\ regions},\\
&\mathbf{U}(z)\widetilde{S}_{U_2}, &\mathrm{for}\ &z\ \mathrm{in\ the\ upper\ part\ of\ the\ first\ lens\ region},\\
&\mathbf{U}(z)\widetilde{S}_{U_1}^{-1}, &\mathrm{for}\ &z\ \mathrm{in\ the\ lower\ part\ of\ the\ first\ lens\ region},\\
&\mathbf{U}(z)\widetilde{S}_{L_2}^{-1}, &\mathrm{for}\ &z\ \mathrm{in\ the\ upper\ part\ of\ the\ second\ lens\ region},\\
&\mathbf{U}(z)\widetilde{S}_{L_1}, &\mathrm{for}\ &z\ \mathrm{in\ the\ lower\ part\ of\ the\ second\ lens\ region},\\
&\mathbf{U}(z)\widetilde{S}_{L_4}^{-1}, &\mathrm{for}\ &z\ \mathrm{in\ the\ upper\ part\ of\ the\ third\ lens\ region},\\
&\mathbf{U}(z)\widetilde{S}_{L_3}, &\mathrm{for}\ &z\ \mathrm{in\ the\ lower\ part\ of\ the\ third\ lens\ region}\\
&\mathbf{U}(z)\widetilde{S}_{U_4}^{-1}, &\mathrm{for}\ &z\ \mathrm{in\ the\ upper\ part\ of\ the\ fourth\ lens\ region},\\
&\mathbf{U}(z)\widetilde{S}_{U_3}, &\mathrm{for}\ &z\ \mathrm{in\ the\ lower\ part\ of\ the\ fourth\ lens\ region},
\end{aligned}
\right.
\end{equation}
where the lens regions are illustrated in Figure \ref{Deforma2} and the same notations are used to denote the analytic extensions of the corresponding jump matrices.
\begin{figure}[t]
\centering
\includegraphics[width=13cm,height=7cm]{Trans2}\\
\caption{The second deformation of the jump contour} \label{Deforma2}
\end{figure}
Immediately, $\mathbf{T}(z)$ solves a RH problem whose jump contour and jump matrices are shown in Figure \ref{Deforma2}.
We proceed to
blow up the four lens in Figure \ref{Deforma2}. Consequently, we obtain the following equivalent RH problem for $\mathbf{T}(z)$.
\begin{figure}[t]
\centering
\includegraphics[width=13cm,height=8cm]{Trans3}\\
\caption{The jump contour $\Sigma_\mathbf{T}$ } \label{Deforma3}
\end{figure}
\subsection*{RH problem for $\mathbf{T}(z)$}
\begin{description}
\item{(1)} $\mathbf{T}(z)$ is analytic for $z\in \mathbb{C}\setminus\Sigma_{\mathbf{T}}$, where $\Sigma_\mathbf{T}$ is shown in Figure \ref{Deforma3}.
\item{(2)} We have $\mathbf{T}_+(z)=\mathbf{T}_-(z)J_k(z)$ for $z\in\pi_k$, $k=1,\cdots,20$, where
\begin{align*}
&J_1(z)=\begin{pmatrix}1 & -s_0^{-1}e^{2x^2g(z)}\\ 0 & 1 \end{pmatrix},
&&
J_2(z)=\begin{pmatrix}1 & s_*s_0^{-1}e^{2x^2g(z)}\\ 0 & 1 \end{pmatrix},\\
&J_3(z)=\begin{pmatrix}1 & 0\\ -\frac{\overline{s}_*e^{-2\pi i\alpha}e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix}, &&
J_4(z)=\begin{pmatrix}1 & 0\\ \frac{(e^{-2\pi i\alpha}-1)e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix},\\
&J_5(z)=\begin{pmatrix}1 & 0\\ \frac{s_*e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix}, &&
J_6(z)=\begin{pmatrix}1 & -\overline{s}_*s_0^{-1}e^{-2\pi i\alpha}
e^{2x^2g(z)} \\ 0 & 1 \end{pmatrix},\\
&J_7(z)=J_9(z)=\begin{pmatrix}1 &
\frac{e^{2x^2g(z)}}{s_0e^{2\pi i\alpha}} \\ 0 & 1 \end{pmatrix},&&
J_8(z)=\begin{pmatrix}1 & 0\\ -s_0e^{2\pi i\alpha}e^{-2x^2g(z)} & 1 \end{pmatrix},\\
&J_{10}(z)=\begin{pmatrix}1 & -\frac{s_*
e^{2x^2g(z)}}{s_0e^{2\pi i\alpha}}\\ 0 & 1 \end{pmatrix}
,&&
J_{11}(z)=\begin{pmatrix}1 & 0\\ \frac{\overline{s}_*e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix},\\
&J_{12}(z)=\begin{pmatrix}1 & 0\\ \frac{(e^{2\pi i\alpha}-1)e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix},&&
J_{13}(z)=\begin{pmatrix}1 & 0\\ -\frac{s_*e^{2\pi i\alpha}e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix},\\
&J_{14}(z)=\begin{pmatrix}1 & \overline{s}_*s_0^{-1}e^{-4\pi i\alpha}e^{2x^2g(z)} \\ 0 & 1 \end{pmatrix},
&&
J_{15}(z)=\begin{pmatrix}1 & -s_0^{-1}e^{-4\pi i\alpha}e^{2x^2g(z)} \\ 0 & 1 \end{pmatrix},\\
&J_{16}(z)=\begin{pmatrix}e^{-2\pi i\alpha} & 0\\ s_0e^{2\pi i\alpha}e^{-2x^2g(z)} & e^{2\pi i\alpha} \end{pmatrix},&&
J_{17}(z)=J_{20}(z)^{-1}=\begin{pmatrix}0 & s_0^{-1}e^{-2\pi i\alpha}\\ -s_0e^{2\pi i\alpha} & 0 \end{pmatrix},
\end{align*}
and
\begin{equation*}
J_{19}(z)=J_{18}(z)^{-1}=\begin{pmatrix}0 & -s_0^{-1}e^{-2\pi i\alpha}(1-|s_*|^2) \\ s_0e^{2\pi i\alpha}(1-|s_*|^2)^{-1} & 0\end{pmatrix}.
\end{equation*}
\item{(3)} As $z\rightarrow \infty$, $\mathbf{T}(z)=\left(\mathbf{I}+O(z^{-1})\right)z^{\alpha\sigma_3}$, where $\arg z\in(0,2\pi)$.
\item{(4)} As $z\to 0$, $\mathbf{T}(z)$ has the following asymptotic behavior
\begin{equation}\label{T0}
\mathbf{T}(z)=\mathbf{T}_0(z)z^{\alpha\sigma_{3}}
E_0S_1e^{-x^2g(z)\sigma_{3}}J_3(z),\quad \arg z\in (\pi/4, {\pi}/{2}),
\end{equation}
where $\mathbf{T}_0(z)$ is analytic in the neighborhood of $z=0$. The behaviors of $\mathbf{T}(z)$ in other regions are determined by \eqref{T0} and the jump relations satisfied by $\mathbf{T}(z)$.
\end{description}
Taking into account the lower and upper triangular structures of the jump matrices, the properties of $\mathop{\rm Re}\nolimits g(z)$ on the anti-Stokes curves (see Figure \ref{ASC}), and the fact that $\mathop{\rm Re}\nolimits g(z)>0$ on the imaginary axis, we see that the jump matrices on $\pi_k$, $k=1,\cdots,15$ of the RH problem for $\mathbf{T}(z)$ tend to the identity matrix exponentially fast when $x\rightarrow -\infty$. Therefore, it is expected that the
dominant contribution to the asymptotics of $\mathbf{T}(z)$ as $x\rightarrow -\infty$ will come from the remaining jumps along the line $(z_{2,-},+\infty)$ and the neighborhoods of the origin $z=0$ and the four saddle points $z_{1,\pm}=\pm\sqrt{2/3}$, $z_{2,\pm}=\pm\sqrt{8/3}$. We are in a position to construct the global parametrix solving the constant jumps along $(z_{2,-},+\infty)$ and local parametrices near the origin and the four saddle points.
\subsection{Global parametrix }
We need to solve the following RH problem.
\subsection*{RH problem for $\mathbf{P}^{(\infty)}(z)$}
\begin{description}
\item{(1)} $\mathbf{P}^{(\infty)}(z)$ is analytic for $z\in \mathbb{C}\setminus [z_{2,-},+\infty)$.
\item{(2)} $\mathbf{P}^{(\infty)}(z)$ satisfies the following jump conditions
\begin{equation}\label{Pinftyjump}
\mathbf{P}^{(\infty)}_{+}(z)=\mathbf{P}^{(\infty)}_{-}(z)J^{(\infty)}(z),
\end{equation}
where
$$
J^{(\infty)}(z)=\left\{\begin{aligned}
&\begin{pmatrix}0 & -e^{-2\pi i\alpha}s^{-1}_0 \\ e^{2\pi i\alpha} s_0& 0\end{pmatrix},& z&\in(z_{2,-},z_{1,-}), \\
&\begin{pmatrix}0 & -e^{-2\pi i\alpha} s_0^{-1}(1-|s_*|^2)\\ e^{2\pi i\alpha} s_0(1-|s_*|^2)^{-1} & 0\end{pmatrix}, & z&\in(z_{1,-},z_{1,+}), \\
&\begin{pmatrix}0 & -e^{-2\pi i\alpha} s^{-1}_0\\ e^{2\pi i\alpha} s_0& 0 \end{pmatrix}, & z&\in(z_{1,+},z_{2,+}),\\
&\begin{pmatrix}e^{-2\pi i\alpha} & 0 \\ 0 & e^{2\pi i\alpha} \end{pmatrix}, & z&\in(z_{2,+},+\infty).
\end{aligned}\right.
$$
\item{(3)} As $z\rightarrow \infty$, we have
\begin{equation}\label{Pinftyinfty}
\mathbf{P}^{(\infty)}(z)=\left(\mathbf{I}+O(z^{-1})\right)z^{\alpha\sigma_3},
\end{equation}
where the branch cut of $z^{\alpha}$ is taken along $\mathbb{R}_+$ so that $\arg z\in(0,2\pi)$.
\end{description}
The solution of the above RH problem can be constructed as follows
\begin{equation}\label{Pinfty}
\mathbf{P}^{(\infty)}(z)=s_0^{-\frac{\sigma_3}{2}}f^{-\sigma_3}_{\infty}\mathbf{X}(z)
f(z)^{\sigma_3}
s_0^{\frac{\sigma_3}{2}},
\end{equation}
where $f(z)$ is the Szeg\H{o} function
\begin{equation}\label{f(z)}
f(z)= \left(\frac{3}{8}\right)^{\frac{\alpha}{2}}
\left(z+\sqrt{z^2-\frac{8}{3}}\right)^{\alpha}
\left(\frac{\left((2-\sqrt{3})z-i\sqrt{z^2-\frac{8}{3}}\right)^2-\frac{8}{3}}
{\left((2-\sqrt{3})z+i\sqrt{z^2-\frac{8}{3}}\right)^2-\frac{8}{3}}\right)^{{\beta}},
\end{equation}
$\mathbf{X}(z)$ is given by
\begin{equation}\label{X(z)}
\mathbf{X}(z)=\begin{pmatrix}1 & 1 \\ -i & i\end{pmatrix} \omega(z)^{\sigma_{3}}
\begin{pmatrix} 1 & 1 \\ -i & i\end{pmatrix}^{-1},\quad \omega(z)=\left(\frac{z-\sqrt{\frac{8}{3}}}
{z+\sqrt{\frac{8}{3}}}\right)^{\frac{1}{4}},
\end{equation}
and $f_{\infty}$ is defined by
\begin{equation}\label{finfty}
f_{\infty}=2^{-\frac{\alpha}{2}}3^{\frac{\alpha}{2}}e^{\frac{\pi i{\beta}}{3}}.
\end{equation}
The branches of the multi-valued functions in \eqref{Pinfty}-\eqref{X(z)} are chosen by specifying
$$
\arg z\in(-\pi,\pi),\quad
\arg\left(z\pm\textstyle\sqrt{\frac{8}{3}}\right)\in(-\pi,\pi),\qquad
\arg \left(\textstyle z+\sqrt{z^2-\frac{8}{3}}\right)\in(0,2\pi),
$$
and
$$
\arg\left(\frac{\left((2-\sqrt{3})z-i\sqrt{z^2-\frac{8}{3}}\right)^2-\frac{8}{3}}
{\left((2-\sqrt{3})z+i\sqrt{z^2-\frac{8}{3}}\right)^2-\frac{8}{3}}\right)
\in(-\pi,\pi).
$$
The constant ${\beta}$ in \eqref{f(z)} and \eqref{finfty} is given by
\begin{equation}\label{nu}
{\beta}=-\frac{1}{2\pi i}\ln(1-|s_*|^2),
\end{equation}
where the logarithm takes the principal value. In the case $0<|s_*|<1$, it is easily seen that
\begin{equation}\label{nuimag}
{\beta}\neq0\quad \mathrm{and}\quad{\beta}\in i\mathbb{R}.
\end{equation}
From \eqref{f(z)} and \eqref{X(z)}, as $z\rightarrow\infty$, straightforward computation yields
\begin{align}\label{X(z)atinfty}
\mathbf{X}(z)&=\mathbf{I}+\sqrt{\frac{2}{3}}\sigma_2z^{-1}
+\frac{\mathbf{I}}{3}z^{-2}+O(z^{-3}),\\
f(z)&=2^{-\frac{\alpha}{2}}3^{\frac{\alpha}{2}}
e^{\frac{\pi i{\beta}}{3}}\left[1-\frac{2}{3}\left(\alpha+\sqrt{3}i{\beta}\right)z^{-2}
+O(z^{-4})\right]z^{\alpha}.\label{fatinfty}
\end{align}
\subsection{Local parametrices near the saddle points $z_{2,\pm}$}\label{sec:AiParametrix}
We will construct two parametrices $\mathbf{P}^{(2,\pm)}(z)$ satisfying the same jumps as $\mathbf{T}(z)$ on $\Sigma_\mathbf{T}$ (see Figure \ref{Deforma3}), respectively in the neighborhoods $U(z_{2,\pm},\delta)=\{z\in\mathbb{C}\mid|z-z_{2,\pm}|<\delta\}$ of the saddle points $z_{2,\pm}$, and matching with $\mathbf{P}^{(\infty)}(z)$ on the boundaries $\partial U(z_{2,\pm},\delta)=\{z\in \mathbb{C}\mid|z-z_{2,\pm}|=\delta\}$.
\subsection*{RH problem for $\mathbf{P}^{(2,+)}(z)$}
\begin{description}
\item{(1)} $\mathbf{P}^{(2,+)}(z)$ is analytic for $z\in U(z_{2,+},\delta)\setminus\Sigma_\mathbf{T}$.
\item{(2)} $\mathbf{P}^{(2,+)}(z)$ satisfies the same jumps as $\mathbf{T}(z)$ on $U(z_{2,+},\delta)\cap\Sigma_\mathbf{T}$.
\item{(3)} On the circular boundary $\partial U(z_{2,+},\delta)$, it holds that
\begin{equation}\label{P2+matching}
\mathbf{P}^{(2,+)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)\mathbf{P}^{(\infty)}(z),\quad \mathrm{as}\quad x\rightarrow -\infty.
\end{equation}
\end{description}
To solve the above RH problem, we define a conformal mapping
\begin{equation}\label{varphi1}
\varphi_1(z)=\left(\frac{3}{2}g(z)\right)^{\frac{2}{3}},
\end{equation}
where the branch is specified by the asymptotic condition
\begin{equation}\label{varphi1beha}
\varphi_1(z)=2^{\frac{5}{6}}3^{-\frac{1}{6}}
\textstyle\left(z-\sqrt{\frac{8}{3}}\right)
\left(1+o(1)\right),\quad \mathrm{as} \quad z\rightarrow \sqrt{\frac{8}{3}}.
\end{equation}
Then, the solution to the above RH problem can be explicitly constructed in terms of the Airy function:
\begin{equation}\label{P2+}
\mathbf{P}^{(2,+)}(z)=\mathbf{E}^{(2,+)}(z)\Phi^{\mathrm{(Ai)}}
\left(|x|^{\frac{4}{3}}\varphi_1(z)\right)
\left(s_0e^{2\pi i\alpha}\right)^{-\frac{\sigma_3}{2}}\sigma_{1}e^{\mp\pi i\alpha\sigma_3}e^{-x^2g(z)\sigma_3},\quad \pm\mathop{\rm Im}\nolimits z>0,
\end{equation}
where $\Phi^{\mathrm{(Ai)}}$ is the Airy model parametrix (see Appendix \ref{AP}), and $\mathbf{E}^{(2,+)}(z)$ is given by
\begin{equation}\label{E2+}
\mathbf{E}^{(2,+)}(z)=\mathbf{P}^{(\infty)}(z)e^{\pm\pi i\alpha\sigma_3}\sigma_{1}\left(s_0e^{2\pi i\alpha}\right)^{\frac{\sigma_3}{2}}
\frac{1}{\sqrt{2}} \begin{pmatrix}1 & -i\\ -i &1\end{pmatrix}|x|^{\frac{\sigma_3}{3}}\varphi_1(z)^{\frac{\sigma_3}{4}},\ \pm\mathop{\rm Im}\nolimits z>0.
\end{equation}
Here, the branch of $\varphi_1(z)^{\frac{1}{4}}$ is chosen such that $\arg\varphi_1(z)\in(-\pi,\pi)$. This means that
\begin{equation}\label{etajump}
\left(\varphi_1(z)^{\frac{\sigma_3}{4}}\right)_+=
\left(\varphi_1(z)^{\frac{\sigma_3}{4}}\right)_-e^{\frac{\pi i}{2}\sigma_3},\quad z\in (z_{2,-},z_{2,+}).
\end{equation}
From \eqref{Pinftyjump} and \eqref{etajump}, it is readily verified that $\mathbf{E}^{(2,+)}(z)$ is analytic in the neighborhood $U(z_{2,+},\delta)$.
Finally, combining \eqref{P2+}, \eqref{E2+} with the asymptotic behavior \eqref{AiryAsyatinfty} gives us the matching condition \eqref{P2+matching}.
\subsection*{RH problem for $\mathbf{P}^{(2,-)}(z)$}
\begin{description}
\item{(1)} $\mathbf{P}^{(2,-)}(z)$ is analytic for $z\in U(z_{2,-},\delta)\setminus\Sigma_\mathbf{T}$.
\item{(2)} On $U(z_{2,-},\delta)\cap\Sigma_\mathbf{T}$, $\mathbf{P}^{(2,-)}(z)$ shares the same jumps as $\mathbf{T}(z)$.
\item{(3)} On the boundary $\partial U(z_{2,-},\delta)$, we have
\begin{equation}\label{P2-matching}
\mathbf{P}^{(2,-)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)\mathbf{P}^{(\infty)}(z),\quad \mathrm{as}\quad x\rightarrow -\infty.
\end{equation}
\end{description}
By symmetry, the solution to the above RH problem is given by
\begin{equation}\label{P2-}
\mathbf{P}^{(2,-)}(z)=\mathbf{E}^{(2,-)}(z)\Phi^{(\mathrm{Ai})}
\left(|x|^{\frac{4}{3}}\varphi_1(-z)\right)
\left(s_0e^{2\pi i\alpha}\right)^{-\frac{\sigma_3}{2}}\sigma_2e^{-x^2g(z)\sigma_3},
\end{equation}
where $\varphi_1(z)$ is the conformal mapping defined by \eqref{varphi1} and $\mathbf{E}^{(2,-)}(z)$ is given by
\begin{equation}\label{E2-}
\mathbf{E}^{(2,-)}(z)=\mathbf{P}^{(\infty)}(z)\sigma_2\left(s_0e^{2\pi i\alpha}\right)^{\frac{\sigma_3}{2}}
\frac{1}{\sqrt{2}} \begin{pmatrix}1 & -i\\ -i &1\end{pmatrix}|x|^{\frac{\sigma_3}{3}}\varphi_1(-z)^{\frac{\sigma_3}{4}}.
\end{equation}
\subsection{Local parametrices near the saddle points $z_{1,\pm}$}
In this subsection, we seek two parametrices $\mathbf{P}^{(1,\pm)}(z)$ satisfying the same jumps as $\mathbf{T}(z)$ on $\Sigma_\mathbf{T}$ (see Figure \ref{Deforma3}) in the neighborhoods $U(z_{1,\pm},\delta)=\{z\in\mathbb{C}\mid|z-z_{1,\pm}|<\delta\}$ of the saddle points $z_{1,\pm}$, and matching with $\mathbf{P}^{(\infty)}(z)$ on the boundaries $\partial U(z_{1,\pm},\delta)=\{z\in \mathbb{C}\mid|z-z_{1,\pm}|=\delta\}$.
\subsection*{RH problem for $\mathbf{P}^{(1,+)}(z)$}
\begin{description}
\item{(1)} $\mathbf{P}^{(1,+)}(z)$ is analytic for $z\in U(z_{1,+},\delta)\setminus\Sigma_\mathbf{T}$.
\item{(2)} $\mathbf{P}^{(1,+)}(z)$ shares the same jumps as $\mathbf{T}(z)$ on $U(z_{1,+},\delta)\cap\Sigma_\mathbf{T}$.
\item{(3)} On the circle $\partial U(z_{1,+},\delta)$, $\mathbf{P}^{(1,+)}(z)$ satisfies
\begin{equation}\label{P1+matching}
\mathbf{P}^{(1,+)}(z)=\left(\mathbf{I}+O(|x|^{-1})\right)\mathbf{P}^{(\infty)}(z),\quad \mathrm{as}\quad x\rightarrow -\infty.
\end{equation}
\end{description}
Let us define the conformal mapping
\begin{equation}\label{varphi2}
\varphi_2(z)=\left\{\begin{aligned}
&2\sqrt{-\textstyle\frac{\sqrt{3}i}{6}-g(z)},\quad& \mathop{\rm Im}\nolimits z>0,\\
&2\sqrt{-\textstyle\frac{\sqrt{3}i}{6}+g(z)},\quad& \mathop{\rm Im}\nolimits z<0,
\end{aligned}\right.
\end{equation}
where the branches of the square roots are specified by choosing
\begin{equation}\label{varphi2beha}
\varphi_2(z)=e^{-\frac{\pi i}{4}}2\cdot3^{-\frac{1}{4}}\textstyle \left(z-\sqrt{\frac{2}{3}}\right)
\left(1+o(1)\right),\quad \mathrm{as}\quad z\rightarrow\sqrt{\frac{2}{3}}.
\end{equation}
Let $\Phi^{\mathrm{(PC)}}$ be the parabolic cylinder model parametrix with parameter ${\beta}$ given by \eqref{nu}; see Appendix \ref{PCP}. The solution to the above RH problem can be constructed as follows:
\begin{align}\label{P1+}
\mathbf{P}^{(1,+)}(z)=\mathbf{E}^{(1,+)}(z)\Phi^{\mathrm{(PC)}}\left(|x|\varphi_2(z)\right)
\left(\frac{s_*}{h_0}\right)^{\frac{\sigma_3}{2}}
\left\{\begin{aligned}
&s_0^{-\frac{\sigma_3}{2}}(-\sigma_{1})e^{-x^2g(z)\sigma_3},& \mathop{\rm Im}\nolimits z>0,\\
&s_0^{\frac{\sigma_3}{2}}\sigma_3e^{2\pi i(\alpha+{\beta})\sigma_3}
e^{-x^2g(z)\sigma_3},& \mathop{\rm Im}\nolimits z<0,
\end{aligned}\right.
\end{align}
where $h_{0}$ is the Stokes multiplier in \eqref{h0} and $\mathbf{E}^{(1,+)}(z)$ is given by
\begin{equation}\label{E1+}
\mathbf{E}^{(1,+)}(z)=\mathbf{W}^{(+)}(z)\left(\frac{s_*}{h_0}\right)^{-\frac{\sigma_3}{2}}
|x|^{{\beta}\sigma_3}e^{i\frac{\sqrt{3}}{6}x^2
\sigma_{3}}2^{-\frac{\sigma_3}{2}}\begin{pmatrix}|x|\varphi_2(z) & 1 \\ 1 & 0\end{pmatrix}
\end{equation}
with
\begin{equation}\label{W+}
\mathbf{W}^{(+)}(z)=\mathbf{P}^{(\infty)}(z)\left\{\begin{aligned}
&(-\sigma_1)s_0^{\frac{1}{2}\sigma_3}\varphi_2(z)^{{\beta}\sigma_3},
&\mathop{\rm Im}\nolimits z&>0,\\
&e^{-2\pi i(\alpha+{\beta})\sigma_3}\sigma_3s_0^{-\frac{1}{2}\sigma_3}
\varphi_2(z)^{{\beta}\sigma_3}, & \mathop{\rm Im}\nolimits z&<0.
\end{aligned}\right.
\end{equation}
The branch of $\varphi_2(z)^{{\beta}}$ is chosen by specifying $\arg\varphi_2(z)\in(0,2\pi)$, this implies the jump relations
\begin{equation}\label{varphijump}
\left\{\begin{aligned}
&\left(\varphi_2(z)^{{\beta}}\right)_+=
\left(\varphi_2(z)^{{\beta}}\right)_-e^{2\pi i{\beta}}, \quad &z&\in(z_{1,+},z_{2,+}),\\
&\left(\varphi_2(z)^{{\beta}}\right)_+
=\left(\varphi_2(z)^{{\beta}}\right)_-, \quad &z&\in(z_{2,-},z_{1,+}).
\end{aligned}\right.
\end{equation}
From the jump conditions \eqref{Pinftyjump} and \eqref{varphijump}, it is readily seen that $\mathbf{W}^{(+)}(z)$ is analytic in the neighborhood $U(z_{1,+},\delta)$. In particular, it follows from \eqref{Pinfty}, \eqref{f(z)}, \eqref{X(z)}, \eqref{varphi2beha} and \eqref{W+} that
\begin{equation}\label{W+atz+}
\mathbf{W}^{(+)}(\textstyle\sqrt{\frac{2}{3}})=-s_0^{-\frac{\sigma_3}{2}}
f_{\infty}^{-\sigma_3}
\begin{pmatrix}1 & 1 \\ -i & i\end{pmatrix} 3^{-\frac{\sigma_{3}}{4}}
e^{\frac{\pi i}{4}\sigma_3}\begin{pmatrix} 1 & 1 \\
-i & i\end{pmatrix}^{-1}2^{-\frac{{\beta}}{2}\sigma_3}
3^{-\frac{{\beta}}{4}\sigma_3}e^{\frac{\pi i{\beta}}{4}\sigma_3}
e^{\frac{\pi i\alpha}{3}\sigma_3}\sigma_1 .
\end{equation}
Finally, a combination of \eqref{Pinfty}, \eqref{nuimag}, \eqref{PCAsyatinfty} and \eqref{P1+} gives
\begin{align}\label{matchingcondition+}
&\mathbf{P}^{(1,+)}(z)\mathbf{P}^{(\infty)}(z)^{-1}\nonumber\\
&=\mathbf{W}^{(+)}(z)\begin{pmatrix}
1+\frac{{\beta}({\beta}+1)}{2|x|^2\varphi_2^2}+O\left(|x|^{-4}\right) & \frac{e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2{\beta}}}{h_0^{-1}s_*}
\left(\frac{{\beta}}{|x|\varphi_2}
+O\left(|x|^{-3}\right)\right) \\ \frac{h_0^{-1}s_*}{e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2{\beta}}}
\left(\frac{1}{|x|\varphi_2}+O\left(|x|^{-3}\right)\right) & 1-\frac{{\beta}({\beta}-1)}{2|x|^2\varphi_2^2}+O\left(|x|^{-4}\right)
\end{pmatrix}\mathbf{W}^{(+)}(z)^{-1}\nonumber\\
&=\mathbf{I}+\mathbf{G}^{(+)}(z)|x|^{-1}+O\left(|x|^{-2}\right),
\end{align}
where
\begin{equation}\label{G+}
\mathbf{G}^{(+)}(z)=\mathbf{W}^{(+)}(z)\begin{pmatrix}0 & \frac{\beta h_0e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2\beta}}
{s_*\varphi_2(z)} \\ \frac{s_*|x|^{-2\beta}}
{h_0e^{ i\frac{\sqrt{3}}{3}x^2}\varphi_2(z)} & 0 \end{pmatrix}\mathbf{W}^{(+)}(z)^{-1}.
\end{equation}
\subsection*{RH problem for $\mathbf{P}^{(1,-)}(z)$}
\begin{description}
\item{(1)} $\mathbf{P}^{(1,-)}(z)$ is analytic for all $z\in U(z_{1,-},\delta)\setminus\Sigma_\mathbf{T}$.
\item{(2)} $\mathbf{P}^{(1,-)}(z)$ satisfies the same jump conditions as $\mathbf{T}(z)$ on $U(z_{1,-},\delta)\cap\Sigma_\mathbf{T}$.
\item{(3)} $\mathbf{P}^{(1,-)}(z)$ fulfills the matching condition on the boundary $\partial U(z_{1,-},\delta)$:
\begin{equation}\label{P1-matching}
\mathbf{P}^{(1,-)}(z)=\left(\mathbf{I}+O(|x|^{-1})\right)\mathbf{P}^{(\infty)}(z),\quad \mathrm{as}\quad x\rightarrow -\infty.
\end{equation}
\end{description}
Similarly, we introduce the conformal mapping
\begin{equation}\label{varphi3}
\varphi_3(z)=\left\{\begin{aligned}
&2\sqrt{-\textstyle\frac{\sqrt{3}i}{6}+g(z)},\quad& \mathop{\rm Im}\nolimits z>0,\\
&2\sqrt{-\textstyle\frac{\sqrt{3}i}{6}-g(z)},\quad& \mathop{\rm Im}\nolimits z<0,
\end{aligned}\right.
\end{equation}
where the branches of square roots are taken so that
\begin{equation}\label{varphi3beha}
\varphi_3(z)=e^{\frac{3\pi i}{4}}2\cdot3^{-\frac{1}{4}}\textstyle \left(z+\sqrt{\frac{2}{3}}\right)
\left(1+o(1)\right),\quad \mathrm{as}\quad z\rightarrow-\sqrt{\frac{2}{3}}.
\end{equation}
The solution to the above RH problem can be also constructed in terms of the parabolic cylinder functions as follows:
\begin{align}\label{P1-}
&\mathbf{P}^{(1,-)}(z)=\mathbf{E}^{(1,-)}(z)\Phi^{\mathrm{(PC)}}\left(|x|\varphi_3(z)\right)
\left(\frac{s_*}{h_0}\right)^{\frac{\sigma_3}{2}}\left\{\begin{aligned}
&e^{2\pi i{\beta}\sigma_3}\left(s_0e^{2\pi i\alpha}\right)^{\frac{\sigma_3}{2}}e^{-x^2g(z)\sigma_3},& \mathop{\rm Im}\nolimits z>0,\\
&\left(s_0e^{2\pi i\alpha}\right)^{-\frac{\sigma_3}{2}}
\sigma_{3}\sigma_{1}e^{-x^2g(z)\sigma_3},
& \mathop{\rm Im}\nolimits z<0,\end{aligned}\right.
\end{align}
where $\Phi^{\mathrm{(PC)}}$ is the parabolic cylinder model parametrix, $h_{0}$ is given in \eqref{h0}, ${\beta}$ is defined by \eqref{nu} and $\mathbf{E}^{(1,-)}(z)$ is given by
\begin{equation}\label{E1-}
\mathbf{E}^{(1,-)}(z)=\mathbf{W}^{(-)}(z)\left(\frac{s_*}{h_0}\right)^{-\frac{\sigma_3}{2}}
|x|^{{\beta}\sigma_3}e^{\frac{i\sqrt{3}x^2}{6}
\sigma_{3}}2^{-\frac{\sigma_3}{2}}\begin{pmatrix}|x|\varphi_3(z) & 1 \\ 1 & 0\end{pmatrix}
\end{equation}
with
\begin{equation}\label{W-}
\mathbf{W}^{(-)}(z)=\mathbf{P}^{(\infty)}(z)\left\{\begin{aligned}
&\left(s_0e^{2\pi i\alpha}\right)^{-\frac{\sigma_3}{2}}
e^{-2\pi i{\beta}\sigma_3}\varphi_3(z)^{{\beta}\sigma_3},\ &\mathop{\rm Im}\nolimits z&>0,\\
&\sigma_1\sigma_3\left(s_0e^{2\pi i\alpha}\right)^{\frac{\sigma_3}{2}}\varphi_3(z)^{{\beta}\sigma_3}, \ &\mathop{\rm Im}\nolimits z&<0.
\end{aligned}\right.
\end{equation}
We point out that the branch of the function $\varphi_3(z)^{{\beta}}$ is chosen such that $\arg\varphi_3(z)\in(-\pi,\pi)$. This implies that
\begin{equation}\label{zetajump}
\left\{\begin{aligned}
&\left(\varphi_3(z)^{{\beta}}\right)_+=\left(\varphi_3(z)^{{\beta}}\right)_-e^{2\pi i{\beta}}, \quad &z&\in(z_{2,-},z_{1,-}),\\
&\left(\varphi_3(z)^{{\beta}}\right)_+=\left(\varphi_3(z)^{{\beta}}\right)_-, \quad &z&\in(z_{1,-},z_{2,+}).
\end{aligned}\right.
\end{equation}
Using the jump relations \eqref{Pinftyjump} and \eqref{zetajump}, it is straightforward to check that $\mathbf{W}^{(-)}(z)$ is analytic in the neighborhood $U(z_{1,-},\delta)$. Particularly, by substituting \eqref{Pinfty}, \eqref{f(z)}, \eqref{X(z)} and \eqref{varphi3beha} into \eqref{W-}, we have
\begin{equation}\label{W-atz-}
\mathbf{W}^{(-)}(-\textstyle\sqrt{\frac{2}{3}})=s_0^{-\frac{\sigma_3}{2}}
f_{\infty}^{-\sigma_3}\begin{pmatrix}1 & 1 \\ -i & i\end{pmatrix} 3^{\frac{1}{4}\sigma_{3}}e^{\frac{\pi i}{4}\sigma_3}\begin{pmatrix} 1 & 1 \\
-i & i\end{pmatrix}^{-1}2^{\frac{{\beta}}{2}\sigma_3}
3^{\frac{{\beta}}{4}\sigma_3}
e^{-\frac{\pi i\alpha}{3}\sigma_3}e^{-\frac{\pi i{\beta}}{4}\sigma_3}.
\end{equation}
Furthermore, using \eqref{Pinfty}, \eqref{nuimag}, \eqref{P1-} and the asymptotic behavior \eqref{PCAsyatinfty}, we obtain the matching condition
\begin{align}\label{matchingcondition-}
&\mathbf{P}^{(1,-)}(z)\mathbf{P}^{(\infty)}(z)^{-1}\nonumber \\
&=\mathbf{W}^{(-)}(z)\begin{pmatrix}
1+\frac{{\beta}({\beta}+1)}{2|x|^2\varphi_3^2}+O\left(|x|^{-4}\right) & \frac{e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2{\beta}}}{h_0^{-1}s_*}
\left(\frac{{\beta}}{|x|\varphi_3}
+O\left(|x|^{-3}\right)\right)\\ \frac{h_0^{-1}s_*}{e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2{\beta}}}
\left(\frac{1}{|x|\varphi_3}+O\left(|x|^{-3}\right)\right) & 1-\frac{{\beta}({\beta}-1)}{2|x|^2\varphi_3^2}+O\left(|x|^{-4}\right)
\end{pmatrix}\mathbf{W}^{(-)}(z)^{-1}\nonumber\\
&=\mathbf{I}+\mathbf{G}^{(-)}(z)|x|^{-1}+O\left(|x|^{-2}\right)
\end{align}
where
\begin{equation}\label{G-}
\mathbf{G}^{(-)}(z)=\mathbf{W}^{(-)}(z)\begin{pmatrix}0 & \frac{{\beta} h_0e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2{\beta}}}
{s_*\varphi_3(z)}\\ \frac{s_*|x|^{-2{\beta}}}{h_0e^{i\frac{\sqrt{3}}{3}x^2}\varphi_3(z)} & 0 \end{pmatrix}\mathbf{W}^{(-)}(z)^{-1}.
\end{equation}
\subsection{Local parametrix near the origin}
In this subsection, we intend to find a function $\mathbf{P}^{(0)}(z)$ having the same jumps as $\mathbf{T}(z)$ on $\Sigma_\mathbf{T}$ (see Figure \ref{Deforma3}) in the neighborhood $U(0,\delta)=\{z\in \mathbb{C}\mid|z|<\delta\}$ of the origin, and matching with $\mathbf{P}^{(\infty)}(z)$ on the boundary $\partial U(0,\delta)=\{z\in \mathbb{C}\mid|z|=\delta\}$.
\subsection*{RH problem for $\mathbf{P}^{(0)}(z)$}
\begin{description}
\item{(1)} $\mathbf{P}^{(0)}(z)$ is analytic for $z\in U(0,\delta)\setminus\Sigma_\mathbf{T}$.
\item{(2)} $\mathbf{P}^{(0)}(z)$ satisfies the same jump conditions as $\mathbf{T}(z)$ on $U(0,\delta)\cap\Sigma_\mathbf{T}$.
\item{(3)} On the boundary $\partial U(0,\delta)$, we have
\begin{equation}\label{P0matching}
\mathbf{P}^{(0)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)\mathbf{P}^{(\infty)}(z),\quad \mathrm{as}\quad x\rightarrow -\infty.
\end{equation}
\item{(4)} $\mathbf{P}^{(0)}(z)$ has the same asymptotic behavior as $\mathbf{T}(z)$ near the origin; see \eqref{T0}.
\end{description}
To begin with, we construct a conformal mapping
\begin{equation}\label{varphi4}
\varphi_4(z)=\pm\frac{1}{8}iz\left(z^2-\frac{8}{3}\right)^{\frac{3}{2}},\quad \pm\mathop{\rm Im}\nolimits z>0,
\end{equation}
which behaves like
\begin{equation}\label{varphi4beha}
\varphi_4(z)=2^{\frac{3}{2}}3^{-\frac{3}{2}}z(1+o(1)),\quad \mathrm{as} \quad z\rightarrow0.
\end{equation}
The solution of the above RH problem can be built explicitly in term of the modified Bessel functions. To see this, let $\Phi^{(\mathrm{Bes})}$ be the Bessel model parametrix with parameter $\alpha$ as given in Appendix \ref{Bessel}. The parametrix $\mathbf{P}^{(0)}(z)$ is then constructed as follows:
\begin{equation}\label{P0}
\mathbf{P}^{(0)}(z)=\mathbf{E}^{(0)}(z)\Phi^{(\mathrm{Bes})}
\left(x^2\varphi_4(z)\right)\mathbf{K}(z)
\left[s_1(e^{-2\pi i\alpha}+s_*)\right]^{-\frac{\sigma_3}{2}}e^{-x^2g(z)\sigma_3},
\end{equation}
where $\mathbf{K}(z)$ is a piecewise constant matrix defined in regions $\Lambda_k$ described in Figure \ref{Bes}
\begin{equation}\label{K(z)}
\mathbf{K}(z)=\left\{\begin{aligned}
&\mathbf{I}, \quad &z&\in \Lambda_1\cup\Lambda_8,\\
&\begin{pmatrix}1 & 0\\ -e^{-2\pi i\alpha} &1\end{pmatrix},\quad &z&\in \Lambda_2,\\
&\begin{pmatrix}e^{-\pi i\alpha} & 0 \\ -e^{\pi i\alpha} & e^{\pi i\alpha}\end{pmatrix},\quad &z&\in \Lambda_3,\\
&e^{-\pi i\alpha\sigma_3},\quad &z&\in \Lambda_4,\\
&e^{\pi i\alpha\sigma_3},\quad &z&\in \Lambda_5,\\
&\begin{pmatrix}e^{\pi i\alpha} & 0 \\ e^{-\pi i\alpha} & e^{-\pi i\alpha}\end{pmatrix},\quad &z&\in\Lambda_6,\\
&\begin{pmatrix}1 & 0\\ e^{2\pi i\alpha} & 1\end{pmatrix},\quad &z&\in \Lambda_7,
\end{aligned}\right.
\end{equation}
and $\mathbf{E}^{(0)}(z)$ is given by
\begin{equation}\label{E(0)}
\mathbf{E}^{(0)}(z)=\mathbf{P}^{(\infty)}(z)\left[s_1(e^{-2\pi i\alpha}+s_*)\right]^{\frac{\sigma_3}{2}}\mathbf{Q}(z)e^{-\frac{1}{4}\pi i\sigma_3}\frac{1}{\sqrt{2}}\begin{pmatrix}1 & i\\ i & 1\end{pmatrix}
\end{equation}
with
\begin{equation}\label{Q}
\mathbf{Q}(z)=\left\{\begin{aligned}
&e^{\frac{1}{2}\pi i\alpha\sigma_3}, \quad &\mathop{\rm Im}\nolimits z>0,~~ |z|<\delta,\\
&e^{-\frac{1}{2}\pi i\alpha\sigma_3}\sigma_3\sigma_1 ,\quad &\mathop{\rm Im}\nolimits z<0,~~ |z|<\delta.
\end{aligned}\right.
\end{equation}
First, using the jump conditions \eqref{Besseljump}, it can be easily checked that the function $\mathbf{P}^{(0)}(z)$ constructed by \eqref{P0}-\eqref{Q} satisfies the same jump relations as $\mathbf{T}(z)$ on $\Sigma_\mathbf{T}\cap U(0,\delta) $. Applying \eqref{Pinftyjump}, \eqref{E(0)} and \eqref{Q}, one can check that $\mathbf{E}^{(0)}(z)$ is also analytic in the neighborhood $U(0,\delta)$. Moreover, the matching condition \eqref{P0matching} follows from \eqref{P0}, \eqref{E(0)} and \eqref{BesInfty}. Next, we verify the behavior of $\mathbf{P}^{(0)}(z)$ as $z\to 0$.
Recalling the definition of the connection matrix $E_0$ as given in \eqref{E0}, we may rewrite the asymptotic behavior \eqref{T0} in the form
\begin{equation}\label{T1}
\mathbf{T}(z)=\mathbf{T}_1(z)z^{\alpha\sigma_{3}} \begin{pmatrix}1 & \frac{1}{1+e^{-2\pi i\alpha}}\\ 0&1\end{pmatrix} \begin{pmatrix}1 & 0\\ -e^{-2\pi i\alpha}&1\end{pmatrix}
\left[s_1(e^{-2\pi i\alpha}+s_*)\right]^{-\frac{\sigma_3}{2}}e^{-x^2g(z)\sigma_3},
\end{equation}
where $\mathbf{T}_1(z)$ is analytic in a neighborhood of $z=0$.
Comparing \eqref{T1} with \eqref{P0} and \eqref{BesParaExpand}, one can see that $\mathbf{P}^{(0)}(z)$ satisfies the asymptotic behavior \eqref{T0} as $z\to0$.
\subsection{Final transformation}
The final transformation is defined as
\begin{equation}\label{R}
\mathbf{R}(z)=\left\{\begin{aligned}
&\mathbf{T}(z)\left[\mathbf{P}^{(1,\pm)}(z)\right]^{-1},\quad &z&\in U(z_{1,\pm},\delta)\setminus\Sigma_\mathbf{T},\\
&\mathbf{T}(z)\left[\mathbf{P}^{(2,\pm)}(z)\right]^{-1},\quad &z&\in U(z_{2,\pm},\delta)\setminus\Sigma_\mathbf{T},\\
&\mathbf{T}(z)\left[\mathbf{P}^{(0)}(z)\right]^{-1},\quad &z&\in U(0,\delta)\setminus\Sigma_\mathbf{T},\\
&\mathbf{T}(z)\left[\mathbf{P}^{(\infty)}(z)\right]^{-1},\quad &\mathrm{e}&\mathrm{lsewhere}.\\ \end{aligned}
\right.
\end{equation}
Immediately, $\mathbf{R}(z)$ fulfills the following RH problem.
\subsection*{RH problem for $\mathbf{R}(z)$}
\begin{figure}[t]
\centering
\includegraphics[width=13cm,height=8cm]{sigmaR}\\
\caption{The jump contour $\Sigma_\mathbf{R}$}\label{Rjump}
\end{figure}
\begin{description}
\item{(1)} $\mathbf{R}(z)$ is analytic for $z\in \mathbb{C}\setminus\Sigma_\mathbf{R}$, where the contour $\Sigma_\mathbf{R}$ is illustrated in Figure \ref{Rjump}.
\item{(2)} On the contour $\Sigma_\mathbf{R}$, we have $\mathbf{R}_+(z)=\mathbf{R}_-(z)J_{\mathbf{R}}(z)$, where
\begin{equation}\label{JumpR}
J_{\mathbf{R}}(z)=\left\{\begin{aligned}
&\mathbf{P}^{(1,\pm)}(z)\mathbf{P}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(z_{1,\pm},\delta),\\
&\mathbf{P}^{(2,\pm)}(z)\mathbf{P}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(z_{2,\pm},\delta),\\
&\mathbf{P}^{(0)}(z)\mathbf{P}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(0,\delta),\\
&\mathbf{P}^{(\infty)}_-(z)J_\mathbf{T}(z)\mathbf{P}^{(\infty)}_+(z)^{-1},\quad &\mathrm{e}&\mathrm{lsewhere}.\\ \end{aligned}
\right.
\end{equation}
\item{(3)} As $z\rightarrow\infty$, we have
\begin{equation}\label{Rexpan}
\mathbf{R}(z)=\mathbf{I}+\frac{\mathbf{R}_1}{z}+\frac{\mathbf{R}_2}{z^2}+O(z^{-3}).
\end{equation}
\end{description}
In view of the matching conditions \eqref{P2+matching}, \eqref{P2-matching}, \eqref{P1+matching}, \eqref{P1-matching} and \eqref{P0matching}, it is now readily seen that as $x\rightarrow -\infty$,
\begin{equation}\label{JRestimation}
J_\mathbf{R}(z)=\left\{\begin{aligned}
&\mathbf{I}+O(|x|^{-1}), &z&\in\partial U(z_{1,\pm},\delta),\\
&\mathbf{I}+O(|x|^{-2}), &z&\in\partial U(0,\delta)\cup\partial U(z_{2,\pm},\delta),\\
&\mathbf{I}+O(e^{-c_1|x|^2}),&\mathrm{e}&\mathrm{lsewhere},
\end{aligned}\right.
\end{equation}
where $c_1$ is some positive constant. We thus have
\begin{equation}\label{Restimation}
\mathbf{R}(z)=\mathbf{I}+O(|x|^{-1})\quad \mathrm{as} \quad x\rightarrow -\infty,
\end{equation}
where the error bound is uniform for $z\in \mathbb{C}\setminus\Sigma_{\mathbf{R}}$.
\subsection{Case $|s_*|=0$}\label{sec:szero}
In this section, we consider the special case $|s_*|=0$. It follows from the definitions \eqref{sstar}, \eqref{sstarbar} of $s_*$, $\bar s_*$ that
\begin{equation*}
\left\{\begin{aligned}
&1+s_0s_1=0,\\
&1+s_0s_1e^{2\pi i\alpha}=0.
\end{aligned}\right.
\end{equation*}
Substituting the first equation into the second one gives $e^{2\pi i\alpha}=1$. Therefore, $\alpha$ is an integer. Recalling the jump matrices on the contour $\Sigma_\mathbf{T}$ (see Figure \ref{Deforma3}), it is readily seen that in this case, there are no jumps on the infinite branches $\pi_k$, $k=2,\cdots,6,10,\cdots,14$. As a result, the RH problem for $\mathbf{T}(z)$ given in Section \ref{sec:RHT} is reduced to the following RH problem for $\widehat{\mathbf{T}}(z)$.
\subsection*{RH problem for $\widehat{\mathbf{T}}(z)$}
\begin{description}
\item{(1)} $\widehat{\mathbf{T}}(z)$ is analytic for $z\in \mathbb{C}\setminus (\Sigma_{\mathbf{T}}\setminus\bigcup\pi_k)$, where $k=2,\cdots,6,10,\cdots,14$.
\item{(2)} $\widehat{\mathbf{T}}(z)$ satisfies the jump relations
\begin{equation*}
\widehat{\mathbf{T}}_+(z)=\widehat{\mathbf{T}}_-(z)\left\{\begin{aligned}
&\begin{pmatrix}1 & -s_0^{-1}e^{2x^2g(z)}\\ 0 & 1 \end{pmatrix},\quad &z&\in \pi_1,\pi_{15}\\
&\begin{pmatrix}1 & s_0^{-1}e^{2x^2g(z)} \\ 0 & 1 \end{pmatrix},\quad &z&\in \pi_7,\pi_9,\\
&\begin{pmatrix}1 & 0\\ -s_0e^{-2x^2g(z)} & 1 \end{pmatrix},\quad &z&\in \pi_8,\pi^-_{16},\\
&\begin{pmatrix}0 & -s_0^{-1}\\ s_0 & 0 \end{pmatrix},\quad &z&\in (z_{2,-},z_{2,+}),
\end{aligned}\right.
\end{equation*}
\item{(3)} $\widehat{\mathbf{T}}(z)=\left(\mathbf{I}+O(z^{-1})\right)z^{\alpha\sigma_3}$ as $z\rightarrow \infty$.
\item{(4)} $\widehat{\mathbf{T}}(z)$ has the following behavior at the origin
\begin{equation}\label{That0}
\widehat{\mathbf{T}}(z)=\widehat{\mathbf{T}}_0(z)z^{\alpha\sigma_3}
E_0S_1e^{-x^2g(z)\sigma_3}, \qquad \arg z\in(\pi/4,\pi/2),
\end{equation}
where $\widehat{\mathbf{T}}_0(z)$ is analytic in a neighborhood of the origin. The asymptotic behaviors of $\widehat{\mathbf{T}}(z)$ in other regions are determined by \eqref{That0} and the jump relations satisfied by $\widehat{\mathbf{T}}(z)$.
\end{description}
The following RH analysis is similar to what we have done in the case $0<|s_*|<1$.
Firstly, neglecting the exponentially small entries in the jump matrices for $\widehat{\mathbf{T}}(z)$, we arrive at an approximate RH problem with constant jump on $(z_{2,-},z_{2,+})$. The solution to this RH problem, namely the global paramatrix, can be constructed as \eqref{Pinfty} by taking the parameter $\beta=0$ therein. Next, we build two local parametrices near $z_{2,\pm}=\pm\sqrt{8/3}$, satisfying exactly the same RH problems for
$\mathbf{P}^{(2,\pm)}(z)$, of which the solutions have been constructed in Section \ref{sec:AiParametrix}. Finally, by defining $\mathbf{R}(z)$ as
the ratio of the solution of the RH problem for $\widehat{\mathbf{T}}(z)$ and the parametrices
similarly to \eqref{R}, we show that $\mathbf{R}(z)$ satisfies the asymptotic $\mathbf{R}(z)=\mathbf{I}+O(|x|^{-2})$ as $x\to-\infty$ and thus the parametrices are good approximations of $\widehat{\mathbf{T}}(z)$ as $x\to-\infty$.
\section{RH analysis as $x\to-\infty$ with $|s_*|>1$}\label{Asymptotic-infty2}
When $|s_*|>1$ and as $x\to-\infty$, a detailed RH analysis has been carried out by the current authors in \cite{XXZ} for $\alpha\in\mathbb{R}$ but $\alpha-\frac{1}{2}\notin\mathbb{Z}$. In this section, we will briefly review the RH analysis therein, and discuss the exceptional case $\frac{1}{2}-\alpha\in\mathbb{N}$.
In the case $|s_*|>1$, we have $1-|s_*|^2<0$. From the definition \eqref{nu} of ${\beta}$, it follows that
\begin{equation}\label{nu0}
{\beta}=-\frac{1}{2\pi i}\ln(1-|s_*|^2)=-\frac{1}{2\pi i}\ln(|s_*|^2-1)-\frac{1}{2}:={\beta}_0-\frac{1}{2},\quad {\beta}_0\in i\mathbb{R}.
\end{equation}
This, together with \eqref{G+} and \eqref{G-}, implies that
$\mathbf{P}^{(1,\pm)}(z)\mathbf{P}^{\infty}(z)^{-1}\nrightarrow \mathbf{I}$ as $x\rightarrow-\infty$. Therefore, the matching conditions \eqref{P1+matching} and \eqref{P1-matching} are no longer fulfilled.
Similar problems have also occurred in deriving singular asymptotics for the PII transcendents \cite{BI,Hu}. In \cite{BI}, Bothner and Its develop a certain ``dressing'' technique to tackle this problem. However, we find their method does not apply to our case. Instead, our idea is to modify the global parametrix $\mathbf{P}^{(\infty)}(z)$, so as to satisfy the matching condition. Similar technique was first used in \cite{ZZ} to derive the uniform asymptotics of the Pollaczek polynomials and subsequently to derive the uniform asymptotics of a system of Szeg\H{o} class polynomials \cite{ZXZ}.
\subsection{Modified global parametrix }
We define
\begin{equation}\label{mPinfty}
\widetilde{\mathbf{P}}^{(\infty)}(z)=\mathbf{H}(z)\mathbf{P}^{(\infty)}(z),
\end{equation}
where $\mathbf{P}^{(\infty)}(z)$ is introduced in \eqref{Pinfty}. The meromorphic function $\mathbf{H}(z)$ with $\det{\mathbf{H}}(z)=1$ is given by
\begin{equation}\label{H}
\mathbf{H}(z)=\mathbf{I}+\frac{\widetilde{A}}{z-\sqrt{\frac{2}{3}}}
+\frac{\widetilde{B}}{z+\sqrt{\frac{2}{3}}}.
\end{equation}
The explicit expressions of $\widetilde{A}$ and $\widetilde{B}$ are
\begin{equation}\label{reptildeAB}
\widetilde{A}=\sqrt{\frac 2 3} \begin{pmatrix} \frac{l}{l+\sqrt 2} &
e^{\frac{5\pi i} 6} \frac{ l\, s_0^{-1}f^{-2}_{\infty}}
{ \sqrt 2+l } \\[.3cm]
e^{ \frac{\pi i} 6} \frac{ l\, s_0 f^{ 2}_{\infty}}
{ \sqrt 2-l } & \frac{l}{ l-\sqrt{2} }
\end{pmatrix},\quad
\widetilde{B}=\sqrt{\frac 2 3} \begin{pmatrix} -\frac{l}{l+\sqrt 2} &
e^{\frac{5\pi i} 6} \frac{ l\, s_0^{-1}f^{-2}_{\infty}}
{ \sqrt 2+l } \\[.3cm]
e^{ \frac{\pi i} 6} \frac{ l\, s_0 f^{ 2}_{\infty}}
{ \sqrt 2-l } & -\frac{l}{ l-\sqrt{2} }
\end{pmatrix},
\end{equation}
where $f_{\infty}$ is given by \eqref{finfty} and $l=l(x)$ is defined as
\begin{equation}\label{c}
l=-\frac{i\sqrt{6}e^{i\phi}}{2+e^{i\phi}},\quad \phi=-\frac{\sqrt{3}}{3}x^{2}+i{\beta}_0\ln \left(2 \sqrt{3} x^{2}\right)
+\frac{2\pi\alpha}{3}+\arg\Gamma\left({\beta}_0+ \frac{1}{2}\right)+\arg s_{*},
\end{equation}
with ${\beta}_0$ and $s_*$ given in \eqref{nu0} and \eqref{sstar}, respectively.
We shall assume in \eqref{reptildeAB} that $x$ does not take the zeros of the functions $\sqrt{2}\pm l(x)$. The set of zeros consists of two sequences $\{x_n\}$ and $\{y_n\}$ for $n\in \mathbb{N}$, defined respectively by the equations
\begin{equation}\label{pole1}
\frac{\sqrt{3}}{3}x_n^{2}-i{\beta}_0\ln \left(2 \sqrt{3} x_n^{2}\right)
-\frac{2\pi\alpha}{3}-\arg \Gamma\left({\beta}_0-\frac{1}{2}\right)-\arg s_{*}-\frac{2\pi}{3}-2n\pi=0,
\end{equation}
and
\begin{equation}\label{pole2}
\frac{\sqrt{3}}{3}y_n^{2}-i{\beta}_0\ln \left(2 \sqrt{3} y_n^{2}\right)
-\frac{2\pi\alpha}{3}-\arg \Gamma\left({\beta}_0+ \frac{1}{2}\right)-\arg s_{*}+\frac{2\pi}{3}-2n\pi=0.
\end{equation}
For detailed derivations of $\widetilde{A}$ and $\widetilde{B}$, we refer to \cite[Section 3.3]{XXZ}.
\subsection{Modified local parametrices at $z_{1,\pm}$, $z_{2,\pm}$ }
We first consider the local parametrix at $z_{1,+}=\sqrt{\frac{2}{3}}$. Instead of \eqref{P1+}, we define
\begin{align*}
\widetilde{\mathbf{P}}^{(1,+)}(z)=\widetilde {\mathbf{E}}^{(1,+)}(z)\Phi^{\mathrm{(PC)}}\left(|x|\varphi_2(z)\right)
\left(\frac{s_*}{h_0}\right)^{\frac{\sigma_3}{2}}
\left\{\begin{aligned}&s_0^{-\frac{\sigma_3}{2}}(-\sigma_{1})
e^{-x^2g(z)\sigma_3},& \mathop{\rm Im}\nolimits z>0,\\
&s_0^{\frac{\sigma_3}{2}}\sigma_3e^{2\pi i(\alpha+{\beta})\sigma_3}
e^{-x^2g(z)\sigma_3},& \mathop{\rm Im}\nolimits z<0,
\end{aligned}\right.
\end{align*}
where $\Phi^{\mathrm{(PC)}}$ is the parabolic cylinder model parametrix, $\varphi_2(z)$ is the conformal mapping \eqref{varphi2}, ${\beta}$ is given by \eqref{nu} and $\widetilde{\mathbf{E}}^{(1,+)}(z)$ is defined as
\begin{equation*}
\widetilde{\mathbf{E}}^{(1,+)}(z)=\mathbf{H}(z)\mathbf{W}^{(+)}(z)
\left(\frac{s_*}{h_0}\right)^{-\frac{\sigma_3}{2}}
|x|^{{\beta}\sigma_3}e^{i\frac{\sqrt{3}}{6}x^2\sigma_{3}}
\begin{pmatrix}1 & 0 \\ -\frac{1}{|x|\varphi_2(z)} & 1\end{pmatrix}2^{-\frac{\sigma_3}{2}}
\begin{pmatrix}|x|\varphi_2(z) & 1 \\ 1 & 0\end{pmatrix}
\end{equation*}
with $\mathbf{H}(z)$ and $\mathbf{W}^{(+)}(z)$ given by \eqref{H} and \eqref{W+}, respectively.
It is straightforward to check that $\widetilde{\mathbf{P}}^{(1,+)}(z)$ solves the following RH problem.
\subsection*{RH problem for $\widetilde{\mathbf{P}}^{(1,+)}(z)$}
\begin{description}
\item{(1)} $\widetilde{\mathbf{P}}^{(1,+)}(z)$ is analytic for all $z\in U(z_{1,+},\delta)\setminus\Sigma_\mathbf{T}$.
\item{(2)} $\widetilde{\mathbf{P}}^{(1,+)}(z)$ fulfills the same jump relations as $\mathbf{T}(z)$ on $U(z_{1,+},\delta)\cap\Sigma_\mathbf{T}$.
\item{(3)} On the circle $\partial U(z_{1,+},\delta)$, we have the matching condition
\begin{equation}\label{P1+matchingm}
\widetilde{\mathbf{P}}^{(1,+)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)\widetilde {\mathbf{P}}^{(\infty)}(z),\quad \mathrm{as}\quad x\rightarrow -\infty.
\end{equation}
\end{description}
Similarly, the local parametrix at $z_{1,-}=-\sqrt{\frac{2}{3}}$ is constructed by
\begin{align*}
\widetilde{\mathbf{P}}^{(1,-)}(z)=\widetilde {\mathbf{E}}^{(1,-)}(z)\Phi^{\mathrm{(PC)}}\left(|x|\varphi_3(z)\right)
\left(\frac{s_*}{h_0}\right)^{\frac{\sigma_3}{2}}\left\{\begin{aligned}&e^{2\pi i{\beta}\sigma_3}\left(s_0e^{2\pi i\alpha}\right)^{\frac{\sigma_3}{2}}
e^{-x^2g(z)\sigma_3},& \mathop{\rm Im}\nolimits z>0,\\
&\left(s_0e^{2\pi i\alpha}\right)^{-\frac{\sigma_3}{2}}
\sigma_{3}\sigma_{1}e^{-x^2g(z)\sigma_3},
& \mathop{\rm Im}\nolimits z<0,\end{aligned}\right.
\end{align*}
where $\varphi_3(z)$ is the conformal mapping \eqref{varphi3} and $\widetilde{\mathbf{E}}^{(1,-)}(z)$ is given by
\begin{equation*}
\widetilde{\mathbf{E}}^{(1,-)}(z)=\mathbf{H}(z)\mathbf{W}^{(-)}(z)
\left(\frac{s_*}{h_0}\right)^{-\frac{\sigma_3}{2}}
|x|^{{\beta}\sigma_3}e^{\frac{i\sqrt{3}x^2}{6}
\sigma_{3}}\begin{pmatrix}1 & 0 \\ -\frac{1}{|x|\varphi_3(z)} & 1\end{pmatrix}2^{-\frac{\sigma_3}{2}}\begin{pmatrix}|x|\varphi_3(z) & 1 \\ 1 & 0\end{pmatrix}
\end{equation*}
with $\mathbf{H}(z)$ and $\mathbf{W}^{(-)}(z)$ given by \eqref{H} and \eqref{W-}, respectively.
Then, $\widetilde{\mathbf{P}}^{(1,-)}(z)$ satisfies the following RH problem.
\subsection*{RH problem for $\widetilde{\mathbf{P}}^{(1,-)}(z)$}
\begin{description}
\item{(1)} $\widetilde{\mathbf{P}}^{(1,-)}(z)$ is analytic for all $z\in U(z_{1,-},\delta)\setminus\Sigma_\mathbf{T}$.
\item{(2)} $\widetilde{\mathbf{P}}^{(1,-)}(z)$ satisfies the same jump relations as $\mathbf{T}(z)$ on $U(z_{1,-},\delta)\cap\Sigma_\mathbf{T}$.
\item{(3)} On the boundary $\partial U(z_{1,-},\delta)$, we have
\begin{equation}\label{P1-matchingm}
\widetilde{\mathbf{P}}^{(1,-)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)\widetilde {\mathbf{P}}^{(\infty)}(z),\quad \mathrm{as}\quad x\rightarrow -\infty.
\end{equation}
\end{description}
Since the function $\mathbf{H}(z)$ defined by \eqref{H} is analytic at $z_{2,\pm}=\pm\sqrt{8/3}$, in order to construct the local parametrices $\widetilde {\mathbf{P}}^{(2,\pm)}(z)$ at these points, we only need to replace $\mathbf{P}^{(\infty)}(z)$ by $\widetilde{\mathbf{P}}^{(\infty)}(z)$ in the constructions of local parametrices at these points in the case $0<|s_*|<1$; see \eqref{P2+} and \eqref{P2-}. Moreover, we have the following matching conditions on the circular boundaries:
\begin{align}\label{P2+matchingm}
\widetilde{\mathbf{P}}^{(2,+)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)\widetilde {\mathbf{P}}^{(\infty)}(z),\quad \mathrm{as} \quad x\to-\infty
\end{align}
and
\begin{equation}\label{P2-matchingm}
\widetilde{\mathbf{P}}^{(2,-)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)\widetilde {\mathbf{P}}^{(\infty)}(z),\quad \mathrm{as} \quad x\to-\infty.
\end{equation}
\subsection{Modified local parametrix near the origin}
Accordingly, the parametrix $\widetilde{\mathbf{P}}^{(0)}(z)$ is defined as
\begin{equation}\label{Ptilde0}
\widetilde{\mathbf{P}}^{(0)}(z)=\widetilde{\mathbf{E}}^{(0)}(z)\Phi^{(\mathrm{Bes})}
\left(x^2\varphi_4(z)\right)\mathbf{K}(z)
\left[s_1\left(e^{-2\pi i\alpha}+s_*\right)\right]^{-\frac{\sigma_3}{2}}e^{-x^2g(z)\sigma_3},
\end{equation}
where $\varphi_4(z)$ is the conformal mapping \eqref{varphi4}, $\mathbf{K}(z)$ is defined by \eqref{K(z)} and $\widetilde{\mathbf{E}}^{(0)}(z)$ is given by
\begin{equation}\label{Etilde(0)}
\widetilde{\mathbf{E}}^{(0)}(z)=\widetilde{\mathbf{P}}^{(\infty)}(z)
\left[s_1(e^{-2\pi i\alpha}+s_*)\right]^{\frac{\sigma_3}{2}}\mathbf{Q}(z)e^{-\frac{1}{4}\pi i\sigma_3}\frac{1}{\sqrt{2}}\begin{pmatrix}1 & i\\ i & 1\end{pmatrix}
\end{equation}
with $\mathbf{Q}(z)$ defined by \eqref{Q}.
It is direct to verify that $\widetilde{\mathbf{P}}^{(0)}(z)$ satisfies the same jump conditions as $\mathbf{T}(z)$ and $\widetilde{\mathbf{P}}^{(0)}(z)$ possesses the same asymptotic behavior near the origin as $\mathbf{T}(z)$ for the case $\alpha-\frac{1}{2}\notin\mathbb{Z}$. Moreover, $\widetilde{\mathbf{P}}^{(0)}(z)$ fulfills the following matching condition
\begin{equation}\label{P0matchingm}
\widetilde{\mathbf{P}}^{(0)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)
\widetilde{\mathbf{P}}^{(\infty)}(z)\quad \mathrm{as} \quad x\to-\infty.
\end{equation}
While, in the case $\frac{1}{2}-\alpha\in\mathbb{N}$ and $|s_*|>1$, we need to check that $\widetilde{\mathbf{P}}^{(0)}(z)$ shares the same asymptotic behaviors as $\mathbf{T}(z)$ near the origin. From \eqref{E0} and \eqref{T(z)}, it follows that as $z\to0$ with $\arg z\in (0,\frac{\pi}{2})$
\begin{equation}\label{Ttilde0}
\mathbf{T}(z)=\widetilde{\mathbf{T}}_0(z)z^{\alpha\sigma_{3}}S_1 e^{-x^2g(z)\sigma_{3}}
\begin{pmatrix}1 & 0\\ -\frac{\overline{s}_*e^{-2\pi i\alpha}e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix},
\end{equation}
where $\widetilde{\mathbf{T}}_0(z)$ is analytic in a neighborhood of $z=0$. Using the following factorization
\begin{equation}
\begin{pmatrix}1 & s_1 \\ 0 & 1 \end{pmatrix}=\begin{pmatrix}1 & 0 \\ s_1^{-1} & 1 \end{pmatrix}\begin{pmatrix}0 & s_1 \\ -s_1^{-1} & 0 \end{pmatrix}\begin{pmatrix}1 & 0 \\ s_1^{-1} & 1 \end{pmatrix},
\end{equation}
we can rewrite \eqref{Ttilde0} as
\begin{equation}\label{Ttilde1}
\mathbf{T}(z)=\widetilde{\mathbf{T}}_1(z)z^{-\alpha\sigma_{3}}\begin{pmatrix}1 & 0\\ 1 & 1 \end{pmatrix}\left[s_1(e^{-2\pi i\alpha}+s_*)\right]^{-\frac{\sigma_3}{2}}
e^{-x^2g(z)\sigma_3},
\end{equation}
where $\widetilde{\mathbf{T}}_1(z)$ is analytic in a neighborhood of $z=0$. Comparing \eqref{Ttilde1} with \eqref{Ptilde0} and \eqref{BesParaExpand1}, it is readily seen that $\widetilde{\mathbf{P}}^{(0)}(z)$ satisfies the asymptotic behavior \eqref{Ttilde0} as $z\to0$ for $\frac{1}{2}-\alpha\in\mathbb{N}$.
\subsection{Final transformation}
The final transformation is defined as
\begin{equation}\label{Rtilde}
\widetilde{\mathbf{R}}(z)=\left\{\begin{aligned}
&\mathbf{T}(z)\left[\widetilde{\mathbf{P}}^{(1,\pm)}(z)\right]^{-1},\quad &z&\in U(z_{1,\pm},\delta)\setminus\Sigma_\mathbf{T},\\
&\mathbf{T}(z)\left[\widetilde{\mathbf{P}}^{(2,\pm)}(z)\right]^{-1},\quad &z&\in U(z_{2,\pm},\delta)\setminus\Sigma_\mathbf{T},\\
&\mathbf{T}(z)\left[\widetilde{\mathbf{P}}^{(0)}(z)\right]^{-1},\quad &z&\in U(0,\delta)\setminus\Sigma_\mathbf{T},\\
&\mathbf{T}(z)\left[\widetilde{\mathbf{P}}^{(\infty)}(z)\right]^{-1},\quad &\mathrm{e}&\mathrm{lsewhere}.\\ \end{aligned}
\right.
\end{equation}
As a consequence, $\widetilde{\mathbf{R}}(z)$ solves the following RH problem.
\subsection*{RH problem for $\widetilde{\mathbf{R}}(z)$}
\begin{description}
\item{(1)} $\widetilde{\mathbf{R}}(z)$ is analytic for $z\in \mathbb{C}\setminus\Sigma_{\mathbf{R}}$, where $\Sigma_\mathbf{R}$ is illustrated in Figure \ref{Rjump}.
\item{(2)} On $\Sigma_{R}$, we have $\widetilde{\mathbf{R}}_+(z)=\widetilde{\mathbf{R}}_-(z)J_{\widetilde{\mathbf{R}}}(z)$, where
\begin{equation}\label{Rtildejump}
J_{\widetilde{\mathbf{R}}}(z)=\left\{\begin{aligned}
&\widetilde{\mathbf{P}}^{(1,\pm)}(z)\widetilde{\mathbf{P}}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(z_{1,\pm},\delta),\\
&\widetilde{\mathbf{P}}^{(2,\pm)}(z)\widetilde{\mathbf{P}}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(z_{2,\pm},\delta),\\
&\widetilde{\mathbf{P}}^{(0)}(z)\widetilde{\mathbf{P}}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(0,\delta),\\
&\widetilde{\mathbf{P}}^{(\infty)}_-(z)J_{\mathbf{T}}(z)
\widetilde{\mathbf{P}}^{(\infty)}_+(z)^{-1},\quad &\mathrm{e}&\mathrm{lsewhere}.\\ \end{aligned}
\right.
\end{equation}
\item{(3)} As $z\rightarrow\infty$, we have
\begin{equation}\label{Rtildeexp}
\widetilde{\mathbf{R}}(z)=\mathbf{I}+\frac{\widetilde{\mathbf{R}}_1}{z}+\frac{\widetilde {\mathbf{R}}_2}{z^2}+O(z^{-3}).
\end{equation}
\end{description}
Using the matching conditions \eqref{P1+matchingm}, \eqref{P1-matchingm}, \eqref{P2+matchingm}, \eqref{P2-matchingm} and \eqref{P0matchingm},
we have the following estimates
for the jump matrices \eqref{Rtildejump} as $x\rightarrow-\infty$
\begin{equation}\label{Rtildejumpestima}
J_{\widetilde{\mathbf{R}}}(z)=\left\{\begin{aligned}
&\mathbf{I}+O(|x|^{-2}), &z&\in\partial U(0,\delta)\cup\partial U(z_{1,\pm},\delta)\cup\partial U(z_{2,\pm},\delta),\\
&\mathbf{I}+O(e^{-c_{2}|x|^2}),&z&\in\pi_k,\ k=1,\cdots,16.
\end{aligned}\right.
\end{equation}
where $c_2$ is some positive constant. Consequently, we have that for any $z\in \mathbb{C}\setminus\Sigma_\mathbf{R}$
\begin{equation}\label{Rtildeexpx}
\widetilde{\mathbf{R}}(z)=\mathbf{I}+O(|x|^{-2}),\qquad \mathrm{as}\qquad x\rightarrow-\infty.
\end{equation}
\section{RH analysis as $x\to-\infty$ with $|s_*|=1$ but $s_*\neq1$}\label{Asymptotic-infty3}
In the case when $|s_*|=1$ with $s_*\neq1$, it follows from \eqref{sstar} and \eqref{sstarbar} that
\begin{equation}\label{sstarvalue}
s_*=-e^{-2\pi i\alpha}.
\end{equation}
We need another $g$-function
\begin{equation}\label{gbar}
\widehat{g}(z)=\frac{1}{8}(z^2-2)^2=\frac{1}{8}z^4-\frac{1}{2}z^2+\frac{1}{2}.
\end{equation}
It is direct to see that $\widehat{g}(z)$ has three saddle points $z_0=0$, $z_{\pm}=\pm\sqrt{2}$.
As shown in Figure \ref{Ubarjump}, the topology of the anti-Stokes curves of $\widehat {g}(z)$ is quite different from $g(z)$ depicted in Figure \ref{ASC}.
As before, we begin with the re-scaling transformation \eqref{rescaling}. To proceed, we introduce
\begin{equation}\label{Ubar}
\widehat{\mathbf{U}}(z)=e^{\frac{x^2}{2}\sigma_3}|x|^{\alpha\sigma_3}\Phi(z)
|x|^{-\alpha\sigma_3}e^{-x^2\widehat{g}(z)\sigma_3}.
\end{equation}
Then, $\widehat{\mathbf{U}}(z)$ satisfies a RH problem with jumps along $\gamma_k$, $k=1,3,4,5,7,8$ as shown in Figure \ref{PIVj}.
Subsequently, we deform the jump contours to the anti-Stokes curves of $\widehat {g}(z)$ illustrated in Figure \ref{Ubarjump}. We rewrite the RH problem as the following problem formulated on the anti-Stokes curves of $\widehat {g}(z)$.
\begin{figure}[H]
\centering
\includegraphics[width=13cm,height=6cm]{Trans11}\\
\caption{The jump contour $\Sigma_{\widehat{\mathbf{U}}}$}\label{Ubarjump}
\end{figure}
\subsection*{RH problem for $\widehat{\mathbf{U}}(z)$}
\begin{description}
\item{(1)} $\widehat{\mathbf{U}}(z)$ is analytic for $z\in\mathbb{C}\setminus \Sigma_{\widehat{\mathbf{U}}}$, with $\Sigma_{\widehat{\mathbf{U}}}$ is shown in Figure \ref{Ubarjump}. Note that we have deformed $\gamma_k$ to the anti-Stokes curves of $\widehat{g}(z)$, namely $\widehat{\gamma}_{k}$, $k=1,3,5,7$.
\item{(2)} On the contour $\Sigma_{\widehat{\mathbf{U}}}$, we have $\widehat{\mathbf{U}}_{+}(z)=\widehat{\mathbf{U}}_{-}(z)J_{\widehat{\mathbf{U}}}(z)$, where
$$J_{\widehat{\mathbf{U}}}(z)=\left\{
\begin{aligned}
&\begin{pmatrix}1 & s_k|x|^{2\alpha}e^{2x^2\widehat{g}(z)}\\ 0 & 1 \end{pmatrix},\quad &z&\in\widehat{\gamma}_{k},\ k=1,3,5,7,\\
&\begin{pmatrix}-1 & 0 \\ s_0e^{2\pi i\alpha}|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)} & -1 \end{pmatrix},\quad &z&\in\widehat{\gamma}_{2},\\
&\begin{pmatrix}-e^{-2\pi i\alpha} & 0 \\ s_0e^{2\pi i\alpha}|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)} & -e^{2\pi i\alpha}\end{pmatrix},\quad &z&\in\widehat{\gamma}_{6},\\
&\begin{pmatrix}1 & 0\\ -s_0e^{2\pi i\alpha}|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)} & 1 \end{pmatrix},\quad &z&\in\gamma_{4},\\
&\begin{pmatrix}e^{-2\pi i\alpha} & 0\\ s_0e^{2\pi i\alpha}|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)} & e^{2\pi i\alpha} \end{pmatrix},\quad &z&\in\gamma_{8}.
\end{aligned}
\right.
$$
\item{(3)} $\widehat{\mathbf{U}}(z)=\left(\mathbf{I}+O(z^{-1})\right)z^{\alpha\sigma_3}$ as $z\rightarrow\infty$, where $\arg z\in(0,2\pi)$.
\item{(4)} $\widehat{\mathbf{U}}(z)$ possesses the following asymptotic behavior near the origin
\begin{equation}\label{Ubaratzero}
\widehat{\mathbf{U}}(z)=\widehat{\mathbf{U}}_{0}(z)z^{\alpha\sigma_3}E_0S_1
|x|^{-\alpha\sigma_3}e^{-x^2\widehat{g}(z)\sigma_3},\qquad \arg z\in(\pi/4,\pi/2),
\end{equation}
where $\widehat{\mathbf{U}}_{0}(z)$ is analytic in the neighborhood of $z_0=0$. The asymptotic behaviors of $\widehat{\mathbf{U}}(z)$ in other regions are determined by \eqref{Ubaratzero} and the jump relations satisfied by $\widehat{\mathbf{U}}(z)$.
\end{description}
From the properties of $g$-function $\widehat{g}(z)$, it is readily seen that the jump matrices on $\gamma_4$ and the anti-Stokes curves $\widehat{\gamma}_k$, $k=1,3,5,7$ approach to identity matrix exponentially fast as $x\rightarrow-\infty$. Thus, the task is to construct a global parametrix satisfying the remaining jump along $(z_-,+\infty)$ and three local parametrices in the neighborhoods of the saddle points $z_0$, $z_{\pm}$.
\subsection{Global parametrix }
We need to solve the following RH problem for a $2\times 2$ matrix-valued function $\widehat{\mathbf{P}}^{(\infty)}(z)$.
\subsection*{RH problem for $\widehat{\mathbf{P}}^{(\infty)}(z)$}
\begin{description}
\item{(1)} $\widehat{\mathbf{P}}^{(\infty)}(z)$ is analytic for $z\in \mathbb{C}\setminus (z_{-},+\infty)$.
\item{(2)} $\widehat{\mathbf{P}}^{(\infty)}(z)$ satisfies the following jump relations
\begin{equation}\label{Phatjump}
\widehat{\mathbf{P}}^{(\infty)}_{+}(z)=\widehat{\mathbf{P}}^{(\infty)}_{-}(z)
\left\{\begin{aligned}
&\begin{pmatrix}-e^{-2\pi i\alpha} & 0 \\ 0 & -e^{2\pi i\alpha}\end{pmatrix},\quad& z&\in(z_{-},0), \\
&-\mathbf{I},\quad & z&\in(0,z_{+}),\\
&\begin{pmatrix}e^{-2\pi i\alpha} & 0 \\ 0 & e^{2\pi i\alpha}\end{pmatrix},\quad & z&\in(z_{+},+\infty)\\
\end{aligned}\right.
\end{equation}
\item{(3)} $\widehat{\mathbf{P}}^{(\infty)}(z)$ has at most singularities of order less than $|\alpha|+\frac{3}{2}$ at $z_{\pm}=\pm\sqrt{2}$.
\item{(4)} As $z\rightarrow \infty$, we have
$\widehat{\mathbf{P}}^{(\infty)}(z)=\left(\mathbf{I}+O(z^{-1})\right)z^{\alpha\sigma_3}$.
\end{description}
A solution to the above RH problem is given by
\begin{equation}\label{Pbarinfty}
\widehat{\mathbf{P}}^{(\infty)}(z)=\widehat{\mathbf{H}}(z)z^{-\alpha\sigma_3}
\left(z+\sqrt{2}\right)^{(\alpha+\frac{1}{2})\sigma_3}
\left(z-\sqrt{2}\right)^{(\alpha-\frac{1}{2})\sigma_3},
\end{equation}
where the branches of the powers are chosen such that $\arg z\in(0,2\pi)$ and $\arg(z\pm\sqrt{2})\in(0,2\pi)$. The function $\widehat{\mathbf{H}}(z)$, similar to \eqref{H}, takes the rational form
\begin{equation}\label{H(z)bar}
\widehat{\mathbf{H}}(z)=\mathbf{I}+\frac{\widehat{A}}{z+\sqrt{2}}+\frac{\widehat {B}}{z-\sqrt{2}},\quad \det\widehat{\mathbf{H}}(z)=1,
\end{equation}
where constant matrices $\widehat{A}$ and $\widehat{B}$ are to be determined. We point out that the factor $\widehat{\mathbf{H}}(z)$ in \eqref{Pbarinfty} is introduced to meet the matching conditions \eqref{matchingPr} and \eqref{matchingPl} below.
\subsection{Local parametrices at $z_{\pm}=\pm\sqrt{2}$}
We seek two functions $\widehat{\mathbf{P}}^{(r)}(z)$ and $\widehat{\mathbf{P}}^{(l)}(z)$, satisfying the same jumps as $\widehat{\mathbf{\mathbf{U}}}(z)$ respectively in the neighborhoods $U(z_{\pm},\delta)=\{z\in\mathbb{C}\mid|z-z_{\pm}|<\delta\}$ of the saddle points $z_{\pm}$, and matching with $\widehat{\mathbf{P}}^{(\infty)}(z)$ on the boundaries $\partial U(z_{\pm},\delta)=\{z\in \mathbb{C}\mid|z-z_{\pm}|=\delta\}$.
\subsection*{RH problem for $\widehat{\mathbf{P}}^{(r)}(z)$}
\begin{description}
\item{(1)} $\widehat{\mathbf{P}}^{(r)}(z)$ is analytic for $z\in U(z_{+},\delta)\setminus \Sigma_{\widehat{\mathbf{U}}}$.
\item{(2)} $\widehat{\mathbf{P}}^{(r)}(z)$ shares the same jumps as $\widehat{\mathbf{U}}(z)$ on $\Sigma_{\widehat{\mathbf{U}}}\cap U(z_{+},\delta)$.
\item{(3)} On the boundary $\partial U(z_+,\delta)$, we have
\begin{equation}\label{matchingPr}
\widehat{\mathbf{P}}^{(r)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)
\widehat{\mathbf{P}}^{(\infty)}(z)\quad \mathrm{as}\quad x\rightarrow-\infty.
\end{equation}
\end{description}
Firstly, we introduce a conformal mapping
\begin{equation}\label{varphi5}
\varphi_5(z)=2\widehat{g}(z)^{\frac{1}{2}}=2(z-z_{\pm})(1+o(1)), \quad z\rightarrow z_{\pm}.
\end{equation}
We will make use of the parabolic cylinder function to construct the solution. Let $\Phi^{(\mathrm{PC})}$ be the parabolic cylinder parametrix with parameter $\beta=\frac{1}{2}-\alpha$ as given in Appendix \ref{PCP}. Then, the parametrix $\widehat{\mathbf{P}}^{(r)}(z)$ is defined as
\begin{equation}\label{Pr}
\widehat{\mathbf{P}}^{(r)}(z)=\mathbf{E}^{(r)}(z)\Phi^{(\mathrm{PC})}(|x|\varphi_5(z))
\left(\frac{h_1^{(r)}}{s_1}\right)^{\frac{\sigma_3}{2}}
\mathbf{D}^{(r)}(z)\sigma_3|x|^{-\alpha\sigma_3}e^{-x^2\widehat{g}(z)\sigma_3},
\end{equation}
where $h_1^{(r)}=\frac{\sqrt{2\pi}e^{-(\alpha-\frac{1}{2})\pi i}}{\Gamma(\alpha-\frac{1}{2})}$ is the Stokes multiplier defined in \eqref{h0}, \begin{equation}\label{Dr}
\mathbf{D}^{(r)}(z)=\left\{\begin{aligned}
&e^{2\pi(\alpha+\frac{1}{4})\sigma_3},
&\arg z&\in(-\frac{\pi}{4},0),\\
&e^{\frac{1}{2}\pi i\sigma_3}, &\arg z&\in(0,\pi),\\
&e^{-\frac{1}{2}\pi i\sigma_3}, &\arg z&\in(\pi,\frac{7\pi}{4}).
\end{aligned}\right.
\end{equation}
and $\mathbf{E}^{(r)}(z)$ is given by
\begin{align}\label{Er}
\mathbf{E}^{(r)}(z)=\widehat{\mathbf{P}}^{(\infty)}(z)\sigma_
&\left(\frac{s_1}{h_1^{(r)}}\right)^{\frac{\sigma_3}{2}}|x|^{\frac{\sigma_3}{2}}
e^{\mp\frac{1}{2}\pi i\sigma_3}\nonumber\\
&\times\varphi_5(z)^{(\frac{1}{2}-\alpha)\sigma_3}
\begin{pmatrix}1 & \frac{\alpha-\frac{1}{2}}{|x|\varphi_5(z)} \\ 0 & 1
\end{pmatrix}2^{-\frac{\sigma_3}{2}}\begin{pmatrix} |x|\varphi_5(z) & 1\\
1 & 0 \end{pmatrix},~~\pm\mathop{\rm Im}\nolimits z>0
\end{align}
with $\arg\varphi_5(z)\in(0,2\pi)$.
It is readily seen from \eqref{Phatjump}, \eqref{Er} and \eqref{PCAsyatinfty} that $\mathbf{E}^{(r)}(z)$ is analytic in the deleted neighborhood $U(z_+,\delta)\setminus\{z_+\}$ and the matching condition \eqref{matchingPr} is satisfied. To guarantee that $\mathbf{E}^{(r)}(z)$ is also analytic at the isolated point $z_+=\sqrt{2}$, we find, by computing the Laurent expansion at $z_+=\sqrt{2}$ using \eqref{Pbarinfty}, \eqref{H(z)bar}, \eqref{varphi5}, \eqref{Dr} and \eqref{Er}, that the constant matrices $\widehat{A}$ and $\widehat{B}$ in \eqref{H(z)bar} should satisfy the following algebraic equation
\begin{equation}\label{AbarBbar1}
\left(I+\frac{\widehat{ A}}{2\sqrt{2}}\right)\begin{pmatrix} 0 & c_r \\ 0 & 0\end{pmatrix}=-\widehat{B},
\end{equation}
where the constant $c_r$ is given by
\begin{equation}\label{cr}
c_r=\frac{2s_1\Gamma(\alpha+\frac{1}{2})e^{i\pi(\alpha-\frac{1}{2})}}{\sqrt{\pi}}.
\end{equation}
\subsection*{RH problem for $\widehat{\mathbf{P}}^{(l)}(z)$}
\begin{description}
\item{(1)} $\widehat{\mathbf{P}}^{(l)}(z)$ is analytic for $z\in U(z_{-},\delta)\setminus \Sigma_{\widehat{\mathbf{U}}}$.
\item{(2)} $\widehat{\mathbf{P}}^{(l)}(z)$ satisfies the same jumps as $\widehat{\mathbf{U}}(z)$ on $\Sigma_{\widehat{\mathbf{U}}}\cap U(z_-,\delta)$.
\item{(3)} On the circle $\partial U(z_-,\delta)$, it holds that
\begin{equation}\label{matchingPl}
\widehat{\mathbf{P}}^{(l)}(z)=\left(\mathbf{I}+O(|x|^{-2})\right)
\widehat{\mathbf{P}}^{(\infty)}(z)\quad \mathrm{as}\quad x\rightarrow-\infty.
\end{equation}
\end{description}
Similarly, the solution $\widehat{\mathbf{P}}^{(l)}(z)$ can also be built in terms of the parabolic cylinder function. We choose the parameter $\beta=-\frac{1}{2}-\alpha$ in Appendix \ref{PCP}. More precisely, we define
\begin{equation}\label{Pl}
\widehat{\mathbf{P}}^{(l)}(z)=\mathbf{E}^{(l)}(z)\Phi^{(\mathrm{PC})}(|x|\varphi_5(z))
\left(\frac{h_1^{(l)}}{s_1}\right)^{\frac{\sigma_3}{2}}
\mathbf{D}^{(l)}(z)\sigma_3|x|^{-\alpha\sigma_3}e^{-x^2\widehat{g}(z)\sigma_3},
\end{equation}
where $h_1^{(l)}=\sqrt{2\pi}e^{-(\alpha+\frac{1}{2})\pi i}/\Gamma(\alpha+\frac{1}{2})$ is the Stokes multiplier given in \eqref{h0},
\begin{equation}\label{Dl}
\mathbf{D}^{(l)}(z)=\left\{\begin{aligned}
&e^{2\pi(\alpha+\frac{1}{2})\sigma_3}, &\arg z&\in(-\frac{\pi}{4},0),\\
&\mathbf{I},
&\arg z&\in(0,\frac{7\pi}{4}).
\end{aligned}\right.
\end{equation}
and $\mathbf{E}^{(l)}(z)$ is given by
\begin{align}\label{El}
\mathbf{E}^{(l)}(z)=\widehat{\mathbf{P}}^{(\infty)}(z)\sigma_
&\left(\frac{s_1}{h_1^{(l)}}\right)^{\frac{\sigma_3}{2}}
|x|^{-\frac{\sigma_3}{2}
\varphi_5(z)^{-(\alpha+\frac{1}{2})\sigma_3}\begin{pmatrix}
1 & 0 \\ -\frac{1}{|x|\varphi_5(z)} & 1
\end{pmatrix}2^{-\frac{\sigma_3}{2}}\begin{pmatrix} |x|\varphi_5(z) & 1\\
1 & 0 \end{pmatrix}
\end{align}
with $\arg\varphi_5(z)\in(0,2\pi)$.
Using the jumps \eqref{Phatjump}, it is straightforward to check that $\mathbf{E}^{(l)}(z)$ is analytic in the deleted neighborhood $U(z_-,\delta)\setminus\{z_-\}$ and the matching condition \eqref{matchingPl} is also satisfied. To ensure that $\mathbf{E}^{(l)}(z)$ is also analytic at the isolated point $z_-=-\sqrt{2}$, by calculating the Laurent expansion at $z_-$ using \eqref{Pbarinfty}, \eqref{H(z)bar}, \eqref{varphi5}, \eqref{El} and \eqref{Dl}, we obtain another algebraic equation
\begin{equation}\label{AB2}
\left(I-\frac{\widehat{B}}{2\sqrt{2}}\right)\begin{pmatrix} 0 & 0 \\ c_l & 0\end{pmatrix}=-\widehat{A},
\end{equation}
where the constant $c_l$ is given by
\begin{equation}\label{cl}
c_l=-\frac{4\sqrt{\pi}}
{s_1\Gamma(\alpha+\frac{1}{2})e^{i\pi(\alpha+\frac{1}{2})}}.
\end{equation}
Combining \eqref{AbarBbar1}, \eqref{AB2} with the fact that $c_rc_l=8$, we can now derive explicit expressions of $\widehat{A}$ and $\widehat{B}$ as follows:
\begin{equation}\label{ABexpres}
\widehat{A}=\begin{pmatrix} -\sqrt{2} & 0\\ -\frac{1}{2}c_l & 0 \end{pmatrix},\quad
\widehat{B}=\begin{pmatrix} 0 & -\frac{1}{2}c_r\\ 0 & \sqrt{2} \end{pmatrix},
\end{equation}
where $c_r$ and $c_l$ are given by \eqref{cr} and \eqref{cl}, respectively.
After determining $\widehat{A}$ and $\widehat{B}$ as given in \eqref{ABexpres}, it is straightforward to verify that the determinant condition $\det\widehat{\mathbf{H}}(z)=1$ is also satisfied.
\subsection{Local parametrix near the origin}
Now, we look for a function $\widehat{\mathbf{P}}^{(0)}(z)$ satisfying the same jumps as $\widehat{\mathbf{U}}(z)$ in the neighborhood $U(z_0,\delta)=\{z\in\mathbb{C}\mid|z-z_0|<\delta\}$ of $z_0=0$ with some constant $0<\delta<\sqrt{2}$, and matching with $\widehat{\mathbf{P}}^{(\infty)}(z)$ on the boundary $\partial U(z_0,\delta)=\{z\in \mathbb{C}\mid|z-z_0|=\delta\}$.
\subsection*{RH problem for $\widehat{\mathbf{P}}^{(0)}(z)$}
\begin{description}
\item{(1)} $\widehat{\mathbf{P}}^{(0)}(z)$ is analytic for $z\in U(z_0,\delta)\setminus [z_{-},z_{+}]$.
\item{(2)} $\widehat{\mathbf{P}}^{(0)}(z)$ satisfies the following jump relations
\begin{equation*}
\widehat{\mathbf{P}}^{(0)}_{+}(z)=\widehat{\mathbf{P}}^{(0)}_{-}(z)
\left\{\begin{aligned}
&\begin{pmatrix}-e^{-2\pi i\alpha} & 0 \\ s_0e^{2\pi i\alpha}|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)} & -e^{2\pi i\alpha}\end{pmatrix},\quad& z&\in(-\delta,0), \\
&\begin{pmatrix}-1 & 0 \\ s_0e^{2\pi i\alpha}|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)} & -1\end{pmatrix},\quad & z&\in(0,\delta).
\end{aligned}\right.
\end{equation*}
\item{(3)} On the boundary $\partial U(z_0,\delta)$, we have
\begin{equation}\label{matchingcondition0}
\widehat{\mathbf{P}}^{(0)}(z)=(\mathbf{I}+O(|x|^{-2\alpha}e^{-x^2}))\widehat {\mathbf{P}}^{(\infty)}(z)\quad \mathrm{as}\quad x\to-\infty.
\end{equation}
\item{(4)} $\widehat{\mathbf{P}}^{(0)}(z)$ has the same behavior as $\widehat{\mathbf{U}}(z)$ near the origin; see \eqref{Ubaratzero}.
\end{description}
A solution to the above RH problem can be constructed explicitly as follows:
\begin{equation}\label{Pbar0}
\widehat{\mathbf{P}}^{(0)}(z)=\widehat{\mathbf{H}}(z)\begin{pmatrix}1 & 0\\ r(z) & 1 \end{pmatrix}
z^{-\alpha\sigma_3}
\left(z+\sqrt{2}\right)^{(\alpha+\frac{1}{2})\sigma_3}
\left(z-\sqrt{2}\right)^{(\alpha-\frac{1}{2})\sigma_3},
\end{equation}
where $\widehat{\mathbf{H}}(z)$ is given by \eqref{H(z)bar} and \eqref{ABexpres}. The function $r(z)$ is defined by
\begin{equation}\label{rexpre}
r(z)=\left\{\begin{aligned} &
-\frac{s_0|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)}z^{2\alpha}}{(1+e^{-2\pi i\alpha})(z+\sqrt{2})^{2\alpha+1}(z-\sqrt{2})^{2\alpha-1}}, && |z|<\delta,~\mathop{\rm Im}\nolimits z>0,\\
&
\frac{s_0|x|^{-2\alpha}e^{-2x^2\widehat{g}(z)}e^{2\pi i\alpha}z^{2\alpha}}{(1+e^{-2\pi i\alpha})(z+\sqrt{2})^{2\alpha+1}(z-\sqrt{2})^{2\alpha-1}}, && |z|<\delta,~\mathop{\rm Im}\nolimits z<0,\end{aligned}\right.
\end{equation}
where the branches are chosen such that $\arg z\in(0,2\pi)$ and $\arg(z\pm\sqrt{2})\in(0,2\pi)$.
\subsection{Final transformation}
The final transformation is now defined as
\begin{equation}\label{Rbar}
\widehat{\mathbf{R}}(z)=\left\{\begin{aligned}
&\widehat{\mathbf{U}}(z)\left[\widehat{\mathbf{P}}^{(l)}(z)\right]^{-1},\quad &z&\in U(z_-,\delta)\setminus\Sigma_{\widehat{\mathbf{U}}},\\
&\widehat{\mathbf{U}}(z)\left[\widehat{\mathbf{P}}^{(r)}(z)\right]^{-1},\quad &z&\in U(z_+,\delta)\setminus\Sigma_{\widehat{\mathbf{U}}},\\
&\widehat{\mathbf{U}}(z)\left[\widehat{\mathbf{P}}^{(0)}(z)\right]^{-1},\quad &z&\in U(z_0,\delta)\setminus\Sigma_{\widehat{\mathbf{U}}},\\
&\widehat{\mathbf{U}}(z)\left[\widehat{\mathbf{P}}^{(\infty)}(z)\right]^{-1},\quad &\mathrm{e}&\mathrm{lsewhere}.\\ \end{aligned}
\right.
\end{equation}
Then, $\widehat{\mathbf{R}}(z)$ satisfies the following RH problem.
\subsection*{RH problem for $\widehat{\mathbf{R}}(z)$}
\begin{description}
\item{(1)} $\widehat{\mathbf{R}}(z)$ is analytic for $z\in \mathbb{C}\setminus\Sigma_{\widehat{\mathbf{R}}}$, where $\Sigma_{\widehat{\mathbf{R}}}$ is depicted in Figure \ref{Rbarjumppicture}.
\item{(2)} On the contour $\Sigma_{\widehat{\mathbf{R}}}$, we have $\widehat{\mathbf{R}}_+(z)=\widehat{\mathbf{R}}_-(z)J_{\widehat{\mathbf{R}}}(z)$, where
\begin{equation}\label{Rbarjump}
J_{\widehat{\mathbf{R}}}(z)=\left\{\begin{aligned}
&\widehat{\mathbf{P}}^{(l)}(z)\widehat{\mathbf{P}}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(z_-,\delta),\\
&\widehat{\mathbf{P}}^{(r)}(z)\widehat{\mathbf{P}}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(z_+,\delta),\\
&\widehat{\mathbf{P}}^{(0)}(z)\widehat{\mathbf{P}}^{(\infty)}(z)^{-1},\quad &z&\in \partial U(z_0,\delta)\\
&\widehat{\mathbf{P}}^{(\infty)}_-(z)J_{\widehat{\mathbf{U}}}(z)\widehat {\mathbf{P}}^{(\infty)}_+(z)^{-1},\quad &\mathrm{e}&\mathrm{lsewhere}.\end{aligned}
\right.
\end{equation}
\item{(3)} As $z\rightarrow\infty$, $\widehat{\mathbf{R}}(z)$ admits the expansion
\begin{equation}\label{Rbarexp}
\widehat{\mathbf{R}}(z)=\mathbf{I}+\frac{\widehat{\mathbf{R}}_1}{z}
+\frac{\widehat{\mathbf{R}}_2}{z^2}+O(z^{-3}).
\end{equation}
\end{description}
\begin{figure}[t]
\centering
\includegraphics[width=13cm,height=7cm]{Trans12}\\
\caption{The jump contour $\Sigma_{\widehat{\mathbf{R}}}$}\label{Rbarjumppicture}
\end{figure}
By virtue of the matching conditions \eqref{matchingPr}, \eqref{matchingPl} and \eqref{matchingcondition0}, the jump matrix \eqref{Rbarjump} has the following estimations as $x\rightarrow-\infty$
\begin{equation}\label{Rbarjumpestima}
J_{\widehat{\mathbf{R}}}(z)=\left\{\begin{aligned}
&\mathbf{I}+O(|x|^{-2}), &z&\in\partial U(z_{\pm},\delta),\\
&\mathbf{I}+O(|x|^{-2\alpha}e^{-x^2}), &z&\in\partial U(z_0,\delta)\cup\widehat{\gamma}_2\cup\widehat{\gamma}_6,\\
&\mathbf{I}+O(e^{-c_{3}x^2}),&\mathrm{e}&\mathrm{lsewhere},
\end{aligned}\right.
\end{equation}
where $c_3$ is some positive constant. As a result, we get that for any $z\in\mathbb{C}\setminus\Sigma_{\widehat{\mathbf{R}}}$,
\begin{equation}\label{Rbarexpans}
\widehat{\mathbf{R}}(z)=\mathbf{I}+O(|x|^{-2}),\quad \mathrm{as}\quad x\rightarrow-\infty.
\end{equation}
\section{Proof of Theorems \ref{thm1} - \ref{thm3}}\label{proof1}
To derive the asymptotics for the PIV solution $q(x;\alpha,\kappa)$ and the associated Hamiltonian $\mathcal{H}(x;\alpha,\kappa)$ as $x\rightarrow-\infty$, we use the identities \eqref{solu2} and \eqref{tau1} and the asymptotic analysis of the RH problem for $\Psi(\xi,x)$
performed in Sections \ref{Asymptotic-infty1}-\ref{Asymptotic-infty3}.
\subsection{Derivations of \eqref{q1} and \eqref{H1}}\label{Derivation1}
Let us focus on the case $\kappa(\kappa-\kappa^*)<0$. Tracing back the series of invertible transformations $$\Phi\mapsto \mathbf{U}\mapsto \mathbf{T}\mapsto \mathbf{R}$$
as defined in \eqref{U(z)}, \eqref{T(z)} and \eqref{R}), respectively, we obtain that for large $z$
\begin{equation*}
e^{\frac{x^2}{3}\sigma_3}\Phi(z)e^{-x^2g(z)\sigma_3}
=\mathbf{R}(z)\mathbf{P}^{(\infty)}(z).
\end{equation*}
It follows from the asymptotic expansions \eqref{Asyatinfty1} and \eqref{gatinfty} that
\begin{equation}\label{RHanalysis}
e^{\frac{x^2}{3}\sigma_3}\Phi(z)e^{-x^2g(z)\sigma_3}z^{-\alpha\sigma_3}
=\mathbf{I}+\frac{e^{\frac{x^2}{3}\sigma_3}\Phi_1e^{-\frac{x^2}{3}\sigma_3}}{z}
+\frac{e^{\frac{x^2}{3}\sigma_3}\Phi_2e^{-\frac{x^2}{3}\sigma_3}-\frac{4x^2}{27}
\sigma_3}{z^2}+O(z^{-3}).
\end{equation}
To compute the coefficients $\Phi_1$ and $\Phi_2$, we need to know the asymptotics of $\mathbf{R}(z)$ and $\mathbf{P}^{(\infty)}(z)$ as $z\to\infty$. The large-$z$ asymptotics of $\mathbf{R}(z)$ has already been given in \eqref{Rexpan}. As for $\mathbf{P}^{(\infty)}(z)$, we obtain from \eqref{Pinfty}, \eqref{X(z)atinfty} and \eqref{fatinfty} that
\begin{equation}\label{Pinftyatinfty}
\mathbf{P}^{(\infty)}(z)z^{-\alpha\sigma_3}=\mathbf{I}+\frac{\mathbf{P}_1^{(\infty)}}{z}
+\frac{\mathbf{P}_2^{(\infty)}}{z^2}+O\left(\frac{1}{z^3}\right),~~~\mbox{as}~ ~ z\rightarrow\infty,
\end{equation}
where
\begin{equation}\label{P1P2}
\mathbf{P}_1^{(\infty)}=\sqrt{\frac{2}{3}}s_0^{-\frac{\sigma_3}{2}}f_{\infty}^{-\sigma_3}
\sigma_2f_{\infty}^{\sigma_3}s_0^{\frac{\sigma_3}{2}},\quad \mathbf{P}_2^{(\infty)}=\frac{\mathbf{I}}{3}
-\frac{2}{3}\left(\alpha+\sqrt{3}i{\beta}\right)\sigma_3.
\end{equation}
A combination of \eqref{Rexpan}, \eqref{RHanalysis} and \eqref{Pinftyatinfty} gives
\begin{equation}\label{Phi1}
\Phi_1=e^{-\frac{x^2}{3}\sigma_3}\left[\mathbf{R}_1+\mathbf{P}_1^{(\infty)}\right]
e^{\frac{x^2}{3}\sigma_3}
\end{equation}
and
\begin{equation}\label{Phi2}
\Phi_2=e^{-\frac{x^2}{3}\sigma_3}\left[\mathbf{R}_1\mathbf{P}_1^{(\infty)}
+\mathbf{R}_2+\mathbf{P}_2^{(\infty)}\right]e^{\frac{x^2}{3}\sigma_3}
+\frac{4x^2}{27}\sigma_3.
\end{equation}
Using formulas \eqref{solu2}, \eqref{tau1} and combining them with \eqref{Rexpan}, \eqref{RHanalysis}, \eqref{P1P2}, \eqref{Phi1} and \eqref{Phi2}, we get the following expressions for $q(x;\alpha,\kappa)$ and $\mathcal{H}(x;\alpha,\kappa)$
\begin{equation}\label{qexpre2}
q(x;\alpha,\kappa)=-\frac{2}{3}x-i\sqrt{\frac{2}{3}}x\Big[s_0f_{\infty}^2(\mathbf{R}_1)_{12}
-s_0^{-1}f_{\infty}^{-2}(\mathbf{R}_1)_{21}\Big]-x(\mathbf{R}_1)_{12}(\mathbf{R}_1)_{21}
\end{equation}
and
\begin{align}\label{Hexpres}
\mathcal{H}(x;\alpha,\kappa)=&-\frac{8}{27}x^3+\frac{4}{3}(\alpha
+\sqrt{3}i{\beta})x-x\sqrt{\frac{2}{3}}i\Big[s_0f_{\infty}^2(\mathbf{R}_1)_{12}
+ s_0^{-1}f_{\infty}^{-2}(\mathbf{R}_1)_{21}\Big]\nonumber \\
&-x\big[(\mathbf{R}_2)_{11}-(\mathbf{R}_2)_{22}\big].
\end{align}
Thus, the remaining task is to compute the asymptotics of $\mathbf{R}_1$ and $\mathbf{R}_2$. It is observed
from \eqref{JRestimation} and \eqref{Restimation} that, neglecting a uniform error term $O(|x|^{-2})$, the contribution to $\mathbf{R}(z)-\mathbf{I}$ comes from the jumps on $\partial U(z_{1,\pm},\delta)$. Therefore, it follows from \eqref{matchingcondition+}, \eqref{matchingcondition-} and \eqref{JumpR} that as $x\to-\infty$,
\begin{equation}\label{Rexpanrevisited}
\mathbf{R}(z)=\mathbf{I}+\frac{\mathbf{R}^{(1)}(z)}{|x|}+O\left(\frac{1}{|x|^2}\right),\quad \mathrm{for} \quad z\in\partial U(z_{1,\pm},\delta),
\end{equation}
where the coefficient $\mathbf{R}^{(1)}(z)=O(z^{-1})$ as $z\rightarrow\infty$, and satisfies the jump relation
\begin{equation}
\mathbf{R}^{(1)}_+(z)-\mathbf{R}^{(1)}_-(z)=\mathbf{G}(z),\quad z\in\partial U(z_{1,\pm},\delta),
\end{equation}
in which $\mathbf{G}(z)=\mathbf{G}^{(\pm)}(z)$ are given by \eqref{G+} and \eqref{G-}, respectively. Using the Sokhotskii-Plemelj formula and keeping in mind the clockwise orientations of the boundaries $\partial U(z_{1,\pm},\delta)$, the solution to the above RH problem is explicitly given by
\begin{equation}\label{R(1)(z)}
\mathbf{R}^{(1)}(z)=\left\{\begin{aligned}
&\frac{L_+}{z-\sqrt{\frac{2}{3}}}+\frac{L_-}{z+\sqrt{\frac{2}{3}}},&z&\notin \overline{U(z_{1,+},\delta)}\cup \overline{U(z_{1,-},\delta)},\\
&\frac{L_+}{z-\sqrt{\frac{2}{3}}}+\frac{L_-}{z+\sqrt{\frac{2}{3}}}-\mathbf{G}(z),
&z&\in U(z_{1,+},\delta)\cup U(z_{1,-},\delta),
\end{aligned}\right.
\end{equation}
where $L_{\pm}=\mathrm{Res}\,\left(\mathbf{G}^{(\pm)}(z),z=\pm\sqrt{\frac{2}{3}}\;\right)$ are given by
\begin{align}\label{res1}
L_+&=\mathbf{W}^{(+)}(\textstyle\sqrt{\frac{2}{3}})\begin{pmatrix}0 & -\frac{{\beta} h_0e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2{\beta}}}
{2\cdot 3^{-\frac{1}{4}}e^{\frac{3\pi i}{4}}s_*}\\ -\frac{s_*|x|^{-2{\beta}}}
{2\cdot 3^{-\frac{1}{4}}e^{\frac{3\pi i}{4}}h_0e^{ i\frac{\sqrt{3}}{3}x^2}} & 0 \end{pmatrix}\mathbf{W}^{(+)}(\sqrt{\frac{2}{3}})^{-1},\\
L_-&=\mathbf{W}^{(-)}(-\textstyle\sqrt{\frac{2}{3}})\begin{pmatrix}0 & \frac{{\beta} h_0e^{i\frac{\sqrt{3}}{3}x^2}|x|^{2{\beta}}}
{2\cdot 3^{-\frac{1}{4}}e^{\frac{3\pi i}{4}}s_*}\\ \frac{s_*|x|^{-2{\beta}}}{2\cdot 3^{-\frac{1}{4}}e^{\frac{3\pi i}{4}}h_0e^{i\frac{\sqrt{3}}{3}x^2}} & 0 \end{pmatrix}\mathbf{W}^{(-)}(-\textstyle\sqrt{\frac{2}{3}})^{-1},\label{res2}
\end{align}
with $\mathbf{W}^{(\pm)}(\pm\sqrt{\frac{2}{3}})$ given by \eqref{W+atz+}) and \eqref{W-atz-}, respectively.
Using \eqref{R(1)(z)} to expand $\mathbf{R}^{(1)}(z)$ in \eqref{Rexpanrevisited} into the Taylor series at infinity, we obtain the asymptotics for the coefficients $\mathbf{R}_1$ and $\mathbf{R}_2$ in the expansion \eqref{Rexpan}:
\begin{equation}\label{R1R2}
\mathbf{R}_1=\frac{\mathbf{R}_1^{(1)}}{|x|}+O(|x|^{-2}),\quad \mathbf{R}_2=\frac{\mathbf{R}_2^{(1)}}{|x|}+O(|x|^{-2}),\quad \mathrm{as} \quad x\rightarrow-\infty.
\end{equation}
The coefficients $\mathbf{R}_1^{(1)}=L_++L_-$ and $\mathbf{R}_2^{(1)}=\sqrt{\frac{2}{3}} \left(L_+-L_-\right)$, after a direct computation from \eqref{res1} and \eqref{res2}, are explicitly given by
\begin{equation}\label{R1(1)}
\mathbf{R}_1^{(1)}=\begin{pmatrix}
0 & \left(C_1-C_2+\frac{i}{\sqrt{3}}(C_1+C_2)\right)s_0^{-1}f_{\infty}^{-2}\\
\left(C_2-C_1+\frac{i}{\sqrt{3}}(C_1+C_2)\right)s_0f_{\infty}^{2} & 0
\end{pmatrix},
\end{equation}
and
\begin{equation}\label{R2(1)}
\mathbf{R}_2^{(1)}=\begin{pmatrix}
\frac{2\sqrt{2}}{3}(C_1+C_2) & 0\\
0 & -\frac{2\sqrt{2}}{3}(C_1+C_2)
\end{pmatrix},
\end{equation}
with
\begin{equation}\label{C1C2}
C_1=\frac{{\beta} h_0e^{\frac{i\sqrt{3}}{3}x^2}2^{{\beta}-1}3^{\frac{2{\beta}+1}{4}}}{s_*
e^{\frac{3\pi i}{4}+\frac{2\pi i\alpha}{3}+\frac{\pi i{\beta}}{2}}}|x|^{2{\beta}},\qquad
C_2=\frac{s_*e^{-\frac{3\pi i}{4}+\frac{2\pi i\alpha}{3}+\frac{\pi i{\beta}}{2}}}{h_0e^{\frac{i\sqrt{3}}{3}x^2}
2^{{\beta}+1}3^{\frac{2{\beta}-1}{4}}}|x|^{-2{\beta}}.
\end{equation}
Now, we are ready to derive the asymptotics for $q(x;\alpha,\kappa)$ and $\mathcal{H}(x;\alpha,\kappa)$. Substituting \eqref{R1R2}, \eqref{R1(1)}, \eqref{R2(1)} and \eqref{C1C2} into \eqref{qexpre2} and \eqref{Hexpres} yields
\begin{align}\label{qexpre3}
q(x;\alpha,\kappa)&=-\frac{2}{3}x+2^{\frac{1}{2}}3^{-\frac{1}{4}}e^{-\frac{\pi i}{4}}\Big[{\beta} h_0s_*^{-1}e^{\frac{\sqrt{3}x^2i}{3}-\frac{2\pi i\alpha}{3}-\frac{\pi i{\beta}}{2}+{\beta}\ln\left(2\sqrt{3}x^2\right)}\nonumber\\
&\qquad\qquad\ -h_0^{-1}s_*e^{-\frac{\sqrt{3}x^2i}{3}+\frac{2\pi i\alpha}{3}+\frac{\pi i{\beta}}{2}-{\beta}\ln\left(2\sqrt{3}x^2\right)}\Big]+O(|x|^{-1}),\quad \mathrm{as}\quad x\rightarrow-\infty,
\end{align}
and
\begin{align}\label{Hexpres1}
\mathcal{H}(x;\alpha,\kappa)&=-\frac{8}{27}x^3
+\frac{4}{3}\left(\alpha+\sqrt{3}i{\beta}\right)x
+2^{\frac{1}{2}}3^{-\frac{3}{4}}e^{-\frac{3\pi i}{4}}\Big[{\beta} h_0s_*^{-1}e^{\frac{\sqrt{3}x^2i}{3}-\frac{2\pi i\alpha}{3}-\frac{\pi i{\beta}}{2}+{\beta}\ln\left(2\sqrt{3}x^2\right)}\nonumber\\
&\qquad\qquad\ +h_0^{-1}s_*e^{-\frac{\sqrt{3}x^2i}{3}+\frac{2\pi i\alpha}{3}+\frac{\pi i{\beta}}{2}-{\beta}\ln\left(2\sqrt{3}x^2\right)}\Big]+O(|x|^{-1}),\quad \mathrm{as}\quad x\rightarrow-\infty.
\end{align}
Recalling the definition \eqref{h0} of the Stokes multiplier $h_0$ and using the reflection formula
\begin{equation}\label{refleformu}
\Gamma(1-{\beta})\Gamma({\beta})=\frac{\pi}{\sin(\pi{\beta})},
\end{equation}
we can rewrite \eqref{qexpre3} and \eqref{Hexpres1} in the following symmetric form
\begin{align}\label{qexpre4}
q(x;\alpha,\kappa)=&-\frac{2}{3}x+3^{-\frac{1}{4}}\pi^{-\frac{1}{2}}
e^{\frac{\pi i{\beta}}{2}}{\beta}|s_*|\Big[\Gamma(-{\beta})
e^{\frac{\sqrt{3}x^2i}{3}-\frac{2\pi i\alpha}{3}-\frac{\pi i}{4}-i\arg s_*+{\beta}\ln\left(2\sqrt{3}x^2\right)}\nonumber\\
&- \Gamma({\beta})e^{-\frac{\sqrt{3}x^2i}{3}+\frac{2\pi i\alpha}{3}+\frac{\pi i}{4}+i\arg s_*-{\beta}\ln\left(2\sqrt{3}x^2\right)}\Big]
+O(|x|^{-1}),\quad \mathrm{as}\quad x\rightarrow-\infty,
\end{align}
and
\begin{align}\label{Hexpres2}
\mathcal{H}(x;\alpha,\kappa)=&-\frac{8}{27}x^3
+\frac{4}{3}\left(\alpha+\sqrt{3}i{\beta}\right)x\nonumber\\
&-3^{-\frac{3}{4}}\pi^{-\frac{1}{2}}e^{\frac{\pi i{\beta}}{2}}{\beta}|s_*|\Big[\Gamma(-{\beta})
e^{\frac{\sqrt{3}x^2i}{3}-\frac{2\pi i\alpha}{3}+\frac{\pi i}{4}-i\arg s_*+{\beta}\ln\left(2\sqrt{3}x^2\right)}\nonumber\\
& - \Gamma({\beta})e^{-\frac{\sqrt{3}x^2i}{3}+\frac{2\pi i\alpha}{3}-\frac{\pi i}{4}+i\arg s_*-{\beta}\ln\left(2\sqrt{3}x^2\right)}\Big]
+O(|x|^{-1}),\quad \mathrm{as}\quad x\rightarrow-\infty.
\end{align}
Remember that ${\beta}$ in the case under consideration is purely imaginary (see \eqref{nuimag}), we thus have the following complex conjugate relation
\begin{equation*}
\overline{\Gamma({\beta})}=\Gamma(-{\beta}).
\end{equation*}
Denote by
\begin{equation}\label{a}
a^2=2i{\beta}=-\frac{1}{\pi}\ln(1-|s_*|^2),\quad a>0.
\end{equation}
It is seen from \eqref{nu}, \eqref{refleformu} and \eqref{a} that
\begin{equation}\label{gammanu}
|\Gamma({\beta})|^2=\Gamma({\beta})\Gamma(-{\beta})
=-\frac{\pi}{{\beta}\sin\pi{\beta}}
=-\frac{2\pi i}{{\beta}e^{\pi i\beta}(1-e^{-2\pi i{\beta}})}
=4\pi |s_*|^{-2}a^{-2}e^{-\frac{1}{2}\pi a^2}.
\end{equation}
By substituting \eqref{a}, \eqref{gammanu} into \eqref{qexpre4} and \eqref{Hexpres2}, we arrive at the final formulas \eqref{q1} and \eqref{H1} for the case $0<|s_*|<1$.
Similarly, by using the asymptotic analysis of the RH problem for $\Psi(\xi,x)$
performed in Section \ref{sec:szero}, we obtain in the special case $|s_{*}|=0$, the following asymptotic formulas
\begin{align}\label{q0}
q(x;\alpha,\kappa)&=-\frac{2}{3}x+O(x^{-1}), \\
\mathcal{H}(x;\alpha,\kappa)&=-\frac{8}{27}x^3+\frac{4\alpha}{3}x+O(x^{-1}),\label{H0}
\end{align}
as $x\rightarrow-\infty$. Notice that it follows from \eqref{nu} that ${\beta}=0$ if $|s_*|=0$. Therefore, \eqref{q0} and \eqref{H0} can be respectively regarded as the limits of \eqref{q1} and \eqref{H1} as $b_1\rightarrow0$.
\subsection{Derivations of \eqref{q3} and \eqref{H3}}
Tracing back the above modified RH analysis for the case $\kappa(\kappa-\kappa^*)<0$, it follows from \eqref{U(z)}, \eqref{T(z)} and \eqref{Rtilde} that
\begin{equation}\label{RHanalysis1}
e^{\frac{x^2}{3}\sigma_3}\Phi(z)
e^{-x^2g(z)\sigma_3}=\widetilde{\mathbf{R}}(z)\widetilde{\mathbf{P}}^{(\infty)}(z),
\end{equation}
for large $z$. Accordingly, to compute the coefficients $\Phi_1$ and $\Phi_2$ in expansion \eqref{Asyatinfty1}, we need the asymptotic approximation of $\widetilde{\mathbf{P}}^{(\infty)}(z)$ as $z\rightarrow\infty$. By \eqref{H} and \eqref{Pinftyatinfty}, we have
\begin{equation}\label{Ptildeatinfty}
\widetilde{\mathbf{P}}^{(\infty)}(z)z^{-\alpha\sigma_3}=\mathbf{I}
+\frac{\widetilde{\mathbf{P}}_1^{(\infty)}}{z}
+\frac{\widetilde{\mathbf{P}}_2^{(\infty)}}{z^2}
+O\left(\frac{1}{z^3}\right),\quad \mathrm{as} \quad z\rightarrow\infty,
\end{equation}
where
\begin{equation}\label{Ptilde1Ptilde2}
\widetilde{\mathbf{P}}_1^{(\infty)}=\mathbf{P}_1^{(\infty)}+\widetilde{A}+\widetilde{B},
\qquad \widetilde{\mathbf{P}}_2^{(\infty)}=(\widetilde{A}+\widetilde{B})\mathbf{P}_1^{(\infty)}
+\sqrt{\frac{2}{3}}(\widetilde{A}-\widetilde{B})+\mathbf{P}_2^{(\infty)},
\end{equation}
with $\widetilde{A}$, $\widetilde{B}$ and $\mathbf{P}_1^{(\infty)}$, $\mathbf{P}_2^{(\infty)}$ given by \eqref{reptildeAB} and \eqref{P1P2}, respectively. Using \eqref{Rtildeexp}, \eqref{RHanalysis1} and \eqref{Ptildeatinfty}, we now obtain
\begin{equation}\label{Phi11}
e^{\frac{x^2}{3}\sigma_3}\Phi_1e^{-\frac{x^2}{3}\sigma_3}
=\mathbf{P}_1^{(\infty)}+\widetilde{\mathbf{R}}_1+\widetilde{A}+\widetilde{B},
\end{equation}
and
\begin{equation}\label{Phi22}
e^{\frac{x^2}{3}\sigma_3}\Phi_2e^{-\frac{x^2}{3}\sigma_3}=\frac{4x^2}{27}\sigma_3
+\widetilde{\mathbf{R}}_1(\widetilde{A}+\widetilde{B}+\mathbf{P}_1^{(\infty)})
+\widetilde{\mathbf{R}}_2+(\widetilde{A}+\widetilde{B})\mathbf{P}_1^{(\infty)}
+\sqrt{\frac{2}{3}}(\widetilde{A}-\widetilde{B})+\mathbf{P}_2^{(\infty)}.
\end{equation}
Inserting \eqref{Phi11}, \eqref{Phi22} into the formulas \eqref{solu2} and \eqref{tau1}, we obtain the following expressions for $q(x;\alpha,\kappa)$ and $\mathcal{H}(x;\alpha,\kappa)$
\begin{align}\label{qexpre6}
q(x;\alpha,\kappa)&=-x\,\Big[(\mathbf{P}^{(\infty)}_1)_{12}+\widetilde{A}_{12}
+\widetilde{B}_{12}+(\widetilde{\mathbf{R}}_1)_{12}\Big]\nonumber\\
&\qquad\qquad\qquad\qquad
\times\Big[(\mathbf{P}^{(\infty)}_1)_{21}+\widetilde{A}_{21}+\widetilde{B}_{21}
+(\widetilde{\mathbf{R}}_1)_{21}\Big],\\
\mathcal{H}(x;\alpha,\kappa)&=-x\,\bigg\{\frac{8}{27}x^2+\left[\widetilde {\mathbf{R}}_1(\widetilde{A}+\widetilde{B}+\mathbf{P}_1^{(\infty)})\right]_{11}
-\left[\widetilde {\mathbf{R}}_1(\widetilde{A}+\widetilde{B}+\mathbf{P}_1^{(\infty)})\right]_{22}\nonumber\\
&\qquad\qquad +(\widetilde{\mathbf{R}}_2)_{11}-(\widetilde {\mathbf{R}}_2)_{22}+\left[(\widetilde{A}+\widetilde{B})\mathbf{P}_1^{(\infty)}\right]_{11}
-\left[(\widetilde{A}+\widetilde{B})\mathbf{P}_1^{(\infty)}\right]_{22}\nonumber\\
&\qquad \qquad\qquad +\sqrt{\frac{2}{3}}(\widetilde{A}_{11}-\widetilde{A}_{22}
+\widetilde{B}_{22}-\widetilde{B}_{11})
+(\mathbf{P}_2^{(\infty)})_{11}-(\mathbf{P}_2^{(\infty)})_{22}\bigg\}.\label{Hexpres3}
\end{align}
Next, we compute the asymptotics of $\widetilde{\mathbf{R}}_1$ and $\widetilde{\mathbf{R}}_2$ as $x\rightarrow-\infty$. In view of the error estimate \eqref{Rtildeexpx} and following similar analysis as we have done in Section \ref{Derivation1}, we easily get that
\begin{equation}\label{tildeR1asym}
\widetilde{\mathbf{R}}_1=O(|x|^{-2}),\qquad
\widetilde{\mathbf{R}}_2=O(|x|^{-2}),\quad\mathrm{as}\quad x\rightarrow-\infty.
\end{equation}
Now, substituting \eqref{reptildeAB}, \eqref{P1P2} and \eqref{tildeR1asym} into \eqref{qexpre6} and \eqref{Hexpres3} gives
\begin{align}\label{qexpre7}
q(x;\alpha,\kappa)&=-2x-4\sqrt{\frac{2}{3}}i\cdot\frac{l-\sqrt{\frac{2}{3}}i}{l^2-2}x
+O(x^{-1}),\\
\mathcal{H}(x;\alpha,\kappa)&=-\frac{8}{27}x^3
+\frac{4}{3}\left(\alpha+\sqrt{3}i{\beta}\right)x
+i\frac{4l}{\sqrt{3}}x\cdot\frac{l-\sqrt{\frac{2}{3}}i}{l^2-2}+O(x^{-1}),\label{Hexpres4}
\end{align}
where the error terms are uniform for $x$ bounded away from the zeros of $l^2-2$. From the definition of the function $l$ in \eqref{c}, we may rewrite
\begin{equation*}
\frac{l-\sqrt{\frac {2}{3}}i}{l^2-2}
=\frac{i}{4}\sqrt{\frac{2}{3}}\frac{(2+e^{i\phi})(1+2e^{i\phi})}
{1+e^{i\phi}+e^{2i\phi}}=\frac{i}{4}\sqrt{\frac{2}{3}}
\left(2+\frac{3}{2\cos\phi+1}\right).
\end{equation*}
By inserting the above equation into the expressions \eqref{qexpre7} and \eqref{Hexpres4}, we arrive at the formulas \eqref{q3} and \eqref{H3}.
\subsection{Derivations of \eqref{q2} and \eqref{H2}}
We now consider the case $\kappa=\kappa^*$. By tracing back the transformations \eqref{Ubar} and \eqref{Rbar}, it follows that for large $z$
\begin{equation}\label{largez}
e^{\frac{x^2}{2}\sigma_3}|x|^{\alpha\sigma_3}\Phi(z)|x|^{-\alpha\sigma_3}
e^{-x^2\widehat{g}(z)\sigma_3}=\widehat{\mathbf{R}}(z)\widehat{\mathbf{P}}^{(\infty)}(z).
\end{equation}
To compute the coefficients $\Phi_1$ and $\Phi_2$ in expansion \eqref{Asyatinfty1}, we need to write down the asymptotic approximation of $\widehat{\mathbf{P}}^{(\infty)}(z)$ as $z\rightarrow\infty$. By \eqref{Pbarinfty}, \eqref{H(z)bar} and \eqref{ABexpres}, we get that as $z\rightarrow\infty$
\begin{equation}\label{Pbaratinfty}
\widehat{\mathbf{P}}^{(\infty)}(z)z^{-\alpha\sigma_3}=
\mathbf{I}+\frac{\widehat{\mathbf{P}}_1^{(\infty)}}{z}
+\frac{\widehat{\mathbf{P}}_2^{(\infty)}}{z^2}+O\left(\frac{1}{z^3}\right),
\end{equation}
where
\begin{equation}\label{Pbar1Pbar2}
\widehat{\mathbf{P}}_1^{(\infty)}=\begin{pmatrix} 0 & -\frac{1}{2}c_r \\ -\frac{1}{2}c_l & 0\end{pmatrix},\qquad
\widehat{\mathbf{P}}_2^{(\infty)}=\begin{pmatrix} 1-2\alpha & 0 \\ 0 & 1+2\alpha\end{pmatrix}
\end{equation}
with $c_r$, $c_l$ given by \eqref{cr} and \eqref{cl}, respectively.
From \eqref{largez} and the large-$z$ expansions \eqref{Rbarexp} and \eqref{Pbaratinfty}, we have
\begin{equation}\label{Phi111}
\Phi_1=e^{-\frac{x^2}{2}\sigma_3}|x|^{-\alpha\sigma_3}
\left(\widehat{\mathbf{P}}_1^{(\infty)}
+\widehat{\mathbf{R}}_1\right)|x|^{\alpha\sigma_3}e^{\frac{x^2}{2}\sigma_3}
\end{equation}
and
\begin{equation}\label{Phi222}
\Phi_2=e^{-\frac{x^2}{2}\sigma_3}|x|^{-\alpha\sigma_3}
\left(\widehat{\mathbf{P}}_2^{(\infty)}
+\widehat{\mathbf{P}}_1^{(\infty)}\widehat{\mathbf{R}}_1
+\widehat{\mathbf{R}}_2\right)|x|^{\alpha\sigma_3}e^{\frac{x^2}{2}\sigma_3}.
\end{equation}
Now, recalling \eqref{solu2} and \eqref{tau1}, we obtain from \eqref{Pbar1Pbar2}, \eqref{Phi111} and \eqref{Phi222} that
\begin{equation}\label{qexpre5}
q(x;\alpha,\kappa)=-x\left[-\frac{c_r}{2}+(\widehat{\mathbf{R}}_1)_{12}\right]
\left[-\frac{c_l}{2}+(\widehat{\mathbf{R}}_1)_{21}\right]
\end{equation}
and
\begin{equation}\label{Hexpres5}
\mathcal{H}(x;\alpha,\kappa)=-x\bigg\{-4\alpha-\frac{1}{2}
\left[c_r(\widehat{\mathbf{R}}_1)_{21}-c_l(\widehat{\mathbf{R}}_1)_{12}\right]
+(\widehat{\mathbf{R}}_2)_{11}-(\widehat{\mathbf{R}}_2)_{22}\bigg\}.
\end{equation}
In view of the estimation \eqref{Rbarexpans}, we easily obtain the following asymptotics for $\widehat{\mathbf{R}}_1$ and $\widehat{\mathbf{R}}_2$ as $x\rightarrow-\infty$
\begin{equation}\label{Rbar1Rbar2asymp}
\widehat{\mathbf{R}}_1=O(|x|^{-2}),\quad \widehat{\mathbf{R}}_2=O(|x|^{-2}).
\end{equation}
Substituting the asymptotics \eqref{Rbar1Rbar2asymp} into the expressions \eqref{qexpre5} and \eqref{Hexpres5} and using the fact that $c_rc_l=8$, we obtain \eqref{q2} and \eqref{H2}.
\subsection{Derivation of \eqref{Hasymp+infty} }
It is seen from \eqref{qAsy} and \eqref{eq:dH} that
\begin{equation}\label{eq:dHAsy}
\frac{d}{dx}\mathcal{H}(x;\alpha,\kappa)=\kappa\, 2^{\alpha+\frac{1}{2}}x^{2\alpha}e^{-x^2}\left(1+O\left(x^{-2}\right)\right), \quad \mathrm{as}~~x\to+\infty.
\end{equation}
Moreover, it follows from \eqref{qAsy} and \eqref{eq:Hq} that
\begin{equation}\label{eq:Hlimit} \mathcal{H}(x;\alpha,\kappa)\to0, \quad \mathrm{as}~~x\to+\infty. \end{equation}
Thus, integrating both sides of \eqref{eq:dHAsy} and using the boundary condition
\eqref{eq:Hlimit}, we obtain \eqref{Hasymp+infty}.
\subsection{Derivations of \eqref{integral-1} and \eqref{integral-2}}\label{proof2}
Recalling from \eqref{Fequation} and using the fact that $F(x)\to \mathbf{I}$ as $x\to+\infty$, as can be seen from the asymptotic analysis of $\Psi(\xi,x)$ as $x\to+\infty$ performed in \cite{IK}, we have
\begin{equation}\label{eq:qInt}
F(x):=\Psi^{(0)}(0,x)=\exp\left(\int_{+\infty}^xq(t)dt\sigma_3\right),
\end{equation}
with $\Psi^{(0)}(\xi,x)$ given in \eqref{Asyatzero}.
Hence,
the evaluation of the total integral of $q(x)$ is boiled down to computing the asymptotic of $F(x)$ as $x\to-\infty$.
In the remaining part of this section, we derive the asymptotics of $F(x)$ as $x\to-\infty$ and prove Theorem \ref{thm3} based on the asymptotic analysis of the RH problem for $\Psi(\xi,x)$
performed in Sections \ref{Asymptotic-infty1} and \ref{Asymptotic-infty3}.
\subsubsection{Derivation of \eqref{integral-1}}
In the case $\kappa(\kappa-\kappa^{*})<0$, by inverting the transformations \eqref{rescaling}, \eqref{U(z)}, \eqref{T(z)} and \eqref{R}, we obtain that for $z$ small and $\arg z\in(\frac{\pi}{4},\frac{\pi}{2})$,
\begin{align}\label{Psiexpress2}
\Psi(|x|^{\frac{1}{2}}z,x)
&=e^{-\frac{x^2}{3}\sigma_3}|x|^{\frac{\alpha}{2}\sigma_3}
\mathbf{R}(z)\mathbf{P}^{(0)}(z)\begin{pmatrix}1 & 0\\ \frac{\overline{s}_*e^{-2\pi i\alpha}e^{-2x^2g(z)}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix}e^{x^2g(z)\sigma_3}\nonumber\\
&=e^{-\frac{x^2}{3}\sigma_3}|x|^{\frac{\alpha}{2}\sigma_3}
\mathbf{R}(z)\mathbf{E}^{(0)}(z)\Phi^{(\mathrm{Bes})}(x^2\varphi_4(z))\nonumber\\
&\qquad\qquad\qquad \times\begin{pmatrix}1 & 0\\ -e^{-2\pi i\alpha} & 1 \end{pmatrix}\left[s_1(e^{-2\pi i\alpha}+s_*)\right]^{-\frac{\sigma_3}{2}}\begin{pmatrix}1 & 0\\ \frac{\bar{s}_*e^{-2\pi i\alpha}}
{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix}.
\end{align}
It is readily seen from \eqref{Asyatzero} and \eqref{Psiexpress2} that
\begin{align}\label{Fexpress6}
F(x)&=e^{-\frac{x^2}{3}\sigma_3}|x|^{\frac{\alpha}{2}\sigma_3}
\mathbf{R}(0)\mathbf{E}^{(0)}(0)\lim\limits_{z\to0}\Phi^{(\mathrm{Bes})}(x^2\varphi_4(z))
\begin{pmatrix}1 & 0\\ -e^{-2\pi i\alpha} & 1 \end{pmatrix}\nonumber\\
&\qquad\qquad\qquad \times\left[s_1(e^{-2\pi i\alpha}+s_*)\right]^{-\frac{\sigma_3}{2}}\begin{pmatrix}1 & 0\\ \frac{\bar{s}_*e^{-2\pi i\alpha}}{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix}S_1^{-1}E_0^{-1}z^{-\alpha\sigma_3}|x|^{-\frac{\alpha}{2}\sigma_3}.
\end{align}
By \eqref{E0}, \eqref{varphi4beha} and \eqref{BesParaExpand}, we have
\begin{align}\label{limit3}
&\lim\limits_{z\to0}\Phi^{(\mathrm{Bes})}(x^2\varphi_4(z))\begin{pmatrix}1 & 0\\ -e^{-2\pi i\alpha} & 1 \end{pmatrix}[s_1(e^{-2\pi i\alpha}+s_*)]^{-\frac{\sigma_3}{2}}\begin{pmatrix}1 & 0\\ \frac{\bar{s}_*e^{-2\pi i\alpha}}{s_1(e^{-2\pi i\alpha}+s_*)} & 1 \end{pmatrix}S_1^{-1}E_0^{-1}z^{-\alpha\sigma_3}\nonumber\\
&\qquad=C_{\alpha}^{\sigma_3}\begin{pmatrix}0 & -1\\ 1 & 0 \end{pmatrix}2^{\frac{3}{2}\alpha\sigma_3}3^{-\frac{3}{2}\alpha\sigma_3}
s_1^{-\frac{\sigma_3}{2}}(e^{-2\pi i\alpha}+s_*)^{\frac{\sigma_3}{2}}(1+e^{-2\pi i\alpha})^{-\sigma_3}|x|^{2\alpha\sigma_3},
\end{align}
where the constant $C_{\alpha}$ is given by \eqref{Calpha}. Recalling the definition \eqref{E(0)} of $\mathbf{E}^{(0)}(z)$, it follows from \eqref{Pinfty}, \eqref{finfty} and \eqref{nu} that
\begin{align}\label{E(0)at0}
\mathbf{E}^{(0)}(0)&=\frac{1}{2}s_0^{-\frac{\sigma_3}{2}}2^{\frac{\alpha}{2}\sigma_3}
3^{-\frac{\alpha}{2}\sigma_3}e^{-\frac{\pi i{\beta}}{3}\sigma_3}
\begin{pmatrix}1 & -1\\ 1 & 1 \end{pmatrix}e^{-\frac{3\pi i}{4}\sigma_3}
\begin{pmatrix}1 & i\\ i & 1 \end{pmatrix}.
\end{align}
Substituting \eqref{limit3} and \eqref{E(0)at0} into \eqref{Fexpress6} yields
\begin{align}\label{Fexpress++}
e^{\frac{x^2}{3}\sigma_3}|x|^{-\frac{\alpha}{2}\sigma_3}F(x)|x|^{-\frac{3\alpha}{2}\sigma_3}
=\mathbf{R}(0)2^{(\alpha-\frac{1}{2})\sigma_3}3^{-2\alpha\sigma_3}\pi^{-\frac{\sigma_3}{2}}
\Gamma(\textstyle\frac{1}{2}-\alpha)^{\sigma_3}[(1-s_*)e^{\pi i\alpha}]^{-\sigma_3}e^{-\frac{4}{3}\pi i{\beta}\sigma_3}.
\end{align}
We choose some constants $c<0$ and $d>0$ such that all real poles of $q(x;\alpha,\kappa)$ lie in the $x$-interval $(c,d)$. Using the estimation \eqref{Restimation}, the expression \eqref{eq:qInt} and letting $x\to-\infty$ in \eqref{Fexpress++}, we obtain
\begin{align}\label{explimit1}
&\exp\Bigg\{\int^{d}_{+\infty}q(t)dt+\int_{\Upsilon}q(t)dt
+\int^{-\infty}_{c}\left(q(t)+\frac{2t}{3}-\frac{2\alpha}{t}\right)dt
+\frac{c^2}{3}-2\alpha\ln|c|\Bigg\}\nonumber\\
&\qquad\qquad=2^{\alpha-\frac{1}{2}}
3^{-2\alpha}\pi^{-\frac{1}{2}}\Gamma\left(\textstyle\frac{1}{2}-\alpha\right)[(1-s_*)
e^{\pi i\alpha}]^{-1}e^{-\frac{4}{3}\pi i{\beta}}.
\end{align}
where $\Upsilon$ is any contour in the complex plane from $d$ to $c$ that avoids the real poles of $q(x;\alpha,\kappa)$. Indeed, we may take for $\Upsilon$ a path along the real axis with infinitesimal semicircular indentations centered at each real pole of $q(x;\alpha,\kappa)$ in the lower half complex plane. Using the fact that all poles of $q(x;\alpha,\kappa)$ are simple with residues $\pm1$, we have
\begin{equation}\label{qintegral1}
\int_{\Upsilon}q(t)dt=\mathrm{P.V.}\int^{c}_{d}q(t)dt+\pi i(N_+-N_-),
\end{equation}
where $\mathrm{P.V.}$ denotes the Cauchy principal value and $N_{\pm}$ are the numbers of real poles of $q(x;\alpha,\kappa)$ of residue $\pm1$, respectively.
By inserting \eqref{qintegral1} into \eqref{explimit1} and using the definitions of $\kappa$, ${\beta}$
as given by \eqref{kapparep}, \eqref{nu}, respectively, we arrive at the total integral \eqref{integral-1}.
\subsubsection{Derivation of \eqref{integral-2}}
Now, we concentrate on the case $\kappa=\kappa^{*}$. Tracing back the transformations \eqref{rescaling}, \eqref{Ubar} and \eqref{Rbar} gives
\begin{align}\label{Psiexpress3}
\Psi(|x|^{\frac{1}{2}}z,x)
&=e^{-\frac{x^2}{2}\sigma_3}|x|^{-\frac{\alpha}{2}\sigma_3}
\widehat{\mathbf{R}}(z)\widehat{\mathbf{P}}^{(0)}(z)|x|^{\alpha\sigma_3}
e^{x^2\widehat{g}(z)\sigma_3}.
\end{align}
It follows from \eqref{Asyatzero}, \eqref{Fequation}, \eqref{Pbar0} and \eqref{Psiexpress3} that
\begin{align}\label{Fexpress7}
F(x)=e^{-\frac{x^2}{2}\sigma_3}|x|^{-\frac{\alpha}{2}\sigma_3}
&\widehat{\mathbf{R}}(0)\widehat{\mathbf{H}}(0)\lim\limits_{z\to0}
\begin{pmatrix}1 & 0\\ r(z) & 1 \end{pmatrix}
z^{-\alpha\sigma_3}(z+\sqrt{2})^{(\alpha+\frac{1}{2})\sigma_3}\nonumber\\
&\qquad\qquad \times (z-\sqrt{2})^{(\alpha-\frac{1}{2})\sigma_3}|x|^{\alpha\sigma_3}
e^{x^2\widehat{g}(z)\sigma_3}S_1^{-1}E_0^{-1}z^{-\alpha\sigma_3}
|x|^{-\frac{\alpha}{2}\sigma_3},
\end{align}
for $z\to0$ with $\arg{z}\in(\frac{\pi}{4},\frac{\pi}{2})$.
Notice that by \eqref{H(z)bar}, \eqref{ABexpres} and \eqref{rexpre},
\begin{equation}\label{H(z)barat0}
\widehat{\mathbf{H}}(0)=(2\pi)^{-\frac{\sigma_3}{2}}e^{\pi i(\alpha-\frac{1}{2})\sigma_3}\Gamma(\alpha+\textstyle\frac{1}{2})^{\sigma_3}
\begin{pmatrix}0 & s_1 \\ -s_1^{-1} & 0 \end{pmatrix}
\end{equation}
and
\begin{align}\label{limit4}
&e^{-x^2\widehat{g}(z)\sigma_3}|x|^{-\alpha\sigma_3}(z+\sqrt{2})^{-(\alpha+\frac{1}{2})\sigma_3}
(z-\sqrt{2})^{-(\alpha-\frac{1}{2})\sigma_3}
z^{\alpha\sigma_3}\begin{pmatrix}1 & 0\\ r(z) & 1 \end{pmatrix}\nonumber\\
&\qquad\qquad\qquad\times z^{-\alpha\sigma_3}(z+\sqrt{2})^{(\alpha+\frac{1}{2})\sigma_3}
(z-\sqrt{2})^{(\alpha-\frac{1}{2})\sigma_3}|x|^{\alpha\sigma_3}
e^{x^2\widehat{g}(z)\sigma_3}=\begin{pmatrix}1 & 0\\ -\frac{s_0e^{2\pi i\alpha}}{1+e^{2\pi i\alpha}} & 1 \end{pmatrix}.
\end{align}
Substituting \eqref{limit4} and \eqref{H(z)barat0} into \eqref{Fexpress7} and combining with the facts
\begin{equation*}
\begin{pmatrix}0 & s_1 \\ -s_1^{-1} & 0 \end{pmatrix}=\begin{pmatrix}1 & 0 \\ -s_1^{-1} & 1 \end{pmatrix}\begin{pmatrix}1 & s_1 \\ 0 & 1 \end{pmatrix}\begin{pmatrix}1 & 0 \\ -s_1^{-1} & 1 \end{pmatrix}
\end{equation*}
and
\begin{equation*}
E_0=\begin{pmatrix}1 & 0\\ \frac{s_0e^{2\pi i\alpha}}{1+e^{2\pi i\alpha}} & 1 \end{pmatrix}=\begin{pmatrix}1 & 0 \\ -s_1^{-1} & 1 \end{pmatrix},
\end{equation*}
which follows from \eqref{E0} and \eqref{sstarvalue}, we have
\begin{align}\label{Fexpress+++}
e^{\frac{x^2}{2}\sigma_3}&|x|^{\frac{\alpha}{2}\sigma_3}F(x)|x|^{\frac{3\alpha}{2}\sigma_3}
e^{\frac{x^2}{2}\sigma_3}=
\widehat{\mathbf{R}}(0)(2\pi)^{-\frac{\sigma_3}{2}}e^{\pi i(\alpha-\frac{1}{2})\sigma_3}
\Gamma(\alpha+\textstyle\frac{1}{2})^{\sigma_3}
2^{-\alpha\sigma_3}e^{\pi i(\frac{1}{2}-\alpha)\sigma_3}.
\end{align}
Similarly, take $c<0$ and $d>0$ such that all real poles of $q(x;\alpha,\kappa)$ lie in the interval $(c,d)$.
Using the estimation \eqref{Rbarexpans} and letting $x\to-\infty$ in \eqref{Fexpress+++}, we
obtain \eqref{integral-2}.
\section{Proof of Theorem \ref{thm:IntRep}}\label{sec:proof of det}
Denote
\begin{align}\label{def:f}
\mathbf{f}(\lambda)&=\begin{pmatrix}f_1(\lambda)\\ f_2(\lambda) \end{pmatrix}=\begin{pmatrix}D_{\nu}\left(\sqrt{2}(\lambda+x)\right) \\
-D_{\nu-1}\left(\sqrt{2}(\lambda+x)\right) \end{pmatrix}, \\
\mathbf{h}(\lambda)&=\begin{pmatrix}h_1(\lambda)\\ h_2(\lambda) \end{pmatrix}=\gamma\begin{pmatrix} D_{\nu-1}\left(\sqrt{2}(\lambda+x)\right) \\
D_{\nu}\left(\sqrt{2}(\lambda+x)\right) \end{pmatrix}, \label{def:h}
\end{align}
where $D_{\nu}$ is the parabolic cylinder function.
Then, the parabolic cylinder kernel \eqref{eq:PCKernel} can be written as
\begin{equation}\label{eq:PCKernel1}
\gamma K_{\nu,x}(\lambda,\mu)=\frac{\mathbf{ f}(\lambda)^{T}\mathbf{ h}(\mu)}{\lambda-\mu}.\end{equation}
For general parameter $\nu\in \mathbb{R}$, the logarithmic derivative of $\gamma K_{\nu,x}(\lambda,\mu)$ can be expressed in terms of the solution of the following RH problem
for $Y(\lambda):=Y(\lambda, x)$.
\subsection*{RH problem for $Y(\lambda)$}
\begin{itemize}
\item[\rm (1)]
$Y(\lambda)$ is analytic in
$\mathbb{C}\setminus [0,\infty)$.
\item[\rm (2)] $Y(\lambda)$ satisfies the jump condition
\begin{equation}\label{eq:YJump}
Y_+(\lambda)=Y_-(\lambda) \left(\mathbf{I}-2\pi i \mathbf{ f}(\lambda) \mathbf{ h}(\lambda)^{T}\right),
\end{equation}
where $\mathbf{ f}$ and $\mathbf{ h}$ are defined by \eqref{def:f} and \eqref{def:h}, respectively.
\item[\rm (3)] The behavior of $Y(\lambda)$ at infinity is
\begin{equation}\label{eq:YInfinity}Y(\lambda)=\mathbf{I}+\frac {Y_{1}}{\lambda}+O\left (\frac 1 {\lambda^2}\right).\end{equation}
\item[\rm (4)] The behavior of $Y(\lambda)$ at the origin is
\begin{equation}\label{eq:Y0}Y(\lambda)=O\left (\ln\lambda\right).\end{equation}
\end{itemize}
Actually, the solution to the RH problem for $Y(\lambda)$ can be expressed as follows \cite{IIKS}
\begin{equation}\label{eq:Ysolution}Y(\lambda)=\mathbf{I}-\int_{0}^{+\infty} \frac{\mathbf{ V}(\mu)\mathbf{ h}(\mu)^{T}}{\mu-\lambda}d\mu,\end{equation}
where
\begin{equation}\label{eq:YFH}
\mathbf{ V}(\mu)=(\mathbf{I}-\gamma K_{\nu,x})^{-1} \mathbf{ f}(\mu) .\end{equation}
The equation \eqref{eq:PCKernel1} implies
\begin{equation}\label{eq:DF1}
\frac{d}{dx} \ln \det(\mathbf{I}-\gamma K_{\nu,x})=-\mathrm{ tr} \left(\left(\mathbf{I}-\gamma K_{\nu,x}\right)^{-1} \gamma \frac{d}{dx}
K_{\nu,x}\right).\end{equation}
Using the recurrence relations satisfied by
the parabolic cylinder function (cf. \cite[Equation (12.8)]{NIST})
\begin{equation}\label{eq:PCRelation1}
D_{\nu}'(\lambda)=\frac{\lambda}{2}D_{\nu}(\lambda)-D_{\nu+1}(\lambda),\end{equation}
and
\begin{equation}\label{eq:PCRelation2}
D_{\nu}'(\lambda)=-\frac{\lambda}{2}D_{\nu}(\lambda)+\nu D_{\nu-1}(\lambda), \end{equation}
we find after some direct calculation that
\begin{equation}\label{eq:dK}
\gamma\frac{d}{dx} K_{\nu,x}(\lambda,\mu) =-\mathbf{ h}(\mu)^{T}\sigma_3 \mathbf{ f}(\lambda)=-\left(f_1(\lambda)h_1(\mu)
-f_2(\lambda)h_2(\mu)\right).\end{equation}
Substituting \eqref{eq:dK} into \eqref{eq:DF1} and applying \eqref{eq:Ysolution}-\eqref{eq:YFH}, we arrive at
the differential identity
\begin{equation}\label{eq:FY}
\frac{d}{dx} \ln \det(\mathbf{I}-\gamma K_{\nu,x})=(Y_1)_{11}-(Y_1)_{22}=2(Y_1)_{11},
\end{equation}
with $Y_1$ given in \eqref{eq:YInfinity}.
To proceed, we define
\begin{equation}\label{eq:PCM0}
P(z)=\begin{pmatrix}e^{i\frac{\pi}{2}\nu} D_{-\nu}(e^{i\frac{\pi}{2}}z)& D_{\nu-1}(z) \\
\nu e^{i\frac{\pi}{2}(\nu+1)} D_{-\nu-1}(e^{i\frac{\pi}{2}}z) & D_{\nu}(z) \end{pmatrix}
\end{equation}
for $\arg z\in(-\frac{1}{4}\pi,0)$.
In view of the asymptotic behavior \eqref{eq:DAsy} for $D_{\nu}$, we see that $P(z)$ satisfies the following asymptotic behavior at infinity
\begin{equation}\label{eq:PAsy}
P(z)=\left(\mathbf{I}+\frac{P_1}{z}+O\left(\frac 1{z^{2}}\right)
\right)z^{-\nu \sigma_3}\exp\left(\frac{1}{4}z^2\sigma_3\right),
\end{equation}
where the diagonal entries of $P_1$ are zero. Define
\begin{equation}\label{eq:Psi}
\widehat{\Psi}(\lambda)=e^{-\frac{1}{2}x^2\sigma_3}2^{\frac{\nu}{2}\sigma_3} Y(\lambda)^{-T} P\left(\sqrt{2}(\lambda+x)\right)\prod_{i=0}^{k-1}J_i,
\end{equation}
for $\arg\lambda\in (\frac{k-1}{2}\pi,\frac{k}{2}\pi)$, $k=1,2,3$.
Here, the constant matrices
\begin{equation}\label{eq:J_k}
J_0= \left(
\begin{array}{cc}
1& 0\\
j_0 & 1 \\
\end{array}
\right), \quad J_1= \left(
\begin{array}{cc}
1& j_1\\
0 & 1 \\
\end{array}
\right), \quad
J_2= \left(
\begin{array}{cc}
1& 0\\
-j_0e^{-2\pi i\nu} & 1 \\
\end{array}
\right), \end{equation}
with $j_0=-i\frac{\sqrt{2\pi}}{\Gamma(\nu)} $ and $j_1=-\frac{\sqrt{2\pi}}{\Gamma(1-\nu)}e^{i\pi \nu}$.
We also denote
\begin{equation}\label{eq:J3}
J_3= (J_0J_1J_2)^{-1}=\left(
\begin{array}{cc}
1& -j_1e^{2\pi i\nu}\\
0 & 1 \\
\end{array}
\right)e^{2\pi i\nu\sigma_3}.
\end{equation}
Then, it is direct to see that $\widehat\Psi(\lambda)$ satisfies the following RH problem.
\subsection*{RH problem for $\widehat{\Psi}(\lambda)$}
\begin{itemize}
\item[\rm (1)]
$\widehat\Psi(\lambda)$ is analytic in
$\mathbb{C}\setminus \Xi_k$,
where
$\Xi_k=e^{i\frac{k}{2}\pi}\mathbb{R}_+$, $k=0,1,2,3$.
\item[\rm (2)] $\widehat\Psi(\lambda)$ satisfies the jump conditions
\begin{equation}\label{eq:PsiJump1}\widehat\Psi_+(\lambda)=\widehat\Psi_-(\lambda)
\left(\begin{array}{cc}1& 0\\ 2\pi i \gamma-i\frac{\sqrt{2\pi}}{\Gamma(\nu)} & 1 \\
\end{array}\right), \quad \lambda\in\Xi_0,\end{equation}
and
\begin{equation}\label{eq:PsiJump2}\widehat\Psi_+(\lambda)=\widehat\Psi_-(\lambda)
J_k, \quad \lambda\in\Xi_k,\end{equation}
with $J_k$, $ k=1,2,3$ given in \eqref{eq:J_k} and \eqref{eq:J3}.
\item[\rm (3)] $\widehat\Psi(\lambda)$ has the asymptotic behavior as $\lambda\to\infty$
\begin{equation}\label{eq:PsiInfinity}\widehat\Psi(\lambda)=\left(\mathbf{I}+\frac {\widehat\Psi_{1}}{\lambda}+O\left (\frac 1 {\lambda^2}\right)\right)\lambda^{-\nu \sigma_3}\exp\left((\lambda^2/2+x\lambda)\sigma_3\right).\end{equation}
\item[\rm (4)] $\widehat\Psi(\lambda)$ possesses the behavior near the origin
\begin{equation}\label{eq:Psi0}\widehat\Psi(\lambda)=O\left (\ln\lambda\right).\end{equation}
\end{itemize}
It comes out that the RH problem for $\widehat\Psi(\lambda)$ is the same as the RH problem for the Jimbo-Miwa Lax pair of the PIV equation with the parameters $\theta_0=0$ and $\theta_{\infty}=\nu$; cf. \cite[(C.30)-(C.31)]{JM}, see also
\cite[Chapter 5.1]{FIKN}.
It is shown in \cite[(C.34)-(C.37)]{JM} that
\begin{equation}\label{eq:PsiExpand1}
\sigma_{\nu}(x)=-2(\widehat\Psi_1)_{11}-2\nu x
\end{equation}
satisfies the equation \eqref{eq:sPIV}.
Using \eqref{eq:PAsy} and \eqref{eq:Psi}, we find
\begin{equation}\label{eq:PsiExpand0}
2(Y_1)_{11} =-2(\widehat\Psi_1)_{11}-2\nu x
\end{equation}
Thus, we have
\begin{equation}\label{eq:PsiExpand}
2(Y_1)_{11} =\sigma_{\nu}(x).
\end{equation}
Recalling \eqref{eq:FY}, we have derived \eqref{eq:IntRep}.
Next, we prove \eqref{eq:SigmaAsy}.
It is seen from \eqref{eq:DAsy} that the parabolic cylinder function $D_{\nu}(x)$ decay exponentially fast as $x$ tends to positive infinity. We get from the series expansion of
the Fredholm determinant that
\begin{equation}\label{eq:DExpand}
\det(\mathbf{I}-\gamma K_{\nu,x})\sim 1-\gamma\int_0^{+\infty}K_{\nu,x}(\lambda,\lambda)d\lambda. \end{equation}
Thus, we have
\begin{equation}\label{eq:FAsy1}
\frac{d}{dx} \ln \det(\mathbf{I}-\gamma K_{\nu,x})\sim \gamma K_{\nu,x}(0,0),~~~x\to +\infty.
\end{equation}
The asymptotic \eqref{eq:SigmaAsy} then follows from \eqref{eq:DAsy}, \eqref{eq:PCKernel} and \eqref{eq:FAsy1}.
We complete the proof of Theorem \ref{thm:IntRep}.
\section*{Acknowledgements}
The authors are grateful to the editor and the referee for their valuable suggestions and comments.
The work of Shuai-Xia Xu was supported in part by the National Natural Science Foundation of China under grant numbers 11571376 and 11971492, and by Guangdong Basic and Applied Basic Research Foundation (Grant No. 2022B1515020063). Yu-Qiu Zhao was supported in part by the National Natural Science Foundation of China under grant numbers 11571375 and 11971489.
\begin{appendices}
\section{Local parametrix models}
\subsection{Airy parametrix}\label{AP}
Introduce
\begin{equation}
\Phi^{(\mathrm{Ai})}(\zeta)=\mathbf{N}\left\{
\begin{aligned}
&\begin{pmatrix}
\mathrm{Ai}(\zeta) & \mathrm{Ai}(w^{2}\zeta) \\
\mathrm{Ai}^{\prime}(\zeta) & w^{2} \mathrm{Ai}^{\prime}(w^{2}\zeta)
\end{pmatrix} e^{-i \frac{\pi}{6} \sigma_{3}}, &\zeta&\in \mathrm{I},\\
&\begin{pmatrix}
\mathrm{Ai}(\zeta) & \mathrm{Ai}(w^{2}\zeta) \\
\mathrm{Ai}^{\prime}(\zeta) & w^{2} \mathrm{Ai}^{\prime}(w^{2}\zeta)
\end{pmatrix} e^{-i \frac{\pi}{6} \sigma_{3}}\begin{pmatrix}
1 & 0 \\
-1 & 1
\end{pmatrix},&\zeta&\in \mathrm{II},\\
&\begin{pmatrix}
\mathrm{Ai}(\zeta) & -w^{2}\mathrm{Ai}(w\zeta) \\
\mathrm{Ai}^{\prime}(\zeta) & -\mathrm{Ai}^{\prime}(w\zeta)
\end{pmatrix} e^{-i \frac{\pi}{6} \sigma_{3}}\begin{pmatrix}
1 & 0 \\
1 & 1
\end{pmatrix},&\zeta&\in \mathrm{III},\\
&\begin{pmatrix}
\mathrm{Ai}(\zeta) & -w^{2} \mathrm{Ai}(w\zeta) \\
\mathrm{Ai}^{\prime}(\zeta) & -\mathrm{Ai}^{\prime}(w\zeta)
\end{pmatrix} e^{-i \frac{\pi}{6} \sigma_{3}},&\zeta&\in \mathrm{IV},
\end{aligned}
\right.
\end{equation}
where $\mathrm{Ai}(\zeta)$ denotes the Airy function (cf. \cite[Chapter 9]{NIST}), $w=e^{2\pi i/3}$ and
$$
\mathbf{N}=\sqrt{2\pi}\,e^{\frac{1}{6}\pi i}\begin{pmatrix}1 & 0\\ 0 & -i \end{pmatrix}.
$$
The regions I-IV are illustrated in Figure \ref{Airy}. Then, $\Phi^{(\mathrm{Ai})}(\zeta)$ solves the following RH problem (cf. \cite{Deft}).
\subsection*{RH problem for $\Phi^{(\mathrm{Ai})}(\zeta)$}
\begin{description}
\item{(1)} $\Phi^\mathrm{(Ai)}(\zeta)$ is analytic for $\zeta\in \mathbb{C}\setminus \bigcup^4_{k=1}\Sigma_{k}$, where $\Sigma_1=\mathbb{R}_+$, $\Sigma_2=e^{\frac{2\pi i}{3}}\mathbb{R}_+$, $\Sigma_3=\mathbb{R}_-$ and $\Sigma_4=e^{-\frac{2\pi i}{3}}\mathbb{R}_+$ with orientations indicated in Figure \ref{Airy}.
\item{(2)} We have the jump relation
\begin{equation*}
\Phi^\mathrm{(Ai)}_+(\zeta)=\Phi^\mathrm{(Ai)}_-(\zeta)\left\{
\begin{aligned}
&\begin{pmatrix}1 & 1 \\ 0 & 1 \end{pmatrix}, && \zeta\in\Sigma_{1},\\
&\begin{pmatrix}1 & 0 \\ 1 & 1 \end{pmatrix}, && \zeta\in\Sigma_{2}\cup\Sigma_4,\\
&\begin{pmatrix}0 & 1 \\ -1 & 0 \end{pmatrix}, && \zeta\in\Sigma_{3}.
\end{aligned}\right.
\end{equation*}
\item{(3)} $\Phi^\mathrm{(Ai)}(\zeta)$ possesses the following asymptotic behavior as $\zeta\to \infty$
\begin{equation}\label{AiryAsyatinfty}
\Phi^{(\mathrm{Ai})}(\zeta)=\zeta^{-\frac{\sigma_{3}}{4} }\frac{1}{\sqrt{2}} \begin{pmatrix}1 & i\\ i &1\end{pmatrix}\left(\mathbf{I}+O\left(\zeta^{-\frac{3}{2}}\right)\right) e^{-\frac{2}{3} \zeta^{\frac{3}{2}} \sigma_{3}}.
\end{equation}
\end{description}
\begin{figure}[t]
\centering
\includegraphics[width=8cm,height=6cm]{Airy}\\
\caption{The jump contour and regions for $\Phi^{(\mathrm{Ai})}$}\label{Airy}
\end{figure}
\subsection{Parabolic cylinder parametrix}\label{PCP}
Let ${\beta}$ be a fixed real or complex number. Define
\begin{equation}\label{eq:Dpra}
\mathbf{D}(\zeta)=2^{-\frac{\sigma_{3}}{2}}\begin{pmatrix}
D_{-{\beta}-1}(i\zeta) & D_{{\beta}}(\zeta) \\iD_{-{\beta}-1}'(i\zeta) & D_{{\beta}}'(\zeta)\end{pmatrix}\begin{pmatrix}
e^{i \frac{\pi}{2}({\beta}+1)} & 0 \\ 0 & 1 \end{pmatrix},
\end{equation}
where $D_{{\beta}}(\zeta)$ denotes the standard parabolic cylinder function (cf. \cite[Chapter 12]{NIST}).
Denote
$$
H_{0}=\begin{pmatrix}1 & 0 \\ h_{0} & 1\end{pmatrix}, \quad
H_{1}=\begin{pmatrix}1 & h_{1} \\ 0 & 1\end{pmatrix}, \quad
H_{k+2}=e^{i \pi\left({\beta}+\frac{1}{2}\right) \sigma_{3}} H_{k} e^{-i \pi\left({\beta}+\frac{1}{2}\right) \sigma_{3}}, \ k=0,1,
$$
where
\begin{equation}\label{h0}
h_{0}=-i \frac{\sqrt{2 \pi}}{\Gamma({\beta}+1)}, \quad h_{1}=\frac{\sqrt{2 \pi}}{\Gamma(-{\beta})} e^{i \pi {\beta}}, \quad 1+h_{0} h_{1}=e^{2 \pi i {\beta}}.
\end{equation}
We consider the function
$$
\Phi^{(\mathrm{PC})}(\zeta)=\left\{\begin{aligned}
&\mathbf{ D}(\zeta), \quad & \arg \zeta \in&\left(-\frac{\pi}{4}, 0\right), \\
&\mathbf{ D}(\zeta)H_0,\quad & \arg \zeta \in&\left(0, \frac{\pi}{2}\right), \\
&\mathbf{ D}(\zeta)H_1,\quad & \arg \zeta \in&\left(\frac{\pi}{2}, \pi\right), \\
&\mathbf{ D}(\zeta)H_2,\quad & \arg \zeta \in&\left(\pi, \frac{3 \pi}{2}\right), \\
&\mathbf{ D}(\zeta)H_3,\quad & \arg \zeta \in&\left(\frac{3 \pi}{2}, \frac{7 \pi}{4}\right).
\end{aligned}\right.
$$
It is direct to check that $\Phi^{(\mathrm{PC})}(\zeta)$ solves the following RH problem (cf. \cite{BI,FIKN}).
\subsection*{RH problem for $\Phi^{(\mathrm{PC})}(\zeta)$}
\begin{description}
\item{(1)} $\Phi^\mathrm{(PC)}(\zeta)$ is analytic for all $\zeta\in \mathbb{C}\setminus \bigcup^4_{k=0}\Upsilon_{k}$, where
$$
\Upsilon_{k}=\left\{\zeta\in\mathbb{C}~\Big|~\arg\zeta=\frac{k\pi}{2}\right\},~k=0,1,2,3,
\quad \Upsilon_{4}=\left\{\zeta\in\mathbb{C}~\Big|~\arg\zeta=-\frac{\pi}{4}\right\};
$$ see Figure \ref{PC}.
\item{(2)} $\Phi^\mathrm{(PC)}(\zeta)$ satisfies the jump conditions
$$
\Phi^{(\mathrm{PC})}_+(\zeta)=\Phi^{(\mathrm{PC})}_-(\zeta)\left\{\begin{aligned}
&H_{k},\quad && \zeta \in\Upsilon_k,~k=0,1,2,3, \\
&e^{2\pi i\beta\sigma_3},\quad && \zeta \in\Upsilon_4.
\end{aligned}\right.
$$
\item{(3)} $\Phi^\mathrm{(PC)}(\zeta)$ satisfies the following asymptotic behavior as $\zeta\to\infty$
\begin{align}\label{PCAsyatinfty}
\Phi^\mathrm{(PC)}(\zeta)=\begin{pmatrix}0 &1 \\ 1 & -\zeta\end{pmatrix}2^{\frac{\sigma_3}{2}}\begin{pmatrix}
1+O\left(\zeta^{-2}\right) & \frac{{\beta}}{\zeta}+O\left(\zeta^{-3}\right) \\ \frac{1}{\zeta}+O\left(\zeta^{-3}\right) & 1+O\left(\zeta^{-2}\right)\end{pmatrix}
e^{\frac{\zeta^{2}}{4}\sigma_{3}}\zeta^{-{\beta}\sigma_3}.
\end{align}
\end{description}
\begin{figure}[t]
\centering
\includegraphics[width=7cm,height=7cm]{PC}\\
\caption{The jump contour and jump matrices for $\Phi^{(\mathrm{PC})}$}\label{PC}
\end{figure}
\subsection{Bessel parametrix}\label{Bessel}
Consider the following RH problem.
\subsection*{RH problem for $\Phi^{(\mathrm{Bes})}(\zeta)$}
\begin{description}
\item{(1)} $\Phi^\mathrm{(Bes)}(\zeta)$ is analytic for all $\zeta\in \mathbb{C}\setminus \bigcup^8_{k=1}\Gamma_{k}$, where $$
\Gamma_{k}=\left\{\zeta\in\mathbb{C}~\Big|~\arg\zeta=\frac{(k-1)\pi}{4}\right\},~k=1,\cdots,8;
$$ see Figure \ref{Bes}.
\item{(2)} We have the jump conditions
\begin{equation}\label{Besseljump}
\Phi_{+}^{(\mathrm{Bes})}(\zeta)=\Phi_{-}^{(\mathrm{Bes})}(\zeta)\left\{\begin{aligned} &\begin{pmatrix}0 & 1 \\ -1 & 0\end{pmatrix}, \quad &\zeta&\in\Gamma_1\cup\Gamma_5,\\
&\begin{pmatrix}1 & 0 \\ e^{-2\pi i\alpha} & 1\end{pmatrix}, \quad &\zeta&\in\Gamma_2\cup\Gamma_6,\\
&\begin{pmatrix}e^{\pi i\alpha} & 0 \\ 0 & e^{-\pi i\alpha}\end{pmatrix}, \quad &\zeta&\in\Gamma_3\cup\Gamma_7,\\
&\begin{pmatrix}1 & 0 \\ e^{2\pi i\alpha} & 1\end{pmatrix}, \quad &\zeta&\in\Gamma_4\cup\Gamma_8.
\end{aligned}\right.
\end{equation}
\item{(3)} As $\zeta\rightarrow\infty$,
\begin{equation}\label{BesInfty}
\Phi^\mathrm{(Bes)}(\zeta)=\frac{1}{\sqrt{2}}\begin{pmatrix}
1 & -i \\ -i & 1 \end{pmatrix}(\mathbf{I}+O(\zeta^{-1}))e^{\frac{\pi i\sigma_3}{4}}e^{-i\zeta\sigma_3}\left\{
\begin{aligned}
&e^{-\frac{\alpha\pi i\sigma_3}{2}}, & \zeta &\in \Lambda_1\cup\Lambda_2,\\
&e^{\frac{\alpha\pi i\sigma_3}{2}}, & \zeta &\in \Lambda_3\cup\Lambda_4,\\
&e^{\frac{\alpha\pi i\sigma_3}{2}}\sigma_1\sigma_3, & \zeta &\in \Lambda_5\cup\Lambda_6,\\
&e^{-\frac{\alpha\pi i\sigma_3}{2}}\sigma_1\sigma_3, & \zeta &\in\Lambda_7\cup\Lambda_8.
\end{aligned}\right.
\end{equation}
\end{description}
\begin{figure}[t]
\centering
\includegraphics[width=7cm,height=7cm]{Besseljump}\\
\caption{The jump contour and regions for $\Phi^{(\mathrm{Bes})}$}\label{Bes}
\end{figure}
From \cite{Van}, we see that the above RH problem can be constructed in terms of the modified Bessel function $I_{\alpha\pm\frac{1}{2}}(\zeta)$ and $K_{\alpha\pm\frac{1}{2}}(\zeta)$:
\begin{equation}\label{BesPara}
\Phi^\mathrm{(Bes)}(\zeta)=\begin{pmatrix}
\sqrt{\pi} \zeta^{\frac{1}{2}}I_{\alpha+\frac{1}{2}}(\zeta e^{-\frac{\pi i}{2}}) & -\frac{1}{\sqrt{\pi}}\zeta^{\frac{1}{2}}K_{\alpha+\frac{1}{2}}(\zeta e^{-\frac{\pi i}{2}})\\
-i\sqrt{\pi} \zeta^{\frac{1}{2}}I_{\alpha-\frac{1}{2}}(\zeta e^{-\frac{\pi i}{2}})&-\frac{i}{\sqrt{\pi}}\zeta^{\frac{1}{2}}K_{\alpha-\frac{1}{2}}(\zeta e^{-\frac{\pi i}{2}})
\end{pmatrix}e^{-\frac{1}{2}\alpha\pi i\sigma_3}\end{equation}
for $\zeta\in\Lambda_2$.
The explicit expressions of $\Phi^\mathrm{(Bes)}(\zeta)$ in other sectors are determined by \eqref{BesPara} and the jump relation \eqref{Besseljump}.
Using the series expansion of the modified Bessel function \cite[(10.25.2)]{NIST}
$$I_{\mu}(\zeta)=\left(\frac{\zeta}{2}\right )^{\mu}\sum_{k=0}^{\infty}\frac{(\frac{ \zeta^2}{4})^k}{k!\,\Gamma(\mu+k+1)},$$
and the connection formula \cite[(10.27.4)]{NIST}
$$K_{\mu}(\zeta)=\frac{\pi}{2}\frac{I_{-\mu}(\zeta)-I_{\mu}(\zeta) }{\sin(\pi \mu)}, \quad \mu \not\in \mathbb{Z},$$
it is seen from \eqref{BesPara} that
\begin{equation}\label{BesParaExpand}
\Phi^\mathrm{(Bes)}(\zeta)=(I+O(\zeta))C_\alpha^{\sigma_3}\begin{pmatrix}0 & -1 \\ 1 & 0\end{pmatrix} \zeta^{\alpha \sigma_3}\begin{pmatrix}
1 & \frac{1}{1+e^{-2\pi i \alpha}} \\ 0& 1
\end{pmatrix},\quad \zeta\to 0, ~~ \zeta\in \Lambda_2, \end{equation}
where $\alpha-\frac{1}{2} \not\in \mathbb{Z} $ and the constant $C_{\alpha}$ is given by
\begin{equation}\label{Calpha}
C_{\alpha}=\pi^{-\frac{1}{2}}2^{\alpha-\frac{1}{2}}e^{\pi i(\alpha+\frac{1}{4})}\Gamma(\alpha+\frac{1}{2}).
\end{equation}
If $\frac{1}{2}-\alpha\in\mathbb{N}$, using the connection formulas \cite[(10.27.1)]{NIST} and \cite[(10.27.3)]{NIST}
\begin{equation*}
I_{-n}(\zeta)=I_n(\zeta),\quad K_{-\mu}(\zeta)=K_{\mu}(\zeta),
\end{equation*}
and the small-$z$ asymptotics \cite[(10.30.1)-(10.30.3)]{NIST}
\begin{equation*}
I_{\mu}(\zeta)\sim \frac{\zeta^{\mu}}{2^{\mu}\Gamma(\mu+1)},\quad
K_{\mu}(\zeta)\sim \frac{2^{\mu-1}\Gamma(\mu+1)}{\zeta^{\mu}},\quad
K_0(\zeta)\sim -\ln \zeta,
\end{equation*}
we obtain
\begin{equation}\label{BesParaExpand1}
\Phi^\mathrm{(Bes)}(\zeta)=\Phi^\mathrm{(Bes)}_{0}(\zeta)\zeta^{-\alpha \sigma_3},\quad \zeta\to 0, \quad \zeta\in \Lambda_2,
\end{equation}
where $\Phi^\mathrm{(Bes)}_{0}(\zeta)$ is analytic near the origin.
The behaviors of $\Phi^\mathrm{(Bes)}(\zeta)$ near the origin in the other regions can be determined by \eqref{BesParaExpand}, \eqref{BesParaExpand1} and the jump relations \eqref{Besseljump}.
\end{appendices}
|
\section{INTRODUCTION}
Depth estimation is a fundamental computer vision task, which plays an influential role in many practical applications, such as autonomous driving, robot grasping and augmented reality.
The traditional geometric approaches~\cite{schonberger2016colmap, gherardi2010improvingsfm, mur2015orb} solve this problem by exploiting the key point matching relationship between adjacent frames and recovering the depth information of 2D pixels through triangulation.
Although traditional geometric methods have achieved promising performance, they are unable to handle low-texture regions and non-stationary scenes.
In addition, the huge computational cost of feature matching limits their practical application.
Recently, deep learning techniques have improved this research area by training networks to directly predict depth from single image.
Despite the fact that supervised learning methods have achieved good results~\cite{adabins, li2022depthformer, simipu}, the cost of labeling ground truth limits the practical application of such methods.
To solve this problem, unsupervised depth estimation~\cite{monodepth2,manydepth,epc++,zhou2017unsupervised,zhan2018unsupervised} is gradually becoming a novel research trend, which treats depth estimation as a view synthesis problem.
Although self-supervised depth estimation has made great progress in recent years, there are still some unresolved issues.
First, without the feature matching step, the unsupervised depth estimation cannot leverage the semantic information from adjacent frames, which leads to inaccurate depth estimation results.
Second, it is known that the photometric loss is sensitive to dynamic objects, which violates the assumption of the static world. Therefore, the perspective projection will be non-rigid, resulting in unstable training process.
On the one hand, Manydepth\cite{manydepth} leveraged cost-volume based method to aggregate multi-frame information.
However, inaccurate pose estimation introduces additional noise to feature matching, which can sharply decreasing model performance in dynamic scenes.
On the other hand, recent work have tried to introduce instance segmentaion \cite{lee2019instance}, auto-mask \cite{monodepth2} and optical flow \cite{yin2018geonet} to deal with dynamic objects under the self-supervised learning framework.
These methods introduce expensive additional labeling costs, which are therefore difficult to be applied in real-world applications.
\begin{figure}[t!]
\centering
\includegraphics[width=0.35\textwidth]{figures/overview.jpg}
\caption{
\textbf{Overall architecture} of our proposed method. On the one hand, we model the dynamic objects in the scene to improve the accuracy of depth estimation in dynamic regions. On the other hand, considering the presence of dynamic objects, we design a geometry-based cross-view feature fusion approach to obtain discriminative feature representations.
}
\label{fig:overview}
\end{figure}
In this work, we propose a \textbf{Dyna-DepthFormer} network in order to solve the above-mentioned problems.
Specifically, we propose a depth feature enhance module consisting of self-attention and cross-attention module to capture the correlation in multi-view feature map.
Inspired by \cite{zhu2020deformable}, we realize this module by constructing a Transformer-based module, while the deformable attention is adopted to reduce the computational costs and adjust sample point to capture the feature in the local region.
The generated features will be more discriminative and robust, since it focus on the surrounding local context and aggregates relevant information.
Furthermore, we propose a novel Motion Network, which model the dynamic object by leveraging the motion field prediction. Then, We can use the motion field and camera motion to calculate the reprojection loss without using semantic prior. In summary, our contributions are:
\begin{itemize}
\item
We propose a transformer-based Depth Network, which enhances the depth feature representation through multi-view feature aggregating. Considering the high computation of dense attention, we propose a perspective-projection based temporal deformable cross-attention module that efficiently reduces the computation.
\item
In addition to camera ego motion, we proposed to estimate the 3D motion field of moving objects, and improve its completeness through iterative refinement. This eliminates the necessity of the scene-static assumption and enables unsupervised method to use more unlabeled data for training.
\item
We demonstrate that our proposed method achieves promising results in monocular depth estimation on benchmark datasets like KITTI and Cityscapes while reducing computing cost.
\end{itemize}
\section{RELATED WORK}
\subsection{Self-Supervised Depth Estimation}
Self-supervised depth estimation methods, in contrast to supervised depth estimation methods, can estimate depth from unlabeled video training data.
The work of Monodepth~\cite{godard2017unsupervised} introduced a self-supervised depth estimation method by leveraging the left-right consistency.
SFM-learner~\cite{zhou2017unsupervised} estimated the relative poses between adjacent frames simultaneously with the addition of a pose network.
Then, Monodepth2~\cite{monodepth2} absorbed the advantages of the prior methods and proposed a universal framework, which can be trained with stereo pairs and videos data.
However, this approach was afflicted by issues such as scale ambiguity, dynamic objects, low-texture areas, and non-Lambertian surfaces.
In recent years, several strategies have been developed to improve the performance of depth estimation by tackling the aforementioned difficulties.
Featdepth~\cite{shu2020featdepth} introduced a auto-encoder structure in order to obtain a robust feature representation, which can improve the depth estimation accuracy in low-texture region.
Packnet~\cite{packnet} proposed a novel 3D convolution packing module to overcome the scale ambiguity problem, while it suffered from heavy computational cost.
With the development of cost-volume based multi-view stereo (MVS) methods, some of the work~\cite{manydepth, guizilini2022multi, wang2021deep, monorec} introduced the cost-volume to self-supervised depth estimation field, which enabled the learning of additional geometric cues from feature matching across frames.
However, due to the hand-crafted similarity metrics, the cost volume will be an unreliable source of depth information in locations where objects are moving or surfaces are untextured.
\subsection{Attention Mechanisms in Depth Estimation}
Self-Attention mechanism have shown promising results in the field of natural language processing~\cite{attention_all} and are becoming increasingly popular in computer vision region ~\cite{dosovitskiy2020vit, liu2021swin}.
In the field of depth estimation, current methods can be divided into two categories.
One class of methods apply attention mechanisms to exploit long-range correlation and local information of single view feature representation representation~\cite{lyu2021hrdepth,yan2021channel,li2022depthformer}.
The other class of methods tried to fuse features from multiple sources through an attention mechanism.
Some work \cite{jung2021fsre, li2021learning} leveraged semantic information to guide depth feature representation.
More related to our work, \cite{ruhkamp2021attention} proposed a spatial-temporal attention module to fuse cross-view information through two cascaded attention blocks.
However, their approach did not make sufficient use of geometric information, resulting in high computational cost.
\subsection{Depth estimation in Dynamic Environment}
Video sequence data is usually captured in scenes containing dynamic objects.
However, the non-rigid parts of the scene can affect the self-supervised training.
To alleviate this problem, \cite{monodepth2} introduced an novel auto-mask method to exclude dynamic object areas.
Although auto-mask can exclude dynamic object pixels, it also masks away some pixels that are useful for training.
Some of the work \cite{boulahbal2022instance, lee2019instance} leveraged semantic information to seperate the foreground and background object.
Then, they estimated the object motion and camera ego-motion seperately.
However, their methods require expensive labeling costs.
Recent work attempted to model the moving object explicitly with optical flow~\cite{yin2018geonet} or 3D motion field \cite{li2020dynamic, lee2021attentive}.
Comparing to \cite{lee2021attentive}, we design a one-stage warping based network to estimate the motion field, which directly represent object motion in 3D space.
Furthermore, to improve the quality of the predicted motion fields, we also propose an iterative optimization methods with sparsity regularization loss.
\section{METHOD}
In this section, we propose Dyna-DepthFormer, a self-supervised depth estimation Framework with Transformer.
We first describe the problem formulation, followed by a framework overview of our method in Section \ref{sec:overview}.
In order to predict accurate depth map of the target image, we present a depth estimation network in Section \ref{sec:depth_estimate}, which can fuse multi-view geometric information efficiently.
In Section \ref{sec:motion_estimation}, we give details on the Motion Estimation Network, an module designed to deal with the moving objects and non-rigid region in training process.
In Section \ref{sec:loss_function}, we further elaborate on the self-supervised loss function and regularization to finalize our training strategy.
\subsection{Framework Overview}
\label{sec:overview}
The overall framework, as illustrated in Fig \ref{fig:model_pipeline}, consists of four parts: the pre-trained backbone, the depth estimation network, the motion estimation network and the view synthesis module.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.60\textwidth]{figures/pipeline.pdf}
\caption{
\textbf{Pipeline of our proposed Dyna-Depthformer architecture}. The entire pipeline includes three parts, Depth Network, Motion Network and View Synthesis Module. For Depth Network, cross-view attention and self attention module are adopted to enhance the feature modeling and depth inferring. For Motion Network, we predict ego motion and object motion with two decoupled head. The view synthesis module reconstructs the image by differentiable warping as a way to perform self-supervised learning.
}
\label{fig:model_pipeline}
\vspace{-4mm}
\end{figure*}
The pre-trained backbone takes the scaled images pair $\{\mathbf{I}_t,\mathbf{I}_{s}\}$ as input to generate multi-scale image feature. We first use motion network $\mathbf{M}_{\psi}$ and pose network $\mathbf{P}_{\phi}$ to generate ego motion $\{\mathbf{R}_{t \rightarrow s},\mathbf{t}_{t \rightarrow s}\}$ and object motion field $\mathbf{T}_{res}$. Then, we perform perspective projection based attention to obtain enhanced multi-view depth feature to generate accurate depth prediction $\mathbf{D}_t$ with depth network $\mathbf{D}_{\theta}$.
\begin{equation}
\begin{gathered}
\left\{\mathbf{R}_{t \rightarrow s}, \mathbf{t}_{t \rightarrow s}\right\}=\mathbf{P}_\phi\left(\left\{\mathbf{I}_t, \mathbf{I}_s\right\}\right) \\
\mathbf{T}_{\text {res }}=\mathbf{M}_{\psi}\left(\left\{\mathbf{I}_t, \mathbf{I}_{s \rightarrow t}\right\}\right) \\
\mathbf{D}_t=\mathbf{D}_\theta\left(\mathbf{I}_t\right).
\end{gathered}
\end{equation}
With the predicted depth $\mathbf{D}_t \in \mathbb{R}^{1 \times H \times W}$, relative camera pose $\{ \mathbf{R}_{t \rightarrow s}, \mathbf{t}_{t \rightarrow s} \}$ and object motion field $\mathbf{T}_{res}\in\mathbb{R}^{3 \times W \times H}$, we can synthesize the target frame $\mathbf{I}_t$ using the adjacent source frame $\mathbf{I}_s$ by differentiable view synthesis module~\cite{packnet}.
The difference between the source image and the reconstructed image can provide supervisory signal for self supervised training.
\subsection{Depth Estimation Network}
\label{sec:depth_estimate}
The goal of the depth estimation network ${D}_{\theta}$ is to map the input RGB images $\mathbf{I}_t$ to depth map $\mathbf{D}_t$.
Noted that, monocular depth estimation is inherently an ill-posed problem, which suffers from moving object, textureless region and non-rigid region, etc.
Many previous work~\cite{yan2021channelwise, lyu2021hrdepth,shu2020featdepth} attempted to improve feature representation by applying self-attention module to current picture feature in order to address the aforementioned issue.
These approaches, however, have substantial computational costs and do not leverage geometric previous information.
Instead, we propose an perspective-project-based temporal cross-attention module to aggregate features in a learnable way.
\textbf{Feature Extract.} Given a target image $\mathbf{I_t} \in \mathbb{R}^{3 \times H \times W}$ and its neighboring source images $\mathbf{I_s} \in \{ \mathbf{I_{t+1}},\mathbf{I_{t-1}} \}$, the first step is to extract the multi-scale image features of these images.
A shared backbone~\cite{he2016resnet} is applied, where the images are downscaled $N$ times to produce multi-scale deep features map $\mathbf{F}_{i=\{t-1, t, t+1\}}^{k=0,\ldots,N-1} \in \mathbb{R}^{C_k \times \frac{H}{2^k} \times \frac{W}{2^k}}$.
Then, we feed the extracted feature maps to the subsequent module.
\textbf{Temporal Cross-Attention.}
As single-view depth estimation suffers from ill-posed problem, our method enhance the feature representation of target image by aggregating its neighboring image features with cross attention mechanisms.
We use the features of the target image as \textit{Query} to match the features in the source image (\textit{Keys}).
However, due to the huge number of key-value pairs, the original attention mechanism cannot be applied directly.
Meanwhile, the epipolar sampling method~\cite{wang2022mvster,li2022depthformer} would suffer from camera parameter noise or scale-ambiguous problem as a result of the inaccurate pose estimate.
Inspired by \cite{zhu2020deformable, li2022bevformer}, we obtained the key-value pairs in the neighboring image features with a learnable way to reduce computational cost and increase efficiency.
Fig.~\ref{fig:cross_attention} illustrates the feature warping process and the generation of key-value pairs in our cross-attention module.
Given a query point $\mathbf{p}_t=(u,v)$ in $\mathbf{F}_t$, it can be warped to the source camera coordinate by:
\begin{equation}
\left(\mathbf{p}_s~1 \right)^T = \mathbf{K}\left(\mathbf{R}_{t \rightarrow s} \hat{\mathbf{D}}_t\left(\mathbf{p}_t\right) \mathbf{K}^{-1}\left (\mathbf{p}_t~1)\right)^T+\mathbf{t}_{t \rightarrow s}\right),
\end{equation}
where $\mathbf{p}_s$ is the associated point of the query point on the source plane, $\mathbf{K}$ is the camera intrinsics given in advance, $\{\mathbf{R}_{t \rightarrow s},\mathbf{t}_{res}\}$ indicates the relative camera pose predicted by motion network, and $\hat{\mathbf{D}}_t(p_t)$ is the depth value at point $\mathbf{p}_t$.
Note that, we use an auxiliary single scale depth decoder to generate a coarse predicted initial depth estimate $\hat{\mathbf{D}}_t$.
Then, we search for the most relevant features of the query feature in the neighborhood of $\mathbf{p}_s$.
Specifically, we learn some offsets $\Delta \mathbf{p} \in \mathbb{R}^2$ to the point $\mathbf{p}_s$ with MLP layer.
The deformable attention feature is calculated by:
\begin{equation}
\mathbf{F}_t^{\prime}(\mathbf{p}_t)=\sum_{m=1}^M \mathbf{W}_m\left[\sum_{k=1}^K A_{m k} \cdot \mathbf{W}_m^{\prime} \mathbf{F}_s\left(\mathbf{p}_s+\Delta \mathbf{p}_{m k}\right)\right],
\end{equation}
where $m$ indexes the attention head, $k$ indexes the sampled keys, $K$ is the total sampled points, $\mathbf{F}^{\prime}_t$ indicates the enhanced target feature, and $\mathbf{A}_{mk}$ indicates the scalar attention weight between \textit{Query} and \textit{Key} embedding.
As $\mathbf{p}_s+\Delta \mathbf{p}_{m k}$ can be a fraction, bilinear interpolation is used to compute $\mathbf{F}_s\left(\mathbf{p}_s+\Delta \mathbf{p}_{m k}\right)$.
Note that $\mathbf{p}_s$ can also be regarded as an coarse initial point generated by the geometric prior, which will be optimized in the subsequent training process.
After that, We can fuse multiple aggregated features from source views by linear projection.
\begin{figure}[t!]
\centering
\includegraphics[width=0.3\textwidth]{figures/cross_attention.pdf}
\caption{
The details of a \textbf{Temporal Cross Attention Module.} This module fuses adjacent frame features by deformable cross-attention mechanism to obtain a robust and discriminative depth feature representation.
}
\label{fig:cross_attention}
\vspace{-4mm}
\end{figure}
\textbf{Self-Attention Refinement.}
The enhanced feature $F_t^{\prime}$ are generated following temporal cross-view feature aggregation, which is more robust, but matching noise is maintained.
Inspired by the cost aggregation step in the traditional stereo matching task~\cite{sgm}, we consider leveraging self-attention to refine the local features.
The vectors in the feature map aggregate the discriminant features in the neighborhood based on cosine similarity to smooth the noisy information in cross-view enhanced feature map.
Specifically, Given the query point $\mathbf{p}_q=(u,v)$, the query feature $z_q$, and the enhanced feature $\mathbf{F}_t^{\prime}$, the self-attention refined feature is calculated by:
\begin{equation}
\mathbf{F}_t^{\prime\prime}(\mathbf{p}_q)=\sum_{m=1}^M \mathbf{W}_m\left[\sum_{k=1}^K A_{m k} \cdot \mathbf{W}_m^{\prime} \mathbf{F}_t^{\prime}\left(\mathbf{p}_q+\Delta \mathbf{p}_{m k}\right)\right].
\end{equation}
After obtaining the enhanced feature $\mathbf{F}_t^{\prime\prime}$, we utilize a straightforward decoder mentioned in~\cite{lyu2021hrdepth} to generate the multi-scale depth predictions $\mathbf{D}_t$.
\begin{table*}[t!]
\vspace{2mm}
\centering
\caption{\textbf{Depth estimation results} on the KITTI and Cityscapes datasets. \emph{Multi-Fr.} indicates using multiple frames during test time. \emph{Semantic.} indicates leveraging semantic prior during training process. K and CS denote the KITTI and the Cityscapes datasets.}
\label{tab:depth_result}
\scalebox{0.8}{
\begin{tabular}{lcccccccccc}
\toprule
\textbf{Method} & Multi-Fr & Semantic & Dataset & AbsRel & SqRel & RMSE & RMSE$_{log}$ & $\delta < 1.25$ & $\delta < 1.25^2$ & $\delta < 1.25^3$ \\ \midrule
Struct2depth~\cite{struct2depth} & & \checkmark & K & 0.141 & 1.026 & 5.291 & 0.215 & 0.816 & 0.945 & 0.979 \\
GeoNet~\cite{yin2018geonet} & & & K & 0.155 & 1.296 & 5.857 & 0.233 & 0.793 & 0.931 & 0.973 \\
EPC++~\cite{epc++} & & & K & 0.141 & 1.029 & 5.350 & 0.216 & 0.816 & 0.941 & 0.976 \\
SC-SfM~\cite{bian2019depth} & & & K & 0.119 & 0.857 & 4.950 & 0.197 & 0.863 & 0.957 & 0.981 \\
CC~\cite{competi_colab} & & & K & 0.140 & 1.070 & 5.326 & 0.217 & 0.826 & 0.941 & 0.975 \\
GLNet~\cite{glnet} & & & K & 0.135 & 1.070 & 5.230 & 0.210 & 0.841 & 0.948 & 0.980 \\
Monodepth2~\cite{monodepth2} & & & K & 0.115 & 0.882 & 4.701 & 0.190 & 0.879 & 0.961 & 0.982 \\
PackNet-SFM~\cite{packnet} & & & K & 0.111 & 0.785 & 4.601 & 0.189 & 0.878 & 0.960 & 0.982 \\
Lee \textit{et al.} \cite{lee2021attentive} & & \checkmark & K & 0.112 & 0.777 & 4.772 & 0.191 & 0.872 & 0.959 & 0.982 \\
Gao \textit{et al.} \cite{gao} & & & K & 0.112 & 0.866 & 4.693 & 0.189 & 0.881 & 0.961 & 0.981 \\
FeatDepth~\cite{shu2020featdepth} & & & K & 0.109 & 0.923 & 4.819 & 0.193 & 0.886 & 0.963 & 0.981 \\
TC-Depth~\cite{ruhkamp2021attention} & \textbf{\checkmark} & & K & {0.106} & 0.770 & 4.558 & 0.187 & 0.890 & 0.964 & 0.983 \\
RM-Depth~\cite{rmdepth} & & & K & 0.108 & \textbf{0.710} & 4.513 & 0.183 & 0.884 & 0.964 & 0.983 \\
Manydepth~\cite{manydepth} & \checkmark & & K & \underline{0.098} & 0.770 & \underline{4.459} & \underline{0.176} & \textbf{0.900} & \underline{0.965} & 0.983 \\
\textbf{Ours} & \checkmark & & K & \textbf{0.094} & {\underline{0.734}} & \textbf{4.442} & \textbf{0.169} & \underline{0.893} & \textbf{0.967} & \textbf{0.983} \\ \hline
Struct2Depth~\cite{struct2depth} & \multicolumn{1}{l}{} & \checkmark & CS & 0.145 & 1.737 & 7.280 & 0.205 & 0.813 & 0.942 & 0.976 \\
Monodepth2~\cite{monodepth2} & \multicolumn{1}{l}{} & & CS & 0.129 & 1.569 & 6.876 & 0.187 & 0.849 & 0.957 & 0.983 \\
Manydepth~\cite{manydepth} & \checkmark & \multicolumn{1}{l}{} & CS & 0.114 & 1.193 & 6.223 & 0.170 & 0.875 & 0.967 & 0.989 \\
InstaDM~\cite{lee2019instance} & \multicolumn{1}{l}{} & \checkmark & CS & 0.111 & 1.158 & 6.437 & 0.182 & 0.868 & 0.961 & 0.983 \\
Lee \textit{et al.} \cite{lee2021attentive} & \multicolumn{1}{l}{} & \checkmark & CS & 0.116 & 1.213 & 6.695 & 0.186 & 0.852 & 0.951 & 0.982 \\
RM-Depth~\cite{rmdepth} & \textbf{} & \multicolumn{1}{l}{} & CS & \underline{0.103} & \underline{1.000} & \underline{5.867} & \underline{0.157} & \underline{0.895} & \underline{0.974} & \underline{0.991} \\
\multicolumn{1}{l}{\textbf{Ours}} & \checkmark & \multicolumn{1}{l}{} & CS & \textbf{0.100} & \textbf{0.834} & \textbf{5.843} & \textbf{0.154} & \textbf{0.901} & \textbf{0.975} & \multicolumn{1}{c}{\textbf{0.992}} \\ \bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[t!]
\centering
\caption{\textbf{Depth Estimation Result on Dynamic Object Region.} We evaluate the depth prediction accuracy of dynamic objects (e.g.,Vehicles, Person, Bike) on KITTI and Cityscapes datasets. The instance masks are obtained using Maskformer~\cite{maskformer} pre-trained model.}
\label{tab:dynamic_result}
\scalebox{0.95}{
\begin{tabular}{lcccccccc}
\toprule
Methods & Dataset & AbsRel & SqRel & RMSE & RMSE$_{log}$ & $\delta < 1.25$ & $\delta < 1.25^2$ & $\delta < 1.25^3$ \\ \midrule
Monodepth2~\cite{monodepth2} & K & 0.172 & 1.892 & 5.732 & 0.273 & 0.819 & 0.917 & 0.945 \\
Packnet~\cite{packnet} & K & 0.169 & 1.873 & 5.797 & 0.276 & 0.809 & 0.913 & 0.942 \\
Insta-DM~\cite{lee2019instance} & K & 0.156 & 1.327 & 5.763 & 0.272 & 0.818 & 0.917 & 0.943 \\
Manydepth~\cite{manydepth} & K & 0.175 & 2.014 & 5.837 & 0.280 & 0.773 & 0.910 & 0.942 \\
\textbf{Ours} & K & \textbf{0.148} & \textbf{1.312} & \textbf{5.280} & \textbf{0.257} & \textbf{0.837} & \textbf{0.921} & \textbf{0.945} \\ \hline
Monodepth2~\cite{monodepth2} & CS & 0.159 & 1.948 & 6.492 & 0.217 & 0.820 & 0.948 & 0.981 \\
Packnet~\cite{packnet} & CS & 0.167 & 2.219 & 6.683 & 0.228 & 0.764 & 0.917 & 0.968 \\
Insta-DM~\cite{lee2019instance} & CS & 0.141 & 1.698 & 5.797 & 0.185 & 0.841 & 0.956 & 0.982 \\
Manydepth~\cite{manydepth} & CS & 0.163 & 2.138 & 6.537 & 0.220 & 0.770 & 0.928 & 0.971 \\
\textbf{Ours} & CS & \textbf{0.132} & \textbf{1.281} & \textbf{4.715} & \textbf{0.170} & \textbf{0.853} & \textbf{0.962} & \textbf{0.984} \\ \bottomrule
\end{tabular}
}
\end{table*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.55\textwidth]{figures/depth_result.pdf}
\includegraphics[width=0.35\textwidth]{figures/flow_result.jpg}
\makebox[0.6\textwidth]{\small (a) Comparison of Depth Result }
\makebox[0.3\textwidth]{\small (b) Motion Field and Depth Result}
\caption{
\textbf{Qualitative depth estimation results} of our proposed Dyna-DepthFormer architecture, on the Cityscapes dataset. The results show that our method can achieve more accurate depth estimation results on dynamic objects and low-texture regions, while having sharper borders. In addition, our motion component can decouple object motion and camera ego motion efficiently.
}
\vspace{-4mm}
\label{fig:depth_result}
\end{figure*}
\begin{table}[t]
\vspace{2mm}
\centering
\caption{Ablation study of attention modules on KITTI dataset.}
\label{tab:cross_attention_ablation}
\begin{tabular}{@{}lccccc@{}}
\toprule
\textbf{Method} & AbsRel & SqRel & RMSE & $\delta < 1.25$ & Mem (GB) \\ \midrule
\textbf{full} & \textbf{0.100} & \textbf{0.698} & \textbf{4.236} & \textbf{0.903} & 6.6 \\
W/o cross-attn & 0.106 & 0.714 & 4.562 & 0.887 & 3.8 \\
W/o self-attn & 0.102 & 0.711 & 4.383 & 0.895 & 6.3 \\
baseline & 0.108 & 0.721 & 4.421 & 0.883 & 3.5 \\
\midrule
K=1 & 0.108 & 0.774 & 4.823 & 0.882 & 3.7 \\
K=2 & 0.106 & 0.746 & 4.548 & 0.887 & 4.6 \\
K=4 & 0.100 & 0.714 & 4.320 & 0.895 & 7.8 \\
\textbf{K=8} & \textbf{0.097} & \textbf{0.694} & \textbf{4.216} & \textbf{0.913} & 12.4 \\ \hline
Manydepth & 0.100 & 0.770 & 4.459 & 0.900 & 16.8 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[t]
\vspace{2mm}
\centering
\begin{tabular}{@{}lcccc@{}}
\toprule
\textbf{Method} & AbsRel & SqRel & RMSE & $\delta < 1.25$ \\ \midrule
W Projection & 0.094 & 0.734 & 4.442 & 0.893 \\
W/o Projection & 0.100 & 0.764 & 4.567 & 0.876 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[t]
\centering
\caption{
Ablation study of the motion field estimation network on the dynamic area in Cityscapes Dataset.
\emph{Motion} indicates motion field based warping. \emph{Refine} indicates iterative refine step. \emph{Mask} indicates the auto-mask mechanism.
}
\label{tab:motion_ablation}
\begin{tabular}{@{}cccccc@{}}
\toprule
Motion & Refine & Mask & AbsRel & RMSE & $\delta < 1.25$ \\ \midrule
& & & 0.153 & 6.325 & 0.831 \\
\checkmark & & & 0.142 & 5.835 & 0.837 \\
\checkmark & \checkmark & & 0.138 & 5.331 & 0.847 \\
\checkmark & & \checkmark & 0.136 & 5.218 & 0.850 \\
\checkmark & \checkmark & \checkmark & \textbf{0.132} & \textbf{4.715} & \textbf{0.853} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[t]
\vspace{2mm}
\centering
\caption{Ablation study of the number of iterative refinement on dynamic area in Cityscapes dataset.}
\label{tab:iter_num}
\begin{tabular}{@{}ccccc@{}}
\toprule
\textbf{Iter} & AbsRel & SqRel & RMSE & $\delta < 1.25$ \\ \midrule
1 & 0.136 & 1.576 & 5.342 & 0.849 \\
2 & 0.133 & 1.410 & 4.922 & 0.852 \\
4 & 0.132 & 1.293 & 4.723 & 0.853 \\
6 & \textbf{0.132} & \textbf{1.281} & \textbf{4.715} & \textbf{0.853} \\ \bottomrule
\end{tabular}
\end{table}
\subsection{Motion Estimate Network}
\label{sec:motion_estimation}
Previous research have tended to recover depth and camera motion simultaneously, but have ignored the non-rigid flow~\cite{yin2018geonet} caused by moving objects, resulting in incorrect view synthesis and a direct influence on the performance of self-supervised learning.
In this section, we propose a novel motion estimation network that aims to address moving objects problem in a general manner.
Specifically, We recover the camera and object motion with two separate heads, a motion field decode head and pose decode head, while they share the same motion feature encoder.
Note that we do not utilize any semantic prior during training process.
The camera motion is also referred to as rigid flow.
The shared encoder takes the image pair $\{\mathbf{I}_t, \mathbf{I}_s\}$ as input and outputs the relative pose $\{\mathbf{R}_{t \rightarrow s},\mathbf{t}_{t \rightarrow s}\}$ between them.
Then, we can warp the source images $\mathbf{I}_s$ to target frame by differentiable view synthesis
\begin{equation}
\mathbf{I}_{s \rightarrow t}=\mathbf{I}_{s}\left\langle\operatorname{proj}\left(\mathbf{D}_{t}, \{\mathbf{R}_{t \rightarrow s},\mathbf{t}_{t \rightarrow s}\}, \mathbf{K}\right)\right\rangle,
\end{equation}
where $\left\langle \right\rangle$ is the bilinear interpolation sampling operator, $\operatorname{proj}()$ are the resulting 2D coordinates of the projected depths in $\mathbf{I}_s$, and $\mathbf{D}_t$ is the depth predicted by depth network.
As we synthesize the target view with the predicted camera motion, we suppose that the misaligned component is mainly caused by moving object.
Based on the above assumption, we then feed $\{\mathbf{I}_t, \mathbf{I}_{s \rightarrow t}\}$ into the shared encoder to generate object motion field $\textbf{T}_{res} \in \mathbb{R}^{3 \times H \times W}$.
It is necessary to clarify that the motion field $\mathbf{T}_{res}$ is not exactly the same as scene flow \cite{teed2021raft3d}, as we consider the rotational motion of the moving object to be nearly zero.
\textbf{Iterative Refinement.}
Our motion network are trained to decrease the local errors caused by object motion while learning the motion field.
However, due to the limitation of photometric consistency, motion variation arises during training.
To address this problem, we propose iterative refinement step to improve consistency and completeness of the predicted motion field.
Specifically, we iteratively leverage $\textbf{T}_{res}$ and $\mathbf{I}_{t \rightarrow s}$ to reconstruct target image frame,
\begin{equation}
\begin{gathered}
\mathbf{I}_{s \rightarrow t}=\mathbf{I}_s\left\langle\operatorname{proj}\left(\mathbf{D}_t,\left\{\mathbf{R}_{t \rightarrow s}, \mathbf{t}_{t \rightarrow s}+\mathbf{T}_{\text {res }}\right\}, \mathbf{K}\right)\right\rangle \\
\Delta \mathbf{T}=\mathbf{M}_{\psi}\left(\mathbf{I}_t, \mathbf{I}_{t \rightarrow s}\right),~
\mathbf{T}_{\text {res }}=\mathbf{T}_{\text {res }}+\Delta \mathbf{T},
\end{gathered}
\end{equation}
where $\mathbf{T}_{res}$ is initialized to zero. By calculating the above process in several cycles, we can obtain the complete motion.
\subsection{Self-Supervised Objective}
In this section, we present the details of our self-supervised objective, which can be fomulated as:
\begin{equation}
{L}=\lambda_p {L}_p + \lambda_{st} {L}_{st} + \lambda_{sd} {L}_{sd} + \lambda_{sp} {L}_{sp}.
\end{equation}
where ${\lambda_p,\lambda_{st},\lambda_{sd},\lambda_{sp}}$ are human-set parameters to balance the impact of each loss function on the model training,
\textbf{Photometric consistency loss.}
The photometric loss measures the difference between the target image $\mathbf{I}_t$ and the synthesized image $\mathbf{I}_{s \rightarrow t}$ for self-supervised training. The photometric error is defined as a combination of a L1 and SSIM loss:
\begin{equation}
L_p = \alpha \frac{1-\operatorname{SSIM}\left(\mathbf{I}_{t}, \mathbf{I}_{s \rightarrow t}\right)}{2}+(1-\alpha)\left\|\mathbf{I}_{t}-I_{s \rightarrow t}\right\|_{1},
\end{equation}
where $\alpha$ is set to $0.85$.
To address the occlusion pixel problem, we compute the per-pixel minimum reprojection loss across all source views rather than averaging them, and we also integrate the auto-masking mechanism to filter dynamic pixel.
\textbf{Smoothness regularization.}
Following \cite{monodepth2}, we also use an edge-aware depth smoothness regularization and depth smoothness regularization to improve the performance in low texture regions:
\begin{equation}
{L}_{st}=\sum_{p}\left(\nabla \mathbf{T}_{res}(p) \cdot e^{-\nabla \mathbf{D}_t(p)}\right)^{2}
\end{equation}
\begin{equation}
L_{sd}=\sum_{p}\left(\nabla \mathbf{D}_{t}(p) \cdot e^{-\nabla \mathbf{I}_{t}(p)}\right)^{2}
\end{equation}
\textbf{Motion Field Sparsity loss.}
In order to encourage the sparsity of motion field prediction, we also design a motion field regularization loss:
\begin{equation}
{L}_{sp}= \sum_{p} L_{g1}(\mathbf{T}_{res}),
\end{equation}
where $L_{g1}()$ is chosen to be the sparsity function as it is mentioned in \cite{dynamic_scenes}.
\label{sec:loss_function}
\section{EXPERIMENTS}
\subsection{Datasets}
\textbf{KITTI}~\cite{geiger2013vision} is an autonomous driving dataset collected in urban and highway areas. We adopted the pre-processing method of \cite{zhou2017unsupervised} to remove the static frames in Eigen split~\cite{eigen2014depth}.
Finally, The training set includes 39,810 image triples, while the validation set has 4,424 images.
The depth map ground truth is generated by projecting the LiDAR pointcloud into the camera coordinate.
Note that, we only calculate accuracy and error metrics on these sparse depth points.
The results shown are tested on 697 test images following the steps in \cite{garg2016unsupervised}.
\textbf{Cityscapes}~\cite{cityscapes} is also a popular urban driving dataset.
The training set contains 69,730 image triplets, and the validation set includes 1,525 images.
We used the cropping and evaluation protocol mentioned in~\cite{laina2016deeper}.
\subsection{Implementation Details}
\textbf{Network Details.} We designed Depth Estimation Network with an encoder and decoder structure based on ResNet50 \cite{he2016resnet}, which is pre-trained with ImageNet \cite{deng2009imagenet}.
The decoder has the same structure as HR-Depth ~\cite{lyu2021hrdepth}, and its output is a multi-scale inverse depth map.
The transformer module is realized follow by \cite{zhu2020deformable}. The total attention head M is set to 8, and the total sample points K is set to 4 for efficiency.
\textbf{Training Details.} Our models are implemented using PyTorch~\cite{paszke2017automatic} and trained across four NVIDIA GTX 2080~Ti GPUs.
We use the Adam optimizer~\cite{kingma2014adam} , with $\beta_1 = 0.9$ and $\beta_2=0.999$, and a batch size of 4 per GPU.
The loss weights $\{\lambda_p,\lambda_{st},\lambda_{sd},\lambda_{sp}\}$ are set to $\{1, 0.01, 0.001, 0.01\}$.
We train our networks for 25 epochs, with the learning rate of 2e-4 for the first 15 epochs and reduce the learning rate to 2e-5 for the rest epochs.
To ensure the stability of the training process, we freeze the pose estimation network after 15 epochs.
\subsection{Benchmark Performance}
In order to validate our Dyna-DepthFormer architecture, we performed an in-depth comparison of its performance against other published methods.
Table~\ref{tab:depth_result} shows the depth prediction results on the Cityscapes \cite{cityscapes} and KITTI \cite{gtkitti} testing sets.
We compare with three types of methods: monocular depth estimation methods which predict the depth without using multi-frame information in the same mini-batch, and cost-volume based method including Manydepth~\cite{manydepth} and TC-Depth~\cite{ruhkamp2021attention}, and dynamic object modeling methods including InstaDM~\cite{lee2019instance}, struct2depth~\cite{struct2depth} and RM-Depth~\cite{rmdepth}, etc.
As shown in Table~\ref{tab:depth_result}, Dyna-DepthFormer outperforms the compared methods on both KITTI and Cityscapes datasets.
Cost-volume based methods can achieve promising result, while they suffer from feature matching noise on dynamic area, which has a large impact on depth estimation accuracy.
Differently, we fuse the multi-view features through a more robust cross-attention based approach and therefore achieve more accurate depth estimation results.
Compared to the KITTI dataset, the Cityscapes dataset involves more moving objects, therefore the improvement of our method on the cityscapes dataset is more significant.
Table~\ref{tab:dynamic_result} shows the depth prediction on dynamic objects region (e.g.,Vehicles, Person, Bike).
We obtain the dynamic instance masks using the pre-trained state-of-the-art instance segmentation method~\cite{maskformer}.
To cope with dynamic objects during training process, we did not use two-stage training scheme~\cite{lee2021attentive} or apply instance-level semantic prior, but our approach achieved better results by decoupling object and ego motion explicitly.
Qualitative results are shown in Fig.~\ref{fig:depth_result}.
\subsection{Ablation Analysis}
\textbf{Effectness of Attention Module.}
As shown in Table~\ref{tab:cross_attention_ablation}, the complete model outperforms the baseline model by a large margin. The proposed attention components are effective in improving the depth accuracy. By removing either the cross attention or self attention module, the depth accuracy is decreased. This is to be expected, because without them, models can only predict the depth from a single
view feature map, rather than estimating depth by matching features between adjacent frames.
We also modified the parameters of our architecture and evaluated its performance, obtained by modifying the number of sampling points K.
By observing the trends in the Table~\ref{tab:cross_attention_ablation}, we can learn that increasing the complexity of the cross-attention network leads to better results.
Also, it is further demonstrated that robust feature matching is beneficial for depth estimation task.
\textbf{Effectness of Motion Field Prediction.}
As shown in Table~\ref{tab:motion_ablation}, utilization of the motion field for warping-based view synthesis significantly improves depth estimate accuracy in dynamic regions.
In addition, we optimize the prediction of motion field by iterative refinement. As shown in Table \ref{tab:iter_num}, the accuracy of dynamic object depth estimation generally improves as the number of repetitions grows, but ultimately approaches saturation.
\section{Conclusion}
This paper proposes a novel attention-based architecture for self-supervised monocular depth estimation in dynamic scenes.
First, we designed a set of attention-based modules to perform feature aggregating across adjacent images efficiently, which has proved to be helpful in improving the overall performance of our network.
Second, we designed an object motion field estimation network to model the dynamic object explicitly without leveraging any semantic prior. The effectiveness of our algorithm has been demonstrated on various autonomous driving datasets.
Future work will consider the use of the proposed method for vision-based SLAM system in dynamic environments.
\bibliographystyle{ieeetr}
|
\section{Introduction}\label{intro}
Black hole X-ray binaries (BHXRBs) are binary systems harboring a stellar-mass black hole as the accretor and a donor star, emitting dominantly in the X-ray band of the electromagnetic spectrum. BHXRBs provide the ideal environment to test the behavior of the matter in strong gravitational potentials around black holes. Most BHXRB systems are known to be transients, spending the big majority of their lives in a quiescent state. As these sources transition through different accretion states throughout the outburst (see, e.g., \citet{Fender04} and \citet{Remillard06}), the contributions from thermal and non-thermal spectral components are observed to evolve which can be traced in a hardness-intensity diagram (HID), transitioning through a "q-shaped" path throughout a full outburst \citep{Fender04, Dunn10} covering the low/hard state (LHS), high/soft state (HSS) and the steep powerlaw or intermediate state. For more on the evolution of these accretion states, see the reviews by \citet{Fender04, Homan05, Remillard06, Done07, Dunn10, Belloni11, Dunn11, Fender12, Fender16}.
GRO J1655-40 was discovered in 1994 \citep{Zhang94} by the Burst and Transient Source Experiment (BATSE) on-board the Compton Gamma Ray Observatory and is one of the most extensively studied Galactic low mass black hole X-ray binaries (LMBXRBs) with a black hole with a mass of ${\rm M_{\mathrm{BH}}}=6.3 \pm 0.5 \: {\rm M_{\odot}}$ and a companion star of ${\rm M_{*}}=2.4 \pm 0.4 \: {\rm M_{\odot}}$ based on optical observations \citep{Orosz97}. Based on optical/NIR observations, however, \citet{Beer02} reported a black hole mass of ${\rm M_{\mathrm{BH}}}=5.4 \pm 0.3 \: {\rm M_{\odot}}$ and a companion star of ${\rm M_{*}}=1.45 \pm 0.35 \: {\rm M_{\odot}}$. The inclination angle has been reported as $\rm i=70.2^{\circ} \pm 1.9^{\circ}$ \citep{Kuulkers00} while \citet{Orosz97, Maccarone02} suggested a difference of $15^{\circ}$ between the binary orbital plane inclination angle and the jet inclination angle based on radio observations. We adopt a distance of d$\rm \sim3.2 \pm 0.2$ kpc based on \citet{Hjellming95}, while \citet{Foellmi06} argued that it should be lower than 1.7 kpc. Different methods to measure the spin on GRO J1655-40 revealed a controversy, with $\rm 0.65< a_{*} <0.75$ obtained from spectral continuum fitting with the Advanced Satellite for Cosmology and Astrophysics (ASCA) and the Rossi X-Ray Timing Explorer (RXTE) observations \citep{Shafee06} while Fe emission line profile predicts much higher values, with a lower limit on the spin at $\rm a_{*}=0.9$ \citep{Reis09} based on XMM-Newton observations. On the contrary to these measurements pointing towards a high spin case for GRO J1655-40, timing analysis of the entire set of RXTE data using the relativistic precession model \citep{Stella98, Stella99} resulted in a much smaller spin parameter $\rm a_{*} = 0.290 \pm 0.003$ \citep{Motta14}.
The X-ray continuum fitting has been used extensively to measure the black hole spin and depends on accurately locating the innermost stable circular orbit (ISCO) by modeling the thermal component of the X-ray spectral continuum. In this method, it's a common practice to focus primarily on the HSS to measure the location of $\rm R_{{ISCO}}$ which is determined by the black hole spin and mass as follows
\begin{equation}\label{r_isco}
\rm R_{ISCO}=R_{g}\left\{3+A_{2} \pm\left[\left(3-A_{1}\right)\left(3+A_{1}+2 A_{2}\right)\right]^{1 / 2}\right\}
\end{equation}
\begin{equation}\label{A_1}
\rm A_{1}=1+\left(1-a_{*}^{2}\right)^{1 / 3}\left[\left(1+a_{*}\right)^{1 / 3}+\left(1-a_{*}\right)^{1 / 3}\right]
\end{equation}
\begin{equation}\label{A_2}
\rm A_{2}= \left(3 a_{*}^{2}+A_{1}^{2}\right)^{1 / 2}
\end{equation}
\begin{equation}\label{A_2}
\rm a_{*}=\frac{|\mathbf{J}| c}{G M^{2}}
\end{equation}
where $\rm J$ is the black hole angular momentum, $\rm G$ is the gravitational constant, $\rm c$ is the speed of light, $\rm M$ is the black hole mass, $\rm R_{g}= G M / c^{2}$ is the gravitational radius and the upper and lower signs describe a prograde disk ($\rm a_{*}>0$) and a retrograde disk ($\rm a_{*}<0$), respectively.
In this paper, we test the most widely used relativistic disk models \texttt{KERRBB} and \texttt{KYNBB} by performing an extensive spectral analysis using all of the publicly available RXTE observations during the 2005 outburst of GRO J1655-40. The structure of this paper is as follows: in section \ref{data}, we summarize the RXTE data reduction, provide a review of the relativistic spectral modeling of accretion disks in black hole X-ray binaries and summarize the spectral analysis procedure. In section \ref{results}, we present the results of our analysis and summarize our conclusions in section \ref{conclusions}.
\section{Data Reduction and Analysis}\label{data}
\subsection{RXTE Observations Data Reduction}
RXTE observed GRO J1655-40 extensively throughout the entire mission spanning multiple outbursts between 1996-2011. For the spectral fitting, we used a total of more than 502 observations from the 2005 outburst of GRO J1655-40 \citep{Markwardt05} publicly available on HEASARC (High Energy Astrophysics Science Archive Research Center) archive\footnote{\url{https://heasarc.gsfc.nasa.gov/docs/archive.html}}.
We followed the standard procedure in the RXTE cookbook\footnote{\url{https://heasarc.gsfc.nasa.gov/docs/xte/recipes/cook_book.html}} using \texttt{FTOOLS/HEASOFT} version 6.29\footnote{\url{https://heasarc.gsfc.nasa.gov/docs/software/lheasoft/}} software package for data reduction. We only used the Proportional Counter Unit-2 (PCU-2) of RXTE’s Proportional Counter Array (PCA, \citet{Jahoda06}) with all of its layers included as it was almost always operating throughout the entire RXTE mission and has the best calibration among other PCUs on-board PCA. We removed data lying within 10 minutes of the South Atlantic Anomaly and extracted the source spectra in the \texttt{standard2f} mode which provides the optimal spectral resolution and generated response matrices and background spectra. We then made use of the publicly available tool \texttt{pcacorr} for calibration and applied an additional 0.1\% systematic errors to account for uncertainties in the telescope’s response following \citet{Garcia14}.
\subsection{Relativistic Spectral Modeling of the Accretion Disks in Black Hole X-ray Binaries}\label{modeling}
Proper spectral modeling of the observed thermal spectrum is crucial to developing a better understanding of the inner structure of the accretion disk and the nature of the accretion flow in the presence of strong gravitational potentials around black holes. It is, therefore, important to include a detailed treatment of relativistic effects on the thermal emission from the innermost region of the accretion disk.
To this date, there have been several attempts to include relativistic accretion disk models in \texttt{XSPEC}. \texttt{KERRBB} \citep{Li05} and \texttt{KYNBB} \citep{Dovciak04} played a significant role in the X-ray spectral continuum method to determine the black hole spin. Both \texttt{KERRBB} and \texttt{KYNBB} assume a \citet{Novikov73} geometrically thin, optically thick accretion disk where orbits around the black hole are Keplerian. Both rely on the ray-tracing technique to compute the spectrum of accretion disks around black holes where the local flux is calculated as a sum of black body radiation with the \citet{Novikov73} radial temperature profile. The space-time around the black hole is described by the Kerr metric and both models are constructed to include relativistic effects on the accretion disk due to strong gravitational fields around black holes (i.e. light bending, self-irradiation or the returning radiation, gravitational redshift, frame dragging and Doppler boost) while computing the observed spectrum.
Additional to the relativistic effects, \texttt{KERRBB} and \texttt{KYNBB} also account for any deviations of the observed spectrum from the perfect blackbody emission assumed by non-relativistic models like \texttt{DISKBB} and \texttt{DISKPBB} by introducing the color correction factor $\rm f_{col} = T_{col} / T_{eff}$ where $\rm T_{col}$ is the color temperature and $\rm T_{eff}$ is the effective temperature of the accretion disk. This deviation in the observed color temperature becomes prominent when the gas temperature at the surface of the accretion disk is high enough such that the electron scatterings start to dominate over absorptive processes taking place in the disk atmosphere, usually corresponding to photon energies > 1.0 keV \citep{Shakura73}.
\texttt{KYNBB} is a local model that was developed as an extension to the original \texttt{KY} package\footnote{\url{https://projects.asu.cas.cz/stronggravity/kyn}}. While \texttt{KERRBB} describes an accretion disk with an inner edge extending down to the (ISCO), \texttt{KYNBB} offers an option to have the inner edge either as a free parameter or fixed at a certain value other than $\rm R_{ISCO}$. \texttt{KYNBB} also has multiple options one can employ to include the polarization calculations and/or calculate the spectrum only from a selected section of the disk but these additional options are turned off for the entirety of our spectral analysis. For more details about the other model parameters not listed in TABLE \ref{table:model_pars}, see \url{https://projects.asu.cas.cz/stronggravity/kyn}.
\subsection{Spectral Analysis}\label{spectral_analysis}
We used the \texttt{HEASOFT} (v.6.29) package \texttt{XSPEC (v.12.12)} \citep{Arnaud96} and \texttt{PyXspec}\footnote{\url{https://heasarc.gsfc.nasa.gov/xanadu/xspec/python/html/}}, a Python interface to the \texttt{XSPEC} for spectral analysis. To apply $\rm \chi^{2}$ statistics, we grouped spectra to have 25 counts per bin using the \texttt{FTOOLS} task \texttt{grppha} and restricted the energy range to 3-25 keV for all of the PCA spectra. All of the models used for the spectral analysis consist of a thermal disk component (\texttt{KERRBB} or \texttt{KYNBB}) and a simple powerlaw component accounting for the non-thermal component of the continuum (\texttt{POWERLAW} in \texttt{XSPEC}). We use \texttt{TBABS} to account for the ISM absorption. Throughout this paper, we refer to each model (\texttt{TBABS*(KERRBB+POWERLAW))} and \texttt{TBABS*(KYNBB+POWERLAW)}) by the specific disk model used for the analysis. We summarize all of the parameters and their values used in this paper in Table \ref{table:model_pars}. Since both disk models do not have the disk temperature as a model parameter, we evaluated the position of the peak of the disk component in both models in order to obtain the value of the disk temperature. While this peak corresponds to a disk radius not exactly at $\rm R_{ISCO}$ but slightly further out due to the zero-torque boundary condition adopted by \texttt{KERRBB} at the innermost disk radius, this is still a valid approximation of the accretion disk temperature as measured by \texttt{KERRBB}.
\begin{table}[h!]
\centering
\caption{Model parameter values for \texttt{KERRBB} and \texttt{KYNBB}}
\label{table:model_pars}
\begin{tabular}{p{0.14\textwidth}p{0.14\textwidth}p{0.14\textwidth}}
\hline \hline
Model Parameter & Value & Free/Frozen \\
\hline
&\texttt{KERRBB}\\
\hline
$\eta$ & 0 & Frozen \\
Black Hole Spin & -- & Free \\
Inclination Angle & 70.2 & Frozen \\
Black Hole Mass & 6.3 & Frozen \\
Mass Accretion Rate & -- & Free \\
Distance & 3.2 & Frozen \\
$\rm f_{col}$ & 1.7 & Frozen\\
Self-irradiation Flag & 0 & Frozen \\
Limb-darkening Flag & 0 & Frozen \\
Normalization & 1 & Frozen\\
\hline
&\texttt{KYNBB}\\
\hline
Black Hole Spin & -- & Free \\
Inclination Angle & 70.2 & Frozen \\
Inner Radius & -- & See Section \ref{spectral_analysis} \\
Outer Radius & 1000 & Frozen \\
Black Hole Mass & 6.3 & Frozen \\
Mass Accretion Rate & -- & Frozen (see Section \ref{spectral_analysis} )\\
$\rm f_{col}$ & 1.7 & Frozen \\
Normalization & 9.76 & Frozen\\
\hline
\hline \hline
\end{tabular}
\end{table}
We used disk fractions to distinguish observations based on how much the disk component is dominating over powerlaw following \citet{Kording06, Dunn08, Dunn10} with the disk fraction defined as \\
\begin{equation}\label{disk_fraction}
\rm F_{Disk}=\frac{L_{Disk} \;(0.001-100.0 \; keV)}{ L_{Disk} (0.001-100.0 \; keV)+L_{Power} (1.0-100.0 \; keV)}
\end{equation}\\
where $\rm L_{Disk}$ and $\rm L_{Power}$ are the unabsorbed disk luminosity in 0.001-100.0 keV and unabsorbed powerlaw luminosity in 1.0-100.0 keV energy range, respectively. We used the convolution model \texttt{CFLUX} in \texttt{XSPEC} for flux calculations of each component.
We first analyzed the entire set of observations using \texttt{KERRBB} with black hole spin and accretion rate as free parameters. For further analysis, we adopted a selection criteria by assessing reduced $\rm \chi^{2}$ ($\rm \chi^{2}$/d.o.f.) values and excluded the observations with reduced $\rm \chi^{2}>2.0$ to make sure that a simple model consisting of thermal (\texttt{KERRBB} or \texttt{KYNBB}) and non-thermal (\texttt{POWERLAW}) components was able to fit the spectra well to obtain reasonable values for the physical parameters of the system. Then, using the sample after the reduced $\rm \chi^{2}$ selection criteria applied, we set up \texttt{KYNBB} to match \texttt{KERRBB} by fixing the inner disk radius at $\rm R_{ISCO}$ instead of leaving it as a free and set the accretion rate parameter in \texttt{KYNBB}, the photon index and the normalization of \texttt{POWERLAW} in \texttt{TBABS*(KYNBB+POWERLAW)} to the values obtained by \texttt{TBABS*(KERRBB+POWERLAW)}, leaving the spin as the only free parameter remaining in the model for the spectral fitting.
\section{Results}\label{results}
We present the comparison between the two models in FIGURE \ref{fig:corner-kynbb-kerrbb} and the evolution of other parameters of the system throughout the outburst. We first attempted to fix the spin at $\rm a_{*}=0.7$ following the results from the X-ray spectral continuum fitting method by \citet{Shafee06}. This assumed spin value resulted in significantly worse fit statistics for $\sim$ 89\% of observations which produced reduced $\rm \chi^{2}$ values much larger than 2.0. We tested how \texttt{KERRBB} behaved with different spin values and we did not observe the model favoring a specific value of spin over all possible values. These measured spin values were observed to span a parameter space between $\rm 0.52<a_{*}<0.94$ which corresponds to values that cannot be explained only by uncertainties introduced by the fitting procedure. This behavior in measured spin can be interpreted as a result of the notable evolution of the innermost edge of the accretion disk throughout the entire outburst of GRO J1655-40 (Yilmaz et al. 2022, in preparation).
We compared the two models using calculated disk fractions (Equation \ref{disk_fraction}), disk luminosity $\rm L_{Disk} / \rm L_{Edd}$, black hole spin and temperature values by setting the inner edge of the accretion disk parameter in \texttt{KYNBB} at $\rm R_{ISCO}$. We used the spin values measured by both \texttt{KERRBB} and \texttt{KYNBB} to show the evolution of $\rm R_{in}$ calculated using Equation \ref{r_isco} with respect to other parameters. The change in $\rm R_{in}$ with respect to the disk temperature has a clear trend indicating a decrease of $\rm R_{in}$ with increasing temperature values. This non-linear trend in the evolution of $\rm R_{in}$ was not observed with respect to $\mathrm{L_{Disk}} / \mathrm{L_{Edd}}$. There's a clear overturn in $\rm R_{in}$ values at $\rm R_{in}\sim 3 \; R_{g} $ and $\mathrm{L_{Disk}} / \mathrm{L_{Edd}} \sim \rm 0.08$, after which $\rm R_{in}$ starts to decrease with decreasing $\mathrm{L_{Disk}} / \mathrm{L_{Edd}}$. This behavior after the turning point contradicts the expected evolution of the inner edge of the accretion disk with respect to decreasing mass accretion rates where the disk is expected to be truncated at larger radii compared to higher mass accretion rates. This observed deviation in $\rm R_{in}$ is similar to the previously reported ${L_{\rm {Disk}}}$ - $T$ relations observed for many BHXRBs by \citet{Gier04, Done07, Dunn08,Dunn10, Dunn11}.
On average, both models produced consistent results for the black hole spin (4.5\% difference on average), disk temperature, disk fraction and disk luminosity for the same accretion rates assuming $\rm R_{in}=\rm R_{ISCO}$. There's an anti-correlation between the disk luminosity $\mathrm{L_{Disk}} / \mathrm{L_{Edd}}$ and the disk fraction. This anti-correlation contradicts the expected relation between these parameters. With increasing luminosity, the disk component is expected to start to dominate over the non-thermal component which results in increasing disk fraction (Equation \ref{disk_fraction}). Throughout the 2005 outburst, GRO J1655-40 was observed to enter an unusually soft or hyper-soft state \citep{Uttley15}. In this unexpected state, a strong accretion disc wind was detected with Chandra X-ray HETG \citep{Miller08}. This sudden suppression of the non-thermal component could be explained by this Compton-thick accretion disk wind and it was even suggested that it might indicate intrinsic luminosities above the Eddington limit due to such high obscuration \citep{Neilsen16}. On the other hand, with increasing luminosity from the LHS, the source enters a state where the luminosity is very high but a steep-powerlaw starts to dominate over the thermal component resulting in an inverse relationship between the disk luminosity and the disk fractions calculated from the unabsorbed components of the total model.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{corner-kynbb-kerrbb.png}
\caption{\textit{Top}: Corner plot presenting parameters (disk luminosity $\mathrm{L_{Disk}} / \mathrm{L_{Edd}}$, inner edge of the accretion disk $\rm R_{in}$, disk fraction (see Equation\ref{disk_fraction}) and disk temperature) for fits comparing \texttt{KERRBB} with \texttt{KYNBB} with $\rm R_{in} = \rm R_{ISCO} $.}
\label{fig:corner-kynbb-kerrbb}
\textit{}
\end{figure*}
\section{Conclusions}\label{conclusions}
Our analysis showed that both \texttt{KERRBB} and \texttt{KYNBB} provide identical results when $\mathrm{R_{in}}=\mathrm{R_{ISCO}}$ is assumed. Fitting results, shown in FIGURE \ref{fig:corner-kynbb-kerrbb}, show only small statistical fluctuations in some cases which can be explained by small differences in the reduced $\rm \chi^{2}$ values corresponding to different local minima in \texttt{XSPEC}.
Our analysis showed that a variable black hole spin parameter was necessary to obtain reasonable fit results for the entire 2005 outburst of GRO J1655-40. This behavior in the spin parameter could be interpreted as an actual change in the innermost edge of the accretion disk, possibly not even extending down to $\mathrm{R_{in}}$. While the $\mathrm{R_{in}}=\mathrm{R_{ISCO}}$ assumption in \texttt{KERRBB} has been widely adopted for the X-ray spectral continuum fitting method for over a decade, a variable inner edge of the disk might suggest that this standard assumption might not hold true throughout the entire set of different accretion states. While it is accepted that the inner edge of the accretion disk extends down to $\mathrm{R_{ISCO}}$ in the highest luminosity state, we still observed a notable variance in calculated $\mathrm{R_{in}}$ values. This variability in the innermost edge of the accretion disk, $\mathrm{R_{in}}$, will be further investigated in a follow-up work.
\backmatter
\section*{Acknowledgments}
A.Y., J.S., and P.G.B. acknowledge financial support from the \fundingAgency{Czech Science Foundation} under Project No. \fundingNumber{19-05599Y}. M.B. and M.D. thank for the support
from the \fundingAgency{Czech Science Foundation} under Project No. \fundingNumber{21-06825X}.
A.Y., P.G.B., M.B., M.D., and J.S. also acknowledge the institutional support from \fundingAgency{Astronomical Institute of the Czech Academy of Sciences} with \fundingNumber{RVO:6798581}.
|
\section{Introduction}
The Parker Solar Probe (PSP) mission, launched in 2018 August, is intended to dive below the Alfv\'en surface and make direct in situ measurements of the source regions of the solar wind for the first time \citep{fox2016}. One of the striking features in the observations from the very first encounter is the prevalence of Alfv\'enic flows with reversed magnetic field and enhanced radial velocity, called switchbacks \citep[SBs;][]{bale2019, kasper2019}. Similar structures have been observed in the solar wind by earlier missions but with much reduced frequency \citep[e.g.,][]{balogh1999, gosling2009, horbury2018}. SBs tend to occur in patches, i.e., bursts of the structure separated by quieter intervals \citep[e.g.,][]{bale2021, fargette2021}, so they appear to be modulated. The origin of SBs is highly debated. Models have been proposed to explain the generation and evolution of SBs, including interchange magnetic reconnection \citep[e.g.,][]{fisk2020, zank2020, drake2021}, velocity shear and footpoint motion \citep{landi2006, ruffolo2020, schwadron2021}, expanding waves and turbulence \citep[e.g.,][]{squire2020, shoda2021}, and coronal jets \citep[e.g.,][]{sterling2020}. No sufficient evidence has been provided so far to distinguish the theories.
The Alfv\'en surface marks the critical transition in the solar wind radial velocity from sub-Alfv\'enic to super-Alfv\'enic \citep{weber1967}. The topological nature of this transition is a key issue but remains unclear. \citet{liu2021} estimate the average Alfv\'en radius to be around 10 solar radii from PSP measurements at the first four encounters. They observe a speed dependence in the Alfv\'en radius, which suggests a ``rugged" Alfv\'en surface around the Sun. Similar results are obtained by \citet{verscharen2021}. Based on MHD simulations coupled to a turbulent transport model, \citet{chhiber2022} argue that the transition may be better described as an extended, fragmented zone rather than a wrinkled surface, i.e., a region with mixed parcels of sub-Alfv\'enic and super-Alfv\'enic wind. Again, evidence is needed to distinguish this hypothesis from that of a wrinkled surface.
As another accomplishment, PSP entered a sub-Alfv\'enic solar wind region for about 5 hr around a distance of about 19 solar radii at the eighth encounter on 2021 April 28 \citep{kasper2021}. This first glimpse of the sub-Alfv\'enic wind implies a surprisingly large Alfv\'en radius, which is over 20 solar radii. However, the solar wind density within all the first several sub-Alfv\'enic intervals is unusually low, which is a main factor contributing to the crossings of the Alfv\'enic transition \citep{kasper2021}. This indicates that the observed sub-Alfv\'enic wind is not a typical wind. At the subsequent encounters, crossings of the Alfv\'enic transition have been made on several more occasions \citep[e.g.,][]{bandyopadhyay2022}. Measurements in those sub-Alfv\'enic intervals show predominantly outward propagating Alfv\'enic fluctuations \citep[e.g.,][]{zank2022} and much weaker, fewer SBs than in the super-Alfv\'enic wind \citep[e.g.,][]{kasper2021, bandyopadhyay2022}. The nature of the special, sub-Alfv\'enic wind remains to be investigated, and an intriguing question is what it reveals concerning the morphology of the Alfv\'enic transition and the generation of SBs.
This paper has a threefold focus. First, we identify a special structure in the pristine solar wind, defined as a low Mach-number boundary layer (LMBL), and determine its typical signatures and source region on the Sun. Second, we illustrate that the sub-Alfv\'enic wind observed by PSP is an LMBL flow by nature, and the similar origin and similar properties of the sub-Alfv\'enic LMBLs suggest a wrinkled surface for the morphology of the Alfv\'enic transition. Third, we show how the properties of SBs and those of the LMBLs themselves help clarify the generation and evolution of SBs as well as their modulation. Data and methodology are described in Section 2. Details on LMBLs and their interpretations are given in Section 3. We then discuss in Section 4 important implications on the morphology and crossings of the Alfv\'enic transition, the generation and evolution of SBs, and the SB modulation. The conclusions are summarized in Section 5.
\section{Data, Methodology and Speculations}
We use measurements from the FIELDS instrument suite \citep{bale2016} and the SWEAP package \citep{kasper2016} aboard PSP. The SWEAP ion instruments include an electrostatic analyzer (SPAN-I) on the ram side of PSP, and a Faraday Cup (SPC) that looks directly at the Sun \citep{kasper2016, case2020, whittlesey2020}. With their combined fields of view, SPAN-I and SPC are able to provide plasma parameters of the solar wind throughout key encounter phases of the mission. Parameters of the alpha particles are also available starting from the fourth encounter, as PSP moves fast enough to shift the majority of the particle velocity distribution core into the field of view of SPAN-I. Because of the partial moments in the plasma measurements, the electron density from quasi-thermal noise (QTN) spectroscopy \citep{moncuquet2020} is used as a proxy of the plasma density. This density is considered to be most reliable, since it is derived from measurements of the local plasma frequency. All the solar wind parameters from FIELDS and SWEAP are interpolated to 2-s, evenly spaced averages.
\subsection{Derivation of Deflection Angle}
The deflection angle of the magnetic field is defined as the deviation angle from the average background field direction, which is assumed to be radial or anti-radial depending on the field polarity. This assumption is valid at PSP encounter distances, because at those distances the Parker spiral field is indeed almost radial or anti-radial. In RTN coordinates, for an overall positive field (i.e., $\langle{B_R}\rangle > 0$) the deflection angle can be expresses as
\begin{equation}
\theta = \left\lbrace \begin{array}{ll}
\arccos(B_R/B),~~~~~~~~~~~~~~~~~B_T>0, \\
-\arccos(B_R/B),~~~~~~~~~~~~~~B_T<0.
\end{array} \right.
\end{equation}
For an overall negative field (i.e., $\langle{B_R}\rangle < 0$) the deflection angle is written as
\begin{equation}
\theta = \left\lbrace \begin{array}{ll}
-\arccos(-B_R/B),~~~~~~~~~~~~~B_T>0, \\
\arccos(-B_R/B),~~~~~~~~~~~~~~~~B_T<0.
\end{array} \right.
\end{equation}
An advantage of such a definition is that the sign of $\theta$ indicates the deflection direction. A positive angle ($0<\theta\leqslant\pi$) corresponds to a counterclockwise deflection in the RT plane as viewed from the north, while a negative one ($-\pi\leqslant\theta<0$) a clockwise deflection. It takes a value of $\pi$ or $-\pi$ when the magnetic field is completely reversed. This definition also allows to set up a relationship with the enhancement in the solar wind radial velocity, as will be shown below. Note that our definition should not be confused with the magnetic field cone angle, the angle between the local magnetic field vector and the radial direction \citep[e.g.,][]{horbury2018, kasper2019, woolley2020} that does not provide information on the deflection direction.
An SB by name implies a deflection angle larger than $90^{\circ}$. To look at how the phenomenon forms and evolves, however, we need to include all possible deflection angles, not just those more than $90^{\circ}$ \citep[e.g.,][]{mozer2021, tenerani2021}. Indeed, \citet{wit2020} find a continuum of deflections with no particular threshold that can be used to define an SB. As will be shown below, those small deflections hold important clues on the origin of the phenomenon in the solar atmosphere (below the Alfv\'en critical point), while those large ones indicate key information on more evolved states of the phenomenon at greater distances. We suggest that the terminology is better changed to Alfv\'enic flows with deflected magnetic field and enhanced radial velocity, or Alfv\'enic deflections for short. Section~2.2 provides more support for this statement.
\subsection{Velocity Enhancement and Relationship with Deflection Angle}
For a magnetic field with magnitude $B$ deflected from its original direction (radial or anti-radial depending on the field polarity) by an angle $\theta$, the change in the radial component is $\delta{B_R} = B_{R0}\cos\theta - B_{R0}$, where $B_{R0}$ is the initial radial component of the field. Obviously, $B=|B_{R0}|$ under our assumption. For Alfv\'enic fluctuations, the change in the radial velocity is related to the change in the radial field by $\delta{v_R}/v_{\rm A} = \pm \delta{B_R}/B$ \citep[e.g.,][]{belcher1971}, where $v_{\rm A}$ is the local Alfv\'en speed. Outward propagating Alfv\'enic fluctuations yield a negative sign for the above equation if the background field is positive, and a positive sign if the background field is negative. It can be readily shown that
\begin{equation}
{\delta{v_R} \over v_{\rm A}} = 1 - \cos\theta
\end{equation}
for outward propagating Alfv\'enic fluctuations. This simple analytical development indicates that any deflection of the field would always be associated with an enhancement in the solar wind radial velocity (i.e., $\delta{v_R}>0$). It explains the one-sided nature of the velocity spikes observed within SBs \citep[e.g.,][]{gosling2009, horbury2018, kasper2019}, which is simply a result of outward propagating Alfv\'enic fluctuations. At PSP encounter distances the mean field is almost radial or anti-radial, so the effect is amplified in the encounter measurements compared with those at greater distances where the field has larger spiral angles. This predicts that considerable radial velocity spikes in SBs observed at 1 AU and beyond may correspond to a background field geometry with a large radial component.
Another interesting prediction is that below the Alfv\'en critical point the velocity enhancement and thus the deflection angle must be significantly reduced. This can be shown with the following simple reasoning. Equation~(3) indicates that the velocity is always enhanced and the enhancement can be a large fraction of the local Alfv\'en speed. If it were not reduced below the Alfv\'en critical point, the enhancement would easily lead to a radial velocity higher than the local Alfv\'en speed, which conflicts with the condition of $v_R < v_{\rm A}$. A prerequisite is that Alfv\'enicity still holds for the sub-Alfv\'enic wind, which is indeed the case \citep{zank2022}.
One may expect from Equation~(3) an enhancement of $v_{\rm A}$ in the radial velocity for a $90^{\circ}$ deflection and 2$v_{\rm A}$ for a complete field reversal. Note that, however, the background magnetic field at PSP encounter distances is not exactly radial or anti-radial (i.e., $B \geqslant |B_{R0}|$), so the equation is an upper limit for the velocity enhancement. To quantify the velocity enhancement and how it compares with Equation~(3), we derive the velocity enhancement by applying a low-pass Butterworth filter to the observed radial velocities with a cutoff frequency of 2$\times$$10^{-4}$ Hz. This cutoff frequency is well below the break frequency (about 2$\times$$10^{-3}$ Hz) separating the energy-containing range and inertial range in the power spectrum of solar wind fluctuations \citep[e.g.,][]{kasper2021, zank2022}. Velocity variations with frequencies higher than the cutoff frequency are removed in the filtering. The radial velocity enhancement can then be calculated using $\delta{v_R} = v_R - v_{Rf}$, where $v_{Rf}$ is the filtered value and serves as the ``baseline" for the solar wind radial velocity.
When writing this paper, we notice a similar model proposed by \citet{matteini2014} to explain the correlation between the solar wind speed and the local magnetic field orientation in fast, polar coronal hole wind observed by Ulysses. They also suggest that SBs may naturally arise from Alfv\'enic fluctuations in the context of their model and the observed correlation. Our independent development of Equation~(3) closely parallels their work. New important information will be revealed by looking at how the velocity enhancement varies with the Alfv\'en Mach number as well as the deflection angle. These may constitute a more complete explanation for the generation and evolution of SBs.
\subsection{Estimate of Alfv\'en Radius}
The derivation of the Alfv\'en radius ($r_{\rm A}$) and its uncertainties are described in \citet{liu2021}. Here we rewrite the expression in RTN coordinates
\begin{equation}
r_{\rm A} = \sqrt{v_R \over v_{R\rm A}}{r \over M_{\rm A}},
\end{equation}
where $r$ is the heliocentric distance, $v_{R\rm A}$ the radial velocity at the Alfv\'en critical point (i.e., $r_{\rm A}$), and $M_{\rm A}$ the radial Alfv\'en Mach number. In the classic theory of \citet{weber1967}, the radial Alfv\'en Mach number is defined as $M_{\rm A} = v_R\sqrt{\mu\rho}/|B_R|$, where $\rho$ is the mass density of the solar wind and $\mu$ the permeability constant. We realize that SBs may give rise to a singularity for a deflection around $90^{\circ}$, since $|B_R|=B\cos\theta$. Indeed, we see large spikes in $M_{\rm A}$ \citep{liu2021}. While this could be useful for identifying SBs, we redefine it as $M_{\rm A} = v_R\sqrt{\mu\rho}/B$ taking advantage of $B \simeq |B_R|$ in the background solar wind at PSP encounter distances.
A reasonable approximation is obtained by assuming that $v_R$ does not change much from $r_{\rm A}$ to PSP encounter distances, i.e.,
\begin{equation}
r_{\rm A} \simeq {r \over M_{\rm A}}.
\end{equation}
\citet{weber1967} demonstrate that this approximation, while simple, provides a rigorous estimate of the Alfv\'en radius. This is particularly true for PSP encounter measurements as they are made close to the Sun. The estimate should be considered as a lower limit when PSP is outside $r_{\rm A}$ (i.e., $v_R > v_{R\rm A}$), and an upper limit when inside (i.e., $v_R < v_{R\rm A}$). Readers are directed to \citet{liu2021} for more detailed discussions on the uncertainties of the estimate.
\subsection{Determination of Solar Source}
The magnetic connectivity between PSP and the photosphere is established with a potential field source surface (PFSS) model plus a simple ballistic projection from the location of the spacecraft to the source surface \citep[e.g.,][]{altschuler1969, schatten1969, wang1992, badman2020}. An ideal Parker spiral is assumed for the magnetic field line connecting PSP to the source surface with curvature determined by the solar wind radial velocity observed at PSP. Below the source surface the coronal magnetic fields are constructed using the PFSS model, based on the Air Force Data Assimilative Photospheric Flux Transport (ADAPT) magnetograms provided by the Global Oscillation Network Group (GONG). The ADAPT-GONG synoptic map is updated every two hours. We select a map when the PFSS model becomes stable with time. For encounter 2, which is included in our analysis, the PFSS model results vary significantly between the ADAPT-GONG maps because of the emergence of an active region on the far side of the Sun \citep{wallace2022}. The PFSS model for encounter 2 is thus developed using a magnetogram recorded on 2019 April 12 when the far-side active region is incorporated into the map.
The height of the source surface is usually set to 2.5 solar radii ($R_S$), at which the field lines are forced to be radial. The PFSS modeling gives areas of open fields, which can be compared with EUV imaging observations of coronal holes. This is how we select a value for the height of the source surface. In this study we use EUV synoptic maps from SDO 193 \AA\ observations for such a comparison (available at https://solarflare.njit.edu). The magnetic mapping from the PSP location to the source surface and then to the photosphere suffers different sources of uncertainties, including the radial variation of the solar wind speed and the magnetogram used as the boundary condition. The uncertainties are difficult to quantify, but some of the effects may counteract each other \citep[see more discussions in][]{badman2020, kasper2021, pinto2021}. Our experience with the mapping indicates that the position error on the photosphere is at least a few degrees \citep{chen2021, meng2022}.
\section{Examples of LMBLs and Interpretations}
Figure~1 shows the in situ measurements near perihelion of encounter 2, where a transition from slow to relatively fast wind is observed. We see a clear drop in the radial Alfv\'en Mach number associated with this transition. The boundary of the LMBL is mainly determined from the decrease in $M_{\rm A}$. The primary contribution to this drop is the significantly reduced density plus the relatively low speed. The Alfv\'en radius increases to more than 20 solar radii. If the altitude of PSP were as low as 20 $R_S$, we would see a crossing of the Alfv\'enic transition as early as encounter 2. \citet{liu2021} indicate conditions of reduced densities and low radial velocities that favor crossings of the Alfv\'enic transition. The crossings starting from encounter 8 confirm this prediction \citep[][and see data below]{kasper2021}. Note that we scale the density and magnetic field to values at 1 AU in order to eliminate their radial variations associated with the change in the distance of the spacecraft. The magnetic field strength is roughly constant across the slow-to-fast wind transition. \citet{rouillard2020} examine the same time period and suggest that PSP was measuring high-density streamer flows when it was connected to streamers and tenuous wind when it was not. The low density and steadily increasing velocity inside the LMBL indicate that PSP was likely sampling open magnetic fields from a coronal hole near its boundary, and the even higher velocity and persistent low density after the LMBL are suggestive of deeper penetration into the coronal hole wind. Our magnetic mapping will provide further evidence for this interpretation.
An important thing to note is that the amplitudes of SBs within the LMBL are considerably reduced. A first impression of this reduction is from the variations of $B_{R}$ (Figure~1d). Quantitative results about the reduction are given by the deflection angle of the magnetic field and its hourly rate of over $30^{\circ}$. The solar wind radial velocity exhibits largely one-sided variations (i.e., enhancement) compared with its ``baseline" from the low-pass filtering. Inside the LMBL, the velocity enhancement in Figure~1c does not show an obvious decrease, but when it is calculated in units of the local Alfv\'en speed a clear reduction is seen (Figure~1h). These are consistent with the correlation between the deflection angle and velocity enhancement predicted by Equation~(3). In general, $\delta{v_R} \leqslant v_{\rm A}$ as indicated by Figure~1h (see more discussions in Section~4). Also note that the LMBL has fine structures: whenever $M_{\rm A}$ drops we see a reduction in SBs. This seems true for data outside the LMBL as well. On average, the slower wind ahead of the LMBL has larger deflection angles than the faster wind behind the LMBL, which is probably due to the higher $M_{\rm A}$ of the slower wind. Therefore, SBs are better sorted by $M_{\rm A}$ rather than the velocity \citep[also see][]{mozer2021}. The deflection angle before the LMBL is predominantly negative, indicative of a preferential clockwise rotation in the magnetic field associated with the streamer flows. This agrees with the finding of \citet{meng2022}.
Figure~2 presents the in situ measurements near perihelion of encounter 8. Compared with the case in Figure~1, an opposite transition is observed, i.e., from relatively fast, tenuous to slow, dense wind. PSP was likely transiting from a deeper immersion in a coronal hole wind before the LMBL, to the stream out of the coronal hole boundary in the LMBL, and then to streamer flows after the LMBL. Indeed, we see crossings of the heliospheric current sheet (HCS) embedded in the high-density wind. Significant decreases in the Alfv\'en radius are observed around the HCS crossings, which is predicted by MHD simulations \citep{chhiber2022}. The alpha particles show a radial velocity higher than that of the protons for most of the wind inside the LMBL. This is a typical signature of fast wind \citep[e.g.,][]{asbridge1976, marsch1982}, although the solar wind is slow in our case. The characteristics of being slow while resembling fast wind provide support for our interpretation about the origin of an LMBL, i.e., boundary wind from inside a coronal hole.
The situation is in general akin to what is shown in Figure~1, such as associations of the LMBL with decreased $M_{\rm A}$, increased $r_{\rm A}$ and reduced SBs. In the current case, the Alfv\'en radius rises above the distance of PSP, which leads to the first glimpse of the sub-Alfv\'enic wind \citep{kasper2021}. The fact that an LMBL has a large Alfv\'en radius enabling an easier crossing of the Alfv\'enic transition suggests that the sub-Alfv\'enic wind observed by PSP so far is an LMBL flow by nature. More specifically, the sub-Alfv\'enic wind is a special type of wind flowing out along open magnetic fields from a coronal hole near its boundary. Note that in the sub-Alfv\'enic wind SBs (or Alfv\'enic deflections in our terminology) do not disappear; they are just reduced because of the low $M_{\rm A}$. This observation is consistent with our speculation about the behavior of SBs below the Alfv\'en critical point based on Equation~(3).
Figure~3 displays the in situ measurements near perihelion of encounter 9. Again, we see a transition from relatively fast, tenuous to slow, dense wind. During the transition in the tenuous wind, PSP was likely penetrating deeper inside a coronal hole wind (i.e., out of the boundary wind) temporarily. This is why two LMBLs are identified and relatively fast wind with an SB patch is seen between the two LMBLs. The solar wind velocity is also enhanced behind the second LMBL. This is not a coronal hole wind, but reconnection outflow around the HCS crossing \citep[e.g.,][]{phan2020, chen2021}. Inside the two LMBLs, the radial velocity of the alpha particles is considerably higher than that of the protons, despite the fact that the solar wind is slow. This, once more, supports our interpretation about the origin and nature of an LMBL.
Like the picture that we have seen in Figures~1 and 2, the two LMBLs are associated with decreased $M_{\rm A}$, increased $r_{\rm A}$ and reduced SBs. Again, the sub-Alfv\'enic wind in the present case is an LMBL flow by nature. Interestingly, here we can compare the case when PSP was ``surfing" around the Alfv\'enic transition (i.e., $M_{\rm A} \simeq 1$ inside LMBL1) with the case when PSP was ``diving" well below it (i.e., $M_{\rm A} < 1$ inside LMBL2). The SBs (better called Alfv\'enic deflections) within LMBL2 are further reduced compared with those inside LMBL1 (Figure~3d, h). The magnetic field in LMBL2, however, still shows an indication of deflections (Figure~3d), so the magnetic deflections have an origin well below the Alfv\'en critical point. They develop when $M_{\rm A}$ increases or the plasma accelerates. At a certain $M_{\rm A}$ or distance, their amplitudes are such that the magnetic field begins to be deflected backward. The occurrence of $90^{\circ}$ deflections should be well above the Alfv\'en critical point, as suggested by our results.
More evidence for the above interpretations is provided by the magnetic mapping of the LMBLs, as shown in Figure~4. The LMBL from encounter 2 is connected to the boundary of the equatorial extension of a southern polar coronal hole (Figure~4a). A similar mapping result is obtained by \citet{meng2022} and \citet{wallace2022} for the same time period. This magnetic connectivity agrees with our interpretation of the LMBL as the transition layer between streamer flows and deeper interior of a coronal hole wind. For the LMBL from encounter 8, the magnetic mapping shows connections first to the edge of a low-latitude small coronal hole, then to the boundary of the equatorial extension of a southern polar coronal hole, and finally to the edge of another low-latitude small coronal hole (Figure~4b). The mapping result is quite similar to that in \citet{kasper2021}. Although there are multiple coronal holes in the present case, the connections to their boundaries are consistent with our interpretation of the LMBL being coronal hole boundary wind. It is important to note that the first low-latitude coronal hole is located near a small active region (see the background EUV image in Figure~4b). The magnetic field is enhanced at the footpoints of the field lines, which may result in a higher rate of energy deposition into the wind. This may explain the large fluctuations in the solar wind radial velocity, in particular the radial velocity of the alphas (Figure~2c). As for the LMBLs from encounter 9, connections to the boundary of a low-latitude coronal hole are seen once more (Figure~4c). We have indicated earlier that the interval between the two LMBLs is connected to deeper interior of a coronal hole (red lines in Figure~4c). How deep this is cannot be resolved by the magnetic mapping given its uncertainties.
Figure~5 illustrates the origin and nature of an LMBL, based on information collected from PSP measurements and the magnetic mapping. The solar wind constituting the LMBL emanates from the peripheral region inside a coronal hole along rapidly diverging open field lines. The origin from within the coronal hole explains the observed low density and considerable differential velocity between alphas and protons. It is well known that there is an inverse relationship between the solar wind speed and coronal field line expansion factor \citep[e.g.,][]{wang1990}. The rapidly diverging field lines have a large expansion factor, which explains the low velocity of the wind. \citet{wang2000} suggest that the area just inside the coronal hole boundary is a source of slow wind. Although slow, the wind may bear signatures of fast wind given its origin. Slowly diverging open field lines are rooted deeper into the coronal hole. They have a smaller expansion factor, so the wind speed is higher. The LMBL can thus be considered as a transition layer between slow and fast wind. The width of the transition layer is expected to be a few degrees on the Sun \citep[also see][]{zurbuchen1999}, i.e., comparable to or smaller than the uncertainty of the magnetic mapping. Although this is a small region on the Sun, it can map to a considerable volume in the heliosphere.
This simple picture clarifies many observations associated with the LMBLs and sub-Alfv\'enic wind. The decrease in the radial Alfv\'en Mach number is a direct consequence of the reduced density and relatively low velocity, which in turn leads to the increase in the Alfv\'en radius and the suppression of SBs. The enhanced Alfv\'en radius allows an easier crossing of the Alfv\'enic transition, so the sub-Alfv\'enic wind is observable even at a distance of about 20 $R_S$ from the Sun. The substantial volume that the LMBL maps to in the inner heliosphere implies that detection of the sub-Alfv\'enic wind is not occasional. An important thing to keep in mind though is that the sub-Alfv\'enic wind is not a typical wind, but a special wind emanating from a coronal hole near its boundary along rapidly diverging open field lines. The low Mach number will inhibit the amplitudes of SBs given the Alfv\'enicity in the velocity and magnetic field fluctuations, as predicted by Equation~(3). Note that there may be other sources on the Sun for LMBLs, such as patches of rapidly diverging open fields near an active region or from the entire interior of a very small coronal hole; the field line geometry may also lead to a reduced density and relatively low speed.
We summarize typical signatures of an LMBL that can be used as criteria to identify the phenomenon from solar wind measurements. The primary signatures include a decreased radial Alfv\'en Mach number, a reduced density, a relatively low velocity, and suppressed SBs. Although the decreased Alfv\'en Mach number is a result of the reduced density and relatively low velocity, it is always helpful to look at all the three parameters together. Suppression in the amplitudes of SBs is useful in the identification of the boundaries of an LMBL. Specific thresholds for these parameters are difficult to quantify because of variations of the quantities with distance. The timescale of an LMBL also varies depending on how long the magnetic footpoint of the spacecraft lingers at the source and the size of the source on the Sun.
\section{Important Implications}
\subsection{Morphology and Crossings of Alfv\'enic Transition}
A first implication concerns the morphology and crossings of the Alfv\'enic transition. \citet{chhiber2022} argue that fluctuations in the plasma and magnetic field, which they expect to increase toward the Alfv\'enic transition, would lead to enhanced variations in the Alfv\'en Mach number approaching the transition. Based on this reasoning, they suggest that the Alfv\'enic transition should not be a simple surface (either smooth or corrugated), but an extended, fragmented zone with mixed parcels of sub-Alfv\'enic and super-Alfv\'enic wind. They demonstrate this possibility using MHD simulations coupled to a turbulent transport model. The possibility is fascinating in view of the turbulent nature of the corona and solar wind. As a result of their theory, one would see subvolumes or blobs of sub-Alfv\'enic wind of random origin and random character in regions that are predominantly super-Alfv\'enic. Our results, however, indicate that the LMBLs including the sub-Alfv\'enic wind all have a similar origin (i.e., coronal hole boundaries with rapidly diverging open fields) and similar properties (i.e., reduced density, relatively low velocity, and signatures resembling fast wind). These are difficult to explain in the frame of their theory, but favor the picture of a ``rugged" Alfv\'en surface that \citet{liu2021} have proposed earlier.
The typical or average Alfv\'en radius is 10 - 12 solar radii. This value is slightly enhanced by the modification of the radial Alfv\'en Mach number, compared with that of \citet{liu2021}. With this average Alfv\'en radius, we echo the suggestion of \citet{liu2021} that substantial diving below the Alfv\'en surface is plausible only for relatively slow solar wind given the orbital design of PSP. The sub-Alfv\'enic wind observed by PSP is, again, not a typical wind. Its detection is simply a result of an enhanced Alfv\'en radius that enables an easier crossing of the Alfv\'enic transition. We do anticipate higher frequency of sub-Alfv\'enic intervals as PSP descends to lower perihelia. For example, flows with magnetic footpoints located deeper inside a coronal hole will have a higher possibility that their Alfv\'en radius is above PSP distance, as the spacecraft descends. This can be easily seen in the context of our results.
\subsection{Generation and Evolution of Switchbacks}
Another implication pertains to the generation and evolution of SBs, a highly debated issue in the field. Our results indicate a dependence of the amplitudes of SBs on the radial Alfv\'en Mach number. Figure~6 shows the magnetic field deflection angle versus the Alfv\'en Mach number using encounter 9 measurements as a representative. The measurements were made from 15:00 UT on August 5 to the end of August 9, 2021 when the distance of PSP ranged from about 16 to 45 $R_S$. This time period is chosen mainly because it covers a relatively broad spectrum of $M_{\rm A}$; during the time period, there are no HCS crossings or transient phenomena such as coronal mass ejections.
The distribution resembles a ``herringbone" structure, where a larger deflection angle tends to be associated with a higher $M_{\rm A}$ (Figure~6 left). Note that the trend with $M_{\rm A}$ is contaminated by mixing of different solar wind streams of different origins or properties. Striations are discernible in the distribution, which are likely formed by flows of similar origin or properties. The increasing trend of the deflection angle with $M_{\rm A}$ is better seen along a striation. It is most clear in the sub-Alfv\'enic wind (within the dashed blue box) where mixing of different flows is reduced. If the increase in $M_{\rm A}$ is considered as the solar wind acceleration process from low to high altitudes, then we would expect larger deflection angles at greater distances. This is indeed found by \citet{mozer2020}. Deflection angles larger than $90^{\circ}$ seem to occur only when $M_{\rm A} \gtrsim 2$. Obviously, a true SB takes place well above the Alfv\'en critical point. For the majority of data the deflection angles are below $90^{\circ}$, so most magnetic field lines are not deflected backward but just ``sideways". The distribution is not symmetric around the $\theta=0$ line; more negative angles are observed. Specifically, the data show a preferential clockwise deflection in the RT plane. This may be caused by the Parker spiral geometry of the mean field, which rotates clockwise.
Effects associated with the original definition of $M_{\rm A}$ are shown in the right panel of Figure~6. Similar features are present as in the left panel, including striations in the distribution, increasing trend of $\theta$, predominance of deflections less than $90^{\circ}$, and preferential clockwise deflections. Large values are seen in the $M_{\rm A}$ as a result of the deflection of the magnetic field from the radial or anti-radial direction. A singularity occurs at $90^{\circ}$, which separates the deflections of $\theta > 90^{\circ}$ from the rest. Note that the apparent decreasing trend in the deflection angles larger than $90^{\circ}$ is an illusion caused by the mathematical singularity. In the current case we also see a critical $M_{\rm A}$ value above which deflections larger than $90^{\circ}$ occur, although the value is changed.
Figure~7 displays the radial velocity enhancement in units of the local Alfv\'en speed versus the magnetic field deflection angle using data from the same time period as in Figure~6. The velocity enhancement in units of the local Alfv\'en speed generally increases with the deflection angle. This is what we expect from Equation~(3). As mentioned earlier, the red curve represents an upper limit of the data, because the mean magnetic field at PSP encounter distances is not exactly radial or anti-radial. Note that, while there are quite some deflection angles larger than $90^{\circ}$, the data points with $\delta{v_R} > v_{\rm A}$ are very few. The fact of $\delta{v_R} \leqslant v_{\rm A}$ in general can also be seen in Figures~1-3. This is surprising since previous studies indicate a typical velocity change of the order of 2$v_{\rm A}$ for a complete field reversal \citep[e.g.,][]{horbury2018, kasper2019, woolley2020}. It seems to suggest a nonlinearly evolved, saturated state where the local Alfv\'en speed is roughly an upper bound for the velocity enhancement. This nonlinearly evolved state, which is missed in previous studies, may hold crucial information on the evolution of SBs.
With the results on the distribution of the amplitudes of SBs as well as the LMBL properties, we are now in a position to test the theories proposed to explain the origin of SBs. We will see how they fit, or are challenged by, our results. Note that the following testing should not be considered as conclusive. More studies are needed to further constrain or distinguish the theories.
(i) Interchange reconnection. A propagating kink in the coronal magnetic field can be produced by reconnection between adjacent open and closed fields \citep[e.g.,][]{fisk2020, zank2020, liang2021}. \citet{drake2021} suggest that reconnection between open and closed flux with a strong guide field injects flux ropes, which can maintain their field geometry over long distances and reproduce some features of SBs observed in the solar wind. Interchange reconnection is an intriguing theory, but a major difficulty it faces is that one would expect decreasing deflection angles with distance. Our results actually reveal increasing deflections with $M_{\rm A}$ or distance (in the solar wind acceleration process $M_{\rm A}$ increases with distance). In addition, an LMBL represents the open-closed boundary, forming an ideal environment for interchange reconnection. Strong SBs are anticipated in an LMBL according to the interchange reconnection theory. We observe the opposite.
(ii) Velocity shear and footpoint motion. The interaction between Alfv\'enic fluctuations and velocity shears is proposed as a possible mechanism to generate SBs \citep[e.g.,][]{landi2006, ruffolo2020}. Shear-driven dynamics are triggered only above the Alfv\'en critical point \citep{ruffolo2020}, so this can be viewed as an in situ formation process in the solar wind. However, we do see magnetic deflections below the Alfv\'en critical point (i.e., $M_{\rm A} < 1$), although the deflection angles are small. Also, we expect enhanced velocity shears inside an LMBL since it is a transition layer between slow and fast wind. Indeed, \citet{pinto2021} find that strong shears develop at such boundaries using MHD simulations. The velocity shear theory would predict enhanced SBs inside an LMBL, but again the opposite is observed.
\citet{schwadron2021} suggest that footpoint motion from the source of slow to fast wind, the so-called super-Parker spiral case, is able to create SBs via the velocity shear between the slow and fast wind. When footpoint motion is reversed (i.e., from the source of fast to slow wind), the so-called sub-Parker spiral case, the field line is straightened, i.e., no SBs. Our Figures~1 and 2 can be considered as the super-Parker and sub-Parker spiral cases, respectively. However, SBs are reduced in both transition layers (i.e., the LMBLs) due to the low Alfv\'en Mach number. No difference is seen between the two LMBLs. They find paucity of SBs within plasma rarefaction regions and use this as a support for their theory. Our results indicate that this is just a consequence of the decreased $M_{\rm A}$ in the rarefaction regions because of the reduced density.
(iii) Expanding waves and turbulence. MHD simulations indicate that Alfv\'en waves in the expanding solar wind can naturally develop into SBs as observed at PSP \citep[e.g.,][]{squire2020, shoda2021}. This is the most promising theory in the context of our work, i.e., the increasing trend in the amplitudes of SBs with the Alfv\'en Mach number, and the association with velocity enhancement for any deflection of the field as a result of outward propagating Alfv\'enic fluctuations. Although \citet{squire2020} stress that this is an ``in situ" formation process in the solar wind, the theory can be modified to include the growth of magnetic deflections in the solar atmosphere, such as the work by \citet{shoda2021}. Our results suggest that SBs have an origin in the solar atmosphere. Their amplitudes are just suppressed below the Alfv\'en critical point, and those that have deflection angles larger than $90^{\circ}$ represent a well evolved state above the Alfv\'en critical point.
(iv) Coronal jets. Another possibility is that SBs originate from coronal jets \citep[e.g.,][]{sterling2020}. This remains a speculation so far, and no much information can be compared with our results. Our work suggests that, however, outward propagating Alfv\'enic fluctuations naturally produce magnetic deflections as well as velocity spikes observed in SBs. Therefore, it is not clear if coronal jets are necessary for (or how they contribute to) the generation of SBs. The concern is valid, in particular when we consider that convection motions in the photosphere would readily drive Alfv\'en waves filling the corona and heliosphere \citep[][and references therein]{cranmer2009}.
\subsection{Modulation of Switchbacks}
Our results about the LMBLs and dependence of SBs on the Alfv\'en Mach number can be used to explain the modulation of SBs. Previously the patchy distribution of SBs has been attributed to modulation by supergranulation on the Sun \citep{bale2021, fargette2021}; the patches are thought to arise from the diverging magnetic funnels associated with supergranulation, as their angular scales are comparable, i.e., a few degrees. \citet{shi2022} examine measurements from encounters 1 and 10 when the longitudinal span of PSP is only several degrees, and also find similar patchy distributions. They argue that those patches are likely not related with the spatial scale of supergranules and temporal effects have to be taken into account. Here we suggest that between adjacent patches SBs are inhibited by a reduction in the Alfv\'en Mach number. More specifically, the patchy distribution is simply a result of modulation by reductions in $M_{\rm A}$.
To test this idea, we show in Figure~8 the data from the same time period as in the Figure~2 of \citet{bale2021}. The shaded regions separating patches that have been studied by \citet{bale2021} are indeed associated with a reduced Alfv\'en Mach number. Clear decreases are seen in the amplitudes of SBs inside the shaded intervals, including the deflection angle and velocity enhancement in units of the local Alfv\'en speed. In general, we observe a correlation between the drop in $M_{\rm A}$ and the reduction in SBs throughout the time period, except the transient structure on September 26. The correlation is also observed in Figures~1-3. The shaded intervals have similar properties to an LMBL, such as a reduced density, velocity valley, and increased Alfv\'en radius. It is likely that the funnels associated with supergranulation may lead to a higher $M_{\rm A}$ of the wind, but this cannot be determined by the magnetic mapping as its uncertainty is comparable to or larger than the angular size of supergranulation. Our results indicate that a repeated in-and-out motion of the PSP footpoint with respect to one or more LMBLs can readily explain the patchy distribution of SBs, no matter whether it is a spatial or temporal effect.
\section{Conclusions}
We have highlighted a special structure in the young solar wind, termed a low Mach-number boundary layer (LMBL). Key findings are revealed concerning the nature of the sub-Alfv\'enic wind observed by PSP, the morphology of the Alfv\'enic transition, and the generation and evolution of SBs. We summarize the results as follows.
(1) An LMBL, by name, is characterized by a reduced radial Alfv\'en Mach number resulting from its decreased solar wind density and relatively low velocity. The low Mach number, in turn, leads to an increased Alfv\'en radius, which enables an easier crossing of the Alfv\'enic transition. Inside an LMBL, the amplitudes of SBs are suppressed by the low Alfv\'en Mach number. A probable source on the Sun for an LMBL is the peripheral region inside a coronal hole with rapidly diverging open fields, so an LMBL can be considered as a transition layer between slow and fast wind. The typical angular width of an LMBL is expected to be a few degrees on the Sun, but its timescale may vary depending on how long the magnetic footpoint of PSP lingers at the source.
(2) The sub-Alfv\'enic wind detected by PSP is a special type of wind emanating from a coronal hole near its boundary along rapidly diverging open fields, i.e., an LMBL flow by nature. It is observable even at a distance of about 20 solar radii from the Sun because of its enhanced Alfv\'en radius. The height of $\sim$20 solar radii should not be regarded as the typical Alfv\'en radius. The average value is 10 - 12 solar radii. In the sub-Alfv\'enic wind, Alfv\'enic deflections do not disappear, although they are reduced by the low Alfv\'en Mach number. A preliminary look at some other sub-Alfv\'enic intervals from the latest PSP encounters seems to confirm the typical characteristics identified here (e.g., decreased Alfv\'en Mach number, reduced density, relatively low velocity, and suppressed SBs).
(3) The LMBLs including the sub-Alfv\'enic wind all have a similar origin (i.e., coronal hole boundaries with rapidly diverging open fields) and similar properties (i.e., reduced density, relatively low velocity, and signatures resembling fast wind). These favor a wrinkled surface for the morphology of the Alfv\'enic transition, rather than an extended, fragmented zone with mixed parcels of sub-Alfv\'enic and super-Alfv\'enic wind. In the context of our results, the picture of a ``rugged" Alfv\'en surface predicts higher frequency of sub-Alfv\'enic intervals as PSP descends to lower perihelia.
(4) We find a dependence of the amplitudes of SBs on the radial Alfv\'en Mach number. The distribution, which resembles a ``herringbone" structure, indicates that a larger deflection angle tends to be associated with a higher Mach number. The majority of deflection angles are below $90^{\circ}$, so most magnetic field lines are not deflected backward but just ``sideways". We suggest that the term of SBs is better changed to Alfv\'enic flows with deflected magnetic field and enhanced radial velocity, or Alfv\'enic deflections for short. The magnetic deflections have an origin well below the Alfv\'en critical point (i.e., $M_{\rm A}<1$), and deflection angles larger than $90^{\circ}$ seem to occur only when $M_{\rm A} \gtrsim 2$, i.e., well above the Alfv\'en critical point.
(5) We present a new approach to calculating the deflection angle, which indicates the deflection direction and allows to establish a relationship with the enhancement in the solar wind radial velocity. A preferential clockwise deflection is observed in the RT plane as viewed from the north. This is attributed to the Parker spiral geometry of the mean magnetic field, or the rotation of the Sun.
(6) A simple analytical model is developed for the relationship between the deflection angle and velocity variation. As indicated by the model, the velocity change is always positive for any deflection of the field, which explains the one-sided nature of the velocity spikes observed in SBs. It suggests that below the Alfv\'en critical point the velocity enhancement and the deflection angle must be significantly reduced. This agrees with the observations in the LMBLs including the sub-Alfv\'enic wind. The model also predicts that obvious SB remnants at 1 AU and beyond may correspond to a high Alfv\'en Mach number and a background field geometry with a large radial component. These are simply consequences of outward propagating Alfv\'enic fluctuations.
(7) The velocity enhancement in units of the local Alfv\'en speed generally increases with the deflection angle. This is consistent with our analytical model. A surprising finding is that, while there are quite some deflection angles larger than $90^{\circ}$, the data points with $\delta{v_R} > v_{\rm A}$ are very few. This finding likely indicates a nonlinearly evolved, saturated state where the local Alfv\'en speed is roughly an upper bound for the velocity enhancement.
(8) We test the theories on the origin of SBs using our results about SBs and the LMBLs. Among the theories (i.e., interchange reconnection, velocity shear and footpoint motion, expanding waves and turbulence, and coronal jets), the most promising one in the context of our work is the model of expanding waves and turbulence. Others do not quite fit our results. Note that the theory of expanding waves and turbulence should not be considered as only an in situ formation process in the solar wind. We find that SBs have an origin well below the Alfv\'en critical point.
(9) The results about the LMBLs and dependence of SBs on the Alfv\'en Mach number indicate that the patchy distribution of SBs is due to modulation by reductions in the Mach number. In general, a correlation is observed between the drop in the Mach number and the reduction of SBs. A repeated in-and-out motion of the PSP footpoint with respect to one or more LMBLs can readily explain the patchy distribution, no matter whether it is a spatial or temporal effect.
(10) Finally, we create a picture on the generation and evolution of SBs based on our results. Photospheric motions shake magnetic field lines and drive Alfv\'en waves that propagate outward \citep[][and references therein]{cranmer2009}. As the solar plasma expands and accelerates to become the solar wind, the outward propagating Alfv\'en waves are able to produce deflections in the magnetic field and enhancements in the plasma radial velocity. Although weak below the Alfv\'en critical point, their amplitudes develop when the Alfv\'en Mach number increases. At a certain Mach number or distance well above the Alfv\'en critical point, the amplitudes are such that the magnetic field begins to be deflected backward (i.e., true SBs form). Deflections larger than $90^{\circ}$ represent a well evolved state, where the local Alfv\'en speed can be regarded as an upper bound for the velocity enhancement. At a certain stage, a decay process may come into play \citep[e.g.,][]{tenerani2021}. SBs begin to fade before reaching 1 AU. Clear remnants observed at 1 AU or beyond correspond to only certain circumstances.
\acknowledgments The research was supported by NSFC under grants 42274201 and 42004145, and the Specialized Research Fund for State Key Laboratories of China. We acknowledge the NASA Parker Solar Probe mission and the SWEAP and FIELDS teams for use of data. The PFSS extrapolation is performed using the \emph{pfsspy} Python package \citep{stansby2020}. The data used for PFSS modeling are courtesy of GONG and SDO/AIA.
|
\section{Introduction}
In the asymmetric limited magnitude error model, a symbol $a\in\mathbb{Z}$ may be modified to $b$ during transmission, and the error magnitude $|a-b|$ is likely to be bounded by certain threshold. One of applications of asymmetric limited magnitude error model is flash memory \cite{CSBB10}. Moreover, for an information $\bf{a}\in\mathbb{Z}^{n}$, a common noise affects only some of the entries. Hence, for integers $n\ge t\ge1$ and $k_{+}\ge k_{-}\ge0$, we define the $(n,t,k_{+},k_{-})$-error ball as
\[\mathcal{B}(n,t,k_{+},k_{-}):=\{{\bf{a}}=(a_1,a_2,\dots,a_n)\in\mathbb{Z}^n:\ a_i\in[-k_{-},k_{+}],\ \text{wt}({\bf{a}})\le t\},\]
where $\text{wt}({\bf{a}})$ denotes the Hamming weight of ${\bf{a}}$. Under this setting, it is easy to see that an error correcting code is equivalent to a packing of $\mathbb{Z}^n$ by $\mathcal{B}(n,t,k_{+},k_{-})$, and a perfect code is equivalent to a tiling of $\mathbb{Z}^{n}$ by $\mathcal{B}(n,t,k_{+},k_{-})$. Moreover, a linear perfect code is equivalent to a lattice tiling of $\mathbb{Z}^{n}$ by $\mathcal{B}(n,t,k_{+},k_{-})$.
When $n=t$, it is easy to see that $\mathcal{B}(n,t,k_{+},k_{-})$ is a cube $[-k_{+}-k_{-},k_{+}+k_{-}]^n$, hence we will only consider $1\le t\le n-1$.
Tilings (or packings) of $\mathbb{Z}^{n}$ by $\mathcal{B}(n,1,k_{+},k_{-})$ have been intensively studied in recent years due to their own interests and applications in flash memory. See \cite{HS86,KBE11,KLNY11,KLY12,S67,S84,SS94,S86,S87,T98,W95} for the researches on cross $\mathcal{B}(n,1,k,k)$ and semi-cross $\mathcal{B}(n,1,k,0)$. Later, these two shapes are extended to quasi-cross $\mathcal{B}(n,1,k_{+},k_{-})$ by Schwartz \cite{S12}, and immediately, quasi-cross was received a lot of attentions \cite{S14,XL20,XL21,YKB13,YZZG20,YZ20,ZG16,ZG18,ZZG17}. For $t\ge2$, there are only a few results. Tilings (or packings) of $\mathbb{Z}^{n}$ by $\mathcal{B}(n,n-1,k,0)$ are considered in \cite{BE13,KLNY11,S90}. In \cite{WS22,WWS21}, the authors studied the tiling and packing problem in the general case. In particular, Wei and Schwartz \cite{WS22} gave a complete classification of the lattice tilings with $\mathcal{B}(n,2,1,0)$ and $\mathcal{B}(n,2,2,0)$. Recently, Wei and Schwartz \cite{WS2023} considered perfect burst-correcting codes for the limited magnitude error channel.
The goal of this paper is to continue this research. Since the lattice tilings with $\mathcal{B}(n,2,1,0)$ and $\mathcal{B}(n,2,2,0)$ have been completely classified, and $\mathcal{B}(n,2,1,1)$ is the next open case. In this paper, we will solve this case. The main result of this paper is the following one.
\begin{theorem}\label{mainthm}
For $n\ge3$, there exists a lattice tiling of $\mathbb{Z}^n$ by $\mathcal{B}(n,2,1,1)$ if and only if $n=11$.
\end{theorem}
This paper is organized as follows. In Section~\ref{pre}, we will give the group ring representation of lattice tilings of $\mathbb{Z}^n$ by $\mathcal{B}(n,2,1,1)$. In Section~\ref{mainsec}, we prove our main result.
\section{Preliminaries}\label{pre}
Let $\mathbb{Z}[G]$ denote the group ring of $G$ over $\mathbb{Z}$, where $G$ is a finite abelian group (written multiplicatively). For any $A\in\mathbb{Z}[G]$, $A$ can be written as formal sum $A=\sum_{g\in G}a_gg$, where $a_g\in\mathbb{Z}$. For any $A=\sum_{g\in G}a_gg$ and $t\in\mathbb{Z}$, we define
\[A^{(t)}=\sum_{g\in G}a_gg^t.\]
Addition, subtraction, multiplication and scalar multiplication in group ring are defined as:
\[\sum_{g\in G}a_gg\pm\sum_{g\in G}b_gg=\sum_{g\in G}(a_g\pm b_g)g,\]
\[\sum_{g\in G}a_gg\sum_{g\in G}b_gg=\sum_{g\in G}(\sum_{h\in G}a_{h}b_{h^{-1}g})g,\]
and
\[\lambda\sum_{g\in G}a_gg=\sum_{g\in G}(\lambda a_g)g,\]
where $\lambda\in\mathbb{Z}$ and $\sum_{g\in G}a_gg,\sum_{g\in G}b_gg\in\mathbb{Z}[G]$. For any multiset $A$ of $G$, we can identify $A$ with the group ring element $\sum_{g\in G}a_gg$, where $a_g$ is the multiplicity of $g$ appearing in $A$. We also define $\text{supp}(A)$ to be the support of $A=\sum_{g\in G}a_gg$, i.e., $\text{supp}(A)=\{g\in G: a_g\ne0\}$. For any $A=\sum_{g\in G}a_gg\in\mathbb{Z}[G]$, we denote
\[A^{*}=\sum_{g\in G\backslash\{e\}}a_gg.\]
The following theorem establishes the connection between lattice tilings of $\mathbb{Z}^n$ and finite abelian groups.
\begin{theorem}{\rm{\cite{HA12}}}\label{tilinggroup}
Let $V$ be a subset of $\mathbb{Z}^{n}$. Then there is a lattice tiling $\mathcal{T}$ of $\mathbb{Z}^n$ by $V$ if and only if there are both an abelian group $G$ of order $|V|$ and a homomorphism $\phi:\mathbb{Z}^n\rightarrow G$ such that the restriction of $\phi$ to $V$ is a bijection.
\end{theorem}
Now we translate the existence of lattice tilings of $\mathbb{Z}^n$ by $\mathcal{B}(n,2,1,1)$ into group ring equations.
\begin{theorem}\label{groupring}
Let $n\ge3$, then there exists a lattice tiling of $\mathbb{Z}^n$ by $\mathcal{B}(n,2,1,1)$ if and only if there exists a finite abelian group $G$ of order $2n^2+1$ and a subset $T\subseteq G$ satisfying
\begin{enumerate}
\item [(1)] $|T|=2n+1$,
\item [(2)] $e\in T$,
\item [(3)] $T=T^{(-1)}$,
\item [(4)] $T^2=2G+T^{(2)}+(2n-2)e$.
\end{enumerate}
\end{theorem}
\begin{proof}
Suppose that there exists a lattice tiling of $\mathbb{Z}^n$ by $\mathcal{B}(n,2,1,1)$.
Let $e_i$, $i=1,2,\dots,n$, be a fixed orthonormal basis of $\mathbb{Z}^{n}$.
By Theorem~\ref{tilinggroup}, there are both an abelian group $G$ of order $2n^2+1$ and a homomorphism $\phi:\mathbb{Z}^n\rightarrow G$ such that the restriction of $\phi$ to $\mathcal{B}(n,2,1,1)$ is a bijection. Since the homomorphism $\phi$ is determined by the values $\phi(e_i)$, $i=1,\dots,n$, then there exists an $n$-subset $\{a_1,a_2,\dots,a_n\}\subset G$ (let $\phi(e_i)=a_i$) such that
\begin{align*}
G=\{e\}\cup\{a_i,a_{i}^{-1}:\ 1\le i\le n\}\cup\{a_ia_j,a_ia_{j}^{-1},a_{i}^{-1}a_j,a_{i}^{-1}a_{j}^{-1}:\ 1\le i<j\le n\}.
\end{align*}
In the language of group ring, the above equation can be written as
\begin{align*}
G=e+\sum_{i=1}^{n}(a_i+a_{i}^{-1})+\sum_{1\le i<j\le n}(a_i+a_{i}^{-1})(a_j+a_{j}^{-1}).
\end{align*}
Let $T=e+\sum_{i=1}^{n}(a_i+a_{i}^{-1})$, then $|T|=2n+1$, $e\in T$ and $T^{(-1)}=T$. Moreover,
\begin{align*}
T^2=&(e+\sum_{i=1}^{n}(a_i+a_{i}^{-1}))^2\\
=&e+2\sum_{i=1}^{n}(a_i+a_{i}^{-1})+\sum_{i=1}^{n}(a_{i}^{2}+a_{i}^{-2})+2\sum_{1\le i<j\le n}(a_i+a_{i}^{-1})(a_j+a_{j}^{-1})+2ne\\
=&2G+T^{(2)}+(2n-2)e.\qedhere
\end{align*}
\end{proof}
\section{Proof of the main theorem}\label{mainsec}
In this section, we will prove our main result. The main method was developed by Leung and Zhou \cite{LZ20}, which completely solved the existence of linear perfect Lee codes with radius 2. This method was also used to determine the existence of almost perfect linear Lee codes of radius 2 \cite{XZ2022,ZZ2022} and linear diameter perfect Lee codes with diameter 6 \cite{ZG2022}.
The sketch of our proof is: we will first investigate $T^{(2)}T\pmod{3}$, we can get some restrictions on the coefficients of elements of $T^{(2)}T$. This will solve the cases $n\equiv2\pmod{3}$ except $n=5,11$. By studying $TT^{(3)}\pmod{3}$, we solve the case $n\equiv1\pmod{3}$ except $n=4$.
For the case $n\equiv0\pmod{3}$, we will also study $TT^{(4)}\pmod{5}$ and $TT^{(5)}\pmod{5}$. Small dimensions will be solved by symmetric polynomial method, which was developed by Kim \cite{K17}.
Let $n\ge3$ be an integer. Suppose that $G$ is a finite abelian group with order $2n^2+1$ and $T\subset G$ satisfying conditions in Theorem~\ref{groupring}. Recall that $T=e+\sum_{i=1}^{n}(a_i+a_{i}^{-1})$ and
\begin{align}\label{eq1}
G=e+\sum_{i=1}^{n}(a_i+a_{i}^{-1})+\sum_{1\le i<j\le n}(a_i+a_{i}^{-1})(a_j+a_{j}^{-1}).
\end{align}
The following lemma is directly from Condition (4) of Theorem~\ref{groupring}.
\begin{lemma}\label{lemma1}
For any $t\ne e$, we have
\[|\{(t_1,t_2)\in T\times T: t_1t_2=t\}|= \begin{cases}
3, & \mbox{if } t\in T^{(2)}; \\
2, & \mbox{if }t\notin T^{(2)}.
\end{cases}\]
\end{lemma}
From Condition (4) in Theorem~\ref{groupring}, we have
\begin{align}\label{eq4}
TT^{(2)}=T^{3}-(4n+2)G-(2n-2)T.
\end{align}
We write
\begin{align}\label{eq2}
TT^{(2)}=\sum_{i=0}^{M}iX_i,
\end{align}
where $X_i$ $(i=0,1,\dots,M)$ form a partition of group $G$, i.e.,
\begin{align}\label{eq3}
G=\sum_{i=0}^{M}X_i.
\end{align}
Then we have the following lemma.
\begin{lemma}\label{lemma2}
\begin{enumerate}
\item [(1)] $\sum_{i=1}^{M}i|X_i|=4n^2+4n+1$,
\item [(2)] $\sum_{i=0}^{M}|X_i|=2n^2+1$,
\item [(3)] $\sum_{i=1}^{M}|X_i|=1-\beta+\sum_{i\ge3}\frac{(i-1)(i-2)}{2}|X_i|$, where $2\beta=|T^{*}\cap T^{(2)*}|$.
\end{enumerate}
\end{lemma}
\begin{proof}
The first two equations are directly from Equations~(\ref{eq2}) and (\ref{eq3}). Now we count the size of $\text{supp}(T^{(2)}T)$.
We write $T^{(2)}=\sum_{i=0}^{2n}b_i$, where $b_0=e$ and $b_{i}^{-1}=b_{n+i}$ for $i=1,2,\dots,n$.
By Lemma~\ref{lemma1}, we have
\[|b_iT\cap b_jT|= \begin{cases}
3, & \mbox{if } b_ib_{j}^{-1}\in T^{(2)}; \\
2, & \mbox{if }b_ib_{j}^{-1}\notin T^{(2)}.
\end{cases}\]
Let $\alpha=|\{(i,j): 0\le i<j\le 2n, b_ib_{j}^{-1}\in T^{(2)}\}|$ and $N(t)=|\{(i,j): 0\le i<j\le 2n, b_ib_{j}^{-1}=t\}|$, then $\alpha=\sum_{t\in T^{(2)}}N(t)$. Since $\gcd(2n^2+1,2)=1$, then $T^{(2)}$ also satisfies conditions in Theorem~\ref{groupring}. By the definition of $b_i$ and Lemma~\ref{lemma1}, we have
\[N(t)+N(t^{-1})= \begin{cases}
3, & \mbox{if } t\in T^{(2)*}\cap T^{(4)*}; \\
2, & \mbox{if } t\in T^{(2)}\backslash T^{(4)}.
\end{cases}\]
Hence, $\alpha=|T^{(2)}\backslash T^{(4)}|+\frac{3|T^{(2)*}\cap T^{(4)*}|}{2}=|T^{*}\backslash T^{(2)*}|+\frac{3|T^{*}\cap T^{(2)*}|}{2}=2n+\beta$.
By the inclusion-exclusion principle, we have
\begin{align*}
\sum_{i=1}^{M}|X_i|=&|\text{supp}(T^{(2)}T)|\\
=&\sum_{i=0}^{2n}|b_iT|-\sum_{0\le i<j\le 2n}|b_iT\cap b_jT|+\sum_{r\ge3}\sum_{0\le i_1<\dots<i_r\le 2n}(-1)^{r-1}|b_{i_1}T\cap\dots\cap b_{i_r}T|\\
=&4n^2+4n+1-3\alpha-2(\binom{2n+1}{2}-\alpha)+\sum_{r\ge3}\sum_{0\le i_1<\dots<i_r\le 2n}(-1)^{r-1}|b_{i_1}T\cap\dots\cap b_{i_r}T|.
\end{align*}
Suppose that $g\in b_{i_1}T\cap\dots\cap b_{i_r}T$ for some $r\ge 3$, then $g\in X_i$ for some $i\ge3$. The contribution for $g$ in the sum $\sum_{r\ge3}\sum_{1\le i_1<\dots<i_r\le 2n}(-1)^{r-1}|b_{i_1}T\cap\dots\cap b_{i_r}T|$ is
\[(-1)^{3-1}\binom{i}{3}+(-1)^{4-1}\binom{i}{4}+\cdots=\binom{i}{2}-\binom{i}{1}+\binom{i}{0}=\frac{(i-1)(i-2)}{2}.\]
Hence, we have
\begin{align*}
\sum_{i=1}^{M}|X_i|=&4n^2+4n+1-3\alpha-2(\binom{2n+1}{2}-\alpha)+\sum_{r\ge3}\sum_{1\le i_1<\dots<i_r\le 2n}(-1)^{r-1}|b_{i_1}T\cap\dots\cap b_{i_r}T|\\
=&4n^2+4n+1-3\alpha-2(\binom{2n+1}{2}-\alpha)+\sum_{i\ge3}|X_i|\frac{(i-1)(i-2)}{2}\\
=&1-\beta+\sum_{i\ge3}\frac{(i-1)(i-2)}{2}|X_i|.\qedhere
\end{align*}
\end{proof}
\begin{lemma}\label{lemma6}
$e$ appears $2\beta+1$ times in $T^{(3)}$, where $2\beta=|T^{*}\cap T^{(2)*}|$.
\end{lemma}
\begin{proof}
If there exist $a_i,a_j\in T^{*}$ with $j\ne i$ such that $a_i=a_{j}^{2}\in T^{*}\cap T^{(2)*}$, then $a_{j}=a_{i}a_{j}^{-1}$. By Equation~(\ref{eq1}), it can be seen that $a_{j}$ appears twice in $G$, which is a contradiction.
If there exists $a_{i}\in T^{*}$ such that $a_i=a_{i}^{2}\in T^{*}\cap T^{(2)*}$, then $a_{i}=e$, which is a contradiction again.
If there exists $a_{i}\in T^{*}$ such that $a_i=a_{i}^{-2}\in T^{*}\cap T^{(2)*}$, then $a_{i}^{3}=e$.
On the other hand, if there exists $a_{i}\in T^{*}$ such that $a_{i}^3=e$, then $a_i=a_{i}^{-2}\in T^{*}\cap T^{(2)*}$. Hence $e$ appears $2\beta+1$ times in $T^{(3)}$.
\end{proof}
\subsection{$n\equiv2\pmod{3}$}
\begin{proposition}\label{prop1}
Theorem~\ref{mainthm} holds for $n\equiv2\pmod{3}$ except for $n=5,11$.
\end{proposition}
\begin{proof}
By Equation~(\ref{eq4}), we obtain
\begin{align}\label{eq23}
T^{(2)}T\equiv T^{(3)}+2G+T\pmod{3}.
\end{align}
This implies that
\begin{align*}
&\sum_{i\ge0}|X_{3i}|=a_1,\\
&\sum_{i\ge0}|X_{3i+1}|=a_2,\\
&\sum_{i\ge0}|X_{3i+2}|=2n^2+1-a_1-a_2,
\end{align*}
where
\begin{align*}
&a_1=|T\backslash T^{(3)}|+b_1+b_2,\\
&a_2=b_3+b_4,\\
&b_1=|\{t\in T\cap T^{(3)}: t\text{ appears }3i\text{ times in }T^{(3)}\text{ for some }i\}|,\\
&b_2=|\{t\in T^{(3)}\backslash T: t\text{ appears }3i+1\text{ times in }T^{(3)}\text{ for some }i\}|,\\
&b_3=|\{t\in T\cap T^{(3)}: t\text{ appears }3i+1\text{ times in }T^{(3)}\text{ for some }i\}|,\\
&b_4=|\{t\in T^{(3)}\backslash T: t\text{ appears }3i+2\text{ times in }T^{(3)}\text{ for some }i\}|.
\end{align*}
Hence $a_1+a_2\le4n+1$. Since for any $t\in T^{*}$, $t=t^{-1}\cdot t^2$, $t^3=t\cdot t^2$, then Equation (\ref{eq23}) implies that for any $e\ne g\in G$, $g$ appears at least twice in $TT^{(3)}$. Hence $|X_0|=0$ and $|X_1|\le1$.
{\bf{Claim: $|X_1|=0$.}}
Assume to the contrary, $|X_1|=1$, then $e\in X_1$ and $\beta=0$. By Lemma~\ref{lemma2}, we have
\begin{equation}\label{eq7}
\begin{split}
\sum_{i=3}^{M}(i-2)|X_i|=&\sum_{i=1}^{M}i|X_i|-2\sum_{i=1}^{M}|X_i|+|X_1|\\
=&4n^2+4n+1-2(2n^2+1)+1\\
=&4n,
\end{split}
\end{equation}
and
\begin{equation}\label{eq8}
\begin{split}
2n^2=&\sum_{i=3}^{M}\frac{(i-1)(i-2)}{2}|X_i|\\
\le&\frac{M-1}{2}\sum_{i=3}^{M}(i-2)|X_i|\\
=&\frac{M-1}{2}4n.
\end{split}
\end{equation}
This leads to $M\ge n+1$. Note that for any $t\in T^{*}$, $t$ appears at least $3$ times in $TT^{(2)}$. Then Equation (\ref{eq7}) implies that
\begin{align*}
4n=&\sum_{i=3}^{M}(i-2)|X_i|\\
\ge&(2n-|X_M|)+(M-2)|X_M|.
\end{align*}
This leads to $(M-3)|X_M|\le2n$. So $|X_M|\le2$. Since $X_{M}=X_{M}^{(-1)}$, then $|X_M|=2$ and $M\le n+3$. Note that for any $t\in T^{*}\cup T^{(3)*}$, if $t$ appears $a$ times in $T^{*}\cup T^{(3)*}$, then $t$ appears $a+2$ times in $TT^{(2)}$. By Equation~(\ref{eq7}) again, we have $4n\ge(3-2)(4n-2(M-2))+(M-2)2$. This implies that $|X_3|=4n-2(M-2)$, $|X_M|=2$ and $|X_i|=0$ for $4\le i\le M-1$. From Equation (\ref{eq8}), we obtain
\begin{align*}
2n^2=&4n-2(M-2)+2\frac{(M-1)(M-2)}{2}\\
=&4n+(M-2)(M-3)\\
\le&4n+n(n+1).
\end{align*}
This is possible only for $n=5$. This finishes the proof of claim.
By Lemma~\ref{lemma2}, we have
\begin{equation}\label{eq5}
\begin{split}
\sum_{i=3}^{M}(i-2)|X_i|=&\sum_{i=2}^{M}i|X_i|-2\sum_{i=2}^{M}|X_i|\\
=&4n^2+4n+1-2(2n^2+1)\\
=&4n-1,
\end{split}
\end{equation}
and
\begin{equation}\label{eq6}
\begin{split}
2n^2+\beta=&\sum_{i=3}^{M}\frac{(i-1)(i-2)}{2}|X_i|\\
\le&\frac{M-1}{2}\sum_{i=3}^{M}(i-2)|X_i|\\
=&\frac{M-1}{2}(4n-1).
\end{split}
\end{equation}
This leads to $M\ge\frac{4n^2+2\beta}{4n-1}+1$. Since $M\in\mathbb{Z}$, then $M\ge n+2$.
{\bf{Case 1: $|X_{M}|\ge2$.}}
Note that for any $t\in T$, $t$ appears at least $3$ times in $TT^{(2)}$.
From Equation~(\ref{eq5}), we obtain
\begin{align*}
4n-1=&\sum_{i=3}^{M}(i-2)|X_i|\\
\ge&(3-2)(2n+1-|X_M|)+n|X_M|\\
=&2n+1+(n-1)|X_M|.
\end{align*}
This leads to $|X_{n+2}|=2$, $|X_{3}|=2n-1$, $e\in X_3$, $\beta=1$ and $|X_i|=0$ for all $4\le i\le n+1$. By Lemma~\ref{lemma2}, we have
\begin{align*}
2n^2+1=&\sum_{i=3}^{M}\frac{(i-1)(i-2)}{2}|X_i|\\
=&|X_3|+\frac{(n+1)n}{2}|X_{n+2}|\\
=&2n-1+n^2+n\\
=&n^2+3n-1.
\end{align*}
This implies that $n^2-3n+2=0$, which is a contradiction.
{\bf{Case 2: $|X_M|=1$.}}
For this case, we have $X_M=\{e\}$ and $M=2\beta+1$.
Note that for any $t\in T^{*}\cup T^{(3)*}$, if $t$ appears $a$ times in $T^{*}\cup T^{(3)*}$, then $t$ appears $a+2$ times in $TT^{(2)}$. By Equation~(\ref{eq5}), we have
\begin{align*}
4n-1=&\sum_{i=3}^{M}(i-2)|X_i|\\
\ge&(3-2)(4n+1-M)+(M-2)\\
=&4n-1.
\end{align*}
This implies that $|X_3|=4n+1-M$ and $|X_i|=0$ for all $4\le i\le M-1$. By Lemma~\ref{lemma2}, we have
\begin{align*}
2n^2+\frac{M-1}{2}=&\sum_{i=3}^{M}\frac{(i-1)(i-2)}{2}|X_i|\\
=&|X_3|+\frac{(M-1)(M-2)}{2}\\
=&4n+1-M+\frac{(M-1)(M-2)}{2}.
\end{align*}
This leads to $M^2-6M-(4n^2-8n-5)=0$, then $M=2n+1$, and so $T=T^{(2)}$. Hence $(T^{*})^2=2G-T^{*}+(2n-2)e$ and $T^{*}$ is a $(v,k,\lambda,\mu)$ regular partial difference set with $v=2n^2+1$, $k=2n$, $\lambda=1$ and $\mu=2$. By \cite[Theorem 6.1]{M1994}, this is only possible for $n=11$.
\end{proof}
\subsection{$n\equiv1\pmod{3}$}
By Equation (\ref{eq4}), we obtain
\begin{align*}
T^4=&(8n^2+8n+2)G+(2n-2)T^2+T^2T^{(2)}\\
=&(8n^2+8n+2)G+(2n-2)(2G+T^{(2)}+(2n-2)e)+(2G+T^{(2)}+(2n-2)e)T^{(2)}\\
=&(8n^2+16n)G+(4n-4)T^{(2)}+(4n^2-8n+4)e+(T^{(2)})^2\\
=&(8n^2+16n+2)G+(4n-4)T^{(2)}+T^{(4)}+(4n^2-6n+2)e.
\end{align*}
This implies
\begin{align}
TT^{(3)}\equiv2G+T^{(4)}\pmod{3}.\label{eq11}
\end{align}
We write $TT^{(3)}=\sum_{i=0}^{L}iZ_i$, where $Z_i\ (i=0,1,\dots,L)$ form a partition of group $G$, i.e., $G=\sum_{i=0}^{L}Z_i$. Then it is easy to get the following equations.
\begin{align}
&4n^2+4n+1=\sum_{i=1}^{L}i|Z_i|,\label{eq9}\\
&2n^2+1=\sum_{i=0}^{L}|Z_i|.\label{eq10}
\end{align}
By Equation~(\ref{eq11}), we see that for any $t^4\in T^{(4)}$, $t^4=t\cdot t^3$, so $t^4$ appears at least 3 times in $TT^{(3)}$, and for any $g\in G\backslash T^{(4)}$, $g$ appears at least 2 times in $TT^{(3)}$. Hence $|Z_0|=|Z_1|=0$. Moreover, we have
\begin{align}
&\sum_{i\ge1}|Z_{3i}|=2n+1,\label{eq12}\\
&\sum_{i\ge0}|Z_{3i+2}|=2n^2-2n.\label{eq13}
\end{align}
\begin{lemma}\label{lemma3}
For any $g\in G$, $g$ appears at most 2 times in $T^{(3)}$.
\end{lemma}
\begin{proof}
Assume to the contrary, suppose that $g$ appears at least $3$ times in $T^{(3)}$, then for any $t\in T$, $g\cdot t$ appears at least 3 times in $TT^{(3)}$. By Lemma~\ref{lemma1}, for any $g'\in T^{(3)}$, $g'\ne g$, we have $|gT\cap g'T|\ge2$. This implies that the elements in $gT$ appear at least $3(2n+1)+(2n-2)2$ times in $TT^{(3)}$. By Equation~(\ref{eq9}), we obtain
\begin{align*}
&4n^2+4n+1\\
=&\sum_{i=1}^{L}i|Z_i|\\
\ge&2(2n^2-2n)+3(2n+1)+2(2n-2)\\
=&4n^2+6n-1,
\end{align*}
which is a contradiction.
\end{proof}
Since $e$ appears odd times in $T^{(3)}$, then $e$ appears once in $T^{(3)}$.
\begin{lemma}\label{lemma4}
For any $e\ne g\in G$, $g$ appears at most $n+2$ times in $TT^{(3)}$.
\end{lemma}
\begin{proof}
If there exists $g\ne e$ such that $g$ appears at least $n+3$ times in $TT^{(3)}$, then $g^{-1}$ also appears at least $n+3$ times in $TT^{(3)}$. By Equation~(\ref{eq9}), we obtain
\begin{align*}
&4n^2+4n+1\\
=&\sum_{i=1}^{L}i|Z_i|\\
\ge&2(2n^2-2n)+3(2n+1)+2(n+3)-3\cdot2\\
=&4n^2+4n+3,
\end{align*}
which is a contradiction.
\end{proof}
\begin{lemma}\label{lemma7}
If $n\ge7$, then $|\text{supp}(T^{(3)})|=2n+1.$
\end{lemma}
\begin{proof}
If there exists $g\in G$, $g\ne e$ such that $g$ appears twice in $T^{(3)}$, then $g^{-1}$ also appears twice in $T^{(3)}$. By Lemma~\ref{lemma1}, for any $g'\in T^{(3)}$, $g'\ne g,g^{-1}$, we have $|gT\cap g'T|\ge2$ and $|g^{-1}T\cap g'T|\ge2$. By Lemma~\ref{lemma4}, we have that the elements in $gT\cup g^{-1}T$ appear at least $2(2n+1)+2(2n+1)+2(2n+1-4)+(2n+1-4-(n+2-2))=13n-5$ times in $TT^{(3)}$. By Equation~(\ref{eq9}), we have
\begin{align*}
&4n^2+4n+1\\
=&\sum_{i=1}^{L}i|Z_i|\\
\ge&2(2n^2+1-|gT\cup g^{-1}T|)+13n-5\\
\ge&2(2n^2+1-4n)+13n-5\\
=&4n^2+5n-3,
\end{align*}
this is possible only for $n=4$.
\end{proof}
With a similar discussion as Lemma~\ref{lemma2}, we have the following lemma.
\begin{lemma}\label{lemma5}
If $n\ge7$, then we have $\sum_{i=1}^{L}|Z_i|=1-\delta+\sum_{i=3}^{L}\frac{(i-1)(i-2)}{2}|Z_i|$, where $2\delta+1=|T\cap T^{(3)}|$.
\end{lemma}
\begin{proposition}\label{prop2}
Theorem~\ref{mainthm} holds for $n\equiv1\pmod{3}$ except $n=4$.
\end{proposition}
\begin{proof}
By Equations (\ref{eq9}), (\ref{eq10}), (\ref{eq12}) and (\ref{eq13}), we get
\begin{align}
&\sum_{i=3}^{L}(i-2)|Z_i|=\sum_{i=2}^{L}i|Z_i|-2\sum_{i=2}^{L}|Z_i|=4n-1,\\
&\sum_{i\ge2}(3i-3)|Z_{3i}|+\sum_{i\ge1}3i|X_{3i+2}|=\sum_{i=3}^{L}(i-2)|Z_i|-\sum_{i\ge1}|Z_{3i}|=2n-2.\label{eq14}
\end{align}
From Lemma~\ref{lemma5}, we obtain
\begin{align*}
2n^2+\delta-4n+1=&\sum_{i=3}^{L}\frac{(i-1)(i-2)}{2}|Z_i|-\sum_{i=3}^{L}(i-2)|Z_i|\\
=&\sum_{i=3}^{L}\frac{(i-3)(i-2)}{2}|Z_i|\\
\le&\frac{L-2}{2}\sum_{i=3}^{L}(i-3)|Z_i|\\
\le&\frac{L-2}{2}(2n-2).
\end{align*}
This leads to $L\ge2n$. From Equation (\ref{eq14}), we have $|Z_L|=1$ and then $Z_L=\{e\}$. Equation (\ref{eq12}) implies that $L\equiv0\pmod{3}$, so $L=2n+1$. Hence $T=T^{(3)}$. Comparing Equation (\ref{eq11}) with Condition (4) in Theorem~\ref{groupring}, we obtain $T^{(2)}=T^{(4)}$, which contradicts to Lemmas~\ref{lemma6} and \ref{lemma7}.
\end{proof}
\subsection{$n\equiv0\pmod{3}$}
Since $\gcd(3,2n^2+1)=1$, then $|\text{supp}(T^{(3)})|=2n+1$. So $\beta=0$, i.e., $|T\cap T^{(2)}|=1$. Hence $e\in X_1$. By Equation~(\ref{eq4}), we obtain
\begin{align*}
T^{(2)}T\equiv T^{(3)}+G+2T\pmod{3}.
\end{align*}
This implies that for any $g\in G\backslash(T\cup T^{(3)})$, $g$ appears at least once in $T^{(2)}T$, and for any $t\in T$, $t=t^{-1}t^2\in T^{(2)}T$, $t^3=tt^2\in T^{(2)}T$. Hence $|X_0|=0$. Moreover,
\begin{align}
&\sum_{i\ge1}|X_{3i}|=a,\label{eq16}\\
&\sum_{i\ge0}|X_{3i+2}|=a,\\
&\sum_{i\ge0}|X_{3i+1}|=2n^2+1-2a,\label{eq15}
\end{align}
where $a=|T\backslash(T\cap T^{(3)})|=|T^{(3)}\backslash(T\cap T^{(3)})|$. Then $a\le2n$. By Lemma~\ref{lemma2}, we have
\begin{align*}
\sum_{i\ge1}3i(|X_{3i}|+|X_{3i+1}|+|X_{3i+2}|)=&\sum_{i=1}^{M}i|X_i|-\sum_{i\ge0}|X_{3i+1}|-2\sum_{i\ge0}|X_{3i+2}|\\
=&(4n^2+4n+1)-(2n^2+1-2a)-2a\\
=&2n^2+4n,
\end{align*}
and
\begin{align*}
2n^2=&\sum_{i=3}^{M}\frac{(i-1)(i-2)}{2}|X_i|\\
\ge&-2|X_3|+\sum_{i\ge1}3i(|X_{3i}|+|X_{3i+1}|+|X_{3i+2}|)\\
=&-2|X_3|+2n^2+4n.
\end{align*}
This leads to $|X_3|\ge2n$. Combining with Equations (\ref{eq16})-(\ref{eq15}) and $a\le2n$, we have
\begin{align*}
|X_1|=\frac{4}{3}n^2-\frac{10}{3}n+1,\ |X_2|=2n,\ |X_3|=2n,\ |X_4|=\frac{2}{3}n^2-\frac{2}{3}n\text{ and }|X_i|=0\text{ for all }i\ge5.
\end{align*}
Moreover
\begin{align*}
T^{(2)}T=X_1+2T^{(3)*}+3T^{*}+4X_4.
\end{align*}
In order to get a contradiction, we need to investigate $TT^{(4)}$. By Theorem~\ref{groupring}, we have
\begin{align*}
(T^{(2)})^2=&(2G+(2n-2)e-T^2)^2\\
=&(8n^2+8n-4)G+(4n^2-8n+4)e-(4G+(4n-4)e)T^2+T^4\\
=&(8n^2+8n-4)G+(4n^2-8n+4)e-(4G+(4n-4)e)(2G+T^{(2)}+(2n-2)e)+T^4\\
=&(-8n^2-16n)G+(-4n^2+8n-4)e-(4n-4)T^{(2)}+T^4.
\end{align*}
On the other hand, $T^{(2)}$ also satisfies conditions in Theorem~\ref{groupring}, then
\[(T^{(2)})^2=2G+T^{(4)}+(2n-2)e.\]
Combining above two equations, we obtain
\[T^4=(8n^2+16n+2)G+(4n^2-6n+2)e+T^{(4)}+(4n-4)T^{(2)}.\]
Multiplying $T$ to both sides, we can get
\begin{align}\label{eq17}
TT^{(4)}=T^{5}-(16n^3+40n^2+20n+2)G-(4n^2-6n+2)T-(4n-4)TT^{(2)}.
\end{align}
Writing $TT^{(4)}$ as
\begin{align*}
TT^{(4)}=\sum_{i=0}^{N}iY_i,
\end{align*}
where $N\le2n+1$, and $Y_i$, $i=0,1,\dots,N$ form a partition of group $H$, i.e.,
\begin{align*}
H=\sum_{i=0}^{N}Y_i.
\end{align*}
Then we have the following lemma.
\begin{lemma}\label{lemma8}
\begin{enumerate}
\item [(1)] $\sum_{i=0}^{N}|Y_i|=2n^2+1$,
\item [(2)] $\sum_{i=1}^{N}i|Y_i|=4n^2+4n+1,$
\item [(3)] $\sum_{i=1}^{N}|Y_i|=1-\gamma+\sum_{i\ge3}|Y_i|\frac{(i-1)(i-2)}{2}$, where $|T^{*}\cap T^{(4)*}|=2\gamma$.
\end{enumerate}
\end{lemma}
\begin{proof}
The proof is similar to that for Lemma~\ref{lemma2}.
\end{proof}
\begin{proposition}\label{prop3}
Theorem~\ref{mainthm} holds for $n\equiv6\pmod{15}.$
\end{proposition}
\begin{proof}
By Equation~(\ref{eq17}), we obtain
\[TT^{(4)}\equiv T^{(5)}+2G\pmod{5}.\]
This implies
\begin{align*}
&\sum_{i\ge0}|Y_{5i+3}|=2n+1,\\
&\sum_{i\ge0}|Y_{5i+2}|=2n^2-2n,\\
&|Y_i|=0\text{ for all }i\equiv0,1,4\pmod{5}.
\end{align*}
By Lemma~\ref{lemma8}, we have
\begin{align}
&\sum_{i\ge3}(i-2)|Y_i|=\sum_{i=1}^{N}i|Y_i|-2\sum_{i=1}^{N}|Y_i|=4n-1,\\
&\sum_{i\ge1}5i(|Y_{5i+2}|+|Y_{5i+3}|)=\sum_{i=1}^{N}i|Y_i|-2\sum_{i\ge0}|Y_{5i+2}|-3\sum_{i\ge0}|Y_{5i+3}|=2n-2,\label{eq18}
\end{align}
and
\begin{align*}
2n^2+\gamma-4n+1=&\sum_{i=3}^{N}\frac{(i-1)(i-2)}{2}|Y_i|-\sum_{i\ge3}(i-2)|Y_i|\\
=&\sum_{i=3}^{N}\frac{(i-3)(i-2)}{2}|Y_i|\\
\le&\frac{N-2}{2}\sum_{i=3}^{N}(i-3)|Y_i|\\
\le&\frac{N-2}{2}\sum_{i\ge1}5i(|Y_{5i+2}|+|Y_{5i+3}|)\\
=&\frac{N-2}{2}(2n-2).
\end{align*}
Since $N\in\mathbb{Z}$, then $N\ge2n$. Equation~(\ref{eq18}) implies that $|Y_N|=1$. Since $Y_N=Y_{N}^{(-1)}$, then $Y_N=\{e\}$. Thus $N$ is odd, so $N=2n+1$. Hence $T=T^{(4)}$. By Equation (\ref{eq17}), we obtain
\[T^{2}=T^{5}-(16n^3+40n^2+20n+2)G-(4n^2-6n+2)T-(4n-4)TT^{(2)}.\]
Multiplying $T$ to both sides, we obtain
\[T^{3}=T^{6}-(2n+1)(16n^3+40n^2+20n+2)G-(4n^2-6n+2)T^2-(4n-4)T^2T^{(2)}.\]
Taking modulo 3, we have
\begin{align*}
T^{(3)}\equiv&(T^{(3)})^2+G+T^2+T^2T^{(2)}\pmod{3}\\
\equiv& (2G+T^{(6)}+e)+G+(2G+T^{(2)}+e)+(2G+T^{(2)}+e)T^{(2)}\pmod{3}\\
\equiv &2G+T^{(6)}+T^{(2)}+2e+2G+(T^{(2)})^2+T^{(2)}\pmod{3}\\
\equiv &G+T^{(6)}+2T^{(2)}+2e+(2G+T^{(4)}+e)\pmod{3}\\
\equiv &T^{(6)}+2T^{(2)}+T\pmod{3}.
\end{align*}
Since $T\cap T^{(2)}=\{e\}$, then this is only possible for $T=T^{(3)}$ and $T^{(2)}=T^{(6)}$, which contradicts to $|T\cap T^{(3)}|=1$.
\end{proof}
\begin{proposition}\label{prop4}
Theorem~\ref{mainthm} holds for $n\equiv3\pmod{15}$ except for $n=3$.
\end{proposition}
\begin{proof}
By Equation~(\ref{eq17}), we obtain
\begin{align*}
TT^{(4)}\equiv& T^{(5)}+G+2TT^{(2)}\pmod{5}\\
\equiv &T^{(5)}+G+2X_1+4T^{(3)*}+T^{*}+3X_4\pmod{5}\\
\equiv&4e+T^{(5)*}+3X_{1}^{*}+2T^{*}+4X_4\pmod{5}.
\end{align*}
This implies that
\begin{align*}
&\sum_{i\ge0}|Y_{5i}|=|T^{(3)*}\backslash T^{(5)*}|+|T^{(5)*}\cap X_4|,\\
&\sum_{i\ge0}|Y_{5i+1}|=|T^{(5)*}\cap T^{(3)*}|,\\
&\sum_{i\ge0}|Y_{5i+2}|=|T^{*}\backslash T^{(5)*}|,\\
&\sum_{i\ge0}|Y_{5i+3}|=|X_{1}^{*}\backslash T^{(5)*}|+|T^{(5)*}\cap T^{*}|,\\
&\sum_{i\ge0}|Y_{5i+4}|=|X_4\backslash T^{(5)*}|+1+|T^{(5)*}\cap X_{1}^{*}|.
\end{align*}
Note that for any $g\in T^{(5)*}$, $g$ appears at least once in $TT^{(4)}$. By Lemma~\ref{lemma8}, we have
\begin{align*}
4n^2+4n+1=&\sum_{i=1}^{N}i|Y_i|\\
\ge&\sum_{i\ge0}|Y_{5i+1}|+2\sum_{i\ge0}|Y_{5i+2}|+3\sum_{i\ge0}|Y_{5i+3}|+4\sum_{i\ge0}|Y_{5i+4}|+5\sum_{i\ge1}|Y_{5i}|\\
\ge&2|T^{*}|+3|X_{1}^{*}|+4(|X_4|+1)+|T^{(5)*}|\\
=&2\cdot2n+3(\frac{4}{3}n^2-\frac{10}{3}n)+4(\frac{2}{3}n^2-\frac{2}{3}n+1)+2n.
\end{align*}
This leads to $8n^2-32n+9\le0$, which is only possible for $n=3$.
\end{proof}
\begin{proposition}\label{prop5}
Theorem~\ref{mainthm} holds for $n\equiv0\pmod{15}$.
\end{proposition}
\begin{proof}
By Equation~(\ref{eq17}), we obtain
\begin{align*}
TT^{(4)}\equiv& T^{(5)}+3G+3T+4TT^{(2)}\pmod{5}\\
\equiv &T^{(5)}+3G+3T++4X_1+3T^{(3)*}+2T^{*}+X_4\pmod{5}\\
\equiv&e+T^{(5)*}+2X_{1}^{*}+T^{(3)*}+3T^{*}+4X_4\pmod{5}.
\end{align*}
This implies that
\begin{align*}
&\sum_{i\ge0}|Y_{5i}|=|T^{(5)*}\cap X_4|,\\
&\sum_{i\ge0}|Y_{5i+1}|=|T^{(3)*}\backslash T^{(5)*}|+1,\\
&\sum_{i\ge0}|Y_{5i+2}|=|T^{(5)*}\cap T^{(3)*}|+|X_{1}^{*}\backslash T^{(5)*}|,\\
&\sum_{i\ge0}|Y_{5i+3}|=|T^{*}\backslash T^{(5)*}|+|T^{(5)*}\cap X_{1}^{*}|,\\
&\sum_{i\ge0}|Y_{5i+4}|=|T^{(5)*}\cap T^{*}|+|X_4\backslash T^{(5)*}|.
\end{align*}
Note that for any $g\in T^{(5)*}$, $g$ appears at least once in $TT^{(4)}$. By Lemma~\ref{lemma8}, we have
\begin{align*}
4n^2+4n+1=&\sum_{i=1}^{N}i|Y_i|\\
\ge&\sum_{i\ge0}|Y_{5i+1}|+2\sum_{i\ge0}|Y_{5i+2}|+3\sum_{i\ge0}|Y_{5i+3}|+4\sum_{i\ge0}|Y_{5i+4}|+5\sum_{i\ge1}|Y_{5i}|\\
\ge&|T^{(3)*}|+1+2|X_{1}^{*}|+3|T^{*}|+4|X_4|+|T^{(5)*}|\\
=&2n+1+2(\frac{4}{3}n^2-\frac{10}{3}n)+3\cdot 2n+4(\frac{2}{3}n^2-\frac{2}{3}n)+2n.
\end{align*}
This leads to $4n^2-10n\le0$, which is impossible.
\end{proof}
In order to solve the cases $n\equiv12\pmod{15}$ and $n\equiv9\pmod{15}$, we need to study $TT^{(5)}$.
By Equation~(\ref{eq17}), we have
\begin{equation}\label{eq19}
\begin{split}
T^6=&(2n+1)(16n^3+40n^2+20n+2)G+T^2T^{(4)}+(4n^2-6n+2)T^2+(4n-4)T^2T^{(2)}\\
=&(32n^4+96n^3+80n^2+24n+2)G+(2G+T^{(2)}+(2n-2)e)T^{(4)}+(4n^2-6n+2)(2G+T^{(2)}\\
&+(2n-2)e)+(4n-4)(2G+T^{(2)}+(2n-2)e)T^{(2)}\\
=&(32n^4+96n^3+104n^2+8n)G+T^{(2)}T^{(4)}+(2n-2)T^{(4)}+(12n^2-22n+10)T^{(2)}+(8n^3-20n^2+16n-4)e\\
&+(4n-4)(2G+T^{(4)}+(2n-2)e)\\
=&(32n^4+96n^3+104n^2+16n-8)G+(6n-6)T^{(4)}+(12n^2-22n+10)T^{(2)}+T^{(2)}T^{(4)}+(8n^3-12n^2+4)e.
\end{split}
\end{equation}
\begin{proposition}\label{prop6}
Theorem~\ref{mainthm} holds for $n\equiv12\pmod{15}$.
\end{proposition}
\begin{proof}
By Equation~(\ref{eq19}), we have
\begin{equation}\label{eq20}
\begin{split}
TT^{(5)}\equiv& T^{(4)}+4T^{(2)}+T^{(2)}T^{(4)}\pmod{5}\\
\equiv& e+T^{(4)*}+X_{1}^{(2)*}+2T^{(6)*}+2T^{(2)*}+4X_{4}^{(2)}\pmod{5}.
\end{split}
\end{equation}
Writing $TT^{(5)}$ as
\[TT^{(5)}=\sum_{i=0}^{L}iZ_i,\]
where $Z_i$, $(i=0,1,\dots,L)$ form a partition of group $G$. Let
\[|T^{(4)*}\cap X_{1}^{(2)*}|=a_1,\ |T^{(4)*}\cap T^{(6)*}|=a_2,\ |T^{(4)*}\cap X_{4}^{(2)}|=a_3.\]
Then $a_1+a_2+a_3=2n.$ From Equation~(\ref{eq20}), we get
\begin{align*}
&\sum_{i\ge0}|Z_{5i}|=a_3,\\
&\sum_{i\ge0}|Z_{5i+1}|=\frac{4}{3}n^2-\frac{10}{3}n+1-a_1,\\
&\sum_{i\ge0}|Z_{5i+2}|=4n-a_2+a_1,\\
&\sum_{i\ge0}|Z_{5i+3}|=a_2,\\
&\sum_{i\ge0}|Z_{5i+4}|=\frac{2}{3}n^2-\frac{2}{3}n-a_3.
\end{align*}
Note that for any $t\in T$, $t^4=t^{-1}\cdot t^5$, then $|Z_0|=0$.
We can compute to get that
\begin{align*}
4n^2+4n+1=&\sum_{i=1}^{L}i|Z_i|\\
\ge&\sum_{i\ge0}|Z_{5i+1}|+2\sum_{i\ge0}|Z_{5i+2}|+3\sum_{i\ge0}|Z_{5i+3}|+4\sum_{i\ge0}|Z_{5i+4}|+5\sum_{i\ge1}|Z_{5i}|\\
=&\frac{4}{3}n^2-\frac{10}{3}n+1-a_1+2(4n-a_2+a_1)+3a_2+4(\frac{2}{3}n^2-\frac{2}{3}n-a_3)+5a_3\\
=&4n^2+2n+1+a_1+a_2+a_3.
\end{align*}
This leads to $|Z_1|=\frac{4}{3}n^2-\frac{10}{3}n+1-a_1$, $|Z_2|=4n-a_2+a_1$, $|Z_3|=a_2$, $|Z_4|=\frac{2}{3}n^2-\frac{2}{3}n-a_3$, $|Z_5|=a_3$ and $|Z_i|=0$ for all $i\ge6$. We can also get $|T\cap T^{(5)}|=1$. With a similar discussion as Lemma~\ref{lemma2}, we have
\begin{align*}
2n^2+1=&\sum_{i=1}^{L}|Z_i|\\
=&1+\sum_{i=3}^{L}\frac{(i-1)(i-2)}{2}|Z_i|\\
=&1+a_2+3(\frac{2}{3}n^2-\frac{2}{3}n-a_3)+6a_3.
\end{align*}
This leads to $a_2+3a_3=2n$. Recall that $T^{(2)}T^{(4)}=X_{1}^{(2)}+2T^{(6)*}+3T^{(2)*}+4X_{4}^{(2)}$. On one hand, the elements in $T^{(4)*}$ appear $a_1+2a_2+4a_3=(a_1+a_2+a_3)+a_2+3a_3=4n$ times in $T^{(2)}T^{(4)}$. On the other hand, since $|T^{(4)*}\cap t^2T^{(4)*}|=|T^{(4)}\cap t^2T^{(4)}|\ge2$, then the elements in $T^{(4)*}$ appear at least $2n+2\cdot2n=6n$ times in $T^{(2)}T^{(4)}$, which is a contradiction.
\end{proof}
\begin{proposition}\label{prop7}
Theorem~\ref{mainthm} holds for $n\equiv9\pmod{15}$.
\end{proposition}
\begin{proof}
By Equation~(\ref{eq19}), we have
\begin{equation}\label{eq21}
\begin{split}
TT^{(5)}\equiv& G+3T^{(4)}+4T^{(2)}+T^{(2)}T^{(4)}+4e\pmod{5}\\
\equiv& 3e+3T^{(4)*}+2X_{1}^{(2)*}+3T^{(6)*}+3T^{(2)*}\pmod{5}.
\end{split}
\end{equation}
Writing $TT^{(5)}$ as
\[TT^{(5)}=\sum_{i=0}^{L}iZ_i,\]
where $Z_i$, $(i=0,1,\dots,L)$ form a partition of group $G$. Let
\[|T^{(4)*}\cap X_{1}^{(2)*}|=a_1,\ |T^{(4)*}\cap T^{(6)*}|=a_2,\ |T^{(4)*}\cap X_{4}^{(2)}|=a_3.\]
Then $a_1+a_2+a_3=2n.$ From Equation~(\ref{eq21}), we get
\begin{align*}
&\sum_{i\ge0}|Z_{5i}|=\frac{2}{3}n^2-\frac{2}{3}n-a_3+a_1,\\
&\sum_{i\ge0}|Z_{5i+1}|=a_2,\\
&\sum_{i\ge0}|Z_{5i+2}|=\frac{4}{3}n^2-\frac{10}{3}n-a_1,\\
&\sum_{i\ge0}|Z_{5i+3}|=4n+1-a_2+a_3,\\
&|Z_i|=0\text{ for }i\equiv4\pmod{5}.
\end{align*}
{\bf{Claim: $a_3\le\frac{4}{3}n.$}}
Recall that $T^{(2)}T^{(4)}=X_{1}^{(2)}+2T^{(6)*}+3T^{(2)*}+4X_{4}^{(2)}$. If $|T^{(4)*}\cap X_{4}^{(2)}|=a_3>\frac{4}{3}n$, then there exist $t_1,t_2,t_3\in T$ such that $t_{1}^{4}=t^2t_{4}^{4}$, $t_{2}^{4}=t^2t_{5}^{4}$ and $t_{3}^{4}=t^2t_{6}^{4}$ for some $t,t_4,t_5,t_6\in T$. This implies that $t^2=t_{1}^{4}t_{4}^{-4}=t_{2}^{4}t_{5}^{-4}=t_{3}^{4}t_{6}^{-4}$, and so $t_1=t_{5}^{-1}=t_{6}^{-1}$, $t_2=t_{4}^{-1}=t_{6}^{-1}$, $t_3=t_{4}^{-1}=t_{5}^{-1}$, which is a contradiction. This completes the proof of claim.
Note that for any $g\in T^{(4)*}\cap T^{(6)*}$, $g$ appears at least $6$ times in $TT^{(5)}$. Hence $|Z_1|=0$.
With a similar discussion as Lemma~\ref{lemma2}, we have
\begin{align*}
&4n^2+4n+1=\sum_{i=1}^{L}i|Z_i|,\\
&\sum_{i=1}^{L}|Z_i|=1-\frac{|T^{*}\cap T^{(5)*}|}{2}+\sum_{i=3}^{L}\frac{(i-1)(i-2)}{2}|Z_i|.
\end{align*}
Then we can get
\begin{equation}\label{eq22}
\begin{split}
&4n^2+4n+2-\frac{|T^{*}\cap T^{(5)*}|}{2}+\sum_{i=6}^{L}\frac{i(i-5)}{2}|Z_i|\\
=&\sum_{i=1}^{L}i|Z_i|+\sum_{i=1}^{L}|Z_i|-\sum_{i=3}^{L}\frac{(i-1)(i-2)}{2}|Z_i|+\sum_{i=6}^{L}\frac{i(i-5)}{2}|Z_i|\\
=&2|Z_1|+3|Z_2|+3|Z_3|+2|Z_4|\\
=&3|Z_2|+3|Z_3|\\
\le&3(\frac{4}{3}n^2-\frac{10}{3}n-a_1+4n+1-a_2+a_3)\\
=&4n^2+2n+3+3(a_3-a_1-a_2)\\
\le&4n^2+4n+3.
\end{split}
\end{equation}
By Equation (\ref{eq21}), we see that $|T\cap T^{(5)}|=3$ or $|T\cap T^{(5)}|\ge13$. If $|T\cap T^{(5)}|\ge13$, then $4n^2+4n+2-\frac{|T^{*}\cap T^{(5)*}|}{2}+\sum_{i=6}^{L}\frac{i(i-5)}{2}|Z_i|\ge4n^2+4n+2-6+\frac{13(13-5)}{2}=4n^2+4n+48$, which is a contradiction. Hence $|T\cap T^{(5)}|=3$. Then $|Z_i|=0$ for all $i\ge6$. From Equation~(\ref{eq22}), we obtain $4n^2+4n+1=3|Z_2|+3|Z_3|$. This implies $3\mid(4n^2+4n+1)$, which is a contradiction.
\end{proof}
\subsection{Some sporadic cases}
For $n=11$, Wei and Schwartz \cite{WS22} have constructed a lattice tiling of $\mathbb{Z}^n$ by $\mathcal{B}(11,2,1,1)$.
For $n=3,4,5$, the nonexistence of lattice tiling with $\mathcal{B}(n,2,1,1)$ follows from the following proposition.
\begin{proposition}\label{prop8}
Suppose that $2n^2+1=mp$ where $p$ is a prime and $p>2n+1$. Define $a=\min\{k\in\mathbb{Z}_{\ge0}:\ p\mid(4^k-4n-2)\}$ and $b$ is the order of $4$ modulo $p$ (If there is no $k$ with $p\mid(4^k-4n-2)$, then let $a=\infty$). Assume that there is a lattice tiling of $\mathbb{Z}^n$ by $\mathcal{B}(n,2,1,1)$, then there exists at least one $\ell\in\{0,1,\dots,\lfloor\sqrt{\frac{m-1}{2}}\rfloor\}$ such that the equation
\[a(x+1)+by=n-\ell\]
has nonnegative integer solutions.
\end{proposition}
\begin{proof}
The proof is similar as \cite{K17}, for readers convenience, we keep the proof.
Let $G$ be a finite abelian group (written additively) with order $2n^2+1$, and $0$ be its identity element.
By Theorem~\ref{tilinggroup}, there exists $T=\{t_i: i=1,\dots, n\}\subseteq G$ such that
\[ \{0\}, \{\pm t_i: i=1,\dots, n\}, \{\pm t_i\pm t_j: 1\leq i<j\leq n\} \]
form a partition of $G$.
Let $H$ be a subgroup of $G$ of index $p$. Let $\rho:G \rightarrow G/H$ be the canonical homomorphism and $a_i=\rho(t_i)$. Then the multisets
\[\{0\}, \{\pm a_i: i=1,\dots, n\}, \{\pm a_i\pm a_j: 1\leq i<j\leq n\} \]
form a partition of $mG/H$.
Let $k$ be an integer. By calculation,
\begin{align*}
&\sum_{i=1}^{n}\left((a_i^{2k} + (-a_i)^{2k} \right)+\sum_{1\leq i < j\leq n}\left( (a_i+a_j)^{2k} + (a_i-a_j)^{2k} + (-a_i+a_j)^{2k} +(-a_i-a_j)^{2k} \right)\\
=& (-4^{k} + 4n + 2)S_{2k} + 2\sum_{t=1}^{k-1}\binom{2k}{2t}S_{2t}S_{2(k-t)}
\end{align*}
where $S_t = \sum_{i=1}^{n}a_i^t$. Since this is also the sum of the $2k$-th powers of every element in $mG/H$,
\begin{equation}\label{eq:kim_main}
(-4^{k} + 4n + 2)S_{2k} + 2\sum_{t=1}^{k-1}\binom{2k}{2t}S_{2t}S_{2(k-t)}=
\begin{cases}
0, & p-1 \nmid 2k,\\
-m, & p-1 \mid 2k.
\end{cases}
\end{equation}
Let $a$ and $b$ be the least positive integers satisfying $p\mid (-4^a+4n+2)$ and $p\mid (4^b-1)$. Define
\[ X = \{ax+by : x\geq 1, y\geq 0\}. \]
We prove the following two claims by induction on $k$.
\textbf{Claim 1:} If $1\le k < \frac{p-1}{2}$ is not in $X$, then $S_{2k}=0$.
Suppose that $S_{2k}=0$ for each $k\le k_0-1$ that is not in $X$. Assume that $k_0\notin X$. As $X$ is closed under addition, for any $t$, at least one of $t$ and $k_0-t$ is not in $X$.
For any integer $k$, if $p\mid (-4^k+4n+2)$, then $k$ must be of the form $a+by$ whence $k\in X$. This implies that $p\nmid (-4^{k_0}+4n+2)$. By \eqref{eq:kim_main} and the induction hypothesis,
\[ 0=(-4^{k_0} + 4n + 2)S_{2k_0} + 2\sum_{t=1}^{k_0-1}\binom{2k_{0}}{2t}S_{2t}S_{2(k_0-t)}=(-4^{k_0} + 4n + 2)S_{2k_0}.\]
Thus $S_{2k_0}= 0$.
Let $e_k$ be the elementary symmetric polynomials with respect to $a_1^2$, $a_2^2$, $\cdots$, $a_n^2$.
\textbf{Claim 2}: If $1\le k\le n< \frac{p-1}{2}$ is not in $X$, then $e_k=0$.
Suppose that $e_k=0$ for each $k\leq k_0-1$ not in $X$ and $k_0\notin X$. As $X$ is closed under addition, for each $0<t<k_0$, at least one of $t$ and $k_0-t$ is not in $X$. By Claim 1 and the inductive hypothesis, $e_t=0$ or $S_{2(k_0-t)}=0$.
Together with Newton identities on $a_1^2,\dots, a_n^2$, we have
\[ k_0e_{k_0} = e_{k_0-1}S_2 + \dots + (-1)^{i+1} e_{k_0-i}S_{2i}+\dots +(-1)^{k_0-1}S_{2k_0}=(-1)^{k_0-1}S_{2k_0}=0.\]
Therefore $e_{k_0}=0$.
Suppose that $0$ appears $\ell$ times in $\rho(T)$, and let $a_{i_1}=a_{i_2}=\dots=a_{i_\ell}=0$, then $-a_{i_j}=0$ and $\pm a_{i_j}\pm a_{i_k}=0$ for $1\le j<k\le \ell$. This means that the number of $0$ in multiset $\rho(G)$ is at least $1+2\ell+4\binom{\ell}{2}=2\ell^{2}+1$.
As $0$ appears exactly $m$ times in $mG/H$, then $2\ell^{2}+1\le m$, i.e., $\ell\le\sqrt{\frac{m-1}{2}}$. Note that $e_{n-\ell}$ is the product of those nonzero $a_i^2$'s, whence $e_{n-\ell}\neq 0$. By Claim 2, $n-\ell$ is in $X$. This completes the proof.
\end{proof}
Theorem~\ref{mainthm} is a combination of Propositions \ref{prop1}, \ref{prop2}, and \ref{prop3}---\ref{prop8}.
\bibliographystyle{IEEEtranS}
|
\section{Introduction}
\input{introduction.tex}
\section{Physics-Informed Neural Networks}
\input{pinn.tex}
\section{Mesh Movement}
\input{mesh.tex}
\section{Results}
\input{results.tex}
\section{Conclusion and Future Work}
In this work, we solved mesh deformation problems with physics-informed neural networks. The selected method uses the linear elastic model since PINN can give accurate results for solving this type of PDE. We note that vertex nodes are moved according to boundary movement. Moreover, exact boundary values are enforced to satisfy the Dirichlet boundary conditions exactly. We test this approach with translation and rotation tests and compared it with finite element solutions. We showed that the PINN solution is comparable with the FEM solutions. The deformation is performed in numerous steps instead of a sudden movement. This prevents vertex collision and edge overlapping. We showed that as the number of steps is increased, the deformed mesh quality gets higher. For a greater mesh quality, the number of steps can be increased.
The mesh movement method in this paper only includes linear elastic equations, although it can be extended to other techniques. Other commonly used methods such as the Laplacian or biharmonic equations are also applicable to PINN formulation. Our future work aims to use other methods that prevent mesh overlapping in the training of the PINN. The network parameters and formulation can be extended in a way that vertex collisions and edge overlapping is prevented.
\bibliographystyle{ieeetr}
\subsection{Mesh Quality Metrics}
To be able to compare the effectiveness of different mesh movement techniques after a deformation, we use a mesh quality metric based on \cite{stein2003lineMesh}. In these metrics, the area and shape changes are considered by checking the element area and the aspect ratio. Both metric uses the initial mesh elements as reference elements and measures the change according to them. The element area change $f_A^e$ and shape change $f_{AR}^e$ is defined as :
\begin{subequations}
\label{eq:meshMetric}
\begin{align}
f_A^e &= \left\lvert \log \left(\frac{A^e}{A_o^e}\right) / \log (2.0) \right\rvert, \\
f_{AR}^e &= \left\lvert \log \left(\frac{AR^e}{AR_o^e}\right) / \log (2.0) \right\rvert.
\end{align}
\end{subequations}
Here, the superscript $e$ represents the specific element, and the subscript $o$ is the initial mesh element before the deformation occurs. $AR^e$ is the element aspect ratio defined in \cite{stein2003lineMesh} as:
\begin{equation}
AR^e = \frac{(l^e_{max})^2}{A^e}.
\end{equation}
Here, $l^e_{max}$ is the maximum edge length for the specific element. For comparison of different techniques, we use the global area and shape changes by considering the maximum values of element area and shape changes, respectively.
\subsection{Fully Connected Neural Networks}
A basic, fully connected deep neural network architecture can be used to solve differential equations \cite{lagaris1998nnPde}. Given an input vector $\mathbf{x} \in \mathbb{R}^d$, a single layer neural network gives an output $\hat{\mathbf{u}}$ by the following form:
\begin{equation}
\label{eq:singleLayerNN}
\hat{\mathbf{u}} = \sigma(\mathbf{W}_1\mathbf{x} + \mathbf{b}_1)\mathbf{W}_2 + \mathbf{b}_2,
\end{equation}
where $\mathbf{W}$ are the weight matrices and $\mathbf{b}$ are the bias vectors. $\sigma(\cdot)$ is a nonlinear function known as the activation function. In general, Sigmoid, hyperbolic tangent, and rectified linear unit (ReLU) are popular choices for the activation function. The hyperparameters $\theta = [\mathbf{W}, \mathbf{b}]$ are estimated by the following optimization problem
\begin{equation}
\label{eq:optimization}
\mathbf{\theta}^* = \argmin_\mathbf{\theta} J(\mathbf{\theta}; \mathbf{x}).
\end{equation}
Here, $J$ is the objective function to be minimized. In this work, this function is defined as the mean squared error of the prediction. This minimization problem in Equation \ref{eq:optimization} can be solved by using first-order stochastic gradient descent (SGD) algorithms \cite{ruder2016sgd}. In each iteration, the hyperparameters are updated in such a way,
\begin{equation}
\label{eq:SGD}
\theta^{i+1} = \theta^i - \eta^i \nabla_{\theta}J(\theta; \mathbf{x}),
\end{equation}
where $i$ being the current iteration and $\eta$ is the learning rate. The gradient of the loss function, $\nabla_{\theta}J(\theta; \mathbf{x})$, is calculated by backpropagation \cite{rumelhart1986backprop}.
For the physics-informed a neural network, we consider the general form of partial differential equations:
\begin{subequations}
\label{eq:PDEsystem}
\begin{align}
&\mathbf{u}_t + \mathcal{N}[u] = 0, \quad &\mathbf{x} \in \Omega,\; t\in[0,T]\label{eq:PDE}\\
&\mathbf{u}(\mathbf{x},0) = f(\mathbf{x}), \quad &\mathbf{x} \in \Omega\label{eq:PDE IC}\\
&\mathbf{u}(\mathbf{x},t) = g(\mathbf{x},t), \quad &\mathbf{x} \in \partial\Omega, \; t \in [0,T]\label{eq:PDE BC}
\end{align}
\end{subequations}
where $\mathcal{N}$ is a generalized differential operator that can be linear or nonlinear, $\mathbf{x} \in \mathbb{R}^d$ and $t \in [0, T]$ are the spatial and temporal coordinates. $\Omega \;\text{and}\; \partial\Omega$ represent the computational domain and the boundary respectively. $\mathbf{u}(\mathbf{x},t)$ is the general solution of the PDE with $f(\mathbf{x})$ is the initial condition and $g(\mathbf{x},t)$ is the boundary condition. The hidden solution, $\mathbf{u}(\mathbf{x}, t)$, can be approximated under the PINN framework proposed by Raissi et al. \cite{raissi_pinn_2019}, by a feedforward neural network $\hat{\mathbf{u}}(\mathbf{x}, t; \theta)$ with parameters $\mathbf{\theta}$. For the supervised training the only labeled data comes from the boundary/initial points. Inside the domain, the loss is determined by the PDE residual. By utilizing automatic differentiation (AD) \cite{baydin_automatic_2017}, PINNs can differentiate the network output w.r.t the input layer. AD applies the chain rule repeatedly to the elementary functions and arithmetic operations to achieve
the derivative of the overall composition. AD is well implemented in popular deep learning frameworks such as TensorFlow \cite{abadi2016tensorflow} and PyTorch \cite{paszke2019pytorch}.\
In classical PINN implementations, the loss term is a composite term including supervised data loss on the boundaries and initial points and the PDE loss. The total loss term can be written such that,
\begin{equation}
\label{eq:pinnLoss}
\mathcal{L} = w_{R}\mathcal{L}_{R} + w_{BC}\mathcal{L}_{BC} + w_{IC}\mathcal{L}_{IC}.
\end{equation}
Here the terms represent the boundary loss $\mathcal{L}_{BC}$, the initial condition loss $\mathcal{L}_{IC}$, and the PDE residual loss $\mathcal{L}_{R}$. The $w$ terms are specific weights of each loss term that can be user-specified or tuned manually or automatically \cite{wang2021_gradientflow, wang2022ntk_pinn}. Each loss term can be written as,
\begin{subequations}
\label{eq:LossFunctions}
\begin{align}
\mathcal{L}_R &= \frac{1}{N_R}\sum_{i=1}^{N_R}|\mathbf{u}_t + \mathcal{N}[\mathbf{u}(\mathbf{x}^i,t^i)]|^2\\
\mathcal{L}_{BC} &= \frac{1}{N_{BC}}\sum_{i=1}^{N_{BC}}|\mathbf{u}(\mathbf{x}^i,t^i) - g(\mathbf{x}^i,t^i)|^2\\
\mathcal{L}_{IC} &= \frac{1}{N_{IC}}\sum_{i=1}^{I_{BC}}|\mathbf{u}(\mathbf{x}^i,0) - f(\mathbf{x}^i)|^2.
\end{align}
\end{subequations}
Here $N_R, \; N_{BC}$, and $N_{IC}$ are the total number of points used for calculating the mean squared error used here as the loss function. The schematic of a classical PINN can be seen in the left part of Figure \ref{fig:PINN}.
\begin{figure}[htbp!]
\centering
\includegraphics[width=\textwidth]{figures/meshPinn2.pdf}
\caption{Schematic of PINN approach with exact boundary enforcement. The first PINN on the left shows the original formulation with weakly enforced Dirichlet boundary conditions. The second network uses the particular solution with exact boundary enforcement to satisfy Dirichlet boundaries exactly}
\label{fig:PINN}
\end{figure}
\noindent The hyperparameters $\mathbf{\theta}=[\mathbf{W}, \mathbf{b}]$ can be optimized by a chosen optimization algorithm to find the minimum total loss defined in Equation \ref{eq:pinnLoss}. As mentioned above, stochastic gradient descent algorithms are commonly used in neural network implementations \cite{ruder2016sgd}. This method aims to find new parameters $\theta$ in the opposite direction of the gradient of the objective function. The gradient of the loss function w.r.t. hyperparameters is calculated by backpropagation. In this work, we used the ADAM algorithm \cite{kingma2014adam} as the SGD optimizer.
\subsection{Exact Boundary Enforcement}
The optimization algorithm used in PINN tries to minimize the physics-based loss, $\mathcal{L}_R$. Using proper boundary and initial conditions can regularize the physics loss in deep neural networks. This classical PINN boundary condition implementation in Equation \ref{eq:LossFunctions} is named soft boundary enforcement \cite{sun2020hardbcPinn}. In this approach, the boundary prediction is minimized in the composite loss function. Although the SGD algorithms can minimize these loss functions, they do not satisfy the boundary values exactly. However, some PDE applications, such as mesh movement, need exact boundary values. For this purpose, we apply exact boundary enforcement. Sun et al. \cite{sun2020hardbcPinn} used this boundary condition enforcement to exactly satisfy the velocity and pressure values on the boundaries of internal flow cases with Navier-Stokes equations. Sukumar and Srivastava \cite{sukumar2022exact} introduced geometry-aware trial functions. They multiply the neural network output with these functions and use its generalization to exactly satisfy boundary conditions on complex geometries. In this work, we use this idea with multiple physics-informed neural networks to exactly satisfy Dirichlet boundary conditions. First, we trained a PINN with soft boundaries. For the mesh movement problem, the displacement vector ${\mathbf{u}} = [X, Y]^T$ will give the new coordinates of the nodes from the first neural network prediction. This solution is then changed on the boundaries with the exact values. This new solution is the particular solution of our approach. Then, a new PINN is trained with an output $\hat{\mathbf{u}}(\mathbf{x};\theta)$. This output is modified with the following equation.
\begin{equation}
\label{eq:hardBC}
\tilde{\mathbf{u}}(\mathbf{x}; \theta) = \mathbf{u}_{par}(\mathbf{x}) + D(\mathbf{x})\hat{\mathbf{u}}(\mathbf{x};\theta).
\end{equation}
Here, $\mathbf{u}_{par}$ is a particular solution that is a globally defined smooth function that only satisfies the boundary conditions. Any smooth function can be used for the particular solution such as radial basis functions (RBF) or linear functions \cite{sun2020hardbcPinn}. In this work, we use the classical PINN predictions with the soft boundary condition implementation as the particular solution. $D$ is a specified distance function from the boundary. Equation \ref{eq:hardBC} states that on the boundaries $D(\mathbf{x})=0$, the particular solution satisfies the exact boundary values, $\mathbf{u} = g$ on $\partial\Omega$. For a general approach, we used the shortest distance between the residual points and the boundaries. Since the geometric domains used in this paper are not too complex, this approach is not very time consuming. For complex geometries, approximate distance functions using R-functions \cite{sukumar2022exact} or pre-trained deep neural networks \cite{berg2018unified} can be used. This modified output contributes to the physics loss of the new PINN. In this network, the objective function is only consisting of the PDE residual $\mathcal{L}_R$ and trained with the same PDE. This approach allows us to exactly satisfy the Dirichlet boundary conditions using PINN.
\subsection{Deformed Square}
In this test case, a square domain is deformed from its boundaries. The square domain is $x,\: y\in [0,1]\times[0,1]$ and the unstructured mesh consists of 2744 triangular elements. The initial mesh can be seen in Figure \ref{fig.square_deformTop}. We want to find a deformed mesh where the position of the top boundary becomes $\hat{y} = y - 0.25\sin(\pi x)$. On the top surface, we implement this condition as a Dirichlet boundary condition as well as $\hat{x} = x$. All the other boundaries have the same Dirichlet boundary condition as $\hat{y}=y$, and $\hat{x}=x$. The deformed mesh can be seen in Figure \ref{fig.square_deformTop}. The figure in the middle shows the results obtained by only using classical PINN. This shows the boundaries, especially the corners, are not in the exact position and are deformed in an undesired way. The figure on the right shows the solution after exact boundary enforcement. The boundary values are corrected with the exact positions with the proposed approach. The $L_2$ error on the boundary nodes is calculated as 0.031. For this test case, we increased the specific weight of the boundary loss of the composite loss function in Equation \ref{eq:pinnLoss}. Since the deformation of the boundary is higher than the deformation of the computational domain, the boundary weight is increased. The weight ratio of the boundary loss and the residual loss is set to 25 to capture the boundary values more precisely.
\begin{figure}[hbtp!]
\begin{center}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/square/square_initMesh.pdf} \end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/square/square_soft.pdf}
\end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/square/square_hard.pdf}
\end{subfigure}
\end{center}
\caption{Initial and deformed meshes of the deformed square case with its deformed top boundary. The unstructured mesh consists of 2744 triangular elements The first deformed figure shows the solution with classical PINN. The last figure represents the solution with exact boundary enforcement.}
\label{fig.square_deformTop}
\end{figure}
The mesh quality measure of the deformed mesh based on the element area and shape changes can be seen in \ref{fig.squareMetric}. The top surface is deformed according to a sinusoidal function. The elements near the deformed boundary have the most change in size and shape as expected. Especially in the middle where the deformation is the largest, the elements are squeezed and get smaller. In the corners where the element vertices have two boundary conditions in each direction, the element area change is not significantly large. However, the shape of the corner elements changes more than the other elements on the boundary. These elements are bounded by the two boundaries and therefore the aspect ratios get larger. The deformation of the inner elements is relatively low, especially near the bottom boundary. The mesh deformation metrics get lower as the elements' position moves away from the deformed boundary. The global area and shape change metrics are calculated as $|f_A^{\infty}| = 0.744$, $|f_{AR}^{\infty}| = 1.264$, respectively.
\begin{figure}
\begin{center}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/square/square_size.pdf}
\caption{Area change}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/square/square_shape.pdf}
\caption{Shape change}
\end{subfigure}
~
\end{center}
\caption{Element quality metrics of the square with deformed top boundary. The figure on the left shows the element area change and the figure on the right shows the element shape change with respect to the initial mesh elements.}
\label{fig.squareMetric}
\end{figure}
To see the capabilities of our approach we further deform the bottom boundary with its coordinates $\hat{y} = y + 0.25\sin(\pi x)$. The Dirichlet boundary conditions on the stationary boundaries are the same as the other, $\hat{x}=x$, $\hat{y}=y$. The same specific weight ratio for the loss function of the PINN formulation is used. The deformed configuration can be seen in Figure \ref{fig.square_squeeze}. The figure in the middle is the solution with the classical PINN approach. The vertices on the boundaries are not in exact positions. Especially on the corners, the classical PINN solution has difficulty satisfying the positions. The $L_2$ error of the boundary positions is calculated as 0.076 for this case.
\begin{figure}
\begin{center}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/square/square_initMesh.pdf} \end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/square/squeeze_soft.pdf}
\end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/square/squeeze_hard.pdf}
\end{subfigure}
\end{center}
\caption{Initial and deformed meshes of the deformed square case. The square is squeezed from its top and bottom boundary. The first deformed figure shows the solution with classical PINN. The last figure represents the solution with exact boundary enforcement.}
\label{fig.square_squeeze}
\end{figure}
The elementwise quality measures of this case can be seen in Figure \ref{fig.squeezeMetric}. the elements on the top and the bottom boundaries are deformed the most, the same as in the previous case. The elements in the middle collapsed more than the case before. The global area and shape change values are $|f_A^{\infty}| = 1.701$, $|f_{AR}^{\infty}| = 1.845$, respectively. The element shape and size change significantly as the deformation is increased.
\begin{figure}
\begin{center}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/square/squeeze_size.pdf}
\caption{Area change}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/square/squeeze_shape.pdf}
\caption{Shape change}
\end{subfigure}
~
\end{center}
\caption{Element quality metrics of the square deformed from the top and bottom boundaries. The figure on the left shows the element area change and the figure on the right shows the element shape change with respect to the initial mesh elements.}
\label{fig.squeezeMetric}
\end{figure}
\subsection{Translation and Rotation tests}
To test the accuracy of our approach, translation and rotation tests in \cite{stein2003lineMesh} are performed. The original mesh can be seen in Figures \ref{fig.translationTest} and \ref{fig.rotationTest}. There is a line object located in $(-L, 0) \times (L,0)$ in a $(-2L, -2L)\times(2L, 2L)$ domain. A total of 2182 triangles are generated for the mesh.
For the translation tests, the object is moved $0.5L$ upwards. The movement is performed in 10 steps with $0.05L$ and in 5 steps with $0.1L$ movement upwards in two different training settings. The last step of the movement can be seen in Figure \ref{fig.translationTest}. In Figure \ref{fig.testPlots}, the PINN method is compared with the approach in \cite{stein2003lineMesh}. The area and shape change metrics of two PINN solutions are presented alongside the classical finite element solutions and solutions with Jacobian-based stiffening. The authors applied a stiffening power to prevent the deformation of the smaller elements. The stiffened approach represents the best value obtained in \cite{stein2003lineMesh} with different applied stiffening power. The two PINN solutions are representing the overall motion in 5 and 10 steps. The total number of steps is represented in parentheses in the figure. As seen in the first row of Figure \ref{fig.testPlots}, the PINN solutions are comparable with the FEM solutions with Jacobian-stiffening. As mentioned before, the PINN approach does not have any criteria to prevent mesh overlapping and sudden movements move the vertex nodes in an undesired way. Therefore, the quality of the deformed mesh improves as the number of steps increases.
\begin{figure}[hbpt!]
\begin{center}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/Line/initial_mesh.pdf}
\end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/Line/translation.pdf}
\end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width= \textwidth]{figures/Line/trans_5it.pdf}
\end{subfigure}
\end{center}
\caption{Initial mesh and deformed mesh after a total translation of $5$ units. The solution in the middle is performed in 10 steps while the solution on the right is performed in 5 steps.}
\label{fig.translationTest}
\end{figure}
For the rotation tests, the object is rotated $0.25\pi$ counterclockwise. Again, to prevent overlapping of edges and collision of vertices, the movement is performed in steps with $0.025\pi$ and $0.05\pi$ counterclockwise movement in each step in two different training. The last step of the rotation can be seen in Figure \ref{fig.rotationTest}. The deformed mesh differs especially on the boundaries between different PINN solutions. The small elements near the moving boundary start to collapse in the PINN solution with 5 steps. As the number of steps increases, the mesh quality increases. The comparison of the rotation tests with the same finite element solution of the translation tests is presented in Figure \ref{fig.testPlots}. The PINN approach again lies between the classical solution and the solution with Jacobian-based stiffening.
\begin{figure}[hbpt!]
\begin{center}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/Line/initial_mesh.pdf}
\end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/Line/rotation.pdf}
\end{subfigure}
~
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width= \textwidth]{figures/Line/rot_5it.pdf}
\end{subfigure}
\end{center}
\caption{Initial mesh and deformed mesh after a total rotation of $0.25\pi$. The solution in the middle is performed in 10 steps while the solution on the right is performed in 5 steps.}
\label{fig.rotationTest}
\end{figure}
\begin{figure}[hbpt!]
\begin{center}
\begin{subfigure}[b]{\textwidth}
\includegraphics[width=\textwidth]{figures/Line/translation_metric_bw.pdf}
\end{subfigure}
~
\begin{subfigure}[b]{\textwidth}
\includegraphics[width=\textwidth]{figures/Line/rotation_metric_bw.pdf}
\end{subfigure}
\end{center}
\caption{Global area and shape change metrics of the translation and rotation tests compared with the FEM solution in \cite{stein2003lineMesh}. The first row shows the comparison of the translation test, while the second row shows the comparison of the rotation tests.}
\label{fig.testPlots}
\end{figure}
\begin{table}[htb!]
\centering
\caption{Global area and shape changes of translation tests. The solution is performed in 10 steps. The values are given in every step.}
\begin{tabular}{ c | c c c c c c c c c c}
\hline \hline
$\Delta y$& 0.05 & 0.1 & 0.15 & 0.2 & 0.25 & 0.3 & 0.35 & 0.4 & 0.45 & 0.5 \\ \hline
{$\lvert f_A\rvert_{\infty}$} & 0.667 & 1.196 & 1.189 & 1.291 & 1.518 & 1.627 & 1.697 & 1.715 & 1.791 & 1.998 \\
{$\lvert f_{AR}\rvert_{\infty}$} & 0.596 & 1.125 & 1.118 & 1.220 & 1.447 & 1.556 & 1.626 & 1.663 & 1.921 & 2.258 \\
\hline \hline
\end{tabular}
\label{tab:translationMetric}
\end{table}
\begin{table}[htbp!]
\centering
\caption{Global area and shape changes of rotation tests. The solution is performed in 10 steps. The values are given in every step.}
\begin{tabular}{ c | c c c c c c c c c c}
\hline \hline
$\Delta\theta(\pi)$& 0.025 & 0.05 & 0.075 & 0.1 & 0.125 & 0.15 & 0.175 & 0.2 & 0.225 & 0.25 \\ \hline
{$\lvert f_A\rvert_{\infty}$} & 0.274 & 0.432 & 0.663 & 0.881 & 0.882 & 1.152 & 1.093 & 1.256 & 1.295 & 1.324 \\
{$\lvert f_{AR}\rvert_{\infty}$} & 0.355 & 0.508 & 0.753 & 1.034 & 1.266 & 1.554 & 1.914 & 2.137 & 2.517 & 2.573 \\
\hline \hline
\end{tabular}
\label{tab:rotationMetric}
\end{table}
In both tests, the global mesh quality metric presented in section \ref{ch:meshMotion} is used. The $\lvert f_A \rvert_{\infty}$ and $\lvert f_{AR} \rvert_{\infty}$ are calculated as the maximum area and shape change of the values in Equation \ref{eq:meshMetric} in every step. The area change and shape change values are presented in Tables \ref{tab:translationMetric} and \ref{tab:rotationMetric}, for the translation and rotation tests, respectively.
\subsection{Flexible Beam}
This test case consists of a mesh movement due to a motion of a flexible beam adapted from the problem in \cite{shamanskiy2021mesh}. The beam is fixed on its left end and sits in the center of the domain. Domain dimensions are $(-10,10)\times(-10,10)$ and the structure's position is $(-5,5)\times(-0.5,0.5)$ The deformation is based on a sinusoidal function $\sin (\frac{\pi}{2}\frac{x}{L})$ with varying amplitude. The initial mesh can be seen in Figure \ref{fig.wingInitMesh}. This unstructured mesh consists of 2098 triangular elements. The right end of the structure first moves to 4 units upwards, then 8 units downwards, following a 4-unit upward motion to return to its initial state. The movements are performed in steps with 2-unit motions, upwards or downwards.
\begin{figure}[hbtp!]
\begin{center}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/wing/original_mesh.pdf}
\end{subfigure}
~
~
\end{center}
\caption{Initial mesh of the flexible beam test case with 2098 triangular elements. The elements are concentrated on the moving boundary to track the deformation in a precise way.}
\label{fig.wingInitMesh}
\end{figure}
In Figure \ref{fig.wing2}, the deformed mesh after two steps of movement is presented with the mesh quality presented with the global area and shape change metric. Using exact boundary enforcement gives the true boundary position and therefore fixes the vertices on the boundaries. Therefore, on the outer boundaries, elements are stretched and squeezed more than the inner elements. Especially elements near the tip of the moving boundary have the most area and shape changes.
\begin{figure}[hbtp!]
\begin{center}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/wing/wing_2_elastic_size.pdf}
\caption{Area change}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/wing/wing_2_elastic_shape.pdf}
\caption{Shape change}
\end{subfigure}
~
\end{center}
\caption{Element quality metrics when the structure tip moves to $y=4$.}
\label{fig.wing2}
\end{figure}
In figure \ref{fig.wing8}, the mesh after one cycle of motion is presented. The structure returns to its original place after eight iterations. By looking at the area change, the sinusoidal motion of the structure can be observed. The most deformed elements are located at the top and bottom boundaries and near the moving tip of the structure. These elements are squeezed first and cannot recover themselves after the relative stretching.
\begin{figure}[hbtp!]
\begin{center}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/wing/wing_8_elastic_size.pdf}
\caption{Area change}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/wing/wing_8_elastic_shape.pdf}
\caption{Shape change}
\end{subfigure}
~
\end{center}
\caption{Element quality metrics when the structure returns its original position.}
\label{fig.wing8}
\end{figure}
|
\section{Introduction}
The {\it IJCAI--23 Proceedings} will be printed from electronic
manuscripts submitted by the authors. These must be PDF ({\em Portable
Document Format}) files formatted for 8-1/2$''$ $\times$ 11$''$ paper.
\subsection{Length of Papers}
All paper {\em submissions} to the main track must have a maximum of seven pages, plus at most two for references / acknowledgements / contribution statement / ethics statement.
The length rules may change for final camera-ready versions of accepted papers and
differ between tracks. Some tracks may disallow any contents other than the references in the last two pages, whereas others allow for any content in all pages. Similarly, some tracks allow you to buy a few extra pages should you want to, whereas others don't.
If your paper is accepted, please carefully read the notifications you receive, and check the proceedings submission information website\footnote{\url{https://proceedings.ijcai.org/info}} to know how many pages you can use for your final version. That website holds the most up-to-date information regarding paper length limits at all times.
\subsection{Word Processing Software}
As detailed below, IJCAI has prepared and made available a set of
\LaTeX{} macros and a Microsoft Word template for use in formatting
your paper. If you are using some other word processing software, please follow the format instructions given below and ensure that your final paper looks as much like this sample as possible.
\section{Style and Format}
\LaTeX{} and Word style files that implement these instructions
can be retrieved electronically. (See Section~\ref{stylefiles} for
instructions on how to obtain these files.)
\subsection{Layout}
Print manuscripts two columns to a page, in the manner in which these
instructions are printed. The exact dimensions for pages are:
\begin{itemize}
\item left and right margins: .75$''$
\item column width: 3.375$''$
\item gap between columns: .25$''$
\item top margin---first page: 1.375$''$
\item top margin---other pages: .75$''$
\item bottom margin: 1.25$''$
\item column height---first page: 6.625$''$
\item column height---other pages: 9$''$
\end{itemize}
All measurements assume an 8-1/2$''$ $\times$ 11$''$ page size. For
A4-size paper, use the given top and left margins, column width,
height, and gap, and modify the bottom and right margins as necessary.
\subsection{Format of Electronic Manuscript}
For the production of the electronic manuscript, you must use Adobe's
{\em Portable Document Format} (PDF). A PDF file can be generated, for
instance, on Unix systems using {\tt ps2pdf} or on Windows systems
using Adobe's Distiller. There is also a website with free software
and conversion services: \url{http://www.ps2pdf.com}. For reasons of
uniformity, use of Adobe's {\em Times Roman} font is strongly suggested.
In \LaTeX2e{} this is accomplished by writing
\begin{quote}
\mbox{\tt $\backslash$usepackage\{times\}}
\end{quote}
in the preamble.\footnote{You may want to also use the package {\tt
latexsym}, which defines all symbols known from the old \LaTeX{}
version.}
Additionally, it is of utmost importance to specify the {\bf
letter} format (corresponding to 8-1/2$''$ $\times$ 11$''$) when
formatting the paper. When working with {\tt dvips}, for instance, one
should specify {\tt -t letter}.
\subsection{Papers Submitted for Review vs. Camera-ready Papers}
In this document, we distinguish between papers submitted for review (henceforth, submissions) and camera-ready versions, i.e., accepted papers that will be included in the conference proceedings. The present document provides information to be used by both types of papers (submissions / camera-ready). There are relevant differences between the two versions. Find them next.
\subsubsection{Anonymity}
For the main track and some of the special tracks, submissions must be anonymous; for other special tracks they must be non-anonymous. The camera-ready versions for all tracks are non-anonymous. When preparing your submission, please check the track-specific instructions regarding anonymity.
\subsubsection{Submissions}
The following instructions apply to submissions:
\begin{itemize}
\item If your track requires submissions to be anonymous, they must be fully anonymized as discussed in the Modifications for Blind Review subsection below; in this case, Acknowledgements and Contribution Statement sections are not allowed.
\item If your track requires non-anonymous submissions, you should provide all author information at the time of submission, just as for camera-ready papers (see below); Acknowledgements and Contribution Statement sections are allowed, but optional.
\item Submissions must include line numbers to facilitate feedback in the review process . Enable line numbers by uncommenting the command {\tt \textbackslash{}linenumbers} in the preamble \footnote{New in IJCAI--23}.
\item The limit on the number of content pages is \emph{strict}. All papers exceeding the limits will be desk rejected.
\end{itemize}
\subsubsection{Camera-Ready Papers}
The following instructions apply to camera-ready papers:
\begin{itemize}
\item Authors and affiliations are mandatory. Explicit self-references are allowed. It is strictly forbidden to add authors not declared at submission time.
\item Acknowledgements and Contribution Statement sections are allowed, but optional.
\item Line numbering must be disabled. To achieve this, comment or disable {\tt \textbackslash{}linenumbers} in the preamble.
\item For some of the tracks, you can exceed the page limit by purchasing extra pages.
\end{itemize}
\subsection{Title and Author Information}
Center the title on the entire width of the page in a 14-point bold
font. The title must be capitalized using Title Case. For non-anonymous papers, author names and affiliations should appear below the title. Center author name(s) in 12-point bold font. On the following line(s) place the affiliations.
\subsubsection{Author Names}
Each author name must be followed by:
\begin{itemize}
\item A newline {\tt \textbackslash{}\textbackslash{}} command for the last author.
\item An {\tt \textbackslash{}And} command for the second to last author.
\item An {\tt \textbackslash{}and} command for the other authors.
\end{itemize}
\subsubsection{Affiliations}
After all authors, start the affiliations section by using the {\tt \textbackslash{}affiliations} command.
Each affiliation must be terminated by a newline {\tt \textbackslash{}\textbackslash{}} command. Make sure that you include the newline after the last affiliation, too.
\subsubsection{Mapping Authors to Affiliations}
If some scenarios, the affiliation of each author is clear without any further indication (\emph{e.g.}, all authors share the same affiliation, all authors have a single and different affiliation). In these situations you don't need to do anything special.
In more complex scenarios you will have to clearly indicate the affiliation(s) for each author. This is done by using numeric math superscripts {\tt \$\{\^{}$i,j, \ldots$\}\$}. You must use numbers, not symbols, because those are reserved for footnotes in this section (should you need them). Check the authors definition in this example for reference.
\subsubsection{Emails}
This section is optional, and can be omitted entirely if you prefer. If you want to include e-mails, you should either include all authors' e-mails or just the contact author(s)' ones.
Start the e-mails section with the {\tt \textbackslash{}emails} command. After that, write all emails you want to include separated by a comma and a space, following the order used for the authors (\emph{i.e.}, the first e-mail should correspond to the first author, the second e-mail to the second author and so on).
You may ``contract" consecutive e-mails on the same domain as shown in this example (write the users' part within curly brackets, followed by the domain name). Only e-mails of the exact same domain may be contracted. For instance, you cannot contract <EMAIL>" and <EMAIL>" because the domains are different.
\subsubsection{Modifications for Blind Review}
When submitting to a track that requires anonymous submissions,
in order to make blind reviewing possible, authors must omit their
names, affiliations and e-mails. In place
of names, affiliations and e-mails, you can optionally provide the submission number and/or
a list of content areas. When referring to one's own work,
use the third person rather than the
first person. For example, say, ``Previously,
Gottlob~\shortcite{gottlob:nonmon} has shown that\ldots'', rather
than, ``In our previous work~\cite{gottlob:nonmon}, we have shown
that\ldots'' Try to avoid including any information in the body of the
paper or references that would identify the authors or their
institutions, such as acknowledgements. Such information can be added post-acceptance to be included in the camera-ready
version.
Please also make sure that your paper metadata does not reveal
the authors' identities.
\subsection{Abstract}
Place the abstract at the beginning of the first column 3$''$ from the
top of the page, unless that does not leave enough room for the title
and author information. Use a slightly smaller width than in the body
of the paper. Head the abstract with ``Abstract'' centered above the
body of the abstract in a 12-point bold font. The body of the abstract
should be in the same font as the body of the paper.
The abstract should be a concise, one-paragraph summary describing the
general thesis and conclusion of your paper. A reader should be able
to learn the purpose of the paper and the reason for its importance
from the abstract. The abstract should be no more than 200 words long.
\subsection{Text}
The main body of the text immediately follows the abstract. Use
10-point type in a clear, readable font with 1-point leading (10 on
11).
Indent when starting a new paragraph, except after major headings.
\subsection{Headings and Sections}
When necessary, headings should be used to separate major sections of
your paper. (These instructions use many headings to demonstrate their
appearance; your paper should have fewer headings.). All headings should be capitalized using Title Case.
\subsubsection{Section Headings}
Print section headings in 12-point bold type in the style shown in
these instructions. Leave a blank space of approximately 10 points
above and 4 points below section headings. Number sections with
Arabic numerals.
\subsubsection{Subsection Headings}
Print subsection headings in 11-point bold type. Leave a blank space
of approximately 8 points above and 3 points below subsection
headings. Number subsections with the section number and the
subsection number (in Arabic numerals) separated by a
period.
\subsubsection{Subsubsection Headings}
Print subsubsection headings in 10-point bold type. Leave a blank
space of approximately 6 points above subsubsection headings. Do not
number subsubsections.
\paragraph{Titled paragraphs.} You should use titled paragraphs if and
only if the title covers exactly one paragraph. Such paragraphs should be
separated from the preceding content by at least 3pt, and no more than
6pt. The title should be in 10pt bold font and to end with a period.
After that, a 1em horizontal space should follow the title before
the paragraph's text.
In \LaTeX{} titled paragraphs should be typeset using
\begin{quote}
{\tt \textbackslash{}paragraph\{Title.\} text} .
\end{quote}
\subsection{Special Sections}
\subsubsection{Appendices}
You may move some of the contents of the paper into one or more appendices that appear after the main content, but before references. These appendices count towards the page limit and are distinct from the supplementary material that can be submitted separately through CMT. Such appendices are useful if you would like to include highly technical material (such as a lengthy calculation) that will disrupt the flow of the paper. They can be included both in papers submitted for review and in camera-ready versions; in the latter case, they will be included in the proceedings (whereas the supplementary materials will not be included in the proceedings).
Appendices are optional. Appendices must appear after the main content.
Appendix sections must use letters instead of Arabic numerals. In \LaTeX, you can use the {\tt \textbackslash{}appendix} command to achieve this followed by {\tt \textbackslash section\{Appendix\}} for your appendix sections.
\subsubsection{Ethical Statement}
Ethical Statement is optional. You may include an Ethical Statement to discuss the ethical aspects and implications of your research. The section should be titled \emph{Ethical Statement} and be typeset like any regular section but without being numbered. This section may be placed on the References pages.
Use
\begin{quote}
{\tt \textbackslash{}section*\{Ethical Statement\}}
\end{quote}
\subsubsection{Acknowledgements}
Acknowledgements are optional. In the camera-ready version you may include an unnumbered acknowledgments section, including acknowledgments of help from colleagues, financial support, and permission to publish. This is not allowed in the anonymous submission. If present, acknowledgements must be in a dedicated, unnumbered section appearing after all regular sections but before references. This section may be placed on the References pages.
Use
\begin{quote}
{\tt \textbackslash{}section*\{Acknowledgements\}}
\end{quote}
to typeset the acknowledgements section in \LaTeX{}.
\subsubsection{Contribution Statement}
Contribution Statement is optional. In the camera-ready version you may include an unnumbered Contribution Statement section\footnote{New in IJCAI--23}, explicitly describing the contribution of each of the co-authors to the paper. This is not allowed in the anonymous submission. If present, Contribution Statement must be in a dedicated, unnumbered section appearing after all regular sections but before references. This section may be placed on the References pages.
Use
\begin{quote}
{\tt \textbackslash{}section*\{Contribution Statement\}}
\end{quote}
to typeset the Contribution Statement section in \LaTeX{}.
\subsubsection{References}
The references section is headed ``References'', printed in the same
style as a section heading but without a number. A sample list of
references is given at the end of these instructions. Use a consistent
format for references. The reference list should not include publicly unavailable work.
\subsubsection{Order of Sections}
Sections should be arranged in the following order:
\begin{enumerate}
\item Main content sections (numbered)
\item Appendices (optional, numbered using capital letters)
\item Ethical statement (optional, unnumbered)
\item Acknowledgements (optional, unnumbered)
\item Contribution statement (optional, unnumbered)
\item References (required, unnumbered)
\end{enumerate}
\subsection{Citations}
Citations within the text should include the author's last name and
the year of publication, for example~\cite{gottlob:nonmon}. Append
lowercase letters to the year in cases of ambiguity. Treat multiple
authors as in the following examples:~\cite{abelson-et-al:scheme}
or~\cite{bgf:Lixto} (for more than two authors) and
\cite{brachman-schmolze:kl-one} (for two authors). If the author
portion of a citation is obvious, omit it, e.g.,
Nebel~\shortcite{nebel:jair-2000}. Collapse multiple citations as
follows:~\cite{gls:hypertrees,levesque:functional-foundations}.
\nocite{abelson-et-al:scheme}
\nocite{bgf:Lixto}
\nocite{brachman-schmolze:kl-one}
\nocite{gottlob:nonmon}
\nocite{gls:hypertrees}
\nocite{levesque:functional-foundations}
\nocite{levesque:belief}
\nocite{nebel:jair-2000}
\subsection{Footnotes}
Place footnotes at the bottom of the page in a 9-point font. Refer to
them with superscript numbers.\footnote{This is how your footnotes
should appear.} Separate them from the text by a short
line.\footnote{Note the line separating these footnotes from the
text.} Avoid footnotes as much as possible; they interrupt the flow of
the text.
\section{Illustrations}
Place all illustrations (figures, drawings, tables, and photographs)
throughout the paper at the places where they are first discussed,
rather than at the end of the paper.
They should be floated to the top (preferred) or bottom of the page,
unless they are an integral part
of your narrative flow. When placed at the bottom or top of
a page, illustrations may run across both columns, but not when they
appear inline.
Illustrations must be rendered electronically or scanned and placed
directly in your document. They should be cropped outside \LaTeX{},
otherwise portions of the image could reappear during the post-processing of your paper.
When possible, generate your illustrations in a vector format.
When using bitmaps, please use 300dpi resolution at least.
All illustrations should be understandable when printed in black and
white, albeit you can use colors to enhance them. Line weights should
be 1/2-point or thicker. Avoid screens and superimposing type on
patterns, as these effects may not reproduce well.
Number illustrations sequentially. Use references of the following
form: Figure 1, Table 2, etc. Place illustration numbers and captions
under illustrations. Leave a margin of 1/4-inch around the area
covered by the illustration and caption. Use 9-point type for
captions, labels, and other text in illustrations. Captions should always appear below the illustration.
\section{Tables}
Tables are treated as illustrations containing data. Therefore, they should also appear floated to the top (preferably) or bottom of the page, and with the captions below them.
\begin{table}
\centering
\begin{tabular}{lll}
\hline
Scenario & $\delta$ & Runtime \\
\hline
Paris & 0.1s & 13.65ms \\
Paris & 0.2s & 0.01ms \\
New York & 0.1s & 92.50ms \\
Singapore & 0.1s & 33.33ms \\
Singapore & 0.2s & 23.01ms \\
\hline
\end{tabular}
\caption{Latex default table}
\label{tab:plain}
\end{table}
\begin{table}
\centering
\begin{tabular}{lrr}
\toprule
Scenario & $\delta$ (s) & Runtime (ms) \\
\midrule
Paris & 0.1 & 13.65 \\
& 0.2 & 0.01 \\
New York & 0.1 & 92.50 \\
Singapore & 0.1 & 33.33 \\
& 0.2 & 23.01 \\
\bottomrule
\end{tabular}
\caption{Booktabs table}
\label{tab:booktabs}
\end{table}
If you are using \LaTeX, you should use the {\tt booktabs} package, because it produces tables that are better than the standard ones. Compare Tables~\ref{tab:plain} and~\ref{tab:booktabs}. The latter is clearly more readable for three reasons:
\begin{enumerate}
\item The styling is better thanks to using the {\tt booktabs} rulers instead of the default ones.
\item Numeric columns are right-aligned, making it easier to compare the numbers. Make sure to also right-align the corresponding headers, and to use the same precision for all numbers.
\item We avoid unnecessary repetition, both between lines (no need to repeat the scenario name in this case) as well as in the content (units can be shown in the column header).
\end{enumerate}
\section{Formulas}
IJCAI's two-column format makes it difficult to typeset long formulas. A usual temptation is to reduce the size of the formula by using the {\tt small} or {\tt tiny} sizes. This doesn't work correctly with the current \LaTeX{} versions, breaking the line spacing of the preceding paragraphs and title, as well as the equation number sizes. The following equation demonstrates the effects (notice that this entire paragraph looks badly formatted, and the line numbers no longer match the text):
\begin{tiny}
\begin{equation}
x = \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \prod_{i=1}^n \sum_{j=1}^n j_i
\end{equation}
\end{tiny}%
Reducing formula sizes this way is strictly forbidden. We {\bf strongly} recommend authors to split formulas in multiple lines when they don't fit in a single line. This is the easiest approach to typeset those formulas and provides the most readable output%
\begin{align}
x = & \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \nonumber \\
+ & \prod_{i=1}^n \sum_{j=1}^n j_i.
\end{align}%
If a line is just slightly longer than the column width, you may use the {\tt resizebox} environment on that equation. The result looks better and doesn't interfere with the paragraph's line spacing: %
\begin{equation}
\resizebox{.91\linewidth}{!}{$
\displaystyle
x = \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \prod_{i=1}^n \sum_{j=1}^n j_i
$}.
\end{equation}%
This last solution may have to be adapted if you use different equation environments, but it can generally be made to work. Please notice that in any case:
\begin{itemize}
\item Equation numbers must be in the same font and size as the main text (10pt).
\item Your formula's main symbols should not be smaller than {\small small} text (9pt).
\end{itemize}
For instance, the formula
\begin{equation}
\resizebox{.91\linewidth}{!}{$
\displaystyle
x = \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j + \prod_{i=1}^n \sum_{j=1}^n j_i + \prod_{i=1}^n \sum_{j=1}^n i_j
$}
\end{equation}
would not be acceptable because the text is too small.
\section{Examples, Definitions, Theorems and Similar}
Examples, definitions, theorems, corollaries and similar must be written in their own paragraph. The paragraph must be separated by at least 2pt and no more than 5pt from the preceding and succeeding paragraphs. They must begin with the kind of item written in 10pt bold font followed by their number (e.g.: {\bf Theorem 1}),
optionally followed by a title/summary between parentheses in non-bold font and ended with a period (in bold).
After that the main body of the item follows, written in 10 pt italics font (see below for examples).
In \LaTeX{} we strongly recommend that you define environments for your examples, definitions, propositions, lemmas, corollaries and similar. This can be done in your \LaTeX{} preamble using \texttt{\textbackslash{newtheorem}} -- see the source of this document for examples. Numbering for these items must be global, not per-section (e.g.: Theorem 1 instead of Theorem 6.1).
\begin{example}[How to write an example]
Examples should be written using the example environment defined in this template.
\end{example}
\begin{theorem}
This is an example of an untitled theorem.
\end{theorem}
You may also include a title or description using these environments as shown in the following theorem.
\begin{theorem}[A titled theorem]
This is an example of a titled theorem.
\end{theorem}
\section{Proofs}
Proofs must be written in their own paragraph(s) separated by at least 2pt and no more than 5pt from the preceding and succeeding paragraphs. Proof paragraphs should start with the keyword ``Proof." in 10pt italics font. After that the proof follows in regular 10pt font. At the end of the proof, an unfilled square symbol (qed) marks the end of the proof.
In \LaTeX{} proofs should be typeset using the \texttt{\textbackslash{proof}} environment.
\begin{proof}
This paragraph is an example of how a proof looks like using the \texttt{\textbackslash{proof}} environment.
\end{proof}
\section{Algorithms and Listings}
Algorithms and listings are a special kind of figures. Like all illustrations, they should appear floated to the top (preferably) or bottom of the page. However, their caption should appear in the header, left-justified and enclosed between horizontal lines, as shown in Algorithm~\ref{alg:algorithm}. The algorithm body should be terminated with another horizontal line. It is up to the authors to decide whether to show line numbers or not, how to format comments, etc.
In \LaTeX{} algorithms may be typeset using the {\tt algorithm} and {\tt algorithmic} packages, but you can also use one of the many other packages for the task.
\begin{algorithm}[tb]
\caption{Example algorithm}
\label{alg:algorithm}
\textbf{Input}: Your algorithm's input\\
\textbf{Parameter}: Optional list of parameters\\
\textbf{Output}: Your algorithm's output
\begin{algorithmic}[1]
\STATE Let $t=0$.
\WHILE{condition}
\STATE Do some action.
\IF {conditional}
\STATE Perform task A.
\ELSE
\STATE Perform task B.
\ENDIF
\ENDWHILE
\STATE \textbf{return} solution
\end{algorithmic}
\end{algorithm}
\section{\LaTeX{} and Word Style Files}\label{stylefiles}
The \LaTeX{} and Word style files are available on the IJCAI--23
website, \url{https://ijcai-23.org/}.
These style files implement the formatting instructions in this
document.
The \LaTeX{} files are {\tt ijcai23.sty} and {\tt ijcai23.tex}, and
the Bib\TeX{} files are {\tt named.bst} and {\tt ijcai23.bib}. The
\LaTeX{} style file is for version 2e of \LaTeX{}, and the Bib\TeX{}
style file is for version 0.99c of Bib\TeX{} ({\em not} version
0.98i). The {\tt ijcai23.sty} style differs from the {\tt
ijcai22.sty} file used for IJCAI--22.
The Microsoft Word style file consists of a single file, {\tt
ijcai23.docx}. This template differs from the one used for
IJCAI--22.
These Microsoft Word and \LaTeX{} files contain the source of the
present document and may serve as a formatting sample.
Further information on using these styles for the preparation of
papers for IJCAI--23 can be obtained by contacting {\tt
<EMAIL>}.
\section{Introduction}
In the last several years, neural networks have increasingly been employed to provide a way to learn representations that generalize well in dense, low dimensional domains. Specifically in the field of computer graphics, coordinate based multi-layer perceptrons (MLPs) with ReLU activations have been vital for applications such as neural radiance fields (NeRF)~\cite{Mildenhall2020nerf} and shape occupancy~\cite{Mescheder2018occupancy}. These networks pass in dense, low dimensional input coordinates, and output the corresponding scene representation of color, shape, or density for various visual signals. Coordinate based MLPs are intriguing due to their severe spectral bias, meaning they are practically incapable of learning the high frequency components of the target signal~\cite{Mildenhall2020nerf}. This limitation can be overcome through a positional encoding of coordinates \(\boldsymbol{\mathrm{v}} \in \mathbb{R}^d\) comprised of high frequency sinusoids:
\[\gamma(\boldsymbol{\mathrm{v}}) = \left[\sin(2^0\pi \boldsymbol{\mathrm{v}}),\cos(2^0\pi \boldsymbol{\mathrm{v}}),\cdots,\sin(2^L\pi \boldsymbol{\mathrm{v}}),\cos(2^L\pi \boldsymbol{\mathrm{v}})\right],\]
where \(L \in \mathbb{N}\) determines the maximum frequency~\cite{tancik2020fourier,Rahimi2007}. This behavior calls into question the limitations of ReLU in dense settings, its relation to a large spectral bias, and how positional encoding overcomes it.
\begin{figure}
\centering
\def0.1{5}
\subfloat[Coordinates]{%
\includegraphics[width=0.235\textwidth]{raw_xy_fractal.jpg}%
}\hfil
\subfloat[Positional Encoding]{%
\includegraphics[width=0.235\textwidth]{sin_cos_fractal.jpg}%
}
\caption{Visualization of the large spectral bias induced by dense low dimensional 2D coordinates (a), which is overcome by a high frequency positional encoding (b). For this task, the coordinate based MLP is tasked with learning the RGB value corresponding to the 2D pixel location, or its sinusoidal embedding. Coordinate based inputs have difficulty converging to the high frequency components of the image, and can only generate a smooth, low frequency representation.}
\label{fig:visualization}
\end{figure}
Spectral bias is the behavior such that a simpler, lower frequency representation of the data is learned before the high frequency components, or local details. This phenomenon is of interest due to its impact on generalization, since it is believed that an implicit regularization biased towards lower frequency solutions avoids overfitting in over-parameterized networks. While spectral bias is generally believed to aid in generalization, in the coordinate based regime it becomes so severe that the network will essentially underfit the target signal (Figure 1).
Further understanding the nature of spectral bias will be crucial in determining its impact on generalization. So far, Neural Tangent Kernel (NTK)~\cite{jacot2018neural} and Fourier analysis have been the primary tools for analyzing spectral bias, both in the coordinate based regime and more generally~\cite{tancik2020fourier,rahaman2019spectral}. While NTK models gradient descent dynamics via a kernel method (whose kernel matrix becomes a constant in the limit of very wide layers), the dynamic properties of the real network are not accurately captured. Fourier analysis has demonstrated a correlation between the network's Lipschitz upper bound and the slower convergence to high frequency components. However, this method is global in nature, making it difficult to study the network's dynamics locally. Specifically, we want to study the relationship between parameter updates and the network's activation regions, which defines its computations across input space. While the Fourier decomposition contains the total number of linear regions, their direct impact on gradient descent convergence for components of varying frequency has not been clearly demonstrated.
In this paper, we directly model spectral bias through the computations of ReLU networks. Our analysis highlights their limitations with dense low dimensional coordinates, and captures the properties that induce spectral bias in the uniform setting. Specifically, we relate expressive capacity and the ability to speed convergence of gradient descent, using a metric termed gradient confusion~\cite{Sankararaman2020}. We find that confusion is higher (slower convergence) when expressive power between inputs is limited, and reduced (faster convergence) when expressive power is enhanced. This results in slower convergence when using coordinate based inputs, as their density restricts many inputs to the same or nearby activation regions during training. Positional encoding alleviates this density while simultaneously generating more regions, substantially reducing confusion overall. This property also applies to the varying components within the target signal, as we find confusion is more prominent in local regions (high frequency components) compared to spaced out intervals (low frequency components) for both methods, resulting in the observed spectral bias.
We focus on the coordinate based regime since the effects of spectral bias are extreme, and positional encoding provides further validation on the limitations the network overcomes. An additional benefit of this method is that the activation regions can be directly analyzed, and we find they contain distinct properties as the frequency of the encoding increases. A study of these properties and their relation to overfitting may provide insight as to how spectral bias impacts generalization in this setting.
The paper is structured as follows. Section 2 provides related work. In Section 3, we define activation regions and explore how density restricts the practical expressive capacity of the network. Section 4 discusses gradient confusion, and connects confusion with the network's activation regions. In Section 5, we analyze the properties of the activation regions induced by higher frequency encodings, and Section 6 demonstrates how dense coordinates turn off ReLU neurons during training. Section 7 is a conclusion which details future directions for this approach.
\section{Related Work}
In~\cite{rahaman2019spectral,basri2020frequency}, it was shown that neural networks have a bias towards learning low frequency functions first, referred to as spectral bias, and that a sinusoidal mapping can allow for higher frequencies to be learned faster. This method was adopted by NeRF~\cite{Mildenhall2020nerf} in order to speed converge of high frequency components during novel view synthesis. As a consequence of this, the relationship between coordinates and positional encoding was analyzed by the same authors using NTK~\cite{tancik2020fourier}. They found that the eigenspectrum of the NTK decays rapidly with dense coordinates, and widens for positional encoding. This allows for faster convergence along the directions of the corresponding eigenfunctions.
Aside from its severity in coordinate based MLPs, spectral bias has been studied independently due to its link to good generalization in over-parameterized networks.~\cite{rahaman2019spectral} provide the first rigorous exploration of spectral bias using Fourier analysis. In~\cite{Cao2019}, it was shown how the eigenfunctions of the NTK have their own convergence rate given by their corresponding eigenvalue, using inputs with uniform density on \(\mathbb{S}^1\). The authors in~\cite{basri2020frequency} also utilized NTK to give convergence rates for inputs of non-uniform density in \(\mathbb{S}^1\). There has been a line of work referring to spectral bias as the F-principle~\cite{Xu_2020}, in which Fourier analysis was utilized for high dimensional inputs in sigmoidal nets. Other works have attempted to compute spectral bias and determine convergence rates as well~\cite{Kiessling_Thor_2022,Basri2019convergence}.
Our analysis of spectral bias utilizes activation/linear regions. Exploring the complexity of functions computable by piece-wise linear networks initially explored in~\cite{pascanu2013}. This approach was expanded upon in~\cite{Montufar2014}, where upper and lower bounds for the maximal number of linear regions was given, and shown to be exponential with depth. In~\cite{Raghu2017}, activation regions were defined, and bounds were computed by utilizing input trajectories \(x(t)\) across the regions. More recently, ~\cite{Hanin2019complexity,Hanin2019} attempted to compute more practical bounds for the number of activation regions, which was found to be independent on depth.~\cite{Zhang2020EmpiricalSO} provides useful experimental quantities for studying the properties of linear regions in a more practical manner.
\begin{figure}[ht]
\centering
\def0.1{0.5}
\subfloat[Coordinate Activation
\\Regions]{%
\includegraphics[width=0.24\textwidth]{raw_xy_activation_visual.png}%
}\hfil
\subfloat[Encoding Activation
\\Regions]{%
\includegraphics[width=0.24\textwidth]{sin_cos_act_regions.png}%
} \\
\subfloat[Coordinate Hyperplanes]{\includegraphics[width=0.45\textwidth]{raw_xy_neuron_hyperplanes.png}} \\
\subfloat[Encoding Hyperplanes]{\includegraphics[width=0.45\textwidth]{sin_cos_neruon_hyperplanes.png}}
\caption{Visualization of the regions in which each input belongs to (a-b) (colored according to unique region) at the beginning of training. Many coordinates lie in the same region, but positional encoding allows for each input to lie in a unique region. The hyperplane arrangement of neurons in first layer are displayed in (c) and (d).}
\label{fig:activation_regions}
\end{figure}
In this paper, we analyze the models systematically and utilize carefully designed experimental results to study the properties of activation regions and compare them to the correlation of gradients during training. We specifically focus on gradient confusion~\cite{Sankararaman2020}, which was shown to be higher for networks of larger depth, but has not been explored in relation to expressive capacity. Other works have utilized correlations between gradients in different settings as well~\cite{Balduzzi2017,Fort2019stiffness}.
\section{Limited Expressive Power}
In the following sections, we provide an alternative interpretation of spectral bias in coordinate based MLPs by directly analyzing their expressivity in relation to gradient descent dynamics. All experiments were conducted on a two layer MLP with 128 hidden neurons and standard uniform initialization using the Adam optimizer, a learning rate of .001, and mini-batches of size 256. We regress a 64x64 image containing random RGB values which corresponds to a 2D signal with high frequency components, where coordinates are typically normalized between \([0,1]\) or \([-1, 1]\). The architecture and signal were chosen in order to clearly demonstrate the limitations of coordinates, since the target values fluctuate rapidly throughout. Similar experiments are conducted for different architectures and natural images in the appendix, where our results are further validated. Note that certain dynamics such as confusion densities will differ slightly based on the signal being regressed and the amount of high frequency components it contains, but the overarching relationships hold.
We analyze the expressive power of a neural network with dense and uniform inputs through its activation regions. We begin by defining activation regions and patterns. Let \(f_{\theta} : \mathbb{R}^{n_{\mathrm{in}}} \rightarrow \mathbb{R}^{n_{\mathrm{out}}}\) be a continuous piece-wise linear function given by a ReLU network containing \(L\) hidden layers with \(N\) hidden neurons per layer. The network is defined as the composition of affine transformations \(f_l(x) = \boldsymbol{W}^l x + \boldsymbol{b}^l\) with ReLU activations \(\sigma = \max(0, f_l)\) such that
\[f_{\theta} = f_{\mathrm{out}} \circ \sigma \circ f_L \circ \sigma \circ \ldots \circ \sigma \circ f_1(x),\]
where \(\theta\) denotes the vector of trainable parameters. Each neuron \(z^l_i(x) = \boldsymbol{W}^l_i x + \boldsymbol{b}^l_i\) composed with \(\sigma\) denotes a hyperplane equation with the scalar determined by the bias. The collection of these hyperplanes in the first layer gives a hyperplane arrangement in \(\mathbb{R}^{n_{\mathrm{in}}}\), splitting the input space into pieces that compute distinct linear functions. As the image of each distinct linear function in the preceding layer \(l-1\) is uniquely partitioned by the following neurons \(z^l_i(x)\) (or not), the \(N-1\) dimensional hyperplanes in layers \(l > 1\) will appear to bend. Overall, this leads to a ReLU network partitioning the input space into convex polytopes on which unique linear functions are computed. One way to evaluate these polytopes is through activation regions, which are determined by the network's activation patterns, defined as
\begin{figure}
\centering
\def0.1{0.1}
\subfloat[Training Loss]{%
\includegraphics[width=0.25\textwidth]{loss_plot.png}%
}
\subfloat[Activation Region Growth]{%
\includegraphics[width=0.25\textwidth]{act_growth.png}%
}
\caption{Training loss (a) and activation region growth (b) for coordinates and encoding. We only consider the activation regions utilized for this signal (4096 regions) in order to display their growth during training. For this particular architecture, the network is incapable of mapping each coordinate to a unique region.}
\label{fig:act_growth}
\end{figure}
\begin{figure}
\centering
\def0.1{0.1}
\subfloat[Coordinates]{%
\includegraphics[width=0.163\textwidth]{no_encoding_dist.png}%
}
\subfloat[Encoding L=5]{%
\includegraphics[width=0.163\textwidth]{L5_dist.png}%
}
\subfloat[Encoding L=16]{%
\includegraphics[width=0.163\textwidth]{L16_dist.png}%
}
\caption{Matrix of coordinate distances (a) and encoding distances (b-c) for different frequencies. Positional encoding induces a kernel with a strong diagonal since it is stationary. }
\label{fig:distances}
\end{figure}
\begin{definition}[Activation Pattern] Let \(f_{\theta}\) be a ReLU network, and \(z^l_i(x)\) denote the pre-activation of a neuron in the \(l\)th layer. Then, an activation pattern \(\mathcal{A}\) of the network is a vector in \(\{0, 1\}^{\# \mathrm{neurons}}\) composed of neuron activations such that \(0\) is assigned if \(z^l_i(x) < 0\) and \(1\) if \(z^l_i(x) > 0\).
\end{definition}
\begin{definition}[Activation Region] For a ReLU network \(f_{\theta}\), the activation regions \(\mathcal{R}\) of \(f_{\theta}\) are the sets of input samples that correspond to the same \(\mathcal{A}\),
\begin{multline*}
\mathcal{R}(f, \mathcal{A}) = \{ x \in \mathbb{R}^{n_{in}} \;\; | \;\; \mathbbm{1}_{\mathbb{R}^{+}}(\mathrm{ReLU}(f^l(x))) = a_l, \;\; \\
\forall l \in 1,...,k, \;\; \forall a_l \in \mathcal{A} \}
\end{multline*}
where \(a_l\) is the activation pattern of layer \(l\).
\end{definition}
Analyzing these activation regions are important because they provide a measure of the expressive power of the network, or the complexity of functions it can compute. In~\cite{Hanin2019}, a realistic upper bound on the maximum number of activation regions created by a ReLU network architecture was given as
\begin{equation}
\frac{\#\mathrm{Regions \;in \;\mathcal{F} \;intersecting\; \mathcal{C}}}{\mathrm{vol}(\mathcal{C})} \leq \frac{(TN)^{n_{\mathrm{in}}}}{n_{\mathrm{in}}!}
\end{equation}
where \(N\) is the number of neurons, \(T\) is a constant, and \(\mathcal{C}\) is a cube in input space.
Equation (1) is determined by the number of neurons and the input dimension. Therefore, the dimensionality of the inputs themselves show that the network with positional encoding should have more expressive capacity. However, input density is a major factor as well, therefore we instead analyze the number of regions the network utilizes for a particular dataset in addition to their distinctiveness/distance. We can define the set of interest on a dataset \(\mathcal{D} = \{x_i, y_i\}^n_{i=1}\), given in terms of the corresponding activation patterns as \[\mathcal{L_{\mathcal{D}}} = \{ \mathbbm{1}_{\mathbb{R}^{+}}(\mathrm{ReLU}(f^l(x))) \; | \; x \in
\mathcal{D}, \forall l \in 1,...,k \},\]
equipped with the hamming distance \(\sum_{i=1}^n|a_i - b_i|, \;\) for \(a, b \in \mathcal{L_{\mathcal{D}}}\) to determine their distinctiveness. Note that the hamming distance would amount to the total number boundaries in which two inputs lie on opposing sides.
In Figures 2 and 3, we get a glimpse of the limitations imposed by dense, low dimensional inputs. For this given architecture, we can see the number of activation regions utilized for coordinates is lower than the number of elements in the dataset (Figure 3 (b)), meaning many inputs will be restricted to the same linear function throughout training. On the other hand, positional encoding can easily map each element to a unique region using the same architecture. The hyperplane arrangements in the first layer are displayed in Figure 2 (c) and (d), demonstrating the complex higher dimensional split that positional encoding allows in two dimensions as opposed to coordinates. Thus, we can see that coordinate based inputs have very little expressive power during training in comparison to positional encoding.
The reasoning for this reduced expressive power is a result of both density and a lack of available regions in low input dimensions. Coordinates impose a grid over a subset of \(\mathbb{R}^d\) in which the resolution, given by their spacing, is very fine-grained. Therefore, fitting enough neurons in the first layer such that a substantial amount of hyperplanes can separate each input becomes difficult. If the input to layers \(l > 1\) are restricted to the images of similar (or the same) linear functions, then the benefits that arise from the unique bending of each ensuing hidden neuron becomes less apparent, which can lead to many inputs lying in nearby regions. With positional encoding, the density is alleviated since the distance \(\lambda = ||\gamma(x_i) - \gamma(x_j)||\) for each \(x_i, x_j\) can be scaled by the frequency, which simultaneously raises its dimensionality and generates more regions. The density of inputs can be interpreted through the induced kernel function, which is known to be stationary for positional encoding~\cite{tancik2020fourier}. This means that the similarity between inputs is only dependent on their distance, inducing a strong diagonal that is missing with coordinate based inputs (Figure 4).
We can relate the density of inputs to the frequency of the target function. Higher density results in a target function containing higher frequency components, and vice versa. Thus, positional encoding can also be viewed as generating a lower frequency function for the network to learn. This gives a relationship between frequency (or density) and expressive power. The generation of a lower frequency function will disperse inputs across the available regions, which is not possible with the fine-grained sampling of coordinates. While it may seem obvious that inputs in a higher dimensional space will have larger distances, for this problem a proper deterministic encoding of lower dimensional coordinates is necessary in order to alleviate their high density, and it is possible to map them to higher dimensions without accomplishing this. For example, in the appendix we demonstrate that using the same frequency along each encoding dimension will result in a similar fine-grained sampling to that found in coordinate space, and expressive capacity is still limited.
\section{High Gradient Confusion}
In~\cite{Sankararaman2020}, the effect of network architecture on speed of convergence was modeled using gradient confusion, a measure determining the correlation of gradients for differing inputs. It was shown that lower confusion during training can speed convergence of SGD. Confusion occurs when two objective functions \(\mathcal{L}(f_{\theta}(x_i), y_i)\) and \(\mathcal{L}(f_{\theta}(x_j), y_j)\), \(i \neq j\), have gradients such that \(\langle \nabla \mathcal{L}(f_{\theta}(x_i), y_i), \nabla \mathcal{L}(f_{\theta}(x_j), y_j) \rangle < 0\). This creates a disagreement on the direction the parameters need to move, slowing down convergence. Convergence rates of SGD were given through the confusion bound \(\eta \geq 0\), given as
\[\langle \nabla \mathcal{L}(f_{\theta}(x_i), y_i), \nabla \mathcal{L}(f_{\theta}(x_j), y_j) \rangle \geq -\eta\]
for all \(i \neq j\) and fixed \(\theta\). We focus on this metric since we are interested in the speed at which gradient descent converges and its relation to the spatial information of the signal.
\subsection{Intuition for Higher Confusion}
In this section, we aim to build intuition as to why limited expressive power will induce higher amounts of confusion when the target values oscillate rapidly. Assume we are regressing a 1D signal, and let \(f^{\mathcal{A}}\) be a linear function given by a ReLU network corresponding to an activation pattern \(\mathcal{A}\). For this particular region, \(f^{\mathcal{A}}\) is composed of weight matrices \(\boldsymbol{W}^{n_{\mathrm{in}} \times \#\mathrm{act}}\) where each row denotes a neuron in which the corresponding \(a^l_i \in \mathcal{A}\) is 1. Assume two dense inputs \(x_i, x_j \in [0, 1]\) belong to the same \(\mathcal{R}(f, \mathcal{A})\). We can then write \(x_j = c_0x_i\), and \(\mathrm{sgn}(w_i^{(l)T}(c^{l-1}_i \odot x^{l-1}_i) + b_i^l) = \mathrm{sgn}(w_i^{(l)T}x^{l-1}_i + b_i^l)\) must hold for all neurons, where \(c^l\) is a vector of positive values. Therefore,
\begin{equation}
\begin{split}
\left<\nabla_{\theta}f^{\mathcal{A}}(x_i), \nabla_{\theta}f^{\mathcal{A}}(x_j)\right> &= \\
&\left<\nabla_{\theta}f^{\mathcal{A}}(x_i), d \odot \nabla_{\theta}f^{\mathcal{A}}(x_i)\right> > 0,
\end{split}
\end{equation}
where \(d \in \mathbb{R}^{\mathrm{\# weights}}\) are the positive numbers that scale the hidden output.
\begin{figure}
\centering
\subfloat[Confusion Bound Local]{\includegraphics[width=0.23\textwidth]{min_confusion_within.png}}
\subfloat[Confusion Bound Global]{\includegraphics[width=0.23\textwidth]{min_confusion_between.png}}
\vspace{.001mm}
\subfloat[Confusion Density\\ Local]{\includegraphics[width=0.26\textwidth]{local_confusion.pdf}}
\subfloat[Confusion Density\\ Global]{\includegraphics[width=0.242\textwidth]{confusion_global.pdf}}
\caption{Confusion of gradient descent in local 3x3 regions of the signal and globally. (a) and (b) plot confusion bounds, and (c) and (d) plot the overall confusion density. We used 100 local regions for (a) and (c), and 10,000 distant pairs for (b) and (d). Confusion is measured through the cosine similarity between input gradients pairs.}
\end{figure}
We will now evaluate using Mean Squared Error (MSE) loss \[\frac{1}{N}\sum_{i=1}^N (y_i - f(x_i))^2.\]
We have the gradient of MSE with respect to the parameters for a single input \(x\) as
\[\nabla_{\theta} \mathcal{L}(f^{\mathcal{A}}(x), y) = -2(y - f^{\mathcal{A}}(x))\nabla_{\theta}f^{\mathcal{A}}(x).\]
From (2), we can see that confusion will be directly induced by residual \(y - f^{\mathcal{A}}(x)\). If \(y_i\) is highly oscillatory for each \(x_i \in \mathcal{R}(f, \mathcal{A})\), then confusion will occur in any situation where \(y_i > f^{\mathcal{A}}(x_i)\) and \(y_j < f^{\mathcal{A}}(x_j)\). Since the direction of the gradient update for each weight is essentially determined by the residual, confusion can only be reduced if the target values change linearly or possibly if constant.
Now, assume we have inputs \(x_i, x_j \in [0,1]\) such that they are regressed by two separate linear function \(f^{\mathcal{A}}\) and \(f^{\mathcal{B}}\) of patterns \(\mathcal{A}\) and \(\mathcal{B}\). Let \(S = \{\boldsymbol{W_1}^{\mathcal{A}} \cap \boldsymbol{W_1}^{\mathcal{B}}, \boldsymbol{W_2}^{\mathcal{A}} \cap \boldsymbol{W_2}^{\mathcal{B}}, ... , \boldsymbol{W}_L^{\mathcal{A}} \cap \boldsymbol{W_L}^{\mathcal{B}}\}\) denote the set of commonly activated weights between layers. The inner product of the gradients will only depend on the parameters in \(S\) since all other weights will cancel out due to ReLU \[\langle \nabla_{S}\mathcal{L}(f^{\mathcal{A}}(x_i), y_i), \nabla_{S}\mathcal{L}(f^{\mathcal{B}}(x_j), y_j)\rangle.\] While the residual of \(\nabla_{S}\mathcal{L}\) may still be highly oscillatory, it is not necessarily the case that \(\langle \nabla_{S}f^{\mathcal{A}}(x_i), \nabla_{S}f^{\mathcal{B}}(x_j) \rangle > 0\), as backpropagation will utilize weights \(\boldsymbol{W}^\mathcal{A}, \boldsymbol{W}^\mathcal{B} \not\in S\), along with their corresponding hidden outputs. We can view this as the network utilizing information from distinct activation regions when optimizing shared weights. Essentially, the network can configure the regions such that updating shared weights in the same direction, which will now impact each region uniquely, can be beneficial for minimizing the loss. The more these patterns are allowed to differ between inputs, the more flexible the network can be when finding these solutions. Thus, limited expressive power contributes to higher amounts of confusion when the target values fluctuate rapidly, and more expressive capacity allows for a reduction of confusion overall.
\begin{figure}
\subfloat[Hamming Distance Local]{\includegraphics[width=0.246\textwidth]{mean_hamming_within.png}}
\subfloat[Hamming Distance Global]{\includegraphics[width=0.246\textwidth]{mean_hamming_across.png}}
\caption{Mean hamming distance of activation patterns between inputs in local 3x3 neighborhoods of input space (a) and globally across input space (b). The hamming distance allows us to directly relate expressive capacity to gradient confusion, as a larger hamming distance between activation patterns should allow for more flexibility when finding unique gradient solutions.}
\label{fig:hamming_distance}
\end{figure}
\subsection{Spectral Bias}
The relationship between expressive capacity and confusion is validated by the confusion densities and bounds plotted in Figure 5, and the mean hamming distance between activation patterns plotted in Figure 6. The hamming distance allows us to measure both the number of opposing hyperplane splits and the number of non-common weights that can be utilized during backpropagation. From Figure 6, we can see that coordinates utilize less distinct activation patterns overall, resulting in higher amounts of confusion across the signal as shown in Figure 5. Positional encoding can increase the distinctiveness between activation patterns (from the reasoning in section 3), drastically reducing confusion in comparison to coordinates. However, for both methods there are less distinct activation patterns in local regions of the signal, which is where confusion is generally concentrated.
We can see that the density of local regions causes the high frequency components to lie in increasingly similar activation regions in comparison to the low frequency components, which reduces confusion across spaced out intervals of the signal and increases it between neighboring inputs. This behavior generates the observed low frequency representations, since the network converges to the overall structure of the target function before the local details. The reduction of these higher frequency components through positional encoding allows for much faster convergence overall, although the behavior still occurs.
While our reasoning should always apply in the uniform case, these dynamics may differ for inputs of non-uniform density, where local densities and their relation to activation regions will come into play. Additionally, more expressive power may not necessarily be more useful if the signal is dominated by low frequency components, where confusion can be avoided easily. More work will need to be done to apply this method to the non-uniform case, extending the work of~\cite{basri2020frequency}.
\begin{figure}
\centering
\subfloat[L=5 \\(epoch 1)]{\includegraphics[width=0.165\textwidth]{angles_L5_epoch1.png}}
\subfloat[L=8 \\(epoch 1)]{\includegraphics[width=0.165\textwidth]{angles_L8_epoch1.png}}
\subfloat[L=16 \\(epoch 1)]{\includegraphics[width=0.165\textwidth]{angles_L16_epoch1.png}}
\vspace{.01mm}
\subfloat[L=5 \\(epoch 5000)]{\includegraphics[width=0.165\textwidth]{angles_L5_epoch5000.png}}
\subfloat[L=8 \\(epoch 5000)]{\includegraphics[width=0.165\textwidth]{angles_L8_epoch5000.png}}
\subfloat[L=16 \\(epoch 5000)]{\includegraphics[width=0.165\textwidth]{angles_L16_epoch5000.png}}
\caption{Cosine similarity between hyperplane normals during training across layers. The normal to each hyperplane is simply given as the
corresponding weights.}
\end{figure}
\section{Frequency and Linear Regions}
We now explore the distinct activation region dynamics induced by higher frequency encodings. We start by analyzing the correlation between hyperplane directions, given as \(\nabla_x z(x)\), by taking the cosine similarity between all weights (Figure 7). We find that higher frequency encodings allow for an increase in orthogonal hyperplane directions in the beginning of training (higher dimensional), then become increasingly parallel as training progresses. This mostly happens in the first layer, but occurs in the second layer as well.
This behavior may induce two distinct properties. First, it should allow for the gradients to become more positively correlated, which is demonstrated in the confusion densities shown in Figure 5. The higher frequency encodings have more positively correlated gradients, mostly for the high frequency components (Figure 5 (c)), while the other methods tend to center around zero. Second, this can result in wider activation regions, which may reduce the distinctiveness between activation patterns in exchange for positively correlated solutions. This reduction in distinctiveness is shown in Figure 6 for higher frequency encodings during training, as there is a dip in the mean hamming distance. This dip is also larger for the low frequency components in comparison to the high frequency components.
\begin{figure}
\centering
\includegraphics[width=.48\textwidth]{act_regions_encoding.pdf}
\caption{2D slice of activation regions for high and low frequency encoding after 5000 epochs. Images were found by moving along a 2D plane (all other dimensions fixed) at the origin, with "Low' denoting a 2D plane corresponding to the low frequency components of the encoding, and "High" the high frequency components.}
\label{fig:slice}
\end{figure}
To validate that the above behavior results in wider activation regions, we visualize 2D slices of the activation regions for a trained network in Figure 8. Higher frequency encodings result in larger regions compared to lower frequency encodings. Figure 9 further validates this through plotting the distance of each input to the nearest boundary \(d(x, \mathcal{B}_f) = \min_{z(x) \in f}|z(x) - b_z| / ||\nabla z(x)||\)~\cite{Hanin2019complexity}, which can indicate the narrowness of the regions. For lower frequency encodings, there is a larger contraction of this distance during training. In~\cite{Hanin2019complexity}, this contraction was linked to better test accuracy. Depending on the signal, higher frequency encodings are known to overfit~\cite{tancik2020fourier}. Therefore, these results may provide insight on the properties that cause this overfitting with higher frequency inputs.
Overall, higher frequency encodings allow the network to find solutions in which many hyperplanes can face the same direction, which is increasingly difficult for lower frequency encodings. This property may allow for positively correlated gradients and a reduction of confusion on the training data, albeit adverse generalization performance. More insight as to how this behavior impacts generalization is being further investigated.
\begin{figure}
\centering
\includegraphics[width=.4\textwidth]{narrowness.png}
\caption{The mean distance of training points to boundaries \(\mathcal{B}_f\) of ReLU neurons (no mini-batches). This gives a measure of the narrowness of regions in the network.}
\label{fig:narrowness}
\end{figure}
\section{Dying ReLU}
Although coordinate based networks slowly increase the number of activation regions throughout training as shown in Figure 3 (b), there is a surprising alternative that is simultaneously occurring. That is, the network is increasingly \emph{shutting off} ReLU neurons during training, limiting its full expressive potential while simultaneously attempting to utilize different linear functions with available neurons. We believe this is due to the exponential increase in parameter norms that occur when training on high frequency target functions, which can cause issues due to the density of the data.
We computed the parameter norms \(\prod_{n=1}^L ||\boldsymbol{W}^{(k)}||\) across layers, where \(|| \cdot ||\) is the spectral norm found by the maximum singular value, displayed in Figure 10. This norm represents the upper bound of the Lipschitz constant \(L_f\) of the network, and it may be slightly counter-intuitive to see that this bound is much higher for the network with a larger spectral bias. In this setting, it may be the case that larger parameters can easily map coordinates to one side of the bias threshold, causing it to be inactive for the rest of training. We can see some intuition of this idea from Figure 10 (d). The number of dead ReLU neurons decreases as the normalizing interval increases, meaning once the data becomes less dense there are more active neurons. While this may be the case, there is still the overall trend of an increasing amount of dead neurons, which displays another limitation of ReLU networks in this particular setting.
\begin{comment}
\section{Encoding Methodology}
Here, we seek to explore positional encoding using ideas from signal processing. Particularly, we view the neural network as an encoding-decoding process in which the encoding process extracts/learns a sequence of features, while the decoding process reassembles the learned features to reconstruct the signal.
\subsection{Encoding-decoding decomposition}
Given any positional encoding $\gamma:\mathbb{R}^{n_0}\rightarrow \mathbb{R}^{L}$ with $\gamma=[\gamma_1,\gamma_2,\cdots,\gamma_L]$,
suppose that a multilayer perceptron (MLP) $f:\mathbb{R}^{n_0}\rightarrow\mathbb{R}^{n_k}$ with this positional encoding is of the shape $(n_0,L,n_1,\cdots,n_k)$.
Then we can represent the whole neural network as $f=g\circ h$ where $g:\mathbb{R}^{n_1}\rightarrow\mathbb{R}^{n_k}$ is a nonlinear continuous function and $h:\mathbb{R}^{n_0}\rightarrow\mathbb{R}^{n_1}$ is a high dimensional function $h = [h_1,h_2,\cdots,h_{n_1}]$ where each $h_i$ is a linear combination of $[\gamma_1,\cdots,\gamma_L]$:
$$h_i(x)=\sum_{j=0}^L w_i^j\gamma_j(x),$$
where $\gamma_0(x)\equiv 1$.
Positional encoding extracts features $h_i$ from the input by tuning coefficients $w_i^j$. Let $\theta$ denote the parameters of $g$, now we look at how $g(\cdot;\theta)$ reassembles these features $h_i$ to obtain the output.
\begin{lemma}\label{lem:taylor}
Assume $f(x;\theta,w)=g(h(x;w);\theta)$ as above, we have
$$f(x;\theta,w)\approx f(x;\theta_0,w_0)+\frac{\partial g}{\partial\theta}\Delta\theta+\sum_{i,j}\frac{\partial g}{\partial h_i}\Delta w_i^j \gamma_j(x).$$
\end{lemma}
\begin{proof}
This follows directly from the Taylor expansion and chain rule.
\end{proof}
In Lemma~\ref{lem:taylor}, $\frac{\partial g}{\partial\theta}\Delta\theta$ represents the information we can learn from a normal MLP. With positional encodings, the high frequency components among each $\gamma_j$ gets weighted and accumulates to the final output.
Assuming that $w\sim\mathcal{P}_1$ and $\theta\sim\mathcal{P}_2$ (typically the $\mathcal{P}_i$ are standard normal distributions). Then we denote NTKs of $f(x;w,\theta)$ and $g(y;\theta)$ as:
\begin{align*}
K_{NTK}^f(x_1,x_2)&:=\mathbb{E}_{w\sim\mathcal{P}_1,\theta\sim\mathcal{P}_2} \big\langle \frac{\partial f}{\partial(w,\theta)}\big\vert_{x_1}, \frac{\partial f}{\partial(w,\theta)}\big\vert_{x_2} \big\rangle\\
K_{NTK}^g(y_1,y_2)&:=\mathbb{E}_{\theta\sim\mathcal{P}_2}\big\langle\frac{\partial g}{\partial y}\big\vert_{y_1}, \frac{\partial g}{\partial y}\big\vert_{y_2}\big\rangle
\end{align*}
\begin{lemma}\label{lem:ntk}
Assume $f(x;\theta,w)=g(h(x;w);\theta)$ as above, we have
\begin{multline*}
K^f_{NTK}(x_1,x_2)= K^g_{NTK}(h(x_1),h(x_2))+ \\ \mathbb{E}_{\theta\sim\mathcal{P}_2}\big\langle\frac{\partial g}{\partial h}\big\vert_{h(x_2)}, \frac{\partial g}{\partial h}\big\vert_{h(x_1)}\big\rangle \big\langle\gamma(x_1),\gamma(x_2)\big\rangle.
\end{multline*}
\end{lemma}
\begin{proof}
Given $f(x;w,\theta) = g\big(h(x;w);\theta\big)$, we have
\begin{align*}
\frac{\partial f}{\partial\theta}=\frac{\partial g}{\partial \theta}, \,\,\,\,&\,\,\,\,
\frac{\partial f}{\partial w_i^j} =\frac{\partial g}{\partial h_i}\frac{\partial h_i}{\partial w_i^j}=\frac{\partial g}{\partial h_i}\gamma_j.
\end{align*}
Hence,
\begin{align*}
K_{NTK}^f&(x_1,x_2) =\mathbb{E}_{w\sim\mathcal{P}_1,\theta\sim\mathcal{P}_2} \big\langle \frac{\partial f}{\partial\theta}\big\vert_{x_1}, \frac{\partial f}{\partial\theta}\big\vert_{x_2} \big\rangle + \\
& \; \; \; \; \; \; \; \; \; \; \; \; \; \; \; \; \; \; \; \; \mathbb{E}_{w\sim\mathcal{P}_1,\theta\sim\mathcal{P}_2}\big\langle \frac{\partial f}{\partial w}\big\vert_{x_1}, \frac{\partial f}{\partial w}\big\vert_{x_2} \big\rangle\\
&= \mathbb{E}_{\theta\sim\mathcal{P}_2}\big\langle \frac{\partial g}{\partial\theta}\big\vert_{h(x_1)}, \frac{\partial g}{\partial\theta}\big\vert_{h(x_2)}\big\rangle + \\ & \; \; \; \; \; \; \mathbb{E}_{\theta\sim\mathcal{P}_2}\sum_{i,j}\frac{\partial g}{\partial h_i}\big\vert_{h(x_1)} \frac{\partial g}{\partial h_i}\big\vert_{h(x_2)} \gamma_j(x_1)\gamma_j(x_2) \\
&= K_{NTK}^g(h(x_1),h(x_2)) + \\ & \; \; \; \; \; \; \mathbb{E}_{\theta\sim\mathcal{P}_2}\sum_{i}\frac{\partial g}{\partial h_i}\big\vert_{h(x_1)} \frac{\partial g}{\partial h_i}\big\vert_{h(x_2)} \sum_j\gamma_j(x_1)\gamma_j(x_2)\\
&= K_{NTK}^g(h(x_1),h(x_2)) + \\ & \; \; \; \; \; \; \mathbb{E}_{\theta\sim\mathcal{P}_2}\big\langle\frac{\partial g}{\partial h}\big\vert_{h(x_2)}, \frac{\partial g}{\partial h}\big\vert_{h(x_1)}\big\rangle \big\langle\gamma(x_1),\gamma(x_2)\big\rangle
\end{align*}
\end{proof}
In~\cite{yang2019fine,tancik2020fourier,lee2019wide}, the authors use NTK theory to explain frequency decay as eigencomponents of $K_{NTK}^f$ corresponding to larger eigenvalues being learned faster. However, if none of the eigenfunctions of $K_{NTK}^f$ carry high frequency features, then the MLP will not be able to reconstruct the high frequency features regardless of training time. Using high frequency encoding functions, by Lemma~\ref{lem:ntk}, we are able to impose high frequency components into $K_{NTK}^f$.. Even though an MLP with positional encoding is able to learn high frequency features, the performance differs with the choice of information encoded.
\begin{figure*}
\begin{center}
\begin{tabular}{c || c | c || c || c}
& Natural Image $\uparrow$ & Text Image & NeRF & 3D shape $\uparrow$ \\ [.5ex]
\hline
Fourier & 26.19 & 29.52 & 25.56 & 0.960\\ [.5ex]
Fourier-\(\mathcal{N}\) & 27.29 & 31.33 & 25.76 & 0.988\\ [.5ex]
Gabor & \(\boldsymbol{27.47}\) & \(\boldsymbol{32.00}\) & \(\boldsymbol{}\) & \\ [.5ex]
MFN & 26.47 & 31.00 & 25.81 & -\\ [.5ex]
SIREN & 27.17 & 31.30 & &\\ [.5ex]
\end{tabular}
\caption{Generalization performance measured in PSNR for each positional encoding on natural and text images, as well as 3D NeRF.}
\end{center}
\end{figure*}
\section{Gabor Encoding}
Here we introduce a learnable Gabor encoding for coordinate based MLPs. We attempt to minimize the joint uncertainty $\Delta S\Delta F$ in both frequency and spatial domain by localizing the sinusoidal components with a Gaussian. In order to provide the best trade-off, we allow the network to learn the corresponding regions and frequencies needed for the particular signal. We take the traditional manually crafted encoding and instead divide it into two layers at the beginning of the network, given as
\begin{align}
& \Phi(v) = \sin(\omega_s v + \phi_s)\exp{(-\frac{1}{2} \gamma_s^2 ||v - \mu_s ||^2)} \\
& \Psi(v) = \cos(\omega_c v + \phi_c)\exp{(-\frac{1}{2} \gamma_c^2 ||v - \mu_c ||^2)}
\end{align}
where \(\mu \in \mathbb{R}^{n_{\mathrm{in}} \times d}\) denotes the center of the regions, \(\gamma \in \mathbb{R}^d\) denotes the standard deviations, \(\omega \in \mathbb{R}^{n_{\mathrm{in}} \times d}\) are the frequencies and \(\phi \in \mathbb{R}^d\) are the phase offsets. We then concatenate these layers and pass them through ReLU layers. We found that the best initialization involved setting the initial frequencies to \(\omega \sim \mathcal{U}(-B\pi, B\pi)\) where \(B\) scales the frequency, and \(\phi \sim \mathcal{U}(-\pi, \pi) \) offsetting by half a period. The following ReLU layers can be initialized with a standard xavier initialization [] drawn from a uniform distribution.
\subsection{Spatial Localization}
The main benefit of spatial localization is the reduction of parameter sharing in the first layer. Since the sinusoidal component will be scaled down to zero if it is not localized, its corresponding parameter \(w_i^1\) will have minimal effect on the activation of neuron \(z_i^1(\gamma(x))\). This means we are softly creating sets of parameters in the first layer that only contribute to the activation of each neuron depending on the spatial location of the coordinate itself. Thus, the activation regions in the first layer are also induced by spatial information, which is then fed into downstream layers where spatial information is implicitly provided along with frequency information. The reduction of parameter sharing in the first layer should additionally help stabilize training.
For sinusoidal encodings, a Fourier representation will need to integrate over the whole domain in order to decompose the signal, which provides intuition as to its limited ability in this setting. SIREN networks are purely sinusoidal, which contributes to their inherent artifacts in rendering tasks. MFN networks attempt to provide spatial information across layers by element-wise multiplying learnable Gabor filters by linear layers. Not only is this method less efficient computationally, but it does utilize the complexity provided by piece-wise non-linearity.
\section{Results}
We compare Fourier Features, Gaussian FF~\cite{tancik2020fourier}, Gabor MFN, SIREN, and our Gabor encoding for the image regression task, NeRF, and 3D shape regression. The MFN, SIREN, and Fourier Features hyperparameters and network architecture are the same as those proposed by the authors for each task. Experiments were conducted on an NVIDIA A4000 GPU.
\subsection{Image Regression}
For this task, we aim to predict the RGB value of the corresponding pixel value that is passed through the network. For the training set, we used a grid containing 1/4 of the pixels, and evaluated the peak signal to noise ratio (PSNR) on the remaining pixels. Training consists of randomly sampled images from the div2k [] dataset, and a text image dataset. These two benchmarks were provided by the authors in []. The coordinates were assembled into random mini-batches of size 8192. From Table 1, we can see that the Gabor encoding was able to outperform each method we tested against on the image generalization task, and by a larger margin on text images.
\subsection{Novel View Synthesis}
We compared each method on the simple NeRF task provided by the authors of~\cite{tancik2020fourier}. From Table 1, we can see the Gabor encoding can improve results on this task compared to other methods, which demonstrates its usefulness in real applications.
\subsection{3D Shape Occupancy}
We can see how the Gabor encoding is able to outperform the traditional Fourier feature based method on this task, which further demonstrates its benefits even in the 3D setting. One limitation of our method is that the implicit spatial information may drown out as the number of layers increase, since 8 layers were utilized here. This is being further investigated.
\begin{figure}
\begin{center}
\begin{tabular}{ c || c }
& 3D shape $\uparrow$ \\ [.5ex]
& Uniform \;\; Boundary \\ [.5ex]
\hline
Fourier & 0.987 \;\;\;\;\;\; 0.960 \\ [.5ex]
Fourier-\(\mathcal{N}\) & \(\boldsymbol{0.988}\) \;\;\;\;\; \(\boldsymbol{0.973}\) \\ [.5ex]
Gabor & 0.987 \;\;\;\;\;\; 0.968 \\ [.5ex]
SIREN & \\ [.5ex]
\end{tabular}
\caption{Generalization performance measured in IoU for 3D shape occupancy.}
\end{center}
\end{figure}
\end{comment}
\section{Conclusion and Future Work}
In this paper, we provided a thorough study of spectral bias in the coordinate based setting by using a direct model of training dynamics. We found that convergence of gradient descent in local regions of input space, which utilizes less expressive capacity for neighboring inputs, is slower in comparison to inputs sampled globally across the signal. This results in the high frequency details of the signal being learned after a low frequency representation is achieved, with the severity of spectral bias depending on how the network assigns inputs to activation regions. Positional encoding is able to generate a lower frequency function and map previously dense inputs to distinct regions, resulting in the reduction of a large spectral bias.
\begin{figure}
\centering
\subfloat[Coordinates]{\includegraphics[width=0.25\textwidth]{xy_norm_layers.png}}
\subfloat[Positional Encoding \\ L=8]{\includegraphics[width=0.25\textwidth]{sin_cos_layer_norms.png}} \\
\subfloat[Spectral Norm Overall]{\includegraphics[width=0.25\textwidth]{overall_norms.png}}
\subfloat[Dead ReLU Neurons]{\includegraphics[width=0.25\textwidth]{dying_relu.png}}
\caption{Spectral norm for both positional encoding (L=8) (b) and coordinates ([0,1]) (a) throughout training (a-c), along with the amount of dead ReLU neurons during training for coordinates (d). There are no dead ReLU neurons for positional encoding.}
\label{fig:dead_relu}
\end{figure}
Our analysis not only provides a new interpretation of spectral bias, but also exposes its severity in dense settings. We hope that the relationship between expressive capacity and gradient descent convergence can be extended to explore spectral bias in high dimensional, non-uniform settings with different objectives such as classification. We also hope that this approach can help provide more insight as to how spectral bias impacts generalization in ReLU networks, in which analyzing the properties of the activation regions may provide first steps.
\bibliographystyle{named}
|
\section{Acknowledgment}
\section*{Acknowledgment}
The authors would like to acknowledge Deutsches Elektronen-Synchrotron DESY and Helmholtz Gesellschaft Funding for all the support.
The measurements leading to these results have been performed at the test beam facility at DESY Hamburg (Germany), a member of the Helmholtz Association.
The authors would like to thank the technical team at the {\mbox{DESY II}}\xspace accelerator and the {\DESYII Test Beam Facility}\xspace for the smooth operation of the test beam
and the support during the test beam campaign.
The readout electronics of this work was supported in part by the National Natural Science Foundation of China
under Grant 11922510 and Grant 11773027, and in part by the National Key Technologies Research and Development Program under Grant 2016YFE0100900.
\section{The {\DESYII Test Beam Facility}\xspace}
DESY operates a test beam facility at the {\mbox{DESY II}}\xspace synchrotron with three beam lines for detector testing purposes as outlined above.
The {\mbox{DESY II}}\xspace synchrotron operates in a sinusoidal ramping mode with a frequency of \SI{12.5}{\hertz}. A single electron bunch with a length of \SI{30}{ps} hits a thin carbon fibre serving as primary target in the beam and generates Bremsstrahlung photons. They are converted back into electron-positron pairs on a metal plate (secondary target). The particles pass a momentum selecting magnet and are delivered to the test beam areas. Users can chose momenta between 1~and~\SI[parse-numbers=false]{6}{GeV} and the particle polarity.
Each of the three test beam areas is equipped with a \textsc{EUDET}\xspace-style beam telescope. A more detailed description of the {\DESYII Test Beam Facility}\xspace can be found in reference \cite{desytb2018}.
\section{Introduction}
Due to the high complexity of detectors for modern particle physics experiments,
it is crucial to demonstrate and validate their performance during all steps of the development and commissioning.
Test beams with a well defined momentum and particle rate can be used to study the detector performance.
DESY Hamburg operates the {\DESYII Test Beam Facility}\xspace~\cite{desytb2018} with three independent beam lines at the {\mbox{DESY II}}\xspace synchrotron. It is one of the few facilities providing test beams in the \si{\GeV} range worldwide.
A beam telescope is a reference tracking system to reconstruct particle trajectories.
It enables measurements of detector characteristics, such as hit detection efficiency and intrinsic spatial resolution at test beam lines.
The \textsc{EUDET}\xspace-style beam telescopes~\cite{jansen2016} have served as precise reference beam telescopes at the {\DESYII Test Beam Facility}\xspace for more than ten years.
They consist of six telescope planes, grouped into two arms, which are framed by plastic scintillators to generate a trigger signal. A trigger-logic-unit~\cite{tlu} distributes this trigger signal to the telescope and user devices.
The mechanical support for the telescope is optimized to be easily adjustable for detectors of different sizes.
The \textsc{EUDET}\xspace-style beam telescopes are about to reach their end of life and a successor is needed.
The ADENIUM telescope is a key step towards the development of the next generation beam telescope.
It features a modular and modern DAQ system and is based on {\textsc{Alpide}}\xspace~\cite{AglieriRinella:2017lym} sensors, which have been identified as best suited currently available sensors.
ADENIUM can be used interchangeably with EUDET-style telescopes without impacting existing device integration.
In this report an overview of the \textsc{ADENIUM}\xspace telescope implementation is given and its performance at the {\DESYII Test Beam Facility}\xspace is discussed.
\section{Performance Studies\label{sec:Performance}}
A series of measurements was performed to verify the functionality of the \textsc{ADENIUM}\xspace telescope
and to characterize its performance in terms of timing and rate capability, noise, hit detection efficiency, cluster sizes and pointing resolution. The telescope arrangement shown in Figure~\ref{fig:setup} was used. Note that all sensors are currently configured with identical
settings, hence the results can be further optimized with sensor-specific settings.
\subsection{Timing and Trigger Rate}
\label{subsec:trigger}
While the TLU is capable to trigger up to a rate of \SI{10}{\mega\hertz} and provides a precise time stamp with \SI{781}{\pico\second} binning, the overall
trigger rate of the telescope is limited by its sensors. Their analog circuit has a peaking time of about \SI{2}{\micro\second}~\cite{AGLIERIRINELLA2017583}. A readout time of ca. \SI{10}{\micro\second}
is chosen accordingly to ensure that the hit that corresponds to the trigger is read out. Therefore, the maximum processable trigger rate is about \SI{100}{\kilo\hertz} - far more than can be achieved with the {\mbox{DESY II}}\xspace test beams. At the {\DESYII Test Beam Facility}\xspace a maximum trigger rate of \SI{40}{\kilo\hertz} was measured and no rate dependent limitations of \textsc{ADENIUM}\xspace were observed. Note that multiple hits can be recorded per trigger and the telescope is not able to resolve them in time, since no on-chip time stamping is provided. Such ambiguities can be resolved by incorporating one additional tracking plane consisting of a detector with high spatial and time resolution.
\subsection{Noise and Hit Detection Efficiency}
\label{subsec:noise}
Due to variations in CMOS processes, the sensors have slightly different noise levels at
identical configurations. A pixel scoring approach has been used to classify the pixel quality.
\begin{table}[htbp]
\caption{The fraction of pixels per plane belonging to the three noise-occupancy categories}
\begin{center}
\begin{tabular}{c c | c c c c c c c}
Grade & Noise Occupancy & Plane \#0 & Plane \#1 & Plane \#2 & Plane \#3 & Plane \#4 & Plane \#5 \\
\toprule
A & $<10^{-6}$ &99.975\% &99.998\% &99.991\% &99.973\% &99.967\% &99.988\% \\
B & $10^{-6}<N<10^{-3}$ & 0.022\% & 0.002\% & 0.008\% & 0.024\% & 0.029\% & 0.011\% \\
C & $>10^{-3}$ & 0.003\% & 0.000\% & 0.001\% & 0.003\% & 0.004\% & 0.001\% \\
\bottomrule
\end{tabular}
\label{tab:pixel_grade}
\end{center}
\end{table}
With the beam turned off, the TLU issues a million triggers to all telescope planes to sample all pixels.
Ignoring the negligible impact of the natural background, the noise occupancy of all individual pixels can be
measured by counting the noise hits per pixel. Pixels are then categorized depending on their noise occupancy
(see Table ~\ref{tab:pixel_grade}). The grade B and grade C pixels can be added to a configuration file to disable them.
\begin{table}[htbp]
\caption{The hit efficiencies for each telescope plane at two beam energies}
\begin{center}
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{c| c c c c c c c}
Energy & Plane \#0 & Plane \#1 & Plane \#2 & Plane \#3 & Plane \#4 & Plane \#5 \\
\toprule
5.6 \si{GeV} & (99.92$\pm$0.01)\% & (99.76$\pm$0.02)\% & (99.87$\pm$0.01)\% & (99.98$\pm$0.00)\% & (99.92$\pm$0.01)\% & (99.48$\pm$0.02)\% \\
2.0 \si{GeV} & (99.90$\pm$0.01)\% & (99.56$\pm$0.02)\% & (99.82$\pm$0.02)\% & (99.95$\pm$0.01)\% & (99.84$\pm$0.01)\% & (98.79$\pm$0.04)\% \\
\bottomrule
\end{tabular}
}
\label{tab:efficiency_per_layer}
\end{center}
\end{table}
The hit detection efficiency is measured using the ratio of reconstructed tracks with no hit on the plane under test to all reconstructed tracks.
"No hit" is defined here as no cluster being found with the distance less than \SI{0.5}{\mm} to the expected intersection of the track, taking the effect of multiple scattering into account when matching the hit to the tracks. The hit efficiency of the different telescope planes is shown in Table~\ref{tab:efficiency_per_layer}.
\subsection{Hit Cluster}
\label{subsec:cluster}
Due to charge sharing, several neighbouring pixels may respond to a single particle hit at the same time. The number of pixels belonging to a hit cluster,
i.e.~the cluster size, depends on the incident angle of the particle as well as its hit position within the pixel.
\begin{figure}[htbp]
\begin{subfigure}[t]{0.48\textwidth}
\includegraphics[width=0.99\linewidth]{clusterSize_trajHit_DetN1_geo7_gev5p6_e_t.pdf}
\caption{}
\label{fig:clusterSize_trajHit_DetN1_layer}
\end{subfigure}
\begin{subfigure}[t]{0.48\textwidth}
\includegraphics[width=0.99\linewidth]{clusterSize_trajHit_fraction_geo7_gev5p6_e_t.pdf}
\caption{}
\label{fig:clusterSize_trajHit_all_layer}
\end{subfigure}
\caption{(a) Fraction of cluster sizes for one telescope plane (layer ID = 1). (b) Fraction of clusters with sizes 1 to 6 for the six telescope planes. The mean cluster sizes for the individual planes are 2.71, 2.56, 2.61, 2.99, 2.81, and 2.86.}
\end{figure}
For single-pixel clusters, the intrinsic resolution of the sensor can be estimated as \SI{7.76}{\micro\metre} and \SI{8.44}{\micro\meter} in the
$x$ and $y$ dimension, respectively, using the formula: \(d/\sqrt{12}\), where $d$ is the pixel pitch. Charge sharing can improve this resolution when the cluster center is used as the measured hit position.
Figure~\ref{fig:clusterSize_trajHit_all_layer} shows the percentage of clusters with different sizes for each telescope plane.
The mean cluster size is 2.78 pixels per cluster and varies within 0.22 among different sensors.
This can be further optimized and adjusted to a uniform value for all sensors, e.g. via the comparator thresholds.
The average cluster size of all sensors is independent of beam energies ranging from \SI{2.0}{\GeV} to \SI{5.6}{\GeV}, hence the intrinsic resolution at the {\mbox{DESY II}}\xspace beam lines is constant, as expected.
\subsection{Track Interpolation}
\label{subsec:trackext}
With a properly aligned detector geometry, the built-in analysis module of the DAQ software (see Section~\ref{sec:Teleintr:track})
in the stand-alone mode is able to reconstruct the particle trajectories
and their intersection points with DUT planes online.
At each propagation step of the CKF algorithm, the predicted $\chi^2$ is expected to follow a distribution with two degrees of freedom for the each telescope plane. A cut of 13.816 is imposed on the predicted $\chi^2$ when the clusters are associated to the track in the CKF algorithm. This corresponds to accepting hits steeming from the correct particle with a probability of 99.9\%.
The interpolated position of a reconstructed track on the DUT plane can be compared to the measured DUT hit.
Figure~\ref{fig:res_geo7_gev5p6_t3} shows these distributions of unbiased residuals, i.e.~excluding the DUT hits from the track fit, for a sample of 500k electrons at \SI{5.6}{GeV} with the telescope setup as
shown in Figure~\ref{fig:dual_tarjs_display}. Two Gaussian fits to
the unbiased residuals show a width of $\sigma=5.87\pm0.01$~\si{\micro\meter} and $\sigma=5.74\pm0.01$~\si{\micro\meter} in the $x$ and $y$ dimension of the DUT plane, respectively.
\begin{figure}[htbp]
\begin{subfigure}[t]{0.48\textwidth}
\includegraphics[width=0.99\linewidth]{res_geo7_gev5p6_200628235239_t3_X.pdf}
\caption{$x$ direction (pixel pitch is \SI{29.24}{\micro\metre})}
\label{fig:res_geo7_gev5p6_x}
\end{subfigure}
\begin{subfigure}[t]{0.48\textwidth}
\includegraphics[width=0.99\linewidth]{res_geo7_gev5p6_200628235239_t3_Y.pdf}
\caption{$y$ direction (pixel pitch is \SI{26.88}{\micro\metre})}
\label{fig:res_geo7_gev5p6_y}
\end{subfigure}
\caption{Distributions of residual in the (a)$x$ and (b)$y$ dimension of the DUT plane for a sample of 500k electrons at \SI{5.6}{GeV}.
The six telescope planes with additional telescope plane taken as the DUT are grouped as shown in Figure~\ref{fig:setup},
where the distances ($D_{ij}$) between the planes are $D_{10}=D_{21}=D_{43}=D_{54}=D_{t2}=D_{t3}=$ \SI{38}{\milli\meter}. The black dots are data, and the red (blue) lines are Gaussian fits to the data}
\label{fig:res_geo7_gev5p6_t3}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.6\textwidth]{resolScanEnergy_geo313_geo7_xy.pdf}
\end{center}
\caption{Resolution of the telescope at an additional telescope plane acting as DUT with distances
between the planes of $D_{10}=D_{21}=D_{43}=D_{54}=$ \SI{38}{\milli\meter} and $D_{t2}=D_{t3}=D_{32}/2$. The different markers represent different distances to the innermost telescope planes ($D_{t2},D_{t3}$). The slight difference in resolution along x/y is due to the difference in pixel pitch.}
\label{fig:resScan}
\end{figure}
The resolution of the telescope for different spacing as a function of the momentum is shown in Figure~\ref{fig:resScan}.
As expected, the resolution improves with increasing momentum since the impact of multiple scattering is reduced.
The dependency of the tracking precision on the distances between telescope planes can be seen as well, i.e.~smaller distances between the planes result in a higher precision on the
DUT.
This can be explained by less air that has to be passed as well as smaller uncertainties on the propagation. In the optimal setup, a resolution of $\sigma_x = \SI{2.89}{\um}$ and $\sigma_y =\SI{2.84}{\um}$ at \SI{5.6}{\GeV} could be determined.
\section{Summary and Outlook}
A demonstrator of a high resolution telescope, as upgrade and drop-in replacement
of the \textsc{EUDET}\xspace-style telescopes, was developed and tested. It reaches pointing resolution below \SI{3}{\micro\meter} and shows no performance limitations at the maximum trigger rates that can be reached at the DESY II test beams (ca. \SI{40}{\kilo\hertz}).
It is compatible with the existing \textsc{EUDET}\xspace-style infrastructure, so
users can switch to the new telescope easily without additional integration effort.
The compact readout electronics allow a flexible arrangement of the telescope
planes.
The demonstrated performance justifies the choice of the sensor and the DAQ layout.
An upgrade of all \textsc{EUDET}\xspace-style pixel beam telescopes in the near future will be based on the presented system.
The next development for a long-term version of the \textsc{ADENIUM}\xspace telescope is on-going.
A more compact readout electronic based on a Xilinx Zynq SOC has been designed and manufactured.
It will allow the implementation of more advanced features including running the DAQ software on the front-end Zynq SOC itself.
\section{\textsc{ADENIUM}\xspace Telescope Overview}
The ADENIUM beam telescope consists of six ALPIDE planes
that are selected for their good spatial resolution and low material to allow precise particle tracking at
low momenta particle beams such as the {\mbox{DESY II}}\xspace electron beam.
ADENIUM has to be capable of processing particle rates of a few \SI{10}{kHz} to match the {\mbox{DESY II}}\xspace test beam rates.
\subsection{Sensor}
Monolithic Active Pixel Sensors (MAPS) are ideal for high resolution tracking telescopes at low momentum beam lines.
Mimosa26 sensors~\cite{baudot2009,huguo2010} had been chosen for the EUDET-style telescopes, since they provide an
excellent intrinsic resolution that stems from the fine pixel pitch of \SI{18.4}{\micro m}, combined with diffusion based charge collection in an up to \SI{20}{\micro\meter} thick epitaxial layer.
Removal of the inactive silicon minimizes the amount of material the particle has to pass.
As the long term availability of the Mimosa26 is not clear, the highly available ALPIDE sensor~\cite{suljic:2658226} is chosen for the ADENIUM telescope.
The ALPIDE sensor is implemented in an \SI{180}{nm} CIS process and fabricated on wafers with a \SI{25}{\um} thick high resistivity p-type epitaxial layer on a p-type substrate. Table~\ref{tab:sensor:parameters}) compares the main
parameters of the two sensors. ALPIDE has a slightly larger pitch, but offers a larger active area as well as significantly shorter readout times.
\begin{table}[htb]
\caption{Main parameters of the Mimosa26~\cite{baudot2009,huguo2010} and {\textsc{Alpide}}\xspace sensor~\cite{suljic:2658226}.}
\begin{center}
\begin{tabular}{lcc}
\toprule
& {\bf Mimosa} 26 & {\bf {\textsc{Alpide}}\xspace} \\ \hline
Chip size & \SI{21.2 x 10.6}{\mm} &\SI{15 x 30 }{\mm} \\
Chip thickness& \SIrange{50}{70}{\um} & \SIrange{50}{100}{\um} \\
Pixel pitch & \SI{18.4 x 18.4 }{\um} &\SI{26.88 x 29.24 }{\um} \\
Pixel matrix & \num{1152x 576} &\num{512 x 1024}\\
Detection efficiency & \SI{> 99}{\percent} & \SI{> 99}{\percent} \\
Fake-hit rate &\SI{\sim e-6}{pixel^{-1}event^{-1}} & \SI{< e-6}{pixel^{-1}event^{-1}} \\
Typical frame readout time &\SI{115.2}{\us} & \SI{10}{\us}\\
\bottomrule
\end{tabular}
\label{tab:sensor:parameters}
\end{center}
\end{table}
Additionally, the power-efficient design of the chip results in a negligible heat generation, allowing for passive cooling, which reduces the complexity of the mechanical design significantly compared to the existing MIMOSA26 planes.
\subsection{Telescope Plane}
Each telescope plane can be operated standalone as an independent network node, maximizing the system flexibility.
An \textsc{ADENIUM}\xspace reference plane, compare Figure~\ref{fig:layer_lab}, consists of three hardware components: the sensor carrier board, the main readout board and a small passive bridge board:
\begin{itemize}
\item{The {\bf main readout board} is custom-made at USTC. It featues
a Xilinx Kintex-7 field-programmable gate array (FPGA) chip as the core component~\cite{yang8698804} running a custom firmware to operate the chip and run a server on an ethernet network node. In addition, an optical SFP interface is provided.
The connection to the sensor carrier board is realized via a dedicated Mezzanine Card (FMC) connector.
Power and clock signals for the chip are provided by the main readout board to the sensor over the bridge board.
}
\item{A {\bf bridge board} is designed, with FMC and PCIe connectors at each end, to bridge between the main readout board and sensor carrier board.
The bridge board provides an optional interface with the \textsc{AIDA-2020}\xspace Trigger Logic Unit (TLU)~\cite{Baesso_2019} via an HDMI connector.}
\item{The {\bf sensor carrier board} is provided by the {ALICE}\xspace collaboration~\cite{AglieriRinella:2017lym} , with the sensor chip glued and wire bonded to it.
Below the sensor the PCB has an opening to minimise the material to reduce multiple scattering when particles pass through the sensor.
Besides the mechanical support for the sensor, the board provides passive electronics for
powering, noise reduction and signal coupling.}
\end{itemize}
\begin{figure}[ht]
\centering
\includegraphics[angle=0,origin=c,width=0.75\linewidth]{adenium_boards_assembly.png}
\caption{An assembled readout electronic of a telescope plane showing sensor carrier board, bridge board and main readout board.}
\label{fig:layer_lab}
\end{figure}
\subsection{DAQ Software and Data Processing}
The telescope DAQ is realized as a client to all connected telescope planes communicating over a TCP/IP-based interface. The data flow is summarized in Figure~\ref{fig:adenium_dataflow}: the telescope DAQ receives data from all planes, synchronizes it by trigger ID and performs hit clustering.
A \textsc{EUDAQ2}\xspace~\cite{Liu_2019} component, called \texttt{Producer}\xspace, provides the interface between the telescope and EUDAQ2.
\textsc{EUDAQ2}\xspace globally controls all components and logs the data flow including the telescope.
Detector synchronization on hardware level is realised with the \textsc{AIDA-2020}\xspace trigger-logic-unit \cite{Baesso_2019}.
The central \texttt{Run Control}\xspace of EUDAQ2, with an optional graphical user interface, is able to control the telescope and start/stop its readout via standardized commands processed by the \texttt{Producer}\xspace.
In this concept the reference planes of the telescope are treated the same way as any device under test (DUT)~\cite{Ahlburg:2019jyj}.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=1.0\textwidth]{adenium_dataflow_0728c.pdf}
\caption{The data flow in \textsc{ADENIUM}\xspace from the individual sensor planes to the the user analysis.}
\label{fig:adenium_dataflow}
\end{center}
\end{figure}
According to \textsc{EUDAQ2}\xspace specification, a so-called \texttt{DataConverter}\xspace is implemented. A \texttt{DataConverter}\xspace plugin in EUDAQ@ converts the raw \textsc{ADENIUM}\xspace telescope data to a \textsc{EUDAQ2}\xspace native format or the \textsc{LCIO}\xspace format~\cite{Gaede:2003ip}.
Either the built-in tracking discussed below or third party software, like the \textsc{Corryvreckan}\xspace package~\cite{Dannheim:2020jlk}
or the \textsc{EUTelescope}\xspace package~\cite{Bisanz_2020} can be used for the particle trajectory reconstruction.
\subsubsection{Trigger Implementation}
An \textsc{AIDA-2020}\xspace Trigger Logic Unit (TLU)~\cite{Baesso_2019} serves as a global trigger distribution system for all connected devices.
A trigger itself is usually generated by a set of scintillator-PMT modules when a particle passes them.
\textsc{ADENIUM}\xspace uses the {\textit{AIDA-mode-with-id}}. In this mode the TLU sends a trigger, a \SI{40}{\mega\hertz}
clock, and a trigger ID via an HDMI cable. A busy from the telescope vetos potential additional triggers during readout.
\subsection{Telescope Mechanics}
The telescope mechanics reuses the support structure of the EUDET-type telescopes. Rails for mounting the sensor planes are arranged in two arms parallel to the beam axis (see Figure~\ref{fig:setup}). This design allows for maximum flexibility when integrating user setups. Placing DUTs in the center of the telescope yields the best pointing resolution as the particle trajectories are confined on both sides.
\begin{figure}[ht]
\centering
\includegraphics[width=0.7\linewidth]{dual_arm_grouping.pdf}
\caption{Typical arrangement of the \textsc{ADENIUM}\xspace Telescope planes.
The beam direction (red line with arrow) is from right to left and the DUT is indicated as a grey box. The telescope planes are
numbered with consecutive numbers starting from $0$ from right to left. The distances between the planes (green lines with arrows) are denoted by $D_{ij}$. }
\label{fig:setup}
\end{figure}
The distance between the two arms is adjustable and can measure up to \SI{50}{\centi\metre} allowing for larger scale DUT setups. The two outermost telescope layers can span a distance of \SI{1.3}{\metre} maximum. A high precision $xy\phi$-stage table can move the DUT and adjust its position with respect to the active area of \textsc{ADENIUM}\xspace. A photograph of the telescope installed at the {\DESYII Test Beam Facility}\xspace is shown in Figure~\ref{fig:tele_lab_beam}.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.5\textwidth]{telecope_photo_3_3.jpg}
\caption{The \textsc{ADENIUM}\xspace telescope installed at the {\DESYII Test Beam Facility}\xspace}
\label{fig:tele_lab_beam}
\end{center}
\end{figure}
The sensor carrier board is fixed on an aluminium jig with a cutout that minimises the amount of material in the beam.
The openings are covered by 25 \si{um} thin polyimide sheets on both sides to protect the sensor from dust.
This jig is fastened on the EUDET-style rail system.
The main board is mounted at the top of the jig to an aluminium aluminium frame.
Each plane can be moved independently on the rails allowing for a flexible layer arrangement.
No active cooling is implemented in this mechanical support as the overall heat dissipation of the sensor as well as the adjacent electronics is relatively low. The power consumption of ALPIDE is below \SI{40}{mW/cm^2}~\cite{MAGER2016434}. Passive cooling through the jig is sufficient to maintain reasonable temperatures during operation.
\subsection{Integration and synchronization of user detectors}
Correlating the reference trajectories from the telescope with data from DUTs is an essential part of test beam measurements. Hence, it has to be ensured that the telescope and DUT data streams are synchronized. The \textsc{AIDA-2020}\xspace TLU takes care of this on hardware level by sending a global trigger signal, typically generated by a scintillator coincidence, to all connected detectors. Up to four detectors can be connected to a single \textsc{AIDA-2020}\xspace TLU. One channel is occupied by ADENIUM, leaving three free slots for user devices. The physical interface must comply with the specification of the AIDA-2020 TLU~\cite{aida2020}. Alongside the trigger signal, the \textsc{AIDA-2020}\xspace TLU sends a trigger ID, which gets included in the telescope data stream. By including it also in the DUT's data stream the synchronisation gets more robust. The data from both telescope and DUT can be stored in the same file, optionally already sorted by trigger ID.
In addition to the scintillator-based trigger detectors installed by default, user devices can also trigger the AIDA-2020 TLU. It features six trigger inputs with configurable coincidence logic generating the global trigger signal. A precise timestamp of the trigger is recorded by the TLU and can be added to the data stream.
An optional full integration of the user device in \textsc{EUDAQ2}\xspace allows for synchronized configuring,
starting and stopping of the telescope's and DUT's DAQs.
Data can be be saved in the \textsc{EUDAQ2}\xspace native data format, where an event package is the fundamental storage structure, which contains the trigger ID, timestamp, telescope data, and DUT hit data belonging to a single trigger.
The size of an event package per trigger varies and depends on the telescope's occupancy and the DUT data defined by the user. Typically, at the {\DESYII Test Beam Facility}\xspace less than three clusters are recorded per telescope plane and trigger. In this case, without a DUT, the size of the event package is about 1~kByte.
\subsection{Built-in Track Reconstruction Module\label{sec:Teleintr:track}}
The ADENIUM DAQ software features built-in track reconstruction based on
a combinatorial Kalman Filter (CKF) tracking algorithm developed as a part of the ACTS project~\cite{ai2021common}. The latter provides a toolkit for track reconstruction in a generic,
framework- and experiment-independent software package. The fast CKF algorithm of the built-in tracking module can run online and offline (see Section~\ref{sec:Performance}). In online mode each trigger event is processed instantaneously. The offline mode can be used with a refined software alignment of the telescope planes.
Since multiple beam particles could potentially pass through the telescope in the same trigger window, the CKF algorithm is configured to take all clusters in the first telescope plane as seed
of potential track candidates. For each seed, the CKF algorithm performs the forward track propagation into the next telescope plane. Clusters in the following telescope plane are scored using the predicted $\chi^2$ obtained by comparing the Kalman filter prediction and the measured cluster center with the resolution of the cluster taken into account~\cite{Fruhwirth:1987fm}. The cluster with the highest score within a search window is added to the track.
The track parameters are then updated for next forward track propagation. The CKF algorithm is tolerant to imperfect trajectory with missing clusters, due to either sensor inefficiency or that the trajectory is partially outside of the sensitive region (misalignment). A backward track parameter smoothing performed afterwards to further improve the precision.
The reconstructed particle trajectories can be stored in the same file alongside the original telescope pixel hit and cluster data.
The telescope DAQ software has built-in 3D graphic window, implemented using OpenGL, to visualise the reconstructed trajectories in real-time.
The graphical display shows only the last reconstructed event.
Figure ~\ref{fig:dual_tarjs_display} shows two reconstructed trajectories which belongs to a single trigger
with a beam energy of \SI{2.0}{GeV} and high beam rate.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.9\textwidth]{dual_trajs_display.png}
\caption{A 3D visualization of reconstructed double trajectories belonging to a single trigger (beam energy at \SI{2.0}{GeV}) by the DAQ software.
The blue rectangles are the sensitive areas of the telescope. The red segmented lines crossing the sensitive areas denote the measured hit clusters by the telescope planes.
The telescope planes are placed as in Figure~\ref{fig:setup} where the distances between the planes are
$D_{10}=D_{21}=D_{43}=D_{54}=38~mm$ and $D_{t2}=D_{t3}=D_{32}/2=150~mm$.}
\label{fig:dual_tarjs_display}
\end{center}
\end{figure}
To not deteriorate the precision of the reconstructed tracks, the geometry of the telescope, i.e.~the position and orientation of the telescope planes, must be precisely known.
The built-in tracking module provides tools to perform a track based alignment using the \textsc{Millepide-II}\xspace~\cite{millepede2:2006} package, which can significantly improve the precision
of the telescope geometry after several iterative runs. For the time being, the alignment procedure is standalone and supposed to be executed prior to the track reconstruction.
The generated geometry file with misalignment corrected is then fed into the online track reconstruction.
\section{Telescope Overview}
All EUDET-style telescopes and their derivatives consist of two arms with each three reference planes and
space in between to place the Devices-under-Test (DUTs). The sensors for the reference planes in the two arms
are selected for their good spatial resolution and low material to allow precise tracking also for
low momenta particle beams such as the {\mbox{DESY II}}\xspace electron beam is providing.
In addition, good time resolution and high readout speed are required.
In the following the development of a new derivative of the EUDET-style telescope based on a different main readout sensor will be described.
\subsection{Reference Plane Upgrade}
Over last decades, Monolithic Active Pixel Sensors (MAPS) technology has made impressive progress, and gradually becomes more available and affordable via many advanced commercial foundries.
MAPS were chosen as technology for the \textsc{EUDET}\xspace-style beam telescope reference planes as those provide good spatial resolution while being low in the material thickness.
The excellent intrinsic resolution stems from the high granularity achievable, combined with the cluster charge sharing due to diffusion. The the thin sensitive volume enables thinning the substrate down to a few tens of micrometres.
The chosen Mimosa26 sensor~\cite{baudot2009,huguo2010} for the \textsc{EUDET}\xspace-style telescopes feature a pixel size of \SI{18.4}{\micro\meter} and a thickness of \SI{50}{\micro\meter}. Minimum ionising particles (MIPs) are detected by collecting the few thousand electrons generated by the particle within the thin, almost un-depleted, epitaxial layer implemented on top of the wafer substrate.
\begin{table}[htb]
\caption{Main parameters of the Alpide sensor~\cite{suljic:2658226}}
\begin{center}
\begin{tabular}{lp{6cm}}
Parameter & \\
\toprule
Chip size & \SI{15 x 30 }{\mm} \\
Chip thickness & \SIrange{50}{100}{\um} \\
Pixel pitch & \SI{26.88 x 29.24 }{\um} \\
Pixel matrix & \num{512 x 1024} pixel cell \\
Detection efficiency & \SI{> 99}{\percent} \\
Fake-hit rate & \SI{< e-6}{pixel^{-1}event^{-1}} \\
\bottomrule
\end{tabular}
\label{tab:sensor:parameters}
\end{center}
\end{table}
A different sensor as main reference plane sensor is needed as the long term availability of the Mimosa26 is not clear.
The {\textsc{Alpide}}\xspace sensor ~\cite{suljic:2658226} is an excellent candidate as it provides good spatial resolution, high detection efficiency and a low fake hit rate (see Table~\ref{tab:sensor:parameters}) and thus meeting the requirements for a pixel beam telescope to be used with low momentum particles.
The design of this MAPS was originally for the upgrade of the inner tracker of the {ALICE}\xspace experiment at the CERN Large Hadron Collider where the sensor chip is required to read out Pb-Pb interactions at \SI{100}{\kHz} with detection efficiency above \SI{99}{\percent},
fake-hit probability below \num{e-5} and a spatial resolution of \SI{5}{\um}.
The {\textsc{Alpide}}\xspace sensor is implemented with \SI{180}{\nm} CIS process and fabricated on substrates with a high resistivity \SI{25}{\um} p-type epitaxial layer on a p-type substrate.
It measures roughly \SI{15 x 30}{\mm} and contains a matrix of \num{512 x 1024} pixels with amplification,
shaping, discrimination and three-events-buffering inside each individual pixel covering \SI{26.88 x 29.24}{\um}.
The in-pixel circuit works in global shutter mode, which eliminates the rolling shutter effect.
It is a signification improvement comparing to the rolling shutter Mimosa sensor of \textsc{EUDET}\xspace-style beam telescopes.
Due to the power efficient design of its hit driven on chip circuit and very limited number of beam particle hits to telescope planes, the heat generation is negligible. The passive cooling is sufficient and extremely reduces the complexity of the mechanical design.
\subsection{Telescope Plane Readout}
The concept for the readout of telescope components is following the \textsc{EUDAQ2}\xspace concept which enables the flexible integration of various independent DUTs into a top-level framework via their specific data acquisition systems. \textsc{EUDAQ2}\xspace globally controls all components, centrally synchronises the data flow, and logs the data flow.
In this concept the main reference planes of the telescope are treated the same way as the DUTs~\cite{Ahlburg:2019jyj}.
Equipping with its standalone sensor readout, each telescope plane runs as independent network node on a distributed readout system on Ethernet network. Thus also the number of reference planes can be adjusted
easily.
In the following the readout of the chosen {\textsc{Alpide}}\xspace sensors as developed for the \textsc{ADENIUM}\xspace telescope is described.
Three components form the readout for the \textsc{ADENIUM}\xspace reference planes: the sensor carrier board, the main readout board and a small passive bridge board:
\begin{itemize}
\item{The {\bf sensor carrier board} is provided by {ALICE}\xspace ITk upgrade community.
The sensor chip is glued and wire bonded onto the carrier board. Below the sensor the PCB board has an opening to minimise the material to reduce
multiple scattering when particles pass through the sensor. Besides the mechanical support for the sensor the board provides passive electronics for
powering, noise reduction and signal coupling as well as a connector.}
\item{The {\bf main readout board} is custom-mode at USTC, carrying
a Xilinx Kintex-7 field-programmable gate array (FPGA) chip as the core component~\cite{yang8698804}.
The main board offers a dedicated Mezzanine Card (FMC) connector for detector sensor and optical fiber SFP connector for Ethernet connection.}
\item{A {\bf bridge board} is designed, with FMC and PCIe connectors at each end, to bridge between the main readout board and sensor carrier board. The bridge board optionally accepts external trigger thought HDMI interface.}
\end{itemize}
Configured by a custom firmware, the FPGA operates the sensor chip and runs the server as node on Ethernet, via IP/TCP. Power and clock signal are provided by main readout board to sensor and bridge board.
During the earlier phase of the telescope development, a simple DAQ software is implement to the validate the functionality of telescope plane readout. Written in C++, it is basic network capable software working above the POSIX sockets API. It later services as underlying software library for the telescope DAQ software which supervises all telescope planes.
\subsection{Telescope Plane Mechanical Setup}
\begin{figure}[ht]
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=0.88\linewidth]{assambed_plane_daqv1_back.jpg}
\caption{An assembled telescope plane}
\label{fig:layer_lab}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=1.0\linewidth]{alpide_tele_lab_portable.png}
\caption{Six telescope planes mounted on a rail}
\label{fig:six_planes_on_rail}
\end{subfigure}
\caption{Telescope plane assembly}
\label{fig:plane_assembly}
\end{figure}
Each telescope plane with the components described about is housed by a mechanical structure compatible with the standard EUDET-style telescope support.
The sensor carrier board is fixed on an aluminium jig with an opening to minimise the material the test beam particles pass through. The openings are covered by 25 \si{um} thin Kapton polyiamide sheets on both sides to protect the sensor from dust. This jigs is fastened in a base frame positioning
the sensors so that the beam is passing perpendicular through it.
The main board is attached with a simple aluminium frame above the jig.
After the assembly, the telescope plane becomes a solid piece and
feasible to mount in preferred arrangements by the user decision.
Due to very limited heat deposition, mainly from the FPGA, the cooling is just passive.
|
\section{Introduction}
Millisecond pulsars (MSPs) are remarkable objects born via type II supernovae (SNe)
explosions and characterized by short spin periods (P$_{spin} \leq $ 20 ms), weak magnetic fields (B $\leq 10^{9}$ G) extremely old
$ \sim 10^{9}$ yr based on recycling process. ATNF census of Galactic 445 MSPs in both the Galactic
plane and in globular clusters
\citep[see e.g.,][]{2005AJ....129.1993M, 2014Freire, 2012ChA&A..36..137C}. They are frequently found in binaries (about 65\%) with white dwarfs (WDs) in circular orbits with companions with masses ranging from $\rm \sim 0.15 M_{\odot} - 0.45M_{\odot}$ \citep[see e.g.,][]{1982Natur.300..728A, 1991PhR...203....1B, 2012AN....333...53T, 2017JPhCS.869a2090T}.
\begin{figure}[htb]
\centering
\includegraphics[width=80mm,height=90mm]{AIC-track-new-eps-converted-to.pdf}
\caption{The AIC of binary system evolution is tracked, and the mass accretion-rate is expected to be around ($\Delta M \rm \sim 0.2M_{\odot}$) with a timescale shorter than 0.4 Gyr \citep[see e.g.,][]{2002MNRAS.329..897H, 2013IJMPS..23...95J}.} \label{fig1}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=80mm,height=90mm]{Fig2.pdf}
\caption{ Simple diagram illustrating how the companion star, $M_c$, will receive some of the mass ejected by the exploding WD during the AIC. The top panel corresponds to an instant right before the explosion. The bottom panel corresponds to an instant right after it happened. The ejected material has been schematically divided into blobs 1, 2, 3, 4, 5, 6, 7. The sum of the masses of all blobs is asymmetric mass ejection. The motion of each blob has been shown with an arrow. The blobs 1, 2 and 7 will reach the surface of $M_c$. Therefore, the mass of blobs 1, 2 and 7 are received and set to be ($\Delta M \rm \sim 0.18M_{\odot}$).} \label{fig2}
\end{figure}
It was suggested that MSPs were formed in low-mass X-ray
binary systems (LMXBs). This argument called for a recycling
process in which the slowly rotating, old Neutron Star (NS) may be spun-up into an MSP via accretion from a binary companion \citep[see e.g.,][]{1982Natur.300..728A, 1991PhR...203....1B, 2006Tauris, 2019AN....340..847T, 2017PASA...34...24T}.
This link has been more recently reinforced by the discovery of many pulsars like PSR J1023+0038 \citep{2009Sci...324.1411A}, whose radio emission is likely to have switched on only recently after the LMXB phase.
An alternative formation scenario is the accretion induced collapse (AIC) of
a ONeMg white dwarf (WD) \citep[see e.g.,][]{1987ApJ...322..206N, 2010MNRAS.402.1437H, 2013Tauris, 2014Freire, 2017JPhCS.869a2090T, 2017ApJ...846..170T, 2019Schwab, 2020RAA....20..135W}. In this scenario, the mass transfer can
increase the mass of a WD toward the Chandrasekhar mass. When the star reaches this mass limit, degeneracy pressure can no longer support it because the supporting electron pressure is robbed by inverse $\beta$
decays \citep{2005ASPC..334...65K}, consequently losing enough energy in
neutrinos. This causes it to collapse and violently release a substantial
amount of gravitational energy, which might be observable by the gravitational wave observatories such as LIGO-Virgo
Collaboration and GEO \citep{2009MNRAS.396.1659M, 2010MNRAS.409..846D, Mardini_2019a, Mardini_2019b, Mardini_2019c, Mardini2020, 2022MNRAS.510.6011W}.
The NS equation of state, along with the mass-radius relation, plays a role in modeling the baryonic and gravitational masses as well as binding energy \citep{2009PhRvC..80f5809N}. The majority of efforts have been made to calculate the effect of the binding energy by estimating the amount of evolution driven by mass accretion in order to understand the evolutionary tracks of binary models \citep[see e.g.,][]{2011MNRAS.413L..47B, 2018MNRAS.477.2349I, 2019RAA....19...12T, 2019JPhCS1258a2029T, 2021arXiv210313605H, 2021ApJ...910L..22H, 2004ApJ...614..914A}.
\cite{2021arXiv210313605H} constrained the baryonic mass and binding energy associated with the equation of state, of two double pulsars (J0737-3039B and J1756-2251), finding that are thought to have formed from an ultra-stripped progenitor. Also he found that these systems are consistent with forming from the AIC.
This work aims to consider the pre-AIC and post-AIC binding energies
of the binary system. Then, taking into account the non-zero velocity of ejected matter, I derive several elementary analytic formulae for evaluating the variation of orbital energy and the possibility of binary survival.
\begin{figure}[htb]
\centering
\includegraphics[width=80mm,height=90mm]{Final-Binding-Energy.pdf}
\caption{The dynamical evolution of the post-AIC and pre-AIC systems as we can be seen the initial and final binding energies. Note that the typical mass loss taken for this is $0.18M_{\odot}$.} \label{fig3}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=80mm,height=90mm]{seperation-distance-eps-converted-to.pdf}
\caption{The final separation distance as a function of time. An indicator of the presence of a dynamical AIC process in a given system is the strong sensitivity to initial conditions i. e. masses and semi-major axes (see system details in the text).} \label{fig4}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=80mm,height=90mm]{Final-orbital-energy-eps-converted-to.pdf}
\caption{The final orbital energy as a function of the final fate of the mass companion. The correlation is roughly linear, as predicted by the H-R diagram.} \label{fig5}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=80mm,height=90mm]{Final-orbital-energy-final-binding-energy-eps-converted-to.pdf}
\caption{The final orbital energy as a function of the final binding energy through the AIC process.} \label{fig6}
\end{figure}
\section{The Binding Energy}
The impact of kick velocity on the companion depends on the binding energy
of the system and the kinetic energy of the kick exerted on the orbit.
It is noteworthy to mention here that, after AIC, the mass transfer will start once more due to the evolution of the companion star. Fig. ~\ref{fig1}, shows the evolutionary track of the AIC process as it decreases mass. However, given a particular initial mass of the WD $1.1M_{\odot}$ (i.e. Taani et al. 2012a), it will take a time interval shorter than 0.4 Gyr \citep[see e.g.,][]{2002MNRAS.329..897H} to trigger the AIC. I assume the mass transfer model used in this work, is based on the Roche model and ignores the secondary's spin angular momentum. The angular momentum loss due to the gravitational radiation is calculated.
This is valid for the WD primary \citep {2001MNRAS.327..888R}. This will further change the orbital period and the eccentricity \citep{2010ChPhL..27k9801W}. I will deal in this section with the binding energy of
a dynamical system based on the change in orbital energy. This could occur during the pre-SNe orbit and immediately before the explosion, as described by the equation
\textbf{\begin{equation}
E_{orb,i} = \frac{- GM_{WD} M_{c,i}}{2a_{i}}
\end{equation}}
\textbf{\begin{equation}
= - \frac{GM_{WD} M_{c,i}}{a_{i}} + \frac{M_{WD}M_{c,i}}{M_{NS}+M_{c,i}} v^{2}_{i}
\end{equation}}
where the $v_{i}$ is WD's initial velocity relative to its companion
(mass= Chandrasekhar mass $M_{ch}$). M$_{NS}$ is the NS's mass. G is Newton’s
gravitational constant. At the evolution of a binary under instantaneous mass loss, the initial mass of the companion $M_c,i$ is assumed to be at least $4M_{\odot}\leq M_{com} \leq6M_{\odot}$, which may be reasonable for AIC binaries where the companion is on the main sequence of helium. $a_{i}$ is the system's initial semi-major axis, which is approximately (1.5 R$_{\odot}$). For more information, \citep[see e.g.,][]{1983AJ.....88.1857H} and references therein.
The orbital energy after the SNe
\textbf{\begin{equation}
E_{orb,f} = \frac{- GM_{NS} M_{c,f}}{2a_{f}}
\end{equation}}
\textbf{\begin{equation}
= - \frac{GM_{NS} M_{c,f}}{a_{f}} + \frac{M_{NS}M_{c,f}}{M_{NS}+M_{c,f}} v^{2}_{f}
\end{equation}}
where the companion $M_{c,f}$'s final mass is estimated to be at least 1.4M$_{\odot}$ [see, for example \citep[see e.g.,][]{2006Tauris}. The final velocity of WD relative to its companion is given by $v_{f}$. a$_{f}$ is the system's final semi-major axis after the AIC process, which is approximately (3R$_{\odot}$) (see i.e, \cite{10.1093/mnras/staa3701, 2020arXiv200203011T}. The geometry of an asymmetric SNe in the binary system is illustrated in Fig. ~\ref{fig2}. This will depend on the mass loss and angular momentum variations. The low binding energy of the ejecta and the low explosive energy are, of course, related. The slow rotation is caused by the accretion of gas with small angular momentum during this process.
It is worth noting that different amounts of deposition energy into the common envelope should make a noticeable difference in the two different phases (pre- and post AIC). This is due to the dissipating mechanism acts on the dynamical orbital evolution. In addition, I also consider the angular momentum that could affect the separation distance for the system based on the amount of accreted mass (see i.e, \cite{10.1093/mnras/staa3701}.
\begin{equation}
\textit{J}= \frac{M_{NS} M_{c,f}}{M_{NS}+ M_{c,f}} a_{f} \frac{2\pi}{P}
\end{equation}
However, this effect could be observed from the gravitational radiation emitted by the binary motion in a circular orbit \citep{2004ApJ...614..914A}. In addition, the role of the compactness C = GM/(c$^{2}$R), where c is the speed of light and G is the gravitational constant, during accretion has been adopted from \cite{2021arXiv210313605H}, and points out the need for of constraining the equation of state. As a result, more accreted baryonic mass will be transformed into binding energy instead of gravitational mass \citep{2021ApJ...910L..22H}.
Here, I consider the pre-AIC and post-AIC binding energies of the binary system by assuming the difference in final and initial orbital energies ($E_{orb,f}$ and $E_{orb,i}$) respectively (see Fig. ~\ref{fig3}). This could cause the system's orbit to continually circularize and contract due to losing orbital energy.
\begin{equation}
\Delta E_{orb}= E_{orb,f} - E_{orb,i}
\end{equation}
Following the \cite{1998AJ....116.1009R}, I employ the efficiency parameter ($\zeta$) during a common envelope evolution \citep {1976IAUS...73...75P} to obtain the binding energy, which is dependent on the detailed structure (i.e. radius of the companion at an evolutionary stage) of the dynamical process in binaries \citep{2000A&A...360.1043D}. Thus, I have
\begin{equation}
\zeta = \frac{\Delta E_{orb}}{\Delta E_{binding}}
\end{equation}
From this, I calculate the $a_{f}$ of the
new binary after the AIC process (see Fig. ~\ref{fig4}).
If the value of $a_{f}$ is too small, the core will overfill its new Roche lobe \citep{2015MNRAS.447.1713B}, then the cores will merge. As a result, gravitational waves may be produced during the pre-SNe phase \citep{2017ApJ...846..170T}.
I also deal with a system that orbits around a common center of mass \citep[see e.g.,][]{2002MNRAS.329..897H, 2022MNRAS.517.3993M}. Then, $a_{f}$ grows, but the system remains bounded by $a = \frac{r(1+e~cos\theta)}{1-e^{2}}$.
As a result, I can get both
$\Delta E_{orb,i}$ and $ E_{orb,f}$.
The initial binding energy of the system can be
\begin{equation}
E_{binding,i} \simeq {G~M_{b}} (\frac{1}{2a_{f}} - \frac{1}{2a_{i}} )
\end{equation}
where $M_{b} = (M_{WD}+M_{c,i}) = (M_{NS}+M_{c,f}$)
Hence, the final binding energy could be as
\begin{equation}
E_{binding,f}= E_{binding,i} + \frac{\Delta E_{orb}}{\zeta}
\end{equation}
In the picture presented here, I find that the initial binding energy will have a significant effect on the final fate of both evolution of the components (see Fig. ~\ref{fig5}), as well as on the orbital parameters that change during the conservative mass transfer. The majority of the mass is lost in an AIC binary system comes from the WD converting baryonic mass into binding energy during the collapse of the newborn NS, with a mass change of the order of $0.18M_{\odot}$ \citep{2022JHEAp..35...83T}. This can be happened due to their larger compactness \cite{2021arXiv210313605H}. If enough orbital energy continues to be lost, this would cause a merger of the binary components (like in some CO WDs merging). At this stage of evolution, this phenomenon created a common envelope phase by reducing orbital energy by envelope binding energy \citep{2004ApJ...612.1044P, 2015MNRAS.447.1713B} (see Fig. ~\ref{fig6}).
A note should be made concerning the given equation of state, because it will govern the specific binding energy for determining the binary's evolution. \cite{PhysRevD.102.103011} found a strong correlation between
several equations of state and tidal deformability associated with NS binding energies. Furthermore, they discovered that the lower limit of NS mass has a binding energy of 1.52$\times10^{53}$ ergs. Based on these interesting results, \cite{2021MNRAS.506.1462N}
estimated the amount of the gravitational mass of the NS, which is $ \rm \sim 1.2M_{\odot}$. This result is consistent with the NS observations made by \cite{2016ARA&A..54..401O, 2022PASA...39...40T}.
The result from the mass estimation benefits from the mildly recycled pulsar J0453+1559 \citep{2015ApJ...812..143M}. This system consists of a double NS with masses of $1.56M_{\odot}$ and $1.17M_{\odot}$, an orbital period of 4.07-day, a spin period of 45 ms, and an orbital eccentricity of (e =0.11). The amount of gravitational binding energy is expected to be around ($\Delta M \rm \sim 0.18M_{\odot}$) as shown in Fig. ~\ref{fig6}. This can be used to add some additional constraints when given a gravitational mass measurement for double NSs.
\section{Conclusions}
An interesting point should be made about the outcome of the final binding energy within the framework of the AIC process, taking into consideration the non-zero velocity of ejected matter. However, the equation of state influences how much binding energy an NS has. As a result, the larger the binding energy, the more compact the NS. The binding energy, thus dominates the gravitational waves at a given amount of accreted mass associated with the angular momentum loss. This would provide more insight into the AIC observations in the future. I have also investigated the effect of accretion on the orbital period on a NS's binding energy. With the use of the orbital period along with a donor star mass evolution, I have shown that it is possible to add some constraints on the expected values of both mass and binding for NS energy parameters that depend on the accretion rate. However, this will depend on the initial binding energy associated with the two systems' initial orbital energies, whether an envelope is present or not, since the final binding energy will depend on the initial binding energy.
This would help us figure out the accretion-driven evolution modes for binaries.
The corresponding increase in orbital energy propels the binary into a longer-period orbit.
This causes it to collapse and violently release its
gravitational energy, which might be observable by gravitational
wave observatories such as the LIGO-Virgo
Collaboration.
\section*{Acknowledgments}
Part of the content of this manuscript has been presented at the IAU Symposium 366, [doi:10.5281/zenodo.5759007]. Special thanks to Nour AlMusleh for helping the calculations.
The author would like to thank the anonymous referees for the careful reading of the manuscript and for all suggestions and comments which allowed us to improve both the quality and the clarity of the paper.
|
\section{Introduction}
Carlitz began a study of degenerate versions of Bernoulli and Euler polynomials, namely the degenerate Bernoulli and degenerate Euler polynomials (see \cite{02,03}). Recently, explorations for various degenerate versions of some special numbers and polynomials have regained interests of many mathematicians and yielded lots of fascinating and fruitful results. Indeed, this quest for degenerate versions even led to the development of degenerate umbral calculus \cite{07}, degenerate $q$-umbral calculus \cite{11} and degenerate gamma function \cite{09}. \par
Goubi introduced the generalized Euler-Genocchi polynomials (of order $\alpha$) in \cite{05}. A degenerate version of those polynomials, namely the generalized degenerate Euler-Genocchi polynomials, is investigated in \cite{10}. Here we study a`multi-version' of them, namely the multi-Euler-Genocchi and degenerate multi-Euler-Genocchi polynomials. The generating function of the Stirling numbers of the first kind is given by the usual logarithm (see \eqref{9}). Naturally, the multi-Stirling numbers of the first kind (see \eqref{13}) are defined by means of the multiple logarithm (see \eqref{7}). In the same way, both the multi-Euler-Genocchi and the degenerate multi-Euler-Genocchi polynomials are defined by using the multiple logarithm, and the latter is a degenerate version of the former (see \eqref{14}, \eqref{21}). \par
The aim of this note is to study the multi-Euler-Genocchi and degenerate multi-Euler-Genocchi polynomials which generalize respectively the generalized Euler-Genocchi (see \eqref{3}) and generalized degenerate Euler-Genocchi polynomials (see \eqref{6}). \par
The outline of this note is as follows. In Theorem 2.1, the value of the multi-Euler-Genocchi polynomial at 1 is expressed in terms of the multi-Stirling numbers of the first kind and the Stirling numbers of the second kind. The multi-Euler-Genocchi polynomials are represented in terms of the generalized Euler-Genocchi polynomials, the multi-Stirling numbers of the first kind and the Stirling numbers of the second kind in Theorem 2.2. Likewise, the degenerate multi-Euler-Genocchi polynomials are expressed in terms of the generalized degenerate Euler-Genocchi polynomials, the multi-Stirling numbers of the first kind and the Stirling numbers of the second kind in Theorem 2.4. A distribution type formula is derived for the degenerate multi-Euler-Genocchi polynomials in Theorem 2.6. In the rest of this section, we recall the facts that are needed throughout this paper. \par
The {\it{Euler polynomials}} are defined by
\begin{equation}\label{1}
\frac{2}{e^t+1}e^{xt}=\sum_{n=0} ^{\infty} E_n(x)\frac{t^n}{n!},{\text{ (see \cite{01, 04, 14})}}.
\end{equation}
When $x=0$, $E_n=E_n(0)$, $(n\geq0)$ are called the {\it{Euler numbers}}. The {\it{Genocchi polynomials}} are given by
\begin{equation}\label{2}
\frac{2t}{e^t+1}e^{xt}=\sum_{n=0} ^{\infty} G_n(x)\frac{t^n}{n!},{\text{ (see \cite{01, 04, 14})}}.
\end{equation}
When $x=0$, $G_n=G_n(0)$ are called the Genocchi numbers. We note that $G_n \in {\mathbb{Z}}$, $(n\geq0)$. \\
Recently, the {\it{generalized Euler-Genocchi polynomials}} are introduced as
\begin{equation}\label{3}
\frac{2t^r}{e^t+1}e^{xt}=\sum_{n=0} ^{\infty} A_n ^{(r)}(x)\frac{t^n}{n!},{\text{ (see \cite{05})}},
\end{equation}
where $r$ is a nonnegative integer. Note that $A_n ^{(0)}(x)=E_n(x)$, $A_n ^{(1)}(x)=G_n(x)$, $(n\geq0)$. \par
For any nonzero $\lambda\in{\mathbb{R}}$, the {\it{degenerate exponentials}} are defined by
\begin{equation}\label{4}
\e ^x(t)=\left(1+\lambda t\right)^{\frac{x}{\lambda}}=\sum_{n=0} ^{\infty} (x)_{n,\lambda}\frac{t^n}{n!},{\text{ (see \cite{08,12})}},
\end{equation}
where $(x)_{0,\lambda}=1$, $(x)_{n,\lambda}=x(x-\lambda)\cdots(x-(n-1)\lambda)$, $(n\geq1)$. In particular, for $x=1$, $\e(t)=\e ^1(t)= \sum_{n=0} ^{\infty} \frac{(1)_{n,\lambda}}{n!}t^n$. \par
In \cite{02,03}, Carlitz introduced the {\it{degenerate Euler polynomials}} given by
\begin{equation}\label{5}
\frac{2}{\e(t)+1}\e ^x(t)=\sum_{n=0} ^{\infty} {\mathcal{E}}_{n,\lambda}(x)\frac{t^n}{n!}.
\end{equation}
Note that $\lim_{n\rightarrow 0} {\mathcal{E}}_{n,\lambda}(x)=E_n(x)$, $(n\geq0)$.
In view of \eqref{2}, the degenerate Genocchi polynomials are defined by
\begin{equation*}
\frac{2t}{\e(t)+1}\e ^x(t)=\sum_{n=0} ^{\infty} G_{n,\lambda}(x) \frac{t^n}{n!},{\text{ (see \cite{10})}}.
\end{equation*}
Recently, the {\it{generalized degenerate Euler-Genocchi polynomials}} are defined by
\begin{equation}\label{6}
\frac{2t^r}{\e(t)+1}\e ^x (t)=\sum_{n=0} ^{\infty} A_{n,\lambda} ^{(r)}(x)\frac{t^n}{n!},{\text{ (see \cite{10})}}.
\end{equation}
When $x=0$, $A_{n,\lambda} ^{(r)}=A_{n,\lambda} ^{(r)}(0)$ are called the {\it{generalized degenerate Euler-Genocchi numbers}}. Note that $\lim_{\lambda\rightarrow 0} A_{n,\lambda} ^{(r)}(x)=A_n ^{(r)}(x)$, $(n\geq0)$. \par
For $k_1,k_2,\ldots,k_r\in{\mathbb{Z}}$, the {\it{multiple logarithm}} is defined by
\begin{equation}\label{7}
Li_{k_1,k_2,\ldots,k_r}(t)=\sum_{0<n_1<n_2<\cdots<n_r} \frac{t^{n_r}}{n_1 ^{k_1} n_2 ^{k_2} \cdots n_r ^{k_r}},~(|t|<1),{\text{ (see \cite{06})}}.
\end{equation}
The {\it{multi-Bernoulli numbers}} are defined by
\begin{equation}\label{8}
\frac{r!Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)}{(\e(t)-1)^r}=\sum_{n=0} ^{\infty} B_n ^{(k_1,k_2,\ldots,k_r)}\frac{t^n}{n!},{\text{ (see \cite{06,13})}}.
\end{equation}
It is well known that the {\it{Stirling numbers of the first kind}} are given by
\begin{equation}\label{9}
\frac{1}{k!}\left(\log(1+t)\right)^k=\sum_{n=k} ^{\infty} S_1(n,k) \frac{t^n}{n!},~(k\geq0),{\text{ (see \cite{04, 14})}}.
\end{equation}
As the inversion formula of \eqref{9}, the {\it{Stirling numbers of the second kind}} are defined by
\begin{equation}\label{10}
\frac{1}{k!}\left(e^t-1\right)^k =\sum_{n=k} ^{\infty} S_2(n,k)\frac{t^n}{n!},~(k\geq0),{\text{ (see \cite{14})}}.
\end{equation} \par
From \eqref{7}, we note that
\begin{equation}\label{11}
\begin{split}
\frac{d}{dt}Li_{k_1,k_2,\ldots,k_r}(t)=&\frac{d}{dt}\sum_{0<n_1<n_2<\cdots<n_r} \frac{t^n_r}{n_1 ^{k_1} n_2 ^{k_2}\cdots n_r ^{k_r}}\\
=&\frac{1}{t}Li_{k_1,\ldots,k_{r-1},k_r-1}(t),{\text{ (see \cite{08})}}.
\end{split}
\end{equation}
By \eqref{11}, we get
\begin{equation}\label{12}
Li_{\underbrace{1,1,\ldots,1}_{r-times}}(t)=\frac{1}{r!}\left(-\log(1-t)\right)^r=\sum_{n=r} ^{\infty} (-1)^{n-r}S_1(n,r)\frac{t^n}{n!}.
\end{equation} \par
In light of \eqref{9} and \eqref{12}, the {\it{multi-Stirling numbers of the first kind}} are defined by
\begin{equation}\label{13}
Li_{k_1,k_2,\ldots,k_r}(t)=\sum_{n=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(n,r)\frac{t^n}{n!},{\text{ (see \cite{13})}}.
\end{equation}
Note from \eqref{12} and \eqref{13} that
\begin{equation*}
S_1 ^{\overbrace{(1,1,\ldots,1)}^{r-times}}(n,r)=(-1)^{n-r}S_1(n,r), \,\,(n,r\geq0).
\end{equation*}
\section{Multi-Euler-Genocchi and degenerate multi-Euler-Genocchi polynomials}
We consider the {\it{multi-Euler-Genocchi polynomials}} given by
\begin{equation}\label{14}
\frac{2r!}{e^t+1}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)e^{xt}=\sum_{n=0} ^{\infty} A_n ^{(k_1,k_2,\ldots,k_r)}(x)\frac{t^n}{n!},
\end{equation}
where $r$ is a nonnegative integer. When $x=0$, $A_n ^{(k_1,k_2,\ldots,k_r)}=A_n ^{(k_1,k_2,\ldots,k_r)}(0)$ are called the {\it{multi-Euler-Genocchi numbers}}.
From \eqref{12} and \eqref{14}, we note that
\begin{equation}\label{15}
\begin{split}
\sum_{n=0} ^{\infty} A_n ^{\overbrace{(1,1,\ldots,1)}^{r-times}}(x)\frac{t^n}{n!}=&\frac{2r!}{e^t+1}Li_{\underbrace{1,1,\ldots,1}_{r-times}}\left(1-e^{-t}\right)e^{xt}\\
=&\frac{2t^r}{e^t+1}e^{xt}=\sum_{n=0} ^{\infty} A_n ^{(r)}(x)\frac{t^n}{n!}.
\end{split}
\end{equation}
Thus, by \eqref{15}, we get
\begin{equation}\label{16}
A_n ^{(\overbrace{1,1,\ldots,1}^{r-times})}(x) =A_n ^{(r)}(x),~(n\geq0).
\end{equation}
From \eqref{13} and \eqref{15}, we have
\begin{equation}\label{17}
\begin{split}
&\sum_{n=0} ^{\infty} A_n ^{(k_1,k_2,\ldots,k_r)}(1)\frac{t^n}{n!}=\frac{r!}{1+\frac{1}{2}\left(e^{-t}-1\right)}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)\\
=&r!\sum_{l=0} ^{\infty}\left(\frac{1}{2}\right)^l(-1)^l \left(e^{-t}-1\right)^l\sum_{m=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(m,r)(-1)^m\frac{1}{m!}\left(e^{-t}-1\right)^m\\
=&r!\sum_{l=0}^{\infty}\sum_{m=r}^{\infty}\left(\frac{1}{2}\right)^{l}S_{1}^{(k_1,k_2,\dots,k_r)}(m,r)(-1)^{l+m}\frac{(l+m)!}{m!}\frac{1}{(l+m)!}(e^{-t}-1)^{l+m}\\
=&r!\sum_{l=0}^{\infty}\sum_{m=r}^{\infty}\left(\frac{1}{2}\right)^{l}S_{1}^{(k_1,k_2,\dots,k_r)}(m,r)(-1)^{l+m}\frac{(l+m)!}{m!}\sum_{n=l+m}^{\infty}S_{2}(n,l+m)(-1)^{n}\frac{t^n}{n!}\\
=&\sum_{n=r} ^{\infty} \left(r!\sum_{k=r} ^n \sum_{m=r} ^k \left(\frac{1}{2}\right)^{k-m}(-1)^{n-k}S_1 ^{\left(k_1,k_2,\ldots,k_r\right)}(m,r)\frac{k!}{m!}S_2(n,k)\right)\frac{t^n}{n!}.
\end{split}
\end{equation}
Therefore, by comparing the coefficient on both sides of \eqref{17}, we obtain the following theorem.
\begin{theorem}
For $n,r\in{\mathbb{Z}}$ with $n \geq r \geq 0$, we have
\begin{equation*}
A_n ^{(k_1,k_2,\ldots,k_r)}(1)=r!\sum_{k=r} ^n \sum_{m=r} ^k \left(\frac{1}{2}\right)^{k-m}(-1)^{n-k}\frac{k!}{m!}S_1 ^{(k_1,k_2,\ldots,k_r)}(m,r)S_2(n,k),
\end{equation*}
and, for $0 \le n<r$, we have
\begin{equation*}
A_n ^{(k_1,k_2,\ldots,k_r)}(1)=0.
\end{equation*}
\end{theorem}
By \eqref{14}, we get
\begin{equation}\label{18}
\begin{split}
&\sum_{n=0} ^{\infty} A_n ^{(k_1,k_2,\ldots,k_r)}(x)\frac{t^n}{n!}=\frac{2r!}{e^t+1}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)e^{xt}\\
=&r!\frac{2}{e^t+1}e^{xt}\sum_{l=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(l,r)(-1)^l\frac{1}{l!}\left(e^{-t}-1\right)^l\\
=&\frac{r!2e^{xt}}{e^t+1}\sum_{l=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(l,r)(-1)^l\sum_{m=l} ^{\infty} S_2(m,l)(-1)^m\frac{t^m}{m!}\\
=&\frac{r!2e^{xt}}{e^t+1}\sum_{m=r}^{\infty}\left(\sum_{l=r} ^m S_1^{(k_1,k_2,\ldots,k_r)}(l,r)S_2(m,l)(-1)^{m-l}\right)\frac{t^m}{m!}\\
=&\frac{r!2t^re^{xt}}{e^t+1}\sum_{m=0} ^{\infty}\left(\sum_{l=r} ^{m+r}S_1 ^{(k_1,k_2,\ldots,k_r)}(l,r)\frac{S_2(m+r,l)m!}{(m+r)!}(-1)^{m+r-l}\right)\frac{t^m}{m!}\\
=&\sum_{j=0} ^{\infty} A_j ^{(r)}(x)\frac{t^j}{j!}\sum_{m=0} ^{\infty}\left(\sum_{l=r} ^{m+r}\frac{S_1 ^{(k_1,k_2,\ldots,k_r)}(l,r)}{\binom{m+r}{r}}S_2(m+r,l)(-1)^{m+r-l}\right)\frac{t^m}{m!}\\
=&\sum_{n=0} ^{\infty}\left(\sum_{m=0} ^n \binom{n}{m}A_{n-m} ^{(r)}(x)\sum_{l=r} ^{m+r} \frac{S_1 ^{(k_1,k_2,\ldots,k_r)}(l,r)}{\binom{m+r}{r}}S_2(m+r,l)(-1)^{m+r-l}\right)\frac{t^n}{n!}.
\end{split}
\end{equation}
Therefore, by comparing the coefficients on both sides of \eqref{18}, we obtain the following theorem.
\begin{theorem}
For $n \geq0$, we have
\begin{equation*}
A_n ^{(k_1,k_2,\ldots,k_r)}(x)=\sum_{m=0} ^n \binom{n}{m}A_{n-m} ^{(r)}(x)\sum_{l=r} ^{m+r} \frac{S_1 ^{(k_1,k_2,\ldots,k_r)}(l,r)}{\binom{m+r}{r}}S_2(m+r,l)(-1)^{m+r-l}.
\end{equation*}
\end{theorem}
From \eqref{13}, we note that
\begin{equation}\label{19}
\begin{split}
&Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)=\sum_{k=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)\frac{1}{k!}\left(1-e^{-t}\right)^k\\
=&\sum_{k=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)\sum_{n=k} ^{\infty} (-1)^{n-k}S_2(n,k)\frac{t^n}{n!}\\
=&\sum_{n=r} ^{\infty}\left(\sum_{k=r} ^n S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)(-1)^{n-k}S_2(n,k)\right)\frac{t^n}{n!}.
\end{split}
\end{equation}
On the other hand, by \eqref{14}, we get
\begin{equation}\label{20}
\begin{split}
Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)=&\frac{1}{2r!}\frac{2r!}{e^t+1}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)\left(e^t+1\right)\\
=&\frac{1}{2r!}\sum_{n=0} ^{\infty}\left(A_n ^{(k_1,k_2,\ldots,k_r)}(1)+A_n ^{(k_1,k_2,\ldots,k_r)}\right)\frac{t^n}{n!}.
\end{split}
\end{equation}
Therefore, by \eqref{19} and \eqref{20}, we obtain the following theorem.
\begin{theorem}
For $n,r\geq0$, we have
\begin{equation*}
A_n ^{(k_1,k_2,\ldots,k_r)}(1)+A_n ^{(k_1,k_2,\ldots,k_r)}=
\begin{cases}
2r!\sum_{k=r} ^n S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)(-1)^{n-k}S_2(n,k), & {\text{if }}n\geq r,\\
0,& {\text{if }}0 \le n<r.
\end{cases}
\end{equation*}
\end{theorem}
Now, we consider the {\it{degenerate multi-Euler-Genocchi polynomials}} given by
\begin{equation}\label{21}
\frac{2r!}{\e(t)+1}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)\e^x(t)=\sum_{n=0} ^{\infty} A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)\frac{t^n}{n!}.
\end{equation}
When $x=0$, $A_{n,\lambda} ^{k_1,k_2,\ldots,k_r)}=A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(0)$ are called the {\it{degenerate multi-Euler-Genocchi numbers}}.
Thus, by \eqref{21}, we get
\begin{equation}\label{22}
\begin{split}
\sum_{n=0} ^{\infty} A_{n,\lambda} ^{(\overbrace{1,1,\ldots,1}^{r-times})}(x)\frac{t^n}{n!}=&\frac{2r!}{\e(t)+1}Li_{\underbrace{1,1,\ldots,1}_{r-times}}\left(1-e^{-t}\right)\e^x(t)\\
=&\frac{2t^r}{\e(t)+1}\e^x(t)=\sum_{n=0} ^{\infty} A_{n,\lambda} ^{(r)}(x)\frac{t^n}{n!}.
\end{split}
\end{equation}
From \eqref{22}, we have
\begin{equation}\label{23}
A_{n,\lambda} ^{(\overbrace{1,1,\ldots,1}^{r-times})}(x)=A_{n,\lambda} ^{(r)}(x),~(n\geq0).
\end{equation}
By \eqref{21}, we get
\begin{equation}\label{24}
\begin{split}
&\sum_{n=0} ^{\infty} A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)\frac{t^n}{n!}=\frac{2r!}{\e(t)+1}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)\e^x(t)\\
=&\frac{2r!}{\e(t)+1}\e^x(t)\sum_{m=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(m,r)\frac{1}{m!}\left(1-e^{-t}\right)^m\\
=&\frac{2r!}{\e(t)+1}\e^x(t)\sum_{m=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(m,r)\sum_{l=m} ^{\infty} S_2(l,m)(-1)^{l-m}\frac{t^l}{l!}\\
=&\frac{2r!}{\e(t)+1}\e^x(t)\sum_{l=r} ^{\infty}\left(\sum_{m=r} ^l (-1)^{l-m}S_2(l,m)S_1^{(k_1,k_2,\ldots,k_r)}(m,r)\right)\frac{t^l}{l!}\\
=&\sum_{j=0} ^{\infty} A_{j,\lambda} ^{(r)}(x)\frac{t^j}{j!}\sum_{l=0} ^{\infty}\left(\sum_{m=r} ^{l+r}\frac{S_2(l+r,m)}{\binom{l+r}{r}}(-1)^{l-m-r}S_1 ^{(k_1,k_2,\ldots,k_r)}(m,r)\right)\frac{t^l}{l!}\\
=&\sum_{n=0} ^{\infty}\left(\sum_{l=0} ^n \binom{n}{l}A_{n-l,\lambda} ^{(r)}(x)\sum_{m=r} ^{l+r}\frac{S_2(l+r,m)}{\binom{l+r}{l}}(-1)^{l-m-r}S_1 ^{(k_1,k_2,\ldots,k_r)}(m,r)\right)\frac{t^n}{n!}.
\end{split}
\end{equation}
Therefore, by comparing the coefficients on both sides of \eqref{24}, we obtain the following theorem.
\begin{theorem}
for $n \geq 0$, we have
\begin{equation*}
A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)=\sum_{l=0} ^n \binom{n}{l}A_{n-l,\lambda} ^{(r)}(x)\sum_{m=r} ^{l+r}\frac{S_2(l+r,m)}{\binom{l+r}{l}}(-1)^{l-m-r}S_1 ^{(k_1,k_2,\ldots,k_r)}(m,r).
\end{equation*}
\end{theorem}
From \eqref{21}, we note that
\begin{equation}\label{25}
\begin{split}
&\sum_{n=0} ^{\infty} A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)\frac{t^n}{n!}=\frac{2r!Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)}{\e(t)+1}\e^x(t)\\
=&\sum_{k=0} ^{\infty} A_{k,\lambda} ^{(k_1,k_2,\ldots,k_r)}\frac{t^k}{k!}\sum_{m=0} ^{\infty}(x)_{m,\lambda}\frac{t^m}{m!}\\
=&\sum_{n=0} ^{\infty} \left(\sum_{k=0} ^n \binom{n}{k} A_{k,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)_{n-k,\lambda}\right)\frac{t^n}{n!}.
\end{split}
\end{equation}
Thus, by \eqref{25}, we obtain the following theorem.
\begin{theorem}
For $n\geq0$, we have
\begin{equation*}
A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)=\sum_{k=0} ^n \binom{n}{k} A_{k,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)_{n-k,\lambda}.
\end{equation*}
\end{theorem}
Assume that $m \in {\mathbb{N}}$ with $m\equiv1$ $({\text{mod }}2)$. Then we have
\begin{equation}\label{26}
\begin{split}
&\sum_{n=0} ^{\infty} A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)\frac{t^n}{n!}=\frac{2r!}{\e(t)+1}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)\e^x(t)\\
=&\frac{2r!}{\e^m(t)+1}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)\sum_{l=0} ^{m-1}(-1)^l\e^{l+x}(t)\\
=&\frac{r!}{t^r}Li_{k_1,k_2,\ldots,k_r}\left(1-e^{-t}\right)\sum_{l=0} ^{m-1}(-1)^l\frac{2t^r}{e_{\lambda/m}(mt)+1}e_{\lambda/m} ^{\frac{l+x}{m}}(mt)\\
=&\frac{r!}{t^r}\sum_{k=r} ^{\infty} S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)\sum_{j=k} ^{\infty} S_2(j,k)(-1)^{j-k}\frac{t^j}{j!}\sum_{l=0} ^{m-1}(-1)^l\sum_{i=0} ^{\infty}A_{i,\lambda/m} ^{(r)}\left(\frac{l+x}{m}\right)m^{i-r}\frac{t^i}{i!} \\
=&\frac{r!}{t^r}\sum_{j=r} ^{\infty} \left(\sum_{k=r} ^j S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)(-1)^{j-k}S_2(j,k)\right)\frac{t^j}{j!}\sum_{i=0} ^{\infty}\left(\sum_{l=0} ^{m-1} (-1)^lA_{i,\lambda/m} ^{(r)}\left(\frac{l+x}{m}\right)m^{i-r}\right)\frac{t^i}{i!}\\
=&\sum_{j=0} ^{\infty}\left(\sum_{k=r} ^{j+r} S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)(-1)^{j+r-k}S_2(j+r,k)\frac{r!j!}{(j+r)!}\right)\frac{t^j}{j!}\\
&\times \sum_{i=0} ^{\infty} \left(\sum_{l=0} ^{m-1}(-1)^lA_{i,\lambda/m} ^{(r)} \left(\frac{l+x}{m}\right)m^{i-r}\right)\frac{t^i}{i!}\\
=&\sum_{n=0} ^{\infty}\left(\sum_{j=0} ^n \binom{n}{j}\sum_{k=r} ^{j+r}\sum_{l=0} ^{m-1} S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)(-1)^{j+r-k}\right.\\
&\times \left. \frac{S_2(j+r,k)}{\binom{j+r}{r}}(-1)^lA_{n-j,\lambda/m} ^{(r)}\left(\frac{l+x}{m}\right)m^{n-j-r}\right)\frac{t^n}{n!}.
\end{split}
\end{equation}
Therefore, by comparing the coefficients on both sides of \eqref{26}, we obtain the following theorem.
\begin{theorem}
For $m \in {\mathbb{N}}$ with $m\equiv 1$ ({\rm{mod}} $2$), we have
\begin{equation*}
\begin{split}
A_{n,\lambda} ^{(k_1,k_2,\ldots,k_r)}(x)=&\sum_{j=0} ^n \binom{n}{j}\sum_{k=r} ^{j+r}\sum_{l=0} ^{m-1} S_1 ^{(k_1,k_2,\ldots,k_r)}(k,r)(-1)^{j+r-k}\\
&\times \frac{S_2(j+r,k)}{\binom{j+r}{r}}(-1)^lA_{n-j,\lambda/m} ^{(r)}\left(\frac{l+x}{m}\right)m^{n-j-r}.
\end{split}
\end{equation*}
\end{theorem}
\section{Conclusion}
In addition to degenerate versions of many special numbers and polynomials, the degenerate gamma function, degenerate umbral calculus and degenerate $q$-umbral calculus are introduced and a lot of interesting results about them are found in recent years.\par
In this note, we introduced the multi-Euler-Genocchi and degenerate multi-Euler-Genocchi polynomials which are multi-versions of the generalized Euler-Genocchi and generalized degenerate Euler-Genocchi polynomials. Among other things, we expressed the former by the generalized Euler-Genocchi polynomials, the multi-Stirling numbers of the first kind and Stirling numbers of the second kind, and the latter by the generalized degenerate Euler-Genocchi polynomials, the multi-Stirling numbers of the first kind and Stirling numbers of the second kind. \par
It is one of our future projects to continue to study various degenerate versions of some special numbers and polynomials and those of certain transcendental functions, and to find their applications to physics, science and engineering as well as to mathematics.
|
\section{Challenges in learning manipulation}
Manipulation and grasping are among the most fundamental topics in robotics. This classic field has been rejuvenated by the recent boom in embodied AI, wherein an agent (\eg, a robot) is tasked to learn by interacting with its environment. Since then, learning-based methods have been widely applied and have elevated robots' manipulation competence. Often, robots either train on data directly obtained from sensors (\eg, object grasping from a cluster~\cite{pinto2016supersizing,mahler2019learning}, pick-and-place~\cite{zeng2018robotic}, object handover~\cite{cini2019choice}, or door opening~\cite{yahya2017collective}) or learn from human demonstrations (\eg, motor motion~\cite{schaal2003computational,maeda2016acquiring}, affordance~\cite{nguyen2016detecting,kokic2017affordance}, task structure~\cite{mohseni2015interactive,xiong2016robot,liu2019mirroring}, or reward functions~\cite{abbeel2004apprenticeship,prieur2012modeling,ibarz2018reward}).
Learning meaningful manipulation has a unique prerequisite: It must incorporate fine-grained physics to convey an understanding of the complex process that occurs during the interaction. Although we have witnessed the solid advancement of certain embodied AI tasks (\eg, visual-language navigation), these successes are primarily attributed to the readily available plain images and their annotations (pixels, segments, or bounding boxes) that are directly extracted from the existing training platforms~\cite{xie2019vrgym,li2021igibson,szot2021habitat}, while physics information during the interactions is still lacking. Similarly, although modern vision-based sensors and motion-capture systems can collect precise trajectory information, neither can precisely estimate physical properties during interactions. Existing software and hardware systems are insufficient for learning sophisticated manipulation skills for the following three reasons:
\begin{figure*}[t!]
\centering
\begin{overpic}
[width=\linewidth]{system_overview/comparison_update_v2}
\put(10,36.6){\color{black} \textbf{(a) Tactile-sensing mode}~\cite{liu2017glove}}
\put(45,36.6){\color{black} \textbf{(b) VR mode}~\cite{liu2019high}}
\put(75,36.6){\color{black} \textbf{(c) Simulation mode}}
\end{overpic}
\caption{\textbf{Overview of our reconfigurable data glove in three operating modes, which share a unified backbone design of an IMU network that captures the hand gesture.} (a) The tactile-sensing mode records the force exerted by the hand during manipulation. (b) The VR mode supports stable grasping of virtual objects in VR applications and provides haptic feedback via vibration motors. Contact configurations are conveniently logged. (c) The simulation mode incorporates state-of-the-art FEM simulation~\cite{li2020incremental} to augment the grasp data with fine-grained changes in the object's properties.}
\label{fig:summary}
\end{figure*}
First, understanding fine-grained manipulation or human-object interactions requires a joint understanding of both hand gesture\footnote{In this article, the phrase ``hand gesture'' is used to refer to the collective movement of the fingers and palm, whereas ``hand pose'' is used to refer to the position and orientation of the wrist.} and force~\cite{liu2017glove}; distinguishing certain actions purely based on the hand gesture is challenging, if not impossible. For example, in the task of opening a medicine bottle that requires either pushing or squeezing the lid to unlock the childproof mechanism, it is insufficient to differentiate the opening actions by visual information alone, because the pushing and squeezing actions are visually similar (or even identical) to each other~\cite{edmonds2019tale}. Reconstructing hand gestures or trajectories alone has already been shown to be challenging, as severe hand-object occlusion hinders the data collection reliability. To tackle this problem, we introduce a \textit{tactile-sensing glove} to jointly capture hand gestures through a network of inertial measurement units (IMUs) and force exerted by the hand using six customized force sensors during manipulation. The force sensors are constructed from Velostat--a piezoresistive fabric with changing resistance under pressures, which is soft and thin to allow natural hand motions. Together, the force sensors provide a holistic view of manipulation events. A preliminary version of this system has been presented in the work of Liu \textit{et al}\onedot~\cite{liu2017glove}; interested readers can refer to the Appendix for details.
Second, contact points between hand and object play a significant role in understanding why and how a specific grasp is chosen. Such information is traditionally challenging to obtain (\eg, through thermal imaging~\cite{brahmbhatt2019contactdb}). To address this challenge, we devise a \textit{VR glove} and leverage VR platforms to obtain contact points. This design incorporates a caging-based approach to determine a stable grasp of a virtual object based on the collision geometry between fingers and the object. The collisions trigger a network of vibration motors on the glove to provide haptic feedback. The VR glove jointly collects trajectory and contact information that is otherwise difficult to obtain physically. A preliminary version of this system has been presented in the work of Liu \textit{et al}\onedot~\cite{liu2019high}; interested readers can refer to the Appendix for details.
Third, much attention has been paid to collecting hand information during fine manipulation but not to the object being manipulated or its effects caused by actions. This deficiency prohibits the use of collected data for studying complex manipulation events. For example, consider a tool-use scenario. A manipulation event cannot be comprehensively understood without capturing the \textit{interplay} between the human hand, the tool being manipulated, and the action effects. As such, this perspective demands a solution beyond the classic hand-centric view in developing data gloves. Furthermore, since the effects caused by the manipulation actions are traditionally difficult to capture, they are often treated as a task of recognizing discrete, symbolic states or attributes in computer vision~\cite{duan2012discovering,liu2017jointly,nagarajan2018attributes}, losing their intrinsic continuous nature. To overcome these limits of traditional data gloves, we propose to integrate a physics-based simulation using the state-of-the-art FEM~\cite{li2020incremental} to model object fluents--the time-varying states in the event~\cite{newton1736method}--and other physical properties involved, such as contact forces and the stress within the object. This \textit{glove with simulation} captures a human manipulation action and analyzes it in four-dimensional (4D) space by including: (i) the contact and geometric information of the hand gesture and the object in three-dimensional (3D) space, and (ii) the transition and coherence between the object's fluent changes and the manipulation events over time. To the best of our knowledge, this is the first time such 4D data offering a holistic view of manipulation events is used in this field, and its use will open up new avenues for studying manipulations and grasping.
Sharing a unified backbone design that reconstructs hand gestures in real-time, the proposed data glove can be easily reconfigured to (i) capture force exerted by hand using piezoresistive material, (ii) record contact information by grasping stably in VR, or (iii) reconstruct both visual and physical effects during the manipulation by integrating physics-based simulation. Our system extends the long history of developing data gloves~\cite{dipietro2008survey} and endows embodied AI agents with a deeper understanding of hand-object interactions.
This paper makes three contributions compared with prior work~\cite{liu2017glove,liu2019high}. First, we introduce the concept of a reconfigurable glove-based system. The three operating modes tackle a broader range of downstream tasks with distinct features. This extension does not sacrifice the easy-to-replicate nature, as different modes share a unified backbone design. Second, a state-of-the-art FEM-based physical simulation is integrated to augment the grasp data with simulated action effects, thereby providing new opportunities for studying hand-object interactions and complex manipulation events. Third, we demonstrate that the data collected by our glove-based system--either virtually or physically--is effective for learning in a series of case studies.
\subsection{Related work}
\paragraph{Hand gesture sensing}
Recording finger joints' movements is the core of hand gesture sensing. Various types of hardware have been adopted to acquire hand gestures. Although curvature/flex sensors~\cite{kramer2011soft,kamel2008glove}, liquid metal~\cite{oh2021liquid}, a stretchable strain sensor~\cite{wang2020gesture}, and triboelectric material~\cite{wen2020machine} are among proven approaches, these can only measure unidirectional bending angles. Hence, they are less efficient for recording a hand's metacarpophalangeal (MCP) joints with two degrees of freedom (DoFs) for finger abduction and adduction. In addition, by wrapping around bending finger joints, these instruments sacrifice natural hand movements due to their large footprint and rigidness. In comparison, IMUs can measure one phalanx's 6-DoF pose, interfere less with joint motions, and perform more consistently over an extended period of time. As a result, adopting IMUs in data gloves has prevailed in modern design, including IMUs channeled by a Zigbee network~\cite{taylor2013forward}, a circuit board with a 6-DoF accelerometer/gyroscope and a 3-DoF magnetometer placed on each of the 15 phalanxes~\cite{kortier2014assessment}, and a population of IMUs connected through flexible cables~\cite{hu2020flexible}. Often, the raw sensory information requires further filtering~\cite{santaera2015low} and estimation~\cite{kortier2014assessment,ligorio2013extended,kortier2015hand}.
\paragraph{Force sensing}
Sensing the forces exerted by a hand during manipulation has attracted growing research attention and requires a more integrated glove-based system. Here, we highlight some signature designs. An elastomer sensor with embedded liquid-metal material~\cite{hammond2014toward} was able to sense force across a large area (\eg, the palm) and estimate joint movements by measuring skin strain. FlexiForce sensors can acquire hand forces~\cite{gu2015fine}, while an optical-based motion-capture system tracks hand gestures. Forces and gestures can also be estimated using 9-DoF IMUs without additional hardware~\cite{mohammadi2016fingertip}, although the force estimation is crude. Other notable designs involve specialized hardware, including force-sensitive resistors~\cite{lin2019novel} and a specific tactile sensor for fingertips~\cite{battaglia2016thimblesense}. Recently, soft films made from piezoresistive materials whose resistance changes under pressing forces (\eg, Velostat) have become increasingly popular in robotic applications; this type of material permits force sensing without constraining the robots' or human hand's motions~\cite{low2015pressure,pugach2016touch,muller2015smart,jeong2011finger}.
\subsection{Overview: Three modes of the reconfigurable data glove}
To tackle the aforementioned challenges and fill in the gap in the literature, we devised a reconfigurable data glove that is capable of operating in three modes for various downstream tasks with distinct features and goals.
\paragraph{Tactile-sensing mode}
We start with a glove design using an IMU configuration~\cite{kortier2014assessment} to reconstruct hand gestures. Our system's software and hardware designs are publicly available for easy replication. A customized force sensor made from Velostat--a soft fabric whose resistance changes under different pressures--is adopted to acquire the force distributions over large areas of the hand without constraining natural hand motions. \cref{fig:summary}a~\cite{liu2017glove,liu2019high,li2020incremental} summarizes this tactile-sensing glove design.
\paragraph{VR mode}
By reconstructing virtual grasps in VR, this mode provides supplementary contact information (\eg, contact points on an object) during manipulation actions. In contrast to the dominating symbolic grasp methods that directly attach the virtual object to the virtual hand when a grasp event is triggered~\cite{boulic1996multi}, our glove-based system enables a natural and realistic grasp experience with a fine-grained hand gesture reconstruction and force estimated at specific contact points; a symbolic grasp would cause finger penetrations or non-contacting (\eg, see examples in \cref{fig:vr:comparison_touch}), since the attachments between the hand and object are predefined. Although collecting grasp-related data in VR is more convenient and economical than other specialized data-acquisition pipelines, the lack of direct contact between the hand and physical objects inevitably leads to less natural interactions. Thus, providing haptic feedback is critical to compensate for this drawback. We use vibration motors to provide generic haptic feedback to each finger, thereby increasing the realism of grasping in VR. \cref{fig:summary}b~\cite{liu2017glove,liu2019high,li2020incremental} summarizes the VR glove design.
\paragraph{Simulation mode}
Physics-based simulations emulate a system's precise changes over time, thus opening up new directions for robot learning\cite{choi2021use}, including learning robot navigation~\cite{xie2019vrgym}, bridging human and robot embodiments in learning from demonstration~\cite{liu2019mirroring}, soft robot locomotion~\cite{hu2019chainqueen}, liquid pouring~\cite{kennedy2019autonomous}, and robot cutting~\cite{heiden2021disect}. In a similar vein, simulating how an object's fluent changes as the result of a given manipulation action provides a new perspective on hand-object interactions. In this article, we adopt a state-of-the-art FEM simulator~\cite{li2020incremental} to emulate the causes and effects of manipulation events. As shown in \cref{fig:summary}c~\cite{liu2017glove,liu2019high,li2020incremental}, by integrating physical data collected by the data glove with simulated effects, our system reconstructs a new type of 4D manipulation data with high-fidelity visual and physical properties on a large scale. We believe that this new type of data can significantly impact how manipulation datasets are collected in the future and can assist in a wide range of manipulation tasks in robot learning.
\subsection{Structure of this article}
The remainder of this article is organized as follows. We start with a unified design for hand gesture sensing in\cref{sec:gesture}. With different goals, the tactile-sensing mode~\cite{liu2017glove} and the VR mode~\cite{liu2019high} are presented in \cref{sec:tactile} and \cref{sec:vr}, respectively. A new state-of-the-art, physics-based simulation using FEM~\cite{wolper2019cd} is integrated in \cref{sec:sim} to collect 4D manipulation data, which is the very first in the field to achieve such high fidelity, to the best of our knowledge. We evaluate our system in three modes in \cref{sec:application} and conclude the paper in \cref{sec:conclusion}.
\section{A unified backbone design for gesture sensing}\label{sec:gesture}
This section introduces the IMU setup for capturing hand gestures in \cref{sec:gesture:reconst}. As this setup is shared among all three modes of the proposed reconfigurable data glove, we further evaluate the IMU performance in \cref{sec:gesture:imu_eval}.
\subsection{Hand gesture reconstruction}\label{sec:gesture:reconst}
\paragraph{IMU specification}
Fifteen Bosch BNO055 9-DoF IMUs are deployed for hand gesture sensing. One IMU is mounted to the palm, two IMUs to the thumb's distal and intermediate phalanges, and the remaining 12 are placed on the phalanxes of the other four fingers. Each IMU includes a 16-bit triaxial gyroscope, a 12-bit triaxial accelerometer, and a triaxial geomagnetometer. This IMU is integrated with a built-in proprietary sensor fusion algorithm running on a 32-bit microcontroller, yielding each phalanx's pose in terms of a quaternion. The geomagnetometer acquires an IMU's reference frame to the Earth's magnetic field, supporting the pose calibration protocol (introduced later). The small footprint of the BNO055 ($5~cm \times 4.5~cm$) allows easy attachment to the glove and minimizes interference with natural hand motions. A pair of TCA9548A $I^2C$ multiplexers is used for networking the 15 IMUs and connecting them to the $I^2C$ bus interfaces on a Raspberry Pi 2 Model B board (henceforth RPi for brevity); RPi acts as the master controller for the entire glove system.
\paragraph{Hand forward kinematics}
A human hand has about 20 DoFs: both the proximal interphalangeal (PIP) joint and the distal interphalangeal (DIP) joint have one DoF, whereas an MCP joint has two. Based on this anatomical structure, we model each finger by a 4-DoF kinematic chain whose base frame is the palm and the end-effector frame is the distal phalanx. The thumb is modeled as a 3-DoF kinematic chain consisting of a DIP joint and an MCP joint.
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\linewidth]{hand_kinematic/hand_kin}
\caption{\textbf{The kinematic chain of the index finger with coordinate frames attached.} Reproduced from Ref.~\cite{liu2017glove} with permission.}
\label{fig:gesture:hand_kin}
\end{figure}
After obtaining a joint's rotational angle using two consecutive IMUs, the position and orientation of each phalanx can be computed by forward kinematics. \cref{fig:gesture:hand_kin}~\cite{liu2017glove} shows an example of the index finger's kinematic chain and the attached frame. Frame 1 is assigned to the palm, and Frames 2, 3, and 4 are assigned to the proximal, middle, and distal phalanx, respectively. The proximal, middle, and distal phalanx lengths are respectively denoted by $l_1$, $l_2$, and $l_3$, The flexion and extension angles of the MCP, PIP, and DIP joints are denoted as $\theta_1$, $\theta_2$, and $\theta_3$, respectively. In addition, the MCP joint has an abduction and adduction angle denoted as $\beta$. $d_x$ and $d_y$ are the offsets in the $x$ and $y$ directions between the palm's center and the MCP joint. \cref{table:gesture:DH} derives the Denavit-Hartenberg (D-H) parameters for each reference frame, wherein a general homogeneous transformation matrix $\mathbf{T}$ from frames $i-1$ to $i$ can be given by the following:
\begin{equation}
^{i-1}_{i}\mathbf{T} =
\begin{bmatrix}
c\theta_i & -s\theta_i & 0 & a_{i-1} \\
s\theta_{i}c\alpha_{i-1} & c\theta_{i}c\alpha_{i-1} & -s\alpha_{i-1} & -s\alpha_{i-1}d_i \\
s\theta_{i}s\alpha_{i-1} & c\theta_{i}s\alpha_{i-1} & c\alpha_{i-1} & c\alpha_{i-1}d_i \\
0 & 0 & 0 & 1
\end{bmatrix},
\end{equation}
where $c\theta_i$ and $s\theta_i$ denote $\cos{(\theta_i)}$ and $\sin{(\theta_i)}$, respectively.
\begin{table}[ht!]
\centering
\small
\caption{\textbf{Denavit-Hartenberg parameters of a finger.}}
\label{table:gesture:DH}
\begin{tabular}{ccccc}
\toprule
Link ID & $\alpha_{i-1}$ & $a_{i-1}$ & $\theta_{i}$ & $d_{i}$\\
\midrule
1 & 0 & 0 & $\beta$ & 0\\
2 & ${\pi}/{2}$ & $l_1$ & $\theta_1$ & 0\\
3 & 0 & $l_2$ & $\theta_2$ & 0\\
4 & 0 & $l_3$ & $\theta_3$ & 0\\
\bottomrule
\end{tabular}
\end{table}
\cref{table:gesture:trans} lists the homogeneous transformation matrices of each phalanx, which can be used to express each phalanx's pose in the palm's reference frame in the cartesian space. The forward kinematics model keeps better track of the sensed hand gesture by reducing the inconsistency due to IMU fabrication error and anatomical variations among the users' hands.
\begin{table}[ht!]
\centering
\small
\caption{\textbf{Concatenation of transformation matrices.}}
\label{table:gesture:trans}
\begin{tabular}{cc}
\toprule
Phalanx & Transformation \\
\midrule
Proximal & $^{0}_{1}T ^{1}_{2}T$ \\
Middle/Distal for thumb & $^{0}_{1}T ^{1}_{2}T ^{2}_{3}T$\\
Distal & $^{0}_{1}T ^{1}_{2}T ^{2}_{3}T ^{3}_{4}T$\\
\midrule
\end{tabular}
\end{table}
\paragraph{Joint limits}
We adopt a set of commonly used inequality constraints~\cite{lin2000modeling} to limit the motion ranges of the finger joints, thereby eliminating unnatural hand gestures due to sensor noise:
\begin{equation}
\begin{aligned}
\text{MCP joint} &:
\begin{cases}
0^{\circ} \le \theta_{1} \le 90^{\circ} \\
-15^{\circ} \le \beta \le 15^{\circ} \\
\end{cases}
\\
\text{PIP joint} &:
0^{\circ} \le \theta_{2} \le 110^{\circ}
\\
\text{DIP joint} &:
0^{\circ} \le \theta_{3} \le 90^{\circ}
\end{aligned}
\label{eq:joint_limit_F}
\end{equation}
\paragraph{Pose calibration}
Inertial sensors such as IMUs suffer from a common problem of drifting, which causes an accumulation of errors during operations. To overcome this issue, we introduce an IMU calibration protocol. When the sensed hand gesture degrades significantly, the user wearing the glove can hold the hand flat and maintain this gesture (see \cref{fig:gesture:calibration}) to initiate calibration; the system records the relative pose between the IMU and world frames. The orientation data measured by the IMUs are multiplied by the inverse of this relative pose to cancel out the differences, thus eliminating accumulated errors due to drifting. This routine can be performed conveniently when experiencing unreliable hand gesture sensing results.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{calibration/calibration}
\caption{\textbf{The IMU calibration protocol.} The protocol starts by holding the hand flat, as shown by the virtual hand model. The relative pose between the world frame and the IMU's local coordinate system is recorded. The inverse of the recorded relative pose corrects the IMU data. Reproduced from Ref.~\cite{liu2019high} with permission.}
\label{fig:gesture:calibration}
\end{figure}
\subsection{IMU evaluation}\label{sec:gesture:imu_eval}
We evaluated an individual IMU's bias and variance during rotations. Furthermore, we examined how accurately two articulated IMUs can reconstruct a static angle, indicating the performance of an atomic element in sensing the finger joint angle.
\paragraph{Evaluations of a single IMU}
As the reliability of the gesture sensing primarily depends on the IMU performance, it is crucial to investigate the IMU's bias and variance. More specifically, we rotated an IMU using a precise stepper motor controlled by an Arduino microcontroller. Four rotation angles--$90^{\circ}$, $180^{\circ}$, $270^{\circ}$, and $360^{\circ}$--were executed 20 times each at a constant angular velocity of 60 rotations per minute (RPMs). We did not test for a rotation angle exceeding $360^{\circ}$, as this is beyond the fingers' motion range. \cref{tab:gesture:imu_std} summarizes the mean and the standard deviation of the measured angular error. Overall, the IMU performed consistently with a bias between $2^{\circ}$ and $3^{\circ}$ and a $\pm 1.7^{\circ}$ standard deviation, suggesting that post-processing could effectively reduce the sensor bias.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=\linewidth,trim=1.05cm 0cm 1.35cm 0.3cm,clip]{imu_eval/imu_error}
\caption{Error in measurement using single IMU.}
\label{tab:gesture:imu_std}
\end{subfigure}%
\\
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=\linewidth,trim=0.86cm 0cm 1.35cm 0.3cm,clip]{imu_eval/imu_ex3}
\caption{Error in recovering a fixed angle with two articulated IMUs.}
\label{tab:gesture:imu_ex3}
\end{subfigure}%
\\
\begin{subfigure}[b]{0.55\linewidth}
\includegraphics[width=\linewidth]{imu_eval/angle}
\caption{Schematic of articulated IMUs.}
\label{tab:gesture:angle}
\end{subfigure}%
\hfill%
\begin{subfigure}[b]{0.442\linewidth}
\includegraphics[width=\linewidth]{imu_eval/angle_print}
\caption{Physical setup.}
\label{tab:gesture:angle_print}
\end{subfigure}%
\caption{\textbf{Evaluations of IMU performance.} The measurement error is summarized as the mean and standard deviation of (a) a single IMU and (b) two articulated IMUs under different settings. The red horizontal lines, blue boxes, and whiskers indicate the median error, the $25^{th}$ and $75^{th}$ percentiles, and the range of data points not considered to be outliers, respectively. A schematic of the experimental setup for evaluating the angle reconstruction with two articulated IMUs is shown in (c), and its physical setup with a $90^{\circ}$ bending angle is shown in (d). Reproduced from Ref.~\cite{liu2017glove} with permission.}
\end{figure}
\paragraph{Evaluations of articulated IMUs}
Evaluating IMU performance on whole-hand gesture sensing is difficult due to the lack of ground truth. As a compromise, we 3D printed four rigid bends with angles of $0^{\circ}$ $45^{\circ}$, $90^{\circ}$, and $135^{\circ}$ to emulate four specific states of finger bending, which evenly divided a finger joint's motion range as defined in \cref{eq:joint_limit_F}. Using two IMUs to construct a bend, assuming it to be a revolute joint, we tested the accuracy of the reconstructed joint angle by computing the relative poses between the two IMUs. \cref{tab:gesture:angle} shows a schematic of this experimental setup, and \cref{tab:gesture:angle_print} shows the physical setup with a $90^{\circ}$ bending angle. During the test, one IMU was placed $2~cm$ behind the bend, and another was placed $1~cm$ ahead, simulating the IMUs attached to a proximal phalanx and a middle phalanx, respectively. We repeated the test 20 times for each rigid bend. \cref{tab:gesture:imu_ex3} shows the errors of the estimated joint angles. As the bending angle increased, the reconstruction errors increased from $4^{\circ}$ to about $6^{\circ}$, with a slightly expanded confidence interval. Overall, the errors were still reasonable, although the IMUs tended to underperform as the bending angle increased. Through combination with the pose calibration protocol, these errors can be better counterbalanced, and the utilized IMU network can reliably support the collection of grasping data (see \cref{sec:application} for various case studies).
\section{Tactile-sensing mode}\label{sec:tactile}
Our reconfigurable data glove can be easily configured to the tactile-sensing mode, which shares the unified backbone design described in \cref{sec:gesture}. The tactile-sensing mode measures the distribution of forces exerted by the hand during complex hand-object interactions. We start by describing the force sensor specifications in \cref{sec:tactile_sensor}, which is followed by details of prototyping in \cref{sec:tactile_prototype}. We conclude this section with a qualitative evaluation in \cref{sec:tactile_eval}.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{0.482\linewidth}
\includegraphics[width=\linewidth]{sensors/velostat}
\caption{Velostat sensor construction}
\label{tab:tactile:velostat}
\end{subfigure}%
\begin{subfigure}[b]{0.51\linewidth}
\includegraphics[width=\linewidth]{sensors/velo_cir}
\caption{Velostat sensor circuit}
\label{tab:tactile:circuit}
\end{subfigure}%
\\
\begin{subfigure}[b]{\linewidth}
\includegraphics[width=\linewidth,trim=0.5cm 0cm 1cm 0cm,clip]{sensors/force_calibration_v2}
\caption{Force-voltage relation of one constructed Velostat sensing taxel}
\label{tab:tactile:calib}
\end{subfigure}%
\\
\begin{subfigure}[b]{0.196\linewidth}
\centering
\includegraphics[width=\linewidth]{sensors/grasp_b}
\caption{}
\label{tab:tactile:grasping}
\end{subfigure}%
\begin{subfigure}[b]{0.8\linewidth}
\centering
\includegraphics[width=\linewidth]{sensors/force_velo}
\caption{}
\label{tab:tactile:force_velo}
\end{subfigure}%
\caption{\textbf{Characterization of the Velostat force sensor.} (a) The multi-layer structure of a Velostat force sensor. (b) The circuit layout for force data acquisition. (c) The force-voltage relation of one sensing taxel. Instead of using a power law, our choice of a logarithmic law fits the data better. (d) A grasp of the half-full bottle. (e) Force responses of grasping empty, half-full, and full bottles, respectively. Reproduced from Ref.~\cite{liu2017glove} with permission.}
\end{figure}
\subsection{Force sensor}\label{sec:tactile_sensor}
We adopt a network of force sensors made from Velostat to provide force sensing in this tactile-sensing mode. \cref{tab:tactile:velostat} illustrates the Velostat force sensor's multi-layer structure. A taxel (\ie, a single-point force-sensing unit) is composed of one inner layer of Velostat ($2~cm \times 2~cm$) and two middle layers of conductive fabric, stitched together by conductive thread and enclosed by two outer layers of insulated fabric. A force-sensor pad consisting of 2 taxels is placed on each finger, and a sensor grid with $4~cm \times 4~cm$ taxels is placed on the palm. Lead wires to the pads and grid are braided into the conductive thread.
As the Velostat's resistance changes with different pressing forces, the measured voltage across a taxel can be regarded as the force reading at that region. To acquire the voltage readings, we connect these Velostat force-sensing taxels in parallel via analog multiplexers controlled by the RPi's GPIO and output to its SPI-enabled ADS1256 ADC. More specifically, two 74HC4051 multiplexers are used for the palm grid, and a CD74HC4067 multiplexer is used for all the finger pads. A voltage divider circuit, shown in \cref{tab:tactile:circuit}, is constructed by connecting a $200\Omega$ resistor between the RPi's ADC input channel and the multiplexers.
We now characterize the sensor's force-voltage relation~\cite{lee2016feasibility}. A total of 13 standard weights ($0.1~kg$ to $1.0~kg$ with $0.1~kg$ increments, $1.2~kg$, $1.5~kg$, and $2.0~kg$) were applied to a taxel, and the associated voltages across that taxel were measured. The calibration circuit was the same as that in \cref{tab:tactile:circuit}, except that only the taxel of interest was connected. The weights in kilograms were converted to forces in Newtons with a gravitational acceleration $g=10~m/s^2$. We first tested the power law~\cite{lee2016feasibility} for characterizing the force-voltage relation of a taxel. The result was $F = -1.067V^{-0.4798}+3.244$ with $R^2 = 0.9704$, where $F$ is the applied force, and $V$ is the output voltage. However, we further tested a logarithmic law, resulting in a better force-voltage relation: $F = 0.569\log{(44.98V)}$ with a higher $R^2 = 0.9902$. Hence, we adopted the logarithmic fit to establish a correspondence between the voltage reading across a taxel and the force the taxel is subjected to. \cref{tab:tactile:calib} compares these two fits.
\subsection{Prototyping}\label{sec:tactile_prototype}
\cref{fig:summary}a~\cite{liu2017glove,liu2019high,li2020incremental} displays a prototype of the tactile-sensing glove. The capability of force sensing is accomplished by placing one Velostat force-sensing pad on each finger (one taxel in the proximal area and another in the distal area) and a single $4~cm \times 4~cm$ Velostat force-sensing grid over the glove's palm region. Based on the established force-voltage relation, these taxels collectively measure the distribution of forces exerted by the hand. Meanwhile, the 15 IMUs capture the hand gestures in motion. These components are all connected to the RPi, which can be remotely accessed to visualize and subsequently utilize the collected gesture and force data in a local workstation, providing a neat solution to collect human manipulation data.
By measuring the voltage and current across each component, we investigated the power consumption of the prototype. \cref{tab:tactile:power} reports the peak power of each component of interest as the product of its voltage and current in a 10-min operation. The total power consumption was $2.72~W$, which can be easily powered by a conventional Li-Po battery, offering an untethered user experience and natural interactions during data collection.
\begin{table}[htb!]
\centering
\small
\setlength{\tabcolsep}{4pt}
\caption{\textbf{Power consumption of the tactile-sensing glove.}}
\label{tab:tactile:power}
\begin{tabular}{ccccc}
\toprule
\multirow{2}{*}{\textbf{Component}} & gesture sensing & force sensing & computing & \multirow{2}{*}{total} \\
& $15$ IMUs & $6$ Velostat & RPI & \\
\midrule
\textbf{Power (W)} & 0.60 & 0.02 & 2.15 & 2.72 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Qualitative evaluation}\label{sec:tactile_eval}
We evaluated the performance of the tactile-sensing glove in differentiating among low, medium, and high forces by grasping a water bottle in three states, \textit{empty}, \textit{half-full}, and \textit{full}, whose weights were $0.13~kg$, $0.46~kg$, and $0.75~kg$, respectively. The participants were asked to perform the grasps naturally and succinctly--exerting a force just enough to prevent the bottle from slipping out of the hand; \cref{tab:tactile:grasping} shows such an instance. Ten grasps were performed for each bottle state. To simplify the analysis, the force in the palm was the average of all 16 force readings of the palm grid, and the force in each finger was the average reading of the corresponding finger pads. \cref{tab:tactile:force_velo} shows the recorded forces exerted by different hand regions.
\section{VR mode}\label{sec:vr}
Since the different modes of our data glove share a unified backbone design, reconfiguring the glove to the VR mode in order to obtain contact points during interactions can be achieved with only three steps. First, given the sensed hand gestures obtained by the shared backbone, we need to construct a virtual hand model for interactions (see \cref{sec:vr_hand_model}). Next, we must develop an approach to achieve a stable grasp of virtual objects (see \cref{sec:vr_grasp}). Finally, grasping objects in VR introduces new difficulty without a tangible object being physically manipulated; we leverage haptic feedback to address this problem in \cref{sec:vr_haptic}. We conclude this section with an evaluation in \cref{sec:vr_eval}.
\subsection{Virtual hand model}\label{sec:vr_hand_model}
Generating a stable grasp is the prerequisite for obtaining contact points during interactions. Existing vision-based hand gesture sensing solutions, including commercial projects such as LeapMotion~\cite{leapmotion} and RealSense~\cite{realsense}, struggle with stable grasps due to occlusions, sensor noises, and a limited field of view (FoV); interested readers can refer to \cref{fig:vr:comparison_vision} for a comparison in a typical scenario. In comparison, existing VR controllers adopt an alternative approach--the virtual objects are directly attached to the virtual hand when a grasp event is triggered. As illustrated in \cref{fig:vr:comparison_touch}, the resulting experience has minimal realism and cannot reflect the actual contact configuration. The above limitations motivate us to realize a stable virtual grasp by developing a caging-based approach that is capable of real-time computation while offering sufficient realism; an example is provided in \cref{fig:vr:comparison_glove}.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{0.333\linewidth}
\includegraphics[width=\linewidth]{vr_glove_comparison/leapmotion_blur}
\caption{LeapMotion}
\label{fig:vr:comparison_vision}
\end{subfigure}%
\begin{subfigure}[b]{0.333\linewidth}
\includegraphics[width=\linewidth]{vr_glove_comparison/touch}
\caption{Oculus Touch}
\label{fig:vr:comparison_touch}
\end{subfigure}%
\begin{subfigure}[b]{0.333\linewidth}
\includegraphics[width=\linewidth]{vr_glove_comparison/glove}
\caption{Ours}
\label{fig:vr:comparison_glove}
\end{subfigure}%
\caption{\textbf{A comparison of grasp among (a) a LeapMotion sensor, (b) an Oculus Touch controller, and (c) our reconfigurable glove system in the VR mode.} The grasp in (a) is unstable, as reflected by the motion blur, due to occlusion in the vision-based hand gesture sensing approach. While (b) affords a form of ``stable'' grasp (\ie, it removes the gravity from the cup) by directly attaching the object to the hand, this approach is unnatural, with minimal realism. It does not reflect the actual contact between a hand and an object, and sometimes the hand even fails to come into contact with the object. (c) The proposed reconfigurable glove in VR mode offers a realistic and stable grasp, which is crucial for obtaining contact points during interactions. Reproduced from Ref.~\cite{liu2019high} with permission.}
\label{fig:vr:comparison}
\end{figure}
Thanks to the reconfigurable nature of the glove, creating a virtual hand model in VR is simply the reiteration of the hand gesture-sensing module described in \cref{sec:gesture}; \cref{fig:vr:hand} shows the structure of the virtual hand. More specifically, the hand gestures in the local frames are given by the IMUs, and a Vive tracker with HTC Lighthouse provides the precise positioning of the hand in a global coordinate, computed by the time-difference-of-arrival.
\begin{figure}[t!]
\centering
\includegraphics[width=0.8\linewidth]{handconstruction/handconstruction}
\caption{\textbf{Structure of the virtual hand model}. Each phalanx is modeled by a small cylinder whose dimensions are measured by a participant. The pose of each phalanx is reconstructed from the data read by the IMUs. The Vive tracker provides direct tracking of the hand pose. Reproduced from Ref.~\cite{liu2019high} with permission.}
\label{fig:vr:hand}
\end{figure}
\subsection{Stable grasps}\label{sec:vr_grasp}
Methods for realizing a virtual grasp in VR can be roughly categorized into two streams, with their unique pros and cons. One approach is to use a physics-based simulation with collision detection to support realistic manipulations by simulating the contact between a soft hand and a virtual object made from varied materials. Despite its high fidelity, this approach often demands a significant amount of computation, making it difficult--if not impossible--to use in real time. Alternatively, symbolic-based and rule-based grasps are popular approaches. A grasp or release is triggered based on a set of predefined rules when specific conditions are satisfied. This approach is computationally efficient but provides minimal realism.
Our configurable glove-based system must balance the above two factors to obtain contact points during interactions. It must provide a more natural interaction than those of rule-based methods, such that the contact points obtained on the objects are relatively accurate, while ensuring more effective computation than high-fidelity physics-based simulations, such that it can be achieved in real time.
In this work, we devise a caging-based stable grasp algorithm, which can be summarized as follows. First, the algorithm detects all collisions between the hands and objects \eg, the red areas in \cref{fig:vr:collision}b). Next, the algorithm computes the geometric center of all collision points between the hands and objects and checks whether this center is within the object. Supposing that the above situation holds (see \cref{fig:vr:collision}a), we consider this object to be ``caged''; thus, it can be stably grasped. The objects' physical properties are turned off, allowing them to move along with the hand. Otherwise, only standard collisions are triggered between the hand and object. Finally, the grasped object is released when the collision event ends or the geometric center of the collisions is outside the object. This process ensures that a grasp only starts after a caging is formed, offering a more natural manipulation experience with higher realism than rule-based grasps.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{collision/collision}
\begin{tabular}{c}
\hspace{0.2cm}(a) \hspace{3.8cm}(b)
\end{tabular}
\caption{\textbf{Detect stable grasps based on collisions.} (a) When the geometric center (green dashed circle) of all the collision points (red balls) overlaps with the object (yellow cylinder), the object is considered to be stably grasped and will move along with the hand. (b) Various stable grasps of a small green cylinder. Reproduced from Ref.~\cite{liu2019high} with permission.}
\label{fig:vr:collision}
\end{figure}
\subsection{Haptic feedback}\label{sec:vr_haptic}
By default, the participants have no way to feel whether or not their \textit{virtual} hands are in contact with the \textit{virtual} objects while operating the glove in VR mode due to the lack of haptic feedback, which prevents they from manipulating objects naturally. To fill this gap, the VR mode implements a network of shaftless vibration motors that are triggered when the corresponding virtual phalanxes collide with the virtual object; this offers an effective means of providing each finger with vibrational haptic feedback in the \textbf{physical} world that corresponds to the contact feedback that the participants should receive in VR. Connected to a 74HC4051 analog multiplexer and controlled by the RPi's GPIO, these small ($10~mm \times 2~mm$) and lightweight ($0.8~g$) vibration motors provide $14,500$ RPM with a $3~V$ input voltage. Once a finger touches the virtual object, the vibration motors located at that region of the glove are activated to provide continuous feedback. When the hand forms a stable grasp, all motors are powered up, so that the user can maintain the current hand gesture to hold the object.
\subsection{Qualitative evaluation}\label{sec:vr_eval}
\begin{figure}[t!]
\centering
\begin{overpic}
[width=\linewidth]{grasp_result/grasp_result}
\put(8,-3){\color{black} (a)}
\put(24,-3){\color{black} (b)}
\put(42,-3){\color{black} (c)}
\put(58,-3){\color{black} (d)}
\end{overpic}
\vspace{12pt}
\caption{\textbf{Various grasp results for four virtual objects:} (a) a mug, (b) a tennis racket, (c) a bowl, and (d) a goose toy. The top and bottom rows show the approach and release of the target objects, respectively. Reproduced from Ref.~\cite{liu2019high} with permission.}
\label{fig:vr:grasp_result}
\end{figure}
We conducted a case study wherein the participants were asked to wear the VR glove and grasp four virtual objects with different shapes and functions, including a mug, a tennis racket, a bowl, and a goose toy (see \cref{fig:vr:grasp_result}). These four objects were selected because (i) they are everyday objects with a large variation in their geometry, providing a more comprehensive assessment of the virtual grasp; and (ii) each of the four objects can be grasped in different manners based on their functions, covering more grasp types~\cite{feix2016grasp,liu2021synthesizing}. We started by testing different ways of interacting with virtual objects, such as grasping a mug by either the handle or the rim. Such diverse interactions afforded a natural experience by integrating unconstrained fine-grained gestures, which is difficult for existing platforms (\eg, LeapMotion). In comparison, our reconfigurable glove in VR mode successfully balanced the naturalness of the interactions with the stability of the grasp, providing a better realism in VR, which was close to how objects are manipulated in the physical world.
Notably, the reconfigurable glove in VR mode was able to track hand gestures and maintain a stable grasp even when the hand was outside the participant's FoV, thus offering a significant advantage compared with vision-based approaches (\eg, the LeapMotion sensor). In a comparative study in which the participant's hand could be outside of the FoV, the performance using the VR glove significantly surpassed that of LeapMotion (see \cref{tab:vr:result}), thereby demonstrating the efficacy of the VR glove hardware, the caging-based grasp approach, and the haptic feedback.
\begin{table}[htb!]
\centering
\small
\caption{\textbf{Success rates of grasping and moving four different objects using the VR glove (G) and the LeapMotion sensor (L)}.}
\label{tab:vr:result}
\begin{tabular}{c c c c c c}
\toprule
\textbf{task} & \textbf{setup} & \textbf{mug} & \textbf{racket} & \textbf{mug} & \textbf{racket}\\
\midrule
\multirow{2}{*}{grasp} & L & 80\% & 13\% & 27\% & 67\% \\
& G & 100\% & 100\% & 100\% & 93\% \\
\multirow{2}{*}{move} & L & 33\% & 7\% & 0\% & 47\% \\
& G & 100\% & 93\% & 93\% & 87\% \\
\bottomrule
\end{tabular}
\end{table}
\section{Simulation mode}\label{sec:sim}
A manipulation event consists of both hand information and object information. Most prior work has focused on the former without paying much attention to the latter. In fact, objects may be occluded or may even change significantly in shape as a result of a manipulation event, such as through deformation or cracking. Such information is essential in understanding the manipulation event, as it reflects the goals. However, existing solutions, even those with specialized sensors, fall short in handling this scenario, so a solution beyond the conventional scope of data gloves is called for.
To tackle this challenge, we integrate a state-of-the-art FEM simulator~\cite{li2020incremental} to reconstruct the physical effects of an object, in numeric terms, during the manipulation. Given the trajectory data obtained by the proposed glove-based system, both physical and virtual properties and how they evolve over time are simulated and rendered, providing a new dimension for understanding complex manipulation events.
\subsection{Simulation method}
We start with a brief background of solid simulation. Solid simulation is often conducted with FEM~\cite{zienkiewicz2000finite}, which discretizes each object into small elements with a discrete set of sample points as the DoFs. Then, mass and momentum conservation equations are discretized on the mesh and integrated over time to capture the dynamics, in which \textit{elasticity} and \textit{contact} are the most essential yet most challenging components. \textit{Elasticity} is the ability of an object to retain its rest shape under external impulses or forces, whereas \textit{contact} describes the intersection-free constraints on an object's motion trajectory. However, elasticity is nonlinear and non-convex, and contact is non-smooth, both of which can pose significant difficulties to traditional solid simulators based on numerical methods~\cite{li2020robust}. Recently, Li \textit{et al}\onedot~\cite{li2020incremental} proposed incremental potential contact (IPC), a robust and accurate contact-handling method for FEM simulations~\cite{li2020codimensional,fang2021guaranteed,ferguson2021intersection,lan2021medial,choo2021barrier}; it formulates the non-smooth contact condition into smooth approximate barrier potentials so that the non-smooth contact condition can be solved simultaneously with electrodynamics using a line search method~\cite{li2019decomposed,wang2020hierarchical,nocedal2006numerical} with a global convergence guarantee. As it is able to consistently produce high-quality results without numerical instability issues, IPC makes it possible to conveniently simulate complex manipulation events, even with extremely large deformations.
We further extend the original IPC to support object fracture by measuring the displacement
of \textit{every} pair of points; that is, we go through all pairs of points for a triangle and all triangles on the mesh. If the displacement relative to the pair of points' original distance exceeds a certain strain threshold (in this work, we set it to $1.1$), we mark the triangle in between as separated. At the end of every time step, we reconstruct the mesh topology using a graph-based approach~\cite{hegemann2013level}, according to the tetrahedra face separation information. Due to the existence of the IPC barrier, which only allows a positive distance between surface primitives, it is essential to ensure that, after the topology change, the split faces do not exactly overlap. Therefore, we perturb the duplicate nodes on the split faces by a tiny displacement toward the normal direction, which works nicely even when edge-edge contact pairs are ignored for simplicity.
\begin{figure*}[t!]
\centering
\begin{subfigure}[b]{0.25\linewidth}
\includegraphics[width=\linewidth]{simulation_result/raw_materials/uncrack/slow_motion}
\caption{uncracked}
\end{subfigure}%
\begin{subfigure}[b]{0.25\linewidth}
\includegraphics[width=\linewidth]{simulation_result/raw_materials/crack/slow_motion}
\caption{cracked}
\end{subfigure}%
\begin{subfigure}[b]{0.25\linewidth}
\includegraphics[width=\linewidth]{simulation_result/raw_materials/smash/slow_motion}
\caption{smashed}
\end{subfigure}%
\begin{subfigure}[b]{0.25\linewidth}
\includegraphics[width=\linewidth]{simulation_result/raw_materials/cut/slow_motion}
\caption{cut in half}
\end{subfigure}%
\caption{\textbf{Four types of tool-use events captured by a slow motion camera at 120 FPS.}}
\label{fig:sim:slow_motion}
\end{figure*}
\begin{figure*}[b!]
\centering
\begin{subfigure}[b]{0.33\linewidth}
\includegraphics[width=\linewidth]{tactile_result/69}
\caption{\textit{Bottle~1}, no childproof lock}
\label{fig:69}
\end{subfigure}%
\hfill%
\begin{subfigure}[b]{0.33\linewidth}
\includegraphics[width=\linewidth]{tactile_result/64}
\caption{\textit{Bottle~2}, pressing down the lid to unlock}
\label{fig:64}
\end{subfigure}%
\hfill%
\begin{subfigure}[b]{0.33\linewidth}
\includegraphics[width=\linewidth]{tactile_result/68}
\caption{\textit{Bottle~3}, pinching the lid to unlock}
\label{fig:68}
\end{subfigure}%
\caption{\textbf{Visualizations of the hand gesture and force of opening three bottles collected using the tactile-sensing glove.} These visualizations reveal the subtle differences between the actions of opening medicine bottles and opening a conventional bottle; the essence of this task is that visual information alone is insufficient to distinguish between the opening of the various bottles. Reproduced from Ref.~\cite{liu2017glove} with permission.}
\label{fig:app:tactile:vis}
\end{figure*}
\subsection{Prototyping and input data collection}
The simulation-augmented glove-based system is essentially the same as the VR glove, except for the lack of vibration motors; however, it is augmented with the simulated force evolved over time. Compared with the aforementioned two hardware-focused designs, the simulation-augmented glove-based system offers an in-depth prediction of physics with fine-grained object dynamics—that is, how the geometry (\eg, large deformation) and topology (\eg, fracture) evolve. To showcase the efficacy of this system, we focus on a tool-use setting wherein a user manipulates a tool (\eg, a hammer) to apply on a target object (\eg, a nut), causing geometry and/or topology changes. To collect one set of data, the hand gestures and poses are reconstructed similarly using the other two glove-base systems. The tool's movement is further tracked to simulate the interactions between the tool and the object.
More specifically, two Vive trackers track the movements of the glove-based system (\ie, the hand) and the tool, respectively. The third tracker, which serves as the reference point for the target object (\eg, a nut) is fixed to the table. All three Vive trackers are calibrated such that their relative poses and the captured trajectories can be expressed in the same coordinate. The target objects and the tool's meshes are scanned beforehand using a depth camera. By combining the scanned meshes and captured trajectories, we can fully reconstruct a sequence of 3D meshes representing the movements of the hand and tool and simulate the resulting physical effects of the target object. The captured mesh sequences are directly input to the simulation as boundary conditions, and the DoFs being simulated are primarily those on the target object. \cref{fig:sim:slow_motion} shows some keyframes of the data collection for cracking walnuts and cutting carrots. It should be noted that capturing how the object changes and its physical properties over time is extremely challenging--if not impossible--using visual information alone.
\subsection{Simulation setup}
An object's material properties in a simulation are mainly reflected by its stiffness (\ie, the object is more difficult to deform or fracture if it is stiffer), governed by its Young's modulus and Poisson's ratio. These parameters must be set appropriately in the simulation in order to produce effects that match those in the physical world. The Young's modulus and Poisson's ratio of a material can be found in related works~\cite{bourne2002food,williams2005mechanical,kiani2011determination}. Another parameter that must be set is the fracturing strain threshold, which determines the dimension of the segments when fracturing is triggered. This parameter is tuned so that the simulator can reproduce the type of effects observed in the physical world. The time step of the simulation is the inversion of the sampling frequency of the Vive trackers that acquire the trajectories.
\section{Applications}\label{sec:application}
In this section, we showcase a series of applications by reconfiguring the data glove to the tactile-sensing mode (\cref{sec:app:tactile}), VR mode (\cref{sec:app:vr}), and simulation mode (\cref{sec:app:sim}), all of which share the same backbone design. (Interested readers can also refer to the Appendix for video demonstrations.)
\subsection{Tactile-sensing mode}\label{sec:app:tactile}
We evaluated the tactile-sensing mode by capturing the manipulation data of opening three types of medicine bottles. Two of these bottles are equipped with different locking mechanisms and require a series of specific action sequences to remove the lid. More specifically, \textit{Bottle 1} does not have a safety lock, and simply twisting the lid is sufficient to open it. The lid of \textit{Bottle 2} must be pressed simultaneously while twisting it. \textit{Bottle 3} has a safety lock in its lid, which requires a pinching action before twisting to unlock it. Notably, the pressing and pinching actions required to open \textit{Bottle 2} and \textit{Bottle 3} are challenging to recognize without using the force information recorded by the glove.
\cref{fig:app:tactile:vis} shows examples of the recorded data with both hand gesture and force information. The first row of \cref{fig:app:tactile:vis} visualizes the captured manipulation action sequences of opening these three bottles. The second row shows the corresponding action sequences captured by an RGB camera for reference.
Qualitatively, compared with the action sequences shown in the second row, the visualization results in the first row differentiate the fine manipulation actions with additional force information. For example, the fingers in \cref{fig:64} are flat and parallel to the bottle lid, whereas those in \cref{fig:68} are similar to those in the gripping pose. The responses of the force markers are also different due to varying contact points between the human hand and the lid: The high responses in \cref{fig:64} are concentrated on the palm area, whereas only two evident responses on the distal thumb and index finger can be seen in \cref{fig:68}. Taken together, these results demonstrate the significance of accounting for forces when understanding fine manipulation actions.
Quantitatively, \cref{fig:app:tactile:compare} illustrates one taxel's force collected on the palm, the thumb's fingertip, and the flexion angle of the index finger's MCP joint. In combination, these three readings can differentiate among the action sequences of opening the three bottles. More specifically, as opening \textit{Bottle 2} involves a pressing action on the lid, the tactile glove successfully captures the high force response on the palm. In contrast, the force reading in the same region is almost zero when opening the other two bottles. \textit{Bottle 3}'s pinch-to-open lock necessitates a greater force exerted by the thumb. Indeed, the opening actions introduce a high force response at the thumb's fingertip, with a longer duration than the actions involved in opening \textit{Bottle 1} without a safety lock. Without contacting the lid, the thumb yields no force response when opening \textit{Bottle 2}. Since opening both \textit{Bottle 1} and \textit{Bottle 3} involves a similar twist action, the measured flexion angles of the index finger's MCP joint are around $50^{\circ}$ in both of these cases. Since only the palm touches the lid and the fingers remain stretched, a small flexion angle occurs when opening \textit{Bottle 2}.
\begin{figure}[t!]
\centering
\begin{overpic}
[width=.8\linewidth,trim=0cm 0cm 0cm 0.3cm,clip]{force_compare/force_compare_v2}
\put(37,82.5){\color{black} (a)}
\put(27,54.5){\color{black} (b)}
\put(17.5,25.5){\color{black} (c)}
\end{overpic}
\caption{\textbf{Force and joint angle recorded by the tactile-sensing glove.} (a) The forces exerted by the palm, (b) forces exerted by the thumb's fingertip, and (c) the flexion angle of the index finger's MCP joint can disentangle the grasp actions of opening different bottles. Reproduced from Ref.~\cite{liu2017glove} with permission.}
\label{fig:app:tactile:compare}
\end{figure}
A promising application of the proposed glove is learning fine manipulation actions from human demonstrations. The collected tactile data has facilitated investigations into a robot's functional understanding of actions and imitation learning~\cite{edmonds2017seeing,liu2019mirroring}, inverse reinforcement learning~\cite{xie2019vrgrasp}, and learning explainable models that promote human trust~\cite{edmonds2019tale}. \cref{fig:app:tactile:robot_exec} showcases the robot's learned skills of opening different medicine bottles~\cite{edmonds2017seeing}.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{\linewidth}
\includegraphics[width=\linewidth]{robot_open_bottle/69_exe}
\caption{}
\label{fig:app:tactile:69_exe}
\end{subfigure}%
\\
\begin{subfigure}[b]{\linewidth}
\includegraphics[width=\linewidth]{robot_open_bottle/64_exe}
\caption{}
\label{fig:app:tactile:64_exe}
\end{subfigure
\caption{\textbf{A Baxter robot learns to open medicine bottles from the collected manipulation data.} Reproduced from Ref.~\cite{edmonds2017seeing} with permission.}
\label{fig:app:tactile:robot_exec}
\end{figure}
\subsection{VR mode}\label{sec:app:vr}
When operating in VR mode, the reconfigurable glove provides a unique advantage compared with traditional hardware. Below, we showcase two data types that can be collected effectively in this mode.
\paragraph{Trajectories}
\begin{figure}[t!]
\begin{overpic}
[width=\linewidth]{pose_trajectory/trajectory_v3}
\put(23,63.5){\color{black} (a)}
\put(73,63.5){\color{black} (b)}
\put(23,29){\color{black} (c)}
\put(73,29){\color{black} (d)}
\end{overpic}
\caption{\textbf{Examples of hand and object trajectories collected by the reconfigurable glove operating in VR mode.} Red triangles indicate the starting poses. The red line and the blue lines show the recorded hand movement and the trajectories of the fingertips, respectively. Once the contact points (green circles) are sufficient to trigger a stable grasp, the object moves together with the hand, following the black line, until the grasp becomes unstable--that is, until it is released at the orange circles. Reproduced from Ref.~\cite{liu2019high} with permission.}
\label{fig:app:vr:traj}
\end{figure}
Hand and object trajectories are particularly useful in robot learning from demonstration. Diverse object models can be placed in the VR without setting up a physical apparatus to ensure a natural hand trajectory. \cref{fig:app:vr:traj} shows some qualitative results of collected trajectories: the hand movement (red line) and the five fingertips' trajectories (blue lines) by combining global hand pose and hand gesture sensing, and the grasped object's movement (black line) as the result of hand movement and grasp configuration (stable grasp or not). These results demonstrate the reliability of our design and the richness of the collected trajectory information in a manipulation event.
\paragraph{Contact points}
It is extremely challenging to obtain the contact points of the objects being manipulated. Despite relying As they rely heavily on training data, computer vision-based methods~\cite{rautaray2015vision} are still vulnerable to handling occlusion between hands and objects. Our reconfigurable glove operating in the VR mode can elegantly log this type of data. Given the meshes of the virtual hand model and the object, the VR's physics engine can effectively check the collisions between them. These collisions not only determine whether the object can be stably grasped based on the criteria described in \cref{sec:vr_grasp} but also correspond well to the contact points on the grasped object. By treating a collision point as the spatial center of a spherical volume whose radius is set to the diameter of the finger, \cref{fig:app:vr:contact} shows three configurations of contacts collected from different participants grasping diverse objects. To better uncover the general grasp habits for an object, the contact points shown in the bottom row of \cref{fig:app:vr:contact} are obtained by averaging the spatial positions of contacts across different trails, fitted by a Gaussian distribution.
\begin{figure}[t!]
\centering
\hfill
\begin{overpic}
[width=.96\linewidth]{object_heatmap/HeatMap_v6}
\put(-5,63){\color{black} (a)}
\put(-5,51){\color{black} (b)}
\put(-5,38){\color{black} (c)}
\put(-5,26){\color{black} (d)}
\put(26.7,17.1){\color{black} (e)}
\put(-5,5.5){\color{black} (f)}
\end{overpic}
\caption{\textbf{Contact points in grasping various objects.} (a) Objects to be grasped. (b–d) Three configurations of the contact points performed by different participants. (e) The distance from each contact point. (f) The average of contact points aggregated from all participants, indicating the preferred regions of contact, given the objects.}
\label{fig:app:vr:contact}
\end{figure}
A fundamental challenge in robot learning of manipulation is the embodiment problem~\cite{dautenhahn2002imitation,liu2019mirroring}: The human hand (five fingers) and robot gripper (usually two or three fingers) have different morphologies. While this problem demands further research, individual contact points can also indicate a preferred region of contact if aggregated from different participants (see the last row in \cref{fig:app:vr:contact}). Such aggregated data can be used for training robot manipulation policies despite different morphologies~\cite{liu2019mirroring}.
\begin{figure*}[t!]
\centering
\begin{subfigure}[b]{\linewidth}
\includegraphics[width=\linewidth]{simulation_result/rendering_v2}
\caption{\textbf{Reconstructed tool-use events by simulation.} The first/third rows show the contact moments between the tool and the object. The second/fourth rows are the corresponding stress given by the simulator; red indicates greater stress. The fifth row shows the objects' final status.}
\label{fig:app:sim:rendering}
\end{subfigure}%
\\
\begin{subfigure}[b]{\linewidth}
\includegraphics[width=\linewidth]{simulation_result/physics_data_v2}
\caption{The energy imposed on the objects, the number of fractured pieces, and the contact pressure calculated by the simulator at each time step during tool use.}
\label{fig:app:sim:data}
\end{subfigure}%
\caption{\textbf{Reconstructed 4D manipulation events of tool use by integrating trajectories collected by the reconfigurable glove and physics-based simulation.} This high-fidelity 4D data reveals fine-grained object fluent changes and physical properties at each time step. The results are produced with a simulation at $20~Hz$; one time step is $0.05~s$.}
\label{fig:app:sim:results}
\end{figure*}
\subsection{Simulation mode}\label{sec:app:sim}
By incorporating the state-of-the-art physics-based simulation, we empower the data glove to capture fine-grained object dynamics during manipulations. \cref{fig:app:sim:results} showcases simulated objects' fluent changes in tool uses. Even when recorded at 120 fps, it is challenging--if not impossible--to capture an object's fluent changes (\eg, how a walnut smashes) using a vision-based method. By feeding the collected trajectory into the simulation, our system renders object fluent changes that are visually similar to the physical reality (see \cref{fig:app:sim:rendering}), thereby revealing critical physical information (see \cref{fig:app:sim:data}) on what occurs in the process.
\paragraph{Results}
\cref{fig:app:sim:rendering} depicts various processes of hammering a walnut. The first column illustrates that a gentle swing action only introduces a small force/energy to the walnut, resulting in a light stress distribution that is quickly eliminated; as a result, the walnut remains uncracked. When a strong swing is performed (third column in \cref{fig:app:sim:rendering}), the larger internal stress causes the walnut to fracture into many pieces, similar to a smashing event in the physical world. This difference is reflected in \cref{fig:app:sim:data}, which was obtained using the physics-based simulator. It is notable that these physical quantities are challenging to measure in the physical world, even with specialized equipment.
\paragraph{Failure examples}
The fourth column of \cref{fig:app:sim:rendering} shows an example of cutting a carrot. The imposed stress is concentrated along the blade that splits the carrot in half. However, when the cutting action is completed and the knife is lifted, it can be seen that the collision between the blade and the carrot has caused undesired fracturing around the cut, which illustrates the limit of the current simulator.
\section{Discussions}
We now discuss two topics in greater depth: Are simulated results good enough, and how do the simulated results help?
\subsection{Are simulated results good enough?}
A central question regarding simulations is whether the simulated results are helpful, given that they are not numerically identical to those directly measured in the physical world. We argue that simulators are indeed helpful, as a simulation preserves the physical events qualitatively, making it possible to study complex events. As illustrated in \cref{fig:app:sim:data}, the walnut's effects have a clear correspondence to the pressure imposed on the contact. Conversely, although a similar amount of energy is imposed when cracking the walnut with a hammer and cutting the carrot with a knife (see the second and fourth columns of hammer and cutting the carrot with a knife (see the second and fourth columns of \cref{fig:app:sim:results}), the resulting pressures differ in magnitude, as the knife introduces a much smaller contact area than the hammer does, producing distinct deformations and topology changes. Hence, the simulation provides a qualitative measurement of the physical events and the objects' fluent change rather than precise quantities. Similar arguments are found in the intuitive physics literature in psychology: Humans usually only make approximate predictions about how states evolve, sometimes even with violations of actual physical laws~\cite{kubricht2017intuitive}. Such inaccuracy does not prevent humans from possessing an effective object and scene understanding; on the contrary, it is a core component of human commonsense knowledge~\cite{spelke2022babies,spelke2007core,zhu2020dark}. Recent work in robot tool use~\cite{zhang2022understanding,li2022gendexgrasp,zhu2015understanding} and physics-informed scene understanding~\cite{han2022scene,han2021reconstructing,chen2019holistic++,huang2018cooperative,huang2018holistic,li2017earthquake,zhu2016inferring,zheng2015scene,zheng2013beyond} has also demonstrated the essential role of physics in understanding objects and scenes.
\subsection{How do the simulated results help?}
The fine-grained object effects produced by the simulation open up new venues for studying existing AI and robotics problems. For example, combining task planning and motion planning~\cite{jiao2021efficieint,jiao2021consolidating,jiao2022planning} is a grand challenge in the field of planning. Simulation could help with this challenge in two aspects~\cite{zhang2022understanding}: (i) by grounding ambiguous task symbols to desired outcomes (\eg, the action symbol of ``crack''), and (ii) by modeling implicit goal specifications (\eg, the status of ``cracked''). In addition, simulations can be used to augment existing datasets, such as GARB~\cite{taheri2020grab} and GenDexGrasp~\cite{li2022gendexgrasp} in grasping, and HUMANISE~\cite{wang2022humanise}, CHAIRS~\cite{jiang2022chairs}, and LEMMA~\cite{jia2020lemma} in scene understanding with unobservable information. Ultimately, we hope that this type of 4D data empowered by physics-based simulation can shed light on several profound questions in manipulation: What and why an object is chosen (\ie, the physics involved), how to properly operate that object (\ie, its affordance), what effect the actor is trying to achieve (\ie, the actor's task goals), and what happens when the goal is not achieved (\ie, planning and replanning).
\section{Conclusion}\label{sec:conclusion}
In this study, we presented three different configurations of a glove-based system based on a unified backbone design, which differs from most conventional data gloves that only capture hand gestures. Utilizing piezoresistive Velostat material, the glove's tactile-sensing mode can aggregate the hand force information during manipulation events. In VR mode, the sensed hand gestures can be reconstructed into a virtual hand to facilitate hand-object interactions in VR by incorporating a caging-based approach, resulting in stable grasps and providing vibrational haptic feedback. The simulation mode further uses an FEM simulator to produce fine-grained object fluent changes and physical properties based on hand-related movements, resulting in 4D manipulation events.
We evaluated the components of the system, including the IMUs, Velostat force-sensor taxels, and haptic feedback provided by the vibration motors, to demonstrate the capability and efficacy of the proposed design. By (i) capturing spatiotemporal signals of force and gesture, (ii) recording hand trajectories and contact points on objects, and (iii) collecting 4D manipulations in challenging manipulation events (\eg, tool use), we demonstrated that the proposed glove-based system can play a crucial role in robot learning from humans and in facilitating embodied AI-related research.
\section{Acknowledgments}
The authors would like to thank Mr. Matt Millar and Dr. Xu Xie (Meta) for developing earlier versions of the system, Miss Chen Zhen (BIGAI) for making the nice figures, and five anonymous reviews for constructive feedback. This work is supported in part by the National Key R\&D Program of China (2021ZD0150200) and the Beijing Nova Program.
\section{Appendix}
\paragraph{Supplementary data}
Code and video demos are available at \url{https://sites.google.com/view/engr-glove}.
\paragraph{Compliance with ethics guidelines}
Hangxin Liu, Zeyu Zhang, Ziyuan Jiao, Zhenliang Zhang, Minchen Li, Chenfanfu Jiang, Yixin Zhu, and Song-Chun Zhu declare that they have no conflict of interest or financial conflicts to disclose.
\newpage
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\label{s:intro} \noindent
Materials with high strength and high toughness are the idea materials for many industry applications. People are consistently seeking high strength and high toughness materials \citep{sun_highly_2012, gong_double-network_2003}. High strength implies high Young's modulus and yielding stress. High toughness means materials can absorb a lot of mechanical energy before fracture. Fracture of materials corresponds to crack formation in a piece of structure, that is, one piece becomes two pieces. Before occurs of material fracture, a plastic cohesive zone will usually form \citep{ahmadi_experimental_2022}. When plastic deformation become localized, shear bands or neckings will be formed. These shear bands and neckings will be generated at the weakest point in the materials, accompanying with localized heating. This generated thermal energy will act as a destroying force to molecular bonds. When thermal energy is greater than bonding energy of molecular, fracture will be possible \citep{slootman_quantifying_2020}. Part of internal energy dissipates to become surface energy\citep{zhou_high-throughput_2022}.
On the other hand, ultrasonic welding is widely used to bond similar and dissimilar polymers, metals, and woods\citep{unnikrishnan_review_2022,li_ultrasonic_2022, bendikiene_study_2021, jongbloed_differences_2020, wagner_ultrasonic_2013}. When an ultrasonic welding device composed of converter, booster and sonotrode is used to bond two pieces of materials such as thermoplastics, thermoplastic elastomer, and thermosets, mechanical energy generated by ultrasound wave will be propagated, amplified and finally localized to the interface between two pieces of adherends \citep{yang_numerical_2022, rubino_ultrasonic_2020, guo_joining_2019}. Localization of stress wave energy will lead to thermal energy accumulated around the interface of adherends. Due to this thermal energy, the temperature rising rate will depend on thermal capacity of polymer materials since thermal capacities of polymers have a big shift across glass transition temperature \citep{koutras_thermal_2021}. Therefore, a large difference of temperature rising rate was observed across glass transition temperature\citep{zhang_study_2010}. After polymers absorb a large amount of thermal energy, molecular bonds will be broken and materials become flowable liquid shape. Once temperature is cooled down, molecular bonds will be reformed.
In summary, if a high speed impact force is applied on materials, one piece can be broken into two. If a cyclic vibration force is applied on materials, two pieces can be stitched. Both are due to localized plastic deformation and accompanying temperature rising.
Traditionally, heating generation due to ultrasonic welding is attributed to frictional heating followed by viscoelastic heating \citep{zhang_study_2010, bhudolia_advances_2020,levy_modeling_2014}. That is on the right track. But it didn't tell the whole story of the ultrasonic welding process. Firstly, localized heating can hardly be explained by frictional heating and viscoelastic heating \citep{koellhoffer_role_2011, villegas_ultrasonic_2019}. Because frictional heating and viscoelastic heating happen everywhere inside materials, why only materials near the interface get heating up. If a longitudinal vibration is applied to adherends, frictional force on the interface can hardly be calculated. Secondly, with high temperature rising and large plastic deformation during ultrasonic vibration, linear viscoelastic heating plays a minor role on bonding formation. Frictional heating is more related to molecular defect motions. Recall in ultrasonic welding processes, energy directors are usually used to guarantee a good bonding quality. Physically, an energy director is used to direct mechanical energy to a particular location. It is directly related to how localized plastic energy changes to thermal energy. So energy directors act as mechanical impedances to block ultrasonic energy propagation.
In the following section, we will demonstrate this process from physical aspects and mathematical aspects.
The rest of the present paper is organized as follows. Section \ref{s:consti}
illustrates the physics of ultrasonic welding principle. Section \ref{s:math} is to explain a mathematical model of ultrasonic welding, especially temperature rising.
Then we will talk about localized heating by using energy directors. Finally, we offer the concluding remarks, followed by a list of cited references.
\section{Physics of Ultrasonic Welding Principle}
\label{s:consti}
\noindent
It is well known that no materials are perfect. There are always some defects inside the materials. In polymers, molecular chains are either cross-linked or entangled \citep{yang_note_2018}. de Gennes \citep{de_gennes_reptation_1971} gave an idea that reptation of molecular chains is caused by defect movements along a molecular chain. Equivalently, defects in molecular chains can be viewed as the results induced by frictional forces or entangled forces between different molecular chains. Plastic deformation in polymers is due to the accumulation and movement of these defects \citep{yang_mathematical_2019}. In metals, line defects such as dislocations are abundant. Dislocation densities for many metals are ranged from $10^{10}$ to $10^{16}/cm^2$. Dislocation generation and movement is closely related to plastic deformation. Therefore, for both polymers or metals, their defects will be the main causes for plastic deformation, especially localized plastic deformation and temperature rising.
Nowadays, there are a lot of experimental investigations and numerical simulations of ultrasonic welding processes \citep{gohel_effect_2022, charlier_ultrasonic_2022, b_g_brito_effects_2022, frederick_disassembly_2021}. But work on understanding the principle of ultrasonic welding is relatively rare \citep{siddiq_thermomechanical_2008, li_integrated_2020}.
In order to understand the principle of ultrasonic welding, firstly we need to understand energy flow due to ultrasonic wave. Particularly, we need to know the cases when stress wave can propagate \citep{yang_numerical_2010, yu_numerical_2010} and when stress wave will stop propagating.
In general, when stress wave propagates, material particles will vibrate in their equilibrium positions. Mechanical energy will be transported to the next adjacent particle by molecular bonding forces. Wave propagation speed will depend on the stiffness of molecular bonding. Suppose there exists an ideal material which has no defects. Then energy will continue transporting to next particle without any loss. However, defects are always exist. Defects will act as impedances to energy transport. For example, in metals, once energy carried by ultrasonic wave is absorbed by dislocations, immobile dislocations will be disengaged by this mechanical energy. Therefore, mobile dislocation density and mobile dislocation velocity will increase. The kinetic energy level of these mobile dislocations will increase as well. But they cannot stay in the higher energy state for a long time since they will collide with lattice frames. dislocation scattering will happen. Their kinetic energies will fall down to a lower energy level quickly. After mobile dislocations lose their kinetic energy, they become immobile dislocations again. But immobile dislocation density will change in this energy gain and lose procedure. Accumulated immobile dislocation density will cause permanent plastic deformation. At the same time, because of mobile dislocation collision with lattice frames, dislocations lose kinetic energy. But lattice frames will gain kinetic energy. Lattices will vibrate more dramatically. Recall that temperature is a microscopic measurement of gas molecular velocity in idea gas or lattice vibration amplitude in solids. Therefore, most lost kinetic energy of dislocations will change to lattice vibration which is macroscopically shown as a gained thermal energy. This leads to temperature rising. Some of their energy will also radiate out as a sound wave.
Similarly, for polymers, because molecular chains are all entangled or cross-linked, bonding force of cross-links is much larger than entangled force. When energy carried by ultrasonic wave propagates along the molecular chains, entangled molecular chains can be disengaged by the ultrasonic wave energy e.g. many short molecular chains or danglings
\citep{lin_fracture_2021} can be formed. Therefore, mobile short molecular chain density and velocity will increase. Like dislocations, these short molecular chains will collide against main backbone chains. Collision scattering of these short molecular chains can cause them to lose their kinetic energy very quickly. On the other hand, main backbone chains will gain kinetic energy from those discrete short molecular chains and vibrate more violently. This gives rise to temperature rising in the entire areas.
Since energy from ultrasonic wave is firstly absorbed by these short chain defects, they will jump to a high energy level and return to a low energy level in a short time.
With redistribution of short chain density and velocity, plastic deformation, temperature rising and bonding will happen.
To make this physics more clear, mechanical energy flow could be compared with electric energy flow.
Physics of ultrasonic wave propagation in metals or polymers is similar to that of electrical current propagation in conductors. In electric circuits, when voltage is setup between a conductor, electrical current will carry electric energy from positive voltage to negative voltage. Due to existence of electric resistance, this electric energy transport is not lossless. Once voltage sets up an electric field inside conductor, electrons will gain a higher speed. Mobile electron density and velocity will increase. That is, these electrons will jump to a higher energy level. But these electrons can only stay at the higher energy level in a very short time, less than a microsecond. Because electron scatterings happen, high speed electrons will collide to nuclei and slow down. Kinetic energy of electrons will be lost. Some of their energy will dissipate as thermal energy and cause temperature rising. Some other will emit as a light wave.
Furthermore, heat-induced shear band and heat-induced ultrasonic welding in materials can be analogy to electric fuse wire and electric soldering.
For electric fuse wires, their melting temperature is much lower than conductor wires. Therefore, an electric fuse wire is the weakest point of the entire conductor wire when an electric current wave passes through. When a large amplitude electric current passes through an electric fuse, heat generated by a large electric current will melt the fuse wire and break it into two pieces. Similarly, if a large impact force is applied to metals or polymers, a high stress gradient and a large strain rate will be generated in materials. Mechanical energy carried by stress wave will pass through the materials. In the weakest point or the largest mechanical resistance area, mechanical energy will change to localized plastic and thermal energy. This leads to shear band and breaks a structure into two pieces.
In electric soldering process, electric energy is used to melt a solder. The solder will secure the connection of two metal surfaces. In ultrasonic welding process, mechanical energy is used to melt energy directors which will bond two pieces of materials.
\section{Mathematics of Ultrasonic Welding Principles}
\label{s:math}
\noindent
As mentioned before, ultrasonic welding is inverse to shear band and necking formation. Heat generated by impact force can damage materials and create cracks. But heat generated by ultrasonic wave can bond materials and eliminate cracks. Therefore, The same mathematical equation used to model shear band and necking formation can be applied to ultrasonic welding.
Yang and Yang \citep{yang_revisit_2020} proposed
a mathematical model to explain localized plastic deformation such as shear band and necking with defects included in the equation. The mathematical structure can be written as
\begin{equation}
\frac{d \sigma}{d \epsilon} \: = \: E \, - \, R \,
\label{eq:localized}
\end{equation}
where $ \sigma$ is engineering stress, $\epsilon$ is engineering strain, $E$ is Young's modulus, and $R$ is mechanical resistance. $R$ is a function of strain and stress and can be written as \citep{yang_revisit_2020}
\begin{equation}
R\: = q\epsilon^m \sigma^n,
\label{eq:resistance1}
\end{equation}
where $E$ and $q$ are time, temperature, pressure dependent which can be modeled using modified logistic functions \citep{yang_viscoelasticity_2021}. $m$ and $n$ are constants depending on materials.
Eq.(\ref{eq:localized}) can be graphically illustrated as shown in Fig.(\ref{f:f4}).
\begin{figure}
\centering \includegraphics[scale = 0.3]{Figure4.eps}
\caption{
Sketch of a stress-strain relationship. Maximum point on the curve corresponds initiation of localized plastic deformation.
}
\label{f:f4}
\end{figure}
Mechanical resistance $R$ increases with stress and strain, see Eq.(\ref{eq:resistance1}) and Young's modulus is independent of stress and strain. Therefore, the slope of stress and strain curve will gradually decrease with increased stress and strain. As shown in Fig.(\ref{f:f4}), the slope of stress and strain curve will be determined by the value of $E - R$ which can be positive or negative. Before maximum engineering stress is reached, e.g., $E > R$, slope is positive. Homogeneous elastic-plastic deformation will happen. After passing maximum engineering stress, e.g., $E < R$, slope becomes negative. Localized plastic deformation will happen. Localized thermal energy will accumulate in a very short band. Temperature in this narrow band can be higher than melting temperature of materials.
Next,
thermal energy due to ultrasonic wave will be illustrated.
Traditionally, energy transmission in ultrasonic welding is based on elastic wave propagation theory. In the interface between the sonotrode and the adherend, some of energy will transmit and some will reflect back \citep{palardy_study_2018}. Energy calculation is based on power reflection coefficient and transmission coefficient \citep{li_weld_2019, yang_modeling_2011}. With this method, mechanical energy arrived at the surface of adherend can be obtained \citep{rogale_interdependence_2022}. Next, this mechanical energy will enter the materials and change to thermal energy. How this mechanical energy changes to thermal energy is what we will discuss below. Currently most researchers explained this transition from mechanical energy to thermal energy by using linear viscoelastic theory \citep{yang_viscoelasticity_2013, yang_theoretical_2020}. Loss modulus is used to calculate the thermal energy. This approach may be acceptable in some cases such as small material deformation. The limitation of viscoelastic heating in traditional linear viscoelastic theory is that no micro-structures and defects of materials are considered in the model. In ultrasonic welding, most of ultrasonic wave energy will be absorbed by defects such as dislocations in metals or short molecular chains in polymers. These defect movements are related to nonlinear visco-plasticity. Large deformation, permanent deformation, large temperature rising can hardly be explained by linear viscoelasticity \citep{long_experiments_2022}. If only linear viscoelasticity is used to calculate temperature rising, temperature will not be high enough to melt the materials \citep{bakavos_mechanisms_2010}. For example, if ultrasonic frequency is 20000 Hz and bonding time is 1 ms, there are only 20 loading and unloading cycles. Nevertheless, localized temperature rising cannot be explained by linear viscoelasticity. Therefore, a new way to calculate temperature rising in ultrasonic welding is introduced.
As we know, when electric potential is setup on a electric circuit, an electric field will be created. Electrons will drift in the electric field and create electric current $I$. Electric current is related to charge density and charge velocity by
\begin{equation}
I \: =\: nq_ev_eA,
\label{eq:current}
\end{equation}
where $n$ is electron concentration, $q_e=1.6 \times 10^{-19}$ coulombs/charge, $n \times q_e$ is electron density, $v_e$ is electron velocity, $A$ is cross section of conductor wires. The heat, $H$, generated by electric current in the electric resistance can be calculated by using Joule's law
\begin{equation}
H \: =\: ( I ) \,^{2} R_e t,
\label{eq:ohm2}
\end{equation}
where $R_e$ is electric resistance.
For dislocation induced plastic deformation, plastic strain rate, $\dot{\epsilon}_p$, is related to dislocation density and dislocation velocity by Orowan equation \citep{yang_mathematical_2019} as
\begin{equation}
\dot{\epsilon}_p \: = \rho b v,
\label{eq:strainrate}
\end{equation}
where $\rho$ is dislocation density, $b$ is Burger's vector, $v$ is dislocation velocity.
By comparing with electric current equation, e.g., Eq.(\ref{eq:current}), we can define mechanical current due to dislocation drifting as
\begin{equation}
I_m = \dot{\epsilon}_p A_m = \rho b v A_m ,
\label{eq:mechanical_current}
\end{equation}
where $I_m$ is mechanical current and $A_m$ is the cross section of a metal bar.
If an electric potential $U$ is applied between an electric resistance $R_e$, a constant electric current $I$ will be maintained. Ohm's law is written as
\begin{equation}
U \: = R_e I.
\label{eq:ohm}
\end{equation}
If a rod is under dynamical impact, stress inside the bar is not uniform. Consider a case that there is a stress gradient along the bar, e.g. one side is in tension while the other side is in compression, see Fig.\ref{f:f1}. There are more materials or equivalently atoms accumulated in the compression side than in the tension side which will build up a mechanical potential. This mechanical potential will setup a mechanical field. This mechanical field will drive dislocations to move along the field gradient direction. Dislocations are line defects which can be viewed as a combination of point defects. Point defects can be vacancies and interstitials. Dislocations with vacancies will drift to the compression side. Dislocations with interstitials will drift to the tension side. Similarly, for polymer bars, chain density in tension side is smaller than that in compression side. short chains who lose their connections to main chains will drift in the built-up field, see Fig.\ref{f:f2} for schematic illustrations.
To cast, if a mechanical potential $U_m$ is maintained between a rod with a mechanical resistance $R_m$, a constant mechanical current $I_m$ will be maintained. Its relationship is written as
\begin{equation}
U_m \: = R_m I_m.
\label{eq:myohm}
\end{equation}
\begin{figure}
\includegraphics[scale = 0.2]{Figure1.eps}
\caption{
Schematic illustration of propagation of defects in crystalline materials. In compression side, atoms are closely packed. In tensile side, the spaces between atoms are wider. Mechanical potential is built based on this uneven distribution of atoms. Defects such as dislocations will drift along this built-in potential
}
\label{f:f1}
\end{figure}
Heat generated per unit volume can be calculated by
\begin{equation}
H \: =\: ( \, I_m ) \,^{2} R_m t,
\label{eq:ohm1}
\end{equation}
Analogy to electric resistance calculation method, mechanical resistance of materials can be obtained by
\begin{equation}
R_m \: =\: l_m/\sigma_m A_m ,
\label{eq:mech_resistance}
\end{equation}
where $l_m$ is the length of mechanical resistor. $\sigma_m$ and $A_m$ are conductivity and cross section of the mechanical resistor, respectively. Conductivity $\sigma_m$ will depend on dislocation scattering in metals or short molecular chain scattering in polymers.
\begin{figure}
\includegraphics[scale = 0.2]{Figure2.eps}
\caption{
Schematic illustration of propagation of defects in polymer materials. In compression side, polymer chain density is high. In tensile side, polymer chain density is low. Mechanical potential is built based on this uneven distribution of molecular chains. Defects such as danglings will drift along this built-in potential.
}
\label{f:f2}
\end{figure}
\section{Energy Director}
\label{s:Energy}
In ultrasonic welding process, welding quality is controlled by many factors such as welding time, welding amplitude, ultrasonic frequency, design of booster and horn, and energy directors \citep{tao_influence_2019, yang_ultrasonic_2022}. Factors except energy directors are used to control how much ultrasonic wave energy flows to energy directors. This energy will finally go to energy directors, which will be melt and used to bond two adherends \citep{tao_influence_2019}.
Energy directors can be designed as different shapes such as semicircular, triangular, rectangular \citep{bhudolia_advances_2020}. For example, Herrmann Ultrasonics, Inc. designed a triangular energy director, shown in Fig.(\ref{f:f3}). Purpose of energy directors has two. The first one is to create localized plastic deformation and localized temperature rising. Consider Eq.(\ref{eq:localized}), when $E < R$, localized plastic deformation will be generated. Consider Eq.(\ref{eq:resistance1}), $R$ is a function of strain and stress. At the tip of a triangular energy director, strain and stress are the largest which means localized heat will start from the tip. The first melting point will start from the tip. This can also be confirmed by Eq.(\ref{eq:mech_resistance}). Mechanical resistance will increase if cross section of an energy director decreases. Therefore, the tip part of this energy director will have the largest mechanical resistance. Melting will be initiated at the tip.
The second purpose of an energy director is bond two adherends after melting. Ideally designed energy directors should be melted completely. Eq.(\ref{eq:ohm1}) can be used to calculate heat generated by ultrasonic wave. With particular design of energy directors, we can adjust ultrasonic amplitude and frequency, welding time to melt it completely. Since the mechanical resistance of an energy director is related to its length and cross section by Eq.(\ref{eq:mech_resistance}), the amount of heat generated will change with different shapes of energy director design \citep{chuah_effects_2000}. We can improve welding speed and obtain the same bonding strength by properly designed energy directors.
\begin{figure}
\includegraphics[scale = 0.2]{Figure3.eps}
\caption{
Triangle tongue and groove energy director design from Herrmann Ultrasonics, Inc. Energy flows from tongue to energy director. Energy director melting direction will start from the tip of the triangle.
}
\label{f:f3}
\end{figure}
\section{Conclusion}
\label{s:conclusion}
In this article, physics and mathematics of ultrasonic bonding process are illustrated. We are trying to understand the principle of ultrasonic welding process in mesoscale level or even in quantum level. We compared defect motions in mesoscale to electron motions in quantum level. Based on former work and history learning, we showed that they can share the same mathematical structures.
Since many mesoscale mechanisms of ultrasonic welding processes can hardly be captured by experimental measurements and finite element analysis, most of current work are based on physical intuition and literature study. With literature study,
energy of ultrasonic wave is mostly absorbed by defects in the materials. These defects will transfer their energy to lattices in crystalline structures or main backbone chains in amorphous structures. Large amplitude lattice vibration and main backbone chain vibration will lead to high temperature rising. When large amplitude vibration increases to a certain value, it is possible that covalent bonds of lattices or main backbone chains break apart. Materials will be melted.
We also propose that an ultrasonic welding process is an inverse procedure to shear band formation. Both are due to localization of plastic deformation and temperature rising. Shear band will damage materials while ultrasonic bonding will repair materials.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
Suppose $\mathcal P$ is a property that can be attributed to a 2-complex. A relative version of $\mathcal P$ which can be applied to 2-complex pairs $(L,K)$, where $K$ is a subcomplex of $L$, can be a powerful tool, especially if a transitivity law holds: If $(L,K)$ has property $\mathcal P$ relative to $K$ and $K$ has property $\mathcal P$, then $L$ has property $\mathcal P$. A good example of a property that has a relative version and satisfies a transitivity law is \lq\lq vanishing reduced homology\rq\rq. In \cite{HarRose2017} and \cite{HarRose2020} we used relative versions of combinatorial asphericity to prove that injective labeled oriented trees are aspherical. This answered the long standing asphericity question for ribbon disc complements in the alternating case. In this paper we study relative non-positive immersion.
An $n$-cell in a complex is called {\em free} if it is used exactly once in an attaching map of a unique $(n+1)$-cell. A cell that contains a free cell in its boundary can be collapsed without changing the homotopy type of the complex. A 2-complex map $f\colon X\to K$ is {\em combinatorial} if it maps open cells homeomorphically to open cells. It is an {\em immersion} if it is locally injective.
\begin{defn} A 2-complex $K$ has {\em collapsing non-positive immersion} if for every combinatorial immersion $X\to K$, where $X$ is finite, connected, and without free vertices or edges, either $\chi(X)\le 0$ or $X$ is a point.
\end{defn}
A number of variants of non-positive immersion were defined by Wise. See Definition 1.2 in \cite{Wise22}. See also \cite{Wise04}. Among many other things Wise showed that if $K$ has non-positive immersion then $\pi_1(K)$ is locally indicable in case that this group is not trivial.
The {\em essential part} of a 2-complex consists of all the 2-cells together with the edges that are used in the attaching paths for the 2-cells.
\begin{defn}
Let $L$ be a 2-complex and $K$ be a subcomplex. Then $(L,K)$ has {\em relative collapsing non-positive immersion} if for every combinatorial immersion $f\colon X\to L$, where $X$ is finite, connected, and without free vertices or edges, either $\chi(X)\le \chi(Y)$, where $Y$ is the essential part of the pre-image $f^{-1}(K)$, or $X$ is a point.
\end{defn}
The weight test and its stronger precursor, the coloring test, are major tools in the study of asphericity of 2-complexes. See Gersten \cite{Ger87} and Sieradski \cite{S83}. These tests can be applied to angled 2-complexes. The coloring test and one of its relative cousins is explained in detail in the next section. An important observation by Wise \cite{Wise04} states that an angled 2-complex that satisfies the coloring test has collapsing non-positive immersion. We do not know if our version of relative collapsing non-positive immersion is transitive in general. Suppose $(L,K)$ has relative collapsing non-positive immersion and $K$ has collapsing non-positive immersion. We would like to show that $L$ has collapsing non-positive immersion. So suppose $f\colon X\to L$ is a combinatorial immersion where $X$ does not have a free vertex or edge, and $X$ is not a point. Let $Y$ be the essential part of $f^{-1}(K)$. Let $Y=Y_1\cup\ldots\cup Y_n$, where the $Y_i$ are the connected components. Then
$$\chi(X)\le \chi(Y_1)+\ldots +\chi(Y_n).$$
Now $f\colon Y_i\to K$ is a combinatorial immersion, and since we assume that $K$ has collapsing non-positive immersion we have $\chi(Y_i)\le 0$, unless $Y_i$ collapses to a point. That is where the problem lies: In general we have no way to avoid collapsing connected components of $Y$. However, under certain circumstances we can control the situation. Here is one of our main results (see Theorem \ref{thm:mainapp}), vocabulary is defined in detail in the next two sections:\\
{\em Let $L$ be a standard 2-complex, $K\subseteq L$ a subcomplex all of whose 2-cells are attached along paths of exponent sum zero. Let $\bar L$ be obtained from $L$ by folding $K$ to the single edge $y$. Assume that $\bar L$ has a zero/one-angle structure that satisfies the coloring test, and $y^+$ and $y^-$ lie in different components of $lk_0(v,\bar L)$. Then $(L,K)$ has relative collapsing non-positive immersion. If in addition $K$ has collapsing non-positive immersion, then so does $L$. }\\
It was already mentioned that relative notions of combinatorial asphericity were used by the authors to show that reduced injective labeled oriented trees (LOTs) are vertex aspherical (VA) and hence aspherical. The motivation for this work is the question whether reduced injective LOTs have non-positive immersion. This is known to be true in the prime case, when the LOT is without sub-LOTs, and we answer this question affirmatively in an important but special non-prime setting. Relevant language and some history about LOTs is provided in the last section of this paper.
\section{Coloring tests}
Let $K$ be a 2-complex. If $v$ is a vertex of $K$ the {\em link at v}, $lk(v,K)$, is the boundary of a regular neighborhood of $v$. It is a graph whose edges come from the corners of 2-cells. For that reason we refer to the edges of $lk(v,K)$ as {\em corners at $v$}. If we assign numbers $\omega(c)$ to the corners $c$ of the 2-cells of $K$ we arrive at an {\em angled 2-complex}. Curvature in an angled 2-complex is defined in the following way. If $v$ is a vertex of $K$ then $\kappa(v,K)$, the curvature at $v$, is
$$\kappa(v,K)=2-\chi(lk(v,K))-\sum \omega(c_i),$$ where the sum is taken over all the corners at $v$. If $d$ is a 2-cell of $K$ then $\kappa(d,K)$, the curvature of $d$, is
$$\kappa(d,K)=\sum \omega(c_j)-(|\partial d|-2),$$ where the sum is taken over all the corners in $d$ and $|\partial d|$ is the number of edges in the boundary of the 2-cell. The combinatorial Gauss-Bonnet Theorem states that
$$2\chi(K)=\sum_{v\in K} \kappa(v,K) + \sum_{d\in K} \kappa(d,K).$$
This was first proven by Ballmann and Buyalo \cite{BallBuy}, and later observed by McCammond and Wise \cite{McCammondWise}.
A map $X\to K$ between 2-complexes is called {\em combinatorial} if it maps open cells homeomorphically to open cells. Note that if $K$ is an angled 2-complex then the angles in the 2-cells of $K$ can be pulled back to make $X$ into an angled 2-complex. We call this angle structure on $X$ the one {\em induced} by the combinatorial map.
An angled 2-complex where all angles are either 0 or 1 is called a {\em zero/one angled 2-complex}.
We denote by
$lk_0(v,K)$ the subgraph of $lk(v,K)$ consisting of the vertices of $lk(v,K)$ together with the corners with angle $0$. The following coloring test is due to Sieradski \cite{S83}.
\begin{defn}(Coloring test)
Let $K$ be a zero/one-angled 2-complex. Then $K$ {\em satisfies the coloring test} if
\begin{enumerate}
\item the curvature of every 2-cell is $\le 0$;
\item for every vertex $v$: If $c_1\cdots c_n$ is a reduced cycle in $lk(v,K)$, then\\ $2-\sum_{i=1}^n \omega(c_i)\le 0$.
\end{enumerate}
\end{defn}
\begin{thm}\label{thm:npi} A zero/one-angled 2-complex $K$ satisfies the coloring test if and only if
\begin{enumerate}
\item the curvature of every 2-cell is $\le 0$;
\item $lk_0(v,K)$ is a forest for every vertex $v$;
\item a corner with angle $1$ does not have both its vertices in a single connected component of $lk_0(v,K)$.
\end{enumerate}
\end{thm}
The proof is straightforward.
\begin{defn}
Let $\Lambda$ be a graph.
\begin{itemize}
\item A cycle of edges $e_1\cdots e_n$ in $\Lambda$ is {\em reduced} if there does not exist an $e_i$ so that $e_{i+1}=\bar e_i$ ($i$ mod $n$), where $\bar e_i$ is the edge $e_i$ with reversed orientation. A cycle of edges $e_1\cdots e_n$ in $\Lambda$ is {\em homology reduced} if there does not exist a pair $e_i, e_j$ so that $e_j=\bar e_i$.
\end{itemize}
Let $K$ be a 2-complex.
\begin{itemize}
\item Let $S$ be the 2-sphere. A {\em spherical diagram} is a combinatorial map $f\colon S\to K$. It is {\em reduced} if, for every vertex $v\in S$, $f$ maps $lk(v)$ to a reduced cycle. It is {\em vertex reduced} if, for every vertex $v\in S$, $f$ maps $lk(v)$ to a homology reduced cycle.
\item $K$ is {\em diagrammatically reducible (DR)} if there do not exist reduced spherical diagrams over $K$. $K$ is {\em vertex aspherical (VA)} if there do not exist vertex reduced spherical diagrams over $K$.
\end{itemize}
Let $(L,K)$ be a 2-complex pair.
\begin{itemize}
\item The pair is {\em relatively DR} if every reduced spherical diagram over $L$ is a diagram over $K$. It is {\em relatively VA} if every vertex reduced spherical diagram over $L$ is a diagram over $K$.
\item A combinatorial map $f\colon X\to L$ is {\em $K$-thin} if the essential part $Y$ of $f^{-1}(K)$ has no interior vertices; that means $lk(v,Y)\ne lk(v,X)$ for every vertex $v\in X$.
\end{itemize}
\end{defn}
\begin{thm}\label{thm:wise}
If $K$ satisfies the coloring test then it is DR and has non-positive immersion.
\end{thm}
The DR part was shown by Sieradski \cite{S83} and non-positive immersion was shown by Wise \cite{Wise04}.
\begin{defn}\label{def:relforest}
Let $\Lambda$ be a graph and $\Lambda'$ be a subgraph, both are allowed to be disconnected. Then
\begin{itemize}
\item {\em $\Lambda$ is a forest relative to $\Lambda'$} if every reduced cycle in $\Lambda$ is contained in $\Lambda'$. {\em $\Lambda$ is a tree relative to $\Lambda'$} if in addition $\Lambda$ is connected.
\item {\em $\Lambda$ is a strong forest relative to $\Lambda'$} if it is a forest relative to $\Lambda '$ and in addition, for each connected component $C$ of $\Lambda$, either $C\cap \Lambda'$ is empty or connected.
\end{itemize}
\end{defn}
\begin{figure}[htbp]
\centering
\includegraphics[width=2in]{strongforest.pdf}
\caption{A strong forest $\Lambda$ relative to $\Lambda'$ (red).}
\end{figure}
\vfill\eject
\begin{prop}\label{rem:strongrelforest}
If $\Lambda$ is a forest relative to $\Lambda'$ and $C$ is a component of $\Lambda$ such that $C\cap \Lambda'\ne \emptyset$ then at most one component of $C\cap \Lambda'$ can not be a tree.
Let $\Lambda/\Lambda'$ be the quotient graph obtained by identifying all of $\Lambda'$ to a single vertex. Then the following conditions are equivalent:
\begin{enumerate}
\item $\Lambda$ is a strong forest relative to $\Lambda'$;
\item The quotient graph $\Lambda/\Lambda'$ is a forest.
\end{enumerate}
\end{prop}
\begin{proof}
Let $C$ be a component of $\Lambda$. Suppose that $C\cap \Lambda'$ contains two components $A$ and $B$ that are not trees. Let $\alpha$ and $\beta$ be reduced closed paths in $A$ and $B$ respectively. Let $\gamma$ be a reduced path connecting $\alpha$ and $\beta$. A reduced version of $\alpha\gamma\beta\bar\gamma$ is a reduced closed path in $\Lambda$ that is not entirely contained in $\Lambda'$, contradicting the fact that $\Lambda$ is a forest relative to $\Lambda'$.
Assume that $\Lambda$ is a strong forest relative to $\Lambda'$ and let $C$ be a component of $\Lambda$. Since $C\cap \Lambda'$ is connected (or empty) and $C$ is a tree relative to $C\cap \Lambda'$, it is clear that the quotient $T_C=C/C\cap \Lambda'$ is a tree. $\Lambda/\Lambda'$ is now obtained from the forest $\bigcup T_C$ (disjoint union) by identifying single vertices from distinct $T_C$'s. This gives a forest.
Assume next that $\Lambda/\Lambda'$ is a forest. Let $C$ be a component of $\Lambda$. It is clear that $C\cap \Lambda'$ is connected or empty, because $C/C\cap \Lambda'$ is a tree. Let $q\colon C\to C/C\cap \Lambda'$ be the quotient map. Suppose that $\gamma=e_1\ldots e_n$ is a reduced closed path in $C$ not entirely contained in $C\cap\Lambda'$. Then $q(\gamma)=q(e_1)\ldots q(e_n)$ is a closed path in $C/C\cap\Lambda'$ that does contain an edge. Note that if $e_i\in \Lambda'$ then $q(e_i)$ is omitted from the sequence $q(e_1)\ldots q(e_n)$. Since $q(e_1)\ldots q(e_n)$ is a path in a tree, it contains at least two vertices of valency 1, so it contains a vertex of valency 1 that is not $C\cap \Lambda'$.
We may assume, after cyclic reordering, that $q(e_1)$ is an edge that contains a vertex of valency 1 that is not $C\cap \Lambda'$. But then $e_1$ is an edge in $C$ that contains a vertex of valency 1, which implies that $e_1\ldots e_n$ is not reduced. In fact $e_n=\bar e_1$. A contradiction.
\end{proof}
\newpage
\begin{defn}\label{defn:relcol} (Relative coloring test)
Let $K$ be a subcomplex of a zero/one-angled 2-complex $L$. Then $(L,K)$ {\em satisfies the relative coloring test} if
\begin{enumerate}
\item the curvature of every 2-cell $d\in L-K$ is $\le 0$;
\item for every vertex $v$: If $c_1\cdots c_n$ is a reduced cycle in $lk(v,L)$ not entirely contained in $lk(v,K)$, then $$2-\sum_{i=1}^n\omega(c_i)\le 0.$$
\end{enumerate}
\end{defn}
\begin{thm}\label{thm:relcol} A zero/one-angled pair $(L,K)$ satisfies the relative coloring test if and only if
\begin{enumerate}
\item the curvature of every 2-cell $d\in L-K$ is $\le 0$;
\item for every vertex $v$: $lk_0(v,L)$ is a forest relative to $lk_0(v,K)$;
\item if a corner $c$ with angle $1$ in a 2-cell $d\in L$ has its vertices in a single connected component of $lk_0(v,L)$, then $d\in K$ and the vertices of $c$ lie in a single component of $lk_0(v,K)$.
\end{enumerate}
\end{thm}
\begin{proof} Assume $(L,K)$ satisfies the relative coloring test. Then conditions 2 and 3 hold because otherwise one can construct a reduced cycle of corners $\lambda=c_1\ldots c_n$ not entirely contained in $lk(v,K)$ such that $\omega(\lambda)=\sum_{i=1}^n\omega(c_i)=0$ or $\omega(\lambda)=1$.
Assume that the conditions of the theorem hold.
Let $\lambda=c_1\ldots c_n$ be a reduced cycle in $lk(v,L)$ not entirely contained in $lk(v,K)$.
$\omega(\lambda)=\sum_{i=1}^n\omega(c_i)=0$ is not possible because of condition 2 in the statement of Theorem \ref{thm:relcol}. If $\omega(\lambda)=1$ then there is $1\le i\le n$ so that $\omega(c_i)=1$ and $\omega(c_j)=0$ for $j\ne i$. Then by condition 3, $c_i$ is a corner in a 2-cell of $K$ with its vertices $p$ and $q$ in a single component of $lk_0(v,K)$. Let $\lambda'$ be a corner path in $lk_0(v,K)$ connecting $p$ to $q$. Then a reduced version of $c_i\ldots c_{i-1}\lambda'c_{i+1}\ldots c_n$ violates condition 2. Thus $\omega(\lambda)\ge 2$.
\end{proof}
The relative coloring test has implications regarding asphericity and non-positive immersion, but they are not as immediate as in the classical coloring test setting.
\begin{thm}\label{thm:relcolasph}
Let $(L,K)$ be a 2-complex pair with a zero/one angle structure that satisfies the relative coloring test and $\kappa(d,L)\le 0$ for all 2-cells $d\in L$. If $f\colon S\to L$ is a $K$-thin spherical diagram, then it is not reduced.
\end{thm}
\begin{proof} Assume $f\colon S\to L$ is a reduced $K$-thin spherical diagram. Pull the zero/one-angle structure back to $S$. Then we have $\kappa(d,S)\le 0$ for every 2-cell $d$ of $S$. Let $v$ be a vertex in $S$. Let $c_1\cdots c_n$ be the corners that make up the link of $v$. Since $f$ is $K$-thin we have that $f(c_1)\cdots f(c_n)$ is a reduced cycle in $lk(f(v),L)$ not entirely contained in $lk(f(v), K)$. It follows that $\kappa(v,S)\le 0$. We obtain
$$4=2\chi(S)=\sum_{v\in S}\kappa(v,S)+\sum_{d\in S}\kappa(d,S)\le 0$$
and have reached a contradiction.
\end{proof}
Under certain conditions imposed on the pair $(L,K)$ it can be shown that this result implies that $(L,K)$ is relatively DR. For more details see \cite{HarRose2020}.
\section{A strong relative coloring test}
In light of Theorem \ref{thm:relcol} we define a stronger version of the relative coloring test as follows:
\begin{defn}\label{def:strong} A zero/one-angled pair $(L,K)$ satisfies the {\em strong relative coloring test} if
\begin{enumerate}
\item the curvature of every 2-cell $d\in L$ is $\le 0$;
\item for every vertex $v\in L$: $lk_0(v,L)$ is a strong forest relative to $lk_0(v,K)$;
\item if a corner $c$ with angle $1$ in a 2-cell $d\in L$ has its vertices in a single connected component of $lk_0(v,L)$, then $d\in K$.
\end{enumerate}
\end{defn}
Note that if $c$ is a corner as in condition 3 with its vertices in a single component $C$ of $lk_0(v,L)$, then its vertices automatically lie in a single component of $lk_0(v,K)$ because $C\cap lk_0(v,K)$ is connected by condition 2.
Useful throughout is the following observation.
\begin{prop}\label{prop:colim}
Suppose $(L,K)$ is a zero/one-angled 2-complex pair that satisfies the (strong) relative coloring test. Let $f\colon X\to L$ be a combinatorial immersion and let $Y$ be the essential part of $f^{-1}(K)$. If we give $(X,Y)$ the angle structure induced from $(L,K)$, then $(X,Y)$ satisfies the (strong) relative coloring test.
\end{prop}
\begin{proof}
Since $f\colon X\to L$ is a combinatorial immersion it is immediate from Definition \ref{defn:relcol} that $(X,Y)$ satisfies the relative coloring test. Thus, by Theorem \ref{thm:relcol}, conditions 1 and 3 of Definition \ref{def:strong} hold for $(X,Y)$ and, for every vertex $x$, $lk_0(x,X)$ is a forest relative to $lk_0(x,Y)$. Assume that $(L,K)$ satisfies the strong relative coloring test. In order to show that $(X,Y)$ satisfies the strong relative coloring test, the only thing left to do is show that for every vertex $x$, $lk_0(x,X)$ is not just a forest relative to $lk_0(x,Y)$, but a strong forest. Let $\Lambda=lk_0(v,L)$ and $\Lambda_K=lk_0(v,K)$. Let $\Lambda'=lk_0(x,X)$, $\Lambda'_Y=lk_0(x,Y)$, and $\Lambda'_{f^{-1}(K)}=lk_0(x, f^{-1}(K))$, where $f(x)=v$. We can think of $\Lambda'$ is a subgraph of $\Lambda$. We have
$$\Lambda'_Y\subseteq \Lambda'_{f^{-1}(K)}\subseteq \Lambda'\subseteq \Lambda.$$
Note that
$$\Lambda'\cap \Lambda_K=\Lambda'_{f^{-1}(K)}.$$ It is easy to see that $\Lambda'$ is a strong forest relative to $\Lambda_{f^{-1}(K)}$. Indeed,
The quotient graph $\Lambda'/(\Lambda'\cap \Lambda_K)$ is a subgraph of $\Lambda/\Lambda_K$, and the latter is a forest because $\Lambda$ is a strong forest relative to $\Lambda_K$. Here we use Proposition \ref{rem:strongrelforest}. Thus $\Lambda'/(\Lambda'\cap \Lambda_K)$ is a forest and, again by Proposition \ref{rem:strongrelforest}, $\Lambda'$ is a strong forest relative to $\Lambda'_{f^{-1}(K)}$.
Since $Y$ is the essential part of $f^{-1}(K)$ we know that
$$\Lambda'_{f^{-1}(K)}=\Lambda'_Y\ \mbox{or}\ \Lambda'_{f^{-1}(K)}=\Lambda'_Y\cup \{p_1,\ldots, p_n\},$$
where the $p_i$ are points not in $\Lambda_Y'$. Suppose $C$ is a component of $\Lambda'$ and $C\cap \Lambda'_Y\ne \emptyset.$ Then $C\cap \Lambda'_{f^{-1}(K)}\ne \emptyset$ and therefore $C\cap \Lambda'_{f^{-1}(K)}$ is connected because $\Lambda'$ is a strong forest relative to $\Lambda'_{f^{-1}(K)}$. It follows that
$$C\cap \Lambda'_Y=C\cap \Lambda'_{f^{-1}(K)}$$
and so $C\cap \Lambda'_Y$ is connected.
\end{proof}
\begin{thm}\label{thm:relNPI1}
If $(L,K)$ satisfies the strong relative coloring test, then it has relative collapsing non-positive immersion.
\end{thm}
We will prove this theorem in the next section.
\bigskip Let $L$ be a 2-complex. We assume the 1-skeleton is a directed graph. If $e$ is an edge of $L$ we denote by $e^+$ a point close to the start of $e$ and by $e^-$ a point close to the end of $e$. Let $v$ be a vertex of $L$ and let $E(v)$ be the set of edges starting or ending at $v$. We denote by $lk^+(v,L)$ and $lk^-(v,L)$ the subgraphs of $lk(v,L)$ spanned by the $e^+$, and the $e^-$, $e\in E(v)$, respectively. Corners that are neither positive nor negative (i.e. not in $lk^+(v,L)\cup lk^-(v,L)$) are called mixed corners.
There is a {\em standard zero/one-angle assignment} for $L$: Give positive and negative corners angle $0$, and mixed corners angle $1$. Note that in this case we have
$$lk_0(v,L)=lk^+(v,L)\cup lk^-(v,L),\ lk^+(v,L)\cap lk^-(v,L)=\emptyset.$$ The second statement always holds and is independent of angle assignments.
Assume $L$ is a 2-complex and $K$ is a subcomplex all of whose 2-cells are attached along closed paths of exponent sum zero. We attach 2-cells to every closed path of exponent sum zero in $K$, reduced or not, to obtain a larger 2-complex $\hat K$. Let $\hat L$ be $L$ together with these added 2-cells. We call $(\hat L,\hat K)$ the {\em maximal expansion of $(L,K)$}. Note that $lk^+(v, \hat K)$ contains a full subgraph on its vertices, and so does $lk^-(v,\hat K)$. In particular both $lk^+(v, \hat K)$ and $lk^-(v, \hat K)$ are connected.
\begin{thm}\label{thm:relNPI}
Let $L$ is a 2-complex and $K$ is a subcomplex all of whose 2-cells are attached along loops of exponent sum zero. Assume
\begin{enumerate}
\item $(\hat L,\hat K)$ carries a zero/one-angle structure which satisfies the strong relative coloring test;
\item the angle structure on $\hat K$ is standard;
\item $K$ has collapsing non-positive immersion.
\end{enumerate}
Then $L$ has collapsing non-positive immersion.
\end{thm}
Since the maximal expansion $(\hat L,\hat K)$ satisfies the strong relative coloring test it has relative collapsing non-positive immersion by Theorem \ref{thm:relNPI1}. Since this property is hereditary by Proposition \ref{prop:colim}, it follows that $(L,K)$ has collapsing non-positive immersion.
Theorem \ref{thm:relNPI} established a transitivity property for collapsing non-positive immersion in a special setting.
We will prove this theorem in a later section. It will involve a process of \lq\lq thinning\rq\rq\ a given combinatorial immersion.\\
Assume $L$ is a standard 2-complex with single vertex $v$, and $K$ is a subcomplex. Assume that all 2-cells of $K$ are attached along paths of exponent sum zero. Let $y$ be an edge of $K$. Notice that we can ``fold" $K$ onto $y$ by mapping every edge of $K$ to $y$ and extending this map to the 2-cells (see Figure \ref{afoldy}). Let $\bar L$ be the quotient of $L$ obtained by folding $K\subseteq L$ to $y$.
\begin{example} Let
$$P_0=\langle y_1,y_2\ |\ y_1y_2y_1^{-1}y_2^{-1}\rangle \subseteq P=\langle x, y_1,y_2\ |\ y_1y_2y_1^{-1}y_2^{-1}, xy_1xy_2 \rangle.$$ Let $K\subseteq L$ be the presentation complexes of $P_0$ and $P$, respectively. We can fold $K$ to $y$ and obtain $\bar L$, the presentation complex for $\langle x, y\ | xyxy \rangle$.
\end{example}
\begin{figure}[ht]\centering
\begin{tikzpicture}[scale=1.3]
\begin{scope}[decoration={markings, mark=at position 0.5 with {\arrow{>}}}]
\draw [postaction={decorate}] (0,0) -- (1,0) node[midway, below]{$y_2$};
\draw [postaction={decorate}] (1,0) -- (1,1) node[midway, right]{$y_1$};
\draw [postaction={decorate}] (0,0) -- (0,1) node[midway, left]{$y_3$};
\draw [postaction={decorate}] (0,1) -- (1,1) node[midway, above]{$y_2$};
\draw [postaction={decorate}] (2,0) -- (2.8,0.5) node[midway, below]{$y$};
\draw [postaction={decorate}] (2.8,0.5) -- (3.1,1.5) node[midway, right]{$y$};
\draw [postaction={decorate}] (2,0) -- (2.3,1) node[midway, left]{$y$};
\draw [postaction={decorate}] (2.3,1) -- (3.1,1.5) node[midway, above]{$y$};
\draw [postaction={decorate}] (4,0) -- (4.4,1) node[midway, right]{$y$};
\end{scope}
\draw[blue] (2.3,0.2) arc (30:45:1cm);
\draw[blue] (2.8,1.3) arc (210:225:1cm);
\fill (4,0) circle (1.6pt);
\fill (4.4,1) circle (1.6pt);
\fill[blue] (4.08,0.2) circle (1.3pt); \node[right,blue] at (4.08,0.2) {$y^+$};
\fill[blue] (4.3,0.8) circle (1.3pt); \node[left,blue] at (4.3,0.8) {$y^-$};
\end{tikzpicture}
\caption{\label{afoldy} A 2-cell in $K$ is folded to the edge $y$.}
\end{figure}
Here is a key observation. $lk(v,\bar L)$ is obtained from $lk(v,L)$ by the following process:
\begin{enumerate}
\item Remove all mixed corners in $lk(v,K)$ from $lk(v,L)$.
\item Identify all of $lk^+(v,K)$ with $y^+$, and all of $lk^-(v,K)$ with $y^-$.
\end{enumerate}
\noindent Here is our main application of Theorem \ref{thm:relNPI}. It was already mentioned in the introduction.
\begin{thm}\label{thm:mainapp}
Let $L$ be a standard 2-complex, $K\subseteq L$ a subcomplex all of whose 2-cells are attached along paths of exponent sum zero. Let $\bar L$ be obtained from $L$ by folding $K$ to the single edge $y$. Assume that $\bar L$ has a zero/one-angle structure that satisfies the coloring test, and $y^+$ and $y^-$ lie in different components of $lk_0(v,\bar L)$. Then $(L,K)$ has relative collapsing non-positive immersion. If in addition $K$ has collapsing non-positive immersion, then so does $L$.
\end{thm}
\begin{proof}
Give $\bar L$ a zero/one-angle structure that satisfies the coloring test. Assign to the maximal expansion $\hat K$ the standard zero/one-angle structure, and for a 2-cell in
$\hat L$ not in $\hat K$ the angle structure pulled back from $\bar L$.
We will show that $(\hat L,\hat K)$ satisfies the strong relative coloring test and therefore the conditions of Theorem \ref{thm:relNPI} hold. Note first that we also have a map $\hat L\to \bar L$, folding all of $\hat K$ to the single edge $y$.\\
\noindent 1. If $d$ is a 2-cell in $\hat K$ then $\kappa(d,\hat K)\le 0$. This is because 2-cells in $\hat K$ are attached along paths of exponent sum zero and $\hat K$ carries the standard zero/one-angle structure. If $d$ is a 2-cell in $\hat L$ not in $\hat K$, then $\kappa(d,\hat L)\le 0$ because $\kappa(\bar d,\bar L)\le 0$, where $\bar d$ is the image of $d$ under the folding map. \\
\noindent 2. We have a quotient map $$q\colon lk_0(v,\hat L) \to lk_0(v,\bar L),$$
where all of $lk^+(v,\hat K)$ gets mapped to $y^+$, and all of $lk^-(v,\hat K)$ gets mapped to $y^-$. Note that $lk_0(v,\bar L)$ is a forest because $\bar L$ satisfies the coloring test.
Since $y^+$ and $y^-$ lie in different components of $lk_0(v, \bar L)$, $lk^+(v,\hat K)$ and $lk^-(v,\hat K)$ lie in different components of $lk_0(v,\hat L)$. The situation is shown in Figure \ref{fig:linkstrong}. Let $C^+$ be the component of $lk_0(v,\hat L)$ that contains $lk^+(v,\hat K)$. Since $q(C^+)$ is a tree, it follows that $C^+$ is a tree relative to $lk^+(v,\hat K)$. This follows from Proposition \ref{rem:strongrelforest}. Similarly, $C^-$ is a tree relative to $lk^-(v,\hat K)$. It follows that $lk_0(v,\hat L)$ is a forest relative to $lk_0(v,\hat K)=lk^+(v,\hat K)\cup lk^-(v,\hat K)$. Furthermore, if $C$ is any component of $lk_0(v,\hat L)$, then either $C\cap lk_0(v,\hat K)=\emptyset$, or $C\cap lk_0(v,\hat K)=lk_0^+(v,\hat K)$, or $C\cap lk_0(v,\hat K)=lk_0^-(v,\hat K)$. So $C\cap lk_0(v,\hat K)$ is either empty or connected.\\
\noindent 3. Let $c$ be a corner of angle 1. If $c$ comes from a 2-cell in $\hat K$, then $c$ connects $lk^+(v, \hat K)$ to $lk^-(v, \hat K)$, which lie in different components of $lk_0(v,\hat L)$. Assume $c$ comes from a 2-cell $d$ in $\hat L$ not in $\hat K$. The folding map $\hat L\to \bar L$ sends $c$ to $\bar c$ in $\bar d$. $\bar c$ is a corner of angle 1 and hence connects two distinct components of $lk_0(v,\bar L)$, because $\bar L$ satisfies the coloring test. Therefore $c$ connects distinct components of $lk_0(v,\hat L)$.
\end{proof}
\begin{figure}[ht]\centering
\begin{tikzpicture}[scale=1.1]
\node[ellipse, draw, red] (e) at (0,0) {$lk_0^+(v,\hat K)$};
\node[ellipse, draw, red] (e) at (5,0) {$lk_0^-(v,\hat K)$};
\draw (-0.5,-0.5) -- (-1.8,-1.8);
\draw (-1,-2) -- (-1,-1) -- (-2,-1);
\draw (0,-0.5) -- (0,-1.2);
\draw (0,1) -- (-0.5, 0.5) -- (-1, 1);
\draw (0.5, 0.5) -- (1,1) -- (1.5,1);
\draw (1,1) -- (0.8,1.5);
\draw (1.8,-0.5) -- (2.2,-0.5) -- (2.8,0) -- (2.8,0.5);
\draw (2.2,-0.5) -- (2.4, -1);
\draw (2.8,0) -- (3.1,-0.3);
\draw (3,-1.5) -- (3.7,-1.3) -- (3.6, -1.8);
\draw (3.7,-1.3) -- (4,-1);
\draw (5,-0.5) -- (5.6,-1.2);
\draw (4.8,-1.4) -- (5.3,-0.85) -- (5.8,-0.85);
\draw (4.5,-2) -- (4.8,-1.4) -- (5,-2);
\draw (6, -0.3) -- (6.4,-1);
\draw (4, -0.3) -- (3.3, -1);
\draw (5.5, 0.5) -- (6,1) -- (6.5,1);
\draw (6,1) -- (5.8,1.5);
\draw (4, 0.3) -- (3.3, 1);
\end{tikzpicture}
\caption{\label{fig:linkstrong}$lk_0(v,\hat L)$}
\end{figure}
\pagebreak
\begin{cor}\label{cstnpi}
Let $L$ be a standard 2-complex, $K\subseteq L$ a subcomplex all of whose 2-cells are attached along paths of exponent sum zero. Let $\bar L$ be obtained from $L$ by folding $K$ to the single edge $y$. Assume that $\bar L$ satisfies the coloring test with the standard zero/one-angle structure. Then $(L,K)$ has relative collapsing non-positive immersion. If in addition $K$ has collapsing non-positive immersion, then so does $L$.
\end{cor}
\begin{proof}
Note that in this setting $lk_0(v,\bar L)=lk^+(v,\bar L)\cup lk^-(v,\bar L)$. We have $y^+\in lk^+(v,\bar L)$, $y^-\in lk^-(v,\bar L)$, and $lk^+(v,\bar L)\cap lk_0^+(v,\bar L)=\emptyset$. Thus $y^+$ and $y^-$ lie in different components of $lk_0(v, \bar L)$.
\end{proof}
\begin{example}
Consider a presentation $P_0=\langle {\bf y}\ |\ {\bf r}\rangle$, where every $r\in {\bf r}$ has exponent sum zero. Let $P=\langle x, {\bf y}\ |\ {\bf r}, xux^{-1}v^{-1} \rangle$, where $u$ and $v$ are words in ${\bf y}^{\pm 1}$ of equal exponent sum $k>0$. The presentation $P$ typically arises in the context of HNN-extensions of groups, where the associated subgroup is $\mathbb Z$. Let $L$ be the presentation complex of $P$, and let $K$ be the subcomplex coming from $P_0$. Choose $y\in {\bf y}$. We fold $K$ onto $y$ and obtain a quotient $L\to \bar L$. Note that $\bar L$ is the standard 2-complex constructed from $\langle x,y\ |\ xy^kx^{-1}y^{-k}\rangle$. It is easy to check that $\bar L$ satisfies the coloring test with the standard zero/one-angle structure. It follows by Corollary \ref{cstnpi} that $(L,K)$ has relative non-positive immersion.
\end{example}
\section{Proof of Theorem \ref{thm:relNPI1}}
\begin{prop}\label{prop:heredi}
Assume that $(L,K)$ satisfies the strong relative coloring test and that $L$ has no free vertex or edge. Then $\chi(L)\le \chi(K)$. Furthermore, if there exist vertices $v_1,\ldots, v_n\in K$ such that $\kappa(v_i,L)<\kappa(v_i,K)$, then $\chi(L)\le \chi(K)-n$.
\end{prop}
\begin{proof}
Let $v$ be a vertex of $L$. We have
$$\kappa(v,L)=2-\chi(lk(v,L))-\sum \omega(c_i)=2-\chi(lk_0(v,L)),$$
where the sum is taken over all corners in the link of $v\in L$.
Assume first that $v$ is a vertex of $L$ that is not in $K$. Then every component of $lk_0(v,L)$ is a tree. We claim that there has to be more than one component. $lk(v,L)$ can not have vertices of valency 0 or 1 because $L$ does not contain a free vertex or edge. Thus a vertex of valency 0 or 1 in $lk_0(v,L)$ is the vertex of a corner with angle 1. That corner connects distinct components. Therefore $lk_0(v,L)$ has at least two components. It follows that $\chi(lk_0(v,L))\ge 2$ and therefore
$$\kappa(v,L)=2-\chi(lk_0(v,L))\le 0.$$
Next assume that $v$ is a vertex of $K$. Assume $lk_0(v,L)=C_1\cup\dots\cup C_n$, where the $C_i$ are the connected components. Assume first that $C_i\cap lk_0(v,K)=\emptyset$ for $1\le i\le k$ and $C_i\cap lk_0(v,K)\ne \emptyset$ for $k+1\le i \le n$ ($k=0$ and $k=n$ are both possible). Then $C_i$ is a tree for $1\le i\le k$, and $C_i$ is a tree relative to the connected subgraph $C_i\cap lk_0(v,K)$ for $k+1\le i \le n$. Thus $C_i$ is homotopically equivalent to $C_i\cap lk_0(v,K)$ for $k+1\le i \le n$. We have
\begin{equation}\label{eq:first}
\chi(lk_0(v,L))=k+\sum_{i=k+1}^n \chi(C_i\cap lk_0(v,K))\ge k+\chi(lk_0(v,K)).
\end{equation}
Thus
\begin{equation}\label{eq:second}
\kappa(v,L)=2-\chi(lk_0(v,L))\le 2-(\chi(lk_0(v,K))+k)=\kappa(v,K)-k.
\end{equation}
Using $\kappa(v,L)\le 0$, $v\not\in K$, and $\kappa(d,L)\le 0$, $d\not\in K$, we obtain
\begin{equation}\label{eq:third}
2\chi(L)=\sum_{v\in L} \kappa(v, L)+\sum_{d\in L} \kappa(d,L)\le \sum_{v\in K} \kappa(v,K)+\sum_{d\in K} \kappa(d,K)=2\chi(K).
\end{equation}
Note that if there exist vertices $v_1,\ldots v_n\in K$ such that $\kappa(v_i,L)<\kappa(v_i,K)$, then $\chi(L) \le \chi(K)-n$ follows immediately from the last Equation \ref{eq:third}.
\end{proof}
{\bf Proof of Theorem \ref{thm:relNPI1}}:
Let $f\colon X\to L$ be a combinatorial immersion. Assume that $X$ is finite, connected, has no free vertex or edge, and that $X$ is not a single point. Let $Y$ be the essential part of $f^{-1}(K)$. Note that $(X,Y)$ satisfies the strong relative coloring test by Proposition \ref{prop:colim}. Therefore $\chi(X)\le \chi(Y)$ by Proposition \ref{prop:heredi}. \qed
\section{A collection of lemmas needed for Theorem \ref{thm:relNPI}}
Let $(L,K)$ be a 2-complex pair. An {\em interior point $p$ of $K$} is a point so that $lk(p,L)=lk(p,K)$. Define $K^{\circ}$ to be the set of interior points and $\partial K = K-K^{\circ}$. Note that $\partial K=(L-K^{\circ})\cap K$ is a subgraph of the 1-skeleton of $K$.
If all of $K$ is essential, that is every edge is part of the attaching path of some 2-cell, then every edge in $\partial K$ is in the boundary of a 2-cell of $K$ and a 2-cell of $L-K^{\circ}$.
\pagebreak
\begin{lemma}\label{lem:val1}
Suppose $(L,K)$ satisfies the strong coloring test, $L$ has no free vertex or edge. We assume that all of $K$ is essential. Let $v$ be a vertex of $\partial K$. Assume
\begin{enumerate}
\item $v$ is an isolated vertex (valency 0) in $\partial K$, or
\item $v$ has valency 1 in $\partial K$, or
\item $lk_0(v,K)$ is connected.
\end{enumerate}
Then $lk_0(v,L)$ contains a connected component $C$ so that $C\cap lk_0(v,K)=\emptyset$. Thus, in all three cases $\kappa(v,L)<\kappa(v,K)$, and therefore we have $\chi(L)<\chi(K)$.
\end{lemma}
\begin{proof}
Let $C$ be a component of $lk_0(v,L)$ such that $C\cap lk_0(v,K)=\emptyset$. Then $\kappa(v,L)<\kappa(v,K)$. This was shown in the second paragraph of the proof of Proposition \ref{prop:colim}, see Equation \ref{eq:second}.
That this implies that $\chi(L)<\chi(K)$ is part of the same Proposition \ref{prop:colim}.\\
\noindent 1. Here we have $lk(v,L)=lk(v,K)\cup lk(v, L-K^\circ)$ and $lk(v,K)\cap lk(v, L-K^\circ)=\emptyset$. Since $v\in \partial K$ the link $lk(v, L-K^\circ)\ne \emptyset$. Let $p$ be a vertex in $lk(v,L-K^{\circ})$ and let $C$ be the component of $lk_0(v,L)$ that contains $p$. Then $C\subseteq lk(v, L-K^\circ)$ and therefore $C\cap lk_0(v,K)=\emptyset$.\\
\noindent 2. Let $e$ be the unique edge in $\partial K$ which contains $v$. Let $\{ p\} =lk(v,L)\cap e$. We have $$lk(v,L)=lk(v,K)\cup lk(v, L-K^\circ) \ \mbox{and} \ lk(v,K)\cap lk(v, L-K^\circ)=\{ p \}$$ and this implies
$$lk_0(v,L)=lk_0(v,K)\cup lk_0(v, L-K^\circ) \ \mbox{and} \ lk_0(v,K)\cap lk_0(v, L-K^\circ)=\{ p \}.$$
Thus, if $C$ is a component of $lk_0(v,L)$ so that $p\ne C$ and $C\cap lk(v, L-K^\circ)\ne \emptyset$, then $C\cap lk(v,K)=\emptyset$.
Since $e\in \partial K$, there is a corner in $lk(v,L-K^{\circ})$ with endpoints $p$ and $q$. Note that $p\ne q$ because $(L,K)$ satisfies the strong coloring test. Because $v$ has valency 1 in $\partial K$ we have $q\not\in lk(v,K)$. Let $C_p$ and $C_q$ be the components of $lk_0(v,L)$ that contain $p$ and $q$, respectively. See Figure \ref{fig:lemmalink}.
Assume first that that $C_p\ne C_q$.
We have $p\not\in C_q$ and
$C_q\cap lk(v,L-K^{\circ})\ne \emptyset$ because the intersection contains $q$. Thus $C_q\cap lk_0(v,K)=\emptyset$.
Assume next that $C_p=C_q$.
Since $C_p$ contains $q\not\in lk(v,K)$, $C_p$ is not contained in $lk_0(v,K)$. Since $C_p$ is a tree relative to $C_p\cap lk_0(v,K)$, it contains a vertex $r\not\in lk_0(v,K)$ of valency 1. Since $L$ does not have a free edge there is a corner $c$ of angle 1 with vertex $r$. Since $c\not\in lk(v,K)$ (because $r\not\in lk(v,K)$), $c$ connects $C_p$ to a different component $C$ of $lk_0(v,L)$: $C_p\cap C=\emptyset$. Here we used condition 3 in of Definition \ref{def:strong}. In particular $p\not\in C$.
Let $r$ and $s$ be the vertices of the corner $c$. Since $c\in lk(v,L-K^{\circ})$, $s\in lk(v,L-K^{\circ})$. We have $p\not\in C$ and $C\cap lk(v,L-K^{\circ})\ne \emptyset$ because the intersection contains $s$. Thus $C\cap lk_0(v,K)=\emptyset$.\\
\begin{figure}
\centering
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{lemmalink.pdf}
\label{fig:sub1}
\end{subfigure}%
\begin{subfigure}{.4\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{morelemmalink.pdf}
\label{fig:sub2}
\end{subfigure}
\caption{On the left we see a vertex $v$ of valency 1 in $\partial K$. In this case $lk(v,L-K^{\circ}\cap lk(v,K)=\{ p\}$. On the right there is the construction of a component $C$ of $lk_0(v,L)$ such that $C\cap lk_(v,K)=\emptyset$. This construction distinguishes two cases: $C_p\ne C_q$ and $C_p=C_q$.}
\label{fig:lemmalink}
\end{figure}
\noindent 3. Let $C$ be the component of $lk_0(v,L)$ that contains $lk_0(v,K)$. If there is another component $C'$ then $C'\cap lk_0(v,K)=\emptyset$ and we are done. So suppose $C$ is the only component $C=lk_0(v,L)$.
Then $lk(v,L-K^{\circ})$ can not contain corners of angle 1, because such connect distinct components. Thus $lk(v,L)=C\cup lk(v,K)$. Since $C$ is a tree relative to $lk_0(v,K)$, it follows that $lk(v,L)$ is a tree relative to $lk(v,K)$. Since $v\in \partial K$, $lk(v,K)\ne lk(v,L)$. Therefore
$lk(v,L)$ contains a vertex of valency 1, contradicting the fact that $L$ does not have a free edge.
\end{proof}
\begin{lemma}\label{lem:attach}
Suppose $\Lambda$ is a finite connected graph that is not a tree. Then 2-cells can be attached to obtain a 2-complex $K$ so that:
\begin{enumerate}
\item $K$ is collapsible;
\item every edge is in the boundary of a 2-cell;
\item $lk(v,K)$ is connected for every vertex $v$.
\end{enumerate}
\end{lemma}
\begin{proof}
We will do induction on the number of edges in $\Lambda$. If there is only one edge then $\Lambda$ is a circle and the statement is true.
Choose a maximal tree $T$ in $\Lambda$. Remove the interior of an edge $e\not\in T$ from $\Lambda$ to obtain $\Lambda'$. If $\Lambda'$ is a tree then attach a 2-cell along the path $e\gamma$ to $\Lambda $, where $\gamma$ is the path that runs around the tree, from the endpoint of $e$ to its starting point, visiting all vertices. See Figure \ref{fig:2cell}. It is easy to check that this produces a 2-complex with the desired properties.
\begin{figure}[htbp]
\centering
\includegraphics[width=2in]{treepath.pdf}
\caption{\label{fig:2cell}The 2-cell is attached along $e\gamma$, where $\gamma$ goes around a maximal tree covering every edge. Note that $\gamma$ need not be reduced.}
\end{figure}
Now assume $\Lambda'$ is not a tree. Then by induction we can attach 2-cells to $\Lambda'$ and produce a 2-complex $K'$ with the desired properties. Let $p$ and $q$ be the starting and the ending vertex of $e$ ($p=q$ is possible) respectively. We attach a 2-cell $d$ to $K'$ along the path $e\gamma$, where $\gamma$ is a path in $T$ from $q$ to $p$ that covers all of $T$. See Figure \ref{fig:2cell}. This gives a 2-complex $K$. Condition 1 is clearly satisfied. $K$ is collapsible because the last 2-cell we attached has $e$ as a free edge and $K'$ is collapsible. Lets check the links. If $v$ is a vertex different from both $p$ and $q$, then adding the 2-cell $d$ adds edges to $lk(v,K')$, but not any new vertices. Since $lk(v,K')$ is connected, so is $lk(v,K)$. Adding $d$ introduces a new vertex $e^+$ to the link at $p$, but also an edge connecting $e^+$ to $lk(p,K')$. Again, since $lk(p,K')$ is connected, so is $lk(p,K)$. The same argument also shows that $lk(q,K)$ is connected.
\end{proof}
A vertex $v$ in a 2-complex $K$ is called a {\em sink (source)} if all edges in $K$ that contain $v$ point towards (away from) $v$.
\begin{lemma}\label{lem:sink}
Let $K$ be a connected finite 2-complex all of whose 2-cells are attached along paths of exponent sum zero. Assume $\pi_1(K)=1$. Then every closed path in $K$ has exponent sum zero. Furthermore, $K$ has sink and source vertices.
\end{lemma}
\begin{proof}
Because $\pi_1(K)=1$ every closed path is, up to free reductions and expansions, of the form
$$\prod \gamma_i(\partial d_i)^{\pm 1}\gamma_i^{-1}$$
where the $d_i$ are not necessarily distinct 2-cells. Since $\partial d_i$ has exponent sum zero, the first statement follows.
Fix a vertex $v_0$. We define a map $h\colon V(K)\to \mathbb Z$, where $V(K)$ is the set of vertices of $K$, in the following way: If $v$ is a vertex of $K$ and $\gamma$ is a path from $v_0$ to $v$, let $h(v)$ be the exponent sum of $\gamma$. This is well defined by the first statement just shown. Let $v_m$ and $v_M$ be vertices where $h$ is minimal and maximal, respectively. Then $v_m$ is a source and $v_M$ is a sink.
\end{proof}
\section{Proof of Theorem \ref{thm:relNPI}}
Let $(L,K)$ be a zero/one-angled 2-complex pair so that the conditions of Theorem \ref{thm:relNPI} hold:
\begin{enumerate}
\item $L$ is a 2-complex and $K$ is a subcomplex all of whose 2-cells are attached along loops of exponent sum zero.
\item The expansion $(\hat L,\hat K)$ carries a zero/one-angle structure which is standard on $\hat K$ and which satisfies the strong relative coloring test.
\item $K$ has collapsing non-positive immersion.
\end{enumerate}
We will show that $L$ has collapsing non-positive immersion.\\
Let $f\colon X\to L$ be a combinatorial immersion. We assume $X$ does not contain free edges or vertices and is not a point. We will show that $\chi(X)\le 0$. Let $Y$ be the essential part of $f^{-1}(K)$. Give $(X,Y)$ the zero/one-angle structure induced from $(L,K)\subseteq (\hat L,\hat K)$. Since $(\hat L,\hat K)$ satisfies the strong relative coloring test, Proposition \ref{prop:colim} implies that both $(L,K)$ and $(X,Y)$ do as well.
Also note that because the immersion $f$ is combinatorial and cells in $K$ are attached along paths of exponent sum zero, cells of $Y$ are attached along paths of exponent sum zero. And since the angle structure on $K$ is standard, the angle structure on $Y$ is standard as well. Let $Y=Y_1\cup\dots \cup Y_n$ be the disjoint union of connected components. Each $f\colon Y_i\to K$ is an immersion. Since we assume that $K$ has collapsing non-positive immersion it follows that $\chi(Y_i)\le 0$ or $Y_i$ is collapsible. Thus, if no $Y_i$ is collapsible, then, by Proposition \ref{prop:heredi},
$$\chi(X)\le \chi(Y)=\sum_{i=1}^n\chi(Y_i)\le 0$$ and we are done. Assume
$$Y=Y_1\cup\ldots\cup Y_m\cup\ldots\cup Y_n \ \mbox{(connected components)}$$
where the $Y_i$, $1\le i\le m$, are collapsible, and the $Y_i$, $m+1\le i\le n$, are not. The idea now is to find vertices $v_i\in Y_i$, $1\le i\le m$ so that $\kappa(v_i,X)<\kappa(v_i,Y)$. Then, again by Proposition \ref{prop:heredi}, we have
$$\chi(X)\le \chi(Y)-m=\sum_{i=1}^m\chi(Y_i)-m+\sum_{i=m+1}^n\chi(Y_i).$$
Since $\chi(Y_i)=1$, $1\le i\le m$, and $\chi(Y_i)\le 0$, $m+1\le i\le n$, we obtain the desired result $\chi(X)\le 0$. Lemma \ref{lem:val1} will be crucial in locating the vertices $v_i$. There is one case however where this lemma is of little help: If for some $1\le i\le m$ the graph $\partial Y_i$ does not contain a vertex $v$ of valency 0 or 1, or where $lk_0(v,Y)$ is not connected. In order to avoid this case we will carefully replace the combinatorial immersion $f\colon X\to L$ by a more convenient one $f'\colon X'\to \hat L$ such that $\chi(X)\le \chi(X')$.
Assume
$$Y=Y_1\cup\ldots\cup Y_k\cup \ldots\cup Y_m\cup\ldots\cup Y_n \ \mbox{(connected components)},$$
where $Y_i$ is collapsible for $1\le i\le m$, $\partial Y_i$, $1\le i\le k$, does not contain a vertex $v$ of valency 0 or 1, but $\partial Y_j$, $k+1\le j\le m$, does.
Let $$Y_{col}=Y_1\cup\ldots\cup Y_k$$ and
$$X_{rest}=X-Y_{col}^{\circ}.$$
Assume $$\partial Y_{col}=\Delta_1\cup\ldots\cup\Delta_p \ \mbox{(connected components)}.$$
Note that none of the $\Delta_i$, $1\le i\le p$, is a tree, because we assumed that $\partial Y_i$, $1\le i\le k$, does not contain a vertex or valency 0 or 1. $X_{rest}$ might not be connected and the $\Delta_i$ are distributed over the various connected components of $X_{rest}$.
We attach 2-cells to each $\Delta_i$, $1\le i\le p$, to obtain 2-complexes $Z_i$ that satisfy the conditions of Lemma \ref{lem:attach}. Let
$$Z=Z_1\cup\ldots\cup Z_p$$ and
$X'=X_{rest}\cup Z$. Note that $Z$ is essential (meaning every edge is part of the attaching path of a 2-cell in $Z_i$). Thus every edge in $Z$ belongs to a 2-cell of $Z$ and a 2-cell of $X_{rest}$. This implies that $Z^{\circ}=\mbox{\{open 2-cells in $Z$\}}$, that the 1-skeleton of $Z$ is $\partial Z$, and that $X'$ does not have a free vertex or a free edge.
We will next extend the immersion $f|_{X_{rest}}\colon X_{rest}\to L$ to an immersion $f'\colon X'\to \hat L$. We have already mentioned that cells in $Y$ are attached along paths of exponent sum zero. Since the components of $Y_{col}$ are collapsible, every closed path in $Y_{col}$ has exponent sum zero. In particular every closed path in $\partial Y_{col}$ has exponent sum zero. Let $d$ be a 2-cell in $Z_i$ and let $\gamma$ be the attaching path of $d$. Note that $\gamma\subseteq \partial Y_{col}$. Then $f(\gamma)\subseteq K$ is a closed path of exponent sum zero, and therefore there exists a 2-cell $\hat d$ in $\hat K$ with attaching path $f(\gamma)$. Define $f'(d)=\hat d$. This gives the desired extended immersion.
Note that the essential part of $f'^{-1}(\hat K)$ is
$$W'=Z_1\cup\ldots Z_p\cup Y_{k+1}\cup\ldots\cup Y_m\cup\ldots\cup Y_n.$$
Since $(\hat L,\hat K)$ satisfies the strong relative coloring test, it follows from Proposition \ref{prop:colim} that $(X',W')$ does so as well. And since the angle structure on $\hat K$ is standard, so is the induced one on $W'$.
Since $Z_i$, $1\le i\le p$, is contractible and all 2-cells are attached along loops of exponent sum zero, there are sink and source vertices in $Z_i$ by Lemma \ref{lem:sink}. Since the angle structure on $Z_i$ is standard, this implies that there is a vertex $v_i\in Z_i$ so that $lk(v_i,Z_i)=lk_0(v_i,Z_i)$. Since $lk(v_i,Z_i)$ is connected by construction this shows that $lk_0(v_i,Z_i)=lk_0(v_i,W')$ is connected. The last equation is true because
$$W'=Z_1\cup\ldots Z_p\cup Y_{k+1}\cup\ldots Y_m\cup\ldots\cup Y_n$$
is a disjoint union.
We can now use Lemma \ref{lem:val1} to conclude that $\kappa(v_i,X')<\kappa(v_i,W')$, $1\le i\le p$.
We know that $Y_i$, $k+1\le i\le m$, is collapsible and $\partial Y_i$ contains a vertex $v_i$ of valency 0 or 1. By Lemma \ref{lem:val1} it follows that $\kappa(v_i,X')< \kappa(v_i,W')$. We have located vertices $v_1,\ldots, v_p,\ldots ,v_m$ for which $\kappa(v_i,X')< \kappa(v_i,W')$.
Thus, by Proposition \ref{prop:heredi} we have
$$\chi(X')\le \chi(W')-m,$$
and since each $Z_i$, $1\le i\le p$, and each $Y_i$, $k+1\le i \le m$, is collapsible and therefore has Euler characteristic 1, we have
$$\chi(X')\le \chi(W')-m=\chi(Y_{m+1})+\ldots +\chi(Y_n).$$ Since $Y_i$, $m+1\le i\le n$ is not collapsible and $f\colon Y_i\to K$ is a combinatorial immersion and, furthermore, $K$ has collapsing nonpositive immersion, it follows that $\chi(Y_i)\le 0$. In summary we have
$$\chi(X')\le 0.$$
We have left to show that $$\chi(X)\le \chi(X').$$
$$\chi(X)=\chi(X-Y_{col}^{\circ})+\chi(Y_{col})-
\chi(\partial Y_{col})$$
$$\chi(X')=\chi(X'-Z^{\circ})+\chi(Z)-\chi(\partial Z).$$
Note that $$X'=X_{rest}\cup \mbox{\{ 2-cells of $Z$\}}.$$
As observed above, $Z^{\circ}$ are exactly the open 2-cells in $Z$.
Thus $X'-Z^{\circ}=X_{rest}$. So we have
$$X-Y_{col}^{\circ}=X'-Z^{\circ}=X_{rest}$$ and also
$$\partial Y_{col}=\Delta_1\cup\ldots\cup \Delta_p=\partial Z.$$
Thus
$$\chi(X)=\chi(X_{rest})+\chi(Y_{col})-(\chi(\Delta_1)+\ldots +\chi(\Delta_p))$$
$$\chi(X')=\chi(X_{rest})+\chi(Z)-(\chi(\Delta_1)+\ldots +\chi(\Delta_p)).$$ We obtain
$$\chi(X)-\chi(X')=\chi(Y_{col})-\chi(Z).$$
Recall that
$$Y_{col}=Y_1\cup\ldots \cup Y_k$$
and the $Y_i$, $1\le i\le k$, are the collapsible components of $Y$ so that $\partial Y_i$ does not contain a vertex of valency 0 or 1,
and
$$Z=Z_1\cup\ldots \cup Z_p \mbox{ (collapsible connected components)}.$$
Therefore $\chi(Y_{col})=k$ and $\chi(Z)=p$. On one hand we have
$$\partial Y_{col}=\partial Y_1\cup\ldots\cup \partial Y_k$$
and on the other
$$\partial Y_{col}=\Delta_1\cup\ldots\cup \Delta_p \ \mbox{(connected components)}.$$
Thus we have $k\le p$, and it follows that
$$\chi(Y_{col})-\chi(Z)= k-p\le 0.$$ In summary
$$\chi(X)\le \chi(X')\le 0.$$ \qed
\begin{remark}
In the above proof we have found vertices $v_i\in \Delta_i=\partial Z_i$ so that $\kappa(v_i,X')<\kappa(v_i,W')$. Note that since $\partial Z\subseteq \partial Y$, these vertices are also contained in $Y$. Here we remark that for these vertices we also have $\kappa(v_i,X)<\kappa(v_i,Y)$. Note that $lk_0(v_i,X')$ is obtained from $lk_0(v_i,X)$ by the following process: The interior of $lk_0(v_i,Y)$ is removed and replaced with the connected graph $lk_0(v_i,W')$. Since $lk_0(v_i,X')$ contains a connected component $C$ such that $C\cap lk_0(v_i,W')=\emptyset$, $C$ is also a component of $lk_0(v_i,X)$ such that $C\cap lk_0(v_i,Y)=\emptyset$. It follows that $\kappa(v_i,X)<\kappa(v_i,Y)$. This argument shows that every collapsible component of $Y$ contains a vertex $v$ such that $\kappa(v,X)<\kappa(v,Y)$. This provides an alternative proof for the fact that $\chi(X)\le 0$.
\end{remark}
\section{Labeled oriented trees}
A labeled oriented graph (LOG) $\Gamma = (E, V, s, t, \lambda)$ consists of two sets $E$, $V$ of edges and vertices, and three maps $s, t, \lambda\colon E\to V$ called, respectively source, target and label. $\Gamma$ is said to be a labeled oriented tree (LOT) when the underlying graph is a tree. The associated LOG presentation is defined as
$$P(\Gamma)=\langle V\ |\ s(e)\lambda(e)=\lambda(e)t(e),\ e\in E \rangle.$$
The LOG complex $K(\Gamma)$ is the standard 2-complex defined by the presentation, and the group $G(\Gamma)$ presented by $P(\Gamma )$ is equal to $\pi_1(K(\Gamma))$.
It is known that LOT-complexes are spines of ribbon 2-knot complements. See Howie \cite{How83}. So the study of LOTs is an extension of classical knot theory. Asphericity, known for classical knots, is unresolved for LOTs. The asphericity question for LOTs is of central importance to Whitehead's asphericity conjecture: A subcomplex of an aspherical 2-complex is aspherical. See Berrick/Hillman \cite{BerrickHillman}, Bogley \cite{Bogley}, and Rosebrock \cite{Ro18}.
A sub-LOG $\Gamma_0=(E_0, V_0)\subseteq\Gamma$ is a subgraph so that $E_0\ne\emptyset$ and $\lambda\colon E_0\to V_0$. A LOG is called {\em boundary reduced} if whenever $v$ is a vertex of valency 1 then $v=\lambda(e)$ for some edge $e$. It is called {\em interior reduced} if for every vertex $v$ no two edges starting or terminating at $v$ carry the same label. It is called {\em compressed }if for every edge $e$ the label $\lambda(e)$ is not equal to $s(e)$ or $t(e)$. Finally, a LOG is {\em reduced} if it is boundary reduced, interior reduced, and compressed. Given a LOG, reductions can be performed to produce a reduced LOG, and, in case the LOG is a LOT, this process does not affect the homotopy type of the LOT complex. A LOG is called {\em injective} if the labeling map $\lambda\colon E\to V$ is injective.
We collect some known facts about injective LOTs (see\\
Harlander/Rosebrock \cite{HarRose2022})
\begin{itemize}
\item[F1] (see Theorem 3.2 and Theorem 3.3 of \cite{HarRose2022}) If $\Gamma$ is a reduced injective LOT that does not contain boundary reducible sub-LOTs, then $K(\Gamma)$ admits a zero/one-angle structure that satisfies the coloring test. It follows that $K(\Gamma)$ is DR (and hence aspherical), has collapsing non-positive immersion, and $G(\Gamma)$ is locally indicable.
\item[F2] (see Theorem 8.4 of \cite{HarRose2022}) Suppose $\Gamma$ is reduced and injective. Assume that $\Gamma_1,\ldots, \Gamma_n$ are disjoint sub-LOTs and that collapsing each into one of its vertices produces a quotient LOT $\bar\Gamma$ without boundary reducible sub-LOTs. Then $K(\Gamma)$ admits a zero/one-angle structure that satisfies a coloring test relative to $K(\Gamma_1),\ldots ;K(\Gamma_n)$. (This relative coloring test agrees with the one defined in this paper only in case $n=1$).
\item[F3] (see Theorem 8.5 of \cite{HarRose2022}) If $\Gamma$ is reduced and injective, then $K(\Gamma)$ is VA and hence aspherical. It follows that if $\Gamma$ is an injective LOT, reduced or not, then $K(\Gamma)$ is aspherical.
\end{itemize}
We believe that if $\Gamma$ is a reduced injective LOT, then $K(\Gamma)$ has collapsing non-positive immersion. We outline a strategy for proving this. $\Gamma$ contains disjoint maximal sub-LOTs $\Gamma_1,\ldots, \Gamma_n$ so that identifying each into one of its vertices produces a quotient LOT $\bar\Gamma$ without boundary reducible sub-LOTs. Cases that differ from this scenario typically yield to ad hoc considerations. We may assume by induction on the number of vertices that each $K(\Gamma_i)$ has collapsing non-positive immersion. If $n=1$ the methods developed in this paper apply and we obtain a positive result. See Theorem \ref{thm:lotnpi} below. If $n>1$ then we are in the situation of the second point F2 mentioned above. $K(\Gamma)$ does satisfy a relative coloring test, but not the version presented in this paper, and certainly not the strong relative coloring test. Thus the methods developed here fall short in establishing non-positive immersion for all reduced injective LOT complexes $K(\Gamma)$. We do think that our methods can be strengthened to give this result. This is a topic for future work.
\begin{thm}\label{thm:lotnpi}
Suppose $\Gamma$ is a reduced injective LOT that contains a sub-LOT $\Gamma_1$ so that
\begin{enumerate}
\item $K(\Gamma_1)$ has collapsing non-positive immersion;
\item identifying $\Gamma_1$ to one of its vertices produces a reduced quotient LOT $\bar\Gamma$ without boundary reducible sub-LOTs.
\end{enumerate} Then $K(\Gamma)$ has collapsing non-positive immersion.
\end{thm}
\begin{proof}
Let $L=K(\Gamma)$ and $K=K(\Gamma_1)$. Assume that $\Gamma_1$ is collapsed to the vertex $y$. Then $K$ can be folded to the edge $y$ which produces $\bar L=K(\bar\Gamma)$. We will show that the result follows from Theorem \ref{thm:mainapp}. $\bar L$ satisfies the coloring test by fact F1 stated above. In order to show that $y^+$ and $y^-$ lie in different components of $lk_0(\Bar L)$ we need to take a closer look at that link. It was shown in \cite{HarRose2022}, Theorem 3.3, that $\bar L=K(\bar\Gamma)$ has the following local bi-forest property: If $x_1,\ldots, x_n$ are the edges of $\bar L$, then there exists a choice of $\epsilon_i\in \{ +,-\}$ so that
$\Lambda_1=\Lambda(x_1^{\epsilon_1},\ldots, x_n^{\epsilon_n})$ and $\Lambda_2=\Lambda(x_1^{-\epsilon_1},\ldots, x_n^{-\epsilon_n})$ are forests. Here $\Lambda(x_1^{\epsilon_1},\ldots, x_n^{\epsilon_n})$ is the subgraph of $lk(\bar L)=\Lambda$ spanned by the vertices $x_1^{\epsilon_1},\ldots, x_n^{\epsilon_n}$. Furthermore, a zero/one-angle structure can be put on $\bar L$ (see \cite{HarRose2022}, Theorem 3.2) so that
$$lk_0(\bar L)=\Lambda_1\cup\Lambda_2.$$ Since $\Lambda_1$ and $\Lambda_2$ are disjoint, it follows that $y^+$ and $y^-$, the $y$ being one of the $x_i$, lie in different components of $lk_0(\bar L)$.
Theorem \ref{thm:mainapp} now gives the result.
\end{proof}
The {\em core} of a LOT $\Gamma$ is the sub-LOT obtained when performing all boundary reductions on $\Gamma$. Note that if the core of $\Gamma_1$ in Theorem \ref{thm:lotnpi} does not contain boundary reducible sub-Lots, then it follows from Lemma \ref{lem:core} and the above fact F1 that $K(\Gamma_1)$ does have collapsing non-positive immersion.
\begin{lemma}\label{lem:core}
Let $\Gamma$ be a LOT and let $\Gamma_c$ be its core. If $K(\Gamma_c)$ has collapsing non-positive immersion, then so does $K(\Gamma)$.
\end{lemma}
\begin{proof}
Let $f\colon X\to K(\Gamma)$ be a combinatorial immersion. Assume that $X$ has no free vertices or edges and is not a point. Then $f$ maps into $K(\Gamma_c)$, because otherwise it would have a free edge. It follows that $\chi(X)\le 0$
\end{proof}
|
\section{Introduction and a result}
We investigate one-dimensional Schr\"{o}dinger operators with sparse potentials.
It is known that the spectrum of Schr\"{o}dinger operators with sparse potential consists of singular continuous spectrum.
Simon and Spencer $\cite{SimonSpencer}$ show the absence of absolutely continuous spectrum of Schr\"odinger operators with sparse potentials.
Simon and Stoltz $\cite{Simonsparse}$ also show that $-\frac{d^2}{dx^2}f+Vf=Ef$ has no $L^2$-solutions for any $E>0$.
We have the question whether the edge of the singular continuous spectrum is an eigenvalue or not.
We give a sufficient condition for the absence of embedded eigenvalues and give examples.
\begin{Definition}
A function $V:[0,\infty)\rightarrow \mathbb{R}$ is called a sparse potential, if there exist positive sequences $\{x_n\}_{n=1}^{\infty}, \{\alpha_n\}_{n=1}^{\infty}$ and $\{h_n\}_{n=1}^{\infty}$ such that $x_{n+1}>x_n$ for $n=1,2,3,...$,
\begin{enumerate}[$($\rm i$)$]
\item
$\begin{displaystyle}\lim_{n\rightarrow \infty}\frac{x_{n+1}-x_n}{\alpha_{n+1}+\alpha_n+1}=\infty\end{displaystyle}$,
\item
$|V(x)|\leq h_n$, if $x \in [x_n-\alpha_n,x_n+\alpha_n]$ for $n=1,2,3,...$,
\item
$V(x)=0$, if $\begin{displaystyle} x \in \left(\bigcup_{n=1}^{\infty} [x_n-\alpha_n,x_n+\alpha_n] \right)^c\end{displaystyle}$
\end{enumerate}
\end{Definition}
We define $L_n=x_{n+1}-x_n-\alpha_{n+1}-\alpha_n$ for $n\geq1$ and $L_0=x_1-\alpha_1$.
By $\rm(i)$, $L_n\rightarrow \infty$ as $n\rightarrow \infty$.
By Strum-Liouville theory $\cite[{\rm Theorem}\: 9.1.]{Strum}$, there exists a unique solution $f\in AC_{loc}([0,\infty))$ of the equation $-\frac{d^2}{dx^2}f+Vf=0$ with $\frac{d}{dx}f \in AC_{loc}([0,\infty))$ and the boundary condition\\
$f(0)=\alpha,\frac{d}{dx}f(0)=\beta,\alpha,\beta\in\mathbb{C}$.
We give a sufficient condition of the existence of a non $L^2$-integrable solution.
\begin{Theorem}\label{main}
Let $V$ be the sparse potential and $f$ a weak solution of $-\frac{d^2}{dx^2}f+Vf=0$.
If
\begin{eqnarray}\label{assumption}
\cfrac{L_n}{4^n}
\left(
\prod_{m=1}^{n}(L_{m-1}^2+2)
\right)^{-1}
\left(
\prod_{m=1}^{n}(2\alpha_m^2+1)
\right)^{-1}
\exp
\left(-\frac{2}{3}\sum_{m=1}^{n}h_m(4\alpha_m^3+3\alpha_m)
\right)
\rightarrow \infty,
\end{eqnarray}
as $n\rightarrow \infty$, then $f\notin L^2([0,\infty))$.
\end{Theorem}
We give an example for one-dimensional Schr\"{o}dinger operators with singular continuous spectrum which has no embedded eigenvalues.
Let $x_n=\exp(n^n)$ for $n=1,2,3,...$, and
\begin{eqnarray*}
V(x)=
\begin{cases}
e^{n},&\text{if $|x-x_n|\leq\frac{1}{2}$ for $n=1,2,3,...$},\\
0,&\text{otherwise}.
\end{cases}
\end{eqnarray*}
Let $H=-\frac{d^2}{dx^2}+V:L^2([0,\infty))\rightarrow L^2([0,\infty))$ with the domain $D(H)=C^{\infty}_0((0,\infty))$.
We see that $H$ is regular at zero and in the limit point case at infinity.
This implies that $H$ has self-adjoint extensions $H_\theta$ which can be parametrized by boundary conditions.
Hence, $H_{\theta}$ is the restriction of $H^*$ to $D_{\theta}=\{f\in D(H^*)\mid {f(0)}\sin\theta-{\frac{d}{dx}f(0)}\cos\theta=0\}$.
By $\cite{Simonsparse}$, we have $\sigma_{sc}(H_\theta)=[0,\infty)$, $\sigma_{pp}(H_\theta)\cap(0,\infty)=\emptyset$ and $\sigma_{ac}(H_\theta)=\emptyset$ for all $\theta\in (\frac{\pi}{2},\frac{\pi}{2}]$.
See Appendix A for the proof.
Theorem $\ref{main}$ implies the next corollary.
This also implies that $H_\theta$ has purely singular continuous spectrum for some $\theta$.
\begin{Corollary}
It follows that
\begin{enumerate}[$(1)$]
\item
$\sigma_{sc}(H_\theta)=[0,\infty)$, $\sigma_{pp}(H_\theta)\cap[0,\infty)=\emptyset$ and $\sigma_{ac}(H_\theta)=\emptyset$ for all $\theta\in(-\frac{\pi}{2},\frac{\pi}{2}]$,
\item
$H_{\theta}$ has purely singular continuous spectrum for $\theta\in[0,\frac{\pi}{2}]$,
\item
$H_{\theta}$ has a single negative eigenvalue for
$\theta\in (-\frac{\pi}{2},\arctan(-\frac{1+\sqrt{3}}{2})]$.
\end{enumerate}
\end{Corollary}
\section{Proof of Theorem 1.2.}
We calculate a lower bound of Wronskian matrices.
\begin{comment}
Let $
\left|
\left(
\begin{array}{c}
a\\
b
\end{array}
\right)
\right|=\sqrt{a^2+b^2}$ and \\
$
\left(
\begin{array}{c}
a\\
b
\end{array}
\right)\cdot
\left(
\begin{array}{c}
c\\
d
\end{array}
\right)
=ac+bd$ for $a,b,c,d\in \mathbb{R}$.
\end{comment}
For a $2\times2$-real matrix $M$, let
\[
\low M=
\inf
\left\{
\left|
M
\begin{pmatrix}
\cos \theta\\
\sin \theta
\end{pmatrix}
\right| \:
\middle|\:
\theta \in [0,2\pi)
\right\}.
\]
We see that $|Mu|\geq \low M |u|$ for $u \in \mathbb{R}^2$ and $\low M=\sqrt{\inf\sigma({}^t\!MM)}$.
Let $V$ be a sparse potential and $f\in AC_{loc}([0,\infty))$ be a weak solution of $-\frac{d^2}{dx^2}f+Vf=0$ with $\frac{d}{dx}f\in AC_{loc}([0,\infty))$ and the boundary condition $f(0)=\cos\theta,\frac{d}{dx}f(0)=\sin\theta$.
We can represent the weak solution concretely as follows.
Let $J_0=[0,x_1-\alpha_1]$, $I_n=[x_n-\alpha_n,x_n+\alpha_n]$ and $J_n=[x_n+\alpha_n,x_{n+1}-\alpha_{n+1}]$ for $n=1,2,...$.
Let $p_n, q_n :J_n \rightarrow \mathbb{R}$ be defined by $p_n(x)=1, q_n(x)=x-x_n-\alpha_n$, and $p_0, q_0 :J_0 \rightarrow \mathbb{R}$ by $p_0(x)=1, q_0(x)=x$.
Let $\varphi_n,\psi_n:I_n\rightarrow \mathbb{R}$ be the weak solution of $-\frac{d^2}{dx^2}f+Vf=0$ on $I_n$ with the boundary condition $\varphi_n(x_n-\alpha_n)=1,\frac{d}{dx}\varphi_n(x_n-\alpha_n)=0, \psi_n(x_n-\alpha_n)=0$ and $\frac{d}{dx}\psi_n(x_n-\alpha_n)=1$.
We define $\tilde{f}:[0,\infty)\rightarrow \mathbb{R}$ by
\begin{eqnarray*}
\tilde{f}(x)=c_n^{(1)}p_n(x)+c_n^{(2)}q_n(x) \qquad \text{if $x\in J_n$},\\
\tilde{f}(x)=d_n^{(1)}\varphi_n(x)+d_n^{(2)}\psi_n(x)\qquad \text{if $x\in I_n$},
\end{eqnarray*}
where $c_n^{(1)},c_n^{(2)},d_n^{(1)}$ and $d_n^{(2)}$ are inductively determined by for $n=1,2,...$
\begin{eqnarray}\nonumber
\lim_{x\downarrow 0}
\begin{pmatrix}
p_{0}(x)&q_{0}(x)\\
\frac{d}{dx}p_{0}(x)&\frac{d}{dx}q_{0}(x)
\end{pmatrix}
\begin{pmatrix}
c_0^{(1)}\\
c_0^{(2)}
\end{pmatrix}
&=&
\begin{pmatrix}
\cos \theta \\
\sin \theta
\end{pmatrix},
\\\label{c1}
\lim_{x\uparrow x_n-\alpha_n}
\begin{pmatrix}
p_{n-1}(x)&q_{n-1}(x)\\
\frac{d}{dx}p_{n-1}(x)&\frac{d}{dx}q_{n-1}(x)
\end{pmatrix}
\begin{pmatrix}
c_{n-1}^{(1)}\\
c_{n-1}^{(2)}
\end{pmatrix}
&=&
\lim_{x\downarrow x_n-\alpha_n}
\begin{pmatrix}
\varphi_{n}(x)&\psi_{n}(x)\\
\frac{d}{dx}\varphi_{n}(x)&\frac{d}{dx}\psi_{n}(x)
\end{pmatrix}
\begin{pmatrix}
d_{n}^{(1)}\\
d_{n}^{(2)}
\end{pmatrix},
\\\label{c2}
\lim_{x\uparrow x_n+\alpha_n}
\begin{pmatrix}
\varphi_{n}(x)&\psi_{n}(x)\\
\frac{d}{dx}\varphi_{n}(x)&\frac{d}{dx}\psi_{n}(x)
\end{pmatrix}
\begin{pmatrix}
d_{n}^{(1)}\\
d_{n}^{(2)}
\end{pmatrix}
&=&
\lim_{x\downarrow x_n+\alpha_n}
\begin{pmatrix}
p_{n}(x)&q_{n}(x)\\
\frac{d}{dx}p_{n}(x)&\frac{d}{dx}q_{n}(x)
\end{pmatrix}
\begin{pmatrix}
c_{n}^{(1)}\\
c_{n}^{(2)}
\end{pmatrix}
.
\end{eqnarray}
By the definition, $\tilde{f}$ and $\frac{d}{dx}\tilde{f}$ are continuous, and
$\left(
\begin{array}{c}
\tilde{f}(0)\\
\frac{d}{dx}\tilde{f}(0)
\end{array}
\right)
=
\left(
\begin{array}{c}
\cos \theta\\
\sin \theta
\end{array}
\right)$.
The coefficients $c_n^{(1)},c_n^{(2)},d_n^{(1)}$ and $d_n^{(2)}$ are uniquely determined by $\theta$.
We see that
\[
\int_{0}^{\infty}\tilde{f}(x)\frac{d^2}{dx^2}g(x)dx=\int_{0}^{\infty}\tilde{f}(x)V(x)g(x)dx
\]
for $g \in C_0^{\infty}((0,\infty))$ straightforwardly.
This implies that $\tilde{f}$ satisfies $-\frac{d^2}{dx^2}f+Vf=0$ in the sence of the weak derivative.
Therefore $f=\tilde{f}$ by the uniqueness of the solution.
\begin{Lemma}
Let $c_n=
\begin{pmatrix}
c_n^{(1)}\\
c_n^{(2)}
\end{pmatrix}
.$
Then we see that
\begin{eqnarray}\label{lower bound of L^2 integral}
\int_{J_n}|f(x)^2|dx
\geq
\frac{1}{4}
\frac{L_n^4}{L_n^3+3L_n}|c_n|^2.
\end{eqnarray}
\end{Lemma}
\begin{Proof}\rm
We obtain
\begin{eqnarray*}
\int_{J_n}|f(x)|^2dx&=&\int_{J_n}|c_n^{(1)}p_n(x)+c_n^{(2)}q_n(x)|^2dx
\\
&=&
\begin{pmatrix}
c_n^{(1)}&
c_n^{(2)}
\end{pmatrix}
\begin{pmatrix}
L_n&\frac{1}{2}L_n^2\\
\frac{1}{2}L_n^2&\frac{1}{3}L_n^3
\end{pmatrix}
\begin{pmatrix}
c_n^{(1)}\\
c_n^{(2)}
\end{pmatrix}.
\end{eqnarray*}
The matrix
$
\begin{pmatrix}
L_n&\frac{1}{2}L_n^2\\
\frac{1}{2}L_n^2&\frac{1}{3}L_n^3
\end{pmatrix}$ can be diagonalized and its eigenvalues $\lambda_{\pm}$ are
\[
\lambda_\pm=\cfrac{1}{2}
\left(
L_n+\frac{1}{3}L_n^3
\pm
\sqrt{(L_n+\frac{1}{3}L_n^3)^2-\frac{1}{3}L_n^4}
\right).
\]
Since $1-\sqrt{1-t}\geq\frac{1}{2}t$ for $0<t<1$, we have
\begin{eqnarray}\nonumber
\lambda_-=
\cfrac{L_n+\frac{1}{3}L_n^3}{2}
\left(1
-\sqrt{1-\frac{L_n^4}{3}\left(L_n+\frac{1}{3}L_n^3\right)^{-2}}
\right)
\geq
\frac{1}{4}
\frac{L_n^4}{L_n^3+3L_n}.
\end{eqnarray}
This implies our assertion.
\qed
\end{Proof}
By $(\ref{c1})$ and $(\ref{c2})$, we obtain $c_n=R_nW_{n-1}c_{n-1}$ for $n\geq1$, where
\begin{eqnarray*}
R_m&=&\lim_{x\uparrow x_m+\alpha_m}
\begin{pmatrix}
\varphi_{m}(x)&\psi_{m}(x)\\
\frac{d}{dx}\varphi_{m}(x)&\frac{d}{dx}\psi_{m}(x)
\end{pmatrix},
\\
W_m&=&
\lim_{x\uparrow x_{m+1}-\alpha_{m+1}}
\begin{pmatrix}
p_{m}(x)&q_{m}(x)\\
\frac{d}{dx}p_{m}(x)&\frac{d}{dx}q_{m}(x)
\end{pmatrix}
=
\begin{pmatrix}
1&L_m\\
0&1
\end{pmatrix}
.
\end{eqnarray*}
We shall estimate $\low R_m$ and $\low W_m$.
\begin{Lemma}
It follows that
\begin{eqnarray}\label{lower bound of W_m}
\low W_m\geq \frac{1}{\sqrt{L_m^2+2}}.
\end{eqnarray}
\end{Lemma}
\begin{Proof}\rm
The eigenvalues $\xi_{\pm}$ of ${}^tW_mW_m$ are
\[
\xi_{\pm}=
\frac{1}{2}\left(
L_m^2+2\pm
\sqrt{(L_m^2+2)^2-4}
\right).
\]
Since $1-\sqrt{1-t}\geq\frac{1}{2}\:t$ for $0<t<1$, we have
\begin{eqnarray*}
\xi_{-}=\frac{L_m^2+2}{2}\left(
1-
\sqrt{1-4(L_m^2+2)^{-2}}
\right)
\geq
\frac{1}{L_m^2+2}.
\end{eqnarray*}
Since $\low M=\sqrt{\inf\sigma({}^t\!MM)}$, we have our assertion.
\qed
\end{Proof}
Let $\tilde{\varphi}_m,\tilde{\psi}_m:I_m\rightarrow \mathbb{R}$ be defined by $\tilde{\varphi}_m(x)=1,\tilde{\psi}_m(x)=x-x_m+\alpha_m$.
We see that $\tilde{\varphi}_m,\tilde{\psi}_m$ satisfy $-\frac{d^2}{dx^2}f=0.$
There exist $u_m^{(j)},v_m^{(j)}\in AC(I_m)$, $j=1,2$ such that
\begin{eqnarray*}
\begin{pmatrix}
\varphi_m(x)\\
\frac{d}{dx}\varphi_m(x)
\end{pmatrix}
=
u_m^{(1)}(x)
\begin{pmatrix}
\tilde{\varphi}_m(x)\\
\frac{d}{dx}\tilde{\varphi}_m(x)
\end{pmatrix}
+
u_m^{(2)}(x)
\begin{pmatrix}
\tilde{\psi}_m(x)\\
\frac{d}{dx}\tilde{\psi}_m(x)
\end{pmatrix},
\\
\begin{pmatrix}
\psi_m(x)\\
\frac{d}{dx}\psi_m(x)
\end{pmatrix}
=
v_m^{(1)}(x)
\begin{pmatrix}
\tilde{\varphi}_m(x)\\
\frac{d}{dx}\tilde{\varphi}_m(x)
\end{pmatrix}
+
v_m^{(2)}(x)
\begin{pmatrix}
\tilde{\psi}_m(x)\\
\frac{d}{dx}\tilde{\psi}_m(x)
\end{pmatrix}.
\end{eqnarray*}
We see that $u_m^{(1)}(x_m-\alpha_m)=1,u_m^{(2)}(x_m-\alpha_m)=0$,
$v_m^{(1)}(x_m-\alpha_m)=0,v_m^{(2)}(x_m-\alpha_m)=1$, and
\begin{eqnarray}\label{derivative of constants}
\begin{pmatrix}
\varphi_{m}&\psi_{m}\\
\frac{d}{dx}\varphi_{m}&\frac{d}{dx}\psi_{m}
\end{pmatrix}
=
\begin{pmatrix}
\tilde{\varphi}_{m}&\tilde{\psi}_{m}\\
\frac{d}{dx}\tilde{\varphi}_{m}&\frac{d}{dx}\tilde{\psi}_{m}
\end{pmatrix}
\begin{pmatrix}
u_{m}^{(1)}&v_{m}^{(1)}\\
u_m^{(2)}&v_{m}^{(2)}
\end{pmatrix}
.
\end{eqnarray}
Note that $\varphi_n$ and $\psi_n$ satisfy the equation $-\frac{d^2}{dx^2}f+Vf=0$ which is equivalent to
\[
\begin{pmatrix}
\frac{d}{dx}f\\
\frac{d^2}{dx^2}f
\end{pmatrix}
=
\begin{pmatrix}
0&1\\
V&0
\end{pmatrix}
\begin{pmatrix}
f\\
\frac{d}{dx}f
\end{pmatrix}.
\]
Differentiating both sides of $(\ref{derivative of constants})$, we obtain
\begin{eqnarray*}
\begin{pmatrix}
0&0\\
V&0
\end{pmatrix}
\begin{pmatrix}
\tilde{\varphi}_{m}&\tilde{\psi}_{m}\\
\frac{d}{dx}\tilde{\varphi}_{m}&\frac{d}{dx}\tilde{\psi}_{m}
\end{pmatrix}
\begin{pmatrix}
u_{m}^{(1)}&v_{m}^{(1)}\\
u_m^{(2)}&v_{m}^{(2)}
\end{pmatrix}
=
\begin{pmatrix}
\tilde{\varphi}_{m}&\tilde{\psi}_{m}\\
\frac{d}{dx}\tilde{\varphi}_{m}&\frac{d}{dx}\tilde{\psi}_{m}
\end{pmatrix}
\begin{pmatrix}
\frac{d}{dx}u_{m}^{(1)}&\frac{d}{dx}v_{m}^{(1)}\\
\frac{d}{dx}u_m^{(2)}&\frac{d}{dx}v_{m}^{(2)}
\end{pmatrix}.
\end{eqnarray*}
Thus we have
\begin{eqnarray}\label{deribative of constants2}
\begin{pmatrix}
\frac{d}{dx}u_{m}^{(1)}&\frac{d}{dx}v_{m}^{(1)}\\
\frac{d}{dx}u_m^{(2)}&\frac{d}{dx}v_{m}^{(2)}
\end{pmatrix}
=
-V
\begin{pmatrix}
\tilde{\varphi}_{m}\tilde{\psi}_{m}&\tilde{\psi}_{m}^2\\
-\tilde{\varphi}_{m}^2&-\tilde{\varphi}_{m}\tilde{\psi}_{m}
\end{pmatrix}
\begin{pmatrix}
u_{m}^{(1)}&v_{m}^{(1)}\\
u_m^{(2)}&v_{m}^{(2)}
\end{pmatrix}.
\end{eqnarray}
Let $u_m=
\begin{pmatrix}
u_m^{(1)}\\
u_m^{(2)}
\end{pmatrix}$ and $v_m=
\begin{pmatrix}
v_m^{(1)}\\
v_m^{(2)}
\end{pmatrix}$.
By $(\ref{deribative of constants2})$, we see that $u_m$ and $v_m$ satisfy that
\begin{eqnarray}\label{derivative of constants3}
\frac{d}{dx}u_m&=&
-V
\begin{pmatrix}
\tilde{\varphi}_{m}\tilde{\psi}_{m}&\tilde{\psi}_{m}^2\\
-\tilde{\varphi}_{m}^2&-\tilde{\varphi}_{m}\tilde{\psi}_{m}
\end{pmatrix}
u_m,
\\
\label{derivative of constants4}
\frac{d}{dx}v_m&=&
-V
\begin{pmatrix}
\tilde{\varphi}_{m}\tilde{\psi}_{m}&\tilde{\psi}_{m}^2\\
-\tilde{\varphi}_{m}^2&-\tilde{\varphi}_{m}\tilde{\psi}_{m}
\end{pmatrix}
v_m.
\end{eqnarray}
\begin{Lemma}\label{wronskian constancy}
For $x\in I_m$,
$u_{m}^{(1)}(x)v_{m}^{(2)}(x)-v_{m}^{(1)}(x)u_m^{(2)}(x)=1$.
\end{Lemma}
\begin{Proof}\rm
By $(\ref{derivative of constants3})$ and $(\ref{derivative of constants4})$, we obtain
\begin{eqnarray}\nonumber
\frac{d}{dx}(u_{m}^{(1)}v_{m}^{(2)}-v_{m}^{(1)}u_m^{(2)})&=&
\frac{d}{dx}u_{m}^{(1)}v_{m}^{(2)}+u_{m}^{(1)}\frac{d}{dx}v_{m}^{(2)}-\frac{d}{dx}v_{m}^{(1)}u_m^{(2)}-v_{m}^{(1)}\frac{d}{dx}u_m^{(2)}
\\
&=&\nonumber
-V(\tilde{\varphi}_{m}\tilde{\psi}_{m}u_m^{(1)}+\tilde{\psi}_{m}^2u_m^{(2)})v_{m}^{(2)}
+u_{m}^{(1)}V(\tilde{\varphi}_{m}^2v_m^{(1)}+\tilde{\varphi}_{m}\tilde{\psi}_{m}v_m^{(2)})
\\
&&\nonumber
+V(\tilde{\varphi}_{m}\tilde{\psi}_{m}v_m^{(1)}+\tilde{\psi}_{m}^2v_m^{(2)})u_m^{(2)}
-v_{m}^{(1)}V(\tilde{\varphi}_{m}^2u_m^{(1)}+\tilde{\varphi}_{m}\tilde{\psi}_{m}u_m^{(2)})
\\
&=&0.\nonumber
\end{eqnarray}
Since $u_{m}^{(1)}(x_m-\alpha_m)v_{m}^{(2)}(x_m-\alpha_m)-v_{m}^{(1)}(x_m-\alpha_m)u_m^{(2)}(x_m-\alpha_m)=1$, we have our assertion.
\qed
\end{Proof}
\begin{Lemma}\label{lower bound of constants matrix}
Assume $u_j,v_j \in \mathbb{R}$, $j=1,2$ and
$u_1v_2-v_1u_2=1$.
Then
\[
\low
\begin{pmatrix}
u_1&v_1\\
u_2&v_2
\end{pmatrix}
\geq \cfrac{1}{\sqrt{u_1^2+u_2^2+v_1^2+v_2^2}}.
\]
\end{Lemma}
\begin{Proof}\rm
Let $u=
\begin{pmatrix}
u_1\\
u_2
\end{pmatrix}$ and
$v=
\begin{pmatrix}
v_1\\
v_2
\end{pmatrix}
$, and
$u\cdot v=u_1v_1+u_2v_2$.
The eigenvalues $\xi_{\pm}$ of the matrix $
\begin{pmatrix}
u_1&u_2\\
v_1&v_2
\end{pmatrix}
\begin{pmatrix}
u_1&v_1\\
u_2&v_2
\end{pmatrix}
$ are
\[
\xi_{\pm}=\frac{1}{2}
\left(
|u|^2+|v|^2\pm \sqrt{
(|u|^2+|v|^2)^2-4(|u|^2|v|^2-u\cdot v)^2
}
\right).
\]
Since $1-\sqrt{1-t}\geq\frac{1}{2}t$ for $0<t<1$ and $|u|^2|v|^2-u\cdot v=(u_1v_2-u_2v_1)^2=1$, we obtain
\[
\xi_{-}\geq\frac{1}{|u|^2+|v|^2}.
\]
This implies our assertion.
\qed
\end{Proof}
\begin{Lemma}
It follows that
\begin{eqnarray}\label{lower bound of R_m}
\low R_m \geq
\cfrac{1}{2\sqrt{2\alpha_m^2+1}}
\exp
\left(
-\frac{1}{3}h_m(4\alpha_m^3+3\alpha_m)
\right).
\end{eqnarray}
\end{Lemma}
\begin{Proof}\rm
Note that
\[
R_m=\lim_{x\uparrow x_m+\alpha_m}
\begin{pmatrix}
\varphi_{m}(x)&\psi_{m}(x)\\
\frac{d}{dx}\varphi_{m}(x)&\frac{d}{dx}\psi_{m}(x)
\end{pmatrix}=
\lim_{x\uparrow x_m+\alpha_m}
\begin{pmatrix}
\tilde{\varphi}_{m}(x)&\tilde{\psi}_{m}(x)\\
\frac{d}{dx}\tilde{\varphi}_{m}(x)&\frac{d}{dx}\tilde{\psi}_{m}(x)
\end{pmatrix}
\begin{pmatrix}
u_{m}^{(1)}(x)&v_{m}^{(1)}(x)\\
u_m^{(2)}(x)&v_{m}^{(2)}(x)
\end{pmatrix}.
\]
It is straightfoward to see
\begin{eqnarray}
\lim_{x\uparrow x_m+\alpha_m}
\begin{pmatrix}
\tilde{\varphi}_{m}(x)&\tilde{\psi}_{m}(x)\\
\frac{d}{dx}\tilde{\varphi}_{m}(x)&\frac{d}{dx}\tilde{\psi}_{m}(x)
\end{pmatrix}
&=&\nonumber
\begin{pmatrix}
1&2\alpha_m\\
0&1
\end{pmatrix},
\\
\low
\begin{pmatrix}
1&2\alpha_m\\
0&1
\end{pmatrix}
&\geq& \cfrac{1}{\sqrt{4\alpha_m^2+2}}.
\label{lower bound of non-pertubed matrix}
\end{eqnarray}
By Lemmas $\ref{wronskian constancy}$ and $\ref{lower bound of constants matrix}$, we have
\begin{eqnarray*}
\low
\begin{pmatrix}
u_{m}^{(1)}(x)&v_{m}^{(1)}(x)\\
u_m^{(2)}(x)&v_{m}^{(2)}(x)
\end{pmatrix}
\geq
\frac{1}{\sqrt{|u_m(x)|^2+|v_m(x)|^2}}.
\end{eqnarray*}
We see that for $a, b \in \mathbb{R}$,
\begin{eqnarray}
\sup_{\theta\in[0,\pi)}
\left|
\left(
\begin{pmatrix}
\cos \theta\\
\sin \theta
\end{pmatrix}
,
\begin{pmatrix}
ab&b^2\\
-a^2&-ab
\end{pmatrix}
\begin{pmatrix}
\cos \theta\\
\sin \theta
\end{pmatrix}
\right)
\right|=\frac{a^2+b^2}{2}.
\label{q-norm}
\end{eqnarray}
By $(\ref{derivative of constants3})$ and $(\ref{q-norm})$, we have
\begin{eqnarray}\nonumber
\frac{d}{dx}(|u_m|^2)&=&2\left(u_m, \frac{d}{dx} u_m\right)\\
&\leq&\nonumber
2|V|\left|
\left(u_m,
\begin{pmatrix}
\tilde{\varphi}_{m}\tilde{\psi}_{m}&\tilde{\psi}_{m}^2\\
-\tilde{\varphi}_{m}^2&-\tilde{\varphi}_{m}\tilde{\psi}_{m}
\end{pmatrix}u_m \right)\right|
\\ \nonumber
&\leq&
h_m(\tilde{\varphi}_{m}^2+\tilde{\psi}_{m}^2)|u_m|^2.
\end{eqnarray}
Thus, by Gronwall's inequality, we obtain
\begin{eqnarray*}
|u_m(x)|^2
\leq
\exp
\left(
h_m
\int_{x_m-\alpha_m}^{x}
\left(\tilde{\varphi}_{m}(y)^2+\tilde{\psi}_{m}(y)^2\right)
dy
\right).
\end{eqnarray*}
In particular
\begin{eqnarray}\nonumber
|u_m(x_m+\alpha_m)|^2
&\leq&
\exp
\left(
h_m
\int_{x_m-\alpha_m}^{x_m+\alpha_m}
\left(1+(y-x_m+\alpha_m)^2\right)
dy
\right).
\\
\nonumber
&=&
\exp
\left(
h_m
(2\alpha_m+\frac{8}{3}\alpha_m^3)
\right).
\end{eqnarray}
We can estimate $|v_m|^2$ in a similar way:
\begin{eqnarray*}
|v_m(x_m+\alpha_m)|^2
\leq
\exp
\left(
\frac{2}{3}h_m(4\alpha_m^3+3\alpha_m)
\right).
\end{eqnarray*}
Therefore, we see that
\begin{eqnarray}
\low
\left.
\begin{pmatrix}
u_{m}^{(1)}(x)&v_{m}^{(1)}(x)\\
u_m^{(2)}(x)&v_{m}^{(2)}(x)
\end{pmatrix}
\right|_{x=x_m+\alpha_m}
&\geq&
\left.
\frac{1}{\sqrt{|u_m(x)|^2+|v_m(x)|^2}}
\right|_{x=x_m+\alpha_m}
\nonumber
\\
\label{lower bound of constants}
&\geq&
\cfrac{1}{\sqrt{2}}
\exp
\left(
-\frac{1}{3}h_m(4\alpha_m^3+3\alpha_m)
\right).
\end{eqnarray}
By $(\ref{lower bound of non-pertubed matrix})$ and $(\ref{lower bound of constants})$, we have our assertion.
\qed
\end{Proof}
\begin{proof_mainT}\rm
By $(\ref{lower bound of W_m})$, $(\ref{lower bound of R_m})$, and $c_n=R_nW_{n-1}c_{n-1}$ for $n\geq1$,
we obtain
\begin{eqnarray}\label{lower bound of c_n}
|c_n|\geq
\frac{1}{2^n}
\left(
\prod_{m=1}^{n}(L_{m-1}^2+2)(2\alpha_m^2+1)
\right)^{-\frac{1}{2}}
\exp
\left(
-\frac{1}{3}\sum_{m=1}^{n}h_m(4\alpha_m^3+3\alpha_m)
\right).
\end{eqnarray}
If $n$ is sufficiently large, then
\begin{eqnarray*}
\frac{L_n^4}{L_n^3+3L_n}\geq
\frac{L_n}{4}.
\end{eqnarray*}
Thus, by $(\ref{lower bound of L^2 integral})$ and $(\ref{lower bound of c_n})$, if $n$ is sufficiently large, then we have
\begin{eqnarray*}
\int_{J_n}|f(x)^2|dx
\geq
\frac{L_n}{4^{n+2}}
\left(
\prod_{m=1}^{n}(L_{m-1}^2+2)(2\alpha_m^2+1)
\right)^{-1}
\exp
\left(-\frac{2}{3}\sum_{m=1}^{n}h_m(4\alpha_m^3+3\alpha_m)
\right).
\end{eqnarray*}
Suppose $(\ref{assumption})$. Then we obtain
\begin{eqnarray}\nonumber
\int_{J_n}|f(x)^2|dx\rightarrow \infty.
\end{eqnarray}
This implies that any solutions $f$ of $-\frac{d^2}{dx^2}f+Vf=0$ do not belong to $L^2([0,\infty))$.
\qed
\end{proof_mainT}
\section{Proof of Corollary 1.3.}
Note that $x_n=\exp(n^n)$, $\alpha_n=\frac{1}{2}$, $h_n=e^{n}$, $L_0=x_1-\frac{1}{2}$ and $L_n={x_{n+1}}-{x_n}-1$ for $n\geq1$.
\begin{Lemma}\label{AAA}
If there exists $f \in D(H_{\theta})$ such that $(f,H_{\theta}f)<0$, then $H_{\theta}$ has a single negative eigenvalue.
\end{Lemma}
\begin{Proof}\rm
$\mathcal{R}[A]$ denotes the range of a map $A$.
We see that $\dim\mathcal{R}[E_{\theta}((-\infty,0))]\leq1$.
For its proof, see Lemma \ref{appendix}.
Let $f \in D(H_{\theta})$ such that $(f,H_{\theta}f)<0$.
Then this implies $\mathcal{R}[E_{\theta}((-\infty,0))] \neq \{0\}$ and $\dim\mathcal{R}[E_{\theta}((-\infty,0))]=1$.
This implies our assertion.
\qed
\end{Proof}
\begin{Lemma}\label{BBB}
It follows that
\begin{enumerate}[$(1)$]
\item
$H_{\theta}$ has no negative eigenvalues for $\theta \in [0,\frac{\pi}{2}]$,
\item
$H_{\theta}$ has a single negative eigenvalue for $\theta \in (-\frac{\pi}{2},\arctan(-\frac{1+\sqrt{3}}{2})]$.
\end{enumerate}
\end{Lemma}
\begin{Proof}\rm
Let $f \in D(H_{\theta})$.
Then $f$ satisfies the boundary condition $f(0)\sin\theta-\frac{d}{dx}f(0)\cos\theta=0$.
We obtain that
\begin{eqnarray*}
(f,H_{\theta}f)&=&\int_{0}^{\infty}\overline{f(x)}(-\frac{d^2}{dx^2}f(x)+V(x)f(x))dx\\
&=&
\overline{f(0)}\frac{d}{dx}f(0)+
\int_{0}^{\infty}\left(\left|\frac{d}{dx}f(x)\right|^2+V(x)|f(x)|^2\right)dx.
\end{eqnarray*}
If $\theta=0$ or $\frac{\pi}{2}$, then we have $\overline{f(0)}\frac{d}{dx}f(0)=0$ and $(f,H_{\theta}f)\geq0$.
If $0<\theta<\frac{\pi}{2}$, then by the boundary condition, we obtain
\begin{eqnarray*}
(f,H_{\theta}f)=
|f(0)|^2\tan\theta+\int_{0}^{\infty}\left(\left|\frac{d}{dx}f(x)\right|^2+V(x)|f(x)|^2\right)dx
\geq0.
\end{eqnarray*}
Thus $(f,H_{\theta}f)\geq0$ for $\theta\in[0,\frac{\pi}{2}]$.
This implies the first part of our assertion.
We shall prove that there exists $f \in D(H_{\theta})$ such that $(f,H_{\theta}f)<0$ for any $\theta \in (-\frac{\pi}{2},\arctan(-\frac{1+\sqrt{3}}{2})]$.
It is sufficient to prove there exists $f\in L^2([0,\infty))$ such that $(-\frac{d^2}{dx^2}f+Vf,f)<0$ for any boundray conditions $\frac{\frac{d}{dx}f(0)}{f(0)}=-\lambda$, $\lambda\geq\frac{1+\sqrt{3}}{2}$.
Let $\lambda\geq1$.
Define $f_{\lambda}:[0,\infty)\rightarrow \mathbb{R}$ by
\[
f_{\lambda}(x)=
\begin{cases}
\exp\left( \cfrac{\lambda}{x-1} \right),& \text{if $0\leq x <1$}\\
0,&\text{otherwise}.
\end{cases}
\]
We see that $f_{\lambda} \in L^2([0,\infty))$ and that $f_{\lambda}(0)=\exp(-\lambda)$, $\frac{d}{dx}f_{\lambda}(0)=-\lambda\exp(-\lambda)$.
Then we have
\begin{eqnarray*}
\int_{0}^{\infty}|\frac{d}{dx}f_{\lambda}(x)|^2dx&=&
\int_0^{1}
\left|
\cfrac{\lambda}{(x-1)^2}\exp \left(\frac{\lambda}{x-1} \right)
\right|^2dx
\\
&=&
\frac{1}{4\lambda}(2\lambda^2+2\lambda+1)\exp(-2\lambda).
\end{eqnarray*}
Since $V(x)f_{\lambda}(x)=0$ for $x\geq0$, we obtain that, for
$\lambda \geq \frac{1+\sqrt{3}}{2}$,
\begin{eqnarray*}
\left(-\frac{d^2}{dx^2}f_{\lambda}+Vf_{\lambda},f_{\lambda}\right)
=\frac{1}{4\lambda}(-2\lambda^2+2\lambda+1)\exp(-2\lambda)
<0.
\end{eqnarray*}
By Lemma $\ref{AAA}$, we have our assertion.
\qed
\end{Proof}
\begin{Lemma}\label{lem}
For any $p>0$, it follows that
\begin{eqnarray*}
\lim_{n\rightarrow \infty}
x_n
\left(
\prod_{m=1}^{n-1}x_m
\right)^{-p}=\infty.
\end{eqnarray*}
\end{Lemma}
\begin{Proof}\rm
We obtain
\begin{eqnarray*}
x_{n+1}
\left(
\prod_{m=1}^{n}x_m
\right)^{-p}
&=&
\exp
\left(
(n+1)^{n+1}-pn^n-p\sum_{m=1}^{n-1}m^m
\right)
\\
&\geq&
\exp
\left(
(n+1)^{n+1}-pn^n-p(n-1)^n
\right)
\\
&=&
\exp
\left(
(n+1-2p)(n+1)^{n}
\right)
\\
&\rightarrow&\infty,\qquad\text{as $n\rightarrow \infty$.}
\end{eqnarray*}
\qed
\end{Proof}
\begin{proof_main}\rm
\noindent
By Lemma $\ref{BBB}$ it is sufficient to prove that $0\notin \sigma_{pp}(H_\theta)$ for all $\theta$.
We see that for all $n\geq 1$,
\begin{eqnarray*}
\frac{L_n^2+2}{L_n^2}<2 .
\end{eqnarray*}
Thus we have for all $n\geq 1$,
\begin{eqnarray*}
\left(
\prod_{m=1}^{n}(L_{m-1}^2+2)
\right)^{-1}
\geq
\frac{2^{n-1}}{L_0^2+2}
\left(
\prod_{m=1}^{n-1}L_m
\right)^{-2}.
\end{eqnarray*}
We see that $L_n<x_{n+1}$ for all $n\geq1$ and $L_n>\frac{1}{2}x_{n+1}$ for sufficiently large $n\geq1$.
Therefore we have
\begin{eqnarray}
L_n
\left(
\prod_{m=1}^{n-1}L_m
\right)^{-2}
\geq
\cfrac{1}{2}
x_{n+1}
\left(
\prod_{m=1}^{n}x_m
\right)^{-2}.
\nonumber
\end{eqnarray}
By Lemma $\ref{lem}$, we obtain
\begin{eqnarray}
&&
\cfrac{L_n}{4^n}
\left(
\prod_{m=1}^{n}(L_{m-1}^2+2)
\right)^{-1}
\left(
\prod_{m=1}^{n}(2\alpha_m^2+1)
\right)^{-1}
\exp
\left(-\frac{2}{3}\sum_{m=1}^{n}h_m(4\alpha_m^3+3\alpha_m)
\right)
\nonumber
\\
&&\geq
\frac{1}{2(L_0^2+2)3^{n}}L_n
\left(
\prod_{m=1}^{n-1}L_m
\right)^{-2}
\exp
\left(
-\frac{4}{3}
\sum_{m=1}^{n}e^m
\right)
\nonumber
\\
&&\geq
\frac{1}{4(L_0^2+2)3^{n}}x_{n+1}
\left(
\prod_{m=1}^{n}x_m
\right)^{-2}
\exp
\left(
-\frac{4}{3}e^{n+1}
\right)
\nonumber
\\
&&=
\frac{\exp(\frac{1}{3}(n+1)^{n+1})}{4(L_0^2+2)3^{n}}
{x_{n+1}}^{\frac{1}{3}}
\left(
\prod_{m=1}^{n}x_m
\right)^{-2}
\exp
\left(\frac{1}{3}
\left(
(n+1)^{n+1}
-4e^{n+1}
\right)
\right)
\nonumber
\\
&&
\rightarrow \infty,\qquad\text{as $n\rightarrow \infty$.}
\nonumber
\end{eqnarray}
By Theorem $\ref{main}$, we see that $0\notin \sigma(H_\theta)$.
\qed
\end{proof_main}
\begin{appendices}
\section{}
By \cite{Simonsparse}, we see that $\sigma_{ac}(H_{\theta})=\emptyset$ and $\sigma_{pp}(H_{\theta})\cap(0,\infty)=\emptyset$.
In this appendix, we prove that $H_{\theta}$ has a single negative eigenvalue for some $\theta$ and that $\sigma_{sc}(H_{\theta})=[0,\infty)$ for all $\theta$.
Let $V$ ba a sparse potential with
$x_n=\exp(n^n)$ for $n=1,2,...$, and
\begin{eqnarray*}
V(x)=
\begin{cases}
e^{n},&\text{if $|x-x_n|\leq\frac{1}{2}$ for $n=1,2,...$},\\
0,&\text{otherwise}.
\end{cases}
\end{eqnarray*}
By \cite{Simonsparse} and \cite{SimonSpencer}, we see that $\sigma_{pp}(H_{\theta})\cap(0,\infty)=\emptyset$ and $\sigma_{ac}(H_{\theta})=\emptyset$ for all $\theta$.
\begin{Lemma}\label{appendix}
Let $E_{\theta}$ be the spectral measure of $H_{\theta}$.
For all $\theta$, $\dim \mathcal{R}[ E_{\theta}((-\infty,0))]\leq1$.
\end{Lemma}
\begin{Proof}\rm
We prove this by a contradiction.
Suppose that $\dim \mathcal{R}[ E_{\theta}((-\infty,0))]>1$.
Then we can take $\varphi,\psi \in E_{\theta}((-\infty,0))$ such that $\varphi$ and $\psi$ are orthogonal to each other.
Let $\varphi_n=E_{\theta}((-n,-\frac{1}{n}))\varphi$ and $\psi_n=E_{\theta}((-n,-\frac{1}{n}))\psi$.
We see that $\varphi_n,\psi_n\in D(H_{\theta})$, $\varphi_n \rightarrow \varphi$ and $\psi_n\rightarrow \psi$.
Let $N\geq1$ be sufficiently large such that $\varphi_N$ and $\psi_N$ are linearly independent. Since $\alpha\varphi_N+\beta\psi_N\in \mathcal{R}[E_{\theta}((-N,-\frac{1}{N}))]$ for $\alpha,\beta\in \mathbb{C}$, we have, for $(\alpha,\beta) \neq (0,0)$,
\begin{eqnarray}\label{condition1}
(\alpha\varphi_N+\beta\psi_N,H_{\theta}(\alpha\varphi_N+\beta\psi_N))<0.
\end{eqnarray}
On the other hand, since the deficiency indices of $H$ are equal to one, there exists an ismometric operator $U_{\theta}:\ker(H^*-i)\rightarrow \ker(H^*+i)$ and $w \in \ker(H^*-i)$ such that
\[
D(H_\theta)=\left\{
v+\alpha(w+U_{\theta}w)\middle| v\in D(\overline{H}), \alpha \in \mathbb{C}
\right\},
\]
where $\overline{H}$ is the closure of $H$.
Let $u_\theta=w+U_\theta w$.
There exist $v_1,v_2 \in D(\overline{H})$ and $\alpha_1,\alpha_2\in \mathbb{C}$ such that
\begin{eqnarray*}
\varphi_N&=&v_1+\alpha_1u_{\theta},\\
\psi_N&=&v_2+\alpha_2u_{\theta}.
\end{eqnarray*}
If $\alpha_1=0$, then $(\varphi_N,H_{\theta}\varphi_N)=(v_1,\overline{H}v_1)\geq0$.
$(\ref{condition1})$ implies $\alpha_1\neq0$.
Similarly we have $\alpha_2 \neq 0$.
We obtain
\begin{eqnarray*}\label{condition2}
(\alpha_2\varphi_N+\alpha_1\psi_N,H_{\theta}(\alpha_2\varphi_N+\alpha_1\psi_N))
=(\alpha_2v_1+\alpha_1v_2,\overline{H}(\alpha_2v_1+\alpha_1v_2))\geq0.
\end{eqnarray*}
This contradicts with $(\ref{condition1})$.
Thus we have our assertion.
\qed
\end{Proof}Let $E<0$.
We see that $\dim \ker(H^*-E)=1$.
This implies that there exists a boundary condition $\theta(E)$ such that $H_{\theta(E)}$ has a single negative eigenvalue $E$.
\begin{Lemma}
For all $\theta \in (-\frac{\pi}{2}, \frac{\pi}{2}]$, it follows that $\sigma_{sc}(H_{\theta})=[0,\infty)$.
\end{Lemma}
\begin{Proof}\rm
Since $\sigma_{pp}(H_{\theta})\cap(0,\infty)=\emptyset$ and $\sigma_{ac}(H_{\theta})=\emptyset$ for all $\theta$,
we see that $\sigma(H_{\theta})\cap(0,\infty)=\sigma_{sc}(H_{\theta})\cap(0,\infty)$.
We prove $(0,\infty) \subset \sigma(H_{\theta})$ for all $\theta$ by contradition.
Suppose that there exist $\theta\in[0,\pi),E>0$ such that $E \in (0,\infty)\setminus \sigma(H_{\theta})$.
Since $H$ is regular at zero and limit-circle case at infinity, the deficiency indices $\dim \ker(H^*\pm i)$ are equal to one.
Thus $\dim \ker (H^*-E)=1$.
This implies that there exists an $L^2$-solution of $-\frac{d^2}{dx^2}f+Vf=Ef$.
By \cite[Theorem 2.3.]{Simonsparse}, however, $-\frac{d^2}{dx^2}f+Vf=pf$ has no solutions with $f \in L^2([0,\infty))$ for any $p>0$.
This is a contradiction and we get $(0,\infty)\subset\sigma_{sc}(H_{\theta})$ for all $\theta$.
Since $\sigma_{sc}(H_{\theta})\cap(-\infty,0)=\emptyset$, we get our assertion.
\qed
\end{Proof}
\section*{Acknowledgement}
This work was supported by JST SPRING, Grant Number JPMJSP2136.
\end{appendices}
\section*{Declarations}
\subsection*{Ethical Approval }
The author has no competing interests to declare that are relevant to the content of this article.
\subsection*{Competing interests}
The author declares no conflict of interest.
\subsection*{Authors' contributions }
The author confirms sole responsibility for this manuscript .
\subsection*{Funding}
This work was supported by JST SPRING, Grant Number JPMJSP2136.
\subsection*{Availability of data and materials }
Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
|
\section{Introduction}\label{Introduction}
\IEEEPARstart{D}{eep} Neural Networks (DNNs) have gained much success in diverse real-world applications~\cite{2016Deep, huang2017densely, redmon2016you}. This is largely attributed to their promising ability to automatically extract meaningful features from data without heavy feature engineering\cite{elsken2019neural}. Generally speaking, the performance of DNNs largely depends on their architectures, which breed the famous state of the arts such as ResNet~\cite{2016Deep}, MobileNet~\cite{howard2017mobilenets,sandler2018mobilenetv2,howard2019searching}, and Transformer~\cite{vaswani2017attention}. In practice, most DNN architectures including those aforementioned are all manually designed with expertise. However, the design process is often labor-intensive, where various candidate architectures have to be exhaustively tested. Consequently, researchers without extensive knowledge may wish to use DNNs to solve their tasks at hand, but it is challenging, if not impossible, to design satisfied DNNs. These facts have severely limited the further applications of DNNs and in turn promoted the development of Neural Architecture Search (NAS)~\cite{elsken2019neural}. Specifically, NAS aims to automatically discover high-performance DNN architectures, thus allowing researchers without or with rare expertise to conveniently benefit from the success of DNNs. The architectures designed by NAS have recently shown to even outperform those manually designed in some tasks~\cite{real2019regularized, zoph2018learning}, and have become increasingly popular in the field of deep learning~\cite{ren2021comprehensive}.
Mathematically, NAS formulates the design process of DNN architectures as an optimization problem~\cite{elsken2019neural}. In particular, NAS first defines the search space containing all candidates. Then, it adopts a well-designed search strategy to search for the optimal architecture. During the search process, NAS must evaluate the performance of every searched architecture to effectively guide the running of the search strategy. Generally, the NAS problem is difficult to be solved because of facing multiple optimization challenges, such as the prohibitive computational cost, and with multi-conflicting objectives~\cite{liu2020survey}.
The Evolutionary Computation (EC)~\cite{real2017large,xie2017genetic}, the Reinforcement Learning (RL)~\cite{zoph2018learning,zoph2016neural}, and the gradient algorithm~\cite{fil2021darts} are currently the three mainstream optimization techniques for addressing NAS. The EC-based NAS regards the DNN architectures as the individuals, and iteratively generates the population by applying genetic operators and eliminates poorly performed individuals by selection operators. When the stopping condition is satisfied, the best architecture from the surviving population is picked up for use. In the RL-based NAS, a controller is trained to guide the search process. It uses the performance of the architecture as the reward to update the itself to search for a better architecture in the next iteration.
The gradient-based NAS generally relaxes the discrete search space to be continuous and uses the gradient descent algorithm to search for a promising architecture. Please note that the gradient-based NAS has been inappropriately claimed to be more efficient than others. This misunderstanding is mainly caused by the representative in this category, i.e., DARTS~\cite{liu2018darts}, which was collectively designed with an Efficient Evaluation Method (EEM) named weight-sharing (will be discussed in Section~\ref{one-shot}). In principle, all NAS algorithms have similar computation complexity if they maintain the same performance evaluation techniques.
Generally, whatever optimization algorithm is used, many DNN architectures need to be evaluated during the search process. This is because these optimization algorithms are iterative, and the current performance must be known in advance, thus effectively guiding the next-step iteration search~\cite{elsken2019neural}. For the Traditional Evaluation Method (TEM), the performance in NAS is evaluated by fully training the corresponding architecture(s) searched in each iteration. Commonly, training a DNN from scratch until converging on a small-scale dataset such as CIFAR-10~\cite{cifar10dataset} may take hours or even days depending on the scale of the DNN. Consequently, since there are often thousands of DNNs to be trained in NAS, the whole NAS algorithm becomes prohibitively computation-intensive and time-consuming. For example, on CIFAR-10, the LargeEvo algorithm~\cite{real2017large} consumed 250 Graphic Processing Units (GPUs) for 11 days. The NAS-RL algorithm~\cite{zoph2016neural} took 800 GPUs for 28 days. Even more, the RegularizedEvo algorithm~\cite{real2019regularized} ran on 450 GPUs for 7 days. In practice, buying or renting such scales of GPU resources is commonly unaffordable for most researchers~\cite{xie2022benchenas}. As a result, how to accelerate the TEM to reduce the prohibitive computational overhead is essential, which results in the research topic of EEMs in the NAS community~\cite{ren2021comprehensive}. Specifically, EEM refers to the performance evaluation method that consumes less time than TEM in the entire performance evaluation process of NAS.
To the best of our knowledge, the first work of EEMs is designed in the LargeEvo algorithm~\cite{real2017large} preprintly available in Arxiv in 2016, although there are also some earlier works that potentially can achieve the same goal~\cite{domhan2015speeding}. With the development of NAS, the research of EEMs has received much attention and is becoming one of the hottest topics in the current artificial intelligence community. As evidenced by Fig.~\ref{fig_submission} which reports the submissions of EEMs from 2017 to November 2022 (when the paper was made). The number of submissions turned doubled year by year from 2018 to 2021. As of November 2022, although the number of publications in this year is slightly less than that in 2021, with the publication of more conferences proceeding released in December such as NeurIPS2022, it is believed that the number of publications in this year will finally surpass that in 2021. Despite the popularity and criticality of the EEMs, there is a lack of a survey to systematically review these works. This will make it difficult for interested researchers with rare knowledge to quickly grasp the current situation and landmark works.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.48\textwidth]{submission.pdf}}
\caption{The number of ``submissions'' refers to EEMs. We obtain these submissions by searching on Google scholar with the keywords of ``early stopping'' OR ``learning curve'' OR ``network morphism'' OR ``weight sharing'' OR ``one-shot'' OR ``DARTS'' OR ``differentiable'' OR ``predictor'' OR ``population memory'' OR ``zero-cost'' OR ``zero-shot'' OR ``training-free'' AND ``architecture search'' OR ``architecture design'' OR ``CNN'' OR ``deep learning'' OR ``deep neural network''. After that, we carefully examined these papers to record those belonging to EEMs. Please note the submissions include not only the ones peer reviewed, but also those available on ArXiv. }\label{fig_submission}
\end{figure}
In this paper, we survey these papers of EEMs published up to date. Considering the training of DNNs is the root causing the prohibitive cost of NAS, we reasonably categorize existing EEMs to make audiences easily grasp the design features of these methods. In Section~\ref{org}, we show the detail of the categorization and organization. After that, the details of these EEMs are surveyed in Sections~\ref{section_method} to~\ref{direction}, including the method designs, evaluation metrics, and future directions. Finally, we make a conclusion in Section~\ref{conclusion}.
\section{Categorization and Organization}\label{org}
In this section, we first describe the notation and terminology used in this paper in Subsection~\ref{org_pd}. Then, the categorization for the EEMs is detailed in Subsection~\ref{cate}. Finally, the organization of this survey is discussed in Section~\ref{section_org}.
\subsection{Notation and Terminology}\label{org_pd}
To better describe the TEM and the EEMs, we first give the definitions for the \emph{\textbf{initialization time}}, \emph{\textbf{evaluation time}}, and \emph{\textbf{runtime}} involved in a performance evaluation method. Specifically, the initialization time is the time of pre-computation for constructing the method, while the evaluation time refers to the overall time for evaluating the performance of all architectures in $\mathcal{A}^s$ indicating the set of the architectures searched in all iterations of the NAS algorithm. The runtime is the sum of the initialization time and the evaluation time.
The \textbf{TEM} fully trains each architecture searched in each iteration to obtain their performance. For the TEM, the initialization time is zero because there is no pre-computation. The evaluation time for TEM is the sum of the time to train all searched architectures (i.e., all architectures in $\mathcal{A}^s$). As a result, the runtime of a TEM, say $T_{TEM}$, can be represented as Equation~(\ref{equation_trad}):
\begin{equation}\label{equation_trad}
T_{TEM}= \sum_{i=1}^{|\mathcal{A}^s|}\mathcal{T}_{train}(\mathcal{A}^s_i,D_{train},D_{valid},N_{epoch})
\end{equation}
where $\mathcal{T}_{train}(\cdot)$ represents the time for training an architecture on training dataset $D_{train}$ for $N_{epoch}$ epochs and inferencing on validation dataset $D_{valid}$. For TEMs, $N_{epoch}$ is usually set to a number that can train the architecture to convergence.
TEM is time-consuming because a large number (i.e., $|\mathcal{A}^s|$) of architectures require being trained during the search process. An \textbf{EEM} is a performance evaluation method that consumes less runtime than TEM in the same NAS algorithm. Because of the different working principles, various EEMs calculate the initialization time and evaluation time in different ways. For existing EEMs, some do not require initialization and the architecture performance is obtained with efficient training techniques. For others, a number of architectures are first trained to initialize the EEM. Then, the architecture performance is efficiently obtained based on these trained architectures. In practice, for any EEM, training the architecture takes up most of the runtime, while others can be ignored. As a result, the runtime $T_{EMM}$ for an EEM can be approximately represented as Equation~(\ref{eqa_ac}):
\begin{equation}\label{eqa_ac}
T_{EEM}\approx \sum_{i=1}^{|\mathcal{A}^t|}\mathcal{T}_{train}(\mathcal{A}^t_i,D_{train},D_{valid},N_{epoch})
\end{equation}
where $\mathcal{A}^t$ is the set of all architectures trained in the pre-computation and/or evaluation phases of an EMM.
\subsection{Categorization}\label{cate}
We categorize the existing EEMs based on $|\mathcal{A}^t|$ in Equation~(\ref{eqa_ac}) because $|\mathcal{A}^t|$ can directly reflect the degree of efficiency (i.e., time complexity) of various EEMs. Specifically, this results in four different categories shown below:
\begin{itemize}
\item \emph{$N$-shot evaluation method}. For the $N$-shot evaluation method, the number of trained architectures is greater than or equal to the number of searched architectures, i.e., $|\mathcal{A}^t|\geq|\mathcal{A}^s|$. The $N$-shot evaluation methods still need to train every searched architecture (i.e., $\mathcal{A}^s\in \mathcal{A}^t$), and mainly accelerate the training of these architectures to consume less runtime than TEM.
\item \emph{Few-shot evaluation method}. For the few-shot evaluation method, the number of trained architectures is less than the number of searched architectures and greater than one, i.e., $|\mathcal{A}^s|\textgreater|\mathcal{A}^t|\textgreater 1$. Because the number of trained architectures of the few-shot evaluation method is less than the number of searched architectures, the runtime is naturally less than $T_{TEM}$.
\item \emph{One-shot evaluation method}. For the one-shot evaluation method, the number of trained architectures is one, i.e., $|\mathcal{A}^t|=1$. The one-shot evaluation method can consume less runtime than TEM because only one architecture needs to be trained.
\item \emph{Zero-shot evaluation method}. For the zero-shot evaluation method, the number of trained architectures is zero, i.e., $\mathcal{A}^t=\emptyset$. The zero-shot evaluation method involves no training, thus resulting in an extremely low cost.
\end{itemize}
\subsection{Organization}\label{section_org}
The $N$-shot evaluation methods, the few-shot evaluation methods, the one-shot evaluation methods, and the zero-shot evaluation methods are discussed in Section~\ref{section_method}, with the focus on design principle and strength and weakness analyis of these methods. In Section~\ref{evaluation}, we detail the evaluation for these EEMs, including the evaluation metrics, benchmark dataset, and the comparison results of various methods on benchmark datasets. In Section~\ref{direction}, we discuss the challenges and future directions of EEMs. For the convenience of quickly navigating to the interested part, an illustration of above organization is shown in Fig.~\ref{categorization}.
\begin{figure*}[htbp]
\centerline{\includegraphics[width=0.98\textwidth]{categorization.pdf}}
\caption{The organization of this paper.}
\label{categorization}
\end{figure*}
\section{Methods}\label{section_method}
In this section, we discuss the working principle and research status of various EEMs under the four categories.
\subsection{$N$-shot Evaluation Methods}\label{$N$-shot}
The $N$-shot evaluation methods still need to train every searched architecture, and even require additional training of architectures for some specific $N$-shot methods. They mainly accelerate the TEM by speeding up the training process of the searched architectures, involving $\mathcal{A}_i^t$, $D_{train/valid}$, and $N_{epoch}$ in Equation (\ref{eqa_ac}).
For the purpose of better discussion, we first describe the training process for a DNN. During the training process, all samples in the dataset are sent into the model (i.e., architecture) to complete one forward and backpropagation, and the process repeats $N_{epoch}$ times. The training time mainly depends on three factors: the dataset scale, the model size, and the number of epochs. Specifically, the time of completing one forward and backpropagation depends on the dataset scale and the model size. If the dataset scale or the model size is large, the computational complexity becomes higher, which leads to much time spent on the computation. Furthermore, the training time is also positively correlated with the number of epochs ($N_{epoch}$). The larger $N_{epoch}$, the longer the training time.
The $N$-shot evaluation methods mainly consist of \emph{downscaled dataset methods}, \emph{downscaled model methods}, \emph{network morphism}, and \emph{learning curve extrapolation}. Please note that $\mathcal{A}^t$ is equal to $\mathcal{A}^s$ (i.e., $\mathcal{A}^t=\mathcal{A}^s$) for all $N$-shot evaluation methods except for the learning curve extrapolation. As the name says, the downscaled dataset methods aim to reduce the dataset scale to accelerate the training time. The downscaled model methods focus on reducing the model size. Furthermore, network morphism and learning curve extrapolation aim to reduce the number of epochs. Next, we will discuss these $N$-shot evaluation methods.
\subsubsection{Downscaled Dataset Methods}
Similar to TEM, the downscaled dataset methods train every searched architecture to obtain their performance to guide the search process. The difference is that they use a downscaled dataset to replace the original dataset. Because the dataset scale is reduced, the training of architectures becomes quick. Naturally, the runtime of the downscaled dataset methods is less than that of TEM. The downscaled dataset methods can be divided into three categories.
The first category samples a subset of the original dataset as the downscaled dataset. Because the number of samples in the dataset is reduced, the overall dataset scale is reduced. The simplest way to obtain the subset is to randomly select some samples from the original dataset. For example, Liu \textit{et al.}~\cite{liu2019deep} randomly selected a small number of medical images from the entire dataset to train the searched architectures. However, random sampling may lead to the removal of some representative samples, thus greatly weakening the generalization of the architecture trained in the subset~\cite{park2019data, na2021accelerating}. This will lead to an inaccurate evaluation of the performance. To alleviate the problem, some works designed new sampling methods. For example, Park \textit{et al.}~\cite{park2019data} designed a probe network to measure the impact of every sample on the performance ranking of the architectures. Then, they removed the samples that have a small impact on the performance of architectures. Na \textit{et al.}~\cite{na2021accelerating} used the data entropy to analyze five sampling methods (i.e., random, entropy top-$k$, entropy bottom-$k$, forgetting events, and $k$-center). They concluded that the low-entropy samples could help search for competitive architectures when the size of the subset is small. Furthermore, middle-entropy and high-entropy samples can help improve the performance of the searched architecture upon the use of low-entropy. Based on the observation, the researchers proposed a new sampling method that prefers samples in the tail ends of the data entropy distribution.
The second category downsamples each sample in the original dataset to reduce the dataset scale. Specifically, downsampling can reduce the size of samples. Because the size of every sample becomes smaller, the overall dataset scale naturally becomes smaller. A typical way is to reduce the resolution of the images in the dataset for image classification. Specifically, as the resolution of the image is reduced, the pixels in the image become fewer, and the size of the image naturally becomes smaller. As a result, reducing the image resolution in a dataset can reduce the dataset scale, thus speeding up the training. For example, Chrabaszcz \textit{et al.}~\cite{chrabaszcz2017downsampled} proposed ImageNet16x16, ImageNet32x32, and ImageNet64x64. These datasets are all variants of ImageNet~\cite{deng2009imagenet} and have the same number of images and classes as the ImageNet but with different resolutions. As the name says, ImageNet16x16, ImageNet32x32, and ImageNet64x64 downsample all images which are generally clipped to 256x256 in ImageNet to 16x16, 32x32, and 64x64. Specifically, they chose the box technique to downsample the images. The results showed that it was up to $100$ times cheaper to train on these downsampled datasets.
The third category uses a different and smaller proxy dataset with similar properties to the original dataset to train the searched architectures. The proxy dataset may have fewer samples and/or smaller sample size compared with the original dataset. For example, Zoph \textit{et al.}~\cite{zoph2018learning} used CIFAR-10 as the proxy for ImageNet. They train every architecture on CIFAR-10 to obtain its performance during the search process. Because CIFAR-10 has both fewer samples and smaller sample size than ImageNet, they can shorten the training time effectively. The reason for choosing CIFAR-10 as the proxy is that both CIFAR-10 and Imagenet aim to image classification tasks. The results also showed that the architecture searched on CIFAR-10 also performed well on ImageNet.
The downscaled dataset method is straightforward and easy to implement. It can reduce one to several orders of magnitude computational costs depending on the difference between the size of the downscaled dataset and the original dataset. However, the dataset features of the downscaled dataset are different from the original one. For example, the subset has fewer data features than the original dataset because of fewer samples. The samples in the downsampling dataset have different features compared with that in the original dataset because size is an important feature for data. As the features of each sample change, the features of the downsampling dataset also change. The change of dataset features will lead to inconsistencies between the actual performance and the performance obtained on the downscaled dataset. This is because different architectures have different capabilities to handle different features. If the features are changed, the performance is naturally changed.
\subsubsection{Downscaled Model Methods}
The downscaled model methods reduce the model size during the search process. Please note that the optimal model found is enlarged to the original size after the search process. Because the models are downscaled during the search process, the training time becomes less and the runtime spent on performance evaluation is less than $T_{TEM}$. Many works have used this method to accelerate the performance evaluation process. For example, Zoph \textit{et al.}~\cite{zoph2018learning} proposed a cell-based search space (i.e., NASNet search space) which repeatedly stacked the same cell structure to build the architecture, and only the cell structure needs to be found during the search process. To reduce the model size, they transformed each model to one with a fewer number of cell repeats and filters in the initial convolutional cell during the search. Because of its efficiency, many works~\cite{real2019regularized, liu2018darts} used the NASNet search space or its variants and followed its downscaled model method to accelerate the training.
Downscaling methods usually assume that the same downscaling action has the same effect on the performance of different models. This is because the performance ranking of the downscaled model and the original model will remain consistent only if the degree of the performance change is the same for all models. However, this assumption does not always hold. For example, removing a convolution operation from a high-complexity model may improve its performance because the operation may be redundant. Removing the same convolution operation from a low-complexity model may reduce its expression ability because of the removal of the learnable operation. This will lead to the reduction of the performance of the low-complexity model. As a result, the downscaled model methods may result in a lower correlation between the performance of the downscaled model and that of the original model.
\subsubsection{Network Morphism}
Network morphism can accelerate the training process of most searched architectures by reducing the number of epochs. Please note that not every training process of the searched architecture is accelerated, the reason for which is explained in the third paragraph of this subsection. Because the training time for most searched architectures is reduced, the runtime of network morphism is significantly less than that of TEM. Network morphism is first proposed in the background of transfer learning to rapidly transfer knowledge from one fully-trained network (i.e., a parent network) into another network (i.e., a child network). It is soon applied in the field of NAS because it can effectively accelerate the training of the child network. For the sake of understanding, we first explain the acceleration principle of network morphism and then discuss how it is applied in NAS as an EEM.
The term \emph{morphism} mathematically means a structure-preserving map from one structure to another of the same type. Network morphism refers to a function-preserving action to transform a fully-trained parent network into a child one that can completely preserve the function of the parent network~\cite{wei2016network}. Specifically, the parent network is first transformed into a different child network by predefined morphing actions. Then, the child network directly inherits the weights from the parent network and has the same function and output as the parent network. After that, the child network is generally trained on the dataset. Because the child network does not require being trained from scratch, the number of epochs needed to train to convergence is naturally reduced. As the inheriting process does not need intensive resources, the network morphism can accelerate the training time of the child network. The morphing actions (i.e., function-preserving actions) mainly include width morphing, deep morphing, kernel size morphing, etc. Different types of morphing change different parts of the parent network to generate a new child network. For the convenience of understanding, Fig.~\ref{fig_network_morphism} shows an example of width morphing. As the name says, the child network is wider than the parent network for the width network. To be specific, the node represents the operation while the edge represents the connection. In the child network, node\#6 is a copy of node\#3, and the weights of node\#3 are also transferred to node\#6. To preserve the function of the two networks, the value of weight $b$ is divided by two. In this way, if the same input is fed to the parent network and the child network, respectively, the output of both networks will be the same, which means the function of the parent network is preserved in the child network.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.45\textwidth]{network_morphism.pdf}}
\caption{An example of the network morphism. Different colors represent different operations. The labels on the edges represent the value of the associated weights.}\label{fig_network_morphism}
\end{figure}
Generally, network morphism cannot be used to accelerate the evaluation in NAS because there must exist fully-trained parent networks for using network morphism. To gap this, the architectures searched in the first iteration of the search strategy are generally fully trained as the parent of the architectures searched in the next iteration. Specifically, the architectures searched in the first iteration of NAS are fully trained as the parent network. Then, various morphing operations are applied to these parent networks to generate child networks. These child networks inherit weights from the parent network and are trained to obtain their performance with fewer epochs. These child networks are then regarded as the parent network, and the above process is iteratively performed. Because all searched architectures except the ones searched in the first iteration are trained with fewer epochs compared with TEM, network morphism consumes less runtime than $T_{TEM}$.
Many works have used network morphism to accelerate the performance evaluation process.
For example, Cai \textit{et al.}~\cite{cai2018efficient} adopted the RL-based search strategy, and used the controller to generate a morphing action to apply in the current network for the search. The designed morphing actions include a width morphing called Net2Wider and a deep morphing called Net2Deeper. Net2Wider can replace a layer with a wider layer such as more filters for convolutional layers, and Net2Deeper inserts a new layer that is initialized as an identity mapping between two layers. Then, the newly searched architectures inherited weights from the current network and were trained with fewer epochs to obtain their performance.
Cai \textit{et al.}~\cite{cai2018path} proposed the path-level morphing actions because the previous work adopted layer-level morphing which can only add filters or add layers. The path-level morphing can modify the path topologies of the parent networks and preserve their functions. Noting that the previous morphing actions can only increase the size of the networks because the function-preserving property is not guaranteed when decreasing the size of the networks. The child networks generated by these morphing actions will become deeper or wider, which may be not suitable if users wish to search for architectures with constraints on computation resources. To solve the problem, Elsken \textit{et al.}~\cite{elsken2018efficient} designed the approximate network morphism that could decrease the network size and roughly preserve the functions. Then, they adopted the EC-based search strategy and used the approximate network morphism as a mutation operator to search for architectures.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.44\textwidth]{network_morphism_unflexity.pdf}}
\caption{An example of the limit of the network morphism. The network\#1 is a fully-trained network. The network\#2 and network\#3 cannot be transformed from network\#1, and naturally cannot inherit the weights of the network\#1.}\label{fig_network_morphism_limit}
\end{figure}
Although network morphism can accelerate the performance evaluation process, there are prerequisites for using it. Specifically, if we want to use it to accelerate the training of an architecture, there must exist the parent network of it. Furthermore, not all fully-trained architectures can be used as the parent network for this architecture. As is shown in Fig.~\ref{fig_network_morphism_limit}, network\#1 is an existing fully-trained network. Network\#2 and network\#3 are randomly picked from the search space, they cannot be evaluated through the network morphism because they cannot be morphed from network\#1. As a result, network morphism cannot be flexibly applied to accelerate any given architecture.
\subsubsection{Learning Curve Extrapolation}\label{learning_curve}
Different from other $N$-shot evaluation methods that $\mathcal{A}^t$ is equal to $\mathcal{A}^s$, $\mathcal{A}^t$ in the learning curve extrapolation methods composed of $\mathcal{A}^s$ and $\mathcal{A}^{f}$, i.e., $\mathcal{A}^t=\mathcal{A}^s\cup \mathcal{A}^{f}$, where $\mathcal{A}^{f}$ is an architecture set that includes some fully trained architectures for learning curve extrapolation. The learning curve extrapolation trains a model on $\mathcal{A}^{f}$ to predict the actual performance of the architecture trained after only a few epochs. In this way, the training time of the searched architectures is largely reduced because the epochs become fewer. Since the time saved by training the architectures in $\mathcal{A}^s$ is generally greater than the time spent on fully training the architectures in $\mathcal{A}^f$, it can still spend less time on performance evaluation than TEM.
To illustrate the principle of learning curve extrapolation, we first introduce three terms: \emph{learning curve}, \emph{partial learning curve}, and \emph{final learning curve}. Specifically, the term learning curve generally refers to the function of performance with a growing number of iterations for an iterative machine learning algorithm. We use the term final learning curve to represent the entire learning curve $f_{t}=(p_{1},p_{2},\dots,p_{t}$) of an algorithm from the beginning to the end of the training, where $p_{i}$ represents the performance at the iteration of $i$ and $p_{t}$ is the final performance. Furthermore, we use the term partial learning curve to refer to the learning curve observed as of epoch $l$, i.e., $f_{l}=(p_{1},p_{2},\dots,p_{l}) (l<t)$. Take architecture 1 in Fig.~\ref{fig_learning_curve} as an example, the solid line represents the partial learning curve while the solid line with the dotted line represents the final learning curve.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.42\textwidth]{learning_curve}}
\caption{The learning curve to demonstrate the partial learning curve and the final learning curve.}\label{fig_learning_curve}
\end{figure}
The workflow of the learning curve extrapolation method is as the following. First, some architectures are sampled from the predefined search space, and then are fully trained to construct the architecture set $\mathcal{A}^f=\{(x^{1},f_{t}^{1}),(x^{2},f_{t}^{2}),\dots,(x^{k},f_{t}^{k})\}$ where $x^{i}$ denotes the $i$-th architecture and $f_{t}^{i}$ represents its final learning curve. Then, the architecture set is transformed to another form $\mathcal{A}^f=\{(x^{1},f_{l}^{1},p_{t}^{1}),(x^{2},f_{l}^{2},p_{t}^{2}),\dots,(x^{k},f_{l}^{k},p_{t}^{k})\}$ where $f_{l}^{i}$ represents the partial learning curve and $p_{t}^{i}$ denotes the final performance of architecture $x^i$. The learning curve extrapolation method builds a model $p_{t}^{i}=\mathcal{P'}(x^{i},f_{l}^{i})$ by training it on $\mathcal{A}^f$. The model $\mathcal{P'}(\cdot)$ can be used to predict the final performance of the searched architectures by feeding their architectures and partial learning curves.
Finally, the model is used to predict the final performance of the searched architectures after obtaining their partial learning curve.
The learning curve extrapolation was first proposed to automatically find a well-performing hyperparameter configuration for neural networks~\cite{rijn2015fast}. Because both hyperparameter optimization and NAS are faced with the problem of the high cost of network evaluation, the learning curve extrapolation is soon applied to NAS.
For example, Rawal \textit{et al.}~\cite{rawal2018nodes} developed a Long Short Term Memory (LSTM) consisting of an encoder RNN and a decoder RNN to predict the final performance of a partial training model. They fed the validation loss at the first $10$ epochs to the encoder. Then, the validation loss at $40$ epochs, i.e., the final validation loss, is obtained by the decoder. In this way, every searched architecture is partially trained with four times speed up, thus accelerating the entire NAS algorithm.
Baker \textit{et al.}~\cite{baker2017accelerating} extracted the features derived from model architectures, training hyper-parameters, and partial learning curve. Then, these features are used to train a set of support vector machine~\cite{svmhearst} regression models to predict the final performance of the partially trained models. Furthermore, they model the estimation process as a gaussian perturbation to calculate the probability $p(\hat{y}\leq y_{best})$ where $\hat{y}$ represents the estimated final performance, $y_{best}$ denotes the performance of the best architecture. When $p(\hat{y}\leq y_{best})\geq \Delta$ where $\Delta$ is used to balance the trade-off between increased speedups and risk of prematurely terminating good architecture, the training of an architecture terminates.
Wistuba \textit{et al.}~\cite{wistuba2020learning} built a neural network called LCRankNet to rank the architecture with the input of architecture, partial learning curve, dataset, and hyper-parameters. Furthermore, the LCRankNet can leverage learning curves from other datasets to rank the learning curve in the target dataset, thus saving time for collecting final learning curves for the target dataset.
The learning curve extrapolation can decrease the number of epochs, thus accelerating the training process. However, it needs to fully train some architectures as the dataset for building the prediction models. The amount of data should not be too small, otherwise, the generalization ability of the prediction model may be poor. As a result, although the learning curve extrapolation can accelerate the training process, it still requires consuming much time to collect the dataset. For example, Baker \textit{et al.}~\cite{baker2017accelerating} built three architecture datasets for ResNets, MetaQNN Convolutional Neural Networks (CNNs), and LSTM with the data numbers 1,000, 1,000, and 300, respectively. The training of these architectures will consume a lot of time.
It is worth mentioning that the early-stopping strategy, a popular EEM, is considered a special case of learning curve extrapolation. Specifically, the early-stopping strategy directly regarded the obtained performance after just a few epochs as the final performance. We can view that the early-stopping strategy builds a function $p_{t}=f_{l}(-1)=p_{l}$ for a partial learning curve $f_{l}=p_1,p_2,\dots,p_l$ ($l\textless T$) to predict the final performance $p_{t}$. Because no training architecture is needed to build $p_{t}$, $\mathcal{A}^t$ is equal to $\mathcal{A}^s$ for the early-stopping strategy. The early-stopping strategy is easily embedded in the NAS algorithms. For example, Zoph \textit{et al.}~\cite{zoph2018learning} trained every architecture for only 20 epochs to reduce the search cost. The early-stopping strategy assumes that the performance ranking obtained by the partially-trained architectures is consistent with the actual performance ranking. However, the assumption is not always held because the convergence speed of different networks is not consistent.
As shown in Fig.~\ref{fig_learning_curve}, the immediate performance ranking of architecture\#1 and architecture\#2 is not consistent with the final performance ranking. This is because architecture\#1 is close to convergence at epoch $l$, while architecture\#2 still requires further training. Hence, when both architectures converge, architecture\#2 has higher performance than architecture\#1. Accordingly, the early-stopping strategy may offer an inaccurate performance estimation. In fact, Zoph \textit{et al.}~\cite{zoph2018learning} also retrain the top-$250$ architectures with the highest performance until convergence on CIFAR-10 to discover the best architecture after searching. This confirmed the inaccuracy of the early-stopping strategy.
\subsection{Few-shot Evaluation Methods}\label{few-shot}
In the few-shot methods, only a few architectures require being trained (i.e., $|\mathcal{A}^t|\textless |\mathcal{A}^s|$). Naturally, this leads to the fact that the few-shot evaluation methods consume less runtime than TEM. In the literature, the \emph{performance predictor} and \emph{population memory} are the main techniques falling into this category.
\subsubsection{Performance Predictors}\label{pp}
The performance predictor is a kind of regression model, and can directly predict the performance of the architectures in $\mathcal{A}^s$ after being trained on the architecture set $\mathcal{A}^t={\{(A_i,P_i)\}}^{|\mathcal{A}^t|}_{i=1}$ where $A_i$ denotes the $i$-th architecture and $P_i$ represents its actual performance. Because $|\mathcal{A}^t|$ is generally less than $|\mathcal{A}^s|$ and the time of predicting the performance of the architectures in $\mathcal{A}^s$ is negligible, the performance predictors would consume less runtime than TEM. The main steps of building performance predictors are as follows:
\begin{enumerate}[\textbf{Step} 1]
\item Sample and train some architectures in the predefined search space to serve as the architecture set;\label{pp_step1}
\item Encode the architectures in the architecture set;\label{pp_step2}
\item Train a performance predictor to map the encoding of architectures to the corresponding performance values;\label{pp_step3}
\item Use this trained performance predictor to estimate the performance of architectures searched.\label{pp_step4}
\end{enumerate}
We will separately discuss the existing works about steps $1-3$ in the following.
Step $4$ is not detailed because existing performance predictors are used in the same way. Then, the categorization and the analysis of performance predictors are described.
\textbf{Step 1:}
The researchers often randomly sample a lot of architectures from the search space and train them from scratch to construct the architecture set. However, the set obtained by random sampling may not cover all representative architectures, which may result in the weak generalization of the performance predictor. Consequently, some works use other sampling methods to address this aspect. For example, Dai \textit{et al.}~\cite{dai2021fbnetv3} used Quasi Monte-Carlo (QMC)~\cite{niederreiter1992random} to generate the candidate pool, and trained the predictor iteratively by picking architectures from the pool based on predicted performance. Hassantabar \textit{et al.}~\cite{hassantabar2022curious} chose Sobol sequences~\cite{burhenne2011sampling} to sample architectures from the search space because it can generate more evenly distributed samples. The aims of these works are all to generate a small architecture set that can sufficiently represent the entire search space, thus minimizing the number of architectures required to be trained.
\textbf{Step 2:}
The architecture set gained in Step~\ref{pp_step1} cannot be directly fed into the regression model and must be encoded into a form that the regression model can tackle. In practice, there are two popular ways to encode the architectures: the sequence-based scheme and the graph-based scheme~\cite{ning2020generic}.
The sequence-based scheme encodes the specific serialized information of the architectures and flattens the entire architectures to the strings. This kind of encoding scheme is broadly used in various search spaces. For example, Deng \textit{et al.}~\cite{deng2017peephole} employed a layer-based search space, proposed a uniform layer code to encode each layer to numerical vectors, and used the LSTM which is effective in processing the sequential data to integrate the information along a sequence of layers into a final string. Sun \textit{et al.}~\cite{sun2019completely} proposed a block-based search space that is composed of the ResNet block, DenseNet block, and pooling block. They encoded each block based on its kernel size, input channel, and output channel by sequence, and obtained a string at last. For the cell-based search space, Luo \textit{et al.}~\cite{luo2018neural} used the identifier of the input layers and the name of the operation applied to encode each layer to a string token. Then, a sequence of the discrete tokens was used to describe the given architectures. The sequence-based scheme is easy to implement and straightforward. However, it can only implicitly model the topological information of the given architectures. This makes performance predictors hard to capture topological information from the encoding. As a result, this scheme may lead to the reduction of the prediction ability of performance predictors for some architectures with rich topology information.
The graph-based scheme generally regards the architecture as the Directed Acyclic Graph (DAG). It explicitly represents the topological information of the architecture. For example, Wen \textit{et al.}~\cite{wen2020neural} used the adjacency matrix to describe the connections between operations and the one-hot codes to represent the operation type. To encode more information, Xu \textit{et al.}~\cite{xu2021renas} obtained the vectors of the operations, the FLating-point Operations Per second (FLOPs), and the parameter size for each operation. Then, these vectors were broadcasted into the adjacency matrix to generate the type matrix, the FLOP matrix, and the parameter matrix, respectively. At last, these matrices were concatenated as the final encoding. Ning \textit{et al.}~\cite{ning2020generic} modeled the operations as the transformation of the propagating information to mimic the data processing of the architecture. Although the graph-based scheme can better encode the topological information, it is not suitable for the layer-based and block-based search spaces which are almost the linear structure. The introduction of topological information will also bring redundant information for these search spaces~\cite{luo2020accuracy}.
\textbf{Step 3:}
The performance predictor can be essentially treated as the regression model, and its training can be formulated as Equation~(\ref{eq_PP}):
\begin{equation}\label{eq_PP}
\min _{T_{p}} \mathcal{L}\left(R\left(T_{p}, Encoder(X)\right), y\right)
\end{equation}
where $T_{p}$ is the trainable parameters of the regression model $R$, and $\mathcal{L}(\cdot)$ denotes the loss function. $Encoder(\cdot)$ denotes the encoding method. $X$ and $y$ correspond to the architecture and its performance in the architecture set, respectively. Then, we introduce the training process of the performance predictor from two aspects: the regression model and loss function.
In terms of the regression model, we illustrate it based on the encoding scheme because different models are good at handling different types of data. First, aiming at the sequenced-based schemes, the models that are suitable for the sequence data are broadly used such as LSTM~\cite{deng2017peephole,istrate2019tapas}, RNN~\cite{liu2018progressive}, Transformer. Furthermore, some non-neural models are also commonly used such as Random Forest (RF), and Gradient Boosting Decision Tree (GBDT) because of their ability to distinguish feature importance~\cite{luo2020accuracy}. As for the graph-based scheme, the Graph Convolutional Network (GCN) is commonly used because of its superiority in processing graph data~\cite{wen2020neural}.
With regard to the loss function, it can be mainly divided into two types. The first is the element-wise loss function that calculates the distance between the predicted label and the ground-truth label for every sample. The most commonly used is Mean Square Error (MSE) shown as Equation~(\ref{eq_mse}):
\begin{equation}\label{eq_mse}
\mathcal{L}_{MSE}=\frac{1}{N} \sum_{n=1}^{N}(y_{n}- f({x_{n}}))^{2}
\end{equation}
where $x_{n}$ and $y_{n}$ refer to the architecture and its corresponding performance. Furthermore, Huber loss is also a popular element-wise loss function as shown by Equation~(\ref{eq_huber}), and it can prevent the model from being dominated by outliers~\cite{dai2021fbnetv3}.
\begin{eqnarray}\label{eq_huber}
\mathcal{L}_{Huber} = \left\{\begin{array}{ll}
\frac{1}{2}(y_{n}-f(x_{n}))^{2} & \text { If }|y_{n}-f(x_{n})| \leq \delta \\
\delta|y_{n}-f(x_{n})|-\frac{1}{2} \delta^{2} & \text { otherwise }
\end{array}\right.
\end{eqnarray}
In Equation~(\ref{eq_huber}), $\delta$ can be set according to the specific situation. When the prediction deviation is less than $\delta$, $\mathcal{L}_{Huber}$ uses the squared error. Otherwise, it uses linear error.
The second type of loss function is the pair-wise ranking loss function. It is firstly proposed to be used in the performance predictor in~\cite{xu2021renas}. In practice, the performance rankings between different architectures are more important than the absolute performance of every architecture for NAS. The pair-wise ranking loss function just cares more about the performance ranking. Specifically, a typical hinge pair-wise ranking loss is represented by Equation~(\ref{eq_pw}):
\begin{eqnarray}\label{eq_pw}
\mathcal{L}_{pairwise} = \frac{1}{N} \sum_{i = 1}^{N} \frac{1}{\left|T_{i}\right|} \sum_{j \in T_{i}} \operatorname{Max}\left(0, m-\left(y_{i}-y_{j}\right)\right)
\end{eqnarray}
where $T_{i} = \{j | y_{i} > y_{j}\}$. $m$ is the comparison margin. Because the performance ranking of architectures is more important than the absolute performance for NAS, the performance predictor trained by the pair-wise ranking loss generally achieves a better result in searching high-performance architecture.
\textbf{Categorization of performance predictors:}
According to whether the performance predictors are retrained for updating during the process of NAS, the performance predictors can be divided into online and offline predictors.
For the offline predictor, the architecture-performance pairs have been completely collected before training the predictors. Note that the architecture set $\mathcal{A}^t$ of offline performance predictor is not sampled from the searched architecture $\mathcal{A}^s$. Once the predictors begin to work, there will be no new sample added to the training set. Many works employ this way to train the predictors owing to its simplicity. For example, Wen \textit{et al.}~\cite{wen2020neural} directly trained a small number of the architectures to train a GCN model. The model was utilized to rapidly predict the performance of searched architectures. However, the performance of offline predictors largely relies on the quality of the collected data. If the quality is poor, the predictor may also result in poor performance. Furthermore, once new samples are added to the dataset, the offline predictor has to be retrained, which makes the offline predictor inflexible.
Different from offline predictors, online predictors can add new samples to improve their performance after they have been trained and used. As a result, it is more flexible and practical than offline predictors. Most online predictors first train a small number of architectures to build the predictor. Then, they choose some untrained architectures and predict their performance by predictors. Finally, the architectures with the top-$k$ highest predicted performance are trained and added to the architecture set to retrain the predictor. As a result, the architectures in $\mathcal{A}^t$ are generally partially sampled from $\mathcal{A}^s$ because the newly added architecture is usually found by the NAS algorithm. Please note the runtime of the online predictor also includes the update time, in addition to the initialization time and evaluation time mentioned in Subsection~\ref{org_pd}. Specifically, the update time refers to the time for retraining the offline predictor. Many works adopted online performance predictors. For instance, Liu \textit{et al.}~\cite{liu2018progressive} started the training of the predictor with the simple architectures, and continuously used the predictor to evaluate more complex architectures. Then, they picked and trained the most promising architectures as the new samples to finetune the predictor. Wei \textit{et al.}~\cite{wei2022npenas} proposed an EC-based search strategy, and trained the predictor with the architecture-performance pairs of the initial population. Then, they trained the architectures in the next population and added them into the architecture set to retrain the performance predictor.
Depending on if the architectures are fully labeled, the existing works can also be classified into supervised and semi-supervised performance predictors.
The supervised performance predictors completely use the labeled architectures to train the predictors. Most predictors belong to the supervised one because it is simple. However, the number of labeled architectures is generally limited because the annotation of architectures is expensive. As a result, the supervised performance predictors generally try to extract more meaningful features from the limited architectures to improve their performance. For example, Chen \textit{et al.}~\cite{chen2021not} proposed an operation-adaptive attention module in the predictor to capture the relative significance between operations. Ning \textit{et al.}~\cite{ning2020generic} designed an encoding method that can model the calculation process of neural architectures to extract more meaningful information from architectures, thus improving the performance of predictors.
Unlike the supervised predictor, the semi-supervised one not only uses the labeled architectures but also utilizes the unlabeled architectures. This is because the massive unlabeled architectures can provide invaluable information to optimize the performance predictor. For example, Tang \textit{et al.}~\cite{tang2020semi} used an AutoEncoder to learn meaningful features from both the labeled and unlabeled architectures and constructed a related graph to capture the inner similarities. Then, the related graph and the acquired features were mapped into the performance value by a GCN predictor. In this way, the works can make use of the unlabeled architectures, and improve the prediction ability of the predictor. Luo \textit{et al.}~\cite{luo2020semi} trained the performance predictor with a small set of annotated architectures. Then, they used the predictor to predict the performance of a large number of unlabeled architectures. These predicted architectures were added to the architecture set to retrain the performance predictors.
\textbf{Analysis:}
Performance predictors have emerged as one of the research hotspots in EEMs in recent years\cite{deng2017peephole,liu2018progressive,sun2019surrogate}. However, a critical problem for the performance predictor is that many fully-trained architectures are required to build the performance predictor. The contradiction of this problem lies in that training a large number of architectures may conflict with the design intention for EEMs, but a small number of architectures may lead to the overfitting problem~\cite{shorten2019survey}. In practice, the performance predictors always face a lack of data. The existing works solve the problem mainly from three aspects: encoding method, model, and data. In terms of the encoding method, researchers focus on representing more useful information about the architecture by the encoding method. For example, Xu \textit{et al.}~\cite{xu2021renas} proposed a graph-based encoding method to extract more valuable information about the architecture. Specifically, they encoded not only the topological information and operation type of each node but also the FLOPs and parameters of each node. As for the model, many works focus on how to extract more meaningful features from the encoding. For example, Wang \textit{et al.}~\cite{wang2019alphax} proposed a multi-stage MLP model that used different models to predict the architectures in different ranges of performance values. With regard to the data, Liu \textit{et al.}~\cite{liu2021homogeneous} proposed a homogeneous augmentation method for the architectures to generate a large amount of labeled data in an affordable way. Although many solutions are proposed, as a common problem for machine learning, the lack of labeled architectures cannot be solved from the root.
\subsubsection{Population Memory}
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.5\textwidth]{population_memory}}
\caption{A workflow of the population memory.}\label{population_memory}
\end{figure}
Population memory is a popular method to accelerate fitness evaluation in the field of EC~\cite{coello2001multiobjective}. It is then used in the EC-based NAS to avoid evaluating the same architecture repeatedly. This is because the individuals in the previous populations may appear in the latter populations, and it is a waste of computational resources to reevaluate these individuals.
The population memory can be seen as a cache system, and it mainly shortens time by reusing the architectural information that has appeared. For the population memory, the trained architecture set $\mathcal{A}^t$ is a subset of the searched architecture set $\mathcal{A}^s$ (i.e., $\mathcal{A}^t\in \mathcal{A}^s$). The workflow of the population memory is shown in Fig.~\ref{population_memory}. For each searched architecture\#$i$, if it is in the population memory, we can directly gain its performance. Otherwise, we train architecture\#$i$ to obtain its performance and put the architecture-performance pair into the population memory. In this way, we can avoid evaluating the architectures repeatedly when facing the same architectures as the previous architectures. Many EC-based NAS adopted the method to avoid unnecessary costs. For example, Fujino \textit{et al.}~\cite{fujino2017deep} used a population memory to store the performance value of the individuals, and retrieved from memory when the architectures with the same encoding as the previous architectures appear. Similarly, Sun \textit{et al.}~\cite{sun2020automatically} designed a global cache system to record the hash code and the performance of the architectures. They obtained the performance when facing an architecture whose encoding has been stored in the cache.
Compared with other EEMs, population memory allows for a more accurate evaluation of performance because the obtained performance by population memory is equal to the actual performance. However, the population memory lacks flexibility. Specifically, it can not be used to evaluate any architecture searched. This is because the prerequisite for evaluating an architecture with population memory is that the architecture has been trained before. However, in practice, the repetitive architectures only account for a small portion of all architectures searched~\cite{sun2020automatically}.
\subsection{One-shot Evaluation Methods}\label{one-shot}
The one-shot method only requires training one neural architecture during the whole search process (i.e., $|\mathcal{A}^t|=1$). It it is also called the weight-sharing method in the field of NAS.
Because only one network (also called ``supernet'') needs to be trained, the method is cost-saving. Depending on whether architecture search and supernet training are coupled, the existing one-shot methods can be classified into \emph{path-based methods} and \emph{gradient-based methods}.
\subsubsection{Path-based Methods}\label{one-shot-path}
The path-based method trains a supernet that contains all candidate architectures (i.e., subnet). The weights of subnets can directly be extracted from the supernet. Each path refers to a subnet in the supernet. An example is shown in Fig.~\ref{fig_path-based}.
Specifically, only the weights of the supernet are obtained by training. The weights of the subnet are directly extracted from the supernet. Then, the performance of these subnets can be obtained by infering on the validation dataset. Because only one architecture requires being trained (i.e., $|\mathcal{A}^t|=1$) and the inference on validation dataset is quick, path-based methods can largely accelerate the performance evaluation.
The main workflow of the path-based method is as below:
\begin{enumerate}[\textbf{Step} 1]
\item Design the supernet subsuming all candidates;
\item Train the supernet from scratch by a path sampling strategy;\label{path-2}
\item Predict the searched architectures by inheriting the weights from the supernet and inferencing the performance on the validation dataset.
\end{enumerate}
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.47\textwidth]{path-based.pdf}}
\caption{An example of the path-based method. The grey lines and nodes in the subnet represent inactive connections and nodes, respectively. The red line and blue node in the subnet represent active connections and nodes, respectively. The subnet extracts weights from the supernet, and then inferences on the validation dataset to obtain its performance.}\label{fig_path-based}
\end{figure}
Among these steps, the training of the supernet (i.e., Step~\ref{path-2}) is critical. This is because the weights of the supernet are obtained through training and the weights of the subnet is extracted from the supernet. In practice, the weights of the supernet are almost not directly optimized. The reason is that the weights in the supernet are deeply coupled with the architecture of the supernet. Accordingly, the supernet is not robust to the changes in the architecture~\cite{bender2018understanding,guo2020single}. This may lead to a low correlation between the performance obtained by inheriting the weights and the actual performance
Many path sampling strategies are proposed to decide which path(s) to train in every epoch of the supernet training to decouple the weights and architecture of the supernet. For example, Bender \textit{et al.}~\cite{bender2018understanding} used the path dropout strategy during the training. Specifically, the strategy randomly drops a subset of the operations on one batch of data to alleviate the highly coupled problem. The probability of the dropout is determined by the parameter dropout rate. However, experiments in~\cite{bender2018understanding} indicated the training is sensitive to the dropout rate, which makes the training of supernet complicated. Guo \textit{et al.}~\cite{guo2020single} proposed a uniform path sampling strategy that can overcome the drawbacks of the dropout strategy. The uniform path strategy randomly sampled an architecture in the supernet in each iteration of the optimization. Only the weights of the sampled architecture are updated in each iteration. Compared with the path dropout strategy, the uniform path sampling strategy is hyperparameter-free. In this way, it avoids the defect of the path dropout strategy which is sensitive to the dropout rate.
Chu \textit{et al.}~\cite{chu2021fairnas} pointed out that the inherent unfairness in the uniform path sampling strategy may lead to biased evaluation. They thus proposed strict fairness sampling based on the uniform path sampling strategy. The strict sampling strategy samples $m$ ($m$ is the number of choice blocks per layer) at each epoch. In this strategy, all choice blocks are activated only once and are optimized only on one batch of data.
Zhang \textit{et al.}~\cite{zhang2020one} designed the novelty-driven sampling strategy. Only the weights of the architectures sampled by novelty search were optimized.
You \textit{et al.}~\cite{you2020greedynas} proposed a multi-path sampling strategy with rejection to filter the weak paths by evaluating them on a portion of the validation dataset. During the training of supernet, only those potential paths were trained. In summary, these path sampling strategies can optimize the weights of various subnets in a more balanced way to effectively alleviate the weight coupling problem.
\begin{figure*}[htbp]
\centerline{\includegraphics[width=0.9\textwidth]{multi-model_forgetting.pdf}}
\caption{An example of the multi-model forgetting problem.}\label{fig_multi_model}
\end{figure*}
Although the prediction accuracy of the path-based methods is improved with the help of these path sampling strategies, they still face the multi-model forgetting problem~\cite{benyahia2019overcoming}. Specifically, the multi-model forgetting problem refers to the performance decline of the previous-trained subnets when training subsequent subnets. This is because of the overwriting of shared weights when training the subsequent subnets with partially-shared weights one by one. An example of the multi-model forgetting problem is shown in Fig.~\ref{fig_multi_model}. Specifically, model A and model B are both the subnet of the supernet and share weights with the supernet. Model A is firstly trained and the weights of itself are optimized in the supernet. Then, model B which shares partial weights with model A is trained. During the training of model B, the performance of model A declines. The phenomenon is called multi-model forgetting. The multi-model forgetting problem will deteriorate the predictive ability of the supernet. As a result, although the path-based methods can greatly reduce computational time, they may result in unreliable performance ranking.
Many solutions are proposed to alleviate the multi-model forgetting problem. For example, Benyahia \textit{et al.}~\cite{benyahia2019overcoming} proposed a statistically-justified weight plasticity loss to regularize the optimization of the shared parameters. In this way, they can preserve the shared parameters that were important for the previous model. Zhang \textit{et al.}~\cite{zhang2020forgetting} formulated the supernet training as a constrained continual learning optimization problem. Specifically, during the training of the current architecture, they constrained the training loss of all previous architectures in the current step is less than that in the last step. However, the number of constraints increases linearly as the increase of previous architectures. This makes it difficult to consider all constraints. Hence, they proposed a greedy novelty search method to select a subset of existing constraints from previous architectures. In this way, they could prevent the performance decline of previous architectures during the training of the current architecture.
Although the above works can alleviate the phenomenon of multi-model forgetting, they cannot solve the problem from the root. This is because the problem is caused by the mechanism the path-based method trains the supernet to evaluate subnets. As a result, the ranking correlation between the performance obtained by the path-based method and the actual performance may be poor.
\subsubsection{Gradient-based Methods}\label{one-shot-gradient}
Unlike the path-based methods, the gradient-based methods decouple the supernet training and architecture searching. Specifically, they relax the discrete search space to be continuous, and jointly optimize the architecture and the supernet weights by the gradient-based search strategy. The most famous work falling into this category is DARTS~\cite{liu2018darts} proposed in 2019. Then, the DARTS approach became the dominant approach to gradient-based methods owing to its effectiveness. Specifically, DARTS used a cell-based search space and regarded the supernet as a DAG with $N$ nodes. Each node $x^{(i)}$ is a latent representation such as the feature map, and each edge $(i,j)$ refers to the operation $o^{(i,j)}$ that can transform $x^{i}$. Each intermediate node is computed based on all of its predecessors as shown in Equation~(\ref{eq_pre}):
\begin{eqnarray}\label{eq_pre}
x^{(j)} = \sum_{i<j} O^{(i, j)}\left(x^{(i)}\right)
\end{eqnarray}
To relax the discrete search space to be continuous, the categorical choice of a particular operation is continuously relaxed by employing a softmax function over all the possible operations as shown in Equation~(\ref{eq_op}):
\begin{eqnarray}\label{eq_op}
\bar{o}^{(i, j)}(x) = \sum_{o \in \mathcal{O}} \frac{\exp \left(\alpha_{o}^{(i, j)}\right)}{\sum_{o^{\prime} \in \mathcal{O}} \exp \left(\alpha_{o^{\prime}}^{(i, j)}\right)} o(x)
\end{eqnarray}
where $\mathcal{O}$ is a set of candidate operations. $\alpha^{(i,j)}$ denotes the operation mixing weights (i.e., architecture parameter) for a pair of nodes $(i,j)$. An architecture can be obtained by replacing each mixed operation $\overline{o}^(i,j)$ with the most likely operation, i.e., $o^{(i, j)}=\operatorname{argmax}_{o \in \mathcal{O}} \alpha_{o}^{(i, j)}$. Then, DARTS aims to jointly learn the architecture parameters $\alpha$ and the weights $\omega$ which determine the training loss $\mathcal{L}_{train}$ and validation loss $\mathcal{L}_{val}$. This can be formulated as a bilevel optimization with $\alpha$ as the upper-level variable while $\omega$ as the lower-level variable problem, modeled by Equation~(\ref{eq_biop}):
\begin{eqnarray}\label{eq_biop}
\left\{\begin{matrix}
\min _{\alpha} \mathcal{L}_{val}\left(w^{*}(\alpha), \alpha\right) \\
\text { s.t. } w^{*}(\alpha)=\operatorname{argmin}_{w} \mathcal{L}_{train}(w, \alpha)
\end{matrix}\right.
\end{eqnarray}
By this way, we can find $\alpha^*$ that minimize the $\mathcal{L}_{val}(w^{*}(\alpha),\alpha)$ where the weights $\omega^{*}$ minimize the $\mathcal{L}_{train}(w, \alpha)$.
Despite the efficiency of the gradient-based methods, they always face three challenges:
\textbf{Huge memory}: The gradient-based methods always suffer from large memory. This is because the joint optimization which not only trains a supernet but also searches for the architectures. Many works try to solve the problem. For example, Cai \textit{et al.}~\cite{cai2018proxylessnas} leveraged binarized architecture parameters to guarantee only one path of activation is active in memory at runtime. Xu \textit{et al.}~\cite{xu2019pc} randomly sampled a subset of channels as the proxy of all the channels. In this way, memory consumption is naturally reduced. However, the architecture parameters are optimized by randomly sampled channels, which may lead to undesired fluctuation in search. They thus used edge normalization, which adds a new set of edge-level parameters to improve the stability of the search. Similarly, Xue \textit{et al.}~\cite{xue2022partial} saved the memory by adopting the subset of the channels. The difference is that they proposed an attention mechanism, and selected the channels with higher attention weights. The method can better transmit important feature information into the search space and prevent the instability of the search.
\textbf{Poor generalization}: The gradient-based method generally suffers from poor generalization on various datasets of the searched architectures. Specifically, the searched architecture cannot perform as well as on the dataset used in the search on other datasets. To solve the problem, Li \textit{et al.}~\cite{li2020adapting} introduced the idea of domain adaptation. They improved the generalizability of architectures by minimizing the generalization gap between domains. Liu \textit{et al.}~\cite{liu2021mixsearch} proposed a novel method to mix the data from multiple tasks and domains into a composited dataset. Ye \textit{et al.}~\cite{ye2022beta} proposed a regularization method that can maintain a small Lipschitz of the searched architectures. This is because the architectures have good generalization ability if its Lipschitz is small.
\textbf{Performance collapse}: The gradient-based method lacks stability and frequently faces performance collapse. This is because DARTS tends to accumulate parameter-free operations such as skip connection, which often leads to rapid gradient descent~\cite{arber2020understanding,chu2020darts}. However, learnable operations such as convolution are the better choice for improving the expression ability of the neural network.
Existing works to alleviate this problem can be divided into four categories: \emph{limiting the number of skip connections}, \emph{regularizing relevant indicators}, \emph{modifying the cell or the discretization process}, and others. Regarding limiting the number of skip connections, Chen \textit{et al.}~\cite{chen2019progressive} inserted operation-level dropout after each skip connection operation to controll the number of skip connections. Liang \textit{et al.}~\cite{liang2019darts+} stopped the search when meeting two certain criteria. Specifically, the first criterion is that there are two or more skip connections in one normal cell. The second one is that the ranking of the architecture parameters for learnable operations becomes stable for a fixed number of epochs. Despite of their simplicity and effectiveness, simply limiting the number of skip connections may reject some potentially high-performance architectures. Regarding regularizing relevant indicators, Zela \textit{et al.}~\cite{arber2020understanding} observed that the Hessian eigenvalues of the validation loss can be treated as the indicator for the performance collapse. Consequently, they proposed an early-stopping strategy based on tracking the Hessian eigenvalue. Chen \textit{et al.}~\cite{chen2020stabilizing} proposed regularization techniques that were achieved by random smoothing or adversarial attack to implicitly optimize the Hessian eigenvalue. The biggest limitation of the indicator-based methods is that they rely heavily on the quality of the indicator. Regarding modifying the cell or the discretization process, Chu \textit{et al.}~\cite{chu2020darts} added an auxiliary skip connection to the cell structure to ensure a fairer competition for all operations. Chu \textit{et al.}~\cite{chu2020fair} applied an independent sigmoid function for each operation. A zero-one loss was also designed to push architectural weights towards zero or one to make the competition of operations fairer.
There are also some other works not falling into the above categories. For example, Hong \textit{et al.}~\cite{hong2022dropnas} proposed grouped operation dropout. Specifically, they divided existing operations in DARTS into parameterized and non-parameterized groups. Then, they randomly zeroed out the output of the operations in the two groups. The architecture parameter and weights of these dropped operations are also not optimized during backpropagation. Gu \textit{et al.}~\cite{gu2021dots} further proposed a grouped operation dropout. It divided the existing operations into topology-related and topology-agnostic groups. Ye \textit{et al.}~\cite{ye2022beta} proposed a simple-but-efficient regularization method. The method can regularize the activated architecture parameters to promote fair competition in all operations. Although many solutions have been proposed, these methods are difficult to fundamentally overcome the performance collapse. This is because the problem is caused by the working mechanism of DARTS.
In summary, the one-shot method can largely reduce the computational time by only training one supernet. However, the path-based method may experience the phenomenon of multi-path forgetting, and the gradient-based method faces the problem of performance collapse. These problems may lead to inaccurate ranking. Hence, the one-shot method may fail to reflect the true ranking of architectures. This also affects the performance of the discovered architecture in NAS. The evidence is that the architecture searched by the one-shot method performs sometimes worse than those discovered by the random search strategy~\cite{yang2019evaluation}.
\subsection{Zero-shot Evaluation Methods}\label{zero-shot}
The zero-shot methods do not need to train any architectures (i.e., $\mathcal{A}^t=\emptyset$).
It can directly score architectures for their performance without any training. Because it does not require any training and the score of architectures is time-saving, it has an extremely low cost. Please note that the zero-shot evaluation methods cannot directly evaluate the absolute performance of a given architecture (i.e., the score for performance is not equal to the performance). It can only rank the architectures based on the score of the performance. The zero-shot methods can be mainly classified into \emph{parameter-level methods} and \emph{architecture-level methods}.
\subsubsection{Parameter-level Methods}\label{zero_shot_parameter}
The parameter-level methods first use the indicator to measure the saliency (i.e., importance) of certain parameters. Then, they score for the entire architecture by aggregating the saliencies of certain parameters. This kind of method only required a minibatch of data and a forward/backward propagation pass to calculate the indicators for certain parameters. Because it does not need to train the architecture, it can consume much less runtime than TEM. Generally, the saliency indicators generally come from the network pruning literature. They are originally used to evaluate the importance of parameters and thus remove the unimportant ones.
There have been some works proposed in this category. For example, Abdelfattah \textit{et al.}~\cite{abdelfattah2020zero} adopted a series of pruning-at-initialization metrics includes $grad\_norm$, $snip$~\cite{lee2018snip}, $grasp$~\cite{wang2020picking}, $fisher$~\cite{theis2018faster,turner2019blockswap}, $synflow$~\cite{tanaka2020pruning} to measure the saliency of parameters. Then, they summed up the metric value of all parameters to score for the architecture performance.
Specifically, grad\_norm uses the Euclidean norm of the gradient after a forward/backward propagation pass of a minibatch of training data to measure the parameters. Snip calculates the saliency of a specific parameter by approximating the change in cross-entropy loss when the parameter is pruned. Synflow is a modified version of snip with the change that the loss in synflow is the product of all parameters in the network. Grasp calculates the saliency by measuring the approximate change in gradient norm when the parameter is removed. Fisher measures the approximate loss change when activations channels are removed.
Despite the cost-saving of parameter-level methods, they are proven inaccurate. Simply summing up the saliencies of each parameter to measure the performance of architectures is problematic. This is because the saliency only reflects the impact of the parameter on that architecture. In fact, a recent work~\cite{ning2021evaluating} also pointed out that the existing parameter-level method adapted is not suitable for ranking architectures. Their ranking qualities even cannot surpass those of parameter size or FLOPs. This proves the unreliability of the parameter-level methods on the other hand.
\subsubsection{Architecture-level Methods}\label{zero_shot_architecture}
The architecture-level methods score the architectures by measuring the properties positively related to architecture performance. Because they do not need to train any architecture, they can largely save time for performance evaluation.
Many architecture-level methods have been proposed to efficiently evaluate the performance of the architecture. For example, Mellor \textit{et al.}~\cite{mellor2020neural, mellor2021neural} proposed to quantify the activation overlap of different inputs for a network to score the untrained networks. The learnability of architectures is more strong when the activation for different inputs is well separated. They proposed an indicator based on input Jacobian correlation in~\cite{mellor2020neural}. Furthermore, they also devised using the Hamming distance to judge how dissimilar the two inputs are~\cite{mellor2021neural}. Xu \textit{et al.}~\cite{xu2021knas} assumed that the gradients can be used to evaluate the random-initialized networks. This is based on the observation that gradients can directly decide the convergence and generalization results.
They presented the gradient kernel to take each layer as a basic unit. After that, the mean of the Gram matrix for each layer was computed to score the network.
Lin \textit{et al.}~\cite{lin2021zen} averaged the Gaussian complexity of linear function in each linear region to measure the network expressivity.
Zhou \textit{et al.}~\cite{zhou2022training} observed the synaptic diversity of multi-head self-attention in Vision Transformer (ViT) affects the performance notably. Based on the observation, they proposed an indicator to rank ViT architectures from the perspectives of synaptic diversity and synaptic saliency. Chen \textit{et al.}~\cite{chen2021neural} measured the architecture performance by analyzing the spectrum of the neural tangent kernel and the number of linear regions in the input space. Shu \textit{et al.}~\cite{shu2021nasi} used the neural tangent kernel to characterize the performance of the architectures.
The architecture-level methods are usually motivated by some theoretic studies on neural networks. They proposed indicators to judge the trainability, learnability, generalization, or expressivity of the architectures and ranked the architectures. Despite their efficiency, the architecture-level methods always lead to inaccurate performance ranking. This is because they only roughly measure the properties of architectures that are positively related to architecture performance.
In summary, the zero-shot methods can further reduce the computational time compared with other methods. However, the performance of the zero-shot method is usually not good enough in practice. Moreover, the robustness of the zero-shot methods can not be guaranteed. The performance fluctuated dramatically among different tasks.
\begin{table*}[htp]
\caption{Summary of different EEMs in NAS including their principle and relative references.}
\begin{center}
\begin{tabular}{|p{2cm}|p{3.4cm}|p{7.2cm}|p{3.4cm}|}
\hline
Category & Subcategory & Principle & References \\ \hline
\multirow{4}{*}{$N$-shot method} & Downscaled dataset method & Training time reduced by training on the down-sampled dataset. & \cite{liu2019deep,chrabaszcz2017downsampled,hassantabar2022curious,klein2017fast,shu2020automatically,sapra2020constrained,wang2020particle,xu2021partially,liu2017structure} \\ \cline{2-4}
& Downscaled model method & Training time reduced by downscaling architectures during the search process. & \cite{zoph2018learning, real2019regularized, liu2018darts} \\ \cline{2-4}
& Learning curve extrapolation & The performance is extrapolated after only training for a few epochs. & \cite{baker2017accelerating,klein2016learning,domhan2015speeding,rawal2018nodes,kim2022two} \\ \cline{2-4}
& Network morphism & Training time reduced by inheriting weights from the parent model to reduce epochs. & \cite{cai2018efficient,elsken2018efficient,chen2015net2net,wei2016network,cai2018path,elsken2017simple,zhang2022self} \\ \hline
\multirow{2}{*}{Few-shot method} & Performance predictor & Training a regression model on architecture-performance pairs and Directly predict the performance of DNNs using the regression model. & \cite{deng2017peephole,liu2018progressive,sun2019surrogate,istrate2019tapas,tang2020semi,luo2020accuracy,sun2021novel,huang2022arch,white2021bananas,dudziak2020brp,chen2021contrastive,hassantabar2022curious,mauch2020efficient,dai2021fbnetv3,liu2021fox,li2021generic,li2020gp,wen2020neural,lu2021tnasp,chen2021not,wei2022npenas,peng2022pre,wang2021rank,xu2021renas,wu2021stronger,zhang2021neural} \\ \cline{2-4}
& Population memory & Reuse the performance of privious DNNs when discovering the DNNs that have been trained. & \cite{xie2022benchenas,sun2020automatically,fujino2017deep,johner2019efficient,miahi2022genetic} \\ \hline
\multirow{2}{*}{One-shot method} & Path-based method & Only the supernet requires being trained, and the weights of candidate architectures can be extracted from the supernet. & \cite{bender2018understanding,cheng2022scalenas,zhang2022evolutionary,yu2019autoslim,chen2019detnas,guo2020single,peng2020cream,huang2021searching,cheng2020scalenas,peng2021rsbnet,xia2022progressive,dong2022prior,guo2020powering,huang2021ponas,liang2021opanas,chen2021one,dong2019one,zhang2020one,li2022one,cai2020once,chu2020mixpath,yan2021lighttrack,chen2021iterative,li2020improving,ma2021simplify,you2020greedynas,chu2021fairnas,sinha2021evolving,shen2020bs,shi2020bridging,wang2021attentivenas,liu2020comprehensive} \\ \cline{2-4}
& Gradient-based method & Jointly optimize the architecture parameter and the supernet weights by bilevel optimization. & \cite{liu2018darts,ren2022dartsrepair,wang2022eautodet,wang2022fp,zhang2022memory,yue2022effective,zhang2022differentiable,luo2022you,xu2022dnas,yu2022cyclic,zhang2022balenas,huang2022greedynasv2,dong2019searching,zhang2020overcoming,zhang2020forgetting,hu2021improving,pan2022distribution,arber2020understanding,zhang2021neural,zheng2021ad,akimoto2019adaptive,nakai2020att,fil2021darts,chu2020darts,liang2019darts+,ren2022core,gu2021dots,hong2022dropnas,hasan2021darts,zhou2021ec,li2022enhancing,chu2020fair,wu2019fbnet,wan2020fbnetv2,li2020geometry,bi2020gold,nayman2021hardcore,lv2021heart,zhang2021idarts,xue2021idarts,wang2021mergenas,gao2022mr,hsieh2021ms,zhu2021operation,xu2019pc,li2020pd,tang2021probeable,chen2021progressive,chen2019progressive,cai2018proxylessnas,jin2019rc,wang2021rethinking,xue2021rethinking,zhang2021robustifying,zhang2021rs,hundt2019sharpdarts,xie2018snas,hou2021single,bi2019stabilizing,chen2020stabilizing,chen2020darts,li2021differentiable,shi2021darts,shi2021efficient,li2021one} \\ \hline
\multirow{2}{*}{Zero-shot method} & Parameter-level method & Aggregate the saliencies of all parameters to measure the entire DNN without training. & \cite{abdelfattah2020zero,shu2022unifying} \\ \cline{2-4}
& Architecture-level method & Estimate the performance of DNNs by measuring the properties positively related to architecture performance without training. & \cite{mellor2020neural,mellor2021neural,lin2021zen,chen2021neural,shu2021nasi,xu2021knas,zhou2022training} \\ \hline
\end{tabular}
\end{center}\label{categories_AC}
\vspace{-0.3cm}
\end{table*}
\section{Evaluation}\label{evaluation}
In this section, we first discuss the evaluation metrics for the EEMs in Section~\ref{eva_metric}. Then, we introduce the commonly-used benchmark datasets in Section~\ref{eva_benchmark}. Finally, we report the metric value of various EEMs on the benchmark datasets in Section~\ref{eva_comp}.
\subsection{Evaluation Metrics}\label{eva_metric}
Based on the no-free-lunch theory, the EEMs may perform inaccurately compared with TEM though their speed is much faster. As a result, we mainly measure the EEMs in two folds: runtime and prediction ability.
The runtime of an EEM is generally measured by GPU day (GPU day = The number of GPUs $\times$ The number of days)~\cite{sun2019completely}.
As for the measurement of prediction ability, the commonly used metrics can be mainly divided into three categories: correlation-based, ranking-based, and top-based. For the convenience of discussion, we denote the ground-truth performance and the predicted performance of architectures ${\{A_{i}\}}_{i=1}^{N}$ as ${\{y_{i}\}}_{i}^{N}$ and ${\{\hat{y}_{i}\}}_{i}^{N}$, respectively. The real performance ranking of the architecture $A_{i}$ (i.e., the ranking of $y_{i}$ among ${\{y_{i}\}}_{i}^{N}$) is $r_{i}$, while the predicted performance ranking for $A_{i}$ (i.e., the ranking of $\hat{y}_{i}$ among ${\{\hat{y}_{i}\}}_{i}^{N}$) is $\hat{r}_{i}$. Next, we will discuss the three categories of evaluation metrics.
The correlation-based metric mainly calculates the correlation between the ground-truth performance and the predicted performance. It mainly includes the Pearson coefficient and the Coefficient of Determination (i.e., $\mathcal{R}^{2}$), which are formulated by Equation~(\ref{eq_pearson}) and (\ref{eq_r2}), respectively.
\begin{eqnarray}\label{eq_pearson}
\mathcal{M}_{pearson} = \frac{\sum_{i = 1}^{N}\left(y_{i}-\bar{y}\right)\left(\hat{y}_{i}-\bar{\hat{y}}\right)}{\sqrt{\sum_{i = 1}^{N}\left(y_{i}-\bar{y}\right)^{2} \sum_{i = 1}^{N}\left(\hat{y}_{i}-\bar{\hat{y}}\right)^{2}}}
\end{eqnarray}
\begin{equation}\label{eq_r2}
\mathcal{M}_{R^{2}}=1-\frac{\sum_{i=1}^{N}\left(y_{i}-\hat{y}_{i}\right)^{2}}{\sum_{i=1}^{N}\left(y_{i}-\bar{y}_{i}\right)^{2}}
\end{equation}
For both metrics, the closer to one the value is, the more accurate the EEM is.
The ranking-based metric calculates the correlation between the real performance ranking and the predicted one.
Kendall's Tau ranking correlation (KTau)~\cite{sen1968estimates} and Spearman's ranking correlation (SpearmanR)~\cite{hauke2011comparison} are frequently-used ranking-based metrics. The KTau refers to the relative difference of concordant pairs and discordant pairs, and is shown in Equation~(\ref{eq_Ktau}):
\begin{equation}\label{eq_Ktau}
\mathcal{M}_{KTau}=2 \times \frac{\text { number of concordant pairs }}{N(N-1) / 2}-1
\end{equation}
where the concordant pair means that the predicted rankings and the ground-truth rankings of a given pair are the same.
On the other hand, the SpearmanR is calculated based on the difference between the predicted ranking and the real ranking. It is shown in Equation~(\ref{eq_SpearmanR}):
\begin{eqnarray}\label{eq_SpearmanR}
\mathcal{M}_{SpearmanR} = 1-\frac{6 \sum_{i = 1}^{N} (r_i-\hat{r}_i )^{2}}{N\left(N^{2}-1\right)}
\end{eqnarray}
For both KTau and SpearmanR, the value closer to one means that the EEM is more accurate.
The top-based metric measures the ability of the EEM to discover the best architecture, which is important for NAS. The top-based metric is mainly composed of N@K and Precision@K. In specific, N@K refers to the best ground-truth ranking among the predicted top-K architectures. The value of N@K is a positive integer, and greater than or equal to one. A lower value of N@K means that the EEM is more capable of finding the best architecture. The Precision@K is the proportion of the ground-truth top-K architectures in the predicted top-K proportion architectures. Different from N@K, the value closer to one means the EEM is better at discovering good architectures.
\begin{table*}[htp]
\caption{The characteristics of NAS benchmarks. $|V|$ refers to the number of nodes, and $|E|$ denotes the number of edges. Please note that NAS-Bench-301 provides 60,000 annotated architectures and a performance predictor which can predict all $10^{18}$ architecture in the DARTS search space.}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
Benchmark & Year & Task & Dataset & Search space & OON/OOE &\#Architecture \\ \hline
NAS-Bench-101 & 2019 & Image classification & CIFAR-10 & $|V|\leq 7$,\#ops=3,$|E|\leq 9$&OON& 423,624 \\ \hline
NAS-Bench-201 & 2020 & Image classification & \makecell[c]{CIFAR-10, CIFAR-100,\\ ImageNet-16-120} & $|V|=4$,\#ops=5 &OOE& 15,625 \\ \hline
NAS-Bench-301 & 2020 & Image classification & CIFAR-10 & $|V|\leq 7$,\#ops=7 &OOE& 60,000/$10^{18}$ \\ \hline
NAS-Bench-NLP & 2020 & Natural language processing & Penn Tree Bank & $|V|\leq 24$,\#ops=7 &OOE& 14,322 \\ \hline
NAS-Bench-ASR & 2021 & Automatic Speech Recognization & TIMIT audio dataset & $|V|\leq 4$,\#ops=7,$|E|\leq 9$ &OOE& 8,242 \\ \hline
\end{tabular}
\end{center}\label{tbl_benchmark}
\vspace{-0.3cm}
\end{table*}
\subsection{Benchmark Datasets}\label{eva_benchmark}
EEMs are often evaluated on benchmark datasets to perform fair comparisons. Specifically, the benchmark datasets include a lot of architecture-performance pairs in the specific search spaces. The architecture-performance pairs in these benchmark datasets can be used to evaluate the prediction accuracy of EEMs. The popular benchmark datasets includes NAS-Bench-101~\cite{ying2019bench}, NAS-Bench-201~\cite{ying2019bench}, NAS-Bench-301~\cite{siems2020bench} for image classification tasks, NAS-Bench-NLP~\cite{klyuchnikov2022bench} for natural language processing tasks, and NAS-Bench-ASR~\cite{mehrotra2020bench} for automatic speech recognition tasks, etc.
These benchmark datasets all use the cell-based search space. Specifically, they stack cells to form the architecture, and only the structure of the cell can be searched. Each cell can be treated as a DAG. Based on the position of the operation, the cell-based search space can be divided into Operation on Node (OON) and Operation on Edge (OOE) search spaces. Specifically, the node of the DAG is regarded as the operation while the edge is treated as the connections between operations for OON. The edge of DAG represents the operations in architecture while the node is regarded as the connection between operations for OOE. We summarized the mentioned benchmarks in Table.~\ref{tbl_benchmark}.
Take NAS-Bench-101 as an example, it is proposed in 2017 and aims at the tasks of image classification. The search space belongs to OON, and the number of choice operations is three. The number of nodes is up to seven, and the number of edges is equal to or less than nine. The search space includes 423,624 architectures and provides the performance of these architectures on CIFAR-10.
\subsection{Comparisons on Benchmark Datasets}\label{eva_comp}
In order to make audiences more intuitively compare the performance of each method, we collect the KTau results of various methods on NAS-Bench-101, NAS-Bench-201, and NAS-Bench-301. Please note the results are collected from the original paper of these methods or some paper that compare these methods or the results of the performed experiments. We do not report the results of the downscaled model method, network morphism, gradient-based method, and population memory. This is because we do not have enough computing resources to re-evaluate the downscaled models in these benchmark datasets. Furthermore, network morphism, gradient-based method, and population memory cannot be used to accelerate all architectures in these benchmark datasets. Specifically, network morphism can only be applied to accelerate the training of the child networks transformed from the parent network. However, not all architectures can be transformed from the parent network. The gradient-based method is also a search strategy. It can only be used to accelerate the searched architectures. Performance memory can only be applied to query the performance of architectures that have been evaluated. As a result, the experimental results of these methods are not reported. The comparison results are shown in Table.~\ref{analyze}.
\begin{table*}[htp]
\caption{The Kendall's Tau (KTau) value of various methods on NAS-Bench-101 (NB101), NAS-Bench-201 (NB201), and NAS-Bench-301 (NB301).}
\begin{center}
\begin{tabular}{|p{1.9cm}|p{3.2cm}|p{2.2cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{5cm}|}
\hline
Category & Subcategory & Method & NB101 & NB201 & NB301 & Note\\ \hline
\multirow{4}{*}{N-shot method} & \multirow{2}{*}{Downscaled dataset method} & Proxy dataset & - & 0.827 & - & CIFAR-10 is the proxy dataset, and ImageNet-16-120 is the original dataset. \\ \cline{3-7}
& & Subset of dataset & - & 0.867 & - & The KTau between the validation accuracy on CIFAR-10 (the subset of CIFAR-100) and CIFAR-100. \\ \cline{2-7}
& \multirow{2}{*}{Learning curve extrapolation} & Early stopping(1/4) & - & 0.602 & 0.614 & Terminate training when the epoch is one quarter of the fully trained epoch. \\ \cline{3-7}
& & Early stopping(1/2) & 0.438 & 0.663 & 0.662 & Terminate training when the epoch is one half of the fully trained epoch. \\ \hline
\multirow{6}{*}{Few-shot method} & \multirow{4}{*}{Performance predictor} & NeuralPredictor~\cite{wen2020neural} & 0.679 & 0.646 & - & 424 and 1,564 samples are used as the training data in NAS-Bench-101 and NAS-Bench-201, respectively. \\ \cline{3-7}
& & Peephole~\cite{deng2017peephole} & 0.4556 & - & - & 424 samples are used as the training data. \\ \cline{3-7}
& & E2EPP~\cite{sun2019surrogate} & 0.5038 & - & - & 424 samples are used as the training data. \\ \cline{3-7}
& & ReNAS~\cite{xu2021renas} & 0.657 & - & - & 424 samples are used as the training data. \\ \cline{3-7}
& & HOP~\cite{chen2021not} & 0.813 & 0.897 & - & 381 and 781 samples are used as the training data in NAS-Bench-101 and NAS-Bench-201, respectively. \\ \cline{3-7}
& & TNASP~\cite{lu2021tnasp} & 0.722 & 0.726 & - & 424 and 1,564 samples are used as the training data in NAS-Bench-101 and NAS-Bench-201, respectively. \\ \cline{1-7}
\multirow{2}{*}{One-shot method} & \multirow{2}{*}{Path-based method} & OSNAS~\cite{ning2021evaluating} & 0.446 & 0.744 & 0.548 & Using MC sample in supernet training. \\ \cline{3-7}
& & FairNAS~\cite{chu2021fairnas} & - & 0.706 & 0.527 & - \\ \hline
\multirow{8}{*}{Zero-shot method} & \multirow{6}{*}{Parameter-level method} & synflow~\cite{abdelfattah2020zero} & -0.063 & 0.573 & 0.201 & - \\ \cline{3-7}
& & grad\_norm~\cite{abdelfattah2020zero} & -0.276 & 0.401 & 0.070 & - \\ \cline{3-7}
& & snip~\cite{abdelfattah2020zero} & -0.206 & 0.402 & 0.050 & - \\ \cline{3-7}
& & grasp~\cite{abdelfattah2020zero} & -0.266 & 0.348 & 0.365 & - \\ \cline{3-7}
& & fisher~\cite{abdelfattah2020zero} & -0.202 & 0.362 & -0.158 & - \\ \cline{3-7}
& & plain~\cite{abdelfattah2020zero} & 0.240 & 0.311 & 0.394 & - \\ \cline{2-7}
& \multirow{2}{*}{Architecture-level method} & jacob\_cov~\cite{mellor2020neural} & 0.066 & 0.608 & 0.230 & - \\ \cline{3-7}
& & relu\_logdet~\cite{mellor2021neural} & 0.290 & 0.611 & 0.539 & - \\ \hline
\end{tabular}
\end{center}\label{analyze}
\vspace{-0.3cm}
\end{table*}
\section{Challenges and Future Directions}\label{direction}
Despite the effectiveness of the existing EEMs, there are still future directions that require being explored.
\subsection{Effectiveness Validation}
The current effectiveness validation for EEMs is not convincing for most researchers. Specifically, the benchmarking datasets are used for effectiveness validation and fair comparisons. However, the existing architecture dataset cannot conform to the practical sceneries because of their small scales. For example, NAS-Bench-101 only involves three types of operations (i.e., convolution 1$\times$1, convolution 3$\times$3, and max pooling). It limits the max number of nodes and edges to seven and nine, respectively. The NAS-Bench-101 search space is so small that it only consists of $423,624$ architectures. NAS-Bench-201 only includes four different operations, and the architecture only has four vertices. The NAS-Bench-201 search space only includes 15,625 architectures, is even smaller than NAS-Bench-101. In contrast, the search spaces applied in real application scenarios, such as the search space of NASNet, MobileNet, or Transformer, are more complex. They are generally several orders of magnitude more than NAS-Bench-101 and NAS-Bench-201 in quantity. For example, the NASNet search space designs an operation set that contains $15$ operations and five nodes excluding the input and output nodes. The MobileNet search space, as a block-based search space, contains multiple choices for each block. The search space size is about $10^{39}$ with a block size of five. As a result, the results on the existing benchmarking datasets cannot reflect the effect of the EEMs on real scenarios. There exists the need to construct a larger architecture dataset to assist in the validation of the proposed EEMs.
\subsection{Cross-domain Prediction}
The cross-domain prediction mainly refers to predicting the performance of the architectures in different search spaces. In existing EEMs, only the downscaled dataset method, the downscaled model method and population memory have the ability of cross-domain prediction because. This is because these methods do not care about the domains of the architectures. Other EEMs are specific to a target search space and lack cross-domain ability. For example, the designed function-preserving operations in network morphism only can be used in a specific structure. If this structure is not included in the architecture, the network morphism method cannot be used. The performance predictors generally design encoding methods and regression models for specific search spaces. As a result, they cannot be applied or have poor results in other search spaces. The one-shot method can only predict the performance of the subnets contained in the supernet. The cross-domain ability of zero-shot methods also cannot be guaranteed because the measurement of the properties for different types of architectures is different. This means that we have to rebuild an EEM once the search space changes. This is labor-intensive and computationally expensive. There are works that explore the methods of cross-domain prediction. For example, Han \textit{et al.}~\cite{han2021transferable} represented the candidate CNNs as a general computation graph that consists of only primitive operators. Then, they proposed a semi-supervised graph representation learning procedure to predict the architectures from multiple families. However, the method is limited to the cell-based search space. As a result, how to develop a cross-domain EEM to predict the architectures in different domains is also a challenging issue.
\subsection{Multi-task Prediction}
Multi-task prediction refers that EEMs require predicting the multiple performance values in multiple tasks of the same architecture. Moreover, there is an inner correlation between the multiple labels of the same architecture. The multi-task situation often appears in multi-task learning.
Specifically, multi-task learning leverages the useful information between different but related tasks to improve the generalizability of networks~\cite{zhang2021survey}. Multi-task learning has become a hot topic because it can save computational overhead by applying one network to multiple tasks. Furthermore, training one network on multiple tasks can also improve generalization. In multi-task learning, we need to estimate the performance of the architectures on multiple tasks.
Some researchers have observed this problem and work on it. For example, Huang \textit{et al.}~\cite{huang2022arch} embedded the tasks as a part of the input to achieve the prediction of multiple labels. To promote the development of multi-task prediction, Duan \textit{et al.}~\cite{duan2021transnas} proposed a benchmarking dataset (called TransNAS-Bench-101). Specifically, TransNAS-Bench-101 involves the performance of $51,464$ architectures across seven tasks such as classification, regression, pixel-level prediction, and self-supervised tasks. However, the works on this are fairly limited. With the success of multi-task learning in real-world deployment scenarios, we believe that the EEMs for multi-task are a promising future research direction.
\subsection{Uniform Representation}
The performance of the architecture largely depends on the design of the architecture. As a result, how to extract meaningful information from the architectures is critical for the prediction of NAS. To mine the architectures, we first need to provide the representation method to describe the architectures. Although many powerful representation methods have been proposed, they can only represent the architecture in a specific type of search space. For example, the commonly-used adjacency matrix encodings~\cite{white2020study} can only be used to represent the cell-based architectures. This prevents the researchers from learning knowledge of various architectures at the same time.
Furthermore, this is not conducive to improving the transferability of the EEM. To overcome the problem, Sun \textit{et al.}~\cite{sun2021arctext} designed a unified text method to describe the CNN. Concretely, it designed four units to describe the detailed information of each layer in CNNs. Futhermore, it provided a unique order of layers to make the topology information constant. However, this method cannot describe other types of architectures in addition to CNN. The research on the uniform expression for various architectures is still in the early stage for the field of NAS.
\section{Conclusion}\label{conclusion}
This paper gives a comprehensive survey of EEMs of NAS. Specifically, based on the number of architectures trained, we categorize the EEMs into the $N$-shot methods, few-shot methods, one-shot methods, and zero-shot methods. The $N$-shot methods require training every searched architecture and mainly consist of the downscaled dataset methods, downscaled model methods, learning curve extrapolation methods, and network morphisms. The few-shot methods only need to train a smaller number of architectures than TEM. It mainly includes the performance predictor and the population memory. The one-shot methods merely require training one architecture and are generally equal to the weight-sharing methods in practice. Based on whether decouple the search process and the estimation process, they can be divided into the path-based methods and the gradient methods. The zero-shot methods involve no training, thus further reducing the time compared with other methods. They can be divided into the parameter-level method and architecture-level method.
Furthermore, we review the evaluation metrics and benchmark datasets for EEMs. Furthermore, we also report the results on these benchmark datasets to intuitively show the performance of various EEMs. Then, we summarize the challenges and future research directions of the existing EEMs. First, there is a lack of a larger benchmark dataset in the actual scenario to verify the effectiveness of existing EEMs. Second, researchers should pay close attention to cross-domain prediction because rebuilding an estimator for a new search space is expensive. Third, attention should be paid to multi-task prediction as the development of multi-tasks learning in practical application scenarios. Forth, the uniform expression of various architectures is considered a critical issue for the field of NAS.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Neural networks are increasingly used in safety-critical applications \cite{selfdrivingcars,acasxu}.
However, it has become apparent that they are highly susceptible to adversarial examples \cite{intriguingproperties}, i.e., minor and possibly imperceptible input perturbations can cause the output to change significantly.
As such perturbations can occur in the real world either at random or due to malicious actors, it is of utmost importance to analyze the robustness of deep learning based systems in a mathematically rigorous manner before applying them in safety-critical domains.
To this end, a wide range of methods and corresponding software tools have been developed \cite{ehlers,GehrMDTCV18,marta,reluplex}. However, with tools becoming ever more numerous and specialized, it became increasingly difficult for practitioners to decide which tool to use.
In 2020, the inaugural VNN-COMP was organized to tackle this problem and allow researchers to compare their neural network verifiers on a wide set of benchmarks.
Initially conceived as a friendly competition with little standardization, it was increasingly standardized and automated to ensure a fair comparison on cost-equivalent hardware using standardized formats for both properties and networks.
In this work, we outline this development, summarize key rules and results, describe the high-level trends observed over the last three years, and provide an outlook on possible future developments.
\section{Neural Network Verification}
We consider the neural network verification problem defined as follows: Given an input specification $\phi \subseteq \mathds{R}^{d_\text{in}}$, also called pre-condition, an output specification $\psi \subseteq \mathds{R}^{d_\text{out}}$, also called post-condition, and a neural network $N: \mathds{R}^{d_\text{in}} \mapsto \mathds{R}^{d_\text{out}}$, we aim to prove that the pre-condition implies the post-condition, i.e.,
\begin{equation}
\forall x: x \vDash \phi \Rightarrow N(x) \vDash \psi,
\label{eq:verification}
\end{equation}
or provide a counterexample.
Inspired by the notation common in the SAT-solver community, we encode this problem by specifying a constraint set describing an adversarial example, i.e.,
\begin{equation}
\exists x: x \vDash \phi \wedge N(x) \vDash \lnot \psi.
\label{eq:negated}
\end{equation}
Therefore, we call instances where \cref{eq:negated} is satisfiable and thus the property encoded by \cref{eq:verification} does \emph{not} hold \texttt{SAT}\xspace and instances where \cref{eq:negated} is unsatisfiable and the property encoded by \cref{eq:verification} has been shown to hold \texttt{UNSAT}\xspace.
Note that while it is possible to show \texttt{SAT}\xspace by directly searching for counter-examples using adversarial attacks \cite{fgsm,MadryMSTV18}, those approaches are not complete, i.e., if they are not successful in finding a counter-example this does \emph{not} imply that a property holds.
\paragraph{Example Problems}
One particularly popular property is the robustness to adversarial $\ell_\infty$-norm bounded perturbations in image classification.
There, the network $N$ computes a numerical score $y \in \mathds{R}^{d_\text{out}}$ corresponding to its confidence that the input belongs to each of the $d_\text{out}$ classes for each input $x \in \mathds{R}^{d_\text{in}}$.
The final classification $c$ is then computed as $c = \argmax_i N(x)_i$.
In this setting, an adversary may want to perturb the input such that the classification changes.
Therefore, the verification intends to prove that
\begin{align*}
\argmax_i \, & N(x')_i = t, \\
& \forall x' \in \{x' \in \mathds{R}^{d_\text{in}} \mid \|x-x'\|_\infty \leq \epsilon\},
\end{align*}
where $t$ is the target class, $x$ is the original image and $\epsilon$ is the maximal permissible perturbation magnitude.
There, the pre-condition $\phi$ describes the inputs an attacker can choose from ($\phi = \{x' \in \mathds{R}^{d_\text{in}} \mid \|x-x'\|_\infty \leq \epsilon\}$), i.e., an $\ell_\infty$-ball of radius $\epsilon$, and the post-condition $\psi$ describes the output space corresponding to a classification to the target class $t$ ($\psi = \{y \in \mathds{R}^{d_\text{out}} \mid y_t > y_i, \forall i \neq t\}$).
When neural networks are used as controllers, more complex properties can be relevant. For example, in the ACAS Xu setting \cite{acasxu} a neural controller gives action recommendations based on the relative position and heading of the controlled and intruder aircraft. There, we want to, e.g., ensure that for inputs $\mathcal{D}$ corresponding to the intruder aircraft being straight ahead and heading our way, neither of the evasive maneuvers "strong left" (SL) or "strong right" (SR) is considered the worst option. More formally, we want to verify that
\begin{align*}
\argmin_i & N(x')_i \notin \{\text{SL}, \text{SR}\}, \; \forall x' \in \mathcal{D}.
\end{align*}
Here, we obtain a more complex, non-convex post-condition
\begin{align*}
\psi = &\mathds{R}^{d_\text{out}} \setminus \\
&\big(\{y \in \mathds{R}^{d_\text{out}} \mid y_\text{SL} < y_i, \; \forall i \notin \{\text{SL},\text{SR}\}\} \\
&\cup \{y \in \mathds{R}^{d_\text{out}} \mid y_\text{SR} < y_i, \; \forall i \notin \{\text{SL},\text{SR}\}\} \big).
\end{align*}
\section{Competition Goals} \label{sec:goals}
VNN-COMP is organized to further the following goals.
\paragraph{Define Standards}
To enable practitioners to easily use and evaluate a range of different verification approaches and tools without substantial overhead, it is essential that all tools can process both networks and specifications in a standardized file format.
To this end, the second iteration of the VNN-COMP established such a standard. Problem specifications (pre- and post-condition) are defined using the \texttt{VNN-LIB}\xspace \cite{vnnlib} format and neural networks are defined using the \texttt{ONNX}\xspace \cite{onnx} standard.
In 2022, additionally, a standardized format for counterexamples was introduced.
\paragraph{Facilitate Verification Tool Comparison}
Every year, dozens of papers are published on neural network verification, many proposing not only new methods but also new benchmarks.
With authors potentially investing more time into tuning their method to the chosen benchmarks, a fair comparison between all these methods is difficult.
VNN-COMP facilitates such a comparison between a large number of tools on a diverse set of benchmarks, using cost-equivalent hardware, and test instances not available to participants.
Letting participants and industry practitioners propose a wide range of interesting benchmarks, yields not only a ranking on the problems typically used in the field but also highlights which tools are particularly suitable for more specialized problems.
Further, by ensuring a standardized installation and evaluation process is in place, the comparison to a large number of state-of-the-art tools for any publication is enabled.
\paragraph{Shape Future Work Directions}
The visibility VNN-COMP lends to the problems underlying the considered benchmarks has the potential to raise their profile in the community.
As benchmarks are developed jointly by industry and academia, this constitutes a great opportunity to shape future research to be as impactful as possible.
Over the last years, benchmarks have featured ever-increasing network sizes (see \cref{table:comparison}), promoting scalability, more complex networks (including, e.g., residual \cite{resnet} and max-pooling layers \cite{maxpooling}), promoting generalizability, and more complex specifications, enabling more interesting properties to be analyzed.
\paragraph{Bring Researchers Together}
Both the rule and benchmark discussion phase during the lead-up to the competition, as well as the in-person presentation of results at the Workshop on Formal Methods for ML-Enabled Autonomous Systems (FoMLAS)\footnote{\url{https://fomlas2022.wixsite.com/fomlas2022}} provide participants with a great opportunity to meet fellow researches and discuss the future of the field.
Further, the tool and benchmark descriptions participants provide for the yearly report \cite{vnncomp20,vnncomp21,vnncomp22} serve as an excellent summary of state-of-the-art methods, allowing people entering the field to get a quick overview.
\section{Overview of Three Years of VNN-COMP}\label{sec:history}
In this section, we provide a high-level description of how the VNN-COMP evolved from 2020 to 2022, listing all participants and the final rankings in \cref{table:tools}.
Generally, performance is measured on a set of equally weighted \emph{benchmarks}, each consisting of a set of related \emph{instances}. Each instance consists of a trained neural network, a timeout, and input and output constraints.
Below, we group benchmarks into \emph{categories} to enable a quicker comparison between years.
\subsection{VNN-COMP 2020}\label{sec:history20}
The inaugural VNN-COMP\footnote{\url{https://sites.google.com/view/vnn20/vnncomp}} \cite{vnncomp20} was held in 2020 as a \enquote{friendly competition} with no winner. Its main goal was to provide a stepping stone for future iterations by starting the process of defining common problem settings and identifying possible avenues for standardization.
\subsubsection{Benchmarks}
Three benchmark categories were considered with only one of the eight teams participating in all of them:
\begin{itemize}
\item Fully connected networks with ReLU activations -- two benchmarks, based on ACAS Xu and MNIST.
\item Fully connected networks with sigmoid and tanh activation functions -- one benchmark, based on MNIST.
\item Convolutional networks -- two benchmarks, based on MNIST and CIFAR10.
\end{itemize}
\subsubsection{Evaluation}
Teams evaluated their tools using their own hardware.
While this simplified the evaluation process, it made the reported results incomparable, due to the significant hardware differences.
The teams reported that they used between 4 and 40 CPUs and between 16 and 756 GB of RAM.
\subsection{VNN-COMP 2021}\label{sec:history21}
Based upon the insights gained in 2020, the second iteration of VNN-COMP\footnote{\url{https://sites.google.com/view/vnn2021}} was organized with a stronger focus on comparability between the participating tools \cite{vnncomp21}.
\begin{table*}[h]
\centering
\caption{Available AWS instances.}\label{table:awsinstances}
\renewcommand{\arraystretch}{1.1}
\scalebox{0.98}{
\begin{tabular}{lccccc} \toprule
& 2021 & 2022 & vCPUs & RAM [GB] & GPU \\
\midrule
r5.12xlarge & \cmark & \ding{55} & 48 & 384 & \ding{55} \\
p3.2xlarge & \cmark & \cmark & 8 & 61 & V100 GPU with 16 GB memory \\
m5.16xlarge & \ding{55} & \cmark & 64 & 256 & \ding{55} \\
g5.8xlarge & \ding{55} & \cmark & 32 & 128 & A10G GPU with 24 GB memory \\
t2.large & \ding{55} & \cmark & 2 & 8 & \ding{55} \\
\bottomrule
\end{tabular}
}
\end{table*}
\subsubsection{Benchmarks}
Teams were permitted to propose one benchmark with a total timeout of at most six hours split over its constituting instances.
Networks were defined in the \texttt{ONNX}\xspace format \cite{onnx} and problem specifications were given in the \texttt{VNN-LIB}\xspace format \cite{vnnlib}.
To prevent excessive tuning to specific benchmark instances, benchmark proposers were encouraged to provide a script enabling the generation of new random instances for the final tool evaluation.
However, teams were allowed to tune their tools for each benchmark, using the initial set of benchmark instances.
In 2021, the benchmarks could be split into the following categories, with multiple teams participating in all of them:
\begin{itemize}
\item Fully connected networks with ReLU activations -- two benchmarks, based on ACAS Xu and MNIST.
\item Fully connected networks with sigmoid activations -- one benchmark, based on MNIST.
\item Convolutional networks -- three benchmarks, based on CIFAR10.
\item Networks with max-pooling layers -- one benchmark, based on MNIST.
\item Residual networks -- one benchmark, based on CIFAR10.
\item Large networks with sparse matrices -- one benchmark, based on database indexing.
\end{itemize}
\subsubsection{Evaluation}
To allow for comparability of results, all tools were evaluated on equal-cost hardware using Amazon Web Services (AWS).
Each team could decide whether they wanted their tool to be evaluated on a CPU-focused r5.12xlarge or a GPU-focused p3.2xlarge instance (see \cref{table:awsinstances} for more details).
Further, instead of providing results and runtimes themselves, teams had to prepare scripts automating the installation and execution of their tools.
After the submission deadline, the organizers installed and evaluated each tool using the provided scripts.
In many cases, this process required some debugging in a back and forth between the organizers and teams.
\paragraph{Scoring}
For every benchmark, 10 points were awarded for correctly showing the instance to be \texttt{SAT}\xspace/\texttt{UNSAT}\xspace, with a 100 point penalty for incorrect results (see \cref{table:points21}).
A simple adversarial attack was used to identify \enquote{easy} \texttt{SAT}\xspace instances, on which the available points were reduced from 10 to 1.
If tools reported contradicting results on an instance, the ground truth was decided by a majority vote.
Bonus points were awarded to the fastest two tools on every instance (two points for the fastest and one point for the second fastest).
Runtimes differing by less than 0.2 seconds or below one second were considered equal, so multiple teams could receive the two point bonus.
To correct the notable differences in startup overhead, e.g., due to the need to acquire a GPU, it was measured as the runtime on a trivial instance and subtracted from every runtime.
\begin{table}[h]
\begin{center}
\begin{minipage}{174pt}
\caption{Points per instance in 2021. \texttt{SAT} instances were split into simple and complex based on whether a simple adversarial attack was successful.}\label{table:points21}%
\begin{tabular}{lccc} \toprule
& \multicolumn{3}{c}{Returned Result } \\
\cmidrule(lr){2-4}
Ground Truth $\;$ & \texttt{SAT} & \texttt{UNSAT} & Other \\
\midrule
\texttt{SAT}, simple & $+1$ & $-100$ & 0 \\
\texttt{SAT}, complex & $+10$ & $-100$ & 0 \\
\texttt{UNSAT} & $-100$ & $+10$ & 0\\
\bottomrule
\end{tabular}
\end{minipage}
\end{center}
\end{table}
The benchmark score was computed from the points obtained as discussed above by normalizing with the maximum number of obtained points. Consequently, the tool with the most points was assigned a score of $100\%$.
The total competition score was simply the sum of the per benchmark scores, corresponding to equal weighting.
\paragraph{Results}
In 2021, 12 teams participated in the competition.
$\alpha$-$\beta$-CROWN won first place, followed by VeriNet in second, and ERAN/OVAL in third, depending on the overhead measurement and voting scheme used to determine result-correctness.
Except for VeriNet, they all used the GPU instance.
\subsection{VNN-COMP 2022}\label{sec:history22}
In the most recent iteration of VNN-COMP\footnote{\url{https://sites.google.com/view/vnn2022}} \cite{vnncomp22}, the evaluation was fully automated, allowing the number of benchmarks to be increased.
\subsubsection{Benchmarks}
In 2022, each participating team could submit or endorse up to two benchmarks, allowing industry practitioners to propose benchmarks without entering a tool.
Each benchmark had a total timeout of between three and six hours, with randomization of instances being mandatory this year.
Tool tuning was still permitted on a per benchmark level and in practice also per network using the network's statistics.
The submitted benchmarks can be grouped into the following categories:
\begin{itemize}
\item Fully connected networks with ReLU activations -- three benchmarks, based on reinforcement tasks and MNIST.
\item Fully connected networks in TLL format \cite{tll} -- one benchmark.
\item Large networks with sparse matrices -- one benchmark, based on database indexing and cardinality estimation.
\item Convolutional networks -- three benchmarks, based on CIFAR10.
\item Residual networks -- two benchmarks, based on CIFAR10, CIFAR100, and TinyImageNet.
\item Complex U-Net networks with average-pooling and softmax -- one benchmark based on image segmentation.
\end{itemize}
\subsubsection{Evaluation}
Similar to the previous year, teams could choose between a range of AWS instance types (see \cref{table:awsinstances}) providing a CPU, GPU, or mixed focus.
Except for the much weaker t2.large instance, all instances were priced at around three dollars per hour.
In contrast to 2021 where organizers had to manually execute installation scripts and debug with the participants, an automated submission and testing pipeline was set up.
Teams could submit their benchmarks and tools via a web interface by specifying a git repository, commit hash, and post-installation script (enabling, e.g., the acquisition of licenses).
This triggered a new AWS instance to be spawned where all installation scripts were executed.
If the installation succeeded, the tool was automatically evaluated on a previously selected set of benchmarks before the instance was terminated again.
To enable debugging by the participants, all outputs were logged and made accessible live via the submission website, allowing them to monitor the progress.
This automation allowed each team to perform as many tests as necessary without the need to wait for feedback from the organizers.
Furthermore, teams could test on the same AWS instances used during final evaluation without having to pay for their usage, with the costs kindly covered by the SRI Lab of ETH Zurich.
\paragraph{Scoring}
Unlike during the VNN-COMP 2021, \texttt{SAT}\xspace instances were not divided into simple and complex for scoring purposes, leading to 10 points being awarded for all correct results (see \cref{table:points22}).
Further, instead of relying on a voting scheme to determine the ground truth in the presence of dissent among tools, the burden of proof was placed on the tool reporting \texttt{SAT}\xspace, requiring them to provide a concrete counter-example.
If no valid counter-example was provided, the corresponding tool was judged to be incorrect and awarded the 100 point penalty.
\paragraph{Results}
Out of the eleven participating teams, $\alpha$-$\beta$-CROWN placed first, \textsc{MN-BaB} second, and VeriNet third.
For a comparison of all participating tools across all benchmarks, see Figure~\ref{fig:quantPic}.
\begin{figure*}[h]
\centering
\centerline{\includegraphics[width=\linewidth]{cactus/all.pdf}}
\caption{Cactus plot for all tools in the VNN-COMP 2022 across all benchmarks.}
\label{fig:quantPic}
\end{figure*}
\begin{table}[h]
\centering
\caption{Points per instance in 2022.}\label{table:points22}%
\begin{tabular}{lccc} \toprule
& \multicolumn{3}{c}{Returned Result } \\
\cmidrule(lr){2-4}
Ground Truth $\;$ & \texttt{SAT} & \texttt{UNSAT} & Other \\
\midrule
\texttt{SAT} & $+10$ & $-100$ & 0 \\
\texttt{UNSAT} & $-100$ & $+10$ & 0\\
\bottomrule
\end{tabular}
\end{table}
\section{Comparison Across the Years}
In \cref{table:tools} we list all tools participating in any iteration of the VNN-COMP and refer the interested reader to the corresponding VNN-COMP report for a short description of the tools.
In \cref{table:comparison}, we compare the scope of the competition across the last three years.
As can be seen, the number, variety, complexity, and scale of benchmarks increased with every iteration. Starting with 5 benchmarks covering simple fully connected (FC) and convolutional (Conv) networks in 2020, the 2022 competition saw 12 benchmarks including a range of complex residual and U-Net architectures with up to 140 million parameters.
Further, we believe that the increasing number of registered tools clearly shows that the interest in both the field in general and the competition in particular is growing year by year.
However, the large and increasing discrepancy between registered and submitted tools might indicate that many teams feel like they are not able to invest the significant effort required to support not only the standardized network and specification formats but also the wide variety of different benchmarks.
As tools are ranked by their total score with each benchmark providing a score of up to 100$\%$, the final ranking is biased towards tools that support all benchmarks.
While we believe that this is a valuable incentive for tool developers to develop methods that can be easily applied to new problems, it might be daunting for new teams to implement all necessary features, deterring them from participating at all.
\paragraph{Successful Trends}
While all teams started out using only CPUs in 2020, only one of the top four teams relied solely on CPUs in 2021, and all top three teams chose GPU instances in 2022. This transition enabled both the more efficient evaluation of simple bound propagation methods such as DeepPoly \cite{eran3}, CROWN \cite{ZhangWCHD18}, and IBP \cite{GowalIBP} and approximate solutions of the linear programming (LP) problems arising during verification \cite{FerrariMJV22,abcrown4,abcrown3}.
Similarly, the top two teams in 2021 and all top three teams in 2022 relied on a branch-and-bound (BaB) based approach, recursively breaking down the verification problem into easier subproblems until it becomes solvable, thus effectively enabling the use of GPUs to solve tighter mixed integer linear programming (MILP) encodings of the verification problem \cite{oval2,FerrariMJV22,abcrown4,ZhangGCP}.
Both top two teams in the most recent iteration combined this approach with additional multi-neuron \cite{FerrariMJV22} and solver-generated cutting plane constraints \cite{ZhangGCP}, first introduced by the 3rd place ERAN in 2021 \cite{eran5}.
We thus conclude that successful tools leverage hardware accelerators such as GPUs to efficiently handle tight (MI)LP encodings of the verification problem.
\begin{table*}[h]
\begin{center}
\begin{minipage}{\linewidth}
\centering
\caption{Participating tools. \todo{add urls, probably using package just on the toolname, otherwise probably too ugly, e.g., some don't have refs, but do have github, so can find them: https://github.com/formal-verification-research/VERAPAK}}%
\label{table:tools}%
\begin{tabular}{lp{5.0cm}llllll} \toprule
Tool & Organization & \multicolumn{3}{c}{Participation, Place} & References \\
\cmidrule{3-5}
& & 2020 \cite{vnncomp20} & 2021 \cite{vnncomp21} & 2022 \cite{vnncomp22} & \\
\midrule
$\alpha$-$\beta$-CROWN & Carnegie Mellon, Northeastern, Columbia,
UCLA & \ding{55} & \cmark (1/12) & \cmark (1/11) & \cite{abcrown1,abcrown2,abcrown3} \\
AveriNN & Kansas State University & \ding{55} & \ding{55} & \cmark (11/11) & N/A \\
CGDTest & University of Waterloo & \ding{55} & \ding{55} & \cmark (5/11) & N/A \\
Debona & RWTH Aachen University & \ding{55} & \cmark (6/12) & \cmark (8/11) & \cite{debona} \\
DNNF & University of Virginia & \ding{55} & \cmark (12/12) & \ding{55} & \cite{dnnf} \\
\textsc{ERAN} & ETH Zurich, UIUC & \cmark & \cmark (3/12) & \ding{55} & \cite{eran1,eran2,eran3,eran4,eran5,eran6} \\
FastBatLLNN & University of California & \ding{55} & \ding{55} & \cmark (9/11) & N/A \\
Marabou & Hebrew University of Jerusalem, Stanford University, Amazon Web Services, NRI Secure & \ding{55} & \cmark (5/12) & \cmark (7/11) & \cite{marabou} \\
MIPVerify & Massachusetts Institute of Technology & \cmark & \ding{55} & \ding{55} & \cite{mipverify} \\
\textsc{MN-BaB} & ETH Zurich & \ding{55} & \ding{55} & \cmark (2/11) & \cite{FerrariMJV22} \\
nnenum & Stony Brook University & \cmark & \cmark (8/12) & \cmark (4/11) & \cite{nnenum} \\
NNV & Vanderbilt University & \cmark & \cmark (9/12) & \ding{55} & \cite{nnv1,nnv2,nnv3,nnv4,nnv5} \\
NV.jl & Carnegie Mellon, Northeastern & \ding{55} & \cmark (10/12) & \ding{55} & \cite{nvjl} \\
Oval & University of Oxford & \cmark & \cmark (3/12) & \ding{55} & \cite{oval1,oval2,oval3} \\
PeregriNN & University of California & \cmark & \ding{55} & \cmark (6/11) & \cite{peregrinn} \\
RPM & Stanford & \ding{55} & \cmark (11/12) & \ding{55} & \cite{rpm} \\
Venus & Imperial College London & \cmark & \cmark (7/12) & \ding{55} & \cite{venus1,venus2} \\
VeraPak & Utah State University & \ding{55} & \ding{55} & \cmark (10/11) & N/A \\
VeriNet & Imperial College London & \cmark & \cmark (2/12) & \cmark (3/11) & \cite{verinet1,verinet2} \\
\bottomrule
\end{tabular}
\end{minipage}
\end{center}
\end{table*}
\begin{table*}[h]
\begin{center}
\begin{minipage}{\linewidth}
\centering
\caption{Comparison across years.}\label{table:comparison}%
\begin{tabular}{p{4.0cm}p{3.45cm}p{3.45cm}p{3.45cm}} \toprule
$\;$ & 2020 & 2021 & 2022 \\
\midrule
Tools registered & N/A & 15 & 18 \\
Tools submitted & 8 & 13 & 11 \\
Benchmarks submitted & 5 & 8 (+1 unscored) & 12 (+1 unscored)
\\
Max. network depth & 8 & 18 & 27 \\
Max. network parameters & 855,600 & 42,059,431 (sparse) & 138,356,520 \\
Activation functions & ReLU, tanh, sigmoid & ReLU, sigmoid, MaxPool, AveragePool & ReLU, sigmoid, MaxPool \\
Layer types & Fully Connected, Conv & Fully Connected, Conv, Residual & Fully Connected, Conv, Residual, BatchNorm \\
Applications & Image Recognition, Control & Image Recognition, Control, Database Indexing & Image Recognition, Control, Database Indexing, Cardinality Estimation \\
Mean \#benchmarks/tool & 3.0 (min 2, max 5) & 5.5 (min 1, max 9) & 7.3 (min 1, max 13) \\
\bottomrule
\end{tabular}
\end{minipage}
\end{center}
\end{table*}
\section{Outlook}
Below we discuss considerations that could enable future iterations of the VNN-COMP to serve its goals and the community, discussed in \cref{sec:goals}, even better.
\subsection{Tracking Year-on-Year Progress}
While we believe VNN-COMP already provides reasonable mechanisms for comparing the tools submitted in every iteration, the changing benchmarks and tools make it hard to track the year-on-year progress of the field as a whole.
Because some tools are heavily optimized for the specific benchmarks of that year's competition, simply evaluating them on the benchmarks of previous (or future) years (even if they support them) does not yield a meaningful progress metric.
While one benchmark from the inaugural competition was included as an unscored extra benchmark in the two following iterations (\texttt{cifar2020}), only few unsolved instances remain, making it a very insensitive measure for further improvements.
While including all benchmarks from previous years in the (scored) benchmark selection would place an undue burden on participants, choosing one particularly challenging, representative, and interesting benchmark every year to be included as a (scored) extra benchmark in future iterations might be a good compromise. Additionally, a more restrictive stance on tool tuning could enable a much more representative evaluation of new tools on old benchmarks.
\subsection{Tool Tuning}
Many of the most successful tools do not employ a single verification strategy, but a whole portfolio of different modes, all coming with different hyperparameters.
Depending on their choice, tool performance can vary significantly, making it essential for practitioners to get their choice right when applying these tools to new problems.
However, this can be highly challenging given the large number of parameters and their complex interactions, especially without in-depth knowledge of the tool.
For VNN-COMP, tuning tools was allowed explicitly on a per-benchmark basis and implicitly on a per-network basis, enabling teams to showcase the maximum performance of their tools. However, for future iterations it might be interesting to restrict tuning for some or all benchmarks to encourage authors to develop autotuning strategies, making the adaption of their tools to new problems much easier.
This could, for example, be implemented by not only generating random specifications but also random networks.
\subsection{Batch Processing}
Every VNN-COMP benchmark consists of a set of instances that, while typically related, are evaluated in isolation, with the tool being terminated in between.
Unfortunately, this means that any startup overhead such as acquiring a GPU or preprocessing the considered network is incurred for every instance.
This is in contrast to most practical settings where a large number of input-output specifications are considered for the same network.
This discrepancy is accounted for by measuring and subtracting this overhead from each individual runtime.
However, not only is this overhead measurement process flawed and introduces noise, but it can also dominate the evaluation time for easy instances.
In future iterations, tools could be provided with a whole batch of properties at once to more closely relate to their typical application.
Further, currently, timeouts are defined per instance, making a strategy of always attempting verification until timeout optimal.
However, in a practical setting, recognizing instances where verification is likely to fail and stopping early can significantly increase a method's throughput and thus utility.
Switching to per benchmark timeouts for the VNN-COMP would incentivize the development of effective heuristics towards this goal.
Furthermore, tools could benefit from proof-sharing approaches \cite{FischerSDSV22}, where verified sub-problems from one instance are reused for following instances.
\subsection{Continuous Competition}
In addition to a yearly VNN-COMP, tool submissions for the most recent benchmark set could be accepted on a rolling basis, made possible by the automated submission and evaluation process introduced this year.
This would transform the competition from a yearly snapshot of the current research to a centralized repository of the state-of-the-art, updating as teams submit new methods that they publish.
However, if not implemented with great care, this would enable tools to be tuned on the evaluation instances before submission, leading to a skewed comparison.
Further, the question of funding the required cloud compute remains open.
\subsection{Soundness Evaluation}
An inherent requirement for neural network verifiers is that they are sound, i.e., they never claim a safety property holds, when in fact it does not.
However, assessing soundness is difficult as the ground truth for VNN-COMP problem instances is generally only known if it was shown to be \texttt{SAT}\xspace with a valid counter-example.
This is particularly problematic when no instances in a benchmark are \texttt{SAT}\xspace and thus returning \texttt{UNSAT}\xspace for every instance immediately can not be demonstrated to be unsound. Requiring a certain portion of instances of every benchmark to be \texttt{SAT}\xspace (in expectation), could alleviate this issue.
An interesting alternative avenue to tackle this challenge is proof generation \cite{proofgeneration}. An extra category could be introduced where tools are additionally required or incentivized to provide a verifiable proof if they claim a property is \texttt{UNSAT}\xspace.
While big soundness bugs are rare, few or none of the submitted tools are floating point sound, i.e., even tools that would be sound using exact arithmetic might become unsound due to imprecisions introduced by floating-point arithmetic. This is particularly pronounced if tools choose to use single precision computations for performance reasons.
The sensitivity of different tools to such issues could be evaluated on a benchmark specifically designed to uncover floating point soundness issues.
\subsection{Other Competition Modes}
A dedicated falsifier category could be added to encourage teams to develop and submit stronger attacks, going beyond the standard adversarial attacks.
Further, a meta-solver category could be added to investigate whether approaches that heuristically pick from a range of methods, successful in other domains \cite{metasat}, can significantly outperform individual tools. However, it would need to be ensured that these tools provide sufficient value over individual submissions, which already combine different verification strategies.
\subsection{Promote Common Tool Development}
Parsing large and complex \texttt{VNN-LIB}\xspace files or converting \texttt{ONNX}\xspace files to other common formats can be time-consuming to implement.
While many teams implemented their own tools to this end, available, open-source tools for the parsing of \texttt{VNN-LIB}\xspace files \cite{vnnlibparser} and the optimization of \texttt{ONNX}\xspace files (\texttt{DNNV}\xspace \cite{dnnv}) should be highlighted and their continued development encouraged.
\subsection{Remaining Challenges}
We can broadly identify four groups of challenges in neural network verification:
\begin{itemize}
\item Verifying relatively small but only weakly regularized networks, which requires an extraordinarily precise analysis, can still be intractable with current methods.
\item Scaling precise methods to medium-sized networks (e.g. small ResNets) and datasets (e.g. Cifar100 or TinyImageNet) with a large number of neurons is challenging, as the cost of branch-and-bound based algorithms scales exponentially with the required split depth, making branching decisions both harder and more important.
\item Scaling verification to large networks (e.g. VGG-Net 16) and datasets (e.g. ImageNet) in the presence of dense input specifications requires particularly memory-efficient implementations due to a large number of neurons.
\item Verification outside of the classification setting is underexplored leading to a lack of established approaches for, e.g., image segmentation or object detection.
\end{itemize}
Orthogonally, the training of certifiably robust networks remains an open problem. Despite significant progress over recent years \cite{BalunovicV20,GowalIBP,MirmanGV18,MullerSABR,PalmaIBPR,ShiWZYH21,ZhangCXGSLBH20}, networks trained specifically to exhibit provable robustness guarantees still suffer from severely degraded standard accuracy. Therefore, most benchmarks considered in the VNN-COMP are based on networks trained without consideration for later certification.
More broadly in the community, readers may also be interested in the International Competition on Verifying Continuous and Hybrid Systems (ARCH-COMP)\footnote{\url{https://cps-vo.org/group/ARCH/FriendlyCompetition}} category on Artificial Intelligence and Neural Network Control Systems (AINNCS), which has been held annually since 2019~\cite{lopez2019archcomp_ainncs,lopez2022archcomp_ainncs}, and considers neural network verification in closed-loop systems.
\section{Advice for Participants}
In this section, we provide some guidance for teams that are interested in the VNN-COMP but have not participated yet. Note that these are neither rules nor requirements.
\subsection{For Benchmark Authors}
The VNN-COMP intends to highlight areas where neural network verification can be successfully applied and to showcase interesting differences between the participating tools.
Thus, ideally, tasks are not so hard that none of the instances can be solved by any participant but also not so easy that every tool can solve all of them.
For benchmarks related to real-world applications, we recommend including a detailed description of the background, to highlight the benchmark's relevance and the characteristics of the verification problem, e.g. sparseness of the input or some network layers.
Further questions and requests for modifications should be expected while tool authors work on supporting the proposed benchmark.
\subsection{For Tool Authors}
We recommend teams reference past benchmarks to test their tool before the new benchmarks are submitted.
Given the ever-increasing diversity of submitted benchmarks, it may not be feasible to support all benchmarks from the get-go. If this is the case, we recommend focusing on the fully connected and convolutional ReLU networks, which in the past have covered a wide range of benchmarks, while minimizing implementation effort.
Some operations, e.g., max-pooling can also be simplified to multiple ReLU layers using tools such as \texttt{DNNV}\xspace \cite{dnnv}.
Further, we recommend extensive testing against adversarial attacks to minimize the chance for soundness errors.
For tools that are designed for very specific problems, we also want to encourage authors to submit a relevant benchmark highlighting this specialization.
Finally, we recommend reading publications associated with the well-performing tools (see \cref{table:tools}) to gain a better understanding of the techniques used by successful teams.
\section{Conclusions}
In this report, we summarize the main processes and results of the three VNN-COMP held so far from 2020 to 2022.
We highlight the growing interest in the field, expressed in an increasing number of registered teams and considered benchmarks, including some submitted by industry.
Further, we observe that every year, the size and complexity not only of the considered networks but also specifications grew, driving and exemplifying progress in the field.
Finally, we highlight the increase in accessibility of verification methods resulting from the standardized input and output formats and the automated installation and evaluation process required for participation in VNN-COMP.
\begin{acknowledgements}
This material is based upon work supported by the Air Force Office of Scientific Research and the Office of Naval Research under award numbers FA9550-19-1-0288, FA9550-21-1-0121, FA9550-22-1-0019, FA9550-22-1-0450, and N00014-22-1-2156, as well the Defense Advanced Research Projects Agency (DARPA) Assured Autonomy program through contract number FA8750-18-C-0089, and the National Science Foundation (NSF) under grants 1911017, 2028001, 2220401, and 2220426. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the United States Air Force, the United States Navy, DARPA, nor NSF.
\end{acknowledgements}
\bibliographystyle{spmpsci}
|
\section{Introduction}
\subsection{Uniform Diophantine approximation}
The classical metric Diophantine approximation is concerned with the question of how well a real number can be approximated by rationals.
A qualitative answer is provided by the fact that the set of rationals is dense in the reals.
Dirichlet pioneered the quantitative study by showing that, for any $x\in \mathbb{R}$ and $Q>1,$
there exists $(p,q)\in\mathbb{Z}\times\mathbb{N}$ such that
\begin{equation}\label{D1}
|qx-p|\le \frac{1}{Q}~~\text{and } q<Q.
\end{equation}
The result serves as a start point of the metric theory in Diophantine approximation.
An easy application yields the following corollary$\colon$ for any $x\in \mathbb{R},$
there exists infinitely many $(p,q)\in\mathbb{Z}\times\mathbb{N}$ such that
\begin{equation*}
|qx-p|\le \frac{1}{q}.
\end{equation*}
This corollary claims that $|qx-p|$ is small compared to $q$, while Dirichlet's original theorem (\ref{D1}) provides a uniform estimate of $|qx-p|$ in terms of $Q$. These two kinds of approximations are referred to as uniform approximation and asymptotic respectively. See \cite{W12} for more account on the related subject.
In this article we are interested in the numbers which are approached in uniform or asymptotic way by an orbit (in a dynamical system) with a prescribed speed.
Let $(X, T, \mu)$ be a measure-preserving dynamical system, where
$(X,d)$ is a metric space, $T\colon X\to X$ is a Borel transformation, and $\mu$ is a $T$-invariant Borel probability measure on $X$.
As is well known, Birkhoff's ergodic theorem \cite{PW} implies that, in an ergodic dynamical system, for almost all $y\in X$, the set
$$\left\{x\in X\colon \liminf_{n\to\infty}d(T^{n}(x),y)=0\right\}$$
is of full $\mu$-measure.
The result, which gives a qualitative characterization of the distributions of the $T$-orbits in $X$, can be regarded as a counterpart of the density property of rational numbers in the reals. It leads naturally to the quantitative study of the distributions of the $T$-orbits.
The shrinking target problem in dynamical system $(X,T)$ aims at a quantitative study on the Birkhoff's ergodic theorem, which investigates the set
$$W_y(T,\psi)=\Big\{x\in X\colon d(T^{n}(x),y)<\psi(n) \text{ for infinitely many } n\in \mathbb{N}\Big\},$$
where $\psi\colon \mathbb{N}\to \mathbb{R}$ is a positive function such that $\psi(n)\to 0$ as $n\to\infty,$ and $y\in X.$
Hill $\&$ Velani \cite{HV95} studied the Hausdorff dimension of the set
$$\Big\{x\in X\colon d(T^{n}(x),y)<e^{-\tau n} \text{ for infinitely many } n\in \mathbb{N}\Big\}$$
in the system $(X,T)$ with $T$ an expanding rational map of degree greater than or equal to 2 and $X$ the corresponding Julia set, where $\tau>0.$ See \cite{TW} for more information.
Representations of real numbers are often induced by dynamical systems or algorithms,
and thus the related Diophantine approximation problems are in the nature of dynamical system, fractal geometry and number theory. An active topic of research lies in studying the approximation of real numbers in dynamical systems by the orbits of the points. Recently, many researchers have studied the Hausdorff dimension of the set $W_y(T,\psi)$ in the corresponding dynamical system under different expansions, and obtained many significant results \cite{LWWX,SW,STZ,TZ}.
Marked by the famous mass transfer principle established by Beresnevich $\&$ Velani \cite{BV}, studies on the asymptotic approximation properties of orbits in dynamical systems are relatively mature.
However, there are few results on the uniform approximation properties of orbits.
Let $\big(X,T\big)$ be a exponentially mixing system with respect to the probability measure $\mu$, and let $\psi\colon \mathbb{N}\to \mathbb{R}$ be a positive function satisfying that $\psi(n)\to 0$ as $n\to\infty$.
Kleinbock, Konstantous $\&$ Richter \cite{KKR} studied the Lebesgue measure of the set of real numbers $x\in X$ with the property that,
for every sufficiently large integer $N,$ there is an integer $n$ with $1\le n\le N$ such that the distance between $T^{n}(x)$ and a fixed $y$ is at most $\psi(N),$ i.e.,
$$\mathcal{U}(y,\psi)=\Big\{x\in[0,1)\colon \forall N\gg1, \exists~n\in[1,N], \text{ such that } |T^{n}(x)-y|<\psi(N)\Big\}.$$
They gave the sufficient conditions for $\mathcal{U}(y,\psi)$ to be of zero or full measure.
Although the Khintchine type 0-1 law of the set $\mathcal{U}(y,\psi)$ has not been established,
the work has aroused the interest of researchers (see \cite{GP,KA,KKP} for the related studies).
Bugeaud $\&$ Liao \cite{BL} investigated the size of the set
$$\Big\{x\in[0,1)\colon \forall N\gg1, \exists~ n\in[1,N], \text{ such that } T_\beta^{n}(x)<|I_{N}(0)|^{\hat{\nu}}\Big\}$$
in $\beta$-dynamical systems from the perspective of Hausdorff dimension, where $T_\beta$ is the $\beta$-transformation on $[0,1)$ defined by $T_\beta(x)=\beta x \text{ mod }1,$ $I_n(0)$ denotes the basic interval of order $n$ which contains the point 0, and $\hat{\nu}$ is a nonnegative real number. For more information related to the uniform approximation properties, see \cite{KL,ZW} and the references therein.
In this paper, we shall investigate the uniform approximation properties of the orbits under the Gauss transformation.
The Gauss transformation $T\colon [0,1) \to [0,1)$ is defined as
\begin{equation*}
T(0)=0,~ T(x)= \frac{1}{x}\!\!\!\pmod1~\text{for} ~x\in(0,1).
\end{equation*}
And each irrational number $x\in[0,1)$ can be uniquely expanded into the following form$\colon$
\begin{equation}\label{e1}
x=\frac{1}{a_{1}(x)+\frac{1}{a_{2}(x)+\ddots+\frac{1}{a_{n}+T^{n}(x)}}}
=\frac{1}{a_{1}(x)+\frac{1}{a_{2}(x)+\frac{1}{a_{3}(x)+\ddots}}},
\end{equation}
with $a_{n}(x)=\lfloor\frac{1}{T^{n-1}(x)}\rfloor$, called the $n$-th partial quotient of $x$ (here $\lfloor\cdot\rfloor$ denotes the greatest integer less than or equal to a real number and $T^0$ denotes the identity map).
For simplicity of notation, we write $(\ref{e1})$ as
\begin{equation}\label{e2}
x=[a_{1}(x),a_{2}(x),\ldots,a_{n}(x)+T^{n}(x)]=[a_{1}(x),a_{2}(x),a_{3}(x),\ldots].
\end{equation}
As was shown by Philipp \cite{P67}, the system $([0,1),T)$ is exponentially mixing with respect to the Gauss measure $\mu$ given by $d\mu=dx/(1+x)\log 2.$
Thus the above result of \cite{KKR} applies for the Gauss measure of the set $\mathcal{U}(y,\psi)$ in the system of continued fractions.
In consequence, we shall focus on the size of $\mathcal{U}(y,\psi)$ in dimension.
The dimension of sets $\mathcal{U}(y,\psi)$ depend on the choice of the given point $y$. In this paper, we will consider a class of quadratic irrational numbers $y={(\sqrt{i^{2}+4}-i)}/{2}=[i,i,\ldots]$ with $i\in \mathbb{N}$,
and calculate the Hausdorff dimension of the set
$$\mathcal{U}(y,\hat{\nu})=\Big\{x\in[0,1)\colon \forall N\gg1, \exists~ n\in[1,N], \text{ such that } |T^{n}(x)-y|<|I_{N}(y)|^{\hat{\nu}}\Big\}.$$
For $\beta \in[0,1]$, let $s(\beta,{y})$ denote the solution of
$$P\Big(T,-s\Big(\log |T'|+\frac{\beta}{1-\beta}{g(y)}\Big)\Big)=0,$$
where $P(T,\phi)$ is the pressure function with potential $\phi$ in the continued fraction system $([0,1),T)$, $T'$ is the derivative of $T$, {and $g(y)$ is the limit $\lim_{n}\log q_{n}(y)/n$.}
\begin{thm}\label{TZ1}
Given a nonnegative real number $\hat{\nu},$ we have
\begin{equation*}
\dim_{H}\mathcal{U}(y,\hat{\nu})=\left\{
\begin{array}{ll}
s\Big(\frac{4\hat{\nu}}{(1+\hat{\nu})^{2}},{y}\Big), & \ \ \ \text{if }~ 0\leq \hat{\nu}\leq 1; \\
0, & \ \ \ \text{otherwise.}
\end{array}
\right.
\end{equation*}
Throughout the paper, $\dim_{H}$ denotes the Hausdorff dimension of a set.
\end{thm}
We now turn to the discussion of two approximation exponents which are relevant to asymptotic/uniform Diophantine approximation. For $x\in[0,1),$
we define the asymptotic approximation exponent of $x$ by
$$\nu(x)=\sup\Big\{\nu\ge 0\colon |T^{n}(x)-y|<|I_n(y)|^{\nu}\text{ for infinitely many } n\in \mathbb{N}\Big\}$$
and the uniform approximation exponent by
$$\hat{\nu}(x)=\sup\Big\{\hat{\nu}\ge 0\colon \forall N\gg1, \exists~ n\in[1,N], \text{ such that } |T^{n}(x)-y|<|I_N(y)|^{\hat{\nu}}\Big\}.$$
The exponents $\nu(x)$ and $\hat{\nu}(x)$ are analogous to the exponents introduced in \cite{AB}, see also \cite{BL,BL05}.
By the definitions of $\nu(x)$ and $\hat{\nu}(x)$, it is readily checked that $\hat{\nu}(x)\le \nu(x)$ for all $x\in[0,1).$
Actually, applying Philipp's result \cite{P67}, we deduce that $\nu(x)=0$ for Lebesgue almost all $x\in[0,1)$ (see Lemma \ref{full}). Li, Wang, Wu $\&$ Xu \cite{LWWX} studied the multifractal properties of the asymptotic exponent $\nu(x)$ and showed that for $0\le\nu\le+\infty,$
\begin{equation}\label{Asy}
\dim_{H}\{x\in[0,1)\colon \nu(x)\ge \nu\}=s\Big(\frac{\nu}{1+\nu},{y}\Big).
\end{equation}
We will denote by $E(\hat{\nu})$ the level set of the uniform approximation exponent:
$$E(\hat{\nu})=\{x\in[0,1)\colon \hat{\nu}(x)=\hat{\nu}\}.$$
\begin{thm}\label{TZ3}
Given a nonnegative real number $\hat{\nu},$ we have
$$\dim_{H}E(\hat{\nu})=\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}=\dim_{H}\mathcal{U}(y,\hat{\nu}).$$
\end{thm}
Actually, Theorems \ref{TZ1} and \ref{TZ3} follow from the following more general result which gives the Hausdorff dimension of the set
\begin{equation*}
E(\hat{\nu},\nu)=\{x\in[0,1)\colon \hat{\nu}(x)=\hat{\nu},~\nu(x)=\nu\}.
\end{equation*}
\begin{thm}\label{TZ2}
Given two nonnegative real numbers $\hat{\nu}$ and $\nu$ with $\hat\nu\le\nu$, we have
\begin{equation*}
\dim_{H}E(\hat{\nu},\nu)
=\left\{\begin{array}{ll} 1, & \ \ \ \text{if }~ \nu=0;\\
s\Big(\frac{\nu^{2}}{(1+\nu)(\nu-\hat{\nu})},{y}\Big), & \ \ \ \text{if }~ 0\le\hat{\nu}\le\frac{\nu}{1+\nu}<\nu\leq\infty; \\
0, & \ \ \ \text{otherwise.}
\end{array}
\right.
\end{equation*}
Here, we take $\frac{\nu^{2}}{(1+\nu)(\nu-\hat{\nu})}=1$ when $\nu=\infty$.
\end{thm}
Let us make the following remarks regarding Theorems \ref{TZ1}-\ref{TZ2}:
\begin{itemize}
\item These results remain valid for any quadratic irrational number $y$. Indeed, by Lagrange's theorem, any such $y$ is represented by a periodic continued fraction expansion, i.e., $$y=[a_1,a_2,\ldots,a_{k_0},\overline{a_{k_{0}+1},\ldots,a_{k_{0}+h}}]$$
for some positive integers $k_{0}$ and $h.$
A slight change (replacing the block $[i]$ by the periodic block
$[a_{k_{0}+1},\ldots,a_{k_{0}+h}]$) in the proofs actually shows that Theorems \ref{TZ1}-\ref{TZ2} still hold for every quadratic irrational number $y\in[0,1).$
\medskip
\item The fractal sets $\mathcal{U}(y,\hat{\nu}),$ $E(\hat{\nu})$ and $E(\hat{\nu},\nu)$ are not the so-called limsup sets,
and thus we cannot obtain a natural covering to estimate the upper bound of the Hausdorff dimensions of the sets $\mathcal{U}(y,\hat{\nu})$ and $E(\hat{\nu},\nu)$. To overcome this difficulty, we need a better understanding on the fractal structure of these sets; the previous work of Bugeaud $\&$ Liao \cite{BL} helps.
\end{itemize}
Combining (\ref{Asy}) and Theorem \ref{TZ2}, we obtain the dimension of the level set related to the asymptotic exponent $\nu(x)$.
\begin{cor}\label{Cor}
Given a nonnegative real number $\nu,$ we have
$$\dim_{H}\{x\in[0,1)\colon \nu(x)=\nu\}=s\Big(\frac{\nu}{1+\nu},{y}\Big).$$
\end{cor}
\subsection{Run-length function}
Applying the main ideas of the proofs of Theorems $\ref{TZ1}$ and \ref{TZ2}, we characterize the multifractal properties of run-length function in continued fractions.
The run-length function was initially introduced in a mathematical experiment of cion tossing,
which counts the consecutive occurrences of `heads' in $n$ times trials.
This function has been extensively studied for a long time.
For $x\in[0,1],$ let $r_n(x)$ be the dyadic run-length function of $x,$ namely,
the longest run of 0's in the first $n$ digits of the dyadic expansion of $x.$
Erd\"{o}s $\&$ R\'{e}nyi \cite{E} did a pioneer work on the asymptotic behavior of $r_n(x)\colon$ for Lebesgue almost all $x\in[0,1],$
\begin{equation*}
\lim_{n\to\infty}\frac{r_n(x)}{\log_2 n}=1.
\end{equation*}
Likewise, we define the run-length function in the continued fraction expansion:
for $n\geq 1$, the $n$-th maximal run-length function of $x$ is defined as
$$R_{n}(x)=\max\big\{l\geq1\colon a_{i+1}(x)=\cdots=a_{i+l}(x) \text{ for some } 0 \leq i\leq n-l\big\}.$$
Wang $\&$ Wu \cite{W} considered the metric properties of $R_{n}(x)$ and proved that
$$\lim_{n\to\infty}\frac{R_{n}(x)}{\log_{\frac{\sqrt{5}+1}{2}}n}=\frac{1}{2}$$
for almost all $x\in [0,1).$ They also studied the following exceptional sets
\begin{equation*}
F\big(\{\varphi(n)\}_{n=1}^{\infty}\big)=\Big\{x\in[0,1)\colon \lim_{n\to\infty}\frac{R_{n}(x)}{\varphi(n)}=1\Big\},
\end{equation*}
\begin{equation*}
G\big(\{\varphi(n)\}_{n=1}^{\infty}\big)=\Big\{x\in[0,1)\colon \limsup\limits_{n\to\infty}\frac{R_{n}(x)}{\varphi(n)}=1\Big\},
\end{equation*}
where $\varphi\colon \mathbb{N} \to \mathbb{R}^{+}$ is a non-decreasing function. They showed that$\colon$
{\rm{(1)}} if $\lim_{n\to\infty}\frac{\varphi(n+\varphi(n))}{\varphi(n)}=1,$
then $\dim_{H}F\big(\{\varphi(n)\}_{n=1}^{\infty}\big)=1;$
{\rm{(2)}} if $\liminf_{n\to\infty}\frac{\varphi(n)}{n}=\beta \in[0,1],$
then $\dim_{H}G\big(\{\varphi(n)\}_{n=1}^{\infty}\big)=s\big(\beta,{\frac{\sqrt{5}+1}{2}}\big)$.
\noindent In the study of Case (2), Wang $\&$ Wu studied essentially the Hausdorff dimension of the following set
\begin{equation}\label{equ13}
G(\beta)=\left\{x\in[0,1)\colon \limsup_{n \to\infty}\frac{R_{n}(x)}{n}=\beta\right\}.
\end{equation}
Replacing the limsup of the quantity $R_n(x)/n$ in (\ref{equ13}) with liminf, we study the set
$$F(\alpha)=\left\{x\in[0,1)\colon \liminf_{n\to\infty}\frac{R_{n}(x)}{n}=\alpha\right\},$$
and determine the Hausdorff dimension of the intersections of $F(\alpha)\cap G(\beta).$ As a corollary, we obtain the Hausdorff dimension of $F(\alpha).$
\begin{thm}\label{thm1}
For $\alpha, \beta\in [0,1]$ with $\alpha\leq \beta$, we have
\begin{equation*}
\dim_{H}\Big(F(\alpha)\cap G(\beta)\Big)=\left\{ \begin{array}{ll}
1, & \ \ \ \text{if }~ \beta=0;\\
s\Big(\frac{\beta^{2}(1-\alpha)}{\beta-\alpha},{\frac{\sqrt{5}+1}{2}}\Big), & \ \ \ \text{if }~ 0\leq \alpha\leq \frac{\beta}{1+\beta}<\beta\leq1; \\
0, & \ \ \ \text{otherwise.}\end{array}\right.
\end{equation*}
\end{thm}
\begin{thm}\label{thm2}
For $\alpha\in[0,1],$ we have
\begin{equation*}
\dim_{H}F(\alpha)=\left\{\begin{array}{ll}
s\big(4\alpha(1-\alpha),{\frac{\sqrt{5}+1}{2}}\big), & \ \ \ \text{if }~ 0\leq \alpha\leq \frac{1}{2}; \\
0, & \ \ \ \text{otherwise.} \end{array}\right.
\end{equation*}
\end{thm}
\section{Preliminaries}
\subsection{Properties of continued fractions}
This section is devoted to recalling some elementary properties in continued fractions.
For more information on the continued fraction expansion, the readers are referred to \cite{HW,K,S}.
We also introduce some basic techniques for estimating the Hausdorff dimension of a fractal set (see \cite{F,SF}).
For any irrational number $x\in[0,1)$ with continued fraction expansion (\ref{e2}), we write
$\frac{p_{n}(x)}{q_{n}(x)}=[a_{1}(x),\ldots,a_{n}(x)]$
and call it the $n$-th convergent of $x.$ With the conventions
$p_{-1}(x)=1,$ $q_{-1}(x)=0,$ $p_{0}(x)=0$ and $q_{0}(x)=1,$
we know that $p_{n}(x)$ and $q_{n}(x)$ satisfy the recursive relations \cite{K}$\colon$
\begin{equation}\label{equ21}
p_{n+1}(x)=a_{n+1}(x)p_{n}(x)+p_{n-1}(x),~~q_{n+1}(x)=a_{n+1}(x)q_{n}(x)+q_{n-1}(x), ~~n\geq 0.
\end{equation}
Clearly, $q_{n}(x)$ is determined by $a_{1}(x),\ldots,a_{n}(x),$ so we also write
$q_{n}(a_{1}(x),\ldots,a_{n}(x))$
instead of $q_n(x)$. We write $a_{n}$ and $q_n$ in place of $a_{n}(x)$ and $q_n(x)$ for simplicity when no confusion can arise.
\begin{lem} [\cite{K}]\label{lem21}
For $n\geq 1$ and $(a_{1},\ldots,a_{n})\in \mathbb{N}^{n}$, we have$\colon$
{\rm{(1)}} $q_{n}\geq 2^{\frac{n-1}{2}},$ and $\prod\limits_{k=1}^{n}a_{k}\leq q_{n}\leq \prod\limits_{k=1}^{n}(a_{k}+1).$
{\rm{(2)}} For any $k\ge 1,$
\begin{equation*}
1\leq \frac{q_{n+k}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+k})}
{q_{n}(a_{1},\ldots,a_{n})q_{k}(a_{n+1},\ldots,a_{n+k})}\leq 2.
\end{equation*}
{\rm{(3)}}
If $a_1=a_2=\cdots=a_{n}=i$, then
$$\frac{\big(\tau(i)\big)^{n}}{2} \leq q_{n}(i,\ldots,i)
=\frac{\big(\tau(i)\big)^{n+1}-\big(\zeta(i)\big)^{n+1}}{\tau(i)-\zeta(i)}\leq 2\big(\tau(i)\big)^{n},$$
where $\tau(i)=\frac{i+\sqrt{i^{2}+4}}{2}$ and $\zeta(i)=\frac{i-\sqrt{i^{2}+4}}{2}.$
\end{lem}
\begin{proof}
For the convenience of readers, we give the proof.
{\rm{(1)}}
By the recursive relations (\ref{equ21}), we readily check that
$$\prod\limits_{k=1}^{n}a_{k}\leq q_{n}\leq \prod\limits_{k=1}^{n}(a_{k}+1).$$
Since $a_n\ge 1$ for $n\ge 1,$ we have
$$1=q_0\le q_1<q_2<\cdots q_{n-1}<q_n.$$
By induction $q_n\ge2^{\frac{n-1}{2}}$ for all $n\ge1$; similarly $p_n\ge2^{\frac{n-1}{2}}$.
\smallskip
{\rm{(2)}}
Induction on $k$:
assuming that
\begin{equation*}
1\leq
\frac{q_{n+k}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+k})}{q_{n}(a_{1},\ldots,a_{n})q_{k}(a_{n+1},\ldots,a_{n+k})}
\leq 2
\end{equation*}
holds for all $k\in\{1,\ldots,m\}$, we prove that the above inequality holds for $k=m+1.$
Indeed, this is the case because
\begin{align*}
&\ \ \ \ q_{n+m+1}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+m+1})\\
&=a_{n+m+1}q_{n+m}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+m})+q_{n+m-1}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+m-1})\\
&\ge a_{n+m+1}q_{n}(a_{1},\ldots,a_{n})q_{m}(a_{n+1},\ldots,a_{n+m})+q_{n}(a_{1},\ldots,a_{n})q_{m-1}(a_{n+1},\ldots,a_{n+m-1})\\
&=q_{n}(a_{1},\ldots,a_{n})q_{m+1}(a_{n+1},\ldots,a_{n+m+1}),
\end{align*}
and
\begin{align*}
&\ \ \ \ q_{n+m+1}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+m+1})\\
&=a_{n+m+1}q_{n+m}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+m})+q_{n+m-1}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{n+m-1})\\
&\le 2a_{n+m+1}q_{n}(a_{1},\ldots,a_{n})q_{m}(a_{n+1},\ldots,a_{n+m})+2q_{n}(a_{1},\ldots,a_{n})q_{m-1}(a_{n+1},\ldots,a_{n+m-1})\\
&=2q_{n}(a_{1},\ldots,a_{n})q_{m+1}(a_{n+1},\ldots,a_{n+m+1}).
\end{align*}
{\rm{(3)}}
By the recursive relations (\ref{equ21}), we deduce that
\begin{align*}
\left(
\begin{array}{cc}
p_{n+1} & p_{n} \\
q_{n+1} & q_{n} \\
\end{array}
\right)&=\left(
\begin{array}{cc}
p_{n} & p_{n-1} \\
q_{n} & q_{n-1} \\
\end{array}
\right)\left(
\begin{array}{cc}
a_{n+1} & 1 \\
1 & 0 \\
\end{array}
\right)\\ &
=\left(
\begin{array}{cc}
p_0 & p_{-1} \\
q_0 & q_{-1} \\
\end{array}
\right)\left(
\begin{array}{cc}
a_{1} & 1 \\
1 & 0 \\
\end{array}
\right)\cdots\left(
\begin{array}{cc}
a_{n+1} & 1 \\
1 & 0 \\
\end{array}
\right)\\&=
\left(
\begin{array}{cc}
0 & 1 \\
1 & 0 \\
\end{array}
\right)\left(
\begin{array}{cc}
a_{1} & 1 \\
1 & 0 \\
\end{array}
\right)\cdots\left(
\begin{array}{cc}
a_{n+1} & 1 \\
1 & 0 \\
\end{array}
\right).
\end{align*}
Taking $a_1=\cdots=a_n=a_{n+1}=i$ yields that
$$\left(
\begin{array}{cc}
p_{n+1} & p_{n} \\
q_{n+1} & q_{n} \\
\end{array}
\right)=\left(
\begin{array}{cc}
0 & 1 \\
1 & 0 \\
\end{array}
\right)\left(
\begin{array}{cc}
i & 1 \\
1 & 0 \\
\end{array}
\right)\cdots\left(
\begin{array}{cc}
i & 1 \\
1 & 0 \\
\end{array}
\right).$$
The symmetric matrix $A=\left(
\begin{array}{cc}
i & 1 \\
1 & 0 \\
\end{array}
\right)$ is diagonalizable: $$P^{-1}AP=\left(
\begin{array}{cc}
\tau(i) & 0 \\
0 & \zeta(i) \\
\end{array}
\right)$$ with $P=\left(
\begin{array}{cc}
\tau(i) & \zeta(i) \\
1 & 1 \\
\end{array}
\right)$.
A direct calculation yields that
$$q_n(i,\ldots,i)
=\frac{\big(\tau(i)\big)^{n+1}-\big(\zeta(i)\big)^{n+1}}{\tau(i)-\zeta(i)}.$$
Also
$$\frac{\big(\tau(i)\big)^{n+1}-\big(\zeta(i)\big)^{n+1}}{\tau(i)-\zeta(i)}\le\frac{2\big(\tau(i)\big)^{n+1}}{\tau(i)}=2\big(\tau(i)\big)^{n},$$
and, if $n$ is even,
$$\frac{\big(\tau(i)\big)^{n+1}-\big(\zeta(i)\big)^{n+1}}{\tau(i)-\zeta(i)}\ge\frac{\big(\tau(i)\big)^{n+1}}{2\tau(i)}=\frac{\big(\tau(i)\big)^{n}}{2};$$
if $n$ is odd, (since $\zeta(i)\cdot\tau(i)=-1$)
\begin{align*}
\frac{\big(\tau(i)\big)^{n+1}-\big(\zeta(i)\big)^{n+1}}{\tau(i)-\zeta(i)}
=\frac{\big(\tau(i)\big)^{2(n+1)}-1}{\big(\tau(i)\big)^{n+2}+\big(\tau(i)\big)^{n}}
\ge\frac{\big(\tau(i)\big)^{n}}{2}.
\end{align*}
This completes the proof.
\end{proof}
For $n\geq 1$ and $(a_{1},\ldots,a_{n})\in \mathbb{N}^{n}$, we write
$$I_{n}(a_{1},\ldots,a_{n})=\{x\in[0,1)\colon a_{k}(x)=a_{k}, 1\leq k\leq n\},$$
and call it a basic interval of order $n$.
The basic interval of order $n$ which contains $x$ will be denoted by $I_{n}(x)$, i.e.,
$I_{n}(x)=I_{n}(a_{1}(x),\ldots,a_{n}(x))$.
\begin{lem} [\cite{K}]\label{lem22}
For $n\geq 1$ and $(a_{1},\ldots,a_{n})\in \mathbb{N}^{n},$ we have
\begin{equation}\label{e22}
\frac{1}{2q_{n}^{2}}\leq|I_{n}(a_{1},\ldots,a_{n})|=\frac{1}{q_{n}(q_{n}+q_{n+1})}\leq\frac{1}{q_{n}^{2}}.
\end{equation}
Here and hereafter $|\cdot|$ denotes the length of an interval.
\end{lem}
The next lemma describes the distribution of basic intervals $I_{n+1}$ of order $n+1$ inside an $n$-th basic interval $I_{n}.$
\begin{lem}[\cite{K}]\label{lem23}
Let $I_{n}(a_{1},\ldots,a_{n})$ be a basic interval of order $n,$ which is partitioned into sub-intervals $I_{n+1}(a_{1},\ldots,a_{n},a_{n+1})$ with $a_{n+1}\in \mathbb{N}.$
When $n$ is odd, these sub-intervals are positioned from left to right, as $a_{n+1}$ increases;
when $n$ is even, they are positioned from right to left.
\end{lem}
The following lemma displays the relationship between the ball $B(x,|I_{n}(x)|)$ and the basic interval $I_{n}(x)$.
\begin{lem} [\cite{BW}]\label{lem24}
Let $x=[a_{1},a_{2},\ldots].$ We have$\colon$
{\rm{(1)}} if $a_{n}\neq 1,$ then $B(x,|I_{n}(x)|)\subset\bigcup\limits_{j=-1}^{3}I_{n}(a_{1},\ldots,a_{n}+j);$
{\rm{(2)}} if $a_{n}=1$ and $a_{n-1}\neq 1,$ then $B(x,|I_{n}(x)|)\subset\bigcup\limits_{j=-1}^{3}I_{n-1}(a_{1},\ldots,a_{n-1}+j);$
{\rm{(3)}} if $a_{n}=1$ and $a_{n-1}=1,$ then $B(x,|I_{n}(x)|)\subset I_{n-2}(a_{1},\ldots,a_{n-2}).$
\end{lem}
The following two properties, namely, H\"{o}lder property and the mass distribution principle, are often used to estimate the Hausdorff dimension of a fractal set.
\begin{lem} [\cite{F}] \label{lem212}
If $f\colon X \to Y$ is an $\alpha$-H\"{o}lder mapping between metric spaces,
that is, there exists $c>0$ such that for all $x_{1},x_{2}\in X$,
$$d(f(x_{1}),f(x_{2}))\leq cd(x_{1},x_{2})^{\alpha}.$$
Then $\dim_{H}f(X)\leq \frac{1}{\alpha}\dim_{H}X.$
\end{lem}
\begin{lem} [\cite{F}] \label{lem213}
Let $E\subseteq [0,1]$ be a Borel set and $\mu$
be a measure with $\mu(E)> 0.$ If for every $x \in E$,
$$\liminf_{r \to 0}\frac{\log \mu(B(x,r))}{\log r} \geq s,$$
then $ \dim_{H}E\geq s.$
\end{lem}
We conclude this subsection by quoting a dimensional result related to continued fractions, which will be used in the proof of Theorem \ref{TZ2}.
Let $\mathbf{K}=\{k_{n}\}_{n= 1}^{\infty}$ be a subsequence of $\mathbb{N}$ which is not cofinite.
Let $x=[a_{1},a_{2},\ldots]$ be an irrational number in $[0,1)$.
Eliminating all the terms $a_{k_n}$ from the sequence $a_1,a_2,\ldots$,
we obtain an infinite subsequence $c_1,c_2,\ldots$, and put $\phi_{\mathbf{K}}(x)=y$ with $y=[c_1,c_2,\ldots]$.
In this way, we define a mapping $\phi_{\mathbf{K}}\colon [0,1)\cap \mathbb{Q}^{c}\to [0,1)\cap \mathbb{Q}^{c}$.
Let $\{M_{n}\}_{n\geq 1}$ be a sequence with $M_{n}\in\mathbb{N}$, $n\geq 1$. Set
$$S(\{M_{n}\})=\big\{x\in [0,1)\cap \mathbb{Q}^{c}\colon 1\leq a_{n}(x)\leq M_{n}~\text{for all}~n\geq 1\big\}.$$
\begin{lem}[\cite{CC}]\label{lem214}
Suppose that $\{M_{n}\}_{n=1}^{\infty}$ is a bounded sequence.
If the sequence $\mathbf{K}=\{k_{n}\}_{n=1}^{\infty}$ is of density zero in $\mathbb{N}$, then
$$\dim_{H}S(\{M_{n}\})=\dim_{H}\phi_{\mathbf{K}}S(\{M_{n}\}).$$
\end{lem}
\subsection{Pressure function and pre-dimensional number}
We now introduce the notions of the pressure function and pre-dimensional number in the continued fraction dynamical system.
For more details, we refer the reader to \cite{HMU}.
For $\mathcal{A}$ a finite or infinite subset of $\mathbb{N},$ we set
$$X_{\mathcal{A}}=\big\{x\in[0,1)\colon a_{n}(x)\in \mathcal{A} \text{ for all } n\geq 1\big\}.$$
The pressure function restricted to the subsystem $(X_{\mathcal{A}},T)$ with potential $\phi\colon [0,1)\to \mathbb{R} $ is defined as
\begin{equation}\label{e23}
P_{\mathcal{A}}(T,\phi)=\lim_{n\to\infty}\frac{\log\sum\limits_{(a_{1},\ldots,a_{n})\in \mathcal{A}^{n}}\sup\limits_{x\in X_{\mathcal{A}}}\exp{S_n \phi([a_{1},\ldots,a_{n}+x])}}{n},
\end{equation}
where $S_{n}\phi(x)=\phi(x)+\cdots+\phi(T^{n-1}(x))$ denotes the ergodic sum of $\phi$.
When $\mathcal{A}=\mathbb{N}$, we write $P(T,\phi)$ for $P_{\mathbb{N}}(T,\phi)$.
The $n$-th variation $\textrm{Var}_{n}(\phi)$ of $\phi$ is defined as
$$\textrm{Var}_{n}(\phi)=\sup\big\{|\phi(x)-\phi(y)|\colon I_{n}(x)=I_{n}(y)\big\}.$$
The following lemma shows the existence of the limit in (\ref{e23}).
\begin{lem} [\cite{PW}]
The limit defining $P_{\mathcal{A}}(T,\phi)$ in (\ref{e23}) exists. Moreover,
if $\phi\colon [0,1)\rightarrow \mathbb{R}$ satisfies $\textrm{Var}_{1}(\phi)<\infty$ and $\textrm{Var}_{n}(\phi)\rightarrow 0$ as $n\rightarrow \infty$,
the value of $P_{\mathcal{A}}(T,\phi)$ remains the same even without taking the supremum over $x\in X_{\mathcal{A}}$ in (\ref{e23}).
\end{lem}
For $0<\alpha<1$ and $i\in \mathbb{N}$, we define
\begin{equation*}
\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)})
= \inf \left\{\rho \geq 0\colon \sum\limits_{a_{1},\ldots,a_{n}\in \mathcal{A}} \Big(\frac{1}{(\tau(i))^{\frac{n\alpha}{1-\alpha}}q_{n}(a_{1},\ldots,a_{n})}\Big)^{2\rho}\leq 1\right\}.
\end{equation*}
Following \cite{WW}, we call $\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)})$ the $n$-th pre-dimensional number with respect to $\mathcal{A}$ and $\alpha$. The properties of pre-dimensional numbers are presented in the following lemmas; the original ideas for the proofs date back to Good \cite{G} (see also \cite{M1}).
\begin{lem}[\cite{WW}]\label{lem26}
Let $\mathcal{A}$ be a finite or infinite subset of $\mathbb{N}.$ For $0<\alpha<1$ and $i\in \mathbb{N}$,
the limit $\lim_{n\to\infty}\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)})$ exists, denoted by $s(\mathcal{A},\alpha,{\tau(i)})$.
\end{lem}
By (\ref{e22}) and the definition of $\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)}),$ we know
$0\le\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)})\le1.$ Furthermore, Lemma \ref{lem26} implies that $0\le s(\mathcal{A},\alpha,{\tau(i)})\le1.$
\begin{lem}[\cite{WW}]\label{lem27}
For any $B\in \mathbb{N},$ put $\mathcal{A}_{B}=\{1,\ldots,B\}.$
The limit $\lim_{B \to \infty}s(\mathcal{A}_{B},\alpha,{\tau(i)})$ exists, and is equal to $s(\mathbb{N},\alpha,{\tau(i)})$.
\end{lem}
Similarly to pre-dimensional numbers $\{\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)}\},$ we define
\begin{equation*}
s_{n}(\mathcal{A},\alpha,{\tau(i)})=
\inf \left\{\rho \geq 0\colon \sum\limits_{a_{1},\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}
\Big(\frac{1}{q_{n}(a_{1},\ldots,a_{n-\lfloor na\rfloor},i,\ldots,i)}\Big)^{2\rho}\leq 1\right\}.
\end{equation*}
\begin{rem}\label{rem2}
We remark that
$$\sum\limits_{a_{1},\ldots,a_{n}\in \mathcal{A}} \Big(\frac{1}{(\tau(i))^{\frac{n\alpha}{1-\alpha}}q_{n}(a_{1},\ldots,a_{n})}\Big)^{2\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)})}\leq 1$$
and
$$\sum_{a_{1},\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}
\Big(\frac{1}{q_{n}(a_{1},\ldots,a_{n-\lfloor n\alpha\rfloor},i,\ldots,i)}\Big)^{2s_{n}(\mathcal{A},\alpha,{\tau(i)})}\leq1,$$
with equalities holding when $\mathcal{A}$ is finite.
\end{rem}
By Lemmas \ref{lem26} and \ref{lem27}, we have the following result.
\begin{lem} \label{lem29}
Let $\mathcal{A}$ be a finite or infinite subset of $\mathbb{N}.$ For $0<\alpha<1$ and $i\in \mathbb{N},$ we have
$$\lim_{n\to\infty}s_{n}(\mathcal{A},\alpha,{\tau(i)})=s(\mathcal{A},\alpha,{\tau(i)}).$$
In particular, if $\mathcal{A}=\mathbb{N},$ then
$$\lim_{n\to\infty}s_{n}(\mathbb{N},\alpha,{\tau(i)})=s(\mathbb{N},\alpha,{\tau(i)}).$$
\end{lem}
\begin{proof}
For $\varepsilon>0$ and $n$ large enough, we have
\begin{equation}\label{equ2.4}
2^{\frac{n-\lfloor n\alpha\rfloor}{2}\varepsilon}>64,
\end{equation}
\begin{equation}\label{equ2.5}
\frac{3}{(1-\alpha)(n\alpha-1)}+\frac{\log 4}{n\alpha-1}<\varepsilon,
\end{equation}
\begin{equation}\label{equ2.6}
|\widehat{s}_{n}(\mathcal{A},\alpha,{\tau(i)})-s(\mathcal{A},\alpha,{\tau(i)})|<\frac{\varepsilon}{2}.
\end{equation}
On the one hand, by Remark \ref{rem2}, we deduce that
\begin{align*}
1 \ge& \sum_{a_1,\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}\Big(\frac{1}{q_{n}(a_1,\ldots,a_{n-\lfloor n\alpha\rfloor},i,\ldots,i)}\Big)^{2s_{n}(\mathcal{A},\alpha,{\tau(i)})} \\
\ge & \sum_{a_1,\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}
\Big(\frac{1}{2q_{n-\lfloor n\alpha\rfloor}(a_1,\ldots,a_{n-\lfloor n\alpha\rfloor})q_{\lfloor n\alpha\rfloor}(i,\ldots,i)}\Big)^{2s_{n}(\mathcal{A},\alpha,{\tau(i)})}\\
\ge & \sum_{a_1,\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}\Big(\frac{1}{4q_{n-\lfloor n\alpha\rfloor}(a_1,\ldots,a_{n-\lfloor n\alpha\rfloor})
(\tau(i))^{\frac{\alpha}{1-\alpha}(n-\lfloor n\alpha\rfloor)}}\Big)^{2s_{n}(\mathcal{A},\alpha,{\tau(i)})}\\
\ge & \sum_{a_1,\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}\Big(\frac{1}{q_{n-\lfloor n\alpha\rfloor}(a_1,\ldots,a_{n-\lfloor n\alpha\rfloor})
(\tau(i))^{\frac{\alpha}{1-\alpha}(n-\lfloor n\alpha\rfloor)}}\Big)^{2s_{n}(\mathcal{A},\alpha,{\tau(i)})+\varepsilon},
\end{align*}
where the second inequality holds by Lemma \ref{lem21}(2);
the third inequality is right by Lemma \ref{lem21}(3) and the fact that $\frac{\alpha}{1-\alpha}(n-\lfloor n\alpha\rfloor)\ge \lfloor n\alpha\rfloor$ for $n\in \mathbb{N};$
the last inequality is true by Lemma \ref{lem21}(1) and (\ref{equ2.4}). This means that
$$s_{n}(\mathcal{A},\alpha,{\tau(i)})+\frac{\varepsilon}{2}\ge \widehat{s}_{n-\lfloor n\alpha\rfloor}(\mathcal{A},\alpha,{\tau(i)}).$$
On the other hand, we have
\begin{align*}
1 \ge & \sum_{a_1,\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}\Big(\frac{1}{q_{n-\lfloor n\alpha\rfloor}(a_1,\ldots,a_{n-\lfloor n\alpha\rfloor})
(\tau(i))^{\frac{\alpha}{1-\alpha}(n-\lfloor n\alpha\rfloor)}}\Big)^{2\widehat{s}_{n-\lfloor n\alpha\rfloor}(\mathcal{A},\alpha,{\tau(i)})}\\
\ge & \sum_{a_1,\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}\Big(\frac{1}{q_{n-\lfloor n\alpha\rfloor}(a_1,\ldots,a_{n-\lfloor n\alpha\rfloor})
(\tau(i))^{\lfloor n\alpha\rfloor+\frac{1}{1-\alpha}}}\Big)^{2\widehat{s}_{n-\lfloor n\alpha\rfloor}(\mathcal{A},\alpha,{\tau(i)})}\\
\ge & \sum_{a_1,\ldots,a_{n-\lfloor n\alpha\rfloor}\in \mathcal{A}}\Big(\frac{1}{q_{n}(a_1,\ldots,a_{n-\lfloor n\alpha\rfloor},i,\ldots,i)}\Big)^{2\widehat{s}_{n-\lfloor n\alpha\rfloor}(\mathcal{A},\alpha,{\tau(i)})+\varepsilon},
\end{align*}
where the second inequality is obtained by $\frac{\alpha}{1-\alpha}(n-\lfloor n\alpha\rfloor)\le \lfloor n\alpha\rfloor+\frac{1}{1-\alpha}$ for $n\in \mathbb{N}$;
the last inequality holds by Lemma \ref{lem21}(3) and (\ref{equ2.5}). This implies that
$$s_{n}(\mathcal{A},\alpha,{\tau(i)})\le \widehat{s}_{n-\lfloor n\alpha\rfloor}(\mathcal{A},\alpha,{\tau(i)})+\frac{\varepsilon}{2}.$$
Thus, by (\ref{equ2.6}), we obtain that
$$|s_{n}(\mathcal{A},\alpha,{\tau(i)})-s(\mathcal{A},\alpha,{\tau(i)})|<\varepsilon$$
for $n$ large enough. This completes the proof.
\end{proof}
For simplicity, write $s_{n}(\alpha,{\tau(i)})$ for $s_{n}(\mathbb{N},\alpha,{\tau(i)}),$ $s(\alpha,{\tau(i)})$ for $s(\mathbb{N},\alpha,{\tau(i)}).$
\begin{lem} [\cite{WW}] \label{lem210}
For $ 0<\alpha<1$ and $i\in \mathbb{N}$, we have$\colon$
{\rm{(1)}} $s(\alpha,{\tau(i)})>\frac{1}{2}$;
{\rm{(2)}} $s(\alpha,{\tau(i)})$ is non-increasing and continuous with respect to $\alpha$;
{\rm{(3)}} $\lim_{\alpha \to 0}{s(\alpha,{\tau(i)})}=1$ and $\lim_{\alpha \to 1}{s(\alpha,{\tau(i)})}=\frac{1}{2}.$
\end{lem}
From a point of view of dynamical system, $s(\alpha,{\tau(i)})$ can be regarded as the solution to the pressure function \cite{WWX}
$$P\Big(T,-s\Big(\log |T'|+\frac{\alpha}{1-\alpha}\log{\tau(i)}\Big)\Big)=0.$$
Furthermore, by Lemma \ref{lem210}, we may extend $s(\alpha,{\tau(i)})$ to $[0,1]$ as follows$\colon$
\begin{equation}\label{equ2.7}
s(\alpha,{\tau(i)})=\left\{
\begin{array}{ll}
1, & \ \ \ \alpha=0, \\
s(\alpha,{\tau(i)}), & \ \ 0<\alpha <1, \\
\frac{1}{2}, & \ \ \ \alpha=1.
\end{array}
\right.
\end{equation}
\section{Proof of Theorem \ref{TZ2}$\colon$ Upper~bound}\label{S_3}
In this section, we devote to estimating the upper bound of $E(\hat{\nu},\nu)$.
We first consider the case $\nu=0$.
\begin{lem}\label{full}
$\nu(x)=0$ for Lebesgue almost all $x\in[0,1).$
\end{lem}
\begin{proof}
Since $\sum_{n=1}^{\infty}|I_n(y)|^{\frac{1}{m}}<\infty$, we obtain by Theorem 2B in \cite{P67} that the set
$$\Big\{x\in[0,1)\colon |T^{n}(x)-y|<|I_{n}(y)|^{\frac{1}{m}} \text{ for infinitely many }n\in \mathbb{N}\Big\}$$ is of measure zero. Now
\begin{align*}
\{x\in[0,1)\colon \nu(x)>0\}&\subseteq\bigcup_{m=1}^{\infty}\Big\{x\in[0,1)\colon \nu(x)>\frac{1}{m}\Big\}
\\&\subseteq\bigcup_{m=1}^{\infty}\Big\{x\in[0,1)\colon |T^{n}(x)-y|<|I_{n}(y)|^{\frac{1}{m}} \text{ for infinitely many }n\in \mathbb{N}\Big\}.
\end{align*}
Hence $\{x\in[0,1)\colon \nu(x)>0\}$ is a null set.
This completes the proof.
\end{proof}
We now aim to determine the upper bound of $\dim_{H}E(\hat{\nu},\nu)$ for $0<\nu\le+\infty.$
\begin{lem}\label{lem31}
Let $x\in E(\hat{\nu},\nu)$, where $v>0$. If the continued fraction expansion of $x$ is not periodic, there exist two ascending sequences $\{n_k\}_{k=1}^{\infty}$ and $\{m_k\}_{k=1}^{\infty}$ depending on $x$ such that$\colon$
{\rm{(1)}} $n_k<m_k<n_{k+1}<m_{k+1}$ for $k\ge1$;
{\rm{(2)}} $a_{n_k+1}(x)=\cdots=a_{m_k}(x)=i$ for $k\ge1;$
{\rm{(3)}} $\displaystyle\liminf_{k\to\infty}\frac{m_k-n_k}{n_{k+1}}=\hat{\nu},$ $\displaystyle\limsup_{k\to\infty}\frac{m_k-n_k}{n_k}=\nu.$
\end{lem}
\begin{proof}
For $x=[a_{1}(x),a_{2}(x),\ldots]\in E(\hat{\nu},\nu)$, we define two sequences $\{n_{k}'\}_{k\geq1}$ and $\{m_{k}'\}_{k\geq 1}$ as follows$\colon$
\begin{align*}
m_{0}'=0,~ n_k'&=\min\{n\ge m_{k-1}'\colon a_{n+1}(x)=i\},\\
m_k'&=\max\{n\ge n_k'\colon a_{n_k'+1}=\cdots=a_{n}(x)=i\}.
\end{align*}
The fact that $\nu(x)>0$ guarantees the existence of $n_k'$, and thus $m_k'$ is well defined since
the continued fraction expansion of $x$ is not periodic.
Further, for all $k\ge 1,$ we have that $n_k'\le m_k'<n_{k+1}'$, and
$$\frac{1}{2(i+2)^{2}}|I_{m_k'-n_k'}(y)|\le|T^{n_k'}(x)-y|<|I_{m_k'-n_k'}(y)|,$$
where the first inequality holds by Lemma \ref{lem23}.
We also have $\limsup_{k\to\infty}(m_{k}'-n_{k}')= +\infty$ since $\nu(x)>0$. We then choose a subsequence of $\{(n_k', m_k')\}_{k\geq 1}$ as follows: put $(n_{1},m_{1})=(n'_{1},m'_{1});$
having choosen $(n_{k}, m_k)=(n_{j_{k}}', m_{j_{k}}')$, we set
$j_{k+1}=\min\big\{j>j_k\colon m_{j}'-n_{j}'> m_{k}-n_{k}\big\},$
and put $(n_{k+1},m_{k+1})=(n_{j_{k+1}}', m_{j_{k+1}}')$.
We claim that
$$ \liminf_{k\to\infty}\frac{m_k-n_k}{n_{k+1}}=\hat{\nu}(x),\ \ \limsup_{k\to\infty}\frac{m_k-n_k}{n_k}=\nu(x).$$
To prove the first assertion, we write
$\liminf_{n\to\infty}\frac{m_k-n_k}{n_{k+1}}=a.$
For $\varepsilon>0,$ there is a subsequence $\{k_j\}_{j=1}^{\infty}$ such that
$$m_{k_j}-n_{k_{j}}\le (a+\varepsilon)n_{{k_j}+1}.$$
Putting $N=n_{k_j}-1$, we have for all $n\in [1,N]$ that
$$|T^{n}(x)-y|\ge\frac{1}{2(i+2)^{2}}|I_{m_{k_j}-n_{k_j}}(y)|>|I_{n_{k_j+1}}(y)|^{a+2\varepsilon}>|I_{N}(y)|^{a+3\varepsilon},$$
where the second inequality holds by the fact $\lim_{n\to\infty}\frac{-\log|I_n(y)|}{2n}=\log\tau(i).$
We deduce that $\hat{\nu}(x)\le a+3\varepsilon$ by the definition of $\hat{\nu}(x)$.
On the other hand, when $k\gg1$, we have
$$m_{k}-n_{k}\ge (a-\varepsilon)n_{k+1}.$$
For $n_k\le N<n_{k+1}$,
$$|T^{n_{k}}(x)-y|\le|I_{m_k-n_k}(y)|<|I_{n_{k+1}}(y)|^{a-\varepsilon}<|I_{N}(y)|^{a-\varepsilon}.$$
From here we deduce that $\hat{\nu}(x)\ge a-\varepsilon$.
Letting $\varepsilon\to 0$ we complete the proof of the first assertion; the second one can be proved in a similar way.
\end{proof}
\begin{lem}\label{lem32}
If $0<\frac{\nu}{1+\nu}<\hat{\nu}\le\infty,$ $E(\hat{\nu},\nu)$ is at most countable, and $\dim_{H}E(\hat{\nu},\nu)=0.$
\end{lem}
\begin{proof} If $x\in E(\hat{\nu},\nu)$, and its continued fraction expansion is not periodic, then
by Lemma \ref{lem31}(2), there exist two sequences $\{n_k\}_{k=1}^{\infty}$ and $\{m_k\}_{k=1}^{\infty}$ depending on $x$ such that
$$\liminf_{k\to\infty}\frac{m_k-n_k}{n_{k+1}}=\hat{\nu},\ \ \limsup_{k\to\infty}\frac{m_k-n_k}{m_k}=\frac{\nu}{1+\nu}.$$
This yields $\hat{\nu}\le\frac{\nu}{1+\nu}$; the lemma follows.
\end{proof}
We devote to constructing a covering of $E(\hat{\nu},\nu)$ in the case where $0\le\hat{\nu}\le\frac{\nu}{1+\nu}<\infty$ and $0<\nu\le\infty.$ Since $E(0,\nu)$ is a subset of $\{x\in[0,1)\colon \nu(x)=\nu\},$ by Corollary \ref{Cor}, we have $\dim_{H}E(0,\nu)\le s\Big(\frac{\nu}{1+\nu},{\tau(i)}\Big)$, which is the desired upper bound estimate. Hence, we only need to deal with the case $0<\hat{\nu}\le\frac{\nu}{1+\nu}<\nu\le\infty.$
Whence, given any $x$ in the set $E(\hat{\nu},\nu)$ with non-periodic continued fraction expansion, we associate $x$ with two sequences $\{n_k\}, \{m_k\}$ as in Lemma \ref{lem31}.
The following properties hold:
(1) the sequence $\{m_k\}$ grows exponentially, more precisely,
there exists $C>0$, independent of $x$, such that when $k$ is large enough,
\begin{equation}\label{e30}
k \leq C\log{m_{k}}.
\end{equation}
Indeed, we have that $m_{k}-n_{k}\ge(\hat{\nu}/2)n_{k+1}$
for all large $k$, and thus
$$m_{k}\ge (1+\frac{\hat{\nu}}2)n_{k}
\ge (1+\frac{\hat{\nu}}2)m_{k-1}.$$
\medskip
(2) Write $\xi=\frac{\nu^{2}}{(1+\nu)(\nu-\hat{\nu})}$. For any $\varepsilon>0$, there exist infinitely many $k$ such that
\begin{equation}\label{e36}
\sum_{i=1}^{k}(m_{i}-n_{i})\ge m_{k}(\xi-\varepsilon). \end{equation}
To prove this, we apply a general form of the Stolz-Ces\`aro theorem which states that: if $b_n$
tends to infinity monotonically,
$$ \liminf_n\frac{a_n-a_{n-1}}{b_n-b_{n-1}}\le \liminf_n\frac{a_n}{b_n}\le \limsup_n\frac{a_n}{b_n} \le \limsup_n\frac{a_n-a_{n-1}}{b_n-b_{n-1}}.$$
We deduce from Lemma \ref{lem31} that
\begin{equation*}
\limsup_{k\to\infty}\frac{m_{k}}{n_{k}}=1+\nu
\end{equation*}
and
\begin{equation*}
\liminf_{k\to\infty}\frac{m_{k}}{n_{k+1}}
\ge \liminf_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k+1}}
\cdot\liminf_{k\to\infty}\frac{m_{k}}{m_{k}-n_{k}}
=\frac{\hat{\nu}(1+\nu)}{\nu}.
\end{equation*}
Hence
\begin{equation}\label{e34}
\begin{split}
\liminf_k\frac{\sum_{i=1}^{k}(m_i-n_i)}{m_{k+1}}&\ge
\liminf_k\frac{m_k-n_k}{m_{k+1}-m_k}\\&\ge
\liminf_k\frac{m_k-n_k}{n_{k+1}}\cdot\frac{1}{\limsup_k\frac{m_{k+1}}{n_{k+1}}-\liminf_k\frac{m_k}{n_{k+1}}}\\&\ge
\frac{\hat{\nu}\nu}{\big(\nu-\hat{\nu}\big)\big(1+\nu\big)},
\end{split}
\end{equation}
and thus
\begin{align*}
\sum_{i=1}^{k}(m_{i}-n_{i})&
\geq \Big(\frac{\hat{\nu}\nu}{\big(\nu-\hat{\nu}\big)\big(1+\nu\big)}-\frac{\varepsilon}{2}\Big)m_{k}+(m_{k}-n_{k})
\end{align*}
holds for $k$ large enough. On the other hand, there exist infinitely many $k$ such that
$$m_k-n_k\ge \Big(\frac{\nu}{1+\nu}-\frac{\varepsilon}{2}\Big)m_k.$$
We then readily check that (\ref{e36}) holds for such $k$.
\medskip
We now construct a covering of $E(\hat{\nu},\nu)$. To this end,
we collect all sequences $(\{n_{k}\}, \{m_{k}\})$ associated with some $x\in E(\hat{\nu},\nu)$ as in Lemma \ref{lem31} to form a set
\begin{align*}
\Omega
= \Big\{(\{n_{k}\},\{m_{k}\})\ \colon \text{Conditions (1) \& (3) in Lemma \ref{lem31} are fulfilled} \Big\}.
\end{align*}
For $(\{n_{k}\},\{m_{k}\})\in \Omega$, write
\begin{align*}
&H(\{n_{k}\},\{m_{k}\})=\{x\in[0,1)\colon \text{Condition (2) in Lemma \ref{lem31} is fulfilled} \},\\
&\Lambda_{k, m_k}=\Big\{(n_1,m_1; \ldots; n_{k-1},m_{k-1};n_{k})\colon n_{1}<m_{1}<\cdots<m_{k-1}<n_{k}<m_k, (\ref{e36})\text{ holds} \Big\},\\
&\mathcal{D}_{n_1,m_1;\ldots;n_{k}, m_{k}}=\Big\{(\sigma_{1},\ldots,\sigma_{m_{k}})\in \mathbb{N}^{m_{k}}\colon \sigma_{n_{j}+1}=\cdots=\sigma_{m_{j}}=i \text{ for all }1\leq j\leq k \Big\}.
\end{align*}
Based on the previous analysis, we obtain a covering of $E(\hat{\nu},\nu),$ i.e.,
\begin{equation*}
\begin{split}
E(\hat{\nu},\nu) &\subseteq \bigcup_{(\{n_{k}\},\{m_{k}\})\in \Omega}H(\{n_{k}\},\{m_{k}\}) \\&
\subseteq\bigcap_{K=1}^{\infty} \bigcup_{k=K}^{\infty}
\bigcup_{m_{k}\ge e^{\frac{k}{C}}}
\bigcup_{(n_{1},m_{1},\ldots,m_{k-1},n_{k})\in\Lambda_{k, m_k}}
\bigcup_{(a_{1},\ldots,a_{m_{k}})\in \mathcal{D}_{n_1,m_1;\ldots;n_{k}, m_{k}}}I_{m_{k}}(a_{1},\ldots,a_{m_{k}}).
\end{split}
\end{equation*}
For $\varepsilon>0$, putting $t=s(\xi-\varepsilon,{\tau(i)})+\frac{\varepsilon}{2}$, we have that $t>s(\xi,{\tau(i)})$ and $t>\frac{1}{2}$ by Lemma \ref{lem210}(2) and (\ref{equ2.7}).
We are now in a position to estimate $\mathcal{H}^{t+\frac{\varepsilon}{2}}\big(E(\hat{\nu},\nu)\big)$, the $(t+\frac{\varepsilon}{2})$-dimensional Hausdorff measure of $E(\hat{\nu},\nu)$.
\begin{lem}\label{lem35}
For $\varepsilon>0$, we have $\mathcal{H}^{t+\frac{\varepsilon}{2}}\big(E(\hat{\nu},\nu)\big)<+\infty.$
\end{lem}
\begin{proof}
By Lemma \ref{lem29}, there exists $K\in \mathbb{N}$ such that for all $k\geq K,$
\begin{equation}\label{e35}
s_{m_{k}}(\xi-\varepsilon,{\tau(i)})\leq t,
\end{equation}
\begin{equation}\label{e37}
(4^{2t+\varepsilon}k)^{2C\log k}<2^{\frac{k-1}{4}\varepsilon}.
\end{equation}
Writing $\psi(m_{k})=m_{k}-\sum_{i=1}^{k}(m_{i}-n_{i})$ when $m_{k}\geq K$, we have that
\begin{equation*}
\begin{split}
&\sum\limits_{(a_{1},\ldots,a_{m_{k}})\in {\mathcal{D}_{n_1,m_1;\ldots;n_{k}, m_{k}}}}
|I_{m_{k}}(a_{1},\ldots,a_{m_{k}})|^{t+\frac{\varepsilon}{2}}\\
\le&\sum\limits _{(a_{1},\ldots,a_{m_{k}})\in {\mathcal{D}_{n_1,m_1;\ldots;n_{k}, m_{k}}}}
\Big(\frac{1}{q_{m_{k}}(a_{1},\ldots,a_{m_{k}})}\Big)^{2t+\varepsilon}\\
\leq&\sum\limits _{a_{1},\ldots,a_{\psi(m_{k})}\in \mathbb{N}}2^{k(2t+\varepsilon)}\Big(\frac{1}{q_{\psi(m_{k})}(a_{1},\ldots,
a_{\psi(m_{k})})q_{m_{1}-n_{1}}(i,\ldots,i)\cdots q_{m_{k}-n_{k}}(i,\ldots,i)}\Big)^{2t+\varepsilon}\\
\leq&\sum\limits _{a_{1},\ldots,a_{\psi(m_{k})}\in \mathbb{N}}4^{k(2t+\varepsilon)}\Big(\frac{1}{q_{m_{k}}(a_{1},\ldots,a_{\psi(m_{k})},i,\ldots,i)}\Big)^{2t+\varepsilon}\\
\leq&\sum\limits _{a_{1},\ldots,a_{m_{k}-\lfloor m_{k}(\xi-\varepsilon)\rfloor}\in \mathbb{N}} 4^{k(2t+\varepsilon)}
\Big(\frac{1}{q_{m_{k}}(a_{1},\ldots,a_{m_{k}-\lfloor m_{k}(\xi-\varepsilon)\rfloor},i,\ldots,i)}\Big)^{2s_{m_{k}}(\xi-\varepsilon,{\tau(i)})+\varepsilon}\\
\leq&4^{k(2t+\varepsilon)}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{2}\varepsilon},
\end{split}
\end{equation*}
where the first two inequalities hold by Lemmas \ref{lem22} and \ref{lem21};
the penultimate one follows by (\ref{e36}) and (\ref{e35}); the last one follows by Remark \ref{rem2} and Lemma \ref{lem21}(1).
Therefore,
\begin{equation*}
\begin{split}
&\mathcal{H}^{t+\frac{\varepsilon}{2}}\big(E(\hat{\nu},\nu)\big)\\
\leq &\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty}\sum_{m_{k}= e^{\frac{k}{C}}}^{\infty}
\sum_{(n_{1},m_{1},\ldots,m_{k-1},n_{k})\in\Lambda_{k, m_k}}
\sum\limits _{(a_{1},\ldots,a_{m_{k}})\in \mathcal{D}_{n_1,m_1;\ldots;n_{k}, m_{k}} }|I_{m_{k}}(a_{1},\ldots,a_{m_{k}})|^{t+\frac{\varepsilon}{2}}\\
\leq&\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty}
\sum\limits_{m_{k}=e^{\frac{k}{C}}}^{\infty}
\sum_{n_{k}=1}^{m_{k}}\sum_{m_{k-1}=1}^{n_{k}}\cdots\sum\limits_{m_{1}=1}^{n_{2}}
\sum\limits_{n_{1}=1}^{m_{1}}4^{k(2t+\varepsilon)}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{2}\varepsilon}\\
\leq&\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty}
\sum\limits_{m_{k}=e^{\frac{k}{C}}}^{\infty}(4^{2t+\varepsilon}m_{k})^{2C\log m_{k}}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{2}\varepsilon}\\
\leq&\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty} \sum\limits_{m_{k}=e^{\frac{k}{C}}}^{\infty}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{4}\varepsilon}
\leq \frac{1}{1-(\frac{1}{2})^{\frac{\varepsilon}{4}}}
\sum\limits_{k=1}^{\infty} \Big(\frac{1}{2^{\varepsilon}}\Big)^{\frac{e^{\frac{k}{C}}-1}{4}} < +\infty,
\end{split}
\end{equation*}
where the third and fourth inequalities follow from (\ref{e30}) and (\ref{e37}) respectively.
\end{proof}
By Lemma \ref{lem35}, we obtain the desired inequality $\dim_{H}E(\hat{\nu},\nu)\le s(\xi,{\tau(i)})$ by letting $\varepsilon\to 0$.
\section{Proof of Theorem \ref{TZ2}$\colon$ Lower~bound}\label{S_4}
In this section we establish the lower bound of $\dim_{H}E(\hat{\nu},\nu).$ Since $E(0,0)$ is of full Lebesgue measure and $\dim_{H}E(\hat{\nu},\nu)=0$ for $\hat{\nu}>\frac{\nu}{1+\nu},$ we need only consider the cases $0\le\hat{\nu}\le\frac{\nu}{1+\nu}<\nu<\infty$ or $\nu=\infty.$
Let us start by treating the case $0\le\hat{\nu}\le\frac{\nu}{1+\nu}<\nu<\infty$. We claim that there exist two sequences of natural numbers $\{n_{k}\}_{k=1}^{\infty}$ and $\{m_{k}\}_{k=1}^{\infty}$
satisfying the following conditions$\colon$
{\rm{(1)}} $n_{k}< m_{k}< n_{k+1}$ and $(m_{k}-n_{k})\leq(m_{k+1}-n_{k+1})$ for $k\geq 1$;
{\rm{(2)}} $\lim_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k+1}}=\hat{\nu}$;
{\rm{(3)}} $\lim_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k}}=\nu$.
\noindent Indeed, when $\hat{\nu}>0$, we may take
$$n_{1}=2,~ n_{k+1}=\left\lfloor \frac{\nu}{\hat{\nu}}
\Big(n_{k}+\frac{1}{\nu}\Big)\right\rfloor+2,~ m_{k}=\left\lfloor (1+\nu)n_{k}\right\rfloor+1;$$
when $\hat{\nu}=0$, we may take
$$n_{k}=\left\lfloor(1+\nu)2^{2^{2{k}}}\right\rfloor+2,~ m_{k}=\left\lfloor(1+\nu)n_{k}\right\rfloor+1.$$
From now on, we fix two such sequences $\{n_{k}\}, \{m_{k}\}$; for any $B\ge i+1$, we define
\begin{equation*}
E(B)=\big\{x\in[0,1)\colon 1\leq a_{n}(x)\leq B, a_{n_{k}+1}(x)=\cdots=a_{m_{k}}(x)=i,
n\geq1 \text{ and } k\geq1\big\}.
\end{equation*}
The lower bound estimate of $\dim_H E(\hat\nu,\nu)$ will be established in the following way:
we provide a lower bound of $\dim_{H}E(B)$;
build an injective mapping $f$ from $E(B)$ to $ E(\hat\nu,\nu)$ and prove that $f$ is dimension-preserving.
\subsection{Lower bound of $\dim_{H}E(B)$}\label{S_{4.1}}
Before proceeding, we cite an analogous definition of the pre-dimensional numbers.
Let $l_{k}=m_{k}-m_{k-1}$ for $k\geq 1$ ($m_{0}=0$ by convention). Let
$$\widetilde{f}_{k}(s,{\tau(i)})
=\sum\limits_{1\leq a_{m_{k-1}+1},\ldots,a_{n_{k}}\leq B}
\Big(\frac{1}{q_{l_{k}}(a_{m_{k-1}+1},\ldots,a_{n_{k}}, i,\ldots,i)}\Big)^{2s}.$$
We define $\widetilde{s}_{l_{k}}(\mathcal{A}_{B},\xi,\tau(i))$ to be the solution of the equation $\widetilde{f}_{k}(s,\tau(i))=1$.
\begin{lem}\label{lem41}
The limit $\lim_{k\to\infty}\widetilde{s}_{l_{k}}(\mathcal{A}_{B},\xi,{\tau(i)})$ exists, and is equal to $s(\mathcal{A}_{B},\xi,{\tau(i)}).$
\end{lem}
\begin{proof}
By Lemma \ref{lem29} and the fact $l_{k}\to\infty$ as $k\rightarrow\infty$ (cf. Condition (1)), we deduce that, for any $\varepsilon>0,$
when $k\gg 1,$
\begin{equation}\label{e43}
|s_{l_{k}}(\mathcal{A}_{B},\xi+\varepsilon,\tau(i))-s(\mathcal{A}_{B},\xi+\varepsilon,\tau(i))|<\frac{\varepsilon}{2}.
\end{equation}
Further, from Conditions (2) and (3) we have that
\begin{align*}
\lim_{k\to\infty}\frac{m_{k}-n_{k}}{l_{k}}
=\lim_{k\to\infty}\frac{\frac{m_k-n_k}{m_k}\cdot\frac{m_k}{n_k}}{\frac{m_k}{n_k}-\frac{m_{k-1}-n_{k-1}}{n_k}\cdot\frac{m_{k-1}}{m_{k-1}-n_{k-1}}}
=\xi,
\end{align*}
and thus for $k\gg 1$,
\begin{equation}\label{e42}
\lfloor l_{k}(\xi-\varepsilon)\rfloor
\leq m_{k}-n_{k}\leq \lfloor l_{k}(\xi+\varepsilon)\rfloor.
\end{equation}
Hence, by (\ref{e43}) and (\ref{e42}), we obtain that
\begin{align*}
&\sum\limits_{1\leq a_{1},\ldots,a_{n_{k}-m_{k-1}}\leq B} \Big(\frac{1}{q_{l_{k}}(a_{1},\ldots,a_{n_{k}-m_{k-1}},
i,\ldots,i)}\Big)^{2\big(s(\mathcal{A}_{B},
\xi+\varepsilon,\tau(i))-\varepsilon\big)} \\
\geq&\sum\limits_{1\leq a_{1},\ldots,a_{n_{k}-m_{k-1}}\leq B} \Big(\frac{1}{q_{l_{k}}(a_{1},\ldots,a_{n_{k}-m_{k-1}},
i,\ldots,i)}\Big)^{2s_{l_{k}}(\mathcal{A}_{B},
\xi+\varepsilon,\tau(i))-\varepsilon}\\
\geq&\sum\limits_{1\leq a_{1},\ldots,a_{l_{k}-\lfloor l_{k}(\xi+\epsilon)\rfloor}\leq B}
\Big(\frac{1}{q_{l_{k}}(a_{1},\ldots,a_{l_{k}-\lfloor l_{k}(\xi+\varepsilon)\rfloor},
i,\ldots,i)}\Big)^{2s_{l_{k}}(\mathcal{A}_{B},
\xi+\varepsilon,{\tau(i)})-\varepsilon}\\
\geq&\Big(\frac{1}{q_{l_{k}}(B,\ldots,B)}\Big)^{-\varepsilon}\geq\tau(B)^{l_{k}\varepsilon}\geq 1
\end{align*}
and
\begin{align*}
&\sum\limits_{1\leq a_{1},\ldots,a_{n_{k}-m_{k-1}}\leq B} \Big(\frac{1}{q_{l_{k}}(a_{1},\ldots,a_{n_{k}-m_{k-1}},
i,\ldots,i)}\Big)^{2\big(s(\mathcal{A}_{B},
\xi-\varepsilon,\tau(i))+\varepsilon\big)} \\
\leq&\sum\limits_{1\leq a_{1},\ldots,a_{l_{k}-\lfloor l_{k}(\xi-\varepsilon)\rfloor}\leq B}
\Big(\frac{1}{q_{l_{k}}(a_{1},\ldots,a_{l_{k}-\lfloor l_{k}(\xi-\varepsilon)\rfloor},
i,\ldots,i)}\Big)^{2s_{l_{k}}(\mathcal{A}_{B},
\xi-\varepsilon,\tau(i))+\varepsilon}\\
\leq&\Big(\frac{1}{q_{l_{k}}(1,\ldots,1)}\Big)^{\varepsilon}<1.
\end{align*}
By the monotonicity of $\widetilde{f}_{k}(s,{\tau(i)})$ with respect to $s$, we have
$$s(\mathcal{A}_{B},\xi+\varepsilon,\tau(i))-\varepsilon
\le\widetilde{s}_{l_{k}}(\mathcal{A}_{B},\xi,{\tau(i)})
\le s(\mathcal{A}_{B},\xi-\varepsilon,\tau(i))+\varepsilon,$$
which completes the proof.
\end{proof}
\subsubsection{\textbf{Supporting measure.}}
We define a probability measure $\mu$ on $E(B)$ by distributing mass among the basic intervals. We introduce the symbolic space to code these basic intervals: write $\mathcal{A}_{B}=\{1,\ldots,B\}$; for $n\geq 1,$ set
\begin{equation*}
\mathcal{B}_{n}=\big\{(\sigma_{1},\ldots,\sigma_{n})\in \mathcal{A}_{B}^{n}\colon\sigma_{j}=i \text{ for } n_{k}<j\leq m_{k} \text{ with some } k\geq 1\big\}.
\end{equation*}
\noindent Step I$\colon$ For $(a_{1},\ldots,a_{m_{1}})\in \mathcal{B}_{m_{1}}$, we define
$$\mu\big(I_{m_{1}}(a_{1},\ldots,a_{m_{1}})\big)=\Big(\frac{1}{q_{l_{1}}(a_{1},\ldots,a_{m_{1}})}\Big)^{2\widetilde{s}_{l_{1}}(\mathcal{A}_{B},\xi,{\tau(i)})}$$
and for $1\leq n<m_{1}$, set
$$\mu\big(I_{n}(a_{1},\ldots,a_{n})\big)=\sum\limits_{a_{n+1},\ldots,a_{m_{1}}}\mu\big(I_{m_{1}}(a_{1},\ldots,a_{n},a_{n+1},\ldots,a_{m_{1}})\big),$$
where the summation is taken over all $(a_{n+1},\ldots,a_{m_{1}})$ with $(a_{1},\ldots,a_{m_{1}})\in \mathcal{B}_{m_{1}}$.
\noindent Step II$\colon$ Assuming that $\mu\big(I_{m_{k}}(a_{1},\ldots,a_{m_{k}})\big)$ is defined for some $k\geq 1$, we define
$$\mu\big(I_{m_{k+1}}(a_{1},\ldots,a_{m_{k+1}})\big)=\mu\big(I_{m_{k}}(a_{1},\ldots,a_{m_{k}})\big)\cdot \Big(\frac{1}{q_{l_{k+1}}(a_{m_{k}+1},\ldots,a_{m_{k+1}})}\Big)^{2\widetilde{s}_{l_{k+1}}(\mathcal{A}_{B},\xi,{\tau(i)})}$$
and for $m_{k}<n<m_{k+1}$, set
$$\mu\big(I_{n}(a_{1},\ldots,a_{n})\big)=\sum\limits_{a_{n+1},\ldots,a_{m_{k+1}}} \mu\big(I_{m_{k+1}}(a_{1},\ldots,a_{n},a_{n+1},\ldots, a_{m_{k+1}})\big).$$
Likewise, the last summation is taken under the restriction that $(a_{1},\ldots,a_{m_{k+1}})\in \mathcal{B}_{m_{k+1}}.$
\noindent Step III$\colon$ We have distributed the measure among basic intervals. By the definition of $\widetilde{s}_{l_{k}}(\mathcal{A}_{B},\xi,{\tau(i)})$, we readily check the consistency: for $n\geq 1$ and $(a_{1},\ldots,a_{n})\in \mathcal{B}_{n}$,
$$\mu\big(I_{n}(a_{1},\ldots,a_{n})\big)=\sum\limits_{a_{n+1}}\mu\big(I_{n+1}(a_{1},\ldots,a_{n},a_{n+1})\big).$$
We then extend the measure to all Borel sets by Kolmogorov extension theorem. The extension measure is also denoted by $\mu$.
\smallskip
From the construction, we know that $\mu$ is supported on $E(B)$ and
$$\mu\big(I_{m_{k}}(a_{1},\ldots,a_{m_{k}})\big)
=\prod\limits_{j=1}^{k}\Big(\frac{1}{q_{l_{j}}(a_{m_{j-1}+1},\ldots,a_{m_{j}})}\Big)^{2\widetilde{s}_{l_{j}}(\mathcal{A}_{B},\xi,{\tau(i)})},~~\sum\limits_{a_{1}\in \mathcal{B}_{1}}\mu\big(I_{1}(a_{1})\big)=1.$$
\subsubsection{\textbf{H\"{o}lder exponent of $\mu$.}} We shall start with the study of a basic interval.
For $0<\varepsilon<{s(\mathcal{A}_{B},\xi,{\tau(i)})}/{4}$,
by Lemmas \ref{lem29}, \ref{lem41} and the fact that $m_k$ grows exponentially, we can find $K\in \mathbb{N}$ such that for any $k, j\geq K,$
\begin{equation}\label{e441}
|\widetilde{s}_{l_{k}}(\mathcal{A}_{B},\xi,{\tau(i)})-s(\mathcal{A}_{B},\xi,{\tau(i)})|<\varepsilon,
\end{equation}
\begin{equation}\label{e442}
|\widetilde{s}_{l_{k}}(\mathcal{A}_{B},\xi,{\tau(i)})-s_j(\mathcal{A}_{B},\xi,{\tau(i)})|<\frac{\varepsilon\log 2}{2\log (B+1)}:=\varepsilon',
\end{equation}
and \begin{equation}\label{e443}
\max\big\{(B+1)^K,2^k\big\}\le \frac14\big(q_{m_k}(a_1,\ldots,a_{m_k})\big)^\varepsilon.
\end{equation}
\begin{lem}\label{lem42}
Let $n\ge m_K$. For $(a_{1}, \ldots, a_{n})\in \mathcal{B}_{n}$, we have
$$\mu\big(I_{n}(a_{1},\ldots,a_{n})\big)\leq C_0\cdot |I_{n}(a_1,\ldots,a_n)|^{s(\mathcal{A}_{B},\xi,{\tau(i)})-3\varepsilon},$$
where $C_0=(B+1)^{2(l_1+\cdots+l_{K-1})}.$
\end{lem}
\begin{proof}
To shorten notation, we will write $\widetilde{s}_{l_{k}}, s_k$ and $s$ instead of $\widetilde{s}_{l_{k}}(\mathcal{A}_{B},\xi,{\tau(i)}), s_k(\mathcal{A}_{B},\xi,{\tau(i)})$ and $s(\mathcal{A}_{B},\xi,{\tau(i)})$, respectively. Fixing $(a_{1}, \ldots, a_{n})\in \mathcal{B}_{n}$, we also write $I_n$ for $I_n(a_1,\ldots,a_n)$,
$q_n$ for $q_n(a_1,\ldots,a_n)$ and $q_{l_j}$ for ${q_{l_{j}}(a_{m_{j-1}+1},\ldots,a_{m_{j}})}$ when no confusion can arise. The proof falls naturally into three parts according to the range of $n$.
\smallskip
\underline{\textsc{Case 1$\colon$} $n=m_{k}$ for $k\ge K$}
By Lemmas \ref{lem22} and \ref{lem23}, (\ref{e441}), (\ref{e443}) and the fact $q_{l_j}\le(B+1)^{l_j}$, we obtain
\begin{align*}
\mu(I_{m_{k}})&=\prod\limits_{j=1}^{k} q_{l_j}^{-2\widetilde{s}_{l_j}}=\prod\limits_{j=1}^{K}q_{l_j}^{-2\widetilde{s}_{l_j}}\cdot\prod\limits_{j=K+1}^{k}q_{l_j}^{-2\widetilde{s}_{l_j}}\le C_0\prod\limits_{j=1}^{K}q_{l_j}^{-2(s-\varepsilon)}\cdot\prod\limits_{j=K+1}^{k}q_{l_j}^{-2(s-\varepsilon)}\\
&\leq C_{0}2^{2(k-1)}(q_{m_k})^{-2(s-\varepsilon)}\le \frac{C_{0}}4(q_{m_k})^{-2(s-2\varepsilon)}\leq C_{0}|I_{m_{k}}|^{s-2\varepsilon}.
\end{align*}
\smallskip
\underline{\textsc{Case 2$\colon$} $m_{k}<n< n_{k+1}$ for $k\ge K$}
In this case, we have
$$\mu(I_n)=\sum_{a_{n+1},\ldots,a_{m_{k+1}}}\mu(I_{m_{k+1}})=\sum\prod_{j=1}^{k+1}(q_{l_j})^{-2\widetilde{s}_{l_{j}}}= \prod_{j=1}^{k}(q_{l_j})^{-2\widetilde{s}_{l_{j}}}\cdot \sum (q_{l_{k+1}})^{-2\widetilde{s}_{l_{k+1}}}.$$
We have already seen in \textsc{Case 1} that $\prod_{j=1}^{k}(q_{l_j})^{-2\widetilde{s}_{l_{j}}}\le C_{0}2^{2(k-1)}(q_{m_k})^{-2(s-\varepsilon)}$. And
$$\sum (q_{l_{k+1}})^{-2\widetilde{s}_{l_{k+1}}}\le
\big({q_{n-m_{k}}(a_{m_{k}+1},\ldots,a_{n})}\big)^{-2(s-\varepsilon)} \cdot
\sum \big({q_{m_{k+1}-n}(a_{n+1},\ldots,a_{n_{k+1}},i,\ldots,i)}\big)^{-2\widetilde{s}_{l_{k+1}}}.
$$
We then obtain that
$$ \mu(I_n)\le C_02^{2k}(q_n)^{-2(s-\varepsilon)}\cdot \sum_{a_{n+1},\ldots,a_{m_{k+1}}} \big({q_{m_{k+1}-n}(a_{n+1},\ldots,a_{n_{k+1}},i,\ldots,i)}\big)^{-2\widetilde{s}_{l_{k+1}}}.$$
Now we need an upper estimate of the last sum. By the definition of $\widetilde{s}_{l_{k+1}}$, we have that
$$ \sum\limits_{1\leq a'_{m_{k}+1},\ldots,a'_n,a_{n+1},\ldots,a_{n_{k+1}}\leq B}
\big(q_{l_{k+1}}( a'_{m_{k}+1},\ldots,a'_n,a_{n+1},\ldots,a_{n_{k+1}},i,\ldots,i)\big)^{-2\widetilde{s}_{l_{k+1}}}=1.$$
This yields that
$$\sum
\big( q_{n-m_{k}}(a'_{m_{k}+1},\ldots,a'_{n})\big)^{-2\widetilde{s}_{l_{k+1}}}\cdot
\sum
\big(q_{m_{k+1}-n}(a_{n+1},\ldots,a_{n_{k+1}},i,\ldots,i)\big)^{-2\widetilde{s}_{l_{k+1}}}\le 4.$$
We will bound the first sum from below to reach the desired upper estimate of the second sum. We consider two cases.
{\rm{(1)}} If $n-m_{k}< K$,
$$\sum_{a'_{m_{k}+1},\ldots,a'_{n}}\big(q_{n-m_k}(a'_{m_{k}+1},\ldots,a'_{n})\big)^{-2\widetilde{s}_{l_{k+1}}}
\geq \big(q_{n-m_k}(B,\ldots,B)\big)^{-2}\geq( {B+1})^{-2K}.$$
And thus, by (\ref{e443}), we reach that
$$
\mu(I_{n})
\leq C_{0}2^{2k+2}(B+1)^{2K} ({q_{n}} )^{-2(s-\varepsilon)}\le\frac{C_{0}}4(q_{m_k})^{-2(s-3\varepsilon)}
\leq C_{0} |I_{n}|^{s-3\varepsilon}.$$
{\rm{(2)}} If $n-m_{k}\geq K$, then, by (\ref{e442}), (\ref{e443}) and Remark \ref{rem2}, we have
\begin{align*}
&\sum_{a'_{m_{k}+1},\ldots,a'_{n}}\big(q_{n-m_k}(a'_{m_{k}+1},\ldots,a'_{n})\big)^{-2\widetilde{s}_{l_{k+1}}}
\ge \sum_{a'_{m_{k}+1},\ldots,a'_{n}}\big(q_{n-m_k}(a'_{m_{k}+1},\ldots,a'_{n})\big)^{-2{s}_{n-m_k}-\varepsilon'}\\
&\ge\sum_{a'_{m_{k}+1},\ldots,a'_{n-\lfloor(n-m_k)\xi\rfloor}}\Big({q_{n-m_{k}}(a'_{m_{k}+1},\ldots,a'_{n-\lfloor(n-m_k)\xi\rfloor},
i,\ldots,i)}\Big)^{-2s_{n-m_{k}}-\varepsilon'}\\
&\ge\Big({q_{n-m_{k}}(B,\ldots,B)}\Big)^{-\varepsilon'}
\ge({B+1})^{-(n-m_{k})\varepsilon'}
\ge({B+1})^{-n\varepsilon'}
\ge2^{\frac{-n\varepsilon}{2}}.
\end{align*}
Therefore,
$$
\mu(I_{n})
\leq C_{0}2^{2k+2}2^{\frac{n\varepsilon}{2}} ({q_{n}} )^{-2(s-2\varepsilon)}
\leq \frac{C_{0}}2({q_{n}})^{-2(s-3\varepsilon)}
\leq C_{0}|I_{n}|^{s-3\varepsilon}.
$$
\smallskip
\underline{\textsc{Case 3$\colon$} $n_{k+1}\leq n<m_{k+1}$ for $k\ge K$}
In this case, since $(a_{n+1},\ldots,a_{m_{k+1}})=(i,\ldots,i),$ we have
$$\mu(I_{n}(a_{1},\ldots,a_{n}))=\mu(I_{m_{k+1}}(a_{1},\ldots,a_{m_{k+1}})),$$
then
$$
\mu(I_{n})
\leq C_{0}|I_{m_{k+1}}|^{s-2\varepsilon}
\leq C_{0}|I_{n}|^{s-2\varepsilon}.
$$
These conclude the verification of Lemma.
\end{proof}
Now we study the H\"older exponent for the measure of a general ball $B(x,r)$.
\begin{lem}\label{lem43} For $x\in E(B)$ and $r>0$ small enough, we have
$$\mu(B(x,r))\leq C_{0}\cdot r^{s(\mathcal{A}_{B},\xi,{\tau(i)})-4\varepsilon}.$$
\end{lem}
\begin{proof}
Let $x=[a_{1},a_{2},\ldots]$ be its continued fraction expansion.
Let $n\ge K+2$ be the integer such that
$$|I_{n+1}(a_{1},\ldots,a_{n+1})|\leq r<|I_{n}(a_{1},\ldots,a_{n})|.$$
Therefore it follows from Lemmas \ref{lem24} and \ref{lem42} that
\begin{align*}
\mu(B(x,r))&\leq \mu(I_{n-2}(a_{1},\ldots,a_{n-2}))
\leq C_{0}\cdot|I_{n-2}(a_{1},\ldots,a_{n-2})|^{s(\mathcal{A}_{B},\xi,{\tau(i)})-3\varepsilon}
\\& \le C_{0}(B+1)^{6}\cdot |I_{n+1}(a_{1},\ldots,a_{n+1})|^{s(\mathcal{A}_{B},\xi,{\tau(i)})-3\varepsilon}
\\& \le C_{0} \cdot |I_{n+1}(a_{1},\ldots,a_{n+1})|^{s(\mathcal{A}_{B},\xi,{\tau(i)})-4\varepsilon} \le C_{0}\cdot r^{s(\mathcal{A}_{B},\xi,{\tau(i)})-4\varepsilon}.
\end{align*}
\end{proof}
Applying mass distribution principle (see Lemma \ref{lem213}), letting $\varepsilon\to 0$, we conclude that
$$\dim_{H}E(B)\geq s(\mathcal{A}_{B},\xi,{\tau(i)}).$$
\subsection{Lower bound of $\dim_{H}E(\hat{\nu},\nu)$}\label{S_{4.2}}
We build a mapping $f$ from $E(B)$ to $ E(\hat\nu,\nu)$ and prove that $f$ is dimension-preserving.
Fix an integer $d>B$. For $x=[a_{1},a_{2},\ldots]$ in $E(B)$, we remark that the continued fraction of $x$ is the concatenation of $\mathbb B_0=[a_1,\ldots,a_{n_1}]$ and the blocks
$$\mathbb B_k=[\,\underbrace{ i,\ldots, i}_{m_k-n_k}, a_{m_{k}+1}, \ldots, a_{n_{k+1}}] \quad (k=1,2,\ldots).$$
In the block $\mathbb B_k$, from the beginning we insert a digit $d$ after each $m_k-n_k$ digits to obtain a new block $\mathbb B_k'$, that is,
$$\mathbb B'_k=[\, d, i,\ldots, i, d, a_{m_{k}+1}, \ldots, a_{m_{k}+(m_k-n_k)}, d,
\ldots, a_{n_{k+1}}].$$
Concatenating the blocks $\mathbb B_0, \mathbb B'_1, \mathbb B'_2,\ldots$, we get
$[\mathbb B_0, \mathbb B'_1, \mathbb B'_2,\ldots]$, which is a continued fraction expansion of some $\bar{x}$. We then define $f(x)=\bar{x}.$
Let $\mathbf{K}=\{k_n\}\subset \mathbb N$ be the collection of the occurrences of the digit $d$ in the continued expansion of $\bar x$. It is trivially seen that $\mathbf K$ is independent of the choice of $x\in E(B)$, and, in the notation of Lemma \ref{lem214}, $\phi_\mathbf K(\bar x)=x$ for $x\in E(B)$
Let $h_k$ be the length of the block $\mathbb B'_k.$ Noting that the number of the inserted digit $d$ is at most $\frac{n_{k+1}-m_{k}}{m_{k}-n_{k}}+1=o(h_k)$ in the block $\mathbb B'_k$, we readily check that $\mathbf K$ is a subset of $\mathbb N$ of density zero. Hence
by Lemma \ref{lem214} we have
$$\dim_{H}f\big(E(B)\big)=\dim_{H}E(B).$$
It remains to prove that $f\big(E(B)\big)$ is a Cantor subset of $E(\hat{\nu},\nu).$
\begin{lem}\label{lem44}
$f\big(E(B)\big)\subset E(\hat{\nu},\nu).$
\end{lem}
\begin{proof}
Fix $\bar{x}\in f\big(E(B)\big)$.
For $\varepsilon>0$ and $n$ large enough, there exists some $k$ such that $(\sum_{j=0}^{k-1}h_j)\leq n<(\sum_{j=0}^{k}h_j)$.
From the construction we deduce that:
if $n=(\sum_{j=0}^{k-1}h_j)+1$, then
$$|T^{n}(\bar{x})-y|<|I_{m_{k}-n_{k}}(y)|\le|I_{n}(y)|^{\nu-\varepsilon},$$
where the last inequality holds by the fact $\lim_k \frac{\sum_{j=0}^{k-1}h_j}{n_k}=1;$
if $(\sum_{j=0}^{k-1}h_j)< n<(\sum_{j=0}^{k}h_j),$ then
$$|T^{n}(\bar{x})-y|\ge \frac{1}{2(d+2)^{2}}|I_{m_{k}-n_{k}}(y)|\ge|I_n(y)|^{\nu+\varepsilon}.$$
We then prove that $\nu(\bar{x})=\nu$ by the arbitrariness of $\varepsilon.$
On the other hand, for $(\sum_{j=0}^{k-1}h_j)\leq N<(\sum_{j=0}^{k}h_j)$ with $k$ large enough, we pick $n=(\sum_{j=0}^{k-1}h_j)+1$ to obtain that
$$|T^{n}(\bar{x})-y|<|I_{m_{k}-n_{k}}(y)|\le|I_{\sum_{j=0}^{k}h_j}(y)|^{\hat{\nu}-\varepsilon}<|I_{N}(y)|^{\hat{\nu}-\varepsilon}.$$
When $N=(\sum_{j=0}^{k}h_j)$, for all $n\in [1,N],$ we have that
$$|T^{n}(\bar{x})-y|\ge\frac{1}{2(d+2)^{2}}|I_{m_{k}-n_{k}}(y)|\ge|I_{N}(y)|^{\hat{\nu}+\varepsilon}.$$
We prove that $\hat{\nu}(\bar{x})=\hat{\nu}.$
Hence $\bar{x}\in E(\hat{\nu},\nu),$ as desired.
\end{proof}
Consequently, for $0\le\hat{\nu}\le\frac{\nu}{1+\nu}<\nu<\infty,$ we have
$\dim_{H}E(\hat{\nu},\nu)\geq s(\mathcal{A}_{B},\xi,{\tau(i)})$.
Letting $B\rightarrow\infty$ yileds $\dim_{H}E(\hat{\nu},\nu)\geq s(\xi,{\tau(i)}).$
\bigskip
We conclude this section by determining the lower bound of $\dim_{H}E(\hat{\nu}, +\infty).$
We first study the case $0<\hat{\nu}<1$. Let
$$ n_{1}=2,~n_{k+1}=n_{k}^{k}+2n_{k},
~m_0=0,~m_{k}=\Big\lfloor\hat{\nu}n_{k}^{k}\Big\rfloor+n_{k},
~B_{k}=\lfloor m_{k}\log m_{k}\rfloor.$$
And thus
$$\lim_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k+1}}=\hat{\nu},
~~\lim_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k}}=\infty,
~~\lim_{k\to\infty}\frac{m_{k}-n_{k}}{m_{k}}=1.$$
Define
\begin{equation*}
E=\left\{x\in[0,1)\colon
a_{n}(x)\le B_k \text{ if } m_{k}<n\le n_{k+1} \text{ for some } k;
\ a_n(x)=i, \text{ otherwise}\right\}.
\end{equation*}
As before, for any $x=[a_1,a_2,\ldots]$ in $E,$
we construct an element $\bar{x}:=f(x)\colon$ insert a digit $B_k+1$ after positions $n_{k}$ and $m_{k}+i(m_{k}-n_{k}),$ $0\leq i\leq t_{k}$ in the continued fraction expansion of $x$, where
$t_{k}=\max \{t\in \mathbb{N}\colon m_{k}+t(m_{k}-n_{k})< n_{k+1}\}$; the resulted sequence is the continued fraction of $\bar x$.
The method establishing the the lower bound of $\dim_{H}E(B)$ applies to show that $\dim_{H}E\geq 1/2 .$
Moreover, $f(E)$ is a subset of $E(\hat{\nu}, +\infty).$ It remains to prove that the Hausdorff dimension of $f(E)$ coincide with the one of $E$. To this end, we shall show that $f^{-1}$ is a $(1-\varepsilon)$-H\"{o}lder mapping for any $\varepsilon>0$. We remark that Lemma \ref{lem214} may not apply directly here since $\{B_{k}\}$ is an unbounded sequence.
\begin{lem}
For $\varepsilon>0,$ $f^{-1}$ is a $(1-\varepsilon)$-H\"{o}lder mapping.
\end{lem}
\begin{proof}
We write
$$m_{k}'=m_{k}+\sum_{l=1}^{k-1}(t_{l}+2),~n_{k}'=n_{k}+\sum_{l=1}^{k-1}(t_{l}+2),$$
and define the marked set
$$\mathbf{K}=\big\{m_{k}'+i(m_{k}-n_{k})+1\colon 0\leq i \leq t_{k}, k\geq 1\big\}\cup\big\{ n_{k}'\colon k\geq 1\big\}.$$
Let $\Delta_{n}=\sharp\{i\leq n\colon i\in \mathbf{K}\}$, where $\sharp$ denotes the cardinality of a finite set. Let $k\in \mathbb{N}$ such that $m_{k}'\leq n<m_{k+1}'$. We have
\begin{align*}
\frac{\Delta_{n}\log B_{k}}{n}
&\leq\frac{\Big(\sum_{l=1}^{k-1}(t_{l}+2)+\frac{n-m_{k}'}{m_{k}-n_{k}}+1\Big)\log B_{k}}{n}\\
&\leq \frac{\Big(\sum_{l=1}^{k-1}(t_{l}+2)+1\Big)\log B_{k}}{m_{k}'}+\frac{\log B_{k}}{m_{k}-n_{k}}+\frac{\log B_{k}}{m_{k}}\to 0.
\end{align*}
So there exists $K\in \mathbb{N}$, such that for $k\geq K$ and $n\geq m_{K}',$
\begin{equation}\label{e48}
(B_{k}+2)^{2\Delta_{n}+4}<2^{(n-1)\varepsilon}.
\end{equation}
For $\overline{x_{1}}=f(x_{1})$ and $\overline{x_{2}}=f(x_{2})$ in $f(E),$ we assume without loss of generality that
\begin{equation*}
|\overline{x_{1}}-\overline{x_{2}}|<\frac{1}{2(B_{K}+2)^{2}q^{2}_{m_{K}'}(\overline{x_{1}})};
\end{equation*}
otherwise,
$|f^{-1}(\overline{x_{1}})-f^{-1}(\overline{x_{2}})|<C|\overline{x_{1}}-\overline{x_{2}}|^{1-\varepsilon}$ for some $C,$ as desired.
Let
$$n=\min\{j\geq 1\colon a_{j+1}(\overline{x_{1}})\neq a_{j+1}(\overline{x_{2}})\}.$$
By Lemma \ref{lem22}, we have $m_{k}'\leq n<m_{k+1}'$ for some $k\geq K$ and $n+1<n_{k+1}'.$
Assume that $\overline{x_{1}}>\overline{x_{2}}$ and $n$ is even (the same conclusion can be drawn for the remaining cases).
There exist $1\leq \tau_{n+1}(\overline{x_{1}})<\sigma_{n+1}(\overline{x_{2}}) \leq B_{k}+1$ such that
$\overline{x_{1}}\in I_{n+1}(a_{1},\ldots,a_{n},\tau_{n+1}(\overline{x_{1}}))$,
$\overline{x_{2}}\in I_{n+1}(a_{1},\ldots,a_{n},\sigma_{n+1}(\overline{x_{2}})).$
Combining Lemma \ref{lem23} and the construction yields that $\overline{x_{1}}-\overline{x_{2}}$ is greater than the length of basic interval $I_{n+2}(a_{1},\ldots,a_{n},\sigma_{n+1}(\overline{x_{2}}),B_{k}+1).$ This implies that
\begin{align*}
\overline{x_{1}}-\overline{x_{2}}\geq |I_{n+2}(a_{1},\ldots,a_{n},\sigma_{n+1}(\overline{x_{2}}),B_{k}+1)|
\geq \frac{1}{2(B_{k}+2)^{4}q^{2}_{n}}.
\end{align*}
Furthermore, noting that $f^{-1}(\overline{x_{1}}),$ $f^{-1}(\overline{x_{2}})\in I_{n-\Delta_{n}}(c_{1},\ldots,c_{n-\Delta_{n}}),$
where $(c_{1},\ldots,c_{n-\Delta_{n}})$ is obtained by eliminating all the terms $a_i$ with $i\in \mathbf{K}$ from $(a_1,\ldots,a_n)$, we conclude that
\begin{align*}
|f^{-1}(\overline{x_{1}})-f^{-1}(\overline{x_{2}})|&\leq |I_{n-\Delta_{n}}(a_{1},\ldots,a_{n-\Delta_{n}})|
\\&\leq \frac{1}{q^{2}_{n-\Delta_{n}}}
\leq(B_{k}+2)^{2\Delta_{n}}\frac{1}{q^{2}_{n}}
\leq2|\overline{x_{1}}-\overline{x_{2}}|^{1-\varepsilon},
\end{align*}
where the penultimate inequality follows by (\ref{e48}). This completes the proof.
\end{proof}
Now we deduce that $\dim_{H}E(\hat{\nu},\infty)\geq \frac{1-\varepsilon}{2}$ by Lemma \ref{lem212}.
Letting $\varepsilon\to 0$, we establish that $\dim_{H}E(\hat{\nu},\infty)\geq \frac{1}{2}$ when $0<\hat\nu<1$. A slight change in the proof actually shows that the estimate $\dim_{H}E(\hat{\nu},\infty)\geq \frac{1}{2}$ also works for $\hat\nu=0$ or $1$.
Indeed, when $\hat{\nu}=0,$ we may take
$$n_{k}=2^{2^{2k}},~ m_{k}=n_{k}^{2}, ~B_{k}=2^{n_{k}};$$
when $\hat{\nu}=1,$ we may take
$$m_{k}=(k+1)!,~ n_{1}=1,~n_{k+1}=m_{k}+\frac{m_{k}}{\log m_{k}}, ~B_{k}=\lfloor2^{\sqrt{m_{k}}}\rfloor.$$
\begin{rem}
Applying the similar arguments as in Sections \ref{S_3} and \ref{S_4}, we also prove for any $\hat{\nu}, \nu\ge 0$ that
$$\{x\in[0,1]\colon \hat{\nu}(x)\ge\hat{\nu},~\nu(x)=\nu\}$$
and $E(\hat{\nu},\nu)$ share the Hausdorff dimension.
\end{rem}
\section{Proofs of Theorems \ref{TZ1} and \ref{TZ3}}
In this section we study the Hausdorff dimensions of the following sets:
$$E(\hat{\nu})=\{x\in[0,1)\colon \hat{\nu}(x)=\hat{\nu}\},$$
$$\mathcal{U}(y,\hat{\nu})=\Big\{x\in[0,1)\colon \forall N\gg1, \exists~ n\in[1,N], \text{ such that } |T^{n}(x)-y|<|I_N(y)|^{\hat{\nu}}\Big\}.$$
A direct corollary of the definition is: if $\hat{\nu}_1>\hat{\nu}\ge 0$,
\begin{equation*}
E(\hat{\nu}_1)\subseteq \mathcal{U}(y,\hat{\nu})\subseteq \{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}.
\end{equation*}
Hence, the proofs of Theorems \ref{TZ1} and \ref{TZ3} will be divided into two parts: the upper bound of $\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}$ and the lower bound of $\dim_{H}E(\hat{\nu})$.
Lemma \ref{full} combined with the fact $E(0,0)\subset E(0)$ implies that the sets $E(0),$ $\mathcal{U}(y,0) $ and $\{x\in[0,1)\colon \hat{\nu}(x)\ge0\}$ are of full Lebesgue measure; we only need to deal with the case $\hat{\nu}>0.$
We start with the upper bound of $\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}.$
\begin{lem}
If $0<\hat{\nu}\le 1,$ we have $$\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}\le s\Big(\frac{4\hat{\nu}}{(1+\hat{\nu})^{2}},{\tau(i)}\Big).$$ If $\hat{\nu}>1,$ then $\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}=0.$
\end{lem}
\begin{proof}
For $\varepsilon>0$ small enough, we define
$$E_\varepsilon(\hat{\nu},\nu)=\Big\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu},~\nu\le\nu(x)\le\frac{\nu+\varepsilon}{1-\varepsilon}\Big\}.$$
Since $$\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}\subseteq\bigcup_{\nu\in \mathbb{Q}^{+}}E_\varepsilon(\hat{\nu},\nu),$$
where $\mathbb{Q}^{+}$ denotes the set of positive rational numbers, we have $$\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}\le\sup\big\{\dim_{H}E_\varepsilon(\hat{\nu},\nu)\colon \nu\in \mathbb{Q}^{+}\big\}.$$
If $\hat{\nu}>1$, the set $E_\varepsilon(\hat{\nu},\nu)$ is at most countable by Lemmas \ref{lem31} and \ref{lem32}, and $\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}=0.$
If $0<\hat{\nu}\le1,$ we obtain
$$\dim_{H}E_\varepsilon(\hat{\nu},\nu)\leq s\Big(\frac{\nu^{2}}{(\nu-\hat{\nu}+\hat{\nu}\varepsilon+\varepsilon)(1+\nu)},{\tau(i)}\Big)$$
in much the same way as the proof for the upper bound of $\dim_{H}E(\hat{\nu},\nu)$; we sketch the main differences$\colon$
In Lemma \ref{lem32}, $\limsup\frac{m_{k}-n_{k}}{m_{k}}$ is estimated by
\begin{equation*}
\frac{\nu}{1+\nu}\le\limsup\limits_{k\rightarrow\infty}\frac{m_{k}-n_{k}}{m_{k}}\leq\frac{\nu+\varepsilon}{1+\nu}.
\end{equation*}
The formulae (\ref{e36}) through (\ref{e37}) are replaced by
\begin{align*}
\sum_{i=1}^{k}(m_{i}-n_{i})
\ge m_{k}\Big(\frac{(\nu+\varepsilon)^{2}}{(\nu-\hat{\nu}+\hat{\nu}\varepsilon+\varepsilon)(1+\nu)}-\varepsilon\Big),
\end{align*}
$$1+\nu\le\limsup_{k\to\infty}\frac{m_k}{n_k}\le\frac{1+\nu}{1-\varepsilon},$$
$$\liminf_{k\to\infty}\frac{m_k}{n_{k+1}}\ge\frac{\hat{\nu}(1+\nu)}{\nu+\varepsilon},$$
$$\liminf_{k\to\infty}\frac{\sum_{i=1}^{k}(m_i-n_i)}{m_{k+1}}
\ge \frac{\hat{\nu}(\nu+\varepsilon)(1-\varepsilon)}{(\nu-\hat{\nu}+\hat{\nu}\varepsilon+\varepsilon)(1+\nu)},$$
resepctively. The set $\Omega$ is replaced by
\begin{align*}
\Big\{(\{n_{k}\},\{m_{k}\})\colon \liminf_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k+1}}\ge\hat{\nu},&~\nu\leq\limsup_{k\to\infty}
\frac{m_{k}-n_{k}}{n_{k}}\leq\frac{\nu+\varepsilon}{1-\varepsilon},\\&
n_k<m_k<n_{k+1} ~\text {for all} ~k\geq 1 \Big\}.
\end{align*}
Finally, since the function $\frac{\nu^{2}}{(\nu-\hat{\nu}+\hat{\nu}\varepsilon+\varepsilon)(1+\nu)}$ of $\nu$ attains its minimum at $\nu=\frac{2\hat{\nu}-2(\hat{\nu}+1)\varepsilon}{1-\hat{\nu}+(\hat{\nu}+1)\varepsilon}$, we have by Lemma \ref{lem210}(2) that
\begin{align*}
\dim_{H}\{x\in[0,1)\colon \hat{\nu}(x)\ge\hat{\nu}\}
&\leq \sup \left\{s\Big(\frac{\nu^{2}}{(\nu-\hat{\nu}+\hat{\nu}\varepsilon+\varepsilon)(1+\nu)},{\tau(i)}\Big)\colon \nu \in\mathbb{Q}^{+} \right\}\\
&\leq s\Big(\frac{4\big(\hat{\nu}-(\hat{\nu}+1)\varepsilon\big)}{\big(1+\hat{\nu}-(\hat{\nu}+1)\varepsilon\big)^{2}},{\tau(i)}\Big)
\rightarrow s\Big(\frac{4\hat{\nu}}{(1+\hat{\nu})^{2}},{\tau(i)}\Big)
\end{align*}
as $\varepsilon\to 0$.
\end{proof}
We now deal with the lower bound of the $\dim_{H}E(\hat{\nu})$ for $0<\hat{\nu}\le1.$
\begin{lem}
For $0<\hat{\nu}\le1,$ we have $\dim_{H}E(\hat{\nu})\ge s\Big(\frac{4\hat{\nu}}{(1+\hat{\nu})^{2}},{\tau(i)}\Big).$
\end{lem}
\begin{proof}
Noting that $E(\hat{\nu},\nu)$ is a subset of $E(\hat{\nu})$ for $\nu\ge\frac{\hat{\nu}}{1-\hat{\nu}}$ (or $\hat{\nu}\le\frac{\nu}{1+\nu}$), we have
$$\dim_{H}E(\hat{\nu})\geq \dim_{H}E(\hat{\nu},\nu)=s\Big(\frac{\nu^{2}}{(1+\nu)(\nu-\hat{\nu})},{\tau(i)}\Big).$$
Since the function $\frac{\nu^{2}}{(1+\nu)(\nu-\hat{\nu})}$ is continuous for $\nu\in [\frac{\hat{\nu}}{1-\hat{\nu}},\infty],$
and attains its minimum at $\nu=\frac{2\hat{\nu}}{1-\hat{\nu}},$ so by Lemma \ref{lem210}(2), we have
$$\dim_{H}E(\hat{\nu})\geq\dim_{H}E\Big(\hat{\nu},\frac{2\hat{\nu}}{1-\hat{\nu}}\Big)=s\Big(\frac{4\hat{\nu}}{(1+\hat{\nu})^{2}},{\tau(i)}\Big).$$
\end{proof}
\section{Proof of Theorem \ref{thm1}}
In this section, we prove Theorem \ref{thm1} by considering the upper and lower bounds of $\dim_{H}\big(F(\alpha)\cap G(\beta)\big)$ respectively. Recall that
\begin{equation*}
F(\alpha)=\left\{x\in[0,1)\colon \liminf_{n\to\infty}\frac{R_{n}(x)}{n}=\alpha\right\},
G(\beta)=\left\{x\in[0,1)\colon \limsup_{n \to\infty}\frac{R_{n}(x)}{n}=\beta\right\}.
\end{equation*}
The proof of Theorem \ref{thm1} goes along the lines as that of Theorem \ref{TZ2} with some minor modifications.
Noting that $\Big\{x\in [0,1)\colon \lim_{n\to\infty}\frac{R_{n}(x)}{\log_{\frac{\sqrt{5}+1}{2}}n}=\frac{1}{2}\Big\}\subset F(0)\cap G(0),$
we have $F(0)\cap G(0)$ is of full Lebesgue measure. Furthermore, since $F(\alpha)\cap G(1)\subset G(1)$ and $F(0)\cap G(\beta)\subset G(\beta)$, we have
$\dim_{H}F(\alpha)\cap G(1)\leq s(1,\tau(1))$ and $\dim_{H}\big(F(0)\cap G(\beta)\big)\le s(\beta,\tau(1))$.
We only need to consider the case $0<\alpha\le\beta<1.$
\subsection{Upper bound of $\dim_{H}\big(F(\alpha)\cap G(\beta)\big)$}
For $x=[a_{1}(x),a_{2}(x),\ldots]\in F(\alpha)\cap G(\beta)$ with non-periodic continued fraction expansion, we associate $x$ with two sequences $\{n_k\}$ and $\{m_k\}$ that satisfy the following properties$\colon$
{\rm{(1)}} $n_k<m_k<n_{k+1}<m_{k+1}$ for $k\ge1$;
{\rm{(2)}} $a_{n_k}(x)=\cdots=a_{m_k}(x)$ for $k\ge1;$
{\rm{(3)}} $\displaystyle\liminf_{k\to\infty}\frac{m_k-n_k}{n_{k+1}}=\frac{\alpha}{1-\alpha},$ $\displaystyle\limsup_{k\to\infty}\frac{m_k-n_k}{m_k}=\beta;$
{\rm{(4)}} The sequence $\{m_k\}$ grows exponentially;
{\rm{(5)}} Write $\xi=\frac{\beta^{2}(1-\alpha)}{\beta-\alpha}$. For any $\varepsilon>0$, there exist infinitely many $k$ such that
\begin{equation*}
\sum_{i=1}^{k}(m_{i}-n_{i}+1)\ge m_{k}(\xi-\varepsilon).
\end{equation*}
To this end, we define two ascending sequences $\{n_{k}'\}$ and $\{m_{k}'\}$ as follows$\colon$
$$n_{1}'=1,~ a_{n_{k}'}(x)= a_{n_{k}'+1}(x)= \cdots =a_{m_{k}'}(x)\neq a_{m_{k}'+1}(x),~ n_{k+1}'=m_{k}'+1.$$
Since $\beta=\limsup R_n(x)/n>0$, we have that $\limsup_{k\to\infty}(m_{k}'-n_{k}')= +\infty$, which enables us to pick a non-decreasing subsequence of $\{(n_k', m_k')\}_{k\geq 1}\colon$ put $(n_{1},m_{1})=(n'_{1},m'_{1});$
having choosen $(n_{k}, m_k)=(n_{j_{k}}', m_{j_{k}}')$ for $k\ge1$, we set
$j_{k+1}=\min\big\{j>j_k\colon m_{j}'-n_{j}'> m_{k}-n_{k}\big\},$
and put $(n_{k+1},m_{k+1})=(n_{j_{k+1}}', m_{j_{k+1}}')$.
\smallskip
We readily check the following properties$\colon$
\smallskip
\textbf{(a)} the sequence $\{m_{k}-n_{k}\}_{k\geq 1}$ is non-decreasing and $\lim_{k\to\infty} (m_{k}-n_{k})= +\infty.$
\smallskip
\textbf{(b)} If $m_{k}\leq n\leq n_{k+1}+(m_{k}-n_{k})$ for $k\geq 1$, then $R_{n}(x)=m_{k}-n_{k}+1$, and
\begin{equation*}
\frac{m_{k}-n_{k}+1}{n_{k+1}+(m_{k}-n_{k})}\leq \frac{R_{n}(x)}{n}\leq \frac{m_{k}-n_{k}+1}{m_{k}}.
\end{equation*}
\textbf{(c)} If $n_{k+1}+(m_{k}-n_{k})<n< m_{k+1}$ for $k\geq 1$, then $R_{n}(x)=n-n_{k+1}+1$, and
\begin{equation*}
\frac{m_{k}-n_{k}+1}{n_{k+1}+(m_{k}-n_{k})}\leq \frac{R_{n}(x)}{n}\leq \frac{m_{k+1}-n_{k+1}+1}{m_{k+1}}.
\end{equation*}
Properties \textbf{(a)} and \textbf{(b)} imply
\begin{equation}\label{e63}
\alpha=\liminf_{n\to\infty}\frac{R_{n}(x)}{n}
=\liminf_{k\rightarrow\infty}\frac{m_{k}-n_{k}+1}{n_{k+1}+(m_{k}-n_{k})}
\end{equation}
and
\begin{equation}\label{e64}
\beta=\limsup_{n\to\infty}\frac{R_{n}(x)}{n}
=\limsup_{k\rightarrow\infty}\frac{m_{k+1}-n_{k+1}+1}{m_{k+1}}.
\end{equation}
From (\ref{e63}) we obtain that
\begin{equation}\label{e65}
\liminf_{k\to\infty}\frac{R_{n_{k+1}}(x)}{n_{k+1}}=\liminf_{k\to\infty}\frac{m_{k}-n_{k}+1}{n_{k+1}}
=\frac{\alpha}{1-\alpha},
\end{equation}
which combined with (\ref{e64}) yields $\frac\alpha{1-\alpha}\le\beta$, or equivalently $\alpha\leq \frac{\beta}{1+\beta}$.
And thus the set $F(\alpha)\cap G(\beta)$ is at most countable when $\alpha>\frac{\beta}{1+\beta}$. Moreover, the equality (\ref{e65}) implies that $\{m_{k}\}_{k\geq 1}$ grows at least exponentially,
namely, there exists $C>0$, independent of $x,$ such that $k \leq C\log{m_{k}}$ for $k$ large enough. Further, by (\ref{e64}) and (\ref{e65}), we also have
\begin{equation*}
\liminf_{k\to\infty}\frac{m_{k}}{n_{k+1}}
\ge \liminf_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k+1}}
\cdot\liminf_{k\to\infty}\frac{m_{k}}{m_{k}-n_{k}}
=\frac{\alpha}{\beta(1-\alpha)},
\end{equation*}
which combined with the Stolz-Ces\`aro theorem implies that
\begin{equation*}
\begin{split}
\liminf_k\frac{\sum_{i=1}^{k-1}(m_i-n_i+1)}{m_{k}}
\ge \frac{\alpha\beta(1-\beta)}{\beta-\alpha},
\end{split}
\end{equation*}
and thus, for $\varepsilon>0$ and $k$ large enough,
\begin{equation*}
\sum_{i=1}^{k}(m_{i}-n_{i}+1)\geq \Big(\frac{\alpha\beta(1-\beta)}{\beta-\alpha}-\frac{\varepsilon}{2}\Big)m_{k}+(m_{k}-n_{k}+1).
\end{equation*}
Since there exist infinitely many $k$ such that
\begin{equation}\label{e610}
m_{k}-n_{k}+1\ge
m_{k}(\beta-\frac{\varepsilon}{2}),
\end{equation}
Property (5) holds for such $k$.
\medskip
\underline{Covering of $F(\alpha)\cap G(\beta)$.}\quad We collect all sequences $(\{n_{k}\}, \{m_{k}\})$ associated with $x\in F(\alpha)\cap G(\beta)$ as above to form a set
$$\Omega=\Big\{(\{n_{k}\},\{m_{k}\})\colon \text{Properties } (1) ~\& ~(3) \text{ are fulfilled}\Big\}.$$
For $(\{n_{k}\},\{m_{k}\})\in \Omega$ and $\{b_{k}\}\subset \mathbb N$, write
\smallskip
\begin{align*}
&H(\{n_{k}\},\{m_{k}\})=\big\{x\in[0,1)\colon \text{Property } (2) \text{ is fulfilled}\big\},\\&
\Lambda_{k, m_k}=\Big\{(n_1,m_1; \ldots; n_{k-1},m_{k-1};n_{k})\colon n_{1}<m_{1}<\cdots<m_{k-1}<n_{k}, (\ref{e610})\text{ holds} \Big\},\\ &
\mathcal{D}_{n_1,m_1;\ldots;n_{k}, m_{k}}(\{b_{k}\})=\big\{(\sigma_{1},\ldots,\sigma_{m_{k}})\in \mathbb{N}^{m_{k}}\colon \sigma_{n_{j}}=\cdots=\sigma_{m_{j}}=b_{j}\text{ for all }1\leq j\leq k \big\}.
\end{align*}
We obtain a covering of $ F(\alpha)\cap G(\beta)$:
\begin{equation*}
\begin{split}
&F(\alpha)\cap G(\beta) \subseteq \bigcup_{(\{n_{k}\},\{m_{k}\})\in \Omega}H(\{n_{k}\},\{m_{k}\}) \\
&\subseteq\bigcap_{K=1}^{\infty} \bigcup_{k=K}^{\infty}\bigcup_{m_{k}= e^{\frac{k}{C}}}^{\infty} \bigcup_{(n_{1},m_{1},\ldots,m_{k-1},n_{k})\in\Lambda_{k, m_k}}
\bigcup_{(b_{1},\ldots,b_{k})\in \mathbb{N}^{k}}
\bigcup_{(a_{1},\ldots,a_{m_{k}})\in \mathcal{D}_{n_1,m_1;\ldots;n_{k}, m_{k}}(\{b_{k}\})}I_{m_{k}}(a_{1},\ldots,a_{m_{k}}).
\end{split}
\end{equation*}
Writing $t=s(\xi-\varepsilon,{\tau(1)})+\frac{\varepsilon}{2}$, we estimate the $(t+\frac{\varepsilon}{2})$-dimensional Hausdorff measure of $F(\alpha)\cap G(\beta)$. Setting $\psi(m_{k})=m_{k}-\sum_{i=1}^{k}(m_{i}-n_{i}+1),$ $M=512\sum_{i=1}^{\infty}\big(\tau(i)\big)^{-2t}.$ For sufficiently large $k$, we first have the following estimate$\colon$
\begin{equation}\label{e616}
\begin{split}
& \sum\limits_{b_{k}=1}^{\infty}\cdots\sum\limits_{b_{1}=1}^{\infty}\sum\limits_{(a_{1},\ldots,a_{m_{k}})\in \mathcal{D}_{n_{k}, m_{k}}(\{b_{k}\})}|I_{m_{k}}(a_{1},\ldots,a_{m_{k}})|^{t+\frac{\varepsilon}{2}}\\
\le&\sum\limits_{b_{k}=1}^{\infty}\cdots\sum\limits_{b_{1}=1}^{\infty}\sum\limits_{a_{1},\ldots,a_{\psi(m_{k})}\in\mathbb{N}}4^{k(t+\frac{\varepsilon}{2})}\Big(\frac{1}{q_{\psi(m_{k})}(a_{1},\ldots,a_{\psi(m_{k})})}\Big)^{2t+\varepsilon}\prod_{j=1}^{k}\Big(\frac{1}{q_{m_j-n_j+1}(b_j,\ldots,b_j)}\Big)^{2t+\varepsilon}\\
\le&4^{k(t+\frac{\varepsilon}{2})}\sum\limits_{a_{1},\ldots,a_{\psi(m_{k})}\in\mathbb{N}}\Big(\frac{1}{q_{\psi(m_{k})}(a_{1},\ldots,a_{\psi(m_{k})})}\Big)^{2t+\varepsilon}\prod_{j=1}^{k}\left(\sum_{i=1}^{\infty}\Big(\frac{1}{q_{m_j-n_j+1}(i,\ldots,i)}\Big)^{2t+\varepsilon}\right)\\
\le&(4^{t+\frac{\varepsilon}{2}}M)^{k}\sum\limits_{a_{1},\ldots,a_{\psi(m_{k})}\in\mathbb{N}}\Big(\frac{1}{q_{\psi(m_{k})}(a_{1},\ldots,a_{\psi(m_{k})})}\Big)^{2t+\varepsilon}\prod_{j=1}^{k}\Big(\frac{1}{q_{m_j-n_j+1}(1,\ldots,1)}\Big)^{2t+\varepsilon}\\
\le&(16^{t+\frac{\varepsilon}{2}}M)^{k}\sum\limits_{a_{1},\ldots,a_{\psi(m_{k})}\in \mathbb{N}}\Big(\frac{1}{q_{m_{k}}(a_{1},\ldots,a_{\psi(m_{k})},1,\ldots,1)}\Big)^{2t+\varepsilon}\\
\le&(16^{t+\frac{\varepsilon}{2}}M)^{k}\sum \limits_{a_{1},\ldots,a_{m_{k}-\lfloor m_{k}(\xi-\varepsilon)\rfloor}\in \mathbb{N}}\Big(\frac{1}{q_{m_{k}}(a_{1},\ldots,a_{m_{k}-\lfloor m_{k}(\xi-\delta)\rfloor},1,\ldots,1)}\Big)^{2s_{m_{k}}(\xi-\varepsilon,{\tau(1)})+\varepsilon}\\
\le&(16^{t+\frac{\varepsilon}{2}}M)^{k}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{2}\varepsilon},
\end{split}
\end{equation}
where the third inequality holds since
\begin{equation*}
\begin{split}
\sum\limits_{i=1}^{\infty} \Big(\frac{1}{q_{n}(i,\ldots,i)}\Big)^{2t+\varepsilon}
&=\Big(\frac{1}{q_{n}(1,\ldots,1)}\Big)^{2t+\varepsilon} \sum\limits_{i=1}^{\infty}\Big(\frac{q_{n}(1,\ldots,1)}{q_{n}(i,\ldots,i)}\Big)^{2t+\varepsilon}\\
&\le\Big(\frac{1}{q_{n}(1,\ldots,1)}\Big)^{2t+\varepsilon} \sum\limits_{i=1}^{\infty}\Big(\frac{4\tau(1)}{\tau(i)}\Big)^{2t+\varepsilon};
\end{split}
\end{equation*}
the penultimate one follows from (\ref{e610}) and (\ref{e37}) and the last one is by Remark \ref{rem2} and Lemma \ref{lem21}(1).
Hence,
\begin{equation*}
\begin{split}
\mathcal{H}^{t+\frac{\varepsilon}{2}}\big(F(\alpha)\cap G(\beta)\big)
\leq &\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty}\sum_{m_{k}= e^{\frac{k}{C}}}^{\infty}
\sum_{(n_{1},m_{1},\ldots,m_{k-1},n_{k})\in\Lambda_{k, m_k}}\sum\limits_{b_{k}=1}^{\infty}\cdots
\sum\limits_{b_{1}=1}^{\infty}\\
&\times\sum\limits _{(a_{1},\ldots,a_{m_{k}})\in\mathcal{D}_{n_{k}, m_{k}}(\{b_{k}\})}|I_{m_{k}}(a_{1},\ldots,a_{m_{k}})|^{t+\frac{\varepsilon}{2}}\\
\overset{\text{(\ref{e616})}}{\leq}&\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty}\sum\limits_{m_{k}=e^{\frac{k}{C}}}^{\infty}
\sum_{n_{k}=1}^{m_{k}}\sum_{m_{k-1}=1}^{n_{k}}\cdots\sum\limits_{m_{1}=1}^{n_{2}}
\sum\limits_{n_{1}=1}^{m_{1}}(16^{t+\frac{\varepsilon}{2}}M)^{k}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{2}\varepsilon}\\
\leq&\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty} \sum\limits_{m_{k}=e^{\frac{k}{C}}}^{\infty}(16^{t+\frac{\varepsilon}{2}}Mm_{k})^{2C\log m_{k}}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{2}\varepsilon}\\ \leq&\liminf_{K\to\infty}\sum\limits_{k=K}^{\infty} \sum\limits_{m_{k}=e^{\frac{k}{C}}}^{\infty}\Big(\frac{1}{2}\Big)^{\frac{m_{k}-1}{4}\varepsilon}
\leq \frac{1}{1-(\frac{1}{2})^{\frac{\varepsilon}{4}}}\sum\limits_{k=1}^{\infty} \Big(\frac{1}{2^{\varepsilon}}\Big)^{\frac{e^{\frac{k}{C}}-1}{4}} < +\infty,
\end{split}
\end{equation*}
where the penultimate one holds since $(16^{t+\frac{\varepsilon}{2}}Mk)^{2C\log k}<2^{\frac{k-1}{4}\varepsilon}$ for $k$ large enough.
\subsection{Lower bound of $\dim_{H}\big(F(\alpha)\cap G(\beta)\big)$}
Note that $F(\alpha)\cap F(\beta)$ is at most countable for $\alpha>\frac{\beta}{1+\beta}$; we assume that $\alpha\le \frac{\beta}{1+\beta}.$
Let $\{n_{k}\}$ and $\{m_{k}\}$ be two strictly increasing sequences satisfying the following conditions$\colon$
{\rm{(1)}} $(m_{k}-n_{k})\leq(m_{k+1}-n_{k+1})$ and $n_{k}< m_{k}< n_{k+1}$ for $k\geq 1$;
{\rm{(2)}} $\lim_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k+1}}=\frac{\alpha}{1-\alpha}$;
{\rm{(3)}} $\lim_{k\to\infty}\frac{m_{k}-n_{k}}{m_{k}}=\beta$.
With the help of these sequences, we construct a Cantor subset of $F(\alpha)\cap G(\beta)$ to provide a lower bound estimate of its dimension. The set $E(B)$ is defined in much the same way as in Section \ref{S_{4.2}}, the only difference being that the digit $i$ is replaced by the digit $1$; the mapping $f$ is defined in exact the same way. It remains to verify that the set $f\big(E(B)\big)$ is a subset of $F(\alpha)\cap G(\beta).$
\begin{lem}
For any $B\ge 2,$ $f\big(E(B)\big)\subset F(\alpha)\cap G(\beta).$
\end{lem}
\begin{proof}
Recall the definitions of
$$t_{k}=\max \{t\in \mathbb{N}\colon m_{k}+t(m_{k}-n_{k})< n_{k+1}\},$$
$$m_{k}'=m_{k}+\sum_{l=1}^{k-1}(t_{l}+2),~n_{k}'=n_{k}+\sum_{l=1}^{k-1}(t_{l}+2).$$
We know that
$$\lim_{k\to\infty}\frac{\sum_{l=1}^{k}(t_{l}+2)}{n_{k+1}}=0,\quad \lim_{k\to\infty}\frac{n_k}{n_{k}'}=\lim_{k\to\infty}\frac{m_k}{m_k'}=1.$$
For $x\in f(E(B))$, and $m_{k}'\leq n<m_{k+1}'$ with $k\in \mathbb{N},$ we have that
\begin{equation*}
R_{n}(x)=\left\{\begin{array}{ll}
m_{k}-n_{k},& \ \text{if}~ m_{k}'\leq n\leq n_{k+1}'+m_{k}-n_{k}, \\
n-n_{k+1}', & \ \text{if}~n_{k+1}'+m_{k}-n_{k}< n< m_{k+1}' .\end{array}\right.
\end{equation*}
Observing that for $n_{k+1}'+m_{k}-n_{k}< n< m_{k+1}',$
\begin{equation*}
\frac{m_{k}-n_{k}}{n_{k+1}'+m_{k}-n_{k}}\leq\frac{R_{n}(x)}{n}=\frac{n-n_{k+1}'}{n}\leq\frac{m_{k+1}-n_{k+1}}{m_{k+1}'},
\end{equation*}
we deduce that
\begin{align*}
\liminf_{n\to\infty}\frac{R_{n}(x)}{n}
=\lim_{k\to\infty}\frac{R_{n_{k+1}'+m_{k}-n_{k}}(x)}{n_{k+1}'+m_{k}-n_{k}}
=\lim_{k\to\infty}\frac{m_{k}-n_{k}}{n_{k+1}+m_{k}-n_{k}}
=\alpha,
\end{align*}
and
\begin{align*}
\limsup_{n\to\infty}\frac{R_{n}(x)}{n}
= \lim_{k\to\infty}\frac{m_{k+1}-n_{k+1}}{m_{k+1}'}
= \lim_{k\to\infty}\frac{m_{k+1}-n_{k+1}}{m_{k+1}}
=\beta.
\end{align*}
Hence $x\in F(\alpha)\cap G(\beta)$.
\end{proof}
\section{Proof of Theorem \ref{thm2}}
The proof of Theorem \ref{thm2} will be divided into two parts according as $\alpha=0$ or $0<\alpha\le1.$
We first note that $F(\alpha)\cap G(\beta)$ is at most countable for $\alpha>\frac{1}{2}\ge\frac{\beta}{1+\beta}$ by (\ref{e64}) and (\ref{e65}).
Moreover, since $G(0)\subseteq F(0)$ and $G(0)$ is of full Lebesgue measure, we have $\dim_H F(0)=1$.
Hence, we only need to deal with the case $0<\alpha\le\frac{1}{2}.$
\textbf{Lower bound of $F(\alpha)$.} Since $F(\alpha)\cap G(\beta)\subseteq F(\alpha)$ for any $\beta\geq \frac{\alpha}{1-\alpha},$ we have
$$\dim_{H}F(\alpha)\geq \dim_{H}F(\alpha)\cap G(\beta)=s\Big(\frac{\beta^{2}(1-\alpha)}{\beta-\alpha},{\tau(1)}\Big).$$
The function $\frac{\beta^{2}(1-\alpha)}{\beta-\alpha}$ is continuous for $\beta\in [\frac{\alpha}{1-\alpha},1],$
and attains its minimum at the point $\beta=2\alpha\ge \frac{\alpha}{1-\alpha},$ so by Lemma \ref{lem210}(2), we obtain
$$\dim_{H}F(\alpha)\geq\dim_{H}\big(F(\alpha)\cap G(2\alpha)\big)=s\big(4\alpha(1-\alpha),{\tau(1)}\big).$$
\textbf{Upper bound of $F(\alpha)$.} For $x\in F(\alpha),$ there exists $\beta_{0}\in(0,1]$ such that
$$\liminf_{n\to\infty}\frac{R_{n}(x)}{n}=\alpha,~\limsup_{n\to\infty}\frac{R_{n}(x)}{n}=\beta_{0}.$$
Then $0<\alpha\leq\frac{\beta_{0}}{1+\beta_{0}}<\beta_{0}< 1$ or $\beta_{0}=1$.
If $0<\alpha\leq\frac{\beta_{0}}{1+\beta_{0}}<\beta_{0}< 1$, then for $0<\varepsilon<\frac{\alpha(1-2\alpha)}{2(1-\alpha)},$
there exists $\beta\in \mathbb{Q}^{+}$ such that $0<\alpha\le\frac{\beta_0}{1+\beta_0}\le\beta\leq\beta_{0}\leq \beta+\varepsilon< 1.$
Let
$$E_{\alpha,\beta,\epsilon}=\Big\{x\in [0,1)\colon \liminf_{n\to\infty}\frac{R_{n}(x)}{n}=\alpha,
~\beta \leq \limsup_{n\to\infty}\frac{R_{n}(x)}{n} \leq \beta + \varepsilon < 1\Big\},$$
we have
$$F(\alpha)\subseteq\Big(\bigcup_{\beta\in \mathbb{Q}^{+}}E_{\alpha,\beta,\varepsilon}\Big)\cup \big(F(\alpha)\cap G(1)\big).$$
So $\dim_{H}F(\alpha)\leq \max\Big\{\frac{1}{2}, \sup\big\{\dim_{H}E_{\alpha,\beta,\varepsilon}\colon \beta\in \mathbb{Q}^{+}\big\}\Big\}.$
It remains to estimate the upper bound of $\dim_{H}E_{\alpha,\beta,\varepsilon}.$
Following the same line as the proof for the upper bound of $\dim_{H}\big(F(\alpha)\cap G(\beta)\big),$ we obtain that
$$\dim_{H}E_{\alpha,\beta,\varepsilon}\leq s\Big(\frac{\beta^{2}(1-\alpha)}{\beta-\alpha+\varepsilon},{\tau(1)}\Big).$$
Thus, since the function
$\frac{\beta^{2}(1-\alpha)}{\beta-\alpha+\epsilon}$
with respect to $\beta$ attains its minimum at $\beta=2(\alpha-\varepsilon)$, we have that
\begin{align*}
\dim_{H}F(\alpha)&\leq \sup \left\{s\Big(\frac{\beta^{2}(1-\alpha)}{\beta-\alpha+\varepsilon}\Big)\colon\beta \in \mathbb{Q}^{+} \right\}
\leq s\big(4(\alpha-\varepsilon)(1-\alpha),{\tau(1)}\big)\\&
\rightarrow s\big(4\alpha(1-\alpha),{\tau(1)}\big)\quad \text{as }\varepsilon\to 0.
\end{align*}
\medskip
{\noindent \bf Acknowledgements}. This work is supported by NSFC No. 12171172, 12201476. The authors would like to express their gratitude to Professors Bao-Wei Wang and Jian Xu for helpful discussions during the preparation of the paper.
|
\section{Introduction}
The distribution of star clusters' properties (mass, age, physical size) and the variation of this distribution with galactic environment provides crucial clues to the physics of star formation and star cluster evolution. Moreover, because of the processes responsible for setting this distribution depend on galactic environment, cluster demographics can also trace the history of galaxy assembly and evolution, acting as a fossil record of the environments that existed when clusters formed. Given the importance of cluster demographics, it is not surprising that there have been many attempts to measure them, and that cluster demographics figure prominently in the science cases for a number of large surveys of nearby galaxies, such as PHAT \citep{2012PHAT}, PHANGS-HST \citep{2022PHANGS}, and LEGUS \citep{2015LEGUS}; see \citet{2019ARAA} for a detailed review.
Different aspects of cluster demographics probe different physics.
By studying the shape of the cluster mass function (CMF), and in particular the initial CMF (ICMF) that applies to the youngest clusters, we can constrain star formation theories and provide an observational check on simulations.
Studies to date show that over much of its range the ICMF is well-described by a powerlaw $dN/dM\propto M^{\alpha_M}$ with a slope $\alpha_M \approx -2$ across a wide range of galaxy properties, corresponding to equal mass per logarithmic bin and suggesting a scale-free formation process \citep[e.g.,][]{2012Fall_MID,2014Chandar_MID}.
However, the shape of the ICMF at its high-mass end ($\gtrapprox 10^4 M_{\odot}$) remains uncertain. Some authors report that a truncated distribution such as a Schechter function describes the data better than the pure power-law \citep[e.g.,][]{2012Bastian_Schetcher, LEGUS2017, 2017PHAT}, while others question the statistical significance of claimed detections and instead suggest that the dearth of massive clusters is simply a size-of-sample effect \citep[e.g.,][]{Larsen02a, 2020Mok}. If there is a real truncation in the mass function, its location must depend somehow on the galactic environment, since rapidly star-forming galaxies with large cluster populations often harbour clusters with masses larger than the reported truncation masses in more modestly star-forming galaxies \citep[e.g.,][]{2015MID, 2017Linden_inclusive}. The presence of a non-existence of a high-mass break in the ICMF, its value, and its variation with galactic environment provide an important clue to the star formation process.
As clusters evolve and disperse from their birthplaces, they experience mass loss. Various theories govern how they disrupt, ranging from the ``infant mortality'' stage of rapid gas removal after star formation \citep{1980Hills,2007gas_expulsion,2020gas_expulsion} to long-timescale processes such as two-body relaxation \citep{2001twobody,2010two_body,2012two_body} to processes such as external tidal shocking that operate on intermediate timescales \citep{2005Lamers_tidal,2005Bastian_tidal,2006GMC_encounter,2016tidal,2017GMC_encounter,2019tidal,2019Webb_tidal}.
In order to place observational constraints on these theories, we study
the cluster age function (CAF).
As with the ICMF, observations suggest that the CAF is reasonably well-described by a powerlaw form $dN/dT\propto T^{\alpha_T}$, where $T$ is cluster age. However, the value of $\alpha_T$ remains highly controversial, at least in part due to disagreement among observational groups about what constitutes a cluster \citep{2019ARAA}.
Despite
this uncertainty, the CAF encodes the timescales of diverse cluster disruptive processes.
One question of particular interest is the form of
the joint mass-age distribution, which gives a complete description of cluster formation and disruption.
If the mass and age distributions are separable, i.e.~if $d^2 N/dM\,dT \propto (dN/dM)(dN/dT)$, this implies
mass-independent disruption (MID) of clusters \citep{2005MID,2009MID,2015MID}, while if they are not this implies mass-dependent disruption (MDD; \citealt{2005MDD,2007MDD,2021MDD}).
Knowing whether MID or MDD holds in a particular galaxy or sub-galactic region would in turn place strong constraints on the mechanisms by which clusters disrupt.
Part of the reason that observations have not yet resolved questions such as the existence and value of a cutoff in the ICMF, the slope of the CAF, and whether cluster disruption is mass-dependent or -independent, is that determining the properties of clusters from observations is not trivial. In order to sample a wide range of environments,
studies of cluster populations must work with integrated light rather than resolved stellar populations, since in the crowded environments of star clusters it is generally only possible to resolve individual stellar sources for a handful of the most nearby galaxies.
The traditional approach to extracting cluster demographics from this type of data is to convert the integrated light measurements for each cluster to physical properties such as mass and age by comparing the observed photometry to a grid of simple stellar population (SSP) models and generating a set of best-fitting parameters. One then bins the clusters by mass and age to obtain mass and age distributions.
However, this approach encounters several difficulties. First, the binning process usually entails the loss of useful information, and as a result parameters determined by fitting data to histograms are heavily biased \citep{2009Binning,2017Binning}. Second, the conventional approach of using $\chi^2$ fitting to convert photometry to masses and ages implicitly assumes that the uncertainties on cluster mass and age can be approximated as Gaussian. This is often a poor assumption, because the mapping between physical properties and photometry is non-linear and non-monotonic, particularly once one adds the additional complication of dust extinction. Consequently, a particular set of photometric measurements may plausibly fit two (or more) widely-separated loci in physical parameter space, yielding posterior mass and age distributions with complex non-symmetrical and multi-modal shapes \citep{2014deMeu_MTshape,2014MT_shape,2015Powerlaw,2019MT_shape}.
The problem is exacerbated for low-mass clusters, where stochastic sampling of the IMF leads to a wide range of possible photometry even for clusters of fixed mass and age \citep{2009Piskunov}.
However, the most severe issue for measuring cluster demographics is completeness. The problem is that star clusters become fainter as one moves to both lower mass and older age, so a magnitude limit corresponds to a complex shape in the parameter space of age and mass, meaning that both the mass and age distributions are subject to large potential biases. The conventional way of handling this is to discard clusters outside of a limited range of both mass and age over which completeness is expected to be $\gtrsim 90 \% $, e.g., to retain only clusters with estimated masses $5,000$~M$_\odot$ and estimated ages $<200$~Myr. Such a drastic truncation of the sample avoids bias, but at the cost of a large loss of statistical power at both lower masses and old ages. The former is particularly concerning, since the steep mass function means that low-mass clusters form the majority of the available sample; the sample truncation required to avoid bias from survey magnitude limits therefore often involves discarding the majority of the data.
The main objective of this paper is to present a complete analysis pipeline and preliminary results for cluster demographics in the benchmark galaxy NGC 628 using the catalogue of clusters measured by the \emph{Hubble Space Telescope} Treasury Program Legacy Extragalactic Ultraviolet Survey \citep[LEGUS;][]{2015LEGUS} coupled to the novel Bayesian forward modelling method proposed by \citet{SLUG2019} that overcomes the problems identified above. Specifically, this method naturally copes with complex and non-deterministic mappings between photometric measurements and physical properties, and it enables us to use a total of 1178\ clusters catalogued by LEGUS with non-zero completeness values, as compared to earlier modelling where severe completeness cuts reduced the sample to $\approx 300$ \citep{LEGUS2017}.
This paper is structured as follows: in \autoref{cha:data}, we introduce our target galaxy, cluster catalogue, and analysis of completeness. In \autoref{cha:methods}, we summarise our analysis method. We present our results for cluster demographics in \autoref{cha:results}, and discuss their implications in \autoref{cha:discussions}. We summarise the findings and discuss future prospects in \autoref{cha:results}.
\section{Data Description}
\label{cha:data}
Here we summarise the properties of our target galaxy (\autoref{ssec:ngc628}), the star cluster catalogue we use as the basis for our study (\autoref{ssec:classification}), and our treatment of survey completeness (\autoref{ssec:completeness}).
\subsection{NGC 628}
\label{ssec:ngc628}
Our target NGC 628, also known as Messier 74, is a well-studied grand-design spiral galaxy, located at a distance of 9.9 Mpc \citep{2010Olivares}. We choose NGC 628 for this study as its large size (apparent radius of 5.2'; \citealt{2014Gusev}) and face-on orientation allow for a detailed study of the stellar cluster population. We use observations of NGC 268 taken from the LEGUS survey, a Cycle 21 HST Treasury programme that targeted 50 local galaxies ($\lesssim$15~Mpc) with the Hubble Space Telescope with broad band filter coverage from the UV to the near IR. All targets were imaged with either the Wide Field Camera 3 (WFC3) or the Advanced Camera for Surveys (ACS) in the NUV (WFC3 F275W), U-band (F336W), B-band (ACS/F435W or WFC3/F438W), V-band (ACS or WFC3/F555W or F606W), and I-band (ACS or WFC3 F814W). For the remainder of this paper we follow the conventional Johnson passband naming convention UV, U, B, V and I without converting to the Johnson system.
As part of the LEGUS survey, NGC~628 is covered by two pointings; one at the galactic centre (NGC 628c) and the other at the eastern edge (NGC 628e). We show the LEGUS V-band image of NGC~628 in \autoref{fig:NGC628_cl}. Detailed descriptions of the survey and data reduction are provided in \citet{CalzettiLEGUS2015}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/NGC628_cl_mosaic.pdf}
\caption{V-band (F555W) mosaic image of the centre and east fields of NGC 628 from LEGUS-HST. Symbols show the location of the star clusters, colored by the cluster classifications (\autoref{ssec:classification}) : class 1 (red circle), class 2 (blue circle), and class 3 (blue circle). For detailed description of cluster classification method, Section 2.2.}
\label{fig:NGC628_cl}
\end{figure}
\subsection{Cluster Catalogue}
\label{ssec:classification}
The photometric star cluster catalogue for NGC~628 that we use in this study comes from the LEGUS survey, and we refer readers to \citet{Grasha2015} and \citet{LEGUS2017} for a complete description of how it is constructed. Here, we provide a brief description of the parts of the process most pertinent to this study. LEGUS uses a two-step pipeline for cluster classification. First, an automated cluster extraction tool is used to extract potential cluster candidates with at least a 3$\sigma$ detection in a minimum of 5 contiguous pixels from the white light images. The cluster candidates are then further refined and required to have a V-band concentration index (CI; the difference in magnitude between a radius of 1 and 3 pixels) of $>1.4$ mag for the centre pointing and $>1.3$ mag for the east pointing. The CI cut serves to separate point-like sources (i.e., stars) from more extended sources (i.e., star clusters). The star cluster candidates are required to be detected in the V-band and a minimum of at least three of the other filters, with photometric error $\sigma_\lambda\leq 0.3$ mag in each band.
To create the final cluster catalog, the LEGUS team then visually inspects all candidates brighter than $-$6 in the V-band and classifies them into four morphological categories: (1) centrally concentrated clusters with spherically symmetric profiles; (2) clusters with asymmetric radial profiles; (3) multi-peaked clusters with underlying diffuse emission; and (4) non-cluster contaminants, such as background galaxies, stars, bad pixels, or edge artefacts. In this work, we limit our analysis to Class 1 (426), 2 (437), and 3 (413) clusters, for a total sample of 1276; we reduce this to 1275 by removing one cluster that we found to be a duplicate entry in the LEGUS NGC 628c and 628e catalogues, located in the small region where the fields overlap. \autoref{fig:NGC628_cl} shows the mosaic V-band image of NGC~628 overlaid with the cluster positions, coloured by their morphological classification.
\subsection{Observational Completeness Limits}
\label{ssec:completeness}
Because we intend to forward model the cluster population, we require knowledge of the completeness of the observational catalogue. Specifically, we require knowledge of the function $P_\mathrm{obs}(\mathbf{m})$, which describes the probability that a hypothetical star cluster with a vector of magnitudes $\mathbf{m}$ in the various LEGUS filters would be included in the cluster catalogue. To compute this function, we employ the completeness limits reported by \cite{LEGUS2017}, who carry out artificial cluster tests to derive the completeness of the LEGUS automated catalogue generation procedure in each filter independently. For each filter $F$ and a range of magnitudes $m_F$ in that filter, they determine the probability $P_{\mathrm{obs},F}(m_F)$ that a cluster would be recovered by the automated extraction procedure. In what follows we linearly interpolate these tabulated data to obtain a continuous function $P_{\mathrm{obs},F}(m_F)$ that gives the probability that a cluster of arbitrary magnitude $m_F$ in filter $F$ will be recovered.
To determine the completeness for a cluster with a vector of magnitudes $\mathbf{m}$, we use 10,000 Monte Carlo trials. In each trial, we randomly assign each filter $F$ to be a detection or a non-detection with probability $P_{\mathrm{obs},F}(m_F)$ as determined from the interpolated artificial cluster test results. We then determine from this set of detections and non-detections if the cluster would be catalogued following the same criteria used in construction of the actual LEGUS catalogue, i.e., the cluster is catalogued only if it is (1) detected in V-band and at least three other bands, and (2) has a visual magnitude $M_\mathrm{V} \leq -6$. We then take $P_\mathrm{obs}(\mathbf{m})$ to be equal to the fraction of the Monte Carlo trials in which the cluster is catalogued.
We use this method to calculate the completeness both of our library of synthetic clusters (see \autoref{cha:methods}) and of the actual LEGUS catalogue. For the latter, we find a total of 97 clusters for which we estimate $P_{\mathrm{obs},F}(\mathbf{m}) = 0$; these can be present in the catalogue because a small number of clusters with $m_\mathrm{V} > -6$ were added by hand. To avoid introducing errors in our completeness estimate we remove these from the sample. We remove a total of 97 clusters from the sample of NGC~628 due to their zero completeness values, resulting in a final sample size of 1178\ clusters.
\section{Methods}
\label{cha:methods}
In this section, we describe the pipeline used to derive the cluster demographics, motivated by the forward Bayesian modeling approach demonstrated in \citet{SLUG2019}. For reasons of brevity we only summarise the method, and refer readers to \citeauthor{SLUG2019} for details.
\subsection{Overview of the method}
Given a set of unresolved photometric measurements of a cluster population, how do we infer their underlying demographics? To answer this question, we first propose a joint distribution of mass $M$, age $T$, and extinction $A_\mathrm{V}$ for the cluster population, which we denote as $f (M, T, A_\mathrm{V}\mid \boldsymbol{\theta})$, where $\boldsymbol{\theta}$ represents a vector of parameters describing the joint distribution $f$. For example, if we assume the mass distribution of clusters is described by a Schechter function $f(M, T, A_\mathrm{V}\mid \boldsymbol{\theta}) \propto M^{\alpha_M} \exp(-M/M_\mathrm{break})$, then $\boldsymbol{\theta}$ contains the slope $\alpha_M$ and break mass $M_\mathrm{break}$ of the Schechter function. We describe the functional forms we consider for $f(M, T, A_\mathrm{V}\mid\boldsymbol{\theta})$ and the parameters they involve in \autoref{ssec:models and priors}.
We seek to derive the posterior distribution of $\boldsymbol{\theta}$. We do so in the usual way for a Bayesian method, but writing the posterior as the product of a prior and a likelihood function
\begin{equation}
\label{eq:3.1}
p_\mathrm{post}(\boldsymbol{\theta}\mid \{\mathbf{m}\}) \propto \mathcal{L}(\{\mathbf{m}\}\mid\boldsymbol{\theta}) p_\mathrm{prior}(\boldsymbol{\theta}).
\end{equation}
where $\mathcal{L}(\{\mathbf{m}\} \mid \boldsymbol{\theta})$ is the likelihood for our set of $N_\mathrm{obs}$ photometric measurements. We defer a discussion of priors to \autoref{ssec:models and priors}, and present details of our method of calculating the likelihood function in \autoref{ssec:likelihood}, but to summarise the latter here, we compute the likelihood function using a Gaussian mixture model derived from a large library of synthetic star clusters; as we change $\boldsymbol{\theta}$, we adjust the weights applied to this library, which changes the value of $\mathcal{L}(\{\mathbf{m}\}\mid\boldsymbol{\theta})$. Our method can therefore be summarised into the following steps.
\begin{enumerate}
\item We generate a library composed of synthetic clusters with weights based on a proposed distribution of cluster physical properties and the observational completeness of the survey.
\item Using the newly created library, we produce a synthetic distribution in photometric space, which will be used to compare with the observations.
\item We adjust the model parameters $\boldsymbol{\theta}$ to maximise the resemblance between the observations and the photometric distribution, as parameterised by the likelihood function. As we do so, we map out the posterior distribution.
\end{enumerate}
We carry out the final step of this procedure using a Markov Chain Monte Carlo (MCMC) method as implemented in the software package \textsc{emcee} \citep{2013emcee}. For all the calculations presented in this paper we use 4000 iterations of 100 walkers, discarding the first 300 iterations for burn-in; both visual inspection of the posteriors and quantitative evaluation of the auto-correlation time indicates that the chains are well-converged. We describe the first two steps of the method -- generating and re-weighting the library, in \autoref{ssec:likelihood}.
\subsection{Calculation of the likelihood function}
\label{ssec:likelihood}
As discussed above, we compute the likelihood function using the Gaussian mixture model described in \citet{SLUG2019}, which operates on a library of $N_\mathrm{lib}$ synthetic clusters, each of which is characterised by a mass $M$, age $T$, extinction $A_\mathrm{V}$, and a vector of photometric magnitudes $\mathbf{m}$ in each of the filters used in the observations. The full library is further described by a vector of photometric bandwidths $\mathbf{h}$, which we set to 0.1 mag in all filters; see \citet{2015Powerlaw} and \citet{SLUG2019} for detailed discussion of the meaning of the bandwidth and the motivation for choosing this value. We provide details regarding the library we use in this work in \aref{app:library}.
Given the library, \citet{SLUG2019} show that the likelihood function for the parameters $\boldsymbol{\theta}$ can be written as (their equation 12)
\begin{equation}
\label{eq:likelihood}
\mathcal{L}(\{\mathbf{m}\}\mid\boldsymbol{\theta}) \propto \prod_{i=1}^{N_\mathrm{obs}} \left[\mathcal{A}(\boldsymbol{\theta})\sum_{j=1}^{N_\mathrm{lib}} w_j(\boldsymbol{\theta}) \mathcal{N}(\mathbf{m}_i \mid \mathbf{m}_j, \mathbf{h}'_i)\right],
\end{equation}
where $w_j(\boldsymbol{\theta})$ is the statistical weight of the $j$th library cluster (which depends on the parameters $\boldsymbol{\theta}$ as described below), $\mathbf{m}_i$ and $\mathbf{m}_j$ are the vectors of magnitudes for the $i$th observed and $j$th library clusters, respectively, $\mathcal{A}(\boldsymbol{\theta}) = [\sum_{j=1}^{N_\mathrm{lib}} w_j(\boldsymbol{\theta})]^{-1}$ is a normalisation factor, and $\mathcal{N}(\mathbf{x}\mid\mathbf{x}_0, \boldsymbol{\sigma})$ is the standard multidimensional Gaussian distribution centred at $\mathbf{x}_0$ and with width $\boldsymbol{\sigma}$, evaluated at position $\mathbf{x}$. The quantity $\mathbf{h}_i'$ is given by $\mathbf{h}'_i = (\mathbf{h}^2 + \boldsymbol{\sigma}_i^2)^{1/2}$, where $\boldsymbol{\sigma}_i$ the observational error on the magnitude of cluster $i$; the quantities $\mathbf{h}'_i$, $\mathbf{h}$, and $\boldsymbol{\sigma}_i$ are vectors with one element per filter, and the expression for $\mathbf{h}'_i$ should be understood as applying separately to each filter. In practice we evaluate \autoref{eq:likelihood} numerically using the \textsc{cluster\_slug} module of the \textsc{slug} software suite, which implements a tree-based order $N_\mathrm{obs}\ln N_\mathrm{lib}$ algorithm for performing the calculation that is much faster than a naive brute force evaluation, which would have a computational cost of order $N_\mathrm{obs} N_\mathrm{lib}$
The quantity in square brackets in \autoref{eq:likelihood} is the distribution of photometric magnitudes for the cluster library evaluated at the vector of magnitudes $\mathbf{m}_i$ for the $i$th observed cluster. This distribution, and thus the likelihood function as whole, depends on the parameters describing the cluster population $\boldsymbol{\theta}$ only through the weight functions $w_j(\boldsymbol{\theta})$ that describe the statistical weight of each library cluster. The relationship between weights and $\boldsymbol{\theta}$ is given by (equation 9 of \citealt{SLUG2019})
\begin{equation}
w_j(\boldsymbol{\theta}) = P_\mathrm{obs}(\mathbf{m}_j) \frac{f(M_j, T_j, A_{\mathrm{V},j} \mid \boldsymbol{\theta})}{p_\mathrm{lib}(M_j, T_j, A_{\mathrm{V},j})}.
\label{eq:weights}
\end{equation}
The denominator $p_{\mathrm{lib}(M_j, T_j, A_{\mathrm{V}, j})}$ is the distribution function describing the sampling density of the library (see \aref{app:library}), while the numerator $f(M_j, T_j, A_{\mathrm{V}, j} \mid \boldsymbol{\theta})$ is distribution of the physical properties given $\boldsymbol{\theta}$, and the pre-factor $P_\mathrm{obs}(\mathbf{m}_j)$ is the probability that a cluster with vector of magnitude $\mathrm{m}_j$ would be included in the LEGUS catalogue. The terms in \autoref{eq:weights} can be intuitively understood as follows. The factor $P_{\mathrm{obs}}(\mathbf{m}_j)$ accounts for the fact that only a fraction of clusters with magnitudes $\mathbf{m}_j$ will be observed due to the completeness limits; we compute this probability as described in \autoref{ssec:completeness}. The denominator is the probability density for drawing clusters with a particular combination of physical parameters $(M_j, T_j, A_{\mathrm{V}, j})$ while constructing the library. Finally, the numerator represents the true probability density for a cluster population described by the parameter set $\boldsymbol{\theta}$, so that the ratio $f(M_j, T_j, A_{\mathrm{V},j} \mid \boldsymbol{\theta}) / p_{\mathrm{lib}}(M_j, T_j, A_{\mathrm{V},j})$ represents the factor by which we must up- or down-weight the library so that clusters in the library have the same mass, age, and extinction distribution as clusters in reality; if this weight factor is unity, then our library is sampled from the same distribution of cluster properties as the real population.
Intuitively, then, our method consists of iteratively adjusting the parameters $\boldsymbol{\theta}$ and thus the weights $w_j(\boldsymbol{\theta})$ so as to bring the predicted photometric distribution into as close agreement as possible with the observed one. This will in turn adjust the mass and age distributions, since these are determined by the same vector of parameters $\boldsymbol{\theta}$ as the luminosity distribution. Again, we remind readers that this is just an intuitive description of the underlying process; formal proof that \autoref{eq:likelihood} is the correct likelihood function to accomplish this adjustment, along with some details of how we handle complications like clusters where some filters are missing due to the fields of view in the different filters not being perfectly overlapping, is provided in \citet{SLUG2019}.
\subsection{Demographic models and priors}
\label{ssec:models and priors}
Cluster demographics depend on cluster formation and destruction mechanisms, and thus the set of candidate parametric model distributions $f(M,T,A_\mathrm{V}\mid\boldsymbol{\theta})$ we consider is necessarily informed by theoretical expectation. In this work, we consider the two most prominent models: mass-independent destruction (MID; \autoref{sssec:mid}; \citealt{2005MID,2009MID, 2015MID}) and mass-dependent destruction/disruption (MDD; \autoref{sssec:mdd}; \citealt{2005MDD, 2007MDD, 2021MDD}). We couple both of these to a parametric model for the distribution of extinctions (\autoref{sssec:av_dist}).
\subsubsection{The mass-independent disruption (MID) model}
\label{sssec:mid}
In the MID model, the rate at which star clusters are destroyed is independent of cluster mass. The cluster mass function therefore has the same shape at all ages, and the distribution function $f(M,T,A_\mathrm{V}\mid\boldsymbol{\theta})$ can be separated into two distinct functions, one describing the mass distribution $p_M(M)$ and one the age distribution $p_T(T)$. Given the observational evidence that the mass function is a (possibly) truncated powerlaw, we will adopt a functional form for $p_M(M)$ given by a Schechter function,
\begin{equation}
p_M\left(M\right) \propto M^{\alpha_{M}} \exp \left(-\frac{M}{M_{\mathrm{break}}}\right).
\label{eq:schechter}
\end{equation}
For the age distribution, photometry cannot differentiate between a bound an unbound cluster, so the age distribution is required to be flat for times that are shorter than the physical time required for the stars in a cluster to disperse; of course the distribution can also be flat out to older ages if the mechanisms responsible for disruption do not begin until some time after a cluster forms. Regardless of its physical, origin, we call the time at which cluster disruption begins $T_{\mathrm{MID}}$. After this time, clusters will disrupt, and we approximate the age distribution as a powerlaw. We therefore have
\begin{equation}
p_T(T) \propto \begin{cases}1, & \text{ if }T<T_{\text {MID }} \\ \left(\frac{T}{T_{\text {MID }}}\right)^{\alpha_{T}}, &\text{ if } T>T_{\text {MID }}\end{cases}.
\end{equation}
Thus the joint mass-age distribution in the MID model is
\begin{equation}
\frac{\mathrm{d}^{2} N}{\mathrm{~d} M \mathrm{~d} T} \propto M^{\alpha_{M}} \exp \left(-\frac{M}{M_{\text {break }}}\right) \max \left(T, T_{\text {MID }}\right)^{\alpha_{T}}.
\label{eq:mass_age_mid}
\end{equation}
The MID model therefore has 4 free physical parameters that we place in our vector $\boldsymbol{\theta}$: $\alpha_{M}$, $\log M_{\mathrm{break}}$, $\alpha_{T}$, and $\log T_{\mathrm{MID}}$. We adopt flat priors on $\alpha_M$ from $-4$ to 0 and $\alpha_T$ from $-3$ to 0, reflecting a broad range around previous literature values; we will see that these choices have little effect, as our MCMC never approaches these boundaries. The priors on $\log M_\mathrm{break}$ and $\log T_\mathrm{MID}$ require somewhat more thought. As for $\log M_{\mathrm{break}}$, we impose a flat prior from 2 to 7 because with we barely see clusters with masses more massive than $10^{6.5} M_{\odot}$. The lower mass limit of 2 in log scale is to ensure a reasonable MCMC walker range. For $\log (T_\mathrm{MID}/\mathrm{yr})$ we impose flat priors from 5 to 10 based on physical plausibility. At the lower end, clusters cannot disperse on less than a crossing timescale, and even the densest clusters detected in LEGUS have crossing timescales well above $10^5$ yr. The upper limit is roughly the age of the Universe.
\subsubsection{The mass-dependent disruption (MDD) model}
\label{sssec:mdd}
For the MDD model, clusters lose mass at a rate that varies as a pure power-law function of their current mass \citep{2007MDD,2021MDD}, $\mathrm{~d}M/\mathrm{~d}T \propto -M^{\gamma_{\mathrm{MDD}}}$. For such a mass loss rate, the present-day mass of a cluster born with initial mass $M_i$ at age $T$ is
\begin{equation}
\label{MDD eq}
M=M_{\mathrm{i}}\left[1-\gamma_{\mathrm{MDD}}\left(\frac{M_0}{M_{\mathrm{i}}}\right)^{\gamma_{\text {MDD }}} \frac{T}{T_{\mathrm{MDD}, 0}}\right]^{1 / \gamma_{\text {MDD }}},
\end{equation}
where $T_{\mathrm{MDD}, 0}$ represents the time required for a cluster with mass $M_0$ to have fully disrupted, defined as having reached a present-day mass $M=0$. The joint mass-age distribution therefore obeys
\begin{equation}
\frac{\mathrm{d}^{2} N}{\mathrm{~d} M \mathrm{~d} T} \propto \frac{\mathrm{d}^{2} N}{\mathrm{~d} M_{\mathrm{i}} \mathrm{~d} T} \frac{\mathrm{d} M}{\mathrm{~d} M_{\mathrm{i}}}.
\end{equation}
If the distribution of initial masses $M_i$ follows the Schechter function form given by \autoref{eq:schechter}, then we have
\begin{equation}
\frac{\mathrm{d}^{2} N}{\mathrm{~d} M \mathrm{~d} T} \propto M^{\alpha_{M}} \eta^{\alpha_{M}+1-\gamma_{\text {MDD }}} \exp \left(-\eta \frac{M}{M_{\text {break }}}\right),
\label{eq:mass_age_mdd}
\end{equation}
where
\begin{equation}
\eta(M, T) = \left[1+\gamma_{\mathrm{MDD}}\left(\frac{M_0}{M}\right)^{\gamma_{\text {MDD }}} \frac{T}{T_{\mathrm{MDD}, 0}}\right]^{1 / \gamma_{\text {MDD }}}
\end{equation}
is the ratio of the initial and present-day cluster masses for a cluster of present-day mass $M$ and age $T$.
The MDD model therefore has four free parameters: $\alpha_M$, $\log M_\mathrm{break}$, $\gamma_\mathrm{mdd}$, and $T_{\mathrm{MDD,0}}$; note that $M_0$ is not a separate parameter, because the mass-age distribution depends only on the combination of parameters $M_0^{\gamma_\mathrm{MDD}}/T_\mathrm{MDD,0}$. We therefore without loss of generality choose $M_0 = 100$ M$_\odot$ in what follows. However, for reader convenience we will also report the commonly-used $t_4$ parameter, which is simply the disruption time for a cluster of mass $10^4$ M$_\odot$; this is given by $t_4 = 10^{2\gamma_\mathrm{MDD}} T_\mathrm{MDD,0}$. We adopt the same priors on $\alpha_M$ and $\log M_\mathrm{break}$ as in the MID model (\autoref{sssec:mid}). For $\gamma_\mathrm{MDD}$, previous observational estimates and $N$-body simulations give values in the range $0.6-0.7$ \citep{2010Lamers}, and we adopt broad priors that include this range: we take $\gamma_\mathrm{MDD}$ to be flat from 0 to 1. These limits stem from physical considerations: if $\gamma_\mathrm{MDD} < 0$ then low-mass clusters lose mass more slowly than massive ones, contrary to the physical expectations of the model, while if $\gamma_\mathrm{MDD} \geq 1$ then no clusters ever disrupt because there is no $T$ for which $M = 0$. Finally, we adopt flat priors on $\log(T_\mathrm{MDD,0}/\mathrm{yr})$ from 4 to 10; these limits are broad enough not to matter, because none of our walkers ever approach them.
\subsubsection{Dust extinction}
\label{sssec:av_dist}
The exact functional shape of the distribution of dust extinctions $p_{A_\mathrm{V}}(A_\mathrm{V})$ is unknown, so we model it as non-parametrically as possible. Following \citet{SLUG2019}, we adopt a simple piece-wise linear form over the range $A_\mathrm{V}= 0-3$ mag characterised by six free parameters ($p_{A_{\mathrm{V}, i}}, 0\leq i \leq 6$) to be fit, representing the value of the PDF at $A_{\mathrm{V}}$ = $i\,\Delta A_\mathrm{V}$ mag with $\Delta A_\mathrm{V} = 0.5$ mag. Thus the functional form we adopt for the extinction distribution is
\begin{eqnarray}
\lefteqn{
p_{A_\mathrm{V}}(A_\mathrm{V}) \propto
}
\\
& &
\begin{cases}
p_{A_\mathrm{V},i} + \left(p_{A_\mathrm{V},i+1} - p_{A_\mathrm{V},i}\right)\left(\frac{A_\mathrm{V}}{\Delta A_\mathrm{V}}-i\right), &
i < \frac{A_\mathrm{V}}{\Delta A_\mathrm{V}} \leq i+1 \\
0, & A_\mathrm{V} > 3\,\mathrm{mag}
\end{cases}
\nonumber
\end{eqnarray}
We treat $p_{A_{\mathrm{V}, i}}$ for $i = 0 - 5$ as parameters of our model to be fit, with $p_{A_\mathrm{V},6}$ fixed by the requirement that the total area under the PDF be unity. We set the priors on $p_{A_{\mathrm{V}, i}}$ to be flat for all values $> 0$, subject to the requirement that $p_{A_{\mathrm{V}, i}} > 0$ remain positive for all $A_\mathrm{V}$.
Combining this with the mass and age distributions, our final functional form to be fit is
\begin{equation}
f(M,T,A_\mathrm{V}\mid\boldsymbol{\theta}) \propto \frac{\mathrm{d}^{2} N}{\mathrm{~d} M \mathrm{~d} T} \, p_{A_\mathrm{V}}(A_\mathrm{V}),
\end{equation}
with $\mathrm{d}^2N/\mathrm{d}M\mathrm{~d}T$ given by \autoref{eq:mass_age_mid} or \autoref{eq:mass_age_mdd} for the MID or MDD models, respectively.
\subsection{Model selection using Akaike weights}
The method described thus far allows us to compute the posterior PDFs of the model parameters $\boldsymbol{\theta}$ for both the MID and MDD models. To determine whether the MID or MDD model provides a better fit and more accurate description of the data, we use the Akaike information criterion (AIC) to assess the fit quality provided by non-nested models. For models tested using AIC, we first identify the walker with the highest likelihood, then compute,
\begin{equation}
\mathrm{AIC}_{(\text {MID,MDD})}=2 k-2 \ln \hat{\mathcal{L}}_{\text {(MID,MDD)}},
\label{eq:AIC}
\end{equation}
where $k$ represents the number of free parameters in the models, and $\hat{\mathcal{L}}$ is the maximum of the likelihood function. In our case, $k$ is 11 for both the MID and MDD models, with four parameters describing the joint mass-age distribution ($\alpha_M$, $M_\mathrm{break}$, and either $\alpha_T$ and $T_\mathrm{MID}$ or $\gamma_\mathrm{MDD}$ and $T_\mathrm{MDD,0}$, six parameters representing the dust extinction shape, and one extra parameter to describe the total number of clusters present in the galaxy.
For the model comparison, we calculate the Akaike weights for either the MID or MDD model as
\begin{equation}
w_{\mathrm{MID/MDD}} = \frac{\frac{e^{-\Delta_{\mathrm{MID/MDD}}}}{2}}{\frac{e^{-\Delta_{\mathrm{MID}}}}{2} + \frac{e^{-\Delta_{\mathrm{MDD}}}}{2}}.
\label{eq:Akaike_w}
\end{equation}
The AIC measures the amount of information in the data preserved by a given model., with the relative Akaike weight $w$ of one model indicating the confidence level at which we can claim that it preserves more information than the other models considered. The model with the highest Akaike weight is our best fit and thus the preferred model.
\section{Results}
\label{cha:results}
In this section, we report our fit parameters and model comparison results (\autoref{ssec:fullcatalogue}), along with comparisons between models and observed photometry to verify that our best-fitting models do a reasonable job at reproducing the observations (\autoref{ssec:photocomparison}). To search for variations in cluster population demographics with galactocentric radius, we also separately analyse clusters in the inner and outer galaxy (\autoref{ssec:radialsplit}).
\renewcommand{\arraystretch}{1.25
\setlength{\tabcolsep}{2.pt}
\begin{table*}
\begin{tabular}{ccccccccccc}
\hline
Catalogue &Model& $w$ & $\alpha_{M}$& $\log \left(M_{\mathrm{break}}/M_{\odot}\right) $ & $\alpha_{M_4}$ & $\alpha_{T}$&$\log \left(T_{\mathrm{MID}}/\mathrm{yr}\right)$ & $\log (T_{\mathrm{MDD,0}}/\mathrm{yr})$ &$\gamma_{\mathrm{MDD}}$ & $t_4(\mathrm{Myr})$ \\ \hline
\multirow{2}{*}{All} &MID &0.9997 & $-1.03^{+0.10}_{-0.29}$ &$3.88^{+0.08}_{-0.06}$ & $-2.07^{+0.09}_{-0.28}$ &$-0.45^{+0.04}_{-0.07}$ &$6.78^{+0.42}_{-0.06}$ & $-$ &$-$ &$-$ \\
&MDD &0.0003 & $-1.40^{+0.07}_{-0.08}$ &$4.00^{+0.05}_{-0.05}$ & $-$ &$-$ &$-$ &$6.65^{+0.10}_{-0.07}$ & $0.98^{+0.02}_{-0.04}$& $268^{+37}_{-29}$ \\ \hline
\multirow{2}{*}{$\{ r \leq R_{50}\}$} &MID & 1.0000 & $-0.54^{+0.18}_{-0.53}$ &$3.65^{+0.42}_{-0.07}$ & $-1.64^{+0.16}_{-0.45}$ &$-0.70^{+0.06}_{-0.19}$ &$6.78^{+0.29}_{-0.05}$ & $-$ &$-$ &$-$ \\
&MDD & < $10^{-11}$ & $-1.20^{+0.14}_{-0.13}$ &$3.87^{+0.08}_{-0.08}$ & $-$ &$-$ &$-$ &$6.42^{+0.11}_{-0.07}$ & $0.97^{+0.02}_{-0.05}$& $151^{+22}_{-18}$ \\ \hline
\multirow{2}{*}{$\{ r > R_{50}\}$ } &MID &0.9988 & $-1.12^{+0.13}_{-0.35}$ &$4.07^{+0.11}_{-0.08}$ & $-2.10^{+0.12}_{-0.32}$ &$-0.45^{+0.06}_{-1.41}$ &$6.85^{+1.44}_{-0.14}$ & $-$&$-$ &$-$ \\
&MDD &0.0012& $-1.36^{+0.09}_{-0.09}$ &$4.13^{+0.07}_{-0.06}$ & $-$ &$-$ &$-$ &$6.75^{+0.20}_{-0.14}$ & $0.93^{+0.05}_{-0.10}$& $274^{+52}_{-44}$ \\ \hline
\multirow{2}{*}{$\{ r\leq R_c \}$ } &MID &1.0000& $-0.90^{+0.09}_{-0.09}$ &$3.91^{+0.05}_{-0.05}$ & $-1.92^{+0.08}_{-0.07}$ &$-0.57^{+0.03}_{-0.04}$ &$6.76^{+0.05}_{-0.04}$ & $-$ & $-$ &$-$ \\
&MDD & < $10^{-11}$ & $-1.36^{+0.07}_{-0.07}$ &$4.05^{+0.04}_{-0.04}$ & $-$ & $-$ & $-$ &$6.48^{+0.07}_{-0.05}$ & $0.98^{+0.01}_{-0.03}$& $183^{+19}_{-17}$ \\ \hline
\multirow{2}{*}{$\{ r > R_c \}$ } &MID &0.9713& $-1.47^{+0.33}_{-0.27}$ &$4.04^{+0.25}_{-0.17}$ & $-2.46^{+0.30}_{-0.22}$ &$-0.46^{+0.27}_{-1.74}$ &$8.23^{+0.42}_{-0.06}$ & $-$ & $-$ &$-$ \\
&MDD &0.0287 & $-1.58^{+0.23}_{-0.23}$ &$4.06^{+0.31}_{-0.16}$ & $-$ & $-$ & $-$ &$8.58^{+0.88}_{-1.08}$ & $0.62^{+0.29}_{-0.43}$& $2463^{+20082}_{-1905}$ \\ \hline
\multirow{1}{*}{$\{ p_{\mathrm{obs}} \geq 90 \% \} $} &MID & $-$ & $-1.42^{+0.11}_{-0.12}$ &$3.89^{+0.05}_{-0.5}$ & $-2.44^{+0.1}_{-0.1}$ &$-^{+0.03}_{-0.04}$ &$6.75^{+0.04}_{-0.04}$ & $-$ & $-$ &$-$ \\ \hline
\end{tabular}
\caption{Fitting results. The first column indicates the subset of the LEGUS catalogue used in the fit, the second indicates the type of model fit (MID or MDD), the third gives the Akaike weight for that model, and the remaining columns give marginal posterior PDFs derived for each model parameter, reported as ${q_{50}}^{+(q_{84}-q_{50})}_{-(q_{50}-q_{16})}$, where $q_N$ denotes the estimated $N^{\mathrm{th}}$ percentile. Thus our central values are the median of the PDF, and the ranges shown correspond to the 68\% confidence interval. The co-rotation radius is denoted by $R_c$, and the median galactocentric radius is denoted by $R_{50}$.}
\label{Tab:params}
\end{table*}
\subsection{Full Catalogue Fits}
\label{ssec:fullcatalogue}
We first analyse the full LEGUS NGC 628 catalogue using the method described in \autoref{cha:methods}. We summarise the marginal posteriors we derive on all model parameters in the first block in \autoref{Tab:params}. Our model comparison yields $w(\mathrm{MID}) = 0.9997$ and $w(\mathrm{MDD}) = 3\times 10^{-4}$, suggesting that the MID model does a substantially better job capturing the variations in the data, and we will therefore focus on this as our preferred model from this point forward.
\autoref{fig:MID_corner} shows the posterior PDFs of $\alpha_M$, $\log M_{\mathrm{break}}$, $\alpha_T$, and $\log T_{\mathrm{MID}}$ we obtain from our fit. In this plot, the extent of the axes reflects the full range of model parameters allowed by our priors, so we can immediately read off where parameters are well-constrained by the data, versus where they are unconstrained and occupy the full range of values allowed by our priors. These figures have a few noteworthy features. First, they suggest that cluster disruption in NGC 628 is relatively mild; the most likely scenario is that disruption begins after $T_\mathrm{MID} \approx 10^7$ yr but the is relatively slow ($\alpha_T \approx -0.5$), though an alternate scenario where cluster disruption is faster but begins later ($\alpha_T \approx -1$ but $T_\mathrm{MID} \approx 10^8$ yr) is marginally allowed. In the former scenario, our median values of $\alpha_T \approx -0.45$ and $\log (T_\mathrm{MID}/\mbox{yr})\approx 6.8$ correspond to cluster disruption starting within $1-2$ crossing times of formation ($T_\mathrm{MID} \approx 6$ Myr), and thereafter occurring at a rate such that the cluster disruption time is $\approx 2.2$ times the current cluster age \citep{2019ARAA}. Similarly, the fits
provide strong evidence that
a truncated Schechter-form CMF fits the data better than the pure power-law, with $\log ({M_{\text{break}}}/\mathrm{M}_\odot) = 3.88^{+0.08}_{-0.06}$ (68\% confidence interval), but that the powerlaw slope at lower masses is relatively shallow, $\alpha_M \approx -1$. As with cluster disruption, there is a tail of probability at higher $M_\mathrm{break}$ and steeper CMF ($\alpha_M\approx -2$), but this tail contains little probability mass.
Because our analysis method allows us to reach considerably smaller cluster mass than many previous studies, and because we find a relatively small value of $M_\mathrm{break}$, it is helpful to report an alternative measure of the mass function slope that will be more readily comparable to literature results. For this reason,
we also compute the slope of the mass distribution at a mass of $10^4$ M$_\odot$,
$$
\alpha_{M_4} \equiv \left.\frac{d\log N}{d\log M}\right\vert_{M=10^4 M_{\odot}} = \alpha_M - \frac{10^4\,\mathrm{M}_\odot}{M_\mathrm{break}}.
$$
This is more directly comparable than $\alpha_M$ to the CMF slopes reported in previous studies where the cluster sample is truncated at masses slightly below $10^4$ M$_\odot$ \citep[e.g.,][]{2014Chandar_MID, LEGUS2017}, and thus the CMF slope is measured only at higher masses.
Our results imply $\alpha_{M_4} = -2.07_{-0.28}^{+0.09}$ (68\% confidence), in good agreement with the values of $\alpha_M\approx -2$ found in most previous studies (e.g., see Figure 5 of \citealt{2019ARAA} for a summary of observations).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/MID_corner.pdf}
\caption{Corner plot showing the one-dimensional and two-dimensional histograms of the posterior PDFs of the parameters $\alpha_M$, $\log {M_{\text{break}}}$, $\alpha_T$, and log $T_{\text{MID}}$ for the full LEGUS NGC 628 catalogue. We omit the nuisance parameters describing the distribution of extinctions $A_V$. The 1D blue histograms show one-dimensional marginal PDFs for each parameter, while the colour maps show log probability densities in various 2D cuts through parameter space. The outermost contour level in the 2D plots is set so as to enclose 99\% of the samples, and black dots show individual MCMC samples outside this threshold. The extent of each axis is rounded according to the prior range to achieve optimal visualisation. }
\label{fig:MID_corner}
\end{figure}
\subsection{Photometric Comparisons}
\label{ssec:photocomparison}
Before accepting the results of our fits, we must validate that our models adequately recover the observed luminosity distribution, since matching this distribution is the goal of our forward model. In order to visualise the comparisons between the distributions of observations and the library, we first plot the 1D cluster luminosity distributions in the five LEGUS bands, showing both the measured distribution and the luminosity distribution we predict using the model parameters. The model prediction follows immediately from our expression for the likelihood function in terms of our Gaussian mixture model (\autoref{eq:likelihood}), and is simply
\begin{equation}
p(m) \propto \sum_{j=1}^{N_\mathrm{lib}} w_j(\boldsymbol{\theta}) \mathcal{N}(m \mid m_j, h),
\label{eq:lumdist}
\end{equation}
where $m$ is the magnitude in the filter of interest, $m_j$ is the magnitude of the $j$th library cluster in that filter, and $h$ is the library bandwidth. We show this comparison in \autoref{fig:1D_MID}. As the figure shows, our best-fitting model generally matches the observed luminosity distribution except for at the most luminous end of the distribution.
Since our method attempts to fit the full five-dimensional photometric distribution at once, we can also check for agreement between model and observations in multiple dimensions. To this end, we present three colour-colour plots illustrating the agreement between the observed and model-predicted colour distributions in \autoref{fig:UV_U_B}, \autoref{fig:U_B_V} and \autoref{fig:B_V_I}. As with the luminosity distribution, we compute the colour distribution directly from the Gaussian mixture model, using an expression analogous to \autoref{eq:lumdist}. These figures also show reasonably good agreement between the observed and model-predicted colour distributions, and suggest that our best-fit model provides a reasonable representation of the distribution of observations in 5D photometric space.
The largest deviations between our model and the observations is for the most luminous clusters at the bluest colours; we can see this from our slight underprediction of the UV (F275W) and U luminosity functions in \autoref{fig:1D_MID}, and from the tail of observed clusters extending to higher UV - U (F275W-F336W) than our models in
\autoref{fig:UV_U_B}. One possible explanation for this discrepancy is that our model library assumes Solar metallicity, while NGC 628 is slightly sub-Solar \citep[e.g.,][]{Berg13a, Berg15a}. \citet{2015Powerlaw} show that metallicity variations at this level have relatively little effect when inferring cluster masses or ages, but it is conceivable that because our library is somewhat more metal-rich than the data, it does not quite extend as far to the blue as the observations.
In any event, the effects of this discrepancy on population-level statistics should be small, since they involve only a handful of clusters out of our full catalogue of $>1000$.
\begin{figure}
\includegraphics[scale=0.9]{figs/1D_L_mid.pdf}
\caption{Comparison between observed and model-predicted luminosity functions in UV, U, B, V, I bands. Filled circles with error bars show the luminosity distribution of the LEGUS catalogue, divided into 20 uniformly spaced bins; distributions are normalised to have unit integral, and error bars show the Poisson uncertainty in each bin. The black lines in the figure show model-predicted photometric distributions computed using the $50^{\mathrm{th}}$ percentile values for all model parameters, while the grey bands around these lines show the interquartile range for the models.
\label{fig:1D_MID}
}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.9]{figs/color_mag_UV_U_B.pdf}
\caption{Comparison of predicted and observed colour distributions for our best-fitting model. The central panel shows the predicted colour distribution computed using the $50^{\mathrm{th}}$ percentile values of all parameters as the background heatmap, while blue circles show every $15^{\text{th}}$ cluster in the LEGUS catalogue. The flanking panels compare 1D histograms of observed (red bars) and predicted colour PDFs (blue bars). In order to visualize the Poisson error of data, we plot the $84^{\text{th}}$ percentile upper limit on the observational distribution as a solid red line.
}
\label{fig:UV_U_B}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.9]{figs/color_mag_U_B_V.pdf}
\caption{Same as \autoref{fig:UV_U_B}, except we show U-B versus B-V in colour-colour space here.}
\label{fig:U_B_V}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.95]{figs/color_mag_B_V_I.pdf}
\caption{Same as \autoref{fig:UV_U_B}, except we show B-V versus V-I in colour-colour space.}
\label{fig:B_V_I}
\end{figure}
\subsection{Inner versus Outer Galaxy Clusters}
\label{ssec:radialsplit}
Having presented the full catalogue fits, we now search for radial variations in cluster demographics. For this purpose we assign every cluster in the LEGUS catalogue a galactocentric radius, taking the centre of NGC 628 to be at $\mathrm{RA}=24.17^\circ$, $\mathrm{DEC}=15.78^\circ$, and adopting an inclination $i=25.2^\circ$ and position angle $\mathrm{PA}=25^\circ$ from North \citep{Grasha2015,Grasha2017}. We use the centre location, inclination, and position angle to deproject the cluster coordinates from RA/DEC to a galactocentric coordinate system. Once clusters have been assigned radii, we divide the catalogue in two ways: first simply by making two sub-catalogues of equal size containing the inner and outer halves of the sample, and second by making two sub-catalogues containing clusters inside and outside the galactic co-rotation radius. The advantage of the former approach is that it ensures that we have equal statistical power in both regions, and thus do not miss radial variations because our inner or outer galaxy sample is too small to see them. The advantage of the latter approach is that co-rotation marks a physically-motivated radius where we might expect to see a change in cluster behaviour, as opposed to the galactocentric radius that marks an equal number division, which is solely a function of the size and location of the LEGUS pointings.
\subsubsection{Equal Number Division}
For our first analysis we divide the LEGUS cluster sample into two radial bins each containing an equal number of clusters. The median radius that makes this even division is $R_{50} = 3.53$ kpc. We fit MID and MDD models to both the subset of clusters with $r \leq R_{50}$ and those with $r > R_{50}$ and report the results the second and third blocks of rows in \autoref{Tab:params}. As with the full catalogues, our model comparison strongly prefers the MID model in both the inner and outer regions, and so we focus on it.
We compare the marginal posterior PDFs of $\alpha_{M_4}$ and $M_\mathrm{break}$ in the inner and outer regions \autoref{fig:Div_half_am4Mb}.
We notice that the posterior mass distributions of the inner bin (indicated by the red contours) are bimodal, corresponding to two probability peaks in the parameter space, one at $\alpha_{M_4} \approx -1.8$ and $\log (M_\mathrm{break}/\mathrm{M}_\odot) \approx 3.8$, the second at $\alpha_{M_4}\approx -2.5$ and $\log(M_\mathrm{break}/\mathrm{M}_\odot) \gtrsim 5$, while the posterior for the outer region is closer to unimodal, centred at $\alpha_{M_4}\approx -2$ and $\log(M_\mathrm{break}/\mathrm{M}_\odot)\approx 4.2$.
Despite this difference, however, the $2\sigma$ confidence contours for the two regions overlap, so there is weak but not strongly convincing evidence for a difference in mass function between the two regions.
\autoref{fig:Div_half_aTTb} provides an analogous plot for the two parameters -- $\alpha_T$ and $T_\mathrm{MID}$ -- that describe the CAF. The qualitative conclusion to be drawn from this Figure is similar to that for the mass PDFs as shown in \autoref{fig:Div_half_am4Mb}, i.e., the posterior PDFs for the inner and outer radial bins are largely separated, but show a small region of overlap, in this case at $T_\mathrm{break}\approx 6$ Myr and $\alpha_T \approx -0.7$. This result suggests that we cannot rule out the possibility of two regions having identical age distributions, but is suggestive of one. To the extent that there is a difference, it is that the inner galaxy clusters show steeper $\alpha_T$ and/or smaller $T_\mathrm{MID}$, suggesting the clusters closer to the galactic centre undergo more severe disruption than those farther away.
\subsubsection{Co-rotation Radius Division}
An alternative way of separating the cluster population is into those inside and outside the galactic co-rotation radius at $R_c = 6.3$ kpc where the spiral pattern and orbiting clusters move at equal speed \citep{2021CoR}. To the extent that encounters with molecular clouds or other structures associated with the spiral arms influence cluster formation or disruption, we might expect to see changes in cluster population demographics across co-rotation. Making this division, we have 1159 clusters inside $R_c$ and 117 clusters outside $R_c$.
\autoref{fig:Div_galac_am4Mb}
and \autoref{fig:Div_galac_aTTb} show the inferred posteriors for the clusters located inside and outside the co-rotation radius; we again concentrate on the MID model, because model comparison prefers it to MDD. Here we see differences between the inner and outer galaxy posterior PDFs that are qualitatively consistent with those we observed for a division of the sample into two equal parts. However, the small number of clusters in the outer galaxy sample when we divide at co-rotation ensures that the posterior PDFs are very broad, and, particularly for the age distribution, largely just reflect our priors. Given the very broad outer galaxy posteriors, we cannot rule out the possibility that the cluster demographics inside and outside co-rotation are the same. Doing so would likely require significantly more outer-galaxy clusters than the LEGUS catalogue provides.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Div_r50_am4Mb.pdf}
\caption{2D marginal posterior joint PDF of the parameters $\alpha_{M_4}$ and $\log M_{\mathrm{break}}$ that describe the CMF, for clusters inside (red) and outside (blue) the median cluster galactocentric radius. Blue and red contour lines in the central panel correspond to loci that enclose 68\%, 95\%, and 99\% of the MCMC samples, and thus correspond approximately to 1$\sigma$, 2$\sigma$ and 3$\sigma$ significance confidence levels. The histograms flanking the central panel show the corresponding 1D marginal PDFs of $\alpha_{M_4}$ and $\log M_{\mathrm{break}}$ for the inner (red) and outer (blue) galaxy samples.}
\label{fig:Div_half_am4Mb}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Div_r50_aTTmid.pdf}
\caption{Same as \autoref{fig:Div_half_am4Mb}, expect we show the marginal PDFs of the parameters describing the CAF, $\alpha_{T}$ and $\log T_{\mathrm{MID}}$, rather than the parameters describing the CMF.}
\label{fig:Div_half_aTTb}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Div_coR_am4Mb.pdf}
\caption{
Same as \autoref{fig:Div_half_am4Mb}, but for clusters divided into those inside and outside the co-rotation radius $R_c$, rather than inside and outside the median radius $R_{50}$.
}
\label{fig:Div_galac_am4Mb}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Div_CoR_aTTmid.pdf}
\caption{
Same as \autoref{fig:Div_half_aTTb}, but for clusters divided into those inside and outside the co-rotation radius $R_c$, rather than inside and outside the median radius $R_{50}$.
}
\label{fig:Div_galac_aTTb}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Div_comp_am4Mb.pdf}
\caption{Same as \autoref{fig:Div_half_am4Mb}, but comparing the results for the full LEGUS catalogue (blue) to those obtained using only clusters in part of luminosity space where the estimated completeness is $>90\%$.}
\label{fig:Div_comp_am4Mb}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Div_comp_aTTmid.pdf}
\caption{Same as \autoref{fig:Div_comp_am4Mb}, expect we show age distributions illustrated by 2D PDF of $\alpha_{T}$ versus $\log \left(T_{\mathrm{MID}}/\mathrm{yr}\right)$.}
\label{fig:Div_comp_aTTb}
\end{figure}
\section{Discussion}
\label{cha:discussions}
Here we discuss some implications of our findings. We begin with an additional verification check on our method (\autoref{ssec:completeness_robustness}), and then discuss our findings regarding truncation of the cluster mass distribution (\autoref{sec:CMF}) and the cluster age distribution (\autoref{ssec:cluster_disruption}).
\subsection{Method robustness against completeness uncertainty}
\label{ssec:completeness_robustness}
In \autoref{ssec:completeness} we explain our method to derive the completeness function $P_{\mathrm{obs}}(\mathbf{m})$ for the data and the library. Knowledge of this function, coupled to our forward-modelling techniques, allows us to avoid the traditional method of excluding clusters with masses $\lesssim 5000$ M$_{\odot}$ and ages $\gtrsim 200$ Myr, and in turn allowing us to use essentially all of the LEGUS catalogue, rather than only $\approx 25\%$ of it as in previous work \citep[e.g.,][]{LEGUS2017}. However, our approach is sensitive to our completeness estimates, which therefore represent a potential source of systematic error. To test this possibility, we repeat the numerical experiments with a more conservative approach of only including clusters where we estimate the completeness is $\geq 90\%$. This reduces the sample to 1120 clusters, but ensures that $P_\mathrm{obs}(\mathbf{m})$ never varies by more than 10\%, and thus errors in it have little effect.
We compare the mass and age parameters we derive using this reduced catalogue to those we obtain from the full catalogue in \autoref{fig:Div_comp_am4Mb} and \autoref{fig:Div_comp_aTTb}, respectively. We find that the regions of allowed parameter space for both the parameters describing mass ($\alpha_{M_4}$, $\log M_\mathrm{break}$) and those describing age ($\alpha_T$, $\log T_\mathrm{MID}$) are consistent. The probability peaks are slightly more concentrated for the high completeness sample, but are almost entirely overlapping with the probability maxima we obtain for the full catalogue; the results are consistent at the $1\sigma$ level.
We see the similarity of the different physical parameters as shown in Figures \ref{fig:Div_comp_am4Mb} and \ref{fig:Div_comp_aTTb}. Notably, we found consistent overlapping regions in these three plots. In addition, the red contours fitted with high completeness values have close probability peaks but are more concentrated than the entire catalogue consisting of the fainter clusters. We therefore conclude that possible systematic errors in the completeness function have at most a trivial impact on the outcome.
\subsection{CMF with a high mass truncation}
\label{sec:CMF}
Our results favour a CMF that is truncated at a relatively small mass, $M_\mathrm{break} \approx 10^4$ M$_\odot$, and are strongly incompatible with a pure powerlaw with a slope $\alpha_M\approx -2$ -- to the extent that our results leave any room for a non-truncated mass function, they allow this possibility only if the powerlaw slope $\alpha_M$ is significantly steeper than $-2$. It is therefore interesting to compare our $M_\mathrm{break}$ to values reported in the literature. Previous studies of CMF in spiral galaxies such as M83 \citep{2015GEM83}, NGC 628 \citep{LEGUS2017}, and M51 \citep{2018Messa} all found that Schechter functions provided better fits to the CMF than pure powerlaws, but with exponential truncations in mass at $\approx 10^5 M_{\odot}$, roughly an order of magnitude higher than our value. Indeed, our break mass is roughly an order of magnitude smaller than the value obtained by \citet{LEGUS2017} from the same underlying photometric catalogue.
One possible explanation for this difference is that the previous studies that reported a higher $M_\mathrm{break} \sim 10^5 M_{\odot}$ use mass-and age-limited samples, $M > 5000$ M$_{\odot}$ and $T \leq 200$ Myr. These mass and age cuts induce large uncertainties in the truncation mass estimate because they require that the majority of the data be discarded. Indeed, it is worth stressing that it would be essentially \textit{impossible} for \citeauthor{LEGUS2017}'s study of NGC 628 to have detected a break in the mass function at $M_\mathrm{break} \approx 7600$ M$_\odot$ as we find, simply because this value of $M_\mathrm{break}$ is nearly the same as the mass at which they truncate their sample; the could not have detected a change in CMF slope below this mass. Consequently, CMFs reported in earlier studies should be understood as describing only the CMF of the remaining heavy clusters after the bulk of the population has been removed. This procedure may generate a high truncation mass, whereas in our study these massive clusters, which are greatly outnumbered by the low-mass ones, have much less weight in the fitting. Consistent with this hypothesis, we note that the only two other published studies of the extragalactic CMF that have mass sensitivity comparable to ours (because they target galaxies close enough to resolve individual stars) \textit{also} report breaks in the mass function closer to $10^4$ than $10^5$ M$_\odot$ -- \citet{2017PHAT} find $M_\mathrm{break} \approx 8500$ M$_\odot$ in M31, while \citet{2022Wainer} find $1.7\times 10^4$ M$_\odot$ in M33. This strongly suggests that in at least some cases there are real features in CMFs at masses below the range accessible to traditional, non-Bayesian integrated light techniques.
Besides this statistical effect, the difference between our results and the higher break masses found in some previous studies may also reflect real changes in the cluster population due to galaxies' physical conditions. \citet{2017PHAT} propose a relation between the break mass and the star formation rate density $\langle \Sigma_{\mathrm{SFR}} \rangle$, perhaps due to both varying systematically with mean interstellar pressure; \citet{2022Wainer} extend this proposal by adding new data. In order to place our fitted break mass into this context, we add our value for $M_\mathrm{break}$ in NGC 628 to \citeauthor{2022Wainer}'s compilation in \autoref{fig:Mc_SFRD}; for this purpose we adopt a star formation rate density for NGC 628 from \citet{2022Santoro_SFRD}. As the figure shows, our best-fit break mass is generally consistent with the relation fit by \citeauthor{2017PHAT},
\begin{equation}
\log \frac{M_\mathrm{break}}{\mathrm{M}_\odot} = (1.07\pm 0.10)\log\frac{\langle\Sigma_\mathrm{SFR}\rangle}{\mathrm{M}_\odot\;\mathrm{yr}^{-1}\;\mathrm{kpc}^{-2}} + (6.82\pm 0.20).
\end{equation}
This provides a possible physical explanation for why our break mass for NGC 628 is lower than the break masses typically found in galaxies like M51 and M83: NGC 628 has a lower star formation rate per unit area, presumably indicative of a lower interstellar pressure.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/trM.pdf}
\caption{Comparison of break masses $M_\mathrm{break}$ and star formation rate surface densities $\Sigma_{\mathrm{SFR}}$ for a sample of nearby galaxies. Black points show the compilation of \citet{2022Wainer}, while the black line shows the best-fit linear relation obtained by \citet{2017PHAT}; they grey band shows the 84\% confidence interval around the relation. ``M74(A2017)'' denotes the truncated mass derived in \citet{LEGUS2017} to make it distinct from the mass cut-off reported in our study. The red point shows our result for the full NGC 628 catalogue, with error bars indicating the 84\% confidence interval. We take our value of $\Sigma_{\mathrm{SFR}}$ for NGC 628 from \citet{2022Santoro_SFRD}.}
\label{fig:Mc_SFRD}
\end{figure}
\subsection{Implications for cluster disruption}
\label{ssec:cluster_disruption}
We find that mass-independent disruption (MID) is a better description of the cluster age distribution than mass-dependent disruption (MDD) in all regions, but that the disruption rate is relatively modest, $\alpha_T \approx -0.45$. As with our results for the CMF, it is interesting to put this result into the context of previous studies. As summarised by \citet{2019ARAA}, the results of these studies have tended to depend strongly on the method of cluster catalogue construction, with authors who use ``exclusive'' catalogues that exclude objects that do not have compact, round morphologies favouring MDD and weak disruption, while those who use ``inclusive'' catalogues without such morphological filters favouring MID and more rapid cluster disruption, $\alpha_T \approx -1$. This difference arises because the morphological filters tend to preferentially affect the youngest clusters, so applying them flattens the age distribution. There is a great deal of debate in the literature, which we shall not revisit here, about the extent to which morphological filters artificially remove real star clusters from the sample, versus the extent to which a failure to filter leads to catalogues including large numbers of unbound associations that were never bound to begin with.
Because we include LEGUS class 3 sources in our catalogue, and thus are not applying morphological filters, one would naively have expected our results to more closely match those coming from the inclusive catalogues. However, our actual result -- MID but weak disruption -- sits partway between the inclusive and exclusive results, and this
is likely at least in part because our method
allows us to investigate significantly older ages than in previous studies. This makes us much less sensitive to the young ages where the uncertainty about morphological filtering is most acute. Our statistical signal is not dominated by the small number of clusters with ages $\lesssim 100$ Myr, but by the much larger number of clusters at somewhat older ages, where identification of clusters much less ambiguous since at these ages all bound stellar systems are relaxed. Thus the situation here is somewhat analogous to that for the CMF: it is not such much that our result is incompatible with past work, as that our Bayesian forward-modelling approach provides access to parts of parameter space that were not previously accessible.
It is also interesting to compare our results to those for NGC 628 specifically, for which \citet{LEGUS2017} found mass-independent disruption of clusters in the inner pointing NGC 628c, but mass-dependent disruption at the outskirts in pointing NGC 628e. Since we choose to analyse the cluster population as a whole or divide it by galactocentric radius, rather than by pointing, the samples we analyse are not completely identical. However, since the clusters belonging to NGC 628c (the inner region) constitute the great majority of the population, our model selection result for NGC 628 as a whole is consistent with the findings of \citeauthor{LEGUS2017}. There is possibly some tension in the fact that we favour MID even when we divide our catalogues by galactocentric radius, while \citeauthor{LEGUS2017} find MDD in NGC 628e, but since the divisions are not exactly the same we cannot make this statement more quantitatively. However, it does reinforce the point that our method has the advantage that, since we are forward modelling the completeness in each field, we can easily break up the clusters by any criterion of our choice to derive the demographics in different parts of the galaxy.
\section{Summary And Conclusion}
We use a novel method described by \cite{SLUG2019} to derive the demographics of the cluster population of the nearby spiral galaxy NGC 628 from unresolved photometry measured by the LEGUS survey \citep{CalzettiLEGUS2015}. This method has the advantage of integrating diverse observational data imaged using different filters or depths, without the need for binning or imposing extreme completeness cuts. As a result, our method allows us to analyze the roughly 1200 clusters in NGC 628 catalogued by the LEGUS survey, as opposed to the past work that was restricted to include only the most massive and youngest $\sim 30\%$ of the population. Our method also obviates the need to assign a unique mass and age to every individual cluster of the population, and therefore avoids the inevitable information loss incurred by reducing complex, sometimes multiply-peaked posterior probability distributions for individual cluster masses and ages to single values or even Gaussians. We validate the method by comparing the photometric distributions that emerge from our fits to the observed ones and finding good agreement, indicating that our approach can identify models for the cluster population that reproduce observations.
Our analysis results in three primary findings. First, we use the Akaike information criterion to compare two classes of parametric models for the time-evolution of the cluster population, mass-independent disruption (MID) and mass-dependent disruption (MDD). We find that the MID model fits the data better, both for the entire NGC 628 cluster catalogue and for all sub-divisions of it that we consider. We thus conclude that cluster mass and age distributions are at least approximately separable and that the disruption time scales of the clusters have no dependence on the cluster masses.
Second, we find that cluster mass functions are well-described by Schechter functions $dN/dM \propto M^{\alpha_{M}} \exp(-M/M_{\mathrm{break}})$, with a well-constrained value of $M_\mathrm{break} \sim 10^4$ M$_\odot$, indicating that there is significant evidence for truncation in the cluster mass function. Broadly, our finding indicates that the slope of the mass function changes from $\alpha_M \sim -1$ at $10^3$ M$_{\odot}$ to $\sim -2$ at $10^4$ M$_{\odot}$ to much steeper and nearly cut off at $\sim 10^5$ M$_{\odot}$. Our truncation mass is significantly lower than the one determined by \citet{LEGUS2017} for NGC 628 but is in good agreement with those measured for M31 and M33 by \citet{2017PHAT} and \citet{2022Wainer}, respectively. These studies are based on resolved stellar populations, and thus are similar to ours in that they can probe significantly lower cluster masses than earlier studies such as \citet{LEGUS2017} based on integrated light that have found higher truncation masses.
Third, we find that the cluster age distribution in NGC 628 indicates that cluster disruption begins when clusters are relatively young (age $\approx 5-15$ Myr), but that disruption is relatively mild thereafter. The age distribution once disruption starts can be described by a simple power-law with slope $\alpha_T \approx -0.4$. In past studies, the value of $\alpha_T$ has proven highly sensitive to the cluster catalogue type used, with studies using an inclusive cluster catalogues (those that do not include filters based on morphology) generally finding $\alpha_T \sim -1$ while those using exclusive catalogues (which are filtered for morphology) finding $\alpha_T \sim -0.3$. The significance of our study is that even though we are using an inclusive catalogue, we still find a relatively shallow age slope.
The three conclusions above are derived from our analysis of all the clusters in NGC 628 catalogued by LEGUS. However, we also group clusters by galactocentric radius and search for differences in cluster demographics with radius. We do not find strongly convincing evidence for any radial variations, though there are suggestive hints that inner galaxy clusters are subject to more severe disruption than those further out. Our ability to search for radial variations is limited by the fairly small number of clusters available at larger galactocentric radii, since the great majority of the available sample comes from the LEGUS NGC 628c pointing, which targets the galactic centre. Given the suggestive hints in our analysis, increasing the sample of outer-galaxy clusters would be a worthwhile effort.
Looking forward, now that we have demonstrated the performance of our pipeline on NGC 628, we are in a position to apply it to other cluster catalogues, both those originating from LEGUS and from other studies. Analysis of a diverse sample of galaxies will allow us to compare the demographics of cluster populations in various galactic environments. This in turn opens up the possibility of testing a range of theoretical models for cluster formation and disruption and their variation with the galactic environment.
\label{cha:conclusions}
\section*{Acknowledgements}
JT acknowledges support from Kathryn Grasha’s Discovery Early Career Researcher Award (DECRA) Fellowship (DE220100766). KG is supported by the Australian Research Council through the Discovery Early Career Researcher Award (DECRA) Fellowship DE220100766 funded by the Australian Government.
MRK acknowledges support from the ARC Future Fellowship and Laureate Fellowship funding schemes, awards FT180100375 and FL220100020.
This work is supported by the Australian Research Council Centre of Excellence for All Sky Astrophysics in 3 Dimensions (ASTRO~3D), through project number CE170100013.
Based on observations made with the NASA/ESA Hubble Space Telescope, obtained at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. These observations are associated with program \#13364.
\section*{Data Availability}
The cluster catalog data underlying this article are publicly available online at
\href{https://archive.stsci.edu/prepds/legus/dataproducts-public.html}{MAST}. The software suite, tools, and LEGUS catalogues required to perform the MCMC optimization, as well as the output chains, are publicly accessible on \href{https://bitbucket.org/janet_jianling_tang/legus_slug23/src/master/}{Bitbucket}.
\bibliographystyle{mnras}
|
\section{Introduction}
The stochastic motion performed by a point particle when interacting with the quantum vacuum fluctuations of a relativistic field, e.g., scalar or electromagnetic, is also known as Quantum Brownian motion (QBM). This is an example of a phenomena class which arise from quantum vacuum fluctuations and that, over the past several years, has been studied in different scenarios and with different approaches \cite{gour1999will,yu2004vacuum,yu2004brownian,yu2006brownian,seriu2008switching,seriu2009smearing,
bessa2009brownian,de2016probing,de2019remarks,de2014quantum,camargo2018vacuum,camargo2019vacuum,Camargo:2020fxp,mota2020induced,anacleto2021stochastic,ferreira2022quantum, bessa2017quantum}. The quantum vacuum fluctuations are always present but only become observable when the vacuum is somehow perturbed, for instance, by considering elements such as boundary conditions, temperature, nontrivial topology and so on.
Similarly to the classical Brownian motion, the typical quantities for the quantum version that should be investigated are the position and velocity dispersions. However, the analogy between the classical and quantum Brownian motion is limited given that in the quantum scenario the dispersions can assume negative values, something that does not occur in the classical case. In the latter, dispersions are quantities positively defined, that is, $\langle(\Delta A)^{2}\rangle > 0$, where $A$ is some physical observable to be measured. So, negative values for the dispersions in the classical scenario does not make sense. On the other hand, in the quantum context it is possible that $\langle(\Delta A)^{2}\rangle < 0$, which can be interpreted as due to quantum uncertainty reduction \cite{yu2004brownian,yu2004vacuum}, subvacuum effects \cite{camargo2018vacuum,Camargo:2020fxp, camargo2019vacuum} and failure in the renormalization process as a consequence of boundary conditions imposed on the field \cite{de2014quantum}.
In the quantum context, the basic idea is that a point particle (structureless) interacting with quantum vacuum fluctuations of a field has an induced stochastic motion. In the electromagnetic case, for instance, it has been analyzed in Refs. \cite{yu2004vacuum} and \cite{yu2004brownian} the QBM as a consequence of one and two perfectly reflecting parallel planes, respectively. In both cases the position and velocity dispersions are calculated. Moreover, the study of thermal effects for the QBM in the electromagnetic case with one perfectly reflecting plane was also developed in Ref. {\cite{yu2006brownian}, where the magnitude of thermal and quantum contributions are discussed. Thereby, it is shown that for well defined temperature regimes one contribution can be more significant than the other. By seeking to investigate more realistic systems, a wave packet like structure for the particle has been proposed in Ref. \cite{seriu2009smearing}. In addition, switching time effects associated with the interaction between a point particle and the quantum vacuum fluctuations of the field are considered in Refs. \cite{seriu2008switching,de2016probing}. Also, switching time effects at finite temperature are taken into account in Ref. \cite{de2019remarks}.
Regarding the QBM induced by vacuum fluctuations of a quantum scalar field, the investigations conducted follow similarly to the electromagnetic case. In Ref. \cite{de2014quantum}, for instance, it is studied the induced QBM due to a massless scalar field in the presence of a perfectly reflecting plane and in Ref. \cite{camargo2018vacuum} switching time effects are taken into account. The massive scalar field case in $(D+1)$-dimensions, with one perfectly reflecting plane, is studied in Ref. \cite{camargo2019vacuum} and thermal effects in Ref. \cite{Camargo:2020fxp}, both also considering switching time effects. All scalar field cases just mentioned use Dirichlet boundary condition, which set a null value for the field modes on the boundary. As a complement to the study of the QBM, it is worth mentioning that it has also been investigated in the cosmological context, in particular, considering dark matter detection. According to this scenario, in principle, dark matter may induce a stochastic motion in a test particle of ordinary matter, whose observation would offer new insights into the understanding of dark matter properties \cite{Cheng:2019vwy}.
As a contribution to all cases considered in literature so far, for the massless scalar field, we intend to take into consideration two elements, as far as we know, not yet explored in the study of the QBM. The first of them is to consider, in analogy to the electromagnetic case, two perfectly reflecting parallel planes where the scalar field satisfies not only Dirichlet but also Neumann and mixed boundary conditions (BC's). This way, we confine the modes of the field in one direction, something that naturally leads to momentum discretization in the same direction, providing a natural scale for the system. The second element we would like to consider is the effect of a quasiperiodic condition on the QBM of a scalar point particle.
The conditions mentioned in the previous paragraph can also be seen as possible ways of alter the topology of the spatial section of the Minkowski spacetime, which is the background where we are performing our investigations. The consideration of two planes, for instance, breaks the homogeneity and isotropy of space, which can be interpreted as a way of simulating a topological modification in the spatial section of the spacetime. In the case of the induced inhomogeneity, we note that, in the presence of planes, the spatial directions $y$ and $z$ are similar, but differ from the $x$ direction, where the planes are located. In fact, an observer in the $yz$ plane will perceive an infinity bidimensional space, but the same observer on the $xy$ or $xz$ planes will perceive a semi-infinity space, that is, infinity in the $y$ and $z$ directions, but finite in the $x$ direction. On the other hand, the anisotropy, as we shall see, it is shown by the distinction between the velocity dispersions, which is the observable investigated in this work.
It is important to mention that the investigation of the induced Brownian motion considering nontrivial topologies for the spatial section of the Minkowski spacetime is a topic that has been explored for the past several years. Recently, in Ref. \cite{Bessa:2019aar}, it was investigated the Brownian motion of a point particle induced by quantum vacum fluctuations of an electromagnetic field in a flat spacetime whose spatial section has nontrivial topologies. In principle, it is suggested that this effect can be used to indicate the global inhomogeneity of space.
For similar and more recent discussions see also Refs. \cite{Lemos:2020ogj}, \cite{Lemos:2021rya} and \cite{Lemos:2022rms}, where these effects as a function of their time evolution are used as a supposed indicator of spatial orientability. See also Ref. \cite{Lemos:2021jzy} for an example in a conformally expanding flat spacetime.
In addition, we would like to point out that, in this context, the Casimir effect has also been investigated; for more details see for instance Ref. \cite{Sutter:2006dj} and references therein. Therefore, taking into consideration the current status of the subject just described, in the case of the massless scalar field, the present work aims to complement the investigations conducted so far for the QBM.
It is also worth to emphasize that BC's are not merely technical and mathematical details of academic interest, they also can be related to physical properties of the studied systems. Dirichlet and Neumann BC's, for instance, specify the field value and its normal derivate on the boundary, respectively. Typically, we found these conditions in electrostatic systems where either an electrical potential is fixed on the surface (Dirichlet BC) or the corresponding electric field ($\nabla\phi$) is the one fixed on the surface (Neumann BC) \cite{jackson1998classical, arfken2005mathematical}. Also, there exists mixed BC's, in which case both the field and its normal derivative are specified on the boundary. In Ref. \cite{alves2000spontaneous}, for instance, the spontaneous emission of a two-level system between two parallel plates has been investigated taking into consideration that the electromagnetic vector potential obeys BC's similar to the mixed one on the plates. In this case, one of the plates is perfectly conducting and the other one is perfectly permeable. Hence, we can say that mixed BC's simulates plates with distinct physical properties. As to the quasiperiodic condition, it can indicate the existence of an interaction present in the system. As an example, we can mention the Aharonov-Bohm effect \cite{de2012topological, kretzschmar1965must}.
Regarding the structure of this paper, in Section \ref{SecWF} we describe the system to be investigated, indicating some useful simplifications. Then, we exhibit the complete set of normalized solutions for the scalar field for each condition used in this work. This allows us to obtain the positive frequency Wightman function for each case, which is a fundamental element in our calculations. We also obtain a general form for the Wightman function representing Dirichlet, Neumann and mixed BC's in a single expression. In Section \ref{SecVelDis} we calculate the particle velocity dispersion. Finally, in Section \ref{Conclusions}, we present our conclusions summarizing the main results obtained. Note that we have also dedicated Appendices \ref{AppA} and \ref{AppB} to obtain important expressions used to investigate asymptotic limits for the velocity dispersions. In this work we use natural units such that $c=\hbar=1$.
\section{Wightman Functions}\label{SecWF}
\subsection{Model, general field solution and the expression to calculate the Wightman function}
In this section we want to establish some important results that will be used later on in the velocity dispersion computation, namely, the complete set of normalized solutions for the scalar field and the corresponding Wightman functions. In other words, we are interested in investigating the induced QBM of a point particle coupled to a fluctuating quantum massless scalar field, considering different conditions. As it is known, this stochastic motion is induced by the quantum vacuum fluctuations of the field. The classical action that describes this system is written as
\begin{eqnarray}\label{eq01}
S_{\text{tot}} = S_{\text{f}} + S_{\text{p}} + S_{\text{int}},
\end{eqnarray}
where
\begin{eqnarray}
S_{\text{f}} &=& \int dt\int dV \dfrac{(\partial_{\mu}\phi)(\partial^{\mu}\phi)}{2},
\end{eqnarray}
is the massless scalar field part of the action,
\begin{eqnarray}
S_{\text{p}} &=& \int dt \dfrac{m{\bf\dot{ x}}^{2}}{2}
\end{eqnarray}
corresponds to the action describing a point particle of mass $m$ and
\begin{eqnarray}
S_{\text{int}} &=&-g\int dt\int dV\delta^3({\bf x}-{\bf x'})\phi
\end{eqnarray}
stands for the interaction between the particle and the massless scalar field $\phi$, $dV$ is the volume element of the spatial section of the spacetime and $\delta^3({\bf x}-{\bf x'})$ is the spatial three dimensional Dirac delta function. Note that the measure of the strength of the interaction, denoted by $g$, is the charge of the point particle. This is a model widely known in the literature and has been considered in different scenarios \cite{gour1999will,de2014quantum,camargo2018vacuum,de2019remarks,camargo2019vacuum,Camargo:2020fxp}.
The variation of the action \eqref{eq01} with respect to field, $\phi$, provides the massless Klein-Gordon equation with a three dimensional Dirac delta function as a source, that is,
\begin{eqnarray}\label{eq02}
\Box\phi({\bf x},t) = -g\delta^3({\bf x}-{\bf x'}),
\end{eqnarray}
where $\Box =\partial_{\nu}\partial^{\mu}$ is the d'Alembertian differential operator to be considered in Minkowski spacetime described by the line element
\begin{equation}
ds^2 = dt^2 - dx^2 - dy^2 - dz^2.
\end{equation}
Thereby, Eq. \eqref{eq02} is the equation of motion for a massless scalar field coupled to a point particle with mass $m$ and charge $g$. Although this is a non-homogeneous differential equation, we wish to consider that the point particle's influence on the field is negligible \cite{de2014quantum}. This allows us to write Eq. \eqref{eq02} as
\begin{eqnarray}
\Box\phi({\bf x},t) \approx 0,
\end{eqnarray}
which gives a general nonnormalized solution in terms of plane waves, i.e.,
\begin{eqnarray}\label{eq03}
\phi_{\sigma}({\bf x},t)=Ne^{-i\omega t + ik_{x}x +ik_{y}y+ ik_{z}z},
\end{eqnarray}
where $\omega^{2} = k_{x}^{2}+k_{y}^{2}+k_{z}^{2}$ are the eigenfrequencies of the field, with $k_i$ being the momentum in each spatial direction, and $\sigma = (k_x, k_y, k_z)$ stands for the set of quantum numbers.
The constant $N$ can be obtained via normalization condition
\begin{eqnarray}\label{eq04}
2\omega\int dV\phi_{\sigma}(w)\phi_{\sigma '}^{*}(w)=\delta_{\sigma\sigma '},
\end{eqnarray}
where the delta symbol in the r.h.s is understood as Kronecker delta for discrete quantum numbers and Dirac delta function for continuous quantum numbers. Note that we have introduced the notation $w=({\bf x},t)$ to specify spacetime coordinates. As we shall see later, the solution in Eq. \eqref{eq03} is modified when subjected to both the boundary conditions and the quasiperiodic condition, leading to discretization of one of the momenta.
Once subjecting the solution in Eq. \eqref{eq03} to the conditions considered, we must find the normalization constant $N$ by making use of Eq. \eqref{eq04}. This process makes possible to write the complete set of normalized solution and use it to calculate the Wightman function which is a crucial element to our computations. In order to construct the Wightman function, we may first promote the field to an operator and write it in terms of the positive and negative frequency normalized solutions, with coefficients of the expansion being the creation $a_{\sigma}^{\dagger}$ and annihilation $a_{\sigma}$ operators. Mathematically, we use the standard construction \cite{birrell1984quantum}
\begin{eqnarray}\label{eq05qf}
\hat{\phi}(w) = \sum_{\sigma}[a_{\sigma}\phi_{\sigma}(w)+a_{\sigma}^{\dagger}\phi_{\sigma}^{*}(w)],
\end{eqnarray}
where the creation and annihilation operators obey the commutation relation $[a_{\sigma},a_{\sigma'}^{\dagger}]=\delta_{\sigma\sigma'} $. We, thus, are able to obtain the Wightman function by taking into consideration the definition
\begin{eqnarray}\label{eq05wf}
W(w,w')=\langle 0|\hat{\phi}(w)\hat{\phi}(w')|0\rangle = \sum_{\sigma}\phi_{\sigma}(w)\phi_{\sigma}^{*}(w'),
\end{eqnarray}
where $|0\rangle$ is the vacuum state of the scalar field. Hence, the above equation provides the positive frequency Wightman function for the scalar field. In addition, the summation symbol in \eqref{eq05wf} stands for either integrals in the continuous quantum numbers or possible sums over discrete ones.
\subsection{Dirichlet boundary condition}
Firstly we are interested in considering Dirichlet boundary condition on the massless scalar field solution \eqref{eq03}. This means that by confining the field in a region of length $a$ between two perfectly reflecting parallel planes, perpendicular to the $x$-direction, we must have the condition
\begin{eqnarray}\label{eq06}
\phi({\bf x},t)|_{x=0}=\phi({\bf x},t)|_{x=a}=0.
\end{eqnarray}
Therefore, from Eqs. \eqref{eq03}, \eqref{eq06} and \eqref{eq04} we find that the complete set of normalized solutions in this case is given by
\begin{eqnarray}
\phi_{\sigma}({\bf x},t) = \dfrac{1}{\sqrt{4\pi^{2}a\omega_{n}}}\sin(k_{n}x)e^{-i\omega_{n}t+ik_{y}y+ik_{z}z},
\end{eqnarray}
where $\omega_{n}^2=k_{n}^{2}+k_{y}^{2}+k_{z}^2$ are the eigenfrequencies of the field, with momentum in the $x$-direction now discretized, that is, $k_{n}=\frac{n\pi}{a}$ $(n=1,2,3,\ldots)$. The set of quantum numbers in this case is $\sigma=(n, k_y, k_z)$. In Fig.\ref{figtwoparallelplanesA} we give an illustration of the setup described above. This configuration will also be used for the cases of Neumann and mixed boundary conditions later on.
\begin{figure}[h]
\includegraphics[scale=0.2]{twoparallelplanesA.pdf}
\caption{A point particle with mass $m$ and charge $g$ in the presence of two identical and perfectly reflecting parallel planes $p$ placed at $x=0$ and $x=a$, confining the field modes of a massless quantum scalar field.}\label{figtwoparallelplanesA}
\end{figure}
In order to obtain the corresponding Wightman function we make use of Eq. \eqref{eq05wf} with the summation symbol defined as
\begin{equation}
\sum_{\sigma}\equiv\sum_{n=1}^{\infty}\int_{-\infty}^{\infty}dk_{y}\int_{-\infty}^{\infty}dk_{z}.
\label{SSD}
\end{equation}
Consequently, the Wightman function becomes
\begin{eqnarray}\label{eq07}
W^{\text{(D)}} = \dfrac{1}{2\pi a}\sum_{n=1}^{\infty}\int_{0}^{\infty}dk k J_{0}(\Delta\ell k)\sin(k_{n}x)\sin(k_{n}x')\dfrac{e^{-i\omega_{n} \Delta t}}{\omega_{n}},
\end{eqnarray}
where $J_{\mu}(z)$ is the Bessel function \cite{gradshtein2007}, $\Delta\ell = \sqrt{\Delta y^{2}+\Delta z^{2}}$, $\Delta y = y-y'$, $\Delta z = z-z'$ and $\Delta t = t-t'$. Note that in the above expression we have used polar coordinates for the plane defined by the momentum variables $k_{y}$ and $k_{z}$, such that $k^{2}=k_{y}^{2}+k_{z}^{2}$ and $dk_{y}dk_{z}\rightarrow kdkd\theta$, which made possible to perform the angular integral leading to the Bessel function.
The sum in $n$ present in the Wightman function expression in Eq. \eqref{eq07} can be worked out by making use of the Abel-Plana formula \cite{saharian2007generalized}
\begin{eqnarray}\label{eqAbelPlana}
\sum_{n=0}^{\infty}F(n)=\dfrac{1}{2}F(0) + \int_{0}^{\infty}d\xi F(\xi) + i\int_{0}^{\infty}d\xi\dfrac{[F(i\xi)-F(-i\xi)]}{e^{2\pi \xi}-1}.
\end{eqnarray}
This is a very useful expression and it is often used, for example, in the Casimir energy computations (see Ref. \cite{saharian2007generalized} for more details). The function $F(n)$ in the present case is taken to be
\begin{eqnarray}
F(n) = \sin(k_{n}x)\sin(k_{n}x')\dfrac{e^{-i\omega_{n} \Delta t}}{\omega_{n}},
\label{funF}
\end{eqnarray}
where $F(0)=0$ and, consequently, the contribution from the first term in the r.h.s. of \eqref{eqAbelPlana} vanishes. Hence, by using the above expression in Eq. \eqref{eqAbelPlana}, after some algebraic manipulations, Eq. \eqref{eq07} can be written as
\begin{eqnarray}
W^{\text{(D)}}=W_{1}^{\text{(D)}}+W_{2}^{\text{(D)}},
\label{WFDcomp}
\end{eqnarray}
where for mathematical clarity and convenience, after the change of variables $s=\frac{\pi\xi}{a}$, we have defined
\begin{eqnarray}\label{eqW1dirichlet}
W_{1}^{\text{(D)}} =\dfrac{1}{2\pi{^2}}\int_{0}^{\infty}ds\sin(sx)\sin(sx')\int_{0}^{\infty}dk\dfrac{kJ_{0}(\Delta\ell k)e^{-i\Delta t\sqrt{k^{2}+s^{2}}}}{\sqrt{k^{2}+s^{2}}},
\end{eqnarray}
and
\begin{eqnarray}\label{eqW2dirichlet}
W_{2}^{\text{(D)}}=\dfrac{1}{\pi^{2}}\int_{0}^{\infty}dk k J_{0}(\Delta\ell k)\int_{k}^{\infty}ds\dfrac{\sin(isx)\sin(isx')}{e^{2as}-1}\dfrac{\cosh(\Delta t\sqrt{s^{2}-k^{2}})}{\sqrt{s^{2}-k^{2}}}.
\end{eqnarray}
Note that the expression in Eq. \eqref{eqW1dirichlet} stems from the integral in the second term in the r.h.s of the Abel-Plana formula \eqref{eqAbelPlana}, while Eq. \eqref{eqW2dirichlet} stems from the third term. In the latter, we have also used the identity
\begin{equation}
\sqrt{(\pm i s)^2 + k^2}=\left\{ \begin{array}{l}\pm i\sqrt{s^2 - k^2},\qquad\mathrm{for}\,\, s>k\,,\\
\;\;\;\sqrt{k^2 - s^2},\,\,\qquad\mathrm{for}\,\, s<k\,.
\end{array}\right.
\label{identity}
\end{equation}
The integrals in Eqs. \eqref{eqW1dirichlet} and \eqref{eqW2dirichlet} can be solved with the help of Refs. \cite{prudnikov1986integrals,gradshtein2007}, providing the expressions
\begin{eqnarray}\label{eqW1dirichletsolved}
W_{1}^{\text{(D)}} = \frac{1}{4\pi^2}\left\{\frac{1}{[\Delta x^2 + \alpha^2]} - \frac{1}{[\Delta\bar{x}^2 + \alpha^2]}\right\},
\end{eqnarray}
and
\begin{eqnarray}\label{eqW2dirichletsolved}
W_{2}^{\text{(D)}}=-W_{1}^{\text{(D)}} + \dfrac{1}{8\pi a\alpha}\left\{ \dfrac{\sinh\left(\frac{\pi\alpha}{a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta x}{a}\right) \right]}-\dfrac{\sinh\left(\frac{\pi\alpha}{a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta \bar{x}}{a}\right) \right]} \right\}.
\end{eqnarray}
Consequently, by substituting the two results above in Eq. \eqref{WFDcomp} we obtain
\begin{eqnarray}\label{eqW3dirichlet}
W^{\text{(D)}}=\dfrac{1}{8\pi a\alpha}\left\{ \dfrac{\sinh\left(\frac{\pi\alpha}{a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta x}{a}\right) \right]}-\dfrac{\sinh\left(\frac{\pi\alpha}{a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta \bar{x}}{a}\right) \right]} \right\},
\end{eqnarray}
where $\alpha^{2}=\Delta y^{2}+\Delta z^{2}-\Delta t^{2}$, with $\Delta x = x-x'$ and $\Delta\bar{x} = x+x'$. For our purposes, we can further simplify Eq. \eqref{eqW3dirichlet} by using the identity \cite{gradshtein2007}
\begin{eqnarray}\label{eq08}
\dfrac{\sinh\left(\frac{\pi\alpha}{a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta x}{a}\right) \right]} = \dfrac{2a\alpha}{\pi}\sum_{n=-\infty}^{\infty}\dfrac{1}{[(\Delta x - 2an)^{2}+\alpha^{2}]}.
\end{eqnarray}
This is particularly useful since we can separate the Minkowski contribution in a clearer way. This contribution is the term $n=0$ of the sum which is divergent in the coincidence limit $w'\rightarrow w$. As it is known, this divergent contribution must be subtracted from the the calculation of the velocity dispersion in order to obtain a renormalized quantity. Therefore, Eq. \eqref{eqW3dirichlet} takes the form
\begin{eqnarray}\label{eq09WFDirichlet}
W^{\text{(D)}}(w,w')=\dfrac{1}{4\pi^2}\sum_{n=-\infty}^{\infty}\left[f_{n}(\Delta r)-f_{n}(\Delta \bar{r})\right],
\end{eqnarray}
where
\begin{eqnarray}\label{eq09fn}
f_{n}(\Delta r) &=& \dfrac{1}{(\Delta x -2an)^2+\Delta y^2+\Delta z^2 - \Delta t^2},\nonumber\\
f_{n}(\Delta \bar{r}) &=& \dfrac{1}{(\Delta \bar{x} -2an)^2+\Delta y^2+\Delta z^2 - \Delta t^2}.
\end{eqnarray}
Hence, Eq. \eqref{eq09WFDirichlet} correspond to the positive frequency Wightman function in cartesian coordinates for the massless scalar field whose modes are restrict to obey Dirichlet boundary condition on the two perfectly reflecting parallel planes, placed at $x=0$ and $x=a$. Note that the Minkowski contribution comes from the term $n=0$ in the function $f_{n}(\Delta r) $. In contrast, the term $n=0$ is finite in the coincidence limit $w'\rightarrow w$ for the function $f_{n}(\Delta \bar{r})$. It is in fact the one plane contribution of the Wightman function for the Dirichlet boundary condition.
\subsection{Neumann boundary condition}
In the case we use Neumann boundary condition, the normal derivative of the field must vanish in the boundary. In this sense, considering two perfectly reflecting parallel planes, placed at $x=0$ and $x=a$, we have
\begin{eqnarray}\label{eq10}
\left[\partial_{x}\phi({\bf x},t)\right]|_{x=0}=\left[\partial_{x}\phi({\bf x},t)\right]|_{x=a}=0.
\end{eqnarray}
So, from Eqs. \eqref{eq03}, \eqref{eq10} and \eqref{eq04} we obtain the complete set of normalized solutions as follows
\begin{eqnarray}
\phi({\bf x},t) = c_{n}\cos(k_{n}x)e^{-i\omega_{n}t+ik_{y}y+ik_{z}z},
\end{eqnarray}
where the eigenfrequencies are given by $\omega_{n}^2=k_{n}^{2}+k_{y}^{2}+k_{z}^2$, with $k_{n}=\frac{n\pi}{a}$ $(n=0,1,2,\ldots)$ being the discretized momentum in the $x$-direction, and $\sigma = (n, k_y, k_z)$ is the set of quantum numbers. The normalization constante $c_n$ is written as
\begin{eqnarray}
c_{n} = \left\{
\begin{matrix}
\dfrac{1}{\sqrt{8\pi^{2}a\omega_{n}}}, & n = 0, \\
\dfrac{1}{\sqrt{4\pi^{2}a\omega_{n}}}, & n\geq 1.
\end{matrix}\right.
\end{eqnarray}
Similarly to the previous case, we can calculate the Wightman function by making use of Eq. \eqref{eq05wf}. The summation symbol in Eq. \eqref{eq05wf} is now defined as
\begin{equation}
\sum_{\sigma}\equiv\sum_{n=0}^{\infty}\int_{-\infty}^{\infty}dk_{y}\int_{-\infty}^{\infty}dk_{z}.
\label{SSD2}
\end{equation}
Consequently, the Wightman function takes the form
\begin{eqnarray}\label{eq11}
W^{\text{(N)}}&=& W_1^{\text{(N)}} + W^{\text{(D)}}\nonumber\\
&=&\dfrac{1}{4\pi^{2}a}\sum_{n=0}^{\infty}c^{*}_{n}\int_{-\infty}^{\infty}dk_{y}\int_{-\infty}^{\infty}dk_{z}\dfrac{\cos(k_{n}\Delta\bar{x})e^{-i\omega_{n}\Delta t+ik_{y}\Delta y+ik_{z}\Delta z}}{\omega_{n}} + W^{(D)},
\end{eqnarray}
where $c^{*}_{0}=1/2$ and $c^{*}_{n\geq 1}=1$. In the above expression, we have used trigonometric identities in order to be possible to identify two contributions, i.e., the one in the first term in the r.h.s and the one in the second term corresponding to the Wightman function for Dirichlet boundary condition, given by Eq. \eqref{eq09WFDirichlet}. Since the Dirichlet part has previously been calculated, we only need to focus in the first term in the r.h.s. of Eq. \eqref{eq11}. In the end, the Wightman function for the Neumann boundary condition case takes into consideration the sum of both terms in Eq. \eqref{eq11}.
Let us then work out the first term in the r.h.s. of Eq. \eqref{eq11}. This is possible with the help of the identity
\begin{eqnarray}
\dfrac{e^{-\omega_{n}\Delta\tau}}{\omega_{n}}=\dfrac{2}{\sqrt{\pi}}\int_{0}^{\infty}ds e^{-\omega_{n}^{2}s^{2}-\frac{\Delta\tau^{2}}{4s^{2}}},
\end{eqnarray}
where we have performed the Wick rotation, $\Delta\tau=i\Delta t$. The use of the above identity in $ W_1^{(N)} $, along with the help of Ref. \cite{{gradshtein2007}}, leads to
\begin{eqnarray}\label{eq12}
W^{\text{(N)}}_{1}&=&\dfrac{1}{2\pi a\alpha}\sum_{n=0}^{\infty}c_{n}^{*}\cos\left(\dfrac{\pi\Delta\bar{x}}{a}\right)e^{-\left(\frac{\pi\alpha}{a}\right)n}\nonumber\\
&=&\dfrac{1}{4\pi a\alpha}\dfrac{\sinh\left(\frac{\pi\alpha}{a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta \bar{x}}{a}\right) \right]}.
\end{eqnarray}
Hence, in view of Eqs. \eqref{eq11}, \eqref{eq12} and \eqref{eqW3dirichlet}, we conclude that
\begin{eqnarray}\label{eq09WFNeumann}
W^{\text{(N)}}(w,w')=\dfrac{1}{4\pi^2}\sum_{n=-\infty}^{\infty}\left[f_{n}(\Delta r)+f_{n}(\Delta \bar{r})\right],
\end{eqnarray}
where the functions defined in \eqref{eq09fn} have been used. This is the positive frequency Wightman function for the massless scalar field obeying Neumann boundary condition on the two perfectly reflecting parallel planes. Note
that the difference between the expressions for the Wightman function in the Dirichlet and Neumann boundary condition cases consists of only a changing of sign in the second term of Eq. \eqref{eq09WFDirichlet}. Again, the contribution $n=0$ of the sum in the first term in the r.h.s. of Eq. \eqref{eq09WFNeumann} is the divergent Minkowski contribution in the coincidence limit $w'\rightarrow w$, while in the second term is the one plane finite contribution. The latter has the opposite sign when compared to the Dirichlet boundary condition case.
\subsection{Mixed boundary condition}
In the mixed boundary condition case, the general solution of the field in Eq. \eqref{eq03} must obey Dirichlet condition in one plane and Neumann condition in the other. Thus, two configurations are possible on the first and second planes, that is, Dirichlet and Neumann (DN) as well as Neumann and Dirichlet (ND). For the configuration DN, respectively at $x=0$ and $x=a$, the condition obeyed by the field is given by
\begin{eqnarray}\label{eq13}
\phi({\bf x},t)|_{x=0}=\left[\partial_{x}\phi({\bf x},t)\right]|_{x=a}=0.
\end{eqnarray}
By applying the condition \eqref{eq13} on Eq. \eqref{eq03}, with the use of Eq. \eqref{eq04} afterwards, we obtain the complete set of normalized solutions
\begin{eqnarray}\label{eq14}
\phi_{\sigma}({\bf x},t) = \dfrac{1}{\sqrt{4\pi^{2}a\omega_{n}}}\sin(k_{n}x)e^{-i\omega_{n}t+ik_{y}y+ik_{z}z},
\end{eqnarray}
where the eigenfrequencies are now written as $\omega_{n}^2=k_{n}^{2}+k_{y}^{2}+k_{z}^2$, with $k_{n}=\frac{\pi(2n+1)}{2a}$ $(n=0,1,2, \ldots)$. Again, the momentum in the $x$-direction has been discretized as a consequence of Eq. \eqref{eq13} and the set of quantum numbers is specified by $\sigma = (n, k_y, k_z)$.
The Wightman function is computed through Eq. \eqref{eq05wf}, by making use of the normalized solution in Eq. \eqref{eq14} and
\begin{equation}
\sum_{\sigma}\equiv\sum_{n=0}^{\infty}\int_{-\infty}^{\infty}dk_{y}\int_{-\infty}^{\infty}dk_{z}.
\label{SSD3}
\end{equation}
Thereby, similarly to the Dirichlet condition case, it is possible to write the Wightman function in the form
\begin{eqnarray}\label{eq15}
W^{\text{(M)}} = \dfrac{1}{2\pi a}\sum_{n=0}^{\infty}\int_{0}^{\infty}dk k J_{0}(\Delta\ell k)\sin(k_{n}x)\sin(k_{n}x')\dfrac{e^{-i\omega_{n} \Delta t}}{\omega_{n}},
\end{eqnarray}
where we have used again polar coordinates for the plane defined by the momentum variables $k_y$ and $k_z$, following the same steps as in the Dirichlet condition case. By taking into account the structure of the allowed values for $k_{n}$ it is more convenient to use the Abel-Plana formula written in the form \cite{saharian2007generalized}
\begin{eqnarray}\label{eqAbelPlana2}
\sum_{n=0}^{\infty}F\left(n+\frac{1}{2}\right)= \int_{0}^{\infty}d\xi F(\xi) - i\int_{0}^{\infty}d\xi\dfrac{[F(i\xi)-F(-i\xi)]}{e^{2\pi \xi}+1},
\end{eqnarray}
where the function $F\left(n+\frac{1}{2}\right)$ is defined as in Eq. \eqref{funF} but now with $k_{n}=\frac{\pi(2n+1)}{2a}$. The Abel-Plana formula above allows us to write the Wightman function as
\begin{eqnarray}
W^{\text{(M)}}&=&W^{\text{(D)}}_{1} + W^{\text{(M)}}_{1}\nonumber\\
&=&W^{\text{(D)}}_{1} - \dfrac{1}{\pi^{2}}\int_{0}^{\infty}dk k J_{0}(\Delta\ell k)\int_{k}^{\infty}ds\dfrac{\sin(isx)\sin(isx')}{e^{2as}+1}\dfrac{\cosh(\Delta t\sqrt{s^{2}-k^{2}})}{\sqrt{s^{2}-k^{2}}} ,
\label{WFM}
\end{eqnarray}
where $W^{\text{(D)}}_{1}$ is given by Eq. \eqref{eqW1dirichletsolved} and we have again made use of the identity \eqref{identity}. Furthermore, the contribution in the second term in the r.h.s. of the above expression is found to be
\begin{eqnarray}
W^{\text{(M)}}_{1} = -W_{1}^{\text{(D)}} + \dfrac{1}{4\pi a\alpha}\left\{ \dfrac{\sinh\left(\frac{\pi\alpha}{2a}\right)\cos\left(\frac{\Delta x\pi}{2a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta x}{a}\right) \right]}-\dfrac{\sinh\left(\frac{\pi\alpha}{2a}\right)\cos\left(\frac{\Delta \bar{x}\pi}{2a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta \bar{x}}{a}\right) \right]} \right\},
\label{firstC}
\end{eqnarray}
where we have also again used the help of Refs. \cite{gradshtein2007, prudnikov1986integrals} to solve the integrals in $k$ and in $s$.
The complete Wightman function for the mixed boundary condition case is obtained from Eq. \eqref{WFM}, by using the expressions in Eqs. \eqref{eqW1dirichletsolved} and \eqref{firstC}. This gives
\begin{eqnarray}\label{eqWmistas}
W^{\text{(M)}}=\dfrac{1}{4\pi a\alpha}\left\{ \dfrac{\sinh\left(\frac{\pi\alpha}{2a}\right)\cos\left(\frac{\pi\Delta x}{2a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta x}{a}\right) \right]}-\dfrac{\sinh\left(\frac{\pi\alpha}{2a}\right)\cos\left(\frac{\pi\Delta \bar{x}}{2a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta \bar{x}}{a}\right) \right]} \right\}.
\end{eqnarray}
We can still put the above expression in a more convenient form, analogously to what has been done to the Dirichlet and Neumann condition cases. So, let us make use of the identity \cite{gradshtein2007}%
\begin{eqnarray}
\dfrac{\sinh\left(\frac{\pi\alpha}{2a}\right)\cos\left(\frac{\pi\Delta x}{2a}\right)}{\left[ \cosh\left(\frac{\pi\alpha}{a}\right)-\cos\left(\frac{\pi\Delta x}{a}\right) \right]} = \dfrac{\alpha a}{\pi}\sum_{n=-\infty}^{\infty}\dfrac{e^{i\pi n}}{[(\Delta x - 2an)^{2}+\alpha^{2}]}.
\end{eqnarray}
Consequently,
\begin{eqnarray}\label{eq16WFMistas}
W^{\text{(M)}}(w,w')=\dfrac{1}{4\pi^2}\sum_{n=-\infty}^{\infty}(-1)^{n}\left[f_{n}(\Delta r)-f_{n}(\Delta \bar{r})\right],
\end{eqnarray}
where the functions introduced in \eqref{eq09fn} have been used. It is important to observe that in Eq. \eqref{eq13} we use a configuration of boundary conditions such that Dirichlet and Neumann conditions are applied to the planes at $x=0$ and $x=a$, respectively. If the reverse configuration is used, that is, Neumann and Dirichlet such that $(\partial_{x}\phi)|_{x=0}=\phi|_{x=a}=0$, proceeding in a similar way as above, we obtain the same result shown in Eq. \eqref{eq16WFMistas}, but with the opposite sign in the second term in the r.h.s., which becomes positive.
The results in Eqs. \eqref{eq09WFDirichlet}, \eqref{eq09WFNeumann} and \eqref{eq16WFMistas} obtained for the Wightman function in the cases of Dirichlet, Neumann and mixed boundary conditions can be written as a general and compact expression, i.e.,
\begin{eqnarray}\label{FWDNM}
W^{\text{(i)}}(w,w')=\dfrac{1}{4\pi^{2}} \sum_{n=-\infty}^{\infty}\left[ \gamma_{n}^{\text{(i)}}f_{n}(\Delta r)+\delta_{n}^{\text{(i)}}f_{n}(\Delta \bar{r}) \right],
\end{eqnarray}
where we have conveniently defined
\begin{eqnarray}\label{Coefgd}
\gamma_{n}^{\text{(i)}}&=&\left[\gamma_{n}^{\text{(D)}},\gamma_{n}^{\text{(N)}},\gamma_{n}^{\text{(DN)}},\gamma_{n}^{\text{(ND)}}\right]=[+1,+1,(-1)^{n},(-1)^{n}],\nonumber \\
\delta_{n}^{\text{(i)}}&=&\left[\delta_{n}^{\text{(D)}},\delta_{n}^{\text{(N)}},\delta_{n}^{\text{(DN)}},\delta_{n}^{\text{(ND)}}\right]=[-1,+1,(-1)^{n+1},(-1)^{n}].
\end{eqnarray}
We can note that for all three boundary condition cases analyzed so far the contribution $n=0$ in the sum present in the first term in the r.h.s. of Eq. \eqref{FWDNM} correspond to the Minkowski contribution, which as we have already remarked, is divergent in the coincidence limit $w'\rightarrow w$. This term, as usual, must be subtracted from the physical observables. Moreover, the contribution $n=0$ coming from the second term in the r.h.s. of Eq. \eqref{FWDNM} provides the known expression for only one plane, placed at position $x=0$. The way we have organized the obtained Wightman functions in only one compact expression in Eq. \eqref{FWDNM} is very useful in the sense that it allows us to calculate at once the velocity dispersion for all three boundary condition cases since the derivative and integration operations in Eq. \eqref{eq21}, necessary to calculate the velocity dispersion, will only affect the functions $f_{n}$. Hence, after solving the successive operations acting on $f_{n}$ to obtain the velocity dispersion, we may just select the appropriate coefficients $\gamma_{n}^{(i)}$ and $\delta_{n}^{(i)}$ in order to specify which boundary condition result we are interested in.
\subsection{Quasiperiodic condition}
Finally, we now wish to consider a quasiperiodic condition, which generalizes the well known periodic and antiperiodic conditions by introducing a constant phase $\beta$, that is,
\begin{eqnarray}\label{eq17}
\phi(x,y,z,t) = e^{-2\pi\beta i}\phi(x+a,y,z,t).
\end{eqnarray}
The quasiperiodic parameter $\beta$ assumes values in the range $0\leq\beta <1$. Note that, if $\beta=0$ we restore the periodic condition whereas if $\beta=1/2$ we recover the antiperiodic one. Hence, the boundary condition above allows us to obtain a solution for the scalar field which includes besides the well known periodic and antiperiodic condition particular cases, also the cases for which $\beta\neq 0, 1/2$. As it is clear from Eq. \eqref{eq17} we consider that the compactification, of length $a$, is in the $x$-direction. An illustrative representation of this four-dimensional spacetime configuration is shown in Fig.\ref{figR3xS1}. The introduction of the quasiperiodic parameter $\beta$ may be thought of representing possible interactions in the system, as in the case of the well known Aharonov-Bohm effect \cite{de2012topological, kretzschmar1965must}.
\begin{figure}[h]
\includegraphics[scale=0.3]{R3xS1.pdf}
\caption{Illustrative representation of four-dimensional spacetime with a compactified spatial dimension. The spacetime is composed of a compactified spatial dimension $x$, $S_{1}$, and the tridimensional space $R^{3}$ of coordinates $t,y,z$.}\label{figR3xS1}
\end{figure}
By requiring the solution in Eq. \eqref{eq03} to obey the condition \eqref{eq17}, after making use of the normalization condition \eqref{eq04}, we find
\begin{eqnarray}
\phi_{\sigma}({\bf x},t) = \dfrac{1}{\sqrt{8\pi^{2}a\omega_{n}}}e^{-i\omega_{n}t+ik_{n}x+ik_{y}y+ik_{z}z},
\end{eqnarray}
where the eigenfrequencies are written as $\omega_{n}^2=k_{n}^{2}+k_{y}^{2}+k_{z}^2$, $k_{n} = \frac{2\pi(n+\beta)}{a}$ $(n=0,\pm 1,\pm 2,\ldots)$ and the set of quantum numbers is represented by $\sigma = (n, k_y, k_z)$. Similarly to the previous computations, one is able to calculate the Wightman function through Eq. \eqref{eq05wf}, with
\begin{equation}
\sum_{\sigma}\equiv\sum_{n=-\infty}^{\infty}\int_{-\infty}^{\infty}dk_{y}\int_{-\infty}^{\infty}dk_{z}.
\label{SSD3QP}
\end{equation}
Next, we again adopt polar coordinates in the $(k_{y},k_{z})$-plane, such that $dk_{y}dk_{z}\rightarrow kdkd\theta$ and $k^{2}=k_{y}^{2}+k_{z}^{2}$. After solving the angular part we found
\begin{eqnarray}
W(w,w')&=&\dfrac{1}{4\pi a}\sum_{n=-\infty}^{\infty}e^{ik_{n}\Delta x}\int_{0}^{\infty}dk \dfrac{kJ_{0}(\Delta r k)e^{-i\omega_{n}\Delta t}}{\omega_{n}}\nonumber\\
&=&\dfrac{1}{4\pi a}\sum_{n=-\infty}^{\infty}e^{ik_{n}\Delta x-\alpha |k_{n}|},
\label{WFQPn}
\end{eqnarray}
where we have used the help of Ref. \cite{{prudnikov1986integrals}} to solve the integral in $k$. By splitting the summation in $n$ in two parts in order to eliminate the modulus in $k_n$, with the help of Ref. \cite{gradshtein2007}, we can further simplify the expression in the second line of Eq. \eqref{WFQPn} and write it in the convenient form
\begin{eqnarray}\label{eq18WFQuasiperiodicas}
W(w,w')=\dfrac{1}{4\pi^{2}}\sum_{n=-\infty}^{\infty}e^{2\pi\beta ni}g_{n}(\Delta r),
\end{eqnarray}
where
\begin{eqnarray}\label{eq18WFQuasiperiodicas2}
g_n(\Delta r)= \frac{1}{\left[\left(\Delta x - an\right)^2 + \Delta y^2 + \Delta z^2 - \Delta t^2\right]}.
\end{eqnarray}
This is the positive frequency Wightman function for the massless scalar field subjected to a quasiperiodic condition. The Minkowski divergent contribution can now be easily separated to be subtracted in the renormalization process, leading to finite renormalized velocity dispersions. This term again arises from the $n=0$ contribution of the above sum.
Note that in the periodic case, $\beta=0$, Eq. \eqref{eq18WFQuasiperiodicas} corresponds to a spacetime of topology $S^{1}\times R^{3}$, that is, a compactified direction in a circle and a three-dimensional space of coordinates $(t,y,z)$ with $t\geq 0$, $-\infty<y<\infty$ and $-\infty<z<\infty$. This spacetime configuration is shown in Fig.\ref{figR3xS1}. Then, the case $\beta\neq 0$ can be thought of as a generalization, which we called modified $S^{1}\times R^{3}$ spacetime, because of the phase introduced by the quasiperiodic parameter $\beta$.
With the convenient form for the Wightman functions obtained in this section we can proceed to the next section to calculate the renormalized velocity dispersion in each condition scenario.
\section{Velocity despersions}\label{SecVelDis}
\subsection{General expression}
Let us now analyze the dynamics of the point particle coupled to the massless scalar field. Thus, by varying the action \eqref{eq01} with respect to the position, we obtain the following expression for particle's velocity \cite{camargo2019vacuum,camargo2018vacuum,Camargo:2020fxp, de2014quantum}:
\begin{eqnarray}\label{eq19}
v_{i}(\tau,{\bf x})=-\frac{g}{m}\int_{0}^{\tau}dt\frac{\partial\phi({\bf x},t)}{\partial x_{i}},
\end{eqnarray}
where $i=(x,y,z)$ and we have considered a null initial velocity, that is, $v_{i}(t=0)=0$. This is a classical equation, but if we are interested in studying the QBM induced by quantum vacuum fluctuations we should promote the scalar field to an operator as in Eq. \eqref{eq05qf} which, consequently, leads to the quantization of Eq. \eqref{eq19} as well. As a result, we note that $\langle 0|v_{i}|0\rangle\equiv \langle v_{i}\rangle=0$, that is, the velocity mean value of the particle due the quantum vacuum fluctuations vanishes since, by definition, $a|0\rangle =0$ and $\langle 0|a^{\dagger}=0$.
Although the velocity mean value vanishes, the quantum vacuum fluctuations on the velocity can be calculated through the following expression for the renormalized velocity dispersion \cite{mota2020induced, ferreira2022quantum}:
\begin{eqnarray}\label{eq20}
\langle(\Delta v_{i})^{2}\rangle_{\text{ren}} = \lim_{x\rightarrow x'}\left[ \langle v_{i}(x)v_{i}(x')\rangle -\langle v_{i}(x)v_{i}(x')\rangle_{\text{div}} \right],
\end{eqnarray}
where we have introduced the notation $\langle 0|(\ldots)|0\rangle\equiv \langle(\ldots)\rangle$. Note that the Minkowski divergent contribution has been subtracted from the velocity dispersion, something that is standard in the renormalization process.
From Eqs. \eqref{eq19} and \eqref{eq20} the renormalized velocity dispersion is formally given by
\begin{eqnarray}\label{eq21}
\langle(\Delta v_{i})^{2}\rangle_{\text{ren}} = \frac{g^2}{2m^2}\int_{0}^{\tau}dt'\int_{0}^{\tau}dt\frac{\partial^{2}G_{\text{ren}}^{(1)}(x,x')}{\partial x_{i}'\partial x_{i}},
\end{eqnarray}
where $G^{(1)}(x,x')=\langle\{\hat{\phi}(x),\hat{\phi}(x')\}\rangle$ is the Hadamard function that can be obtained from the positive frequency Wightman function by the relation $G^{(1)}(x,x')=2\text{Re}\,W(x,x')$ \cite{fulling1989aspects}. We should point out that the renormalized Hadamard function in Eq. \eqref{eq21} is obtained by subtracting the divergent Minkowski contribution present in the Wightman function already discussed in the previous section. We should also point out that in order to establish the above expression we have symmetrized the fields, a common procedure adopted in quantum field theory \cite{camargo2018vacuum,gour1999will}.
Next, we shall use the Wightman functions obtained in the previous section jointly with Eq. \eqref{eq21} to calculate the renormalized particle velocity dispersion corresponding to each boundary condition.
\subsection{Dirichlet, Neumann and mixed boundary conditions}\label{veldispDNM}
Let us start by taking into consideration the velocity dispersion induced by Dirichlet, Neumann and mixed boundary conditions. To do this, we first consider the direction perpendicular to the planes, i.e., the $x$-direction. Thereby, from Eqs. \eqref{FWDNM} and \eqref{eq21}, after carrying out the integrals and derivatives operations, we find
\begin{eqnarray}\label{eq22}
\langle(\Delta v_{x})^{2}\rangle^{\text{(i)}}_{\text{ren}} &=& -\frac{g^{2}}{16\pi^{2}m^{2}a^{2}}\left[ 2\sum_{n=1}^{\infty} \gamma_{n}^{\text{(i)}} R(n,\tau_{a}) - \sum_{n=-\infty}^{\infty}\delta_{n}^{\text{(i)}}R(x_{a}-n,\tau_{a})\right],
\end{eqnarray}
where we have conveniently defined the dimensionless parameters $x_{a}=x/a$, $\tau_{a}=\tau/a$ and the function
\begin{eqnarray}\label{eq220a}
R(r,\tau_{a}) = P(r,\tau_{a})+Q(r,\tau_{a}),
\end{eqnarray}
with
\begin{eqnarray}\label{eq22a}
P(r,\tau_{a}) &=& \frac{2\tau_{a}^{2}}{r^{2}(4r^{2}-\tau_{a}^{2})},\nonumber\\
Q(r,\tau_{a}) &=& \frac{\tau_{a}}{2r^{3}}\ln\left(\frac{2r+\tau_{a}}{2r-\tau_{a}}\right)^{2}.
\end{eqnarray}
Note that in order to evaluate the integrals in Eq. \eqref{eq21} we have used the identity \cite{de2014quantum,camargo2018vacuum}
\begin{eqnarray}\label{eq22b}
\int_{0}^{\tau}dt'\int_{0}^{\tau}dt f(|t-t'|) = 2\int_{0}^{\tau}d\xi(\tau-\xi)f(\xi).
\end{eqnarray}
The plot for Eq. \eqref{eq22} is shown in Fig.\ref{fig01}, for distinct boundary conditions. In particular, the plot for mixed boundary condition of types DN and ND coincide when one takes the value $x_a=0.5$ and differ for other values.
\begin{figure}
\includegraphics[scale=0.5]{D_perpendicular_direction.pdf}
\includegraphics[scale=0.5]{N_perpendicular_direction.pdf}
\includegraphics[scale=0.5]{M_perpendicular_direction.pdf}
\caption{Graph behavior of the perpendicular velocity dispersion for (a) Dirichlet (D), (b) Neumann (N) and (c) mixed (DN, ND) boundary conditions. Here we have considered the curves in units of $\langle(\Delta v_{x})^{2}\rangle^{\text{(i)}}=\langle(\Delta v_{x})^{2}\rangle^{\text{(i)}}_{\textrm{ren}}\left(\frac{ma}{g}\right)^{2}$. }\label{fig01}
\end{figure}
Similarly, for the velocity dispersion parallel to the planes we obtain
\begin{eqnarray}\label{eq23}
\langle(\Delta v_{y})^{2}\rangle^{\text{(i)}}_{\text{ren}} &=& \dfrac{g^{2}}{32\pi^{2}m^{2}a^{2}}\left\{ 2\sum_{n=1}^{\infty}\gamma_{n}^{\text{(i)}}Q(n,\tau_{a}) + \sum_{n=-\infty}^{\infty}\delta_{n}^{\text{(i)}}Q(x_{a}-n,\tau_{a}) \right\},
\end{eqnarray}
where we have used the function $Q(r,\tau_a)$ defined in Eq. \eqref{eq22a}. The same result is obtained for the $z$ component of the velocity dispersion, also parallel to the planes. The behavior of this expression is depicted in Fig.\ref{fig02}. Again, in the case of mixed boundary condition, the plot shows that the curves for DN and ND coincide for $x_a=0.5$ and differ when taking other values.
\begin{figure}
\includegraphics[scale=0.5]{D_parallel_direction.pdf}
\includegraphics[scale=0.5]{N_parallel_direction.pdf}
\includegraphics[scale=0.5]{M_parallel_direction.pdf}
\caption{Graph behavior of the parallel velocity dispersion curves for (a) Dirichlet (D), (b) Neumann (N) and (c) mixed (DN, ND) boundary conditions. Here we have considered the curves in units of $\langle(\Delta v_{y})^{2}\rangle^{\text{(i)}}=\langle(\Delta v_{y})^{2}\rangle^{\text{(i)}}_{\textrm{ren}}\left(\frac{ma}{g}\right)^{2}$. Note that the shown peaks represent divergent points.}\label{fig02}
\end{figure}
It should be observed that, for Dirichlet boundary condition, the $n=0$ term of the expressions \eqref{eq22} and \eqref{eq23} corresponds to the one plane contribution for the velocity dispersions which has already been investigated in Ref. \cite{camargo2018vacuum}. This contribution is obtained from the second term in the r.h.s of Eq. \eqref{eq09WFDirichlet} for $n=0$. The latter, of course, is the Wightman function for Dirichlet boundary condition considering only one plane. Note that the one plane contribution for the velocity dispersions in the case of Neumann boundary condition is the same as the one for Dirichlet boundary condition, but with the opposite sign. Note also that the mixed boundary condition is not applicable for one single plane.
We now want to discuss the divergencies present in the expressions \eqref{eq22} and \eqref{eq23}. The first of them are the usual divergencies for points on the planes, at $x_{a}=0$ and $x_{a}=1$. They come from the second term in the r.h.s of Eqs. \eqref{eq22} and \eqref{eq23} when $n=0$ and $n=1$, respectively. In addition, for $x_a\neq 0,1$, there also exist divergencies associated with the time, in a round trip, a light signal takes to travel from the planes to a point located at $x_{a}$ \cite{yu2004vacuum,yu2004brownian}. Mathematically this is given by $\tau_{a}=2|x_{a}-n|$, which tells us that each mode of the field contributes with a divergency. Finally, there are also position independent divergencies in the form of $\tau=2na$ coming from the first term in the r.h.s of Eqs. \eqref{eq22} and \eqref{eq23}. These divergencies represent an increasing number (with the field modes) of round trips from one plane to the other taken by a light signal. All these divergencies can be seen in the plots present in Figs.\ref{fig01} and \ref{fig02} for each boundary condition considered so far. For instance, in the plot for Dirichlet boundary condition shown in Fig.\ref{fig01}, the position independent divergency takes place for $\tau_a=2$ when $n=1$, while the position dependent divergencies take place for $\tau_a=1.4$ (when $n=0$) and $\tau_a=0.6$ (when $n=1$). Note that a larger range for $\tau_a$ would show additional divergencies. Note also that the same analysis can be reached for other values of $x_a$. In Ref. \cite{de2014quantum}, similar divergences have been studied in a one-dimensional model. The authors have shown that by assuming that the particle position fluctuates according to a Gaussian distribution the divergencies are smeared out. It has also been shown in Refs. \cite{de2016probing} and \cite{camargo2018vacuum} that implementation of switching functions can eliminate these typical divergences.
Let us now turn to the investigation of the behavior of the expressions \eqref{eq22} and \eqref{eq23} when $\tau_{a}\gg 1$ and $\tau_{a}\ll 1$, that is, for late and short time regimes, respectively. We start with the short time regime which indicates the behavior of the system in its initial moments of observation. In this sense, by considering the results of Appendix \ref{AppBDNMshorttime}, from Eqs. \eqref{eq22} and \eqref{eq23}, we obtain, for the perpendicular direction,
\begin{eqnarray}\label{eq24}
\langle(\Delta v_{x})^{2}\rangle^{\text{(J)}}_{\text{ren}}\simeq - \dfrac{g^{2}\tau_{a}^{2}}{16\pi^{2}m^{2}a^{2}}\left\{ 3\zeta(4)-\delta^{(J)}\frac{\pi^{4}}{2}[2+\cos(2\pi x_{a})]\csc^{4}(\pi x_{a})\right\}
\end{eqnarray}
and
\begin{eqnarray}\label{eq25}
\langle(\Delta v_{x})^{2}\rangle^{\text{(M)}}_{\text{ren}} \simeq \dfrac{g^{2}\tau_{a}^{2}}{128\pi^{2}m^{2}a^{2}}\left\{ 21\zeta(4)-\delta^{(M)}\pi^{4}[11+\cos(2\pi x_{a})]\cot(\pi x_{a})\csc^{3}(\pi x_{a})\right\},
\end{eqnarray}
while for the parallel direction we have
\begin{eqnarray}\label{eq26}
\langle(\Delta v_{y})^{2}\rangle^{\text{(J)}}_{\text{ren}}\simeq \dfrac{g^{2}\tau_{a}^{2}}{32\pi^{2}m^{2}a^{2}}\left\{ 2\zeta(4)+\delta^{(J)}\frac{\pi^{4}}{3}[2+\cos(2\pi x_{a})]\csc^{4}(\pi x_{a})\right\}
\end{eqnarray}
and
\begin{eqnarray}\label{eq27}
\langle(\Delta v_{y})^{2}\rangle^{\text{(M)}}_{\text{ren}} \simeq -\dfrac{g^{2}\tau_{a}^{2}}{128\pi^{2}m^{2}a^{2}}\left\{ 7\zeta(4)+\delta^{(M)}\frac{\pi^{4}}{3}[11+\cos(2\pi x_{a})]\cot(\pi x_{a})\csc^{3}(\pi x_{a})\right\},
\end{eqnarray}
where $\delta^{\text{(J)}} = [\delta^{\text{(D)}},\delta^{\text{(N)}}]=[-1,+1]$ and $\delta^{\text{(M)}} = [\delta^{\text{(DN)}},\delta^{\text{(ND)}}]=[+1,-1]$. We can see that all the expressions above for the short time regime are of order $\tau_a^2$, the leading order of Eqs. \eqref{apBeq07}, \eqref{apBeq09}, \eqref{apBeq18} and \eqref{apBeq20} considered to perform the analysis. Note that in the expressions above only the divergencies on the planes, at $x_{a}=(0,1)$, are preserved.
On the other hand, similarly to the classical Brownian motion, the late time regime, that is, $\tau_{a}\gg 1$, give us an ideia about the behavior of the system close to an equilibrium state reached between the particle and the environment, which in our case is the quantum fluctuating vacuum of the scalar field. Thereby, based on the results of Appendix \ref{AppADNMlatetime} we obtain, for the perpendicular direction,
\begin{eqnarray}\label{eq28}
\langle(\Delta v_{x})^{2}\rangle^{\text{(J)}}_{\text{ren}} \simeq -\frac{g^{2}}{8\pi^{2}m^{2}a^{2}}\left[\frac{\pi^{2}}{3}+\frac{4}{3\tau_{a}^{2}}-\delta^{(J)}\pi^{2}\csc^{2}(\pi x_{a})\right]
\end{eqnarray}
and
\begin{eqnarray}\label{eq29}
\langle(\Delta v_{x})^{2}\rangle^{\text{(M)}}_{\text{ren}} \simeq \frac{g^{2}}{8\pi^{2}m^{2}a^{2}}\left[\frac{\pi^{2}}{6}-\frac{4}{3\tau_{a}^{2}}-\delta^{(M)}\pi^{2}\cot(\pi x_{a})\csc(\pi x_{a})\right],
\end{eqnarray}
while for the parallel direction we have
\begin{eqnarray}\label{eq30}
\langle(\Delta v_{y})^{2}\rangle^{\text{(J)}}_{\text{ren}} \simeq \frac{g^{2}}{8\pi^{2}m^{2}a^{2}}\left[\frac{\pi^{2}}{3}-\frac{4}{3\tau_{a}^{2}}+\delta^{(J)}\pi^{2}\csc^{2}(\pi x_{a})\right]
\end{eqnarray}
and
\begin{eqnarray}\label{eq31}
\langle(\Delta v_{y})^{2}\rangle^{\text{(M)}}_{\text{ren}} \simeq -\frac{g^{2}}{8\pi^{2}m^{2}a^{2}}\left[\frac{\pi^{2}}{6}+\frac{4}{3\tau_{a}^{2}}+\delta^{(M)}\pi^{2}\cot(\pi x_{a})\csc(\pi x_{a})\right],
\end{eqnarray}
where the coefficients $\delta^{\text{(J)}}$ and $\delta^{\text{(M)}}$ have already been previously defined below Eq. \eqref{eq27}. From Eqs. \eqref{eq28}, \eqref{eq29}, \eqref{eq30} and \eqref{eq31}, we see that all the expressions have a term of order $4/3\tau_{a}^{2}$, which is negligible for large time values so that the remainder terms are the dominant ones. In particular, the position dependent term depends on the boundary condition used and also preserves the divergencies on the planes located at $x_{a}=(0,1)$. The velocity dispersions in Eqs. \eqref{eq22} and \eqref{eq23} becoming time independent at much later times is analogous to what happens in the classical Brownian motion for a point particle immersed in a fluid at finite temperature, which also becomes time independent in this regime.
We can show that the expressions for the late time regime obtained above, for Dirichlet boundary condition, when $x_{a}\ll 1$, is consistent with the result presented in Ref. \cite{camargo2018vacuum} where the authors considered a single plane. Thereby, expanding Eqs. \eqref{eq28} and \eqref{eq30} for $x_{a}\ll 1$ we found
\begin{equation}\label{oneplaneasymptotics}
\langle(\Delta v_{x})^{2}\rangle^{\text{(D)}}_{\text{ren}} = \langle(\Delta v_{y})^{2}\rangle^{\text{(D)}}_{\text{ren}}\simeq-\frac{g^{2}}{8\pi^{2}m^{2}x^{2}},
\end{equation}
which is exactly Eq. (4.3) of Ref. \cite{camargo2018vacuum}. The limit $x_{a}\ll 1$ is equivalent to say that the plane placed at $x=a$ is moved far away from the plane at $x=0$, ideally to infinity (see Fig.\ref{figtwoparallelplanesB}). Consequently, the infinitely distant plane has no effect on the particle. Thus, the resulting scenario is a point particle in the presence of a single plane, placed at $x=0$, which is one of the configurations studied in Ref. \cite{camargo2018vacuum} for the late time regime.
In the case of mixed boundary condition we can observe a similar situation in the limit $x_{a}\ll 1$, that is, $a\rightarrow\infty$. In this case, we can show that the expressions for the velocity dispersion, Eqs. \eqref{eq25}, \eqref{eq27}, \eqref{eq29} and \eqref{eq31}, correspond to either Dirichlet or Neumann boundary condition only, depending whether we consider the DN or ND configuration on the planes. For instance, let us consider the DN configuration, where $\delta^{(M)}\equiv\delta^{(DN)}=+1$. This is the configuration in which Dirichlet and Neumann boundary condition are applied to the planes placed at $x=0$ and $x=a$, respectively. In this sense, taking the limit $a\rightarrow\infty$ in the aforementioned expressions we obtain the result for Dirichlet boundary condition in the corresponding limit, namely, Eqs. \eqref{eq24}, \eqref{eq26}, \eqref{eq28} and \eqref{eq30}, with $\delta^{(J)}\equiv\delta^{(D)}=-1$. The explanation is that once we move the plane placed at $x=a$ to infinity, only the plane with Dirichlet boundary condition, at $x=0$, produces some effect on the particle. The argument for the ND configuration is similar.
\begin{figure}[h]
\includegraphics[scale=0.2]{twoparallelplanesB.pdf}
\caption{If we move the plane ($p$) at $x=a$ to infinity everything works like if the plane at $x=a$ did no exist, that is, the resulting configuration is equivalent to a point particle in the presence of a single plane.}\label{figtwoparallelplanesB}
\end{figure}
Finally, to end this subsection we would like to make a brief comment about possible negative values that the velocity dispersions can take. This can be seen from Eq. \eqref{eq20}, which consists of a diference between the dispersion in the presence of two parallel planes and the dispersion without planes, which is divergent. So, a negative value indicates that the presence of the planes creates a reduction in the velocity dispersion, as argued in Ref. \cite{yu2004vacuum}.
\subsection{Quasiperiodic condition}
In order to obtain velocity dispersions corresponding to the quasiperiodic condition in Eq. \eqref{eq17} we make use of Eqs. \eqref{eq18WFQuasiperiodicas} and \eqref{eq21}. So, for the velocity dispersion in the $x$-direction, that is, the compactified direction, we find
\begin{eqnarray}\label{eq32}
\langle(\Delta v_{x})^{2}\rangle^{\beta}_{\text{ren}} = -\dfrac{g^{2}}{\pi^{2}m^{2}a^{2}}\sum_{n=1}^{\infty}U(n,\beta,\tau_{a}),
\end{eqnarray}
while for the $y$-direction (or $z$), the uncompactified direction, we have
\begin{eqnarray}\label{eq33}
\langle(\Delta v_{y})^{2}\rangle^{\beta}_{\text{ren}} = \dfrac{g^{2}}{2\pi^{2}m^{2}a^{2}}\sum_{n=1}^{\infty}T(n,\beta,\tau_{a}),
\end{eqnarray}
where we have defined the function
\begin{eqnarray}\label{eq34}
U(n,\beta,\tau_{a}) = S(n,\beta,\tau_{a}) + T(n,\beta,\tau_{a}),
\end{eqnarray}
with
\begin{eqnarray}\label{eq34a}
S(n,\beta,\tau_{a}) &=& \frac{\tau_{a}^{2}\cos(2\pi\beta n)}{n^{2}(n^{2}-\tau_{a}^{2})}
\end{eqnarray}
and
\begin{eqnarray}\label{eq34b}
T(n,\beta,\tau_{a}) &=& \frac{\tau_{a}\cos(2\pi\beta n)}{2n^{3}}\ln\left(\frac{n+\tau_{a}}{n-\tau_{a}}\right)^{2}.
\end{eqnarray}
Note that to perform the integrals that have lead to the above expressions we have used again the identity \eqref{eq22b}. Similar to the previous cases, the compactification parameter $a$ provides a natural scale to the system, so that we are able to define the dimensionless time parameter $\tau_{a}$. It is important to call attention to the fact that the quasiperiodic condition has the particular periodic and antiperiodic condition cases given by, respectively, $\beta=0$ and $\beta=1/2$. From Eqs. \eqref{eq32} and \eqref{eq33} we observe that the expressions depend exclusively on the quasiperiodic parameter $\beta$, dimensionless time $\tau_{a}$ and length $a$. The graph behavior for these expressions is shown in the Fig.\ref{figcm01}.
A similar result has been obtained in Ref. \cite{Bessa:2019aar} for a point particle in the presence of a quantized electromagnetic field in a spacetime with spatial section of nontrivial topology, known as $E_{16}$ or slab topology, which is essentially defined by Eqs. \eqref{eq18WFQuasiperiodicas} and \eqref{eq18WFQuasiperiodicas2} for the periodic case $(\beta=0)$.
\begin{figure}
\includegraphics[scale=0.5]{QPBC_compactified_direction1.pdf}
\includegraphics[scale=0.5]{QPBC_compactified_direction2.pdf}
\includegraphics[scale=0.5]{QPBC_uncompactified_direction.pdf}
\caption{Graph of the velocity dispersion in the (a)--(b) compactified and (c) uncompactified direction for the Quasiperiodic condition. Here we have considered the curves in units of $\langle(\Delta v_{x,y})^{2}\rangle^{\beta}=\langle(\Delta v_{x,y})^{2}\rangle^{\beta}_{\textrm{ren}}\left(\frac{ma}{g}\right)^{2}$. Note that the shown peaks represent divergent points.}\label{figcm01}
\end{figure}
Differently from Dirichlet, Neumann and mixed boundary conditions, the expressions \eqref{eq32} and \eqref{eq33} do not have any dependency with the spatial coordinate $x$. The reason is that the quasiperiodic condition does not restrict the modes to a particular region as it happens to the parallel planes case. A spacetime in which one of the directions has a finite length $a$, as it is our case, makes possible to the modes to extend themselves throughout the whole $x$-coordinate. In contrast, in the Dirichlet, Neumann and mixed boundary condition cases the modes are also confined into a region of length $a$, but the $x$-component of the field does not exist outside this finite region. As we have already mentioned, the parallel planes break the homogeneity of the spatial section of the spacetime.
Our expressions reveal that for some values of the time parameter $\tau_{a}$ we obtain divergent results to the velocity dispersion. Although the quasiperioric system is different, the interpretation of these singularities have similarity to those of the parallel planes case. Specifically, these divergencies occur for integer values of time, that is, $\tau_{a}=n$, as we can see from Eqs. \eqref{eq32} and \eqref{eq33}. The latter are plotted in Fig.\ref{figcm01} where, for the time range considered, there exist divergencies at $\tau_a=1$ and $\tau_a=2$. These divergencies are similar to the ones arising from the time a light signal takes to travel from a point $x_a$ to the planes in a round trip. However, in the quasiperiodic condition case, it is more intuitive to imagine circumferences of length $a$, so that $n$ values represent complete turns in the ciclic path. Then, we may understand the divergences in this case as due to the time taken by a light signal to travel an increasing number of ciclic paths of length $a$. As reported in Ref. \cite{Lemos:2020ogj}, where the authors analyzed the periodic case, the origin of such integer divergencies is a consequences of the spacetime topology, namely, $S^1\times R^3$.
Similarly to what has been done in the previous subsection, let us obtain the expressions for the short and late time regimes, that is, the velocity dispersions for the asymptotic time limits $\tau_{a}\ll 1$ and $\tau_{a}\gg 1$, respectively. From the results of Appendix \ref{AppBQPshorttime}, for the short time regime, the velocity dispersion in the $x$-direction is written as
\begin{eqnarray}\label{eq35}
\langle(\Delta v_{x})^{2}\rangle^{\beta}_{\text{ren}} \simeq \frac{g^{2}\tau_{a}^{2}\pi^{2}}{ m^{2}a^{2}}B_{4}(\beta),
\end{eqnarray}
where $B_{n}(z)$ is the Bernoulli polynomial of order $n$ in the $z$ variable \cite{gradshtein2007}. The periodic (p) and antiperiodic (ap) cases are obtained as special cases of Eq. \eqref{eq35} for $\beta=0$ and $\beta=1/2$, respectively. These are given by
\begin{eqnarray}\label{eq35a}
\langle(\Delta v_{x})^{2}\rangle^{\text{(p)}}_{\text{ren}} \simeq -\frac{g^{2}\tau_{a}^{2}\pi^{2}}{30 m^{2}a^{2}}
\end{eqnarray}
and
\begin{eqnarray}\label{eq35b}
\langle(\Delta v_{x})^{2}\rangle^{\text{(ap)}}_{\text{ren}} \simeq\frac{7g^{2}\tau_{a}^{2}\pi^{2}}{240m^{2}a^{2}}.
\end{eqnarray}
Likewise, for the velocity dispersion in the $y$ (or $z$) direction, we find
\begin{eqnarray}\label{eq36}
\langle(\Delta v_{y})^{2}\rangle_{\text{ren}} \simeq -\dfrac{g^{2}\tau_{a}^{2}\pi^{2}}{3m^{2}a^{2}}B_{4}(\beta),
\end{eqnarray}
with
\begin{eqnarray}\label{eq36a}
\langle(\Delta v_{y})^{2}\rangle^{\text{(p)}}_{\text{ren}} \simeq \dfrac{g^{2}\tau_{a}^{2}\pi^{2}}{90m^{2}a^{2}}
\end{eqnarray}
and
\begin{eqnarray}\label{eq36b}
\langle(\Delta v_{y})^{2}\rangle^{\text{(ap)}}_{\text{ren}} \simeq-\dfrac{7g^{2}\tau_{a}^{2}\pi^{2}}{720m^{2}a^{2}},
\end{eqnarray}
for the periodic and antiperiodic cases. It is interesting to note that, similar to the cases studied in Section \ref{veldispDNM}, our expressions here also show a second order time dependency. From Eqs. \eqref{eq35} and \eqref{eq36}, we observe that the dispersion for the uncompactified direction is $-1/3$ of the result for the compactified one.
Furthermore, the sign of the velocity dispersions in the short time regime is defined by the Bernoulli polynomials. In fact, as we can see in Fig.\ref{figBernoulli}, $B_{4}(\beta)$ assumes positive values in the range $r_{-}\leq\beta\leq r_{+}$, but it is negative for any other values of $\beta$, where $r_{\pm}=[1\pm (1-4n)^{1/2})]/2$, with $n=1/\sqrt{30}$, are the physical roots taking into consideration the condition $0\leq\beta<1$.
For the periodic case ($\beta=0$), the compactified and uncompactified velocity dispersions achieve their minimum and maximum value, respectively, Eqs. \eqref{eq35a} and \eqref{eq36a}. On the other hand, in the antiperiodic case, Eqs. \eqref{eq35b} and \eqref{eq36b}, the opposite occurs.
\begin{figure}[h]
\includegraphics[scale=0.5]{QPBC_Bernoulli.pdf}
\caption{Bernoulli polinomials $B_{2}(\beta)$, solid line, and $B_{4}(\beta)$, dashed line, as functions of the quasiperiodic parameter $\beta$.}\label{figBernoulli}
\end{figure}
We turn now to the analysis of the late time regime, that is, $\tau_{a}\gg 1$. Hence, by making use of the results in Appendix \ref{AppAQPlatetime}, for the $x$-direction, we have
\begin{eqnarray}\label{eq37}
\langle(\Delta v_{x})^{2}\rangle^{\beta}_{\text{ren}} \simeq -\dfrac{g^{2}}{\pi^{2}m^{2}a^{2}}\left[\pi^{2}B_{2}(\beta)+\frac{1}{6\tau_{a}^{2}}\right],
\end{eqnarray}
with
\begin{eqnarray}\label{eq37a}
\langle(\Delta v_{x})^{2}\rangle^{\text{(p)}}_{\text{ren}} \simeq-\dfrac{g^{2}}{6\pi^{2}m^{2}a^{2}}\left[\pi^{2}+\frac{1}{\tau_{a}^{2}}\right]
\end{eqnarray}
and
\begin{eqnarray}\label{eq37b}
\langle(\Delta v_{x})^{2}\rangle^{\text{(ap)}}_{\text{ren}} \simeq -\dfrac{g^{2}}{6\pi^{2}m^{2}a^{2}}\left[-\frac{\pi^{2}}{2}+\frac{1}{\tau_{a}^{2}}\right].
\end{eqnarray}
Additionally, for the $y$ (or $z$) direction, the velocity dispersion, in the later time regime, is given by
\begin{eqnarray}\label{eq38}
\langle(\Delta v_{y})^{2}\rangle^{\beta}_{\text{ren}} \simeq \dfrac{g^{2}}{\pi^{2}m^{2}a^{2}}\left[\pi^{2}B_{2}(\beta)-\frac{1}{6\tau_{a}^{2}}\right],
\end{eqnarray}
with
\begin{eqnarray}\label{eq38a}
\langle(\Delta v_{y})^{2}\rangle^{\text{(p)}}_{\text{ren}} \simeq\dfrac{g^{2}}{6\pi^{2}m^{2}a^{2}}\left[\pi^{2}-\frac{1}{\tau_{a}^{2}}\right]
\end{eqnarray}
and
\begin{eqnarray}\label{eq38b}
\langle(\Delta v_{y})^{2}\rangle^{\text{(ap)}}_{\text{ren}} \simeq-\dfrac{g^{2}}{6\pi^{2}m^{2}a^{2}}\left[\frac{\pi^{2}}{2}+\frac{1}{\tau_{a}^{2}}\right],
\end{eqnarray}
for the periodic and antiperiodic cases, respectively. The results above for the late time regime show that, from the last term in Eqs. \eqref{eq37} and \eqref{eq38}, the dispersions tend to a constant value. As already mentioned, this characteristic is similar to the classical Brownian motion and indicates a possible equilibrium between the particle and the surrounding medium, which in this case corresponde to a vacuum filled by the quantum vacuum fluctuations of the massless scalar field subjected to the condition in Eq. \eqref{eq17}.
Note that the contribution arising from the second term in the r.h.s of the expressions above for the late time regime is independent of the parameter $a$ and is identical for both compactified and uncompactified directions. Possibly, this suggests some kind of physical process which is independent of the compactification. In the compactified case, Eq. \eqref{eq37}, this small contribution tend to strengthen the dispersions whereas in the uncompactified case, Eq. \eqref{eq38}, it tends to weaken.
In the late time regime the quasiperiodic velocity dispersions can have a change of sign in the compactified and uncompactified cases. This is due to the behavior of $B_{2}(\beta)$ function shown in Fig.\ref{figBernoulli}. Moreover, only in the case when $\beta=(3\pm\sqrt{3})/6$ the time dependent small contribution define the sign of the velocity dispersions. Finally, we emphasize that the negative results for the velocity dispersions can be understood according to the interpretation given at the end of the previous subsection.
\section{Conclusions}\label{Conclusions}
In this paper we have studied the QBM of a point particle induced by the quantum vacuum fluctuations of a massless scalar field, which are modified by both the presence of two reflecting parallel planes and a quasiperiodic condition that causes the $x$-direction to be compactified. We have considered three distinct boundary conditions for the field moldes to obey on the planes placed perpendicular to the $x$-direction at $x=0$ and $x=a$. The boundary conditions are Dirichlet, Neumann, and mixed which lead to the discretization of the momentum in the $x$-direction. Similarly, the quasiperiodic condition also leads to the discretization of the momentum in $x$-direction in the form $k_{x} = k_{n} = \frac{2\pi}{a}(n+\beta)$, with $0\leq\beta < 1$. In all cases, the parameter $a$, related to confinement, provide a natural scale for the system, which enable us to analyze the resulting expressions in asymptotic regimes of interest, namely, short time ($\tau_{a}\ll 1$) and late time ($\tau_{a}\gg 1$) regimes. In cases of Dirichlet, Neumann and mixed boundary conditions this parameter is the distance between the planes and for the quasiperiodic condition it is the quasiperiodicity length of the space or, in other words, the length of the compactification in the $x$ direction.
In the short time regime, for all conditions, we have seen that the most significant contributions for the velocity dispersions are of second order in time. For the late time regime, on the other hand, we have found that the velocity dispersions tend to a constant value, which depends on the conditions imposed on the field. This constante value for the velocity dispersion in the late time regime is a characteristic similar to the classical Brownian motion and suggests an equilibrium value between the particle and quantum vacuum fluctuations of the massless scalar field.
Divergent results for the velocity dispersions have also been identified, which are related to the usual divergencies on the planes at $x=0$ and $x=a$, to the time a light signal takes to travel in a round trip from one of the planes to a point $x$ and to the time a light signal takes to travel throughout the compactified direction. We have also indicated a position independent divergency in the parallel planes case that are related to an increasing number of round trips that a light signal takes to go from one plane to the other. Furthermore, negative velocity dispersions have also been shown to be possible and, based on discussions found in the literature, this can be understood as a reduction in the particle velocity dispersion due to the presence of the planes and the compactification mechanism.
We would like to stress that two planes configuration has been considered in order to complement the investigations for the electromagnetic and scalar fields found in the literature. Hence, the more remarkable contribution of this work has been the analysis of the QBM induced by the massless scalar field with distinct boundary conditions on the two parallel planes, which until now had not been done, besides Dirichlet conditions adopted only for one plane. In fact, all the works so far had focused on Dirichlet boundary condition. Also, the Dirichlet boundary condition has only been considered on two parallel planes in a system considering the electromagnetic field \cite{yu2004brownian}.
The compact form for the positive frequency Wightman function in cartesian coordinates presented in Eq. \eqref{FWDNM} is very interesting because its structure makes possible to write the result for three boundary conditions into a single expression, namely, Dirichlet, Neumann and mixed boundary conditions. This structure is very useful since it allows to extract the divergent Minkowski contribution and, consequently, obtain other finite physical observables besides the velocity dispersion considered here. For instance, the mean value of field squared, $\langle\phi^{2}\rangle=\lim_{x'\rightarrow x}\langle\phi(x)\phi(x')\rangle$, and the mean value of force squared that acts on the particle, $\langle F^{2}\rangle=\lim_{x'\rightarrow x}\langle F(x)F(x')\rangle$. In fact, as it can be easily checked, all these mentioned quantities depends on the Wightman function.
{\acknowledgments}
E.J.B.F would like to thank the Brazilian agency Coordination for the Improvement of Higher Education Personnel (CAPES) for financial support. E.M.B.G thanks financial support from the Brazilian agency National Council for Scientific and Technological Development (CNPq). H.F.S.M is partially supported by CNPq under grant N$\textsuperscript{o}$ 311031/2020-0.
\\
|
\section{Introduction}
Q-balls are pseudo-like particles that could be defined as lumps of a singularity-free scalar field with finite energy. They have been originally discovered in \cite{Rosen:1968mfz} and independently rediscovered in \cite{Coleman:1985ki}. Contrary to solitons, they do not have a topological charge but a Noether charge based originally on the $U(1)$ global symmetry, and therefore they belong to the class of nontopological solitons. The scalar field is captured in some region of space because of non-linear self-interaction, therefore forming a pseudo-like particle carrying charge and energy.
Q-balls can be produced via many mechanisms, which makes them very interesting in particular in cosmology. Indeed, they could be produced from inflationary models, such as natural inflation \cite{Freese:1990rb,Adams:1992bn}, where if a complex scalar field with a global symmetry is spontaneously broken, we end up with the inflaton as the goldstone boson and a naturally flat potential due to the shift symmetry. Also in supersymmetric extensions of the standard model (see e.g. \cite{Kusenko:1997zq}), Q-balls emerge naturally where the global charge could be assumed by the baryon or the lepton number. For example, the Affleck-Dine mechanism \cite{Affleck:1984fy,Dine:1995kz} uses the supersymmetric flat directions to generate baryogenesis. In this context, some of these flat directions (scalar field) can be parametrized as a complex field, which is in general a condensate of squarks, sleptons and Higgs field. This condensate can be unstable and form Q-balls \cite{Enqvist:2003gh}.
Of course, the most interesting property of Q-balls is their stability, because they could then be considered dark matter candidates \cite{Kusenko:1997si,Kusenko:2001vu}. For that reason, it will be our main focus in this paper along with some interesting properties related to their existence. The analysis of the classical stability was studied in \cite{Friedberg:1976me,Lee:1991ax} where they found that considering a Q-ball of frequency $\omega$ and charge $Q$, stability is similar to the condition $dQ/d\omega <0$. It was shown in \cite{Panin:2016ooo} that the stability of gauged Q-balls is not related to this condition. It would be interesting to see the extension of this criteria to global charge Q-balls but in modified gravity theories.
We will study three types of stability conditions that appear in the literature \cite{Tsumagari:2008bv}, namely, classical stability as we have previously mentioned, absolute stability, and stability against fission \cite{Lee:1991ax}.
In most of the papers, a canonical scalar field is considered which appears naturally at low energies of various theories. But studying Q-balls in the early universe might modify this simple picture. Indeed, e.g. higher dimensions naturally produce scalar fields with nonlinear kinetic terms such as D3-brane \cite{Silverstein:2003hf} or in the context of braneworld gravity \cite{Goon:2011qf}. Also in string theory, a rolling tachyon has a Dirac-Born-Infeld (DBI) type of action \cite{Sen:2002an}. It is therefore natural to look to non-canonical scalar fields. Q-balls in DBI type of kinetic term was studied in \cite{Kuniyasu:2016tse} along with its stability using catastrophe theory \cite{Sakai:2007ft}. In this context, we will study Q-balls in the context of a complex K-field also known as K-inflation \cite{Armendariz-Picon:1999hyi} or K-essence \cite{Armendariz-Picon:2000nqq}.
The plan of the paper is as follows. We introduce the model before discussing the stability conditions encountered in the literature. In the next section, we analyze the range of existence of the Q-balls and define the energy conditions for these solutions. Finally, we will study numerically the properties of the Q-balls before studying the equation of perturbation. We analyze the strong hyperbolicity of these equations along with the stability of the Q-ball before conclusions.
\section{Q-balls}
Let us consider the density Lagrangian
\begin{align}
\mathcal{L}= K(|\Phi|^2,X)
\end{align}
where $K$ is a generic function of a complex scalar field $\Phi$ and the kinetic term $X=-\partial_\mu\Phi\partial^\mu\Phi^*$. The equation of motion is
\begin{align}
\nabla_\mu(K_{,X} \partial^\mu \Phi)+\Phi K_{,|\Phi|^2} =0
\label{eq:flat}
\end{align}
where we have used the notation $K_{,A}\equiv {\partial K}/{\partial {A}}$.
The model admits a global U(1) symmetry to which the associated Noether current is
\begin{align}
j^\mu= i K_{,X} \Bigl(\Phi^*\partial^\mu\Phi-\Phi\partial^\mu\Phi^*\Bigr)
\end{align}
This current is conserved $\partial_\mu j^\mu$ on-shell. The corresponding conserved scalar charge (or total particle number) is
\begin{align}
Q=\int {\rm d}^3 x j^0=i\int {\rm d}^3 x K_{,X}(\Phi \dot\Phi^*-\dot \Phi \Phi^*)
\end{align}
To obtain the energy, we define the canonical conjugate momenta to the variables $\Phi$ and $\Phi^*$,
\begin{align}
&\pi_\Phi = \frac{\partial \mathcal{L}}{\partial \dot\Phi}= K_{,X}\dot\Phi^*\\
& \pi_{\Phi^*} = \frac{\partial \mathcal{L}}{\partial \dot\Phi^*}= K_{,X}\dot\Phi
\end{align}
so the Hamiltonian density is
\begin{align}
\mathcal{H}= \pi_\Phi \dot\Phi + \pi_{\Phi^*} \dot\Phi^*-\mathcal{L}=2|\dot\Phi|^2 K_{X}-K
\end{align}
The energy of the system is then
\begin{align}
E= \int {\rm d}^3 x \Bigl(2|\dot\Phi|^2 K_{X}-K\Bigr)
\label{eq:energy}
\end{align}
We are looking for solutions which minimize the energy for a given charge Q. For that, we define the functional
\begin{align}
E_\omega= E + \omega \Bigr[Q-i\int {\rm d}^3 x K_{,X}(\Phi \dot\Phi^*-\dot \Phi \Phi^*)\Bigl]
\end{align}
where $\omega$ is a Lagrange multiplier which enforces the given charge Q. We have
\begin{align}
E_\omega &= \omega Q + \int {\rm d}^3 x \Bigl[K_{X}\Bigl(2|\dot\Phi|^2-i\omega (\Phi \dot\Phi^*-\dot \Phi \Phi^*)\Bigr)-K\Bigr]\nonumber\\
& =\omega Q + \int {\rm d}^3 x \Bigl[K_{X} |\dot\Phi-i\omega \Phi|^2+K_{X}(|\dot\Phi|^2-\omega^2|\Phi|^2)\nonumber\\
&\qquad\qquad\qquad-K\Bigr]
\end{align}
In the case of a canonical scalar field, $K=X-V(|\Phi|^2)$, we have
\begin{align}
E_\omega =\omega Q &+ \int {\rm d}^3 x \Bigl[|\dot\Phi-i\omega \Phi|^2-\omega^2|\Phi|^2+|\vec\nabla\Phi|^2\nonumber\\
&+V(|\Phi|^2)\Bigr]
\end{align}
where we used that $X=-\partial_\mu\Phi\partial^\mu\Phi^*=|\dot\Phi|^2-|\vec\nabla\Phi|^2$. We can therefore conclude that for a given charge Q, the energy is minimized when $\dot\Phi-i\omega \Phi=0$ which means for $\Phi(t,\vec{x})=\phi(\vec{x})e^{i\omega t}$ \cite{Lee:1991ax}. This simple argument for the canonical scalar field can't be easily generalized to the K-field. But we observe that in the general case, if $\Phi(t,\vec{x})=\phi(\vec{x})e^{i\omega t}$
\begin{align}
E_\omega =\omega Q - \int {\rm d}^3 x ~K
\end{align}
which implies that the extrema of the energy (for fixed charge) coincide with the extrema of the action. Therefore solutions of the following type $\Phi(t,\vec{x})=\phi(\vec{x})e^{i\omega t}$ extremize the energy. Even if we don't know of the existence of other solutions which could also extremize the energy functional, we will assume in the future for this paper this time-dependent phase of the solution.
For a given model, the only parameter which characterizes the energy $E$ and the charge $Q$ is the parameter $\omega$. Therefore we can consider that energy and charge are functions of $\omega$, thus differentiating the energy, we get
\begin{align}
\frac{{\rm d}E}{{\rm d}\omega}=\int {\rm d}^3 x\Bigl[2\omega\phi^2K_{,X}+4\omega^3\phi^4K_{,XX}\Bigr]
\end{align}
Performing the same differentiation of the charge $Q$, we found
\begin{align}
\frac{{\rm d}E}{{\rm d}\omega}= \omega \frac{{\rm d}Q}{{\rm d}\omega}
\label{eq:EQ2}
\end{align}
which extends to K-field results from \cite{Friedberg:1976me}. When $\frac{{\rm d}Q}{{\rm d}\omega}=0$ also $\frac{{\rm d}E}{{\rm d}\omega}=0$ which corresponds to the existence of extremum of the charge and the energy at the same time. They will correspond to the cusps in the diagram $E(Q)$. When $\frac{{\rm d}Q}{{\rm d}\omega}\neq 0$, we obtain
\begin{align}
\frac{{\rm d}E}{{\rm d}Q}=\omega
\label{eq:EQ}
\end{align}
which corresponds to the generic relation found for a $U(1)$ Q-ball.
\section{Stability}
Usually, three different stability criteria are discussed in the literature. The first condition considers that a given Q-ball should not decay into smaller Q-balls, sometimes referred to stability against fission \cite{Lee:1991ax}. In that case, the stability translates into
\begin{align}
E(Q_1+Q_2) < E(Q_1) + E(Q_2)
\end{align}
and if taking derivatives wrt to both charges ($Q_1,Q_2$), we obtain the equivalent condition $\frac{{\rm d}^2 E}{{\rm d} Q^2} <0$ and by using eq.(\ref{eq:EQ}) it reduces to $\frac{{\rm d}Q}{{\rm d}\omega}<0$. Notice the similarity with the more generic Vakhitov-Kolokolov stability criterion \cite{VK} (or spectral stability). Of course, because of eq.(\ref{eq:EQ2}), we could equivalently consider $\frac{{\rm d}E}{{\rm d}\omega}<0$.
The second stability criterion considers decay into free particles of mass $M=\sqrt{\frac{V''(0)}{2}}$. In order to avoid the decay of a Q-ball into $Q$ free particles with the rest masses $M$, we need to consider $E(Q)<M Q$.
Finally, the last stability considers the time evolution of small perturbations, the so-called classical stability that we will analyze later.
Notice that from catastrophe theory, a simple criteria of stability has been proved \cite{Schunck:1992}. Indeed, considering the diagram $E(Q)$, the lowest branch corresponds to the stable soliton while the upper branch is unstable. This condition will be found to be equivalent to the linear stability.
\section{Existence}
\label{section:potential}
In this section, we briefly summarize the conditions of existence of Q-balls. These conditions are obtained by constraining the shape of the potential.
Considering a spherically symmetric spacetime, and $\Phi=\phi(r) e^{i\omega t}$ the eq.(\ref{eq:flat}) becomes
\begin{align}
&K_{,X}\Bigl(\phi''(r)+\frac{2}{r}\phi'(r)+\omega^2 \phi(r)\Bigr)+\phi'(r) X'(r) K_{,XX}\nonumber\\
&\qquad +\phi'(r)^2 K_{,\phi X}+\frac{1}{2}K_{,\phi}=0
\end{align}
with $X=\omega^2 \phi(r)^2-\phi'(r)^2$.
Let us first consider the canonical case, namely $K=X-V(\phi)$. The equation of motion reduces to
\begin{align}
\phi''(r)+\frac{2}{r}\phi'(r)+\omega^2 \phi(r)-\frac{V'(\phi)}{2}=0
\end{align}
which can be written as
\begin{align}
\phi''(r)+\frac{2}{r}\phi'(r)-V_{\text{eff}}'(\phi)=0
\end{align}
with $V_{\text{eff}}(\phi)=(V(\phi)-\omega^2\phi^2)/2$. We see that the $\omega^2$ term acts as a tachyonic contribution to the mass of the field, which will produce solitonic solutions otherwise absent for $\omega=0$. Considering only solutions with finite energy, the energy functional (\ref{eq:energy}) $E=\int{\rm d}^3x (\phi'(r)^2+\omega^2 \phi^2+V(\phi))$ implies that $(\phi,\phi') \rightarrow 0$ for $r \rightarrow \infty$ and $V(0)=0$ (we assumed $V(\phi)>0$).
It is easier to use the analogy with a particle in Newtonian mechanics, namely replacing $\phi\rightarrow x$ and $r \rightarrow t$ which gives $ \ddot x + \frac{2}{t} \dot x +W_{\text{eff}}'(x)=0$, where $W_{\text{eff}}(x)=-V_\text{eff}(x)$. Looking for a trajectory $\phi(r)$ or equivalently $x(t)$, we need to impose $x(\infty)=0$ to obtain a finite energy solution. Therefore, the problem reduces to classifying the different trajectories of the equivalent particle giving finite energy. It is easy to show \cite{Coleman:1985ki} that we need to impose $W''_{\text{eff}}(0)<0$ and $W_{\text{eff}}(\phi)>0$ around $\phi(r=0)$. These conditions translate into
$V''(0)>2\omega^2$ as well as $\text{min}\Bigl(\frac{V(\phi)}{\phi^2}\Bigr)\leq \omega^2$. Thus, non-renormalizable potentials have to be considered and the simplest could be $V(\phi)=m^2 \phi^2-b\phi^4+\lambda \phi^6$. The previous constraints reduce to
\begin{align}
0<m^2-\frac{b^2}{4 \lambda}<\omega^2 \leq m^2
\end{align}
The positivity of $m^2-b^2/4 \lambda$ is imposed by demanding that $V(0)$ is a global minimum. In this paper, we will normalize \cite{Volkov:2002aj} the parameters such as $\lambda=1$ and $b=2$ which implies $m>1$. Therefore we will consider $m^2=1.1$ which implies $0.32<\omega \leq 1.05$. The Q-ball will exist only in this range of frequencies. It is important to mention that this range will change for K-fields. For example, in a model where $K=X+\alpha X^2-V(\phi)$, we have around $r=0$, and using the condition $\phi'=0$, $\phi''(r)+W'_{eff}(\phi)\simeq 0$ with
\begin{align}
W'_{eff}=\omega^2\phi-\frac{m^2-2 b \phi^2+3\lambda \phi^4}{1+2\alpha \omega^2\phi^2}\phi
\end{align}
Therefore the condition $W_{eff}>0$ for some range of the scalar field, implies a different value for the minimum of $\omega$. For our parameters, we found that with good accuracy, $\omega_{min}\simeq (1+\alpha/30)/\sqrt{10}$ while $\omega_{max}$ remains unchanged.
An other important condition for the existence of the Q-ball is the nature of the differential equation. We have an equation
\begin{align}
\Bigl(K_{,X}-2\phi'^2 K_{,XX}\Bigr)\phi''+F(\phi,\phi') = 0
\end{align}
To avoid singular points, we need to impose $K_{,X}-2\phi'^2 K_{,XX} \neq 0$. Therefore, for any model, smoothly connected to the canonical case, $K_{,X}-2\phi'^2 K_{,XX} =1$, we should impose $K_{,X}-2\phi'^2 K_{,XX}>0$. Considering the model $K=X+\alpha X^2-V(\phi^2)$, we have $1+2\alpha\omega^2\phi^2-6\alpha\phi'^2>0$. Around the origin, we have $\phi'=0$, which implies the condition $1+2\alpha\omega^2\phi_0^2>0$ and therefore large negative values of $\alpha$ will not be allowed.
\section{Energy conditions}
For these type of models, the fluid interpretation is not suitable because the kinetic term does not have a definite sign. But, it is mostly positive in the interior of the Q-ball and becomes negative near the surface of the Q-ball. Therefore, deep inside the Q-ball, we can use the hydrodynamical interpretation of the scalar field, by defining the energy-momentum tensor
\begin{align}
T_{\mu\nu}=K g_{\mu\nu}+K_{,X} (\partial_\mu\Phi \partial_\nu\Phi^*+\partial_\mu\Phi^* \partial_\nu\Phi)
\end{align}
from which we define the energy density $\rho=2|\dot\Phi|^2 K_{,X}-K=2\omega^2\phi(r)^2 K_{,X}-K$, the radial pressure $P_r=2\phi'(r)^2K_{,X}+K$ and finally the tangential pressure $P_t=K$. These quantities can be converted into the pressure $P=(P_r+2P_t)/3$ and the shear force $S=P_r-P_t$.
The hydrodynamical approach helps to obtain easily the energy conditions such as the strong energy condition (SEC)
\begin{align}
K_{,X}\geq 0\,,~~
K+(\omega^2\phi^2+\phi'^2)K_{,X}\geq 0
\label{SEC}
\end{align}
the dominant energy condition (DEC)
\begin{align}
K_{,X}\geq 0\,,~~(\omega^2\phi^2-\phi'^2)K_{,X}-K \geq 0
\label{DEC}
\end{align}
the weak energy condition (WEC)
\begin{align}
K_{,X}\geq 0\,,~~2\omega^2\phi^2 K_{,X}-K \geq 0
\label{WEC}
\end{align}
and the null energy condition (NEC)
\begin{align}
K_{,X}\geq 0
\label{NEC}
\end{align}
We notice that $K_{,X}\geq 0$ is common to all energy conditions.
\section{Numerical analysis}
As we have mentioned, Q-balls are finite energy objects and therefore with finite space extension, which imposes the asymptotic condition $\phi(\infty)=0$. Therefore we have used a shooting method for each value of the frequency $\omega$ with mixed boundary conditions $\phi'(0)=0$ and $\phi(\infty)=0$. In practice, we have integrated the system from $r=10^{-30}$ to some value, $r_{\text{max}}$, and demanded that the solution remains unchanged if we increase $r_{\text{max}}$. In Figure \ref{Fig:standard}, we have considered the standard model $K(X)=X-V(|\Phi|^2)$ with the potential defined in the section \ref{section:potential}. For lower frequencies, or thin wall limit, the scalar field is constant and at some radius (often considered as the Q-ball radius) the scalar field drops rapidly to zero, while for larger values of $\omega$, also known as thick wall limit, the scalar field is more shallow. The latter will be unstable. In the same graphics, we have represented the energy and the charge. The energy and charge seem to diverge for the frequencies $\omega_{min}$ and $\omega_{max}$. Also $E(\omega)$ and $Q(\omega)$ reach their minimum for the same frequency, defining therefore a cusp in the energy vs charge graphics. We show also the stability conditions of the Q-balls. The stability criteria against decay is stronger than the fission stability condition. In the $(Q,E)$ plot, it is easy to determine the stable Q-ball. Indeed, for every given charge $Q$, two Q-balls exist, the one with the smallest energy corresponds to the solution stable under fission. We will see later, that it corresponds also to the stable solution under linear perturbations.
\onecolumngrid
\begin{figure}[H]
\centering
\includegraphics[scale=0.48]{alpha01.eps}\hspace{0.5cm}
\includegraphics[scale=0.445]{alpha02.eps}\hspace{0.5cm}
\includegraphics[scale=0.48]{alpha03.eps}
\caption{Left: The field $\phi(r)$ is shown as a function of the radial coordinate for different values of $\omega$. For each value of $\omega$, $\phi(0)$ is adjusted such that $\phi(\infty)=0$. Center: The energy $E$ and the charge $Q$ are shown as a function of the frequency $\omega$ with the critical frequency (change of colors) defined by the condition $dQ/d\omega=0$. Right: The energy is shown as a function of the charge. For all graphics, in green we have stable configurations according to the fission stability criteria, while in red we have unstable solutions. In the first figure, the solution for the critical frequency is shown in blue and in the third graphics, we have added the decay stability criteria which is shown by a red solid line and red dashed line for the unstable solutions while the fission unstable configurations are represented only by red solid line.}
\label{Fig:standard}
\end{figure}
\twocolumngrid
Q-balls have also excited states which correspond to solutions with nodes but with the same limit at infinity, namely $\phi(\infty)=0$. In Figure \ref{Fig:standardbis}, we show the first and second excited modes for a given frequency $\omega$. In order to fulfill the boundary conditions, for excited states, the initial conditions must be extremely fine-tuned. The excited states have as expected larger energy but also charge. We found that the frequency corresponding to $dE/d\omega=0$ becomes larger with the number of nodes. For example, for the fundamental mode, we have a minimal energy for $\omega=0.972$, while $\omega=1.015$ for the first excited mode and $\omega=1.025$ for the second excited mode.
\begin{figure}
\centering
\includegraphics[scale=0.458]{alpha0bis1.eps}
\includegraphics[scale=0.45]{alpha0bis2.eps}
\caption{The field $\phi(r)$ is shown as a function of the radial coordinate for the fundamental mode (green), the first (purple) and the second (blue) radial excited mode for $\omega=0.7$. We also show the evolution of the energy as a function of the frequency. The dashed region corresponds to the unstable solutions according to the fission stability criteria.}
\label{Fig:standardbis}
\end{figure}
All these solutions are easily generalized to K-field theories. We will consider the simplest model where the action is modified by a single parameter, $K=X+\alpha X^2-V(|\Phi|)$ where $\alpha$ is the new parameter of the model. Generically, we found that the structure of the solutions will not change. Q-balls exist for a certain range of frequency which depends on the parameter $\alpha$. We see from Figure \ref{Fig:EQ} that for a given frequency, the Q-ball lowers its energy for large positive values of the parameter $\alpha$, because the radius decreases.
Notice that the critical value, $(E'(\omega)=0)$, of the energy and charge is also lowered for larger values of $\alpha$. Therefore, for a given frequency, the modified model with $\alpha>0$ produces Q-balls with lower charge and energy. The modification by the K-field allows to build Q-balls with small charge and energy or on the contrary with larger energy and charge. Finally, we found that for all values of the parameter $\alpha$, in the limit of $\omega\rightarrow \omega_{\text{max}}$, or thick-wall limit, we have the scaling solution $E=\omega Q^{\gamma}$ with $\gamma=1\pm10^{-4}$. This expression generalizes results found in \cite{Tsumagari:2008bv}.
\begin{figure}
\centering
\includegraphics[scale=0.47]{EQ.eps}
\caption{The energy is show as a function of charge for different values of the parameter $\alpha$ which runs from $\alpha=-0.5$ in red to $\alpha=0.5$ in purple with an incrementation of $0.1$}
\label{Fig:EQ}
\end{figure}
In the Figure \ref{Fig:EC}, we show the energy versus the frequency for different values of $\alpha$ but with the information on the violation of the energy conditions. We see that NEC is never violated. This condition corresponds to $1+2\alpha (\omega^2\phi^2-\phi'^2)>0$. It could be violated for very negative values of $\alpha$, but the construction of Q-balls for $\alpha<-0.5$ becomes very challenging and often impossible. In general, the larger (
and positive $\alpha$, the lower the probability to violate an energy condition, except for the SEC which is violated for any $\alpha$.
\begin{figure}
\centering
\includegraphics[scale=0.61]{EC.eps}
\caption{Energy versus frequency for K-field model with $\alpha$ running from $-0.5$ (in red) to $+0.5$ (in purple) with a step of $0.1$. For each panel, we have represented in dotted line the regime where some energy condition is violated. From top left to bottom right, we show the violation of the SEC, DEC, WEC, NEC.}
\label{Fig:EC}
\end{figure}
\section{Perturbations}
To study the mechanical stability, we decompose our field as
\begin{align}
\Phi(t,r,\theta,\varphi)=\phi(r)e^{i\omega t}+\sum_{\ell,m}\delta\Phi_{lm}(t,r) e^{i\omega t} Y_\ell^m(\theta,\varphi)\nonumber
\end{align}
where $\phi(r)$ is the background scalar field studied in the previous sections, $\delta \Phi_{\ell m}$ is the scalar field perturbation, $e^{i\omega t}$ in the second term is included for convenience and $Y_\ell^m$ are spherical harmonics. Because of the symmetries of the Q-balls, the perturbations will be independent of the azimuthal number $m$ and therefore the spherical harmonics reduce to Legendre polynomials. We will fix $m=0$. Notice that the different modes, $\ell$, do not couple and therefore we will omit this index. At second order in perturbations, and after integrating over the angle variables, the action reduces to
\begin{align}
S&=\int {\rm d}t {\rm d}r \Bigl[
r^2 K_{,X} \dot\Psi_1^2-r^2(K_{,X}-2\phi'^2K_{,XX}) \Psi_1'^2
\nonumber\\
&+r^2 (K_{,X}+2\omega^2\phi^2 K_{,XX}) \dot\Psi_2^2-r^2 K_{,X} \Psi_2'^2
\nonumber\\
&-2\omega r^2 \phi \phi' K_{,XX}\Bigl(\dot\Psi_1 \Psi_2'+\Psi_1'\dot\Psi_2\Bigr)+A\Bigl(\dot\Psi_1\Psi_2-\Psi_1\dot\Psi_2\Bigr)\nonumber\\
&-M_1^2\Psi_1^2 -M_2^2\Psi_2^2\Bigr]
\end{align}
where we have decomposed the perturbation into its real and imaginary parts, $\delta\Phi=\Psi_1+i \Psi_2$, and
\begin{align}
A &= -2\omega r^2 \frac{{\rm d}}{\rm d(\phi^2)}\Bigl(\phi^2 K_{,X}\Bigr)-\omega \frac{{\rm d} }{{\rm d}r} \Bigl(r^2\phi\phi' K_{,XX}\Bigr)\label{eqA}\nonumber\\
M_1^2 &=\lambda K_{,X} - \frac{r^2}{2} K_{,\phi\phi}
-\frac{{\rm d}}{{\rm d} r}\Bigl(r^2\phi' K_{,X\phi}\Bigr)\nonumber\\
M_2^2 &=\lambda K_{,X}- r^2 \Bigl(K_{,\phi^2}+\omega^2K_{,X}\Bigr)\nonumber\\
\lambda &= \ell (\ell+1)
\end{align}
From this action, we obtain the two coupled equations for linear perturbations
\begin{align}
& -K_{,X}\ddot\Psi_1+(K_{,X}-2\phi'^2K_{,XX})\Psi_1'' +2\omega \phi\phi' K_{,XX}\dot\Psi_2'\nonumber\\
& \qquad +F_1(r,\Psi_1,\Psi_2,\Psi_1',\dot \Psi_2)=0 \label{eq:pert1}\\
&-(K_{,X}+2\omega^2\phi^2 K_{,XX})\ddot\Psi_2+K_{,X}\Psi_2''+2\omega \phi\phi' K_{,XX}\dot\Psi_1'\nonumber\\
& \qquad +F_2(r,\Psi_1,\Psi_2,\Psi_2',\dot \Psi_1)=0
\label{eq:pert2}
\end{align}
with $F_1,F_2$ some functions defined by the perturbations and their first derivative. In order to study the stability, we need to insure that the problem is well posed. For that, we will derive the conditions of weak and strong hyperbolicity. Broadly speaking, the weak hyperbolicity condition forbids any solution to grow exponentially in time while the strong hyperbolicity condition imposes a stronger bound than the exponential growth and therefore is equivalent to local well-posedness of the Cauchy problem. In case of a strong hyperbolic system, $F_1,F_2$ will not be relevant while they could change the behavior of the system if weakly hyperbolic. We define the vector $u=(\Psi_1,\Psi_2)^T$ and the system (\ref{eq:pert1},\ref{eq:pert2}) becomes
\begin{align}
u_{,tt}=A u''+B u_{,t}'+\cdots
\end{align}
where $\cdots $ indicates the lowest derivative terms, and
\begin{align}
A_{11} &=\frac{K_{,X}-2\phi'^2 K_{,XX}}{K_{,X}}\\
A_{22} &=\frac{K_{,X}}{K_{,X}+2\omega^2 \phi^2 K_{,XX}}\\
B_{12} &=2\omega \phi\phi' \frac{K_{,XX}}{K_{,X}}\\
B_{21} &=2\omega \phi\phi' \frac{K_{,XX}}{K_{,X}+2\omega^2\phi^2K_{,XX}}
\end{align}
while other elements of the matrices are zero. We consider wave solutions $u(t,r)=e^{i k r} \hat u(t,k)$ and obtain
\begin{align}
\hat u_{,tt}=-k^2 A\hat u+ik B \hat u_{,t}+\cdots
\end{align}
This system can be reduced to first order by defining the variable $\hat v=\hat u_{,t}/(i|k|)$
\begin{align}
\begin{pmatrix}
\hat v\\
\hat u
\end{pmatrix}_{,t}=i|k| \hat P\begin{pmatrix}
\hat v\\
\hat u
\end{pmatrix}
\end{align}
with
\begin{align}
\hat P =\begin{pmatrix}
0 & \frac{k}{|k|}B_{12} & A_{11} & 0\\
\frac{k}{|k|}B_{21} & 0 & 0 & A_{22}\\
1 & 0 & 0 & 0\\
0 & 1 & 0 & 0
\end{pmatrix}
\end{align}
The well posedness of this system is reduced to the analysis of the matrix $\hat P$ (see e.g. \cite{Kreiss:2001cu}). If, for all $k$, the eigenvalues of $\hat P$ are real, the system is weakly hyperbolic. The eigenvalues are
\begin{align}
\Bigl\{\pm 1,\pm \sqrt{\frac{K_{,X}-2\phi'^2 K_{,XX}}{K_{,X}+2\omega^2\phi^2 K_{,XX}}}\Bigr\}
\end{align}
Therefore, we conclude that, if $\frac{K_{,X}-2\phi'^2 K_{,XX}}{K_{,X}+2\omega^2\phi^2 K_{,XX}}\geq 0$, the system is weakly hyperbolic. Additionally, when
\begin{align}
\frac{K_{,X}-2\phi'^2 K_{,XX}}{K_{,X}+2\omega^2\phi^2 K_{,XX}} > 0
\end{align}
the system is strongly hyperbolic because the eigenvectors form a complete set. In that case, the two perturbations propagate at the speed
\begin{align}
c_1 =1\,,\qquad
c_2 = \sqrt{\frac{K_{,X}-2\phi'^2 K_{,XX}}{K_{,X}+2\omega^2\phi^2 K_{,XX}}}
\end{align}
As we have shown in section \ref{section:potential}, we consider the condition $K_{,X}-2\phi'^2K_{,XX}>0$ which implies $K_{,X}+2\omega^2 \phi^2K_{,XX}>0$. Summing these two conditions, we find a weaker condition, viz. $K_{,X}>0$ and $K_{,X}+X K_{,XX}>0$.
Notice that the conditions of well possedness of the system are independent of the energy conditions derived previously (\ref{SEC},\ref{DEC},\ref{WEC},\ref{NEC}). In Figure \ref{Fig:strong}, and for the model $K=X+\alpha X^2-V(\phi)$, we have found that for a certain range of the parameters $(\omega,\alpha)$, the Cauchy problem is not well-posed which never corresponds to $\alpha>0$. Also we found that for any $\alpha<0$, the perturbations are superluminal in some region of space. Even if the classical theory is well posed, the superluminal propagation of the perturbations could be an obstacle to a quantum version of the theory. For example, requiring UV completion for K-essence (real scalar field analog to the case studied in this paper) imposes subluminal propagation \cite{Melville:2019wyy}. A similar situation should be expected in our case \cite{Adams:2006sv}. Even if not equivalent, we found numerically, for all parameters $(\omega,\alpha)$ of Figure \ref{Fig:strong}, that a system which violates weak energy condition does not have a well-posed Cauchy problem. The converse is not true.
\begin{figure}
\centering
\includegraphics[scale=0.5]{strong.eps}
\caption{In gray, the region of parameter space $(\omega,\alpha)$ where the Cauchy problem is not well-posed and in cyan the region of superluminal propagation.}
\label{Fig:strong}
\end{figure}
Restricting our analysis to the cases where the Cauchy problem is well-posed, we can study the mechanical stability of our solutions. For that, we assume the following form for the perturbation
\begin{align}
\delta\Phi(t,r)=\frac{\eta(r)}{r^n} e^{i\rho t}+\frac{\chi^*(r)}{r^n} e^{-i\rho^* t}
\label{eq:pertu}
\end{align}
The system (\ref{eq:pert1},\ref{eq:pert2}) reduces to two ordinary coupled differential equations for $\eta(r)$ and $\chi(r)$. We have included a factor $r^n$ for numerical stability. In general, $n=\ell$ provides faster numerical results. In the canonical case where $K_{,X}=1$, the stability analyses shows that any instability corresponds to $\rho=-\rho^*$ \cite{Panin:2016ooo} which implies the condition $\frac{dQ}{d\omega}<0$. We could not extend this analysis to K-field theories and therefore we will study the perturbations by numerical means. For that, our system can be written as four first order differential equations for the variable $\Psi\equiv(\eta,\chi,\eta',\chi')^T$, $\Psi'=B\Psi$ where the matrix $B$ is given in the appendix \ref{pert}. Considering the conditions at $r=0$ on the scalar field, $\phi'=0$, it is easy to show that perturbations behave as
\begin{align}
\eta(r\simeq 0) &= c_0 r^{\ell+n}\\
\chi(r\simeq 0) &= c_1 r^{\ell+n}
\end{align}
which implies
\begin{align}
\Psi(0)=c_0 r^{\ell+n-1}\begin{pmatrix}
r\\
0\\
\ell+n\\
0
\end{pmatrix}+c_1r^{\ell+n-1}\begin{pmatrix}
0\\
r\\
0\\
\ell+n
\end{pmatrix}
\end{align}
Therefore, we can perform two numerical integrations from $r=0$ with initial conditions $\eta=r^{\ell+n}, \chi=0$ and $\eta=0, \chi=r^{\ell+n}$ respectively. The general solution will be a linear combination of these two solutions with coefficients $(c_0,c_1)$. Similarly, we perform an integration from infinity to $r=0$. We have also a system with two free parameters $(c_3,c_4)$. We can integrate it from a large radius with initial conditions
\begin{align}
\eta = \frac{e^{-r \sqrt{-\frac{K_{,\phi^2}(0,0)}{K_{,X}(0,0)}-(\rho+\omega)^2}}}{r^{1-n}}\,,~~\chi=0
\end{align}
or
\begin{align}
\chi = \frac{e^{-r \sqrt{-\frac{K_{,\phi^2}(0,0)}{K_{,X}(0,0)}-(\rho-\omega)^2}}}{r^{1-n}}\,,~~\eta=0
\end{align}
Having the solution integrated from both boundaries with four free parameters $(c_1,c_2,c_3,c_4)$, we can match them at a given radius, using the four continuity conditions of $(\eta,\chi,\eta',\chi')$. Notice that, because our system is linear, we can always fix one of the parameters, e.g. $c_1=1$. Therefore, we end with a system of four conditions and three parameters, the fourth parameter will determine the value of $\rho$. In conclusion, only a certain number of discrete values of $\rho$ can solve our problem.
In the Figure (\ref{Fig:stability}), we show $|\phi+\delta\Phi|^2$, for $\omega=(0.5,1)$ and $\alpha=0$. For each case, we have found the parameter $\rho$ and using eq.(\ref{eq:pertu}), we obtain the time and space dependence of the solution. In the case, where $\omega=0.5$, the radius of the Q-ball is oscillating, $\rho$ is real. The energy of this solution is constant in time, while for $\omega=1$, the energy grows exponentially as well as the radius of the Q-ball. The solution is unstable and $\rho$ is purely imaginary.
\begin{figure}
\centering
\includegraphics[scale=0.75]{stable.eps}
\includegraphics[scale=0.75]{unstable.eps}
\caption{Space-time diagram of $|\Phi|^2$. The upper diagram shows the stability of the background solution with $\omega=0.5$ and the lower case shows an unstable solution for $\omega=1$. For both solutions, we have considered $\alpha=0$.}
\label{Fig:stability}
\end{figure}
Therefore, the strategy is simple, for each Q-ball, we search in the complex plane for values of $\rho$ solution to our previous problem.
For the excited states, all frequencies $\omega$ were unstable. But for various frequencies, the unstable modes were not always purely imaginary but also with nonzero real part.
For the fundamental solution, Figure \ref{Fig:stability} shows two cases where $\alpha=0$ and $\omega=(0.5,1)$. The first frequency corresponds to a stable solution for which we see an oscillation of the radius of the Q-ball while the energy remains perfectly constant in time. The second case, corresponds to an unstable solution for which the radius increases and the energy grows exponentially.
Generically, we found that the stability region corresponds to $dQ/d\omega<0$ for all $\omega$, generalizing results which were known in the canonical case. In the unstable region, the time scale of the instability is of the order $1/\text{Im}(\rho)$. We found that Im$(\rho)$ and therefore the time scale of the instability depends on the mode $\ell$. For example, for $\alpha=0$, Im$(\rho)$ is of the order $10^{-1}$ for $\ell=0$ and of the order $10^{-4}$ for $\ell=1$. Therefore, we will focus mainly on the spherical mode of perturbations $\ell=0$.
In Figure \ref{Fig:frequency}, we show the unstable modes for three values of $\alpha$. For each $\alpha$, the instability starts when $dQ/d\omega=0$. We notice also that even if for a given frequency, such as $\omega=1.03$, the Q-ball is unstable for all values of the parameter $\alpha$, the instability is slower to develop (lower value of Im$(\rho)$), for larger positive values of $\alpha$, which is consistent with the previous section where we found that the energy is lowered.
\begin{figure}
\centering
\includegraphics[scale=.85]{freq.eps}
\caption{Existence of Im$(\rho)$ as a function of $\omega$ for $\alpha=(-0.5,0,+0.5)$. The existence of such mode implies an instability of the background solution. The dotted line corresponds to unstable modes but in a region where the Cauchy problem is not well-posed and therefore should be excluded from the analysis.}
\label{Fig:frequency}
\end{figure}
In Figure \ref{Fig:summary}, we summarize the various stability conditions. The quantum stability condition, namely the stability against fission is, as expected, stronger than the classical stability condition. We have also represented regions where the energy conditions are violated. The NEC is never violated in the region of analysis of the model while the WEC is violated only in region where the Cauchy problem is not well-posed. The violation of the SEC and the DEC are totally independent of the stability conditions.
\begin{figure}
\centering
\includegraphics[scale=.85]{summary.eps}
\caption{Space of parameters $(\omega,\alpha)$ within the region where the Cauchy problem is well-posed. Are represented regions of quantum stability (against fission) and classical stability as well regions where the energy conditions such as the SEC and DEC are violated. We have kept the cyan and white colors for, respectively, superluminal and subluminal propagation.}
\label{Fig:summary}
\end{figure}
\section{Conclusion}
In this work we studied Q-balls in non-canonical scalar field theory. We derived the general equations of existence and stability for these theories. We found that the stability against fission and the linear mechanical stability are equivalent and reduce to $Q'(\omega)<0$. On the other hand, the condition for decay into free particles is stronger. We found that perturbations have a well-posed Cauchy problem if $\frac{K_{,X}-2\phi'^2 K_{,XX}}{K_{,X}+2\omega^2\phi^2 K_{,XX}} > 0$. When the perturbations are strongly hyperbolic, we found that perturbations are superluminal or subluminal. In the particular case, $K=X+\alpha X^2-V(|\Phi|^2)$, perturbations are subluminal and luminal for $\alpha>0$ while superluminal and luminal for $\alpha<0$.
We found that a Q-ball with $\alpha>0$ lowers its energy for larger values of $\alpha$. Even in the unstable region, the time-scale of this instability becomes larger and therefore more stable. The frequency at which Q-balls become unstable increases with $\alpha$. It would be interesting to find models for which all Q-balls are stable irrespectively of their frequency.
Finally, we have studied the different energy conditions such as the SEC, DEC, WEC, NEC. We found that NEC is never violated and none of these conditions can be related to mechanical stability.
\section*{Acknowledgements}
The work of A.F. is supported by ANID/CONICYT No. 21171262 while R.G. is supported by ANID FONDECYT Regular No. 1220965.
|
\section{Introduction}
Establishing dense correspondences between 3D shapes is a fundamental problem in computer vision and computer graphics, as it enables many downstream applications such as statistical shape analysis \cite{Bogo2014,Pishchulin2017}, texture \cite{Dinh2005} and deformation transfer \cite{Sumner2004,Baran2009}, and registration \cite{Zhou2016}, to name a few.
A particularly challenging setting for this task is the computation of correspondences between 3D shapes that undergo non-rigid, non-isometric deformations, and that may exhibit some partiality, such as missing semantic parts, and can be represented as sparse point clouds.
The standard approach is to formulate shape correspondence as a supervised learning problem, by relying on ground truth maps within large datasets of shape pairs. Given such ground truth, it is possible to train neural networks to either produce deformation fields \cite{groueix20183d}, segmentation maps \cite{masci2015geodesic,monti2017geometric,poulenard2018multi} or functional maps \cite{donati2020deep,attaiki2021dpfm,litany2017deep} between the input shapes. However, all such methods rely on the presence of labeled point-to-point correspondences, which are expensive to obtain, and are only available in a handful of cases.
At the same time, \emph{unsupervised} techniques try to solve the matching problem by imposing structural properties on the maps, either in the intrinsic \cite{halimi2019unsupervised,roufosse2019unsupervised,Rodol2016} or extrinsic domains \cite{eisenberger2020deep,Eisenberger2020SmoothSM}. However, the priors used by these methods tend to be purely geometric (e.g., promoting near isometry or divergence-free deformation fields), and, as we demonstrate below, are not always applicable, especially in challenging non-isometric settings.
In this work, we demonstrate that the neural network itself can act as a powerful prior for shape correspondence problems. For this, we first observe that even without training, recent architectures for 3D shape analysis \cite{sharp2021diffusion,wang2019dgcnn,qi2017pointnet++} produce pointwise features that capture local geometry and lead to well-structured maps between shapes. Secondly and perhaps more importantly, we demonstrate an effect that we call \emph{neural correspondence prior}. Specifically, given noisy input maps, we formulate a supervised learning problem, where we aim to learn pointwise features that, when compared, would recover the input maps. Remarkably, we show that the correspondences computed by the trained networks are typically of higher quality than the input. Our results are consistent with the notion of \emph{noise impedance} observed in other works on neural priors pioneered in \cite{UlyanovVL17}, and \emph{neural bias} \cite{rahaman19a} which states that neural networks tend to optimize lower frequency signals before higher-frequency ones. We show that in the context of shape matching, these effects imply that neural networks can learn powerful features by training to overfit to given noisy or incoherent correspondences.
\input{figures/teaser_tex}
Based on these observations, we develop a two-stage algorithm for unsupervised shape matching (see \cref{fig:teaser}). In the first stage, we adapt an existing unsupervised neural network, using variants of standard methods from the literature. We then train a new network from scratch in a supervised manner using the noisy maps as ground truth. We demonstrate that this second stage not only helps to denoise the maps but also improves the matching result overall and achieves state-of-the-art results on multiple tasks. We show that \emph{neural correspondence prior} is applicable within both collections and on individual shape pairs, to provide test time refinement of dense input maps. Since this method makes no assumptions about the geometry of the shapes, it can be used for complex cases such as non-isometric matching.
Overall, our contributions can be summarized as follows:
\begin{itemize}
\item We demonstrate that untrained neural networks for 3D shapes can produce pointwise features that are on par with complex axiomatic local descriptors.
\item We show that when trained using artifact-laden maps for supervision, the features learned by neural networks lead to correspondences that are more coherent and of higher quality than the input. We call this effect \emph{Neural Correspondence Prior}.
\item Based on these insights, we develop a two-stage algorithm for unsupervised 3D non-rigid shape matching that achieves SOTA results for difficult matching scenarios such as non-isometric and point cloud matching.
\item Using our unsupervised point-to-point correspondence method, we propose an approach for few-shot keypoint detection
\end{itemize}
\section{Related work}
\label{sec:related}
Shape matching is a well-studied area in computer vision and graphics, and a full overview is beyond the scope of this paper. We refer the interested readers to recent surveys \cite{Sahillioglu2019RecentAI,Biasotti2015,deng2022survey,Tam2013} for a more in-depth treatment of this field. Below, we review methods that are most related to our work, with a focus on learning-based techniques, both supervised and unsupervised.
\mypara{3D non-rigid shape correspondence}
There is a vast literature on learning non-rigid shape correspondences. In terms of supervised methods,
an important direction is to consider the matching problem as a dense semantic segmentation problem where the labels are vertices on some template shape \cite{masci2015geodesic,monti2017geometric,poulenard2018multi}, or by mapping via template deformation \cite{groueix20183d}. However, these techniques are known to require a considerable amount of data, and may fail when discretization changes \cite{sharp2021diffusion,donati2020deep}. Another approach is to use the functional map framework that was introduced in \cite{ovsjanikov2012functional} and extended in many follow-up works \cite{rodola2017partial, burghard2017embedding,eynard2016coupled,Shoham2019,ren2018continuous,Ren2019,Nogneng2017,marin22_why,li2022srfeat} to name a few (see \cite{Ovsjanikov2017} for an overview). The functional map (fmap) framework is based on encoding and optimizing maps as small matrices in a reduced basis. This formulation has been successfully used in the supervised setting to establish correspondences between complete and partial non-rigid shapes \cite{donati2020deep,attaiki2021dpfm,litany2017deep}. However, it is typically restricted to near-isometric matching between relatively clean discretizations.
The fmap framework has also been adapted to the unsupervised setting, either by imposing structural properties on the functional maps \cite{roufosse2019unsupervised, sharma2020weakly}, by enforcing cycle consistency \cite{Ginzburg2020}, or by combining intrinsic and extrinsic properties \cite{eisenberger2020deep,Eisenberger2020SmoothSM}.
Another line of work proposes to perform unsupervised shape matching by reconstructing the input shapes in canonical order, based on an autoencoder~\cite{cheng2021learning,zheng2020dit,liu2020learning}. However, these methods require a lot of shapes and learning time, and have only successfully been applied to man-made shapes that do not undergo significant non-rigid deformations.
\mypara{Neural prior}
The neural prior was first introduced in the seminal work of Ulyanov \etal~\cite{UlyanovVL17}, where the authors introduced the Deep Image Prior (DIP), and showed that a randomly initialized convolutional neural network can serve as a good prior for many 2D inverse problems, such as inpainting and denoising. Since then, several modifications and improvements have been made to DIP \cite{heckel_deep_2018,Heckel2020CompressiveSW,Mataev2019DeepREDDI,Liu2019,Baguer_2020}, in order to improve its performance or adapt it to new tasks. Using \textit{untrained features} for matching was recently used in the context of 2D keypoint matching in \cite{Hong_2021_ICCV}. Another line of work \cite{Dittmer_2019,bayesian_DIP} has theoretically studied DIP either through the lens of regularization theory or through its connection to Gaussian processes. Despite its success in 2D computer vision, neural prior has been little exploited in the 3D domain, the only application being shape reconstruction. Indeed, \cite{Williams_2019_CVPR,Hanocka2020} attempt to reconstruct a water-tight mesh from a point cloud, either by overfitting several local patches to the point cloud or by shrinking an input mesh by a neural network. The objective of our work is to fill this gap and to propose a neural prior technique for 3D matching.
\mypara{Unsupervised keypoint detection}
Unlike the 2D case~\cite{Jakab2018UnsupervisedLO,Jakab_2020_CVPR,Thewlis_2017_ICCV,Wiles2018SelfsupervisedLO}, unsupervised 3D keypoint detection is relatively under-explored in the literature. The recently introduced KeyPointNet \cite{keypointnet2020} benchmark provides a good testing bed. Prominent unsupervised keypoint detection methods are based on a self-supervised paradigm using an encoder-decoder network, with a reconstruction error. The encoder of Skeleton Merger \cite{skeleton_merger2021} predicts a set of salient keypoints, that the decoder uses to produce a skeleton of the shape. UKPGAN \cite{you2020ukpgan} harnesses the power of GANs to detect keypoints, by forcing the decoder to reconstruct the input shape using solely the set of keypoints produced by the encoder. Finally, KeypointDeformer \cite{keypoint_deformer2021} predicts a set of keypoints that allow for efficient shape manipulation. In particular, given a source and a target shape, the keypoints predicted on the source shape are used to deform it to the target shape using a cage skinning deformer \cite{Yifan:NeuralCage:2020}.
\section*{Background \& Notation}
Our work uses the functional map framework as a first estimator for p2p maps, and feature learning to learn robust features that allow extracting good correspondences using the nearest neighbor. We provide a brief overview in the next section.
\paragraph{Functional Maps}
As we will see in \cref{sec:non-rigid-corr}, the functional map (fmap) framework will be used for the first stage of our algorithm. For this, we follow the general strategy of recent fmap-based techniques \cite{donati2020deep,attaiki2021dpfm,sharma2020weakly,roufosse2019unsupervised,sharp2021diffusion}, as follows: given source and target shapes $\mathcal{M}$ and $\mathcal{N}$, represented as either triangular meshes or point clouds, having $m$ and $n$ vertices respectively, we pre-compute their Laplace-Beltrami operator, and store their first $k$ eigenfunctions in the matrices $\Phi_{\mathcal{M}} \in \mathbb{R}^{m \times k}$ and $\Phi_{\mathcal{N}} \in \mathbb{R}^{n \times k}$ respectively. Using a Siamese network $\mathcal{F}_{\theta}$, we compute for each shape a $d$-dimensional descriptor $\mathcal{F}_{\theta}(\mathcal{M}) = \mathbf{F} \in \mathbb{R}^{m \times d}$ and $\mathcal{F}_{\theta}(\mathcal{N}) = \mathbf{G} \in \mathbb{R}^{n \times d}$ respectively. These descriptors are then projected to the spectral domain to form the spectral features $\mathbf{A} = \Phi_{\mathcal{M}}^{\dagger} \mathbf{F}$ and $\mathbf{B} = \Phi_{\mathcal{N}}^{\dagger} \mathbf{G}$. A functional map is then computed by solving the following linear system:
\begin{align}
\label{eq:fmap_energy}
C_{opt} = \argmin_{C} \| C\mathbf{A} - \mathbf{B} \| + \lambda \|C \Delta_{\mathcal{M}} - \Delta_{\mathcal{N}} C \|.
\end{align}
where $\Delta_{\mathcal{M}},\Delta_{\mathcal{N}}$ are diagonal matrices of Laplace-Beltrami eigenvalues of the corresponding shapes and $\lambda$ is a scalar hyper-parameter.
Following the unsupervised literature \cite{roufosse2019unsupervised,sharma2020weakly,sharp2021diffusion}, the Siamese network $\mathcal{F}_{\theta}$ is trained by imposing structural properties on the fmap $C$ such as bijectivity and orthogonality. In fact, given the fmap $C_{\mathcal{M}\mathcal{N}}$ from $\mathcal{M}$ to $\mathcal{N}$, and $C_{\mathcal{N}\mathcal{M}}$ from $\mathcal{N}$ to $\mathcal{M}$, the bijectivity constraint is such that $C_{\mathcal{M}\mathcal{N}}C_{\mathcal{N}\mathcal{M}} = \mathbb{I}_k$, and the orthogonality constraint requires that $C_{\mathcal{M}\mathcal{N}}^{\top} C_{\mathcal{M}\mathcal{N}} = \mathbb{I}_k$. $\mathbb{I}_k$ denotes the identity matrix of size $k$ and $\bullet^{\top}$ is the matrix transpose operator.
\paragraph{Feature Learning}
The goal of feature learning is to learn robust descriptors that can allow direct nearest-neighbor matching in the descriptor space. In this work, we will use the contrastive loss PointInfoNCE \cite{xie2020pointcontrast}, i.e., given a set of matched points $\mathcal{P}$, and two features of dimension $d$, it is formulated as follows:
\begin{align}
&\mathcal{L}_{\text{NCE}} = - \sum_{(i,j) \in \mathcal{P}} \log \dfrac{\exp( d(\mathbf{F}_i , \mathbf{G}_j) / \tau)}{\sum_{(\cdot, k) \in \mathcal{P}} \exp( d(\mathbf{F}_i , \mathbf{G}_k) / \tau)}
\label{eq:PointInfoNCELoss}
\end{align}
where $\tau$ is a temperature parameter, and $d(\cdot,\cdot)$ is the Euclidean distance between the two features. The purpose of this loss is to force the distance between the features of the matched points to be minimized, while this distance must be maximized between the unmatched points. The NCE loss is applied to each point individually and thus cannot penalize the consistency of the matches at all. To remedy this, the LIE loss \cite{Marin2020CorrespondenceLV} is used. Given the extracted features $\mathbf{F}$ and $\mathbf{G}$, and the coordinate \textit{xyz} of the shape $\mathcal{M}$ represented by the matrix $\mathbf{M} \in \mathbb{R}^{m \times 3}$, we first compute the soft correspondences matrix $S_{\mathcal{M}\mathcal{N}}$ and then formulate the LIE loss as follows
\begin{align}
&\mathcal{L}_{\text{LIE}} = \lVert S_{\mathcal{M}\mathcal{N}} \mathbf{M} - \Pi_{\mathcal{M}\mathcal{N}}^{gt} \mathbf{M}\rVert_2^2,
\label{eq:LIELoss}\\
&(S_{\mathcal{M}\mathcal{N}})_{ij} = \frac{\exp(- \lVert \mathbf{F}_i - \mathbf{G}_j \rVert_2)}{\sum_{k=1}^{n} \exp(- \lVert \mathbf{F}_i - \mathbf{G}_k \rVert_2}
\end{align}
where $\Pi_{\mathcal{M}\mathcal{N}}^{gt}$ is the ground truth p2p correspondence matrix.
\section{Motivation \& Method overview}
Our main goal is to develop an unsupervised learning-based shape correspondence method that would take as input a pair of shapes $\mathcal{M}$ and $\mathcal{N}$ and produce a dense correspondence, also called a map, $T: \mathcal{M} \rightarrow \mathcal{N}$. For this, we explore the direction of neural priors and show how they can be used for 3D shape matching. The motivation behind this direction is twofold. First, as in the 2D case \cite{Jo_2021_ICCV,Mataev2019DeepREDDI}, we hope that the structure of an untrained 3D neural network captures the low-level statistics of a single 3D shape, and thus can enable fine-tunning using a single pair. Second, as demonstrated in DIP \cite{UlyanovVL17}, deep neural networks have high noise impedance, which allows them to learn good features from noisy inputs. This property can be used to guide the network to good local minima while trying to use, in our case, noisy point-to-point (p2p) matches for supervision.
\textbf{Feature embeddings.} In this work, we formulate the shape-matching problem for both man-made and organic shapes, by learning pointwise features. Such features can be used to compute correspondences either through nearest neighbor search or with minimal post-processing.
Specifically, the network $N$ takes as input a 3D shape $\mathcal{M}$ and produces, as output, a feature vector $N(\mathcal{M})_x \in \mathbb{R}^{d}$, for every point of the shape $x \in \mathcal{M}$. When considering the entire object as a whole, we call the set $N(\mathcal{M})_x$ for all $x\in \mathcal{M}$ a \emph{feature embedding} of shape $\mathcal{M}$.
Our method is based on two main observations: \emph{neural bias} for untrained networks, and an effect that we call \emph{neural correspondence prior}. Below we describe each of these effects and then describe our unsupervised shape-matching method in \cref{sec:two_stage_algo}.
\vspace{-2mm}
\subsection{Neural bias for pointwise features}
\label{sec:neural_bias}
Our first observation is that feature embeddings produced by neural networks have a particular structure, which can be exploited in
the context of computing features for shape correspondence. To highlight this, we considered the test sets of the FAUST-Remeshed (\textsc{Faust}) and SCAPE-Remeshed (\textsc{Scape}) datasets introduced in \cite{ren2018continuous}, and used in many recent works \cite{donati2020deep,eisenberger2020deep,sharma2020weakly}, and computed the correspondences using the pointwise features extracted by a DiffusionNet network \cite{sharp2021diffusion} with randomly set weights. We use the default variant of DiffusionNet, which takes as input the XYZ coordinates of the shapes and produces a 128-dimensional feature vector for every point. We emphasize that the weights of the network are set randomly and \textit{without any training}. We then used the extracted features to produce p2p maps either via a simple nearest neighbor (NN) search in the feature space or with the functional map framework \cite{Ovsjanikov2012} (FMAP). We compare the results to maps produced using the same procedure with classical axiomatic features such as SHOT \cite{tombari2010unique}, or using other axiomatic and training-based non-rigid correspondence methods.
As shown in \cref{fig:ncp_demonstration}-Left, remarkably, the features produced by an \textit{untrained} DiffusionNet network perform on par or better than axiomatic features and even outperform the \textit{supervised} learning method, FMNet \cite{Litany2017}, based on training MLPs to refine pre-defined SHOT features. We attribute the fact that a randomly-initialized DiffusionNet performs better than \textit{trained} point-wise MLPs to the spatially-aware nature of the architecture of the network, which uses diffusion to simulate intrinsic convolution, thereby providing a strong neural prior \cite{NEURIPS2019_e9874147,saxe2011,Hanocka2020}.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\input{tables/faust_rand_init}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{0.49\textwidth}
\centering
\includegraphics[width=1.0\textwidth]{figures/noise_impedance_5005.pdf}
\end{subfigure}
\caption{\textbf{Motivation for the NCP effect.} \underline{Left}: Matching accuracy of DiffusionNet \textit{with random weights} (mean and standard deviation of three random runs), compared to baseline shape matching methods, on the test sets of \textsc{Faust} and \textsc{Scape}.
\underline{Right}: The evolution of the training NCE loss and geodesic error during optimization when using the ground truth map (map (a)), and noisy maps with different levels of noise (maps (b), (c) and (d)) as supervision. The legend reports the geodesic error of the input maps. Observe the noise impedance and geodesic error decrease of \textit{intermediate} maps.}
\label{fig:ncp_demonstration}
\vspace{\figmargin}
\end{figure}
\vspace{-2mm}
\subsection{NCP: Neural Correspondence Prior}
\label{sec:npc}
While the previous effect relates to properties of untrained networks, in this work, we also observe another, complementary phenomenon that we call \textit{Neural Correspondence Prior}, and which we exploit in our approach below. This effect can be formulated as follows. Suppose we are given a fixed map $T_{\mathcal{M}\mathcal{N}}$, between some shapes $\mathcal{M}$ and $\mathcal{N}$. We can formulate an optimization problem, where we aim to learn the feature embedding of $\mathcal{M}$ and $\mathcal{N}$ so that the induced map, e.g., computed via nearest neighbors in the feature space: i.e. $x \rightarrow \argmin_{y} \| N(\mathcal{M})_x - N(\mathcal{N})_y \|$, is as close as possible to $T_{\mathcal{M}\mathcal{N}}$. Note that unlike the neural bias mentioned in the previous section, here we formulate an optimization problem, where the parameters of the network $N$ are optimized to fit the given input map. Our key observation is that when the target map $T_{\mathcal{M}\mathcal{N}}$ is noisy or approximate, networks tend to produce well-structured feature embeddings until the very late stages of the optimization. I.e., throughout the early stages of training, the maps computed by the optimized features tend to be of \textit{higher accuracy than the input} noisy map used for supervision.
To demonstrate the effect of NCP quantitatively, we performed an experiment in which we corrupted the ground truth map between a pair of shapes from the \textsc{Faust} dataset, and then trained the DiffusionNet network, with a gradient descent optimizer using the standard NCE loss \cite{xie2020pointcontrast} to overfit to those corrupted maps. Specifically, to supervise this training, we use: (\textit{a}) the ground truth (GT) map between the shapes, (\textit{b}) the GT map where 25\% of its entries have been assigned to random vertices, (\textit{c}) the GT map with 50\% noise, (\textit{d}) the GT map with 75\% noise. The results of this experiment are shown in \cref{fig:ncp_demonstration}-Right. We plot both the NCE loss being optimized as well as the \emph{geodesic error} of the intermediate maps with respect to the ground truth.
As shown in \cref{fig:ncp_demonstration}-Right, while the network can easily adapt to the correct map, it has difficulty minimizing the loss in the case of noisy maps. Moreover, and perhaps more remarkably, observe that the geodesic error starts to decrease \textit{towards lower values than the input map}, in the first iterations of optimization and diverges only in the later stages, when the network starts to overfit to the noise in the training map. This implies that such optimization with a neural network and early stopping can be used as an effective map denoising mechanism (\cref{sec:test_time_denoising}). It should be mentioned that this effect does not depend on the shape pair, and other examples of different pairs are provided in \cref{sec:exp_test_time_optim} and the supplementary.
We observe this effect to hold very broadly in shape correspondence problems. For a single shape pair with a noisy input map, eventually, any map can be learned by the network if the parametrization and the number of training iterations are sufficient. However, the network architecture strongly regularizes the map search space, providing low impedance to `signal' and high impedance to `noise', resulting in an optimization trajectory that either converges to a good local minimum or passes near one \cite{UlyanovVL17}. Furthermore, in the presence of \textit{shape collections}, this effect is even stronger and, as we demonstrate below, significantly helps to regularize computed features so that\textit{ no early stopping} or post-processing of the results becomes necessary (see our \cref{algo1} that we explain in detail in \cref{sec:two_stage_algo}).
In addition to the network architecture itself, we attribute the \textit{Neural Correspondence Prior} to the spectral bias principle \cite{rahaman19a,Cao2021TowardsUT}, which states that neural networks tend to learn low frequencies in the early stages of training and that low frequencies are more robust to random perturbations of network parameters.
Since we formulate shape matching as the problem of computing optimal features, spectral bias, in our context, implies that even given a noisy or artifact-laden map as input, the \textit{feature embeddings} produced by the optimized neural network tend \textit{to be smooth}, especially in the early stages of the optimization. In other words, it is significantly harder for the network to produce a noisy, high-frequency feature embedding than a smooth one. Furthermore, a smooth feature embedding will tend to produce smooth correspondences between shapes, as suggested by the following theorem, which we state here and prove in the supplementary.
\begin{theorem}
\label{thm_1}
Let $\mathcal{M}$ and $\mathcal{N}$ be two compact smooth surfaces (smooth manifolds of dimension 2). Let $\mathbf{M}$ and $\mathbf{N}$ be their feature embeddings in $\mathbb{R}^d$, given by some functions: $\psi: \mathcal{M} \rightarrow \mathbb{R}^d$ and $\phi: \mathcal{N} \rightarrow \mathbb{R}^d$, so that $\mathbf{M} = \psi (\mathcal{M})$ and $\mathbf{N} = \phi (\mathcal{N})$. For example, $\psi$ and $\phi$ can be given by some neural network. Suppose that $\psi$ and $\phi$ are both smooth and injective. Then up to arbitrarily small perturbations of $\phi, \psi$, the map $T_{nn}: \mathcal{M} \rightarrow \mathcal{N}$ given by $T_{nn}(x) = \argmin_{y \in \mathcal{N}} \| \psi(x) - \phi(y)\|$ must be smooth up to sets of measure 0 on $\mathcal{M}$.
\end{theorem}
This result highlights the fact that smooth feature embeddings, generically, translate into smooth maps, which is a desirable property in most typical correspondence scenarios.
\section{Method description}
Based on the observations above, in this section, we introduce contexts in which the NCP can be exploited, resulting in two algorithms: one for unsupervised shape matching, and the second for p2p map denoising.
\vspace{\secmargin}
\subsection{NCP within a shape collection}
\label{sec:two_stage_algo}
Given a collection of shapes $\{\mathcal{N}_i\}$, we propose to exploit the NCP in two ways. The first configuration assumes that we are also given a collection of artifact-laden maps $\{T_j\}$ between some shape pairs in $\{\mathcal{N}_i\}$ as input. In that case, we propose to train neural network to produce feature embedding for each shape, s.t., the induced maps are as close as possible to the target maps $\{T_j\}$. We provide the exact choice of the loss function depending on the nature of the underlying shapes in \cref{sec:application} below.
In the second configuration, no maps are given as input. In this case, we propose a two-stage pipeline where in the first stage, we adapt an existing unsupervised correspondence method to obtain possibly approximate very noisy initial correspondences. We then formulate a supervised learning problem, where we learn feature embeddings that would induce these computed correspondences as a target. Once the network is trained, we establish final correspondences via a simple nearest neighbor search in the feature space. Note that the trained network can be used to establish correspondences both between shapes within the given shape collection, but also \textit{across new, never seen shape pairs.} In our evaluation below, and unless otherwise noted, we always evaluate on a test set, never seen during training. We dubbed the resulting algorithm \textbf{NCP-UN},
and we summarize it in \cref{algo1} and \cref{fig:teaser}. This is the algorithm that is used throughout most of our experiments.
\setlength{\textfloatsep}{5mm}%
\begin{algorithm}[t!]
\caption{NCP-UN: \textbf{N}eural-\textbf{C}orrespondence-\textbf{P}rior based \textbf{UN}supervised Shape Matching}\label{algo1}
\hspace*{\algorithmicindent} \textbf{Input:} Collection of shapes $\{\mathcal{N}_i\}$\\
\hspace*{\algorithmicindent} \textbf{Output:} p2p maps between shapes of the test set of $\{\mathcal{N}_i\}$\\
\begin{algorithmic}[1]
\vspace{2.5mm}
\rlap{\smash{$
\hspace{-4.5em}
\color{red}\begin{tabular}{l}Stage 1\end{tabular}
\color{red}{\tiny{\left\{\begin{array}{@{}l@{}}\\{}\end{array}\right.}}%
$}}
\vspace{-5.5mm}
\State Train a variant of an off-the-shelf unsupervised matching method $\mathbf{UM}$ on the train set of $\{\mathcal{N}_i\}$
\State Predict p2p maps $\{T_i\}$ on the train set of $\{\mathcal{N}_i\}$ using trained $\mathbf{UM}$
\State Use $\{T_i\}$ to supervise the training of a randomly-initialized neural network $\mathbf{RN}$ with a p2p loss
\State Predict Feature Embedding $\{\mathbf{RN}(\mathcal{N}_i)\}$ on the test set of $\{\mathcal{N}_i\}$ using the trained $\mathbf{RN}$
\rlap{\smash{$
\hspace{-40.6em}
\color{red}\begin{tabular}{l}Stage 2\end{tabular}
\color{red}\left\{\begin{array}{@{}l@{}}\\{}\\{}\end{array}\right.%
$}}
\State Use $\{\mathbf{RN}(\mathcal{N}_i)\}$ to establish correspondences either using the nearest neighbor or functional map pipeline.
\end{algorithmic}
\end{algorithm}
We argue that it is advantageous to perform Stage 2 simultaneously on an entire shape collection, instead of doing it for each pair individually for two main reasons: first, it is faster and avoids the need for test-time optimization (see \cref{sec:test_time_denoising}), and second, the collection provides an additional layer of regularization. In fact, since we optimize for features on each shape, it is difficult to compute features that would overfit to all $O(N^2)$ noisy maps in a collection of $N$ shapes, as the artifacts across pairs are typically not consistent. We observe that the noise impedance property is even stronger in the case of a collection, and the produced maps are better than if the refinement is done per pair.
\vspace{-2mm}
\mypara{Intuition behind Stage 2 of \cref{algo1}}
Our intuition for Stage 2 is threefold.
First, and motivated by the results of \cref{sec:npc} (the NCP effect), we showed that neural networks tend to have difficulty overfitting to corrupted inputs, and tend to favor smooth outputs. Thus, in the process of over-fitting to approximate maps by a neural network, the output maps tend to be of \textbf{better quality} than the input. This motivates the use of artifact-laden maps as supervision for a network thus capitalizing on the NCP effect to recover better correspondences.
Second, as we mentioned above, given a collection of approximate maps, their errors tend to be inconsistent, and training a network using such maps for supervision adds an extra layer of regularization, as it is difficult for the network to adapt to all inconsistent errors on all pairs at once.
Finally, in our formulation of Stage 2, we learn \textit{one feature embedding per shape} and compute correspondences via nearest neighbor search between feature embeddings. This also adds strong regularization, since it makes it difficult for the network to learn a feature embedding, which can reproduce errors in the maps between all shape pairs.
For our applications (see \cref{sec:application}), we used the following design choices. For the $\mathbf{UM}$ method, we used the unsupervised functional map method \cite{roufosse2019unsupervised,sharma2020weakly}, with a DiffusionNet \cite{sharp2021diffusion} backbone for 3D meshes, and Point-MLP \cite{rethink_ma_22} backbone for point clouds. The same backbone was used as $\mathbf{RN}$ for step 3. in \cref{algo1}. Concerning the p2p loss in step 3, we used the LIE loss \cite{Marin2020CorrespondenceLV} for 3D meshes and point clouds.
We also used a variant of our approach, dubbed NCP-UN$_{\text{fmap}}$, that uses the FMAP framework and loss from \cite{donati2020deep} in steps 3 and 5 of
\cref{algo1}, especially on organic shapes, where the Laplacian basis tends to be of high quality.
We want to emphasize that our observations regarding NCP are independent of the choice of loss and network architecture. A demonstration of the generality of NCP, as well as more details about the implementation, are provided in the supplementary. Across \textit{all} cases, we observe a significant improvement in results when using the second stage of our NCP-UN pipeline (supervised learning) compared to the initial input maps.
Finally, we want to emphasize that it is possible to repeat Stage 2 several times, and we observed that there is a slight improvement in some cases, but most often it stagnates after the first iteration. We experimented with this and chose to keep the method simple and use only one iteration, thus, avoiding an additional tunable hyperparameter.
\subsection{Test time denoising}
\label{sec:test_time_denoising}
In addition to the previous setting, we also observe that NCP can be used to denoise a correspondence when given a single shape pair. Specifically, given a noisy initial map, we train a neural network to produce feature embeddings that would induce the given input map, used as supervision. Again, we tailor the exact choice of loss to the nature of the shapes and describe one option in \cref{sec:non-rigid-corr} below.
Test time denoising can be advantageous in the absence of a shape collection. At the same time, as mentioned in the original DIP work \cite{UlyanovVL17}, in the case of a single shape pair, early stopping must be used in order to stop training at the best local minimum. For the shape-matching task, we use a cycle loss as the criterion for stopping. Given a pair of shapes $\mathcal{M}, \mathcal{N}$, and two maps between them $\Pi_{\mathcal{M}\mathcal{N}}$ and $\Pi_{\mathcal{N}\mathcal{M}}$ computed via nearest neighbors between optimized features, and represented as soft binary matrices (see
Appendix A
for computation details), we compute $L_{cycle} = \| X_{\mathcal{N}} - \Pi_{\mathcal{N}\mathcal{M}} \Pi_{\mathcal{M}\mathcal{N}} X_{\mathcal{N}} \|_F^2$, where $X_{\mathcal{N}}$ is the matrix of XYZ coordinates of shape $\mathcal{N}$. We overfit the network for a number of iterations and stop when $L_{cycle}$ stops improving after a predetermined patience period. The weights that produce the minimum value of $L_{cycle}$ are used to compute feature embeddings that we then use to compute the p2p map, via nearest neighbor search.
\subsection{Correspondences as a tool for downstream tasks}
\label{subsec:downstream}
Given computed correspondences, we also use them to solve multiple downstream tasks. Indeed, since our proposed algorithm is unsupervised, we can use its output, i.e., p2p maps, to transfer multiple quantities between shapes, such as keypoints. For this, we only need a few labeled examples with the quantities we want to transfer, resulting in a few-shot algorithm.
In the following, we present a method for few-shot keypoint detection, dubbed \textbf{FSKD} hereafter, and based on the p2p maps obtained by our unsupervised algorithm \textbf{NCP-UN}{}. A particular challenge in this setup is that some keypoints and parts appear exclusively on some shapes and not in others, and so to account for the keypoint appearing there, multiple labeled source shapes must be used.
We begin our \textbf{FSKD} by randomly selecting $N$ shapes that cover the largest set of keypoints. These $N$ shapes will be labeled and we transfer their keypoints to the shapes in the test set, in order to ``discover'' their keypoints.
Given a target shape, \textbf{FSKD} is composed of three steps: \textbf{1.} Detection of potential keypoints by transferring keypoints from labeled shapes, \textbf{2.} Filtering to remove keypoints that are likely not to exist on the target shape, \textbf{3.} Combination: merge transferred keypoints if multiple points on the target shape are assigned to the same keypoint ID. We provide the implementation details for these three steps in the supplementary (see
Appendix C
).
\section{Results \& Applications}
\label{sec:application}
In this section, we provide results in a wide range of challenging tasks, showing the efficiency and robustness of our approach to different types of data and tasks. In particular, we consider the tasks of shape correspondence between man-made shapes in \cref{sec:rigid-corr}, non-rigid shape correspondence in \cref{sec:non-rigid-corr},
part segmentation in \cref{sec:part-seg}, and few-shot keypoint detection in \cref{sec:fs-keypoint-dec}. The fact that our model does not rely on a geometric prior allows it to provide good results on different types of data, including isometric and non-isometric, organic, and man-made data.
\mypara{Datasets}
\label{sec-dataset}
We conducted our experiments on four different datasets. To the best of our knowledge, there is no point cloud dataset that provides dense point-to-point ground-truth correspondences. For this purpose, we follow the setup of \cite{cheng2021learning}, and use \textsc{KeyPointNet} dataset \cite{keypointnet2020} for the man-made shape correspondence experiment in \cref{sec:rigid-corr}. We also use this dataset for our few-shot keypoint detection experiment in \cref{sec:fs-keypoint-dec}.
\textsc{KeypointNet} is composed of over 8K models and 100K semantic keypoints labeled by professional humans.
We use the same train/validation/test splits as the original paper \cite{keypointnet2020}. For the part segmentation experiment in \cref{sec:part-seg}, we use the \textsc{PartNet} dataset \cite{partnet_Yi16}, which is composed of 16881 shapes of 16 categories, where each category has a different number of parts ranging from 2 to 6.
For 3D meshes, we use the organic, non-isometric non-rigid \textsc{Smal} \cite{Zuffi:CVPR:2017,marin22_why} dataset and the \textsc{Shrec'20} Non-Isometric benchmark \cite{Dyke2020} for our non-rigid correspondence experiment in \cref{sec:non-rigid-corr}, as these are challenging \textit{non-isometric} datasets. The former dataset is composed of 50 animal shapes represented as 3D meshes. 25 shapes are used for training, and the remaining 25 are used for testing.
It should be noted that the set of animals and poses seen in training is different from the one used in testing. The \textsc{Shrec'20} benchmark consists of 14 scans of different animals, some of which contain holes and topological deformations. This, in addition to the fact that the ground truth maps are given as a set of sparsely annotated keypoints, means that most supervised methods cannot be applied.
\subsection{3D shape correspondences}
\subsubsection{Correspondences on point clouds}
\label{sec:rigid-corr}
We test the quality of the matches produced by NCP-UN on sparse point clouds using the \textsc{KeyPointNet} dataset. We train the first stage of our method using the unsupervised functional map method \cite{roufosse2019unsupervised,sharma2020weakly}, with a PointMLP backbone \cite{ma2022rethinking} and XYZ coordinates as input, and impose the bijectivity loss on the functional map. The second stage is trained using LIE loss, and maps are extracted using the nearest neighbor in the feature space.
For testing, we follow the same setup as \cite{cheng2021learning}, by generating pairs of shapes between all point clouds in a given category and removing the pairs that do not share the same set of semantic keypoints.
We compute the L2 distance between the transferred keypoint and the ground truth one.
Our method is evaluated against state-of-the-art learning-based 3D dense correspondence prediction approaches, including AtlasNetV2 \cite{AtlasNetV2}, FoldingNet \cite{Yang2018}, IDC \cite{liu2020learning}, CPAE \cite{cheng2021learning}, in addition to the nearest neighbor baseline between the XYZ coordinates of the points, which turns out to be a competitive baseline, not considered by previous methods.
In \cref{fig:kpt_matching}, we report the percentage of testing pairs where the distances between predicted and ground truth maps are below a given threshold for 4 categories and report the remaining categories in the supplementary.
Our method outperforms all methods in 13 out of 16 categories.
Note that our method obtains SOTA results although the method used in stage 1 is not fully designed to work on point clouds (the quality of the Laplacian is not good, and the used bijectivity loss is weak).
We believe that using a more adapted method for the first stage will further improve the results.
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth]{figures/keypointnet_4x1.pdf}
\caption{Correspondence accuracy on 4 categories in the \textsc{KeyPointNet} dataset.
\label{fig:kpt_matching}}
\vspace{-2mm}
\end{figure}
\subsubsection{Correspondences between non-rigid meshes}
\label{sec:non-rigid-corr}
For the non-rigid non-isometric application, we trained the unsupervised functional map network from \cite{sharp2021diffusion} and used LIE loss for the second stage (step 3. in \cref{algo1}). P2P maps were obtained using nearest neighbors between feature embeddings.
Because the \textsc{Shrec'20} dataset is limited (only 14 shapes), we follow previous work \cite{Eisenberger2021NeuroMorphUS} by training and testing on the same set, whereas for \textsc{Smal}, the training and testing sets are different. For a fair evaluation, we compared our method to many state-of-the-art supervised and unsupervised methods. For the supervised methods, we compared to FMNet \cite{litany2017deep}, GeomFMaps \cite{donati2020deep}, GeomFMaps + DiffusionNet \cite{sharp2021diffusion} and LIE \cite{Marin2020CorrespondenceLV}. For the unsupervised methods, we compared our method to
WSupFMNet \cite{sharma2020weakly}, Smooth Shells \cite{Eisenberger2020SmoothSM}, Deep Shells \cite{eisenberger2020deep}, NeuroMorph \cite{Eisenberger2021NeuroMorphUS} and WSupFMNet + DiffusionNet \cite{sharp2021diffusion}.
As shown in \cref{fig:noniso-matching}, our method achieves state-of-the-art results among unsupervised methods and
supervised methods (we follow the standard Princeton protocol \cite{Kim2011} for evaluation). The majority of unsupervised methods and supervised methods fail because of the near-isometry assumption they make. Since our method makes no assumptions about the topology of the shape, it can perform very well in difficult scenarios such as non-isometric animals. It should be noted that these results would not be possible without the second stage of our algorithm, which is enabled by NCP. For example, for the \textsc{Smal} dataset, the result of the first stage is \textbf{7.8}, and the second stage improves this result to \textbf{5.8}, which is a \textbf{25\%} improvement, achieving state-of-the-art results.
We provide in \cref{fig:smal_quali} some qualitative results showing the performance of our algorithm against the state-of-the-art NeuroMorph method on the \textsc{Smal} dataset. It can be seen that our method produces visually plausible maps, especially after applying the second stage.
\begin{figure}
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\input{tables/animals_geod_err}
\end{subfigure}%
\hfill
\begin{subfigure}{0.52\textwidth}
\centering
\includegraphics[width=1.0\textwidth]{figures/shrec20_princeton.pdf}
\end{subfigure}
\caption{\textbf{Non-isometric matching results.} \underline{Left}: Matching accuracy on the \textsc{Smal} test set. Values are mean geodesic error $\times 100$ on unit-area shapes. \underline{Right}: Correspondence quality of different methods on the test set of \textsc{Shrec'20} Non-Isometric dataset.}
\label{fig:noniso-matching}
\vspace{-2mm}
\vspace{\figmargin}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figures/animals_test_time.pdf}
\caption{\textbf{Qualitative results on \textsc{Smal} dataset}, showing the maps produced by Stage 1, and refined by Stage 2, either with test time optimization (see \cref{sec:test_time_denoising}) or using NCP-UN (see \cref{algo1})}
\label{fig:smal_quali}
\vspace{-1mm}
\end{figure}
\mypara{Test time optimization}
\label{sec:exp_test_time_optim}
To show the utility of test time optimization, as well as the early stopping criteria we introduced, we performed the following experiment. After training the first stage of our algorithm, the trained model was used to predict p2p on the test set of \textsc{Smal} dataset. Pairs of shapes are then created, and we use our test time optimization method (see \cref{sec:test_time_denoising}) to improve upon the maps produced by the first stage. Results are reported in \cref{fig:noniso-matching}-Left, see row ``Ours (Test-Time Optimization)''. It can be seen that similar to the previous case, training a random network on top of artifact-laden maps helps to denoise them. We also show in \cref{fig:smal_quali} some qualitative results that illustrate the effect of test time optimization and compare it to NCP-UN.
The results of this experiment confirm our hypothesis that doing the second stage on a collection, instead of individual pairs, provides a second layer of regularization, as the results are better, in addition to improving computational complexity. In fact, the second stage using the whole collection takes around \textbf{15 minutes}, meanwhile, doing it on each pair individually takes around \textbf{160 minutes}, which represents more than $10\times$ improvement in running time.
\subsection{Part segmentation transfer}
\label{sec:part-seg}
We further validate our approach on the part label transfer task on the \textsc{PartNet} dataset following the setup of \cite{cheng2021learning}. Results are summarized in \cref{tab:pair_part_seg}, where the average intersection-over-union IOU is reported. Our method outperforms the baselines on 14 out of 16 categories, achieving state-of-the-art results by more than 3\% IOU on average. In classes with large intra-class variations such as Lamps, our method outperforms the baselines by up to 8 \% average IOU points. In addition to being performant, our method is extremely fast. In fact, training both stages of our method takes on average \textbf{1 hour} (see the computational specifications in
Appendix B
), which is more efficient in comparison to the baselines, where CPAE for example takes around 24 hours to be trained for one category.
\input{tables/pair_part_seg}
\subsection{Few-shot 3D keypoint detection}
\label{sec:fs-keypoint-dec}
We evaluated our few-shot keypoint detection algorithm \textbf{FSKD} on the \textsc{KeyPointNet} dataset, following the setup of \cite{you2020ukpgan}, and using only \emph{3 labeled source shapes} (the resulting algorithm is thus a three-shot keypoint detection algorithm). The goal of this task is to predict keypoints that correlate with human annotation. The performance is evaluated by the mean Intersection over Union (mIoU). The intersection is counted if the geodesic distance between a predicted keypoint and its nearest ground truth keypoint is less than some geodesic threshold. The union is simply the union of the detected keypoints and ground truth keypoints.
We compare \textbf{FSKD} against multiple baselines, including SIFT-3D \cite{Rister2017}, HARRIS-3D \cite{Sipiran2011}, ISS \cite{Zhong2009}, D3Feat \cite{bai2020d3feat}, USIP \cite{Li2019} and UKPGAN \cite{you2020ukpgan}. Results are summarized in \cref{fig:few-shot-detection} - left. It can be seen that our simple, correspondence-based method outperforms the baselines tailored specifically for this task. In particular, our method achieves state-of-the-art results for a threshold superior to 0.035, using only \textbf{three} labeled shapes. We provide some qualitative results in \cref{fig:few-shot-detection} - right.
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth]{figures/fksd_big.pdf}
\caption{\textbf{Few-shot keypoint detection results}.\textbf{(a)} mIoU results of our simple few-shot keypoint detection algorithm on \textsc{KeypointNet} using only 3 labeled examples. Our simple method achieves good results and is on par with or better than recent keypoint detection methods. \textbf{(b)} Qualitative results on the table category, showing the shapes used for training (N=3, left), human annotations on the testset (center) as well as our predictions (right) on four examples.}
\label{fig:few-shot-detection}
\end{figure}
\section{Conclusion \& Limitations}
\label{sec:conclusion}
In this work, we showed that using noisy maps as a supervisory neural network training signal for the 3D shape-matching task can lead to significantly higher quality correspondences. We named this effect Neural Correspondence Prior (NCP). Through extensive experiments, we shed light on the properties of NCP and developed a two-stage algorithm for unsupervised 3D shape matching. We demonstrate the effectiveness and generality of our algorithm on a range of shape-matching problems, including unsupervised shape matching on both man-made and organic non-rigid shapes, achieving state-of-the-art results on a wide range of benchmarks.
One limitation of our approach is that it assumes that the p2p maps used to train the second stage of the algorithm still have some structure. In fact, if the unsupervised method used in the first stage fails completely and provides random maps, our method will not converge. This problem can be mitigated by using a good application-specific unsupervised method for the first stage, especially with the recent interest and increase in unsupervised 3D learning methods.
\mypara{Acknowledgements}
The authors would like to thank Fr\'ed\'eric Chazal and Mathijs Wintraecken for many useful discussions related to Theorem 1. We also acknowledge the anonymous reviewers for their valuable suggestions. Parts of this work were supported by the ERC Starting Grant No. 758800 (EXPROTEA) and the ANR AI Chair AIGRETTE.
\section{Background on functional maps \& Notation}
\label{sec:background}
Our work uses the functional map framework as a first estimator for p2p maps, and multiple losses on point-to-point (p2p) maps to learn robust features that allow extracting good correspondences using the nearest neighbor in feature space. We provide a brief overview in the next section.
\paragraph{Functional maps}
\label{sec:fmap_review}
The functional map (fmap) framework was used for the first stage of our \textbf{NCP-UN}{} algorithm. For this, we follow the general strategy of recent fmap-based techniques \cite{donati2020deep,attaiki2021dpfm,sharma2020weakly,roufosse2019unsupervised,sharp2021diffusion}, as follows: given source and target shapes $\mathcal{M}$ and $\mathcal{N}$, represented as either triangular meshes or point clouds, having $m$ and $n$ vertices respectively, we pre-compute their Laplace-Beltrami operator \cite{sharp2020laplacian}, and store their first $k$ eigenfunctions in the matrices $\Phi_{\mathcal{M}} \in \mathbb{R}^{m \times k}$ and $\Phi_{\mathcal{N}} \in \mathbb{R}^{n \times k}$ respectively. Using a siamese network $\mathcal{F}_{\theta}$, we compute for each shape a $d$-dimensional descriptor $\mathcal{F}_{\theta}(\mathcal{M}) = \mathbf{F} \in \mathbb{R}^{m \times d}$ and $\mathcal{F}_{\theta}(\mathcal{N}) = \mathbf{G} \in \mathbb{R}^{n \times d}$ respectively. These descriptors are then projected to the spectral domain to form the spectral features $\mathbf{A} = \Phi_{\mathcal{M}}^{\dagger} \mathbf{F}$ and $\mathbf{B} = \Phi_{\mathcal{N}}^{\dagger} \mathbf{G}$ ($\bullet^{\dagger}$ is the Moore pseudo-inverse). A functional map is then computed by solving the following linear system:
\begin{align}
\label{eq:fmap_energy}
C_{opt} = \argmin_{C} \| C\mathbf{A} - \mathbf{B} \| + \lambda \|C \Delta_{\mathcal{M}} - \Delta_{\mathcal{N}} C \|.
\end{align}
where $\Delta_{\mathcal{M}},\Delta_{\mathcal{N}}$ are diagonal matrices of Laplace-Beltrami eigenvalues of the corresponding shapes and $\lambda$ is a scalar hyper-parameter.
Following the unsupervised literature \cite{roufosse2019unsupervised,sharma2020weakly,sharp2021diffusion}, the siamese network $\mathcal{F}_{\theta}$ is trained by imposing structural properties on the fmap $C$ such as bijectivity and orthogonality on the shape pairs in the training set. In fact, given the fmap $C_{\mathcal{M}\mathcal{N}}$ from $\mathcal{M}$ to $\mathcal{N}$, and $C_{\mathcal{N}\mathcal{M}}$ from $\mathcal{N}$ to $\mathcal{M}$, the bijectivity loss is formulated as $\|C_{\mathcal{M}\mathcal{N}}C_{\mathcal{N}\mathcal{M}} - \mathbb{I}_k \|_2^2$, and the orthogonality loss is $\|C_{\mathcal{M}\mathcal{N}}^{\top} C_{\mathcal{M}\mathcal{N}} - \mathbb{I}_k\|_2^2$. $\mathbb{I}_k$ denotes the identity matrix of size $k$ and $\bullet^{\top}$ is the matrix transpose operator.
\paragraph{Feature learning}
The goal of feature learning is to learn robust descriptors that can allow direct nearest-neighbor matching in the descriptor space. In this work, we use two losses: PointInfoNCE and the LIE loss.
PointInfoNCE \cite{xie2020pointcontrast} is a contrastive loss such that, given a set of matched points $\mathcal{P}$, and two features of dimension $s$, it is formulated as follows:
\begin{align}
&\mathcal{L}_{\text{NCE}} = - \sum_{(i,j) \in \mathcal{P}} \log \dfrac{\exp( d(\mathbf{F}_i , \mathbf{G}_j) / \tau)}{\sum_{(\cdot, k) \in \mathcal{P}} \exp( d(\mathbf{F}_i , \mathbf{G}_k) / \tau)}\\
& d(\mathbf{F}_i , \mathbf{G}_j) = \|\mathbf{F}_i - \mathbf{G}_j\|_2^2
\label{eq:PointInfoNCELoss}
\end{align}
where $\tau$ is a temperature parameter, and $d(\cdot,\cdot)$ is the Euclidean distance between the two features. In all our experiments, we took $\tau = 0.07$. The purpose of this loss is to force the distance between the features of the matched points to be minimized, while this distance must be maximized between the unmatched points. The NCE loss is applied to each point individually and thus cannot penalize the overall consistency of the matches.
To remedy this, especially when the number of vertices of the shapes is moderate, as is the case for sparse point clouds, we use the LIE loss introduced in \cite{Marin2020CorrespondenceLV}. Given the extracted features $\mathbf{F}$ and $\mathbf{G}$, and the coordinate \textit{xyz} of the shape $\mathcal{N}$ represented by the matrix $\mathbf{N} \in \mathbb{R}^{n \times 3}$, we first compute the soft correspondences matrix $S_{\mathcal{M}\mathcal{N}}$, and then formulate the LIE loss as follows:
\begin{align}
&\mathcal{L}_{\text{LIE}} = \lVert S_{\mathcal{M}\mathcal{N}} \mathbf{N} - \Pi_{\mathcal{M}\mathcal{N}}^{gt} \mathbf{N}\rVert_2^2,
\label{eq:LIELoss}\\
&(S_{\mathcal{M}\mathcal{N}})_{ij} = \frac{\exp(- \lVert \mathbf{F}_i - \mathbf{G}_j \rVert_2)}{\sum_{k=1}^{n} \exp(- \lVert \mathbf{F}_i - \mathbf{G}_k \rVert_2}
\end{align}
where $\Pi_{\mathcal{M}\mathcal{N}}^{gt}$ is the ground truth p2p correspondence matrix. This loss forces the soft correspondences matrix to be as close as possible to the ground truth map, by forcing their action (pull-back) on the shape coordinates, thus taking into account the geometry of the shape. Indeed, erroneous predictions that are geometrically close to the ground truth are penalized less than those that are far from it in terms of L2 distance.
\section{Implementation details}
\label{sec:imp_details}
In Sec. 3.1 in the main text, we used a randomly initialized DiffusionNet \cite{sharp2021diffusion} network to predict features on the test set of FAUST-Remeshed (\textsc{Faust}) and SCAPE-Remeshed (\textsc{Scape}) datasets \cite{ren2018continuous}. For that, we used the publicly available implementation of DiffusionNet released by the authors \footnote{\url{https://github.com/nmwsharp/diffusion-net}}. Unless specified otherwise, all our experiments on 3d-triangular meshes use 4 DiffusionNet blocks of width 128, the input to the network is the XYZ coordinates of the shape. For the competing features, both the Heat Kernel Signature (HKS) \cite{sun2009concise} and the Wave Kernel Signature (WKS) \cite{aubry2011wave} were sampled at 100 values of energy t, logarithmically spaced in the range proposed in their respective original papers. SHOT descriptors \cite{tombari2010unique} are 352-dimensional, and we used the implementation provided by the PCL library \cite{Rusu_ICRA2011_PCL}. For the Laplace-Beltrami computation, we used the discretization introduced in \cite{sharp2020laplacian} for both 3D meshes and point clouds.
In Sec. 3.2 of the main text, we train a DiffusionNet to produce feature embeddings that will induce the maps used for supervision, using the NCE loss. For this experiment and all the following learning experiments, ADAM optimizer \cite{kingma2017adam} was used with a learning rate of 0.001.
In Sec. 5.1.1 and 5.2 of the main text, we applied our \textbf{NCP-UN}{} algorithm on the \textsc{KeyPointNet} \cite{keypointnet2020} and the \textsc{PartNet}~\cite{partnet_Yi16} datasets. For the first stage, we used the unsupervised geometric functional map from \cite{sharp2021diffusion} but used PointMLP \cite{rethink_ma_22} as a feature extractor, instead of DiffusionNet, as it is better suited to the point cloud context. We used the default segmentation configuration provided by the authors \footnote{\url{https://github.com/ma-xu/pointMLP-pytorch}}. In \cref{eq:fmap_energy}, we take $\lambda = 0$. The network is trained using the bijectivity loss presented in \cref{sec:background},
as well as a new unsupervised loss based on the chamfer distance that we introduced. Indeed, we compute the chamfer distance between the source shape's XYZ coordinates, and a new version of this shape created by transferring its coordinates into the spectral space of the target shapes and back again. Formally, using the same notation as above, it is as follows:
\begin{align}
\mathcal{M}_{source} &= \Phi_{\mathcal{M}}\Phi_{\mathcal{M}}^{\dagger} \mathbf{M} \\
\mathcal{M}_{target} &= \Phi_{\mathcal{M}} C_{\mathcal{N}\mathcal{M}} C_{\mathcal{M}\mathcal{N}} \Phi_{\mathcal{M}}^{\dagger} \mathbf{M} \\
L_{\text{chamfer}}(\mathcal{M}_{\text{source}}, \mathcal{M}_{\text{target}}) &= \sum_{m \in \mathcal{M}_{\text{source}}} \min_{n \in \mathcal{M}_{\text{target}}} \left\| m - n \right\|_2 + \sum_{n \in \mathcal{M}_{\text{target}}} \min_{m \in \mathcal{M}_{\text{source}}} \left\| m - n \right\|_2
\end{align}
For the second stage, we used the LIE loss, and we trained a randomly initialized PointMLP, which has the same architecture as the first stage, to produce feature embeddings that induce the input maps.
In Sec. 5.1.2 of the main text, \textbf{NCP-UN}{} was applied on the \textsc{Smal} \cite{Zuffi:CVPR:2017,marin22_why} and \textsc{Shrec'20} \cite{Dyke2020} datasets. The first stage was performed using the unsupervised geometric functional map \cite{sharp2021diffusion} with the DiffusionNet backbone while applying the bijectivity and orthogonality losses, as described above in \cref{sec:fmap_review}. In \cref{eq:fmap_energy}, we set $\lambda = 10^{-3}$ for \textsc{Smal}, and $\lambda = 0$ for \textsc{Shrec'20}.
The second stage was performed using the LIE loss and the same backbone as the first stage. P2P maps were extracted using either the nearest neighbor in the space of features or using the functional map pipeline. In fact, given two feature embedding for two shapes $\mathcal{M}$ and $\mathcal{N}$, we compute the functional map $C_{\mathcal{M}\mathcal{N}}$ from $\mathcal{M}$ to $\mathcal{N}$ using \cref{eq:fmap_energy}, and then convert it to a p2p map $T_{fmap}: \mathcal{N} \rightarrow \mathcal{M}$ using (borrowing the same notation from \cref{sec:background}):
\begin{align}
&T_{fmap}(y) = \argmin_{x} \| (\Phi_{\mathcal{N}})_y - (\Phi_{\mathcal{M}} C_{\mathcal{M}\mathcal{N}}^{\top})_x\|
\label{eq:p2p_fmap}
\end{align}
For the test time optimization experiment in Sec. 5.1.2 of the main text, after extracting the maps predicted by the first stage, we construct pairs between all the shapes of the test set, and for each pair, we train a randomly-initialized DiffusionNet network using the NCE loss and Adam optimizer, to produce feature embedding that induces the map from stage 1. We stop the training when the cyclic loss (see Sec. 4.2 of the main text) stops improving giving a patience period of 100 optimization iterations.
In all the experiments of Sec. 5 of the main text, except for test time optimization, data augmentation was used. In particular, we augment the training data on the fly by randomly rotating the input shapes, applying random scaling in the range [0.9, 1.1], and jittering the position of each point by Gaussian noise with zero mean and 0.01 standard deviation.
\paragraph{Computational specifications} All our experiments are executed using Pytorch \cite{pytorch_NEURIPS2019}, on a 64-bit machine, equipped with an Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz and a RTX 2080 Ti Graphics Card. For all competing methods, we use the original code released by the authors and apply the best parameters reported in the respective papers. As mentioned in the main paper, we will release our complete implementation to ensure the full reproducibility of all of our results.
\label{sec:computation}
\section{Implementation details on FSKD}
\label{sec:fskd_imp}
Below we provide the implementation details on our Few-Shot Keypoint Detection method (FSKD), described in Sec. 5.3
of the main manuscript. As mentioned in that section, \textbf{FSKD} is composed of three steps: \textbf{1.} Detection of potential keypoints by transferring keypoints from labeled shapes, \textbf{2.} Filtering to remove keypoints that are likely not to exist on the target shape, \textbf{3.} Combination: merge transferred keypoints if multiple points on the target shape are assigned to the same keypoint ID.
\textbf{Step 1.} is done using our established maps predicted by \textbf{NCP-UN}{}.
For \textbf{step 2.}, we compute the cycle consistency loss of transferred keypoints and only keep the ones that are below a predetermined threshold. I.e., given a pair of shapes $\mathcal{M}, \mathcal{N}$, and two maps between them $\Pi_{\mathcal{M}\mathcal{N}}$ and $\Pi_{\mathcal{N}\mathcal{M}}$ computed via nearest neighbor matching between their feature embeddings, and represented as binary matrices, the cycle consistency loss of keypoints $i$ is computed as $l_i = \| (X_{\mathcal{N}})_i - (\Pi_{\mathcal{N}\mathcal{M}} \Pi_{\mathcal{M}\mathcal{N}} X_{\mathcal{N}})_i \|_F^2,$
where $X_{\mathcal{N}}$ is the matrix of XYZ coordinates of the source shape $\mathcal{N}$. If $l_i$ is bigger than a predefined threshold $\nu$, the keypoint $i$ is considered not to exist on the target shape. In our experiments, we take $\nu=0.05$.
Concerning \textbf{step 3.}, we perform a spatially weighted average of the different filtered keypoints if there are many. We associate for each keypoint $i$ the weight $w_i = \frac{D_i}{\sum_j D_j}$, where $D_i = \exp(- \frac{l_i}{\sigma})$. We use $\sigma = 0.01$.
\section{Proof of smoothness of maps produced by smooth networks}
\label{sec:theorem}
In Sec. 3.2 of the main text, we briefly mentioned a theorem that states that maps based on the nearest neighbor between smooth features are smooth. In this section, we formally restate it and provide a proof.
\begin{customthm}{1}
\label{thm1_supp}
Let $\mathcal{M}$ and $\mathcal{N}$ be two compact smooth surfaces (smooth manifolds of dimension 2). Let $\mathbf{M}$ and $\mathbf{N}$ be their embeddings in $\mathbb{R}^d$, given by some functions: $\psi: \mathcal{M} \rightarrow \mathbb{R}^d$ and $\phi: \mathcal{N} \rightarrow \mathbb{R}^d$, so that $\mathbf{M} = \psi (\mathcal{M})$ and $\mathbf{N} = \phi (\mathcal{N})$. Suppose that $\psi$ and $\phi$ are both smooth and injective. Then up to arbitrarily small perturbations of $\phi, \psi$, the map $T_{nn}: \mathcal{M} \rightarrow \mathcal{N}$ given by $T_{nn}(x) = \argmin_{y \in \mathcal{N}} \| \psi(x) - \phi(y)\|$ must be smooth up to sets of measure 0 on $\mathcal{M}$.
\end{customthm}
\begin{proof}
First, note that the distance function to an embedded manifold is smooth almost everywhere. Indeed, it is well-known that if $\mathbf{N}$ is a $C^k$-continuous manifold embedded in $\mathbb{R}^d$, then
the distance function to $\mathbf{N}$ must be at least $C^k$ continuous on the complement of the medial axis of $\mathbf{N}$. I.e., let $d_{\mathbf{N}}: \mathbb{R}^d \rightarrow \mathbb{R}^d$ be given by $d_{\mathbf{N}}(x) = \argmin_{y \in \mathbf{N}} \| x - y \|$. Let $\rm{cut}(\mathbf{N})$ denote the medial axis of $\mathbf{N}$, which is defined as the set of points in $\mathbb{R}^d$ with more than one nearest neighbor to $\mathbf{N}$. I.e., $\rm{cut}(\mathbf{N}) = \{x \in \mathbb{R}^d ~|~ \exists~ y_1, y_2 \in \mathbf{N}, y_1 \ne y_2, \rm{ s.t. } \|x - y_1\| = \|x - y_2 \| = \min_{y \in \mathbf{N}} \|x - y\| \}$. Then, $d_{\mathbf{N}} $ is at least $C^k$ continuous on $\mathbb{R}^d \backslash \rm{cut}(\mathbf{N})$ (see Lemma 2.5 in \cite{ghomi2022total}, and \cite{foote1984regularity,mantegazza2022Hamilton} for this and related results).
It remains to prove that up to arbitrarily small perturbations of $\phi$ and $\psi$ the intersection between $\mathbf{M}$ and $\rm{cut}(\mathbf{N})$ has measure zero on $\mathbf{M}$. For this, we first use the fact that the medial axis of compact subanalytic manifolds is also subanalytic~\cite{chazal2004stability}. This means that the medial axis can be stratified (decomposed into a finite union of submanifolds of dimension $d-1$ in $\mathbb{R}^d$). Since (by Stone-Weierstrass's theorem) subanalytic manifolds are dense in the space of smooth manifolds, up to an arbitrarily small perturbation of $\phi$, $\mathbf{N}$ is subanalytic. Finally, the intersection between $\mathbf{M}$ (which is an embedded manifold of dimension 2) and any manifold of dimension $d-1$ by Thom's transversality theorem \cite{guillemin2010differential,greenblatt2015introduction}, must generically be of measure zero on $\mathbf{M}$ and thus on $\mathcal{M}$.
\end{proof}
\section{Verification of assumptions of \cref{thm1_supp}}
\label{sec:verification_thm}
The main motivation behind \cref{thm1_supp} is to highlight the fact that, given smooth feature embeddings, if we add the injectivity condition, the maps extracted using the nearest neighbors in the feature space tend to be smooth. As smoothness is a generally desirable property, and the frequency bias shown in prior works, such as \cite{rahaman19a} suggests that neural networks are biased towards low frequency (and thus smooth) functions, we provide this result as a partial explanation for the Neural Correspondence Prior, which we have observed, for the first time, in our work.
Regarding injectivity, although such a property is not trivial, and might need to be specifically enforced (using, for example, invertible networks \cite{invertiblenet}). However, we observe that when training a network using contrastive learning, such as NCE loss, the latter forces the networks to produce embeddings that are unique for each point, in order to minimize the NCE loss, which can be considered a form of infectivity.
We include in this section the results of an experiment we performed in order to examine the smoothness and injectivity of a randomly initialized network. We start by computing the feature embeddings produced by a randomly initialized DiffusionNet network for the 20 test shapes in the \textsc{Faust} \cite{ren2018continuous} dataset.
To evaluate the smoothness of the embedding, we compute the standard Dirichlet energy of both the embedding produced by the network and the original embedding of the shape in $\mathbb{R}^3$ (the 3D coordinates of the shape’s vertices), using the following formula:
\begin{equation}
E_{Dirichlet}(G) = \frac{1}{n} \sum_{i=1}^n \frac{G_i^{\top} W G_i}{G_i^{\top} A G_i}
\end{equation}
where $G$ is the considered embedding of dimension $n$, while $W$ and $A$ are, respectively, the standard stiffness and mass matrices, computed using the classical cotangent discretization scheme of the Laplace-Beltrami operator \cite{Pinkall1993}. We compute the Dirichlet energy over the whole test set and find it equal to \textbf{47.2} (the average) for coordinate functions (original shape embeddings in $\mathbb{R}^3$), while it is equal to \textbf{13.1} for embeddings produced by the network. This shows that the latter are smoother than embeddings in the original space.
For injectivity, we compute for each point of the embedding, the distance to its nearest neighbor, and took the minimum across all points of a shape. To make the comparison between the original embedding and the embedding produced by the network fair, we normalize both embeddings to the unit sphere. We find that on average, the minimum distance between points in the original 3D space is \textbf{0.0004}, while for the feature embeddings given by the network, it is equal to \textbf{0.0015}. This shows that the network is injective and that distances between points are larger than in the original domain.
\section{Neural Correspondence Prior}
\label{sec:ncp_depth}
In Sec. 3 of the main text, we demonstrated the effect of Neural Correspondence Prior (NCP), on a pair of shapes, using the DiffusionNet network. The objective of this section is to show that the NCP is independent of the choice of the shape pair, the choice of the p2p loss, the choice of the architecture, and finally the choice of the first stage.
\subsection{Independence from shape pairs}
In order to show that the NCP effect demonstrated in Figure 1 of the main text does not depend on the chosen shape, we redid the same experiment using new random pairs from \textsc{Faust} and \textsc{Scape} datasets \cite{ren2018continuous}. Following the same setup of Sec.~3.2 of the main text, given a pair of shapes, we corrupt the ground truth map between them with 50\% noise, and then train a randomly initialized DiffusionNet to produce feature embeddings that overfit the noisy map. Examples of four different pairs are shown in \cref{fig:ncp-multiple_pairs}. We see that the same effect is always present, i.e., the network resists overfitting the noisy maps (high noise impedance), and the intermediate maps during optimization are of high quality, compared to the noisy maps used as supervision.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{figures/noise_impedance_supp_5003.pdf}
\caption{Learning curves showing the NCE loss and the geodesic error for different shape pairs from \textsc{Faust} and \textsc{Scape} datasets. Numbers in parentheses represent the geodesic error of the input maps. Observe how the effect of NCP doesn't depend on the shape pair.}
\label{fig:ncp-multiple_pairs}
\end{figure}
\subsection{Independence from the loss function and network architecture}
In order to examine the independence of NCP on both the network architecture and on the p2p loss used in stage 2 of \textbf{NCP-UN}{}, the following experiment was performed on the Chair subset of the \textsc{KeyPointNet} \cite{keypointnet2020} dataset. Since the latter doesn't provide dense ground truth p2p maps between the shapes, we took the p2p maps produced by the first stage of our algorithm (see Sec. 5.1.1 of the main text), and perturbed them with noise. These noisy maps are used as inputs to the second stage of the \textbf{NCP-UN}{} algorithm. Differently, from the experiment performed in Sec. 5.1.1 of the main text, here, we choose different network architectures and losses for training the second stage. In particular, we consider the NCE loss, the LIE loss, and the supervised FMAP loss from \cite{donati2020deep}. Concerning the network architectures, we considered the ResidualMLP network introduced in \cite{litany2017deep}, PointNet++ \cite{qi2017pointnet++}, and DGCNN \cite{wang2019dgcnn}. For the network architectures, we used the official implementation provided by the authors.
In addition to measuring the geodesic error produced by the maps predicted by the second stage, we also measure their smoothness. Given two shapes $\mathcal{M}, \mathcal{N}$, and a map between them $T_{\mathcal{M}\mathcal{N}}$ represented as a binary matrix $\Pi_{\mathcal{M}\mathcal{N}}$, we compute the smoothness of the latter based on the Dirichlet energy using the following (see \cite{Pinkall1993} for more details):
\begin{align}
&E_{smoothness}(T_{\mathcal{M}\mathcal{N}}) = \sum_{(u, v) \in \mathcal{E}_{\mathcal{M}}} w_{uv}\|\psi_{\mathcal{M}\mathcal{N}}(u) - \psi_{\mathcal{M}\mathcal{N}}(v) \|_2^2 \\
& \psi_{\mathcal{M}\mathcal{N}} = \Pi_{\mathcal{M}\mathcal{N}} X_{\mathcal{N}}
\label{eq:smoothness}
\end{align}
where $X_{\mathcal{N}}$ is the matrix of XYZ coordinates of shape $\mathcal{N}$, $\mathcal{E}_{\mathcal{M}}$ is the set of all the edges of the triangular mesh, and $w_{uv}$ are the stiffness weights of the cotangent Laplacian \cite{Pinkall1993} for shape $\mathcal{M}$.
Results of this experiment are summarised in \cref{tab:indep_arch_loss}. It can be seen that all the losses and network architectures perform well and manage to improve the noisy input maps. One can also notice the smoothing effect of NCP. In fact, the second stage of the network produces feature embeddings that result in p2p maps that are not only geometrically correct but also remove noise and outliers from the input maps. It is worth noting that although the ResidualMLP architecture is weak, since it is applied on each vertex individually, and has no global shape awareness, we note that even if the geodesic error is bad, the network tends to produce smooth embeddings.
\input{tables/indep_arch_loss}
\subsection{Independence from the first stage}
As we stated in the main paper, our \textbf{NCP-UN}{} algorithm is independent of the first stage, i.e the method used to extract the artifact-laden p2p is not crucial, and the NCP effect still applies. To demonstrate this, we followed the same setup of Sec.~5.1.2 of the main text, and we perform the same experiment on the \textsc{Smal} dataset but using different methods for stage 1. In particular, we consider as stage 1 the Unsupervised geometric functional map (Unsup GeomFMaps) \cite{sharp2021diffusion}, Smooth Shells \cite{Eisenberger2020SmoothSM}, Deep Shells \cite{eisenberger2020deep} and NeuroMorph \cite{Eisenberger2021NeuroMorphUS}. Results are summarized in \cref{tab:indep_stage1}. As can be seen, despite the method used in the first step, the NCP still applies, and a significant improvement in results is observed, up to 68\% improvement.
\input{tables/indep_stage1}
\section{Shape matching on man-made data}
\label{sec:matching_kpt}
In Sec. 5.1.1 of the main text, we applied our \textbf{NCP-UN}{} algorithm on the \textsc{KeyPointNet} dataset, and only provided quantitative results for 4 classes due to the page limit. We provide in \cref{fig:princeton_kpt_full} quantitative results for all the 16 classes, in addition to the results of the first stage of our algorithm, i.e the unsupervised method described above in \cref{sec:imp_details}. For the bottle, cap, knife, and skate categories, we only compared them to the best-performing baseline. It can be seen that our method achieves state-of-the-art results in 13 out of 16 classes on this benchmark, with an impressive improvement in some classes such as cap and table. We also see that the second stage of \textbf{NCP-UN}{} always improves the result provided by the first stage, which again demonstrates the role of the NCP effect, without which the SOTA result would not be achieved, see for example the laptop category.
Additionally, we provide in \cref{fig:viz_kpt} some qualitative results, showing the p2p maps produced by both stages of \textbf{NCP-UN}{}, visualized using texture transfer, as well as the usage of these maps to transfer keypoints between shapes. It can be seen that the keypoint transferred by the maps of stage 2 are more accurate.
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth]{figures/keypointnet_4x4.pdf}
\caption{\textbf{Correspondence accuracy on the \textsc{KeyPointNet} dataset}. It can be seen that the second stage of our algorithm always improves upon the first stage due to the NCP effect.}
\label{fig:princeton_kpt_full}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth]{figures/viz_keypointnet.pdf}
\caption{ \textbf{Qualitative results on the \textsc{KeyPointNet} dataset} using four categories: airplane, cap, chair, and table. Both the point-to-point map as well as the keypoint transfer are presented. Each row contains the p2p maps, ground truth keypoint annotations, and predictions made by the first and the second stage of our algorithm.}
\label{fig:viz_kpt}
\end{figure}
\section{Shape matching on non-rigid near-isometric data}
\label{sec:matching_iso}
In our work, we focus on difficult non-rigid non-isometric datasets, where existing methods tend to fail. This is because on near-isometric datasets such as \textsc{Faust} or \textsc{Scape} \cite{ren2018continuous}, current unsupervised methods can exploit the assumption of near-isometry and achieve good results.
Nevertheless, we include a comparison of our method on the \textsc{Faust} and \textsc{Scape} datasets, using the same train/test split used in all previous works (e.g. \cite{donati2020deep}). The notation X on Y means the method is trained on X and tested on Y.
As input for our Stage 2, we use WSupFMNet + DiffusionNet \cite{sharp2021diffusion} (referred to as Stage 1 in the table), which is the same method used in the main manuscript. For our method (Stage 2), we use the same implementation as the one used for the \textsc{Smal} dataset, described in Sec.~5.1.2 of the main manuscript.
\input{tables/ncp_faust_scape}
As shown in \cref{tab:ncp_faust}, even on a near-isometric dataset, our Stage 2 improves upon the initial maps in all categories, by \textbf{18.4\%} on average. Nevertheless, we remark that in such settings it is more advantageous to use specialized methods, such as ZoomOut \cite{Melzi_2019}, that directly exploit the near-isometry assumption.
\section{Societal impact}
\label{sec:societal}
Efficient methods for shape matching and analysis have an immediate impact in many areas of science and engineering from medical imaging (for instance detecting anomalies, and performing follow-up analysis) to shape recognition and classification in areas such as computational biology, archaeology, and paleontology to name a few. Our approach can immediately be adapted to such diverse scenarios, due to its strong generalization power, and its generic unsupervised nature, especially in domains where acquiring data is easy, but labeling it is very expensive, e.g in structural or molecular biology. Our work also opens up important avenues for future research, as it can enable geometric deep learning methods without the need to label large-scale datasets, thus potentially allowing small labs to conduct research in this area without the need to hire many annotators, which is an expensive task. Finally, our method paves the way for more accurate results, helping to improve our understanding in many fields, such as biology where shape-matching techniques are used to analyze gene expression patterns to understand the cause of many human syndromes \cite{Klatzow2022}. Since our method attempts to solve a fundamental problem in computer graphics and computer vision, we do not expect negative results. However, one should note that highly accurate shape correspondence methods might have possibly problematic uses, e.g., in surveillance applications, although we advocate against such uses of our technique.
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{}
\item Did you describe the limitations of your work?
\answerYes{See \cref{sec:conclusion}}
\item Did you discuss any potential negative societal impacts of your work?
\answerYes{See supplementary}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerYes{See supplementary}
\item Did you include complete proofs of all theoretical results?
\answerYes{See supplementary}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{Our code can be found online: \url{https://github.com/pvnieo/NCP}}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{See \cref{sec:application} and Supp Mat}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{In some experiments when it's not prohibitively expensive}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{See Supplementary}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{}
\item Did you mention the license of the assets?
\answerNA{}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerNo{}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerNo{The used datasets are publicly available.}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerNA{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
|
\section{Further details on the belief propagation algorithm}
In this section we give a more detailed background on the BP
algorithm of \cRef{AlkabetzArad2021}. The main ideas in the
algorithm are presented in \Fig{fig:Tree-TN-messages}. Given a PEPS
(\Fig{fig:Tree-TN-messages}a), we first compute the double-layer TN
representing $\langle \psi | \psi \rangle$, in which the
corresponding ket/bra tensors are contracted along the physical
legs, \Fig{fig:Tree-TN-messages}b. Then between every neighboring
spins $a,b$, we define a message $m_{a\to b}$ from spin $a$ to spin
$b$ and a message $m_{b\to a}$ in the opposite direction. Every
message is a positive semi-definite matrix, which is graphically
represented by a 2-legs tensor (see \Fig{fig:Tree-TN-messages}c).
Starting from a set of random messages, the messages at round
$\ell+1$ are calculated from the messages at round $\ell$ by
\begin{align}
\label{eq:bp-messages}
m_{a\rightarrow b}^{(\ell+1)} =
\Tr\left(T_aT_a^* \prod_{a'\in N_a \setminus \{b\}}
m^{(\ell)}_{a'\rightarrow a} \right),
\end{align}
which is illustrated in \Fig{fig:Tree-TN-messages}c (Equation~(1) in
the main text). After few iterations the messages typically
converge to a fixed point, from which a separable local environment
can be obtained (\Fig{fig:Tree-TN-messages}d). With the local
environments, one can easily compute local expectation values or
update the tensors when optimizing the PEPS.
When the underlying geometry of the PEPS is a tree, the converged
messages have a simple interpretation. Cutting the double edge that
connects $a$ and $b$, breaks the TN into two branches
(\Fig{fig:Tree-TN-messages}b). The converged $m_{a\to b}$ message is
simply the contraction of the double-layer branch that contains $a$
and $m_{b\to a}$ is the contraction of the other branch. In such
case, the local environment calculation of
\Fig{fig:Tree-TN-messages} becomes exact, since the contraction of
the branches is the only possible fixed point of the BP equations.
\begin{figure}[h!]
\includegraphics[width=0.4\columnwidth]{figures/Tree-TN-messages.pdf}
\caption{(a) PEPS representation of a quantum state, with
emphasis on the bond between spins $a$ and $b$. (b) Producing a
double layer TN by contracting $|\psi\rangle$ with $\langle
\psi\vert$, with emphasis on the virtual legs $x$ and $x'$. (c)
Graphical representation of Eq.(1) of the main text to compute
the message $m_{a\rightarrow b}$ from spin $a$ to spin $b$. (d)
The converged messages are used as environments to approximately
compute the reduce density matrix (RDM) for the
nearest-neighbor spins $a$ and $b$.}
\label{fig:Tree-TN-messages}
\end{figure}
\section{Detailed implementation of blockBP}
\begin{figure}
\includegraphics[width=0.8\columnwidth]{figures/MPOMPS}
\caption{Updating a local site tensor in the zip-up algorithm to
multiply an MPS by an MPO, which corresponds to
Eq.(\ref{eq:update}). }
\label{fig:figS1}
\end{figure}
The two building blocks of our blockBP algorithm are 1) computing
the fixed points of the MPS messages and 2) performing computations
(either computing local observables or updating the site tensors)
inside the center of each block. We use the bMPS algorithm for both
tasks. For example, for computing the left output MPS message of a
block in the case of a square lattice, one first forms a TN made of
the site tensors inside the block together with the input MPS
messages from the top, right and bottom of the block (one also fills
trivial site tensors at the corners so as to form a square TN), and
then one contracts the resulting square TN from right to left using
the bMPS algorithm to obtain the left output MPS message, see
Fig.1(d) in the main text. For computing local observables or
updating local site tensors inside a block during the imaginary time
evolution, we use exactly the same bMPS algorithm as in
\cRef{LubaschBanuls2014b}.
We choose different bond dimensions for the MPSs involved in these
two tasks, namely $\chi_m=D^2$ to evaluate the messages and
$\chi=2D^2+10$ in the second case, so as to reduce the computational
cost of calculating the messages (the messages do not need to be
extremely precise). The most computationally intensive subroutine
involved in both tasks is to multiply an MPS by a matrix product
operator (MPO). While this can be done by using the standard MPO-MPS
arithmetic~\cite{Schollwock2011}, and then compressing the resulting
MPS, we use the following zip-up algorithm that fuses these two
operations into one to reduce both the memory and computational
cost.
We denote the input MPS $\ket{X}$ as
\begin{align}
\ket{X} = \sum_{j=1}^L X^{a_1}_{b_1, b_2}
X^{a_2}_{b_2, b_3} \cdots X^{a_L}_{b_L, b_{L+1}},
\end{align}
and the input MPO $\hat{O}$ as
\begin{align}
\hat{O} = \sum_{j=1}^L O^{a_1', a_1}_{c_1, c_2}
O^{a_2', a_2}_{c_2, c_3} \cdots O^{a_L', a_L}_{c_L, c_{L+1}}.
\end{align}
Our task is to calculate a fixed bond dimension MPS $\ket{Y}$, which
approximates $\hat{O}\ket{X}$. We begin by first initializing
$\ket{Y}$ randomly:
\begin{align}
\ket{Y} = \sum_{j=1}^L Y^{a_1'}_{d_1, d_2} Y^{a_2'}_{d_2, d_3}
\cdots Y^{a_L'}_{d_L, d_{L+1}}.
\end{align}
Then we use an iterative algorithm to minimize the distance
\begin{align}\label{eq:distance}
\norm{\ket{Y} - \hat{O}\ket{X}}^2 = \braket{Y}{Y}
- \bra{Y}\hat{O}\ket{X} - \bra{X}\hat{O}^\dagger\ket{Y}
+ \bra{X}\hat{O}^{\dagger}\hat{O}\ket{X}.
\end{align}
Note that the last term on the right hand side of \Eq{eq:distance}
is a constant and can be neglected during the optimization. To
optimize $\ket{Y}$, we iteratively update each site tensor of it
using DMRG-like sweeps. For a specific site $l$, we first compute
\begin{align}\label{eq:update}
\tilde{Y}^{a_l'}_{d_l, d_{l+1}}
= \sum_{b_l, c_l, b_{l+1}, c_{l+1}, a_l}
A^{d_l}_{c_l, b_l} O^{a_l', a_l}_{c_l, c_{l+1}}
B^{d_{l+1}}_{c_{l+1}, b_{l+1}} X^{a_l}_{b_l, b_{l+1}},
\end{align}
where the rank-$3$ tensors $A$ and $B$ can be computed iteratively
using
\begin{align}
A^{d_l}_{c_l, b_l} = \sum_{b_{l-1}, c_{l-1}, d_{l-1},
a_{l-1}, a_{l-1}'} A^{d_{l-1}}_{c_{l-1}, b_{l-1}}
Y^{a_{l-1}'}_{d_{l-1},d_{l}} O^{a_{l-1}',a_{l-1}}_{c_{l-1},c_l}
X^{a_{l-1}}_{b_{l-1}, b_l}
\end{align}
and
\begin{align}
B^{d_{l+1}}_{c_{l+1}, b_{l+1}}
= \sum_{b_{l+2}, c_{l+2}, d_{l+2}, a_{l+1}, a_{l+1}'}
B^{d_{l+2}}_{c_{l+2}, b_{l+2}}Y^{a_{l+1}'}_{d_{l+1},d_{l+2}}
O^{a_{l+1}', a_{l+1}}_{c_{l+1}, c_{l+2}}
X^{a_{l+1}}_{b_{l+1}, b_{l+2}}
\end{align}
with $A^{d_1}_{c_1, b_1} = B^{d_{L+1}}_{c_{L+1}, b_{L+1}} = 1$.
During the left to right sweep, we perform a QR decomposition of
$\tilde{Y}^{a_l'}_{d_l, d_{l+1}}$ and get
\begin{align}
\mathrm{QR}(\tilde{Y}^{a_l'}_{d_l, d_{l+1}})
= \sum_s Q^{a_l'}_{d_l, s} R_{s, d_{l+1}},
\end{align}
and then we take $Q^{a_l'}_{d_l, s}$ as the new site tensor. During
the right to left sweep, we perform an LQ decomposition of
$\tilde{Y}^{a_l'}_{d_l, d_{l+1}}$ and get
\begin{align}
\mathrm{LQ}(\tilde{Y}^{a_l'}_{d_l, d_{l+1}})
= \sum_s L_{d_l, s} Q^{a_l'}_{s, d_{l+1} },
\end{align}
and then we take $Q^{a_l'}_{s, d_{l+1} }$ as the new site tensor.
\Eq{eq:update} is also demonstrated in \Fig{fig:figS1}. In addition,
from \Eq{eq:update} we have $-|\tilde{Y}^{a_l'}_{d_l, d_{l+1}}|^2 =
|\vert Y\rangle - \hat{O} \vert X\rangle |^2 - \langle X\vert
\hat{O}^{\dagger}\hat{O}\vert X\rangle $, which is exactly the loss
function subtracted by a constant term, therefore it can be used to
monitor the convergence, similar to the ground state energy in DMRG.
In our simulations we have set the convergence criterion to be the
standard deviation of $-|\tilde{Y}^{a_l'}_{d_l, d_{l+1}}|^2$ for a
full sweep (left to right and then right to left) divided by the
mean value, with a tolerance of $10^{-6}$ and the maximum number of
sweeps to be $10$.
For computing the MPS messages of a double layer TN, the MPS
messages are randomly initialized as a matrix product density
operator\cc{VerstraeteCirac2004} using normal distribution, while
for single layer TN which is used for classical models (see the
simulation of the classical Ising model below), the MPS messages are
simply initialized as a random MPS using uniform distribution.
Assuming that the mean square error between the MPS messages in
$l$-th step and those in the $l-1$-step is $\epsilon_l$, the
convergence criterion in this case is chosen to be $\epsilon_l /
\epsilon_1$, with a tolerance of $10^{-5}$ and the maximum number of
iterations to be $10$.
\section{Evaluation of reduced density matrix for finite systems}
\begin{figure}
\includegraphics[width=0.5\columnwidth]{figures/compare_rdm2_obc}
\caption{Trace distances between the RDMs of the horizontal
bonds computed using bMPS and blockBP with block size $7\times
7$. The ground states of the $21\times 21$ transverse Ising
model for different $B$s are used to compute the RDMs, which are
obtained by bMPS update with different bond dimensions $D$s. The
columns from left to right are results for $D=2,3,4$
respectively, while the rows from up down are results for
$B=2.5,3,3.5$ respectively. }
\label{fig:figS2}
\end{figure}
Here we study the quality of the local environments in our blockBP
by computing the local reduced density matrices (RDMs) on each bonds
using blockBP and compare them to the RDMs computed using bMPS. The
comparison is shown in the form of a heat map in \Fig{fig:figS2},
where we computed the trace distance between the blockBP and bMPS
RDMs on the ground state of a $21\times 21$ transverse Ising model
with OBC. The ground state was computed by bMPS full-update. The
blockBP RDMs were computed using $7\times 7$ blocks. We can see that
for $B=2.5,3.5$ which are away from the critical value ($B_c\approx
3.044$~\cite{BloteDeng2002}), the RDMs computed by these two methods
are very close to each other with average distance lower than
$10^{-5}$ for all the $D$s considered, while for $B=3$ we get
$d \approx 10^{-3}$ in average. We also note that the trace
distance between the RDMs computed using simple-update (or,
equivalently, the plain BP algorithm) and the RDMs computed using
bMPS will be of the order $10^{-1}$ on average in these cases.
\section{Convergence of blockBP for finite systems}
In the main text we have benchmarked our blockBP with bMPS for the
anti-ferromagnetic Heisenberg (AFH) model. Here we further
demonstrate that the convergence of our blockBP update towards the
ground state has a very similar trend as bMPS, given the same
parameters and the same initial state. Concretely, we consider the
imaginary time evolution of the $10\times 10$ AFH model using both
bMPS and blockBP update (with block size $5\times 5$) with
$d\tau=0.01$ and $D=4$, starting from the ground state obtained
using bMPS with $D=3$. The results are shown in \Fig{fig:figS3}. We
can see from the inset that the difference between bMPS and blockBP
is around $10^{-5}$ during the imaginary time evolution.
\begin{figure}
\includegraphics[width=0.4\columnwidth]{figures/xxx_convergence}
\caption{The ground state energy $E$ for the PEPS obtained using
bMPS and blockBP as a function of the imaginary time evolution
steps. The inset shows the difference between the energies
computed for the two PEPSs, namely $\Delta E = E_{{\rm blockBP}}
- E_{{\rm bMPS}}$, the initial state for both cases is chosen as
the ground state of the $10\times 10$ AFH model with $D=3$ and
we have used $d\tau=0.01$ in both cases. }
\label{fig:figS3}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\columnwidth]{figures/infinite_ising}
\caption{(a) Local magnetization $m_z$ as a function of the
inverse temperature $\beta$ for the infinite classical Ising
model. The red solid line is Onsager's exact solution. (b) $m_z$
as a function of $B$ for the infinite quantum Ising model for
$D=2$ (orange dashed line with circle) and $D=3$ (cyan dashed
line with triangle), where we have used a $2\times 2$ unit cell
and a block size $4\times 4$. The corresponding solid lines are
results from \cRef{PhienOrus2015}. The inset shows $m_z$
computed using larger block sizes near the critical point for
$D=3$.} \label{fig:fig4}
\end{figure}
\section{More simulations for the infinite classical and quantum
Ising models}
In \Fig{fig:fig4}a we study the infinite classical Ising model and
the transverse Ising model to further demonstrate the versatility of
blockBP. In panel (a) we look at the local magnetization $m_z$ of
the Gibbs state of the classical Ising model (with Hamiltonian $H =
\sum_{k,l}\sigma^z_k\sigma^z_l$) as a function of the inverse temperature
$\beta$. This can be written as a TN with $\chi=2$, which is an
ideal test ground for computing local observables since it does not
require updating. In this case we simply calculate $m_z$ for the
spin in the center of the block, and consider the effect of
different block sizes compared to the Onsager's exact
solution\cc{Onsager1944,Yang1952}. We see that with block size
$5\times 5$ we already obtain very accurate results away from the
critical point (with $\beta_c\approx 0.44$), and increasing the
block size, we obtain more accurate results near $\beta_c$. Here we
note that our results close to $\beta_c$ are not as accurate as
those obtained in Refs.\cc{LevinNave2007, XieXiang2009,
ZhaoXiang2010, XieXiang2012, EvenblyVidal2015}, however, they are
more accurate than the results obtained using variants of the BP
algorithms such as Refs.\cc{ZhouWang2012, WangZhou2013,
ZhouZheng2015}.
In \Fig{fig:fig4}b, we compute the ground state of the infinite
transverse Ising model using blockBP update for different values of
the external field $B$, and compared our results to the corner
transfer matrix (CTM) from \cRef{PhienOrus2015}. We set the system
size (unit cell) and the center size to be $2\times 2$, and use a
block size $4\times 4$. To compute $m_z$ for the final converged
infinite PEPS, we have copied the unit cell into a $52\times 52$
finite PEPS and computed the average $m_z$ at the central $2\times
2$ cell using bMPS. We can see that away from criticality
($B_c\approx 3.044$\cc{BloteDeng2002}) our results agree well with
the CTM results, especially for $D=3$. For $D=2$, our results
approach better the results at $D=3$ for $B < 3.1$. We note that
with $4\times 4$ block size our simulation is extremely efficient
($\approx 0.4$s and $\approx 1.1$s per iteration for $D=2$ and $D=3$
respectively using a single core of $2.3$ GHz frequency). As in the
classical Ising model, with a small block size we are not able to
accurately reproduce $m_z$ near $B_c$. Nevertheless, the results
close to the phase transition can be systematically improved by
increasing the block size near $B_c$.
\section{Computing local observables of infinite PEPS using blockBP}
Until now we have mostly used blockBP in conjunction with imaginary
time evolution to obtain the PEPS ground state, namely blockBP
update. In addition to that, blockBP can be used as a method to
compute local observables since the local environments computed by
blockBP can be used for both purposes. For finite PEPS we have used
bMPS to compute local observables for the examples considered in
this work. In the following, we show the precision of our blockBP
when used for computing local observables of infinite PEPS.
Concretely, taking an iPEPS ground state computed in Fig.~4 using
blockBP for different $D$s, we compared two different ways of
computing its energy. In the first approach, we embedded the iPEPS
with unit cell $2\times 2$ into a finite PEPS of size $(2k+2)\times
(2k+2)$ where the boundaries are randomly initialized, and then we
compute the energy in the $2\times 2$ center using bMPS. In the
second approach, we embedded the iPEPS with unit cell $2\times 2$
into another iPEPS with block size $2k\times 2k$, and computed the
energy in the $2\times 2$ center using blockBP. The results are
shown in \Fig{fig:figS5}. We can see that both approaches converge
to the $5$-th digit for large enough $k$, and that the blockBP
results already converge with a small $k=3$. Therefore blockBP can
also be used as an efficient method to accurately compute local
observables. Figure~\ref{fig:figS5} also shows the ground state
energies computed using a larger $\chi=2D^2+20$, and as we can see,
they are very close to the energies computed using $\chi=2D^2+10$
(the latter is chosen as the default throughout this work).
\begin{figure}
\includegraphics[width=\columnwidth]{figures/xxx_expectation}
\caption{Ground state energies of the infinite AFH model as a
function of $k$ computed using bMPS and blockBP for (a) $D=2$,
(b) $D=3$ and (c) $D=4$. For all those panels the cyan dashed
lines with squares and x are energies computed using bMPS with
$\chi=2D^2+10$ and $\chi=2D^2+20$ respectively, while the green
dashed lines with circle and + are energies computed using
blockBP with $\chi=2D^2+10$ and $\chi=2D^2+20$ respectively. }
\label{fig:figS5}
\end{figure}
\end{document}
|
\section{Introduction}
\label{sec:intro}
{Abstractly,} this paper {constructs an efficient multilevel estimator} \cite{Giles2008MLMC} for {triple nested expectations}, {of the kind which we later show arise in risk estimation involving the Credit Valuation Adjustment (CVA)}. {We consider quantities of} the form
\begin{equation}
\label{eqn:var}
\eta = \E[\bigg]{\Hbig{\E[\bigg]{f\p*{\E[\big]{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1} \given X_{t_0}}}}.
\end{equation}
Here, \(X_t\) is a stochastic process driven by an underlying \(d\)-dimensional SDE which is evaluated at (possibly random) times \(t_0 < t_1 < {T} \), \(f\p{x, y}\) and \(g\p{x}\) are Lipschitz {or twice differentiable} functionals of \(x\) and \(\H{x}\) is the Heaviside function taking value 1 if \(x\ge 0\) and value 0 otherwise. In many applications, {including for CVA}, \(f\) takes the form
\begin{equation}
\label{eqn:varf}
f\p[\bigg]{\E*{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1} = \Lambda \p[\bigg]{\E*{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1} - L_\eta,
\end{equation}
where \(\Lambda\p{\E*{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1}\) represents random future losses conditional on a risk scenario \(X_{t_0}\) and \(L_\eta\) is a constant representing the value-at-risk (VaR). Often it is of interest to fix \(\eta\) and solve \eqref{eqn:var} for the corresponding VaR, \(L_\eta\). Such problems arise frequently in risk analysis involving x-Valuation Adjustments (xVA) \cite{Gregory2020}.
As motivation for this work, we consider an application computing the probability of large loss due to fluctuations in the Credit Valuation Adjustment (CVA) \cite{Gregory2020, BrigoDamiano2013Ccrc}. Within this context, \(X_t\) denotes relevant market and credit factors, whereas \({g\p*{X_{{T}}}}\) {represents future payoffs given the market state at time \({T}\)} and \(f\p{\E{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1}\) represents losses above a threshold \(L_\eta\) due to fluctuations in market and credit risk factors up to the short risk horizon \(t_0\).\\
In practical settings, all three expectations in \eqref{eqn:var} must be approximated by, for example, a Monte Carlo average. Moreover, the {underlying SDE of the} process \(X_t\) will not typically admit an analytical solution and we must resort to Euler-Maruyama or Milstein {approximations} instead. \\
{For traditional Monte Carlo approximation of \eqref{eqn:var}}, we use \({M^\textnormal{o}}\) Monte Carlo samples to compute the outermost expectation, \({M^\textnormal{m}}\) samples for the middle expectation and \({M^\textnormal{i}}\) samples for the innermost expectation. Additionally, \(\{X_t\}_{0\le t \le {T}}\) is approximated using Euler-Maruyama or Milstein discretisation with step-size \({h}\) given by
\[
{h} =
\begin{cases}
{h^\textnormal{o}} & t\in \p{0, t_0}\\
{h^\textnormal{m}} & t\in \p{t_0, t_1}\\
{h^\textnormal{i}} & t\in \p{t_1, {T}}
\end{cases}.
\]
{By taking \({h^\textnormal{o}} \propto \p*{{M^\textnormal{m}}}^{-1}\) and \({h^\textnormal{m}} \propto \p*{{M^\textnormal{i}}}^{-1}\), it follows from the weak convergence properties of the Euler-Maruyama and Milstein schemes \cite{Kloeden:1999} that the bias induced by approximating the underlying SDE is of the same order as that induced by using \({M^\textnormal{m}}\) and \({M^\textnormal{i}}\) samples to approximate the expectations determined by \(X_{t_0}\) and \(X_{t_1}\), respectively.}
Due to the nested structure of the problem, it follows that the total cost of the resulting estimator is {of order}
\(
{{M^\textnormal{o}}{M^\textnormal{m}}{M^\textnormal{i}}}\p*{{h^\textnormal{i}}}^{-1}.
\)
To achieve a root mean square error of \(\textnormal{TOL}\), it follows from \cite[Proposition 1]{Gordy:2010} that we must take \({M^\textnormal{o}} = \Order{\textnormal{TOL}^{-2}} \), \({M^\textnormal{m}}, {M^\textnormal{i}} = \Order{\textnormal{TOL}^{-1}}\) and \({h^\textnormal{i}} = \Order{\textnormal{TOL}}\). Hence, we obtain a total computational cost of order \(\textnormal{TOL}^{-5}\). The probability of large loss due to fluctuations in the CVA is typically small, \(\eta\ll 1\) in \eqref{eqn:var}. {Therefore}, a small value of \(\textnormal{TOL}\) is required to achieve reasonable results. The computational cost of order \(\textnormal{TOL}^{-5}\) is therefore {extremely prohibitive}.\\
To reduce this cost, in the current work we consider an alternative estimator utilising the multilevel Monte Carlo setup \cite{Giles2008MLMC}. A hierarchy of unbiased multilevel Monte Carlo {corrections} \cite{Rhee2015_Unbiased_estimation} are used to replace \(f\p*{\E*{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1}\) with an alternative random variable with mean \(\E{f\p*{\E*{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1}\given X_{t_0}}\) and an \(\Order{1}\) sampling cost and variance. {Multilevel methods are less effective at estimating discontinuous observables such as the Heaviside function in \eqref{eqn:var}. Adaptive refinement around the discontinuity \cite{hajiali2021adaptive, GilesHajiAli:2018} is used to construct an efficient multilevel estimator in this instance}.\\
The paper is structured as follows:
\begin{itemize}
\item {\Cref{sec:mlmc} describes the construction of a novel and efficient estimator for \eqref{eqn:var} using unbiased and adaptive multilevel Monte Carlo techniques.}
\item \Cref{sec:quantile} discusses how the methods used can be incorporated in value-at-risk computations using stochastic root finding algorithms.
\item The task of computing the probability of large loss due to fluctuations in CVA using the method of \Cref{sec:mlmc} is discussed in \Cref{sec:cva}. The theoretical properties of the estimator are supported by numerical studies in \Cref{sec:num}.
\item {\Cref{app:rmlmc} contains an analysis of the unbiased multilevel Monte Carlo estimator proposed in \Cref{sec:umlmcb}.}
\end{itemize}
\section{Hierarchical Multilevel Monte Carlo Estimation}
\label{sec:mlmc}
In this section, we use a hierarchy of multilevel Monte Carlo (MLMC) estimators to significantly improve upon the complexity of standard Monte Carlo approximation of \eqref{eqn:var}. The discussion focuses on {combining} two methods:
\begin{enumerate}
\item Unbiased MLMC which allows us to express
\begin{equation}
\label{eqn:umlmc-prob}
\E{f\p{\E{{g\p*{X_{{T}}}}\given X_{t_1}},\ t_1}\given X_{t_0}} = \E{Y\given X_{t_0}},
\end{equation}
where \(Y\) can be sampled with \(\Order{1}\) cost and variance.
\item Adaptive MLMC to approximate
\[
\E*{\Hbig{\E{Y\given X_{t_0}}}},
\]
with Monte Carlo approximation of the inner expectation
\(
\E{Y\given X_{t_0}}
\)
using exact samples of \(Y\), and refinement around the discontinuity in \(\H{\cdot}\).
\end{enumerate}
\subsection{Unbiased MLMC}
\label{sec:umlmc}
{We first recall the basic idea of unbiased MLMC \cite{Rhee2015_Unbiased_estimation}}. Consider the general problem of approximating
\(
\E{Z},
\)
where \(Z\) is a random variable which cannot be sampled directly. Assume we can sample from a hierarchy of increasingly accurate approximations \(Z_\ell\approx Z\) for all integers \(\ell\ge0\), {with \(\lim_{\ell\to \infty}Z_\ell = Z\)}, admitting a sequence of correction terms \(\dl{Z}\) with the property
\[
\E{\dl{Z}} =
\begin{cases}
\E{Z_0} & \ell = 0\\
\E{Z_\ell - Z_{\ell-1}} & \ell > 0
\end{cases}.
\]
The following condition is required of the correction terms \(\dl{Z}\).
\begin{assumption}
\label{assumpt:umlmc}
There exists constants \(\beta > \gamma > 0\) {and \(a_0, a_1 > 0\),} independent of \(\ell\), such that
\begin{itemize}
\item \(\textnormal{cost}\p{\dl{Z}} {\le a_0}2^{\gamma\ell}\),
\item \(\E{\dl{Z}^2} {\le a_1}2^{-\beta \ell} \).
\end{itemize}
\end{assumption}
Let \({\tilde{\ell}}\) denote a random, non-negative integer with mass function
\[
p_{\ell} \coloneqq \prob*{{\tilde{\ell}} = \ell}\propto 2^{-\zeta\ell},
\]
for \(\gamma < \zeta<\beta\). Then, it follows from \Cref{assumpt:umlmc} that \cite{Rhee2015_Unbiased_estimation}
\begin{equation}
\label{eqn:telsum-umlmc}
\E*{Z}{= \sum_{\ell=0}^\infty\E{\dl{Z}}} = \E*{\dl[{\tilde{\ell}}]{Z}\: p_{\tilde{\ell}}^{-1}},
\end{equation}
and \(\dl[{\tilde{\ell}}]{Z}\: p_{\tilde{\ell}}^{-1}\) has \(\Order{1}\) sampling cost and variance. {It is shown in \cite{Rhee2015_Unbiased_estimation} that to optimally balance the cost and variance of the unbiased Monte Carlo, one should take \(\zeta = \p{\beta+\gamma} / 2\). Alternatively, it may be of interest to ensure \(\dl[{\tilde{\ell}}]{Z}\: p_{\tilde{\ell}}^{-1}\) has finite \(q\)'th moment for some \(q > 2\). Under general assumptions, this can be shown to hold if \(\zeta < q\beta / 2\p{q-1} \) \cite[Remark 3.1]{GilesHajiAli:2019sampling}.} \\
We now consider the application of this method to \eqref{eqn:umlmc-prob}.
\subsubsection{Unbiased Estimation of \(\E*{{g\p*{X_{{T}}}}\given X_{t_1}}\)}
\label{sec:umlmca}
Given (an approximation of) \(X_{t_1}\), we cannot exactly sample \(X_{{T}}\) and must instead discretize {the underlying} SDE on the interval \([t_1, {T}] \). In the context described above, let \(X_{{T}, \ell}\) denote the Milstein approximation to \(X_{{T}}\) given \(X_{t_1}\) and using time-step \(\dtiell \propto 2^{-\gamma\ell}\) on \([t_1, {T}]\), {for \(\gamma \) as in \Cref{assumpt:umlmc}}. Define the correction terms \(\dl{g} \coloneqq {g\p*{X_{{T}, \ell}}} - {g\p*{X_{{T}, \ell - 1}}} \) for \(\ell>0\) and \(\dl[0]{g} \coloneqq \p{X_{{T}, 0}} \). It follows under regularity conditions on {the coefficients of} the underlying SDE \cite[Chapter 10]{Kloeden:1999} {and for Lipschitz \(g\)} that \Cref{assumpt:umlmc} holds for \(\dl{g}\) with \(\beta = 2\gamma\). Thus,
\[
\E*{{g\p*{X_{{T}}}}\given X_{t_1}} = \E*{\dl[{\tilde{\ell}}]{g}\: p_{\tilde{\ell}}^{-1}\given X_{t_1}},
\]
where \({\tilde{\ell}}\) has probability mass function \(p_\ell \propto 2^{-\zeta\ell}\) for \(\gamma<\zeta<2\gamma\) and \(\dl[{\tilde{\ell}}]{g}\: p_{\tilde{\ell}}^{-1}\) has finite sampling cost and variance. {For later analysis (see \Cref{app:rmlmc}), we require the unbiased Monte Carlo term to have finite 3'rd moment, which instead requires \(\zeta < 3\gamma / 2\).} \\
For high dimensional SDEs the Milstein scheme often requires the numerical simulation of L\'evy areas, which can be computationally prohibitive. In this case, the antithetic Milstein scheme proposed in \cite{giles14antmilstein} may be useful to retain the required property on \(\beta\) while avoiding the simulation of L\'evy areas.
\subsubsection{Unbiased Estimation of \eqref{eqn:umlmc-prob}}
\label{sec:umlmcb}
Using the discussion above, we can express the left hand side of \eqref{eqn:umlmc-prob} as
\[
\E*{f\p*{\E*{\dl[{\tilde{\ell}}]{g}\: p_{\tilde{\ell}}^{-1}\given X_{t_1}}, t_1}\given X_{t_0}}.
\]
There are two remaining aspects of approximation which add bias to an estimator of the above expectation:
\begin{itemize}
\item Monte Carlo approximation of \(\E{\dl[{\tilde{\ell}}]{g}\: p_{\tilde{\ell}}^{-1}\given X_{t_1}} \).
\item Approximation of \(X_{t_1}\) given \(X_{t_0}\).
\end{itemize}
For the following discussion, let \(\dl{g} = \dl{g}\p{X_{t_1}, W_{t_1, {T}}}\) be determined by \(X_{t_1}\) along with the Brownian path \({W_{t_1, T}} = \{W_t\}_{t_1 \le t\le {T}}\) for an underlying Brownian Motion \(W_t\). We can approximate \(\E{\dl[{\tilde{\ell}}]{g}\: p_{\tilde{\ell}}^{-1}\given X_{t_1}}\) by
\[
\begin{aligned}
\Eiell\p*{X_{t_1}}\coloneqq \frac{1}{\Miell}\sum_{m=1}^{\Miell}\dl[{{\tilde{\ell}}^{(m)}}]{g}\p{X_{t_1},\ W_{t_1, {T}}^{{(m)}}}\:p_{{\tilde{\ell}}^{(m)}}^{-1},
\end{aligned}
\]
where {\({\tilde{\ell}}^{(m)}\) are i.i.d. realisations of \({\tilde{\ell}}\)}, \(W_{t_1, {T}}^{(m)}\) are independent Brownian paths and \(\Miell \propto 2^{\gamma\ell}\) {for \(\gamma\) as in \Cref{assumpt:umlmc}.} The path of \(X_{t_1}\), given \(X_{t_0}\) {and a discretisation level \(\ell\)}, can be computed using Milstein discretisation with time-step \(\dtmell \propto 2^{-\gamma\ell}\) on \([t_0, t_1]\). \\
A natural choice for the correction terms is then
\[
\begin{aligned}
\dl{f} &= f\p*{\Eiell\p*{X_{t_1, \ell}}, {t_1}} - f\p*{\Eiell[\ell-1]\p*{X_{t_1, \ell-1}},{t_1}},
\end{aligned}
\]
where \(X_{t_1}\) is approximated using Milstein approximation at level \(\ell\) for the fine term and a correlated approximation at level \(\ell-1\) for the coarse term. Independent samples of \({\tilde{\ell}}\) and the Brownian path on \(W_{t_1, {T}}\) are used for the Monte Carlo estimators at \(\Eiell\) and \(\Eiell[\ell-1]\). However, the resulting error between the fine and coarse Monte Carlo estimators in \(\dl{f}\) leads to \(\beta = \gamma\) in the context of \Cref{assumpt:umlmc}, thus failing the conditions for unbiased MLMC. To improve upon this result, let \(\gamma = 1\), giving \(\Miell \propto 2^{\ell}\). Then, for \(j=0, 1\) define
\[
\begin{aligned}
\Eiell[\ell-1, c, j]\p{X_{t_1}} &\coloneqq \frac{1}{\Miell[\ell-1]}\sum_{m=1}^{\Miell[\ell-1]}\dl[{{\tilde{\ell}}^{(m + j\Miell[\ell-1])}}]{g}\p*{X_{t_1},\ W_{t_1, {T}}^{{(m + j\Miell[\ell-1])}}}\:p_{{\tilde{\ell}}^{(m + j\Miell[\ell-1])}}^{-1}\\
\Eiell[\ell, f]\p{X_{t_1}} &\coloneqq \frac{1}{2}\p[\bigg]{\Eiell[\ell-1, c, 0]\p{X_{t_1}} + \Eiell[\ell-1, c, 1]\p{X_{t_1}} },
\end{aligned}
\]
so that the fine Monte Carlo average \(\Eiell[\ell, f]\p{X_{t_1}}\) is composed of the same Brownian paths \(W_{t_1, T}^{(m)}\) and randomised levels \({\tilde{\ell}}^{(m)}\) used for two coarse averages \(\Eiell[\ell-1, c, 0]\p{X_{t_1}}, \Eiell[\ell-1, c, 1]\p{X_{t_1}}\). Now, consider the antithetic estimator
\begin{equation}
\label{eqn:dlantf}
\begin{aligned}
\dl[\ell]{^\textnormal{ant}f} = &f\p[\big]{\Eiell[\ell, f]\p{X_{t_1, \ell}},{t_1}}\\
&\quad - \frac{1}{2}\p[\bigg]{f\p[\big]{\Eiell[\ell-1, c, 0]\p{X_{t_1, \ell-1}},{t_1}} + f\p[\big]{\Eiell[\ell-1, c, 1]\p{X_{t_1, \ell - 1}},{t_1}} }.
\end{aligned}
\end{equation}
For Lipschitz {and piecewise twice differentiable} functions \(f\), it can be shown using the strong rate of convergence of the Milstein scheme combined with \cite[Theorem 2.3]{BujokK2015} and \cite[Theorem 4.1]{GilesHajiAli:2018} that \(\dl[\ell]{^\textnormal{ant}f}\) satisfies \Cref{assumpt:umlmc} with \(\beta = 3\gamma/2 { = 3/2}\). {A more detailed discussion of this result is given in \Cref{app:rmlmc}.} Thus, we have
\[
\E*{f\p*{\E*{\dl[{\tilde{\ell}}]{g}\: p_{\tilde{\ell}}^{-1}\given X_{t_1}}, t_1}\given X_{t_0}} = \E*{\dl[{\hat{\ell}}]{^\textnormal{ant}f}\: q_{{\hat{\ell}}}^{-1} \given X_{t_0}},
\]
where \({\hat{\ell}}\) is a random, non-negative integer with probability mass function {(using \(\zeta = \p{\beta+\gamma} / 2\))} \(q_\ell \propto 2^{-5\ell / 4}\) and \(\dl[{\hat{\ell}}]{^\textnormal{ant}f}\: q_{{\hat{\ell}}}^{-1}\) has \(\Order{1}\) sampling cost and variance. {The process of generating a single sample of \(\dl[\ell]{^\textnormal{ant}f}\), given \(X_{t_0}\), is detailed in \Cref{alg:ant-rmlmc}.}
\begin{center}
\begin{algorithm}
\KwIn{$\ell, X_{t_0}, \gamma, \Miell, \dtmell, \{\dtiell\}_{\ell \ge 0}$}
\KwResult{Sample of $\dl[\ell]{^\textnormal{ant}f}$, given \(X_{t_0}\)}
\SetKwFunction{FMil}{Milstein}
\SetKwFunction{FMean}{Mean}
\SetKwProg{Fn}{Function}{:}{end}
\SetKwArray{Arr}{array}
\Fn{\FMil{\(T_0\), \(T_1\), \(h\), \(x_0\), \(\omega\)}}{\Return Milstein approximation of \(X_{T_1}\) with step-size \(h\) given Brownian path \(\omega\) on the interval \([T_0,T_1]\) with \(X_{T_0} = x_0\)}
Sample Brownian path \(\omega_{t_1} = \{W_{t_0 + nh_\ell}\ |\ n = 0,\dots, (t_1 - t_0) / \dtmell\}\)\;
Set \(X_{t_1, \ell} = \FMil{$t_0$, $t_1$, $\dtmell$, $X_{t_0}$, $\omega_{t_1}$}\)\;
\If{\(\ell > 0\)}{
Set \(X_{t_1, \ell - 1} = \FMil{$t_0$, $t_1$, $h_{\ell-1}^{(1)}$, $X_{t_0}$, $\omega_{t_1}$}\)\;
}
\tcc{Define arrays \(\Delta_j\) to store samples of \(\Delta_{\tilde{\ell}} g\: p_{\tilde{\ell}}^{-1}\) using \(X_{t_1} \approx X_{t_1, j}\) for \(j = \ell, \ell-1\).}
Set \(\Delta_\ell = \Arr{$\Miell$}\)\;
Set \(\Delta_{\ell-1} = \Arr{$\Miell$}\)\;
\For{\(m = 1,\dots, \Miell\)}{
Sample random integer \(k\) with probability mass function \(p_\ell\propto 2^{-3\ell/2}\)\;
Sample Brownian path \(\omega_{{T}} = \{W_{t_1 + n\dtiell[k]}\ |\ n = 0,\dots \p{{T} - t_1} / \dtiell[k]\}\)\;
Set \(X_{{T}, \ell, k} = \FMil{$t_1$, ${T}$, $\dtiell[k]$, $X_{t_1,\ell}$, $\omega_{{T}}$}\)\;
\If{\(k > 0\)}{
Set \(X_{{T}, \ell, k - 1} = \FMil{$t_1$, ${T}$, $\dtiell[k-1]$, $X_{t_1,\ell}$, $\omega_{{T}}$}\)\;
}
\tcc{Convention: When \(k = 0\) set \(g\p{X_{{T}, \cdot, -1}} \equiv 0\)}
Set \(\Delta_\ell^{(m)} = \p*{g\p{X_{{T}, \ell, k}} - g\p{X_{{T}, \ell, k - 1}}}\:p_k^{-1}\)\;
\If{\(\ell > 0\)}{
Set \(X_{{T}, \ell - 1, k} = \FMil{$t_1$, ${T}$, $\dtiell[k]$, $X_{t_1,\ell - 1}$, $\omega_{{T}}$}\)\;
\If{\(k > 0\)}{
Set \(X_{{T}, \ell - 1, k - 1} = \FMil{$t_1$, ${T}$, $\dtiell[k-1]$, $X_{t_1,\ell - 1}$, $\omega_{{T}}$}\)\;
}
Set \(\Delta_{\ell - 1}^{(m)} = \p*{g\p{X_{{T}, \ell - 1, k}} - g\p{X_{{T}, \ell -1, k - 1}}}\:p_k^{-1}\)
}
}
\Return \(f\p*{\FMean{$\Delta_\ell$}} - \frac{1}{2}\sum_{i=0}^{1}f\p*{\FMean{$\Delta_{\ell-1}[i\Miell[\ell-1]+1:\p{i+1}\Miell[\ell-1]]$}}\)
\caption{Sampling \(\dl{^\textnormal{ant} f}\)}
\label{alg:ant-rmlmc}
\end{algorithm}
\end{center}
\subsection{Adaptive MLMC}
\label{sec:discontmlmc}
The problem of computing \eqref{eqn:var} has now been reduced to estimating
\begin{equation}
\label{eqn:mlmc-prob}
\E{\H{\E{Y\given X_{t_0}}}},
\end{equation}
where independent samples of \(Y\) are readily available. Given \(X_{t_0}\), we thus approximate \(\E{Y\given X_{t_0}}\) by
\[
\Emell \p*{X_{t_0}} \coloneqq \frac{1}{\Mmell}\sum_{m=1}^{\Mmell} Y^{(m)}\p{X_{t_0}},
\]
{where \(Y^{(m)}\) are i.i.d. samples of \(Y\) given \(X_{t_0}\).}
In many {financial applications}, \(t_0 \ll t_1\) represents a short risk horizon. It is therefore possible to approximate \(X_{t_0}\) to a much higher degree of accuracy than \(X_{t_1}\) with negligible cost. Therefore, to simplify the following discussion, we assume \(X_{t_0}\) (or at least a suitably accurate approximation) is available with \(\Order{1}\) cost. As before, we define the multilevel correction terms
\[
\dl{\mathbb{H}} = \dl{\mathbb{H}}\p{X_{t_0}} \coloneqq \Hbig{\Emell\p*{X_{t_0}}} - \Hbig{\Emell[\ell-1]\p*{X_{t_0}}}.
\]
However, due to the possibility of making an \(\Order{1}\) error in the Heaviside function when \(\Emell\p{X_{t_0}}\) and \(\E{Y\given X_{t_0}} \) fall on opposite sides of zero, which occurs for \(\Order*{\p*{\Mmell}^{-1/2}}\) samples of \(\Emell\p{X_{t_0}}\), it follows that \Cref{assumpt:umlmc} is satisfied only allowing for \(\beta = \gamma/2\) \cite{GilesHigham:2009,giles:2015review}. Moreover, this property is not improved by taking an antithetic difference \(\Delta^\textnormal{ant}\) as above \cite{GilesHajiAli:2018}. Thus, we are unable to use unbiased MLMC for this case.\\
Instead, we truncate the telescoping sum {appearing in} \eqref{eqn:telsum-umlmc} at some maximul level \(L\) and compute \(\{\E{\dl{\mathbb{H}}\p{X_{t_0}}}_{\ell=0}^L \}\) using a Monte Carlo average, reverting to standard MLMC \cite{Giles2008MLMC}:
\begin{equation}
\label{eqn:mlmc-telsum}
\begin{aligned}
\E*{\Hbig{\E*{Y\given X_{t_0}}}} = \sum_{\ell=0}^{\infty} \E*{\dl{\mathbb{H}}\p*{X_{t_0}}} &\approx \sum_{\ell = 0}^L \E*{\dl{\mathbb{H}}\p*{X_{t_0}}}\\
&\approx \sum_{\ell=0}^L \frac{1}{\Moell}\sum_{m=1}^{\Moell}\dl{\mathbb{H}}^{(m)}\p*{X_{t_0}^{(m)}},
\end{aligned}
\end{equation}
with \(\dl{\mathbb{H}}^{(m)}\p{X_{t_0}^{(m)}} \overset{\textnormal{i.i.d.}}{\sim} \dl{\mathbb{H}}\p{X_{t_0}}\).
The key idea is that by using many samples at the cheaper levels of approximation and since the variance of \(\dl{\mathbb{H}}\) decreases {as \(\ell\) increases}, we require fewer samples at the expensive level of simulation \(L\). Defining \(V_\ell = \E{\dl{\mathbb{H}}^2} \propto 2^{-\beta\ell}, \ C_\ell = \text{cost}\p{\dl{\mathbb{H}}}\propto 2^{\gamma\ell}\) {by \Cref{assumpt:umlmc}}, it can be shown that by optimizing the number of samples \(M_\ell\) per level, the work required for the estimator \eqref{eqn:mlmc-telsum} to achieve root mean square error \(\textnormal{TOL}\) is proportional to \cite{giles:2015review, Cliffe:2011}
\[
\textnormal{TOL}^{-2}\p*{\sum_{\ell=0}^L\sqrt{V_\ell C_\ell}}^2,
\]
with \(L\propto \log\p{\textnormal{TOL}^{-1}}\). In the best case, \(\beta > \gamma\) and the above term is \(\Order{\textnormal{TOL}^{-2}}\), which is frequently referred to as the canonical complexity. Instead, for \(\beta = \gamma/2\), the estimator \eqref{eqn:mlmc-telsum} can be shown to have \(\Order{\textnormal{TOL}^{-5/2}}\) cost \cite{Cliffe:2011, GilesHajiAli:2018}. For comparison, a standard Monte Carlo average used to compute \eqref{eqn:mlmc-prob}, with \(\E{Y\given X_{t_0}} \approx \Emell[L]\p{X_{t_0}}\), has \(\Order{\textnormal{TOL}^{-3}}\) cost \cite{Gordy:2010}. Thus, MLMC is able to improve the cost by a factor of order \(\textnormal{TOL}^{-1/2}\) over standard Monte Carlo. However, the cost remains higher than the canonical \(\Order{\textnormal{TOL}^{-2}}\) complexity.\\
Several methods have been proposed which remedy the effect of the Heaviside function on the performance of MLMC for this problem. In \cite{Giles:2015Smoothing}, the authors approximate $\H{\cdot}$ by a Lipschitz function with better convergence properties as seen above for unbiased MLMC. However, this approximation adds an additional level of bias which is inversely proportional to the Lipschitz constant of the approximation to \(\H{\cdot}\), which in turn affects the cost of MLMC. In \cite{bayerchiheb:2020} an implicit smoothing technique is applied which relies on a change of variables and a pre-integration step which smooths over the discontinuity. An alternative approach, and the one adopted for the remainder of this paper, relies on adaptively refining the number of Monte Carlo samples \(Y\) required for a given \(X_{t_0}\) \cite{Brodie:2011,Dodwell:2021refinement, Elfverson:2016selectiverefinement, GilesHajiAli:2018, GilesHajiAli:2019sampling, hajiali2021adaptive}. In the adaptive framework, we define the correction term
\[
\dl{^\textnormal{ad}\mathbb{H}} \p{X_{t_0}} = \Hbig{\Emell[\ell+\eta_\ell]\p{X_{t_0}}} - \Hbig{\Emell[\ell - 1 + \eta_{\ell-1}]\p{X_{t_0}}},
\]
with adapted Monte Carlo averages \(\Emell[\ell + \eta_\ell]\p{X_{t_0}} \) obtained using \Cref{alg:ad-samp}. Here, \(0\le \eta_\ell\le \ell\) is a random integer, depending on \(X_{t_0}\), {which reflects} uncertainty in the sign of \(\Emell[\ell + \eta_\ell]\p{X_{t_0}} \). Since \(\H{\Emell[\ell + \eta_\ell]\p{X_{t_0}}}\) is incorrect only when \(\Emell[\ell + \eta_\ell]\p{X_{t_0}}\) is estimated to be on the {opposite side of 0 to \(\E{Y\given X_{t_0}}\)}, it is sensible to use a more accurate estimator, with larger \(\eta_\ell\), when \(\E{Y\given X_{t_0}}\) is thought to be close to the origin. The expected cost of sampling \(\dl{^\textnormal{ad}\mathbb{H}}\), \(\E{C_{\ell+\eta_\ell}}\), is shown to remain of order \(\Mmell\), whereas the variance reduces to \(V_\ell \propto \p*{\Mmell}^{-1}\) under general assumptions, giving \(\beta = \gamma\) and reducing the cost of MLMC to \(\Order{\textnormal{TOL}^{-2}\p{\log \textnormal{TOL}}^2}\). The lack of regularity in \(\H{\cdot}\) leads to a high kurtosis for the multilevel correction terms with and without adaptive sampling, reducing the reliability of error estimates of the MLMC estimator. One approach to this problem involves biased estimates of the first and second moments of \(\dl{\mathbb{H}} \) and is discussed in \Cref{sec:mlmcset}.
\begin{center}
\begin{algorithm}
\KwIn{$X_{t_0}, \ell, r>1, c>0$}
\KwResult{Adaptively refined sample $\tilde E_{\ell+\eta_\ell}$}
Set $\eta_\ell = 0$\;
Sample $\Mmell$ independent samples of $Y$ given $X_{t_0}$\;
\tcc{\(\sigma_\ell\p{X_{t_0}}\) is the sample variance given the \(\Mmell\) samples}
Compute $\p{\Emell\p{X_{t_0}}, \sigma_\ell\p{X_{t_0}}}$ from the existing samples\;
\While{$\sigma_{\ell + \eta_\ell}\p{X_{t_0}}^{-1}\abs{\Emell[\ell+\eta_\ell]\p{X_{t_0}}}< c2^{\p{\ell\p{1-r}-\eta_\ell}/r}$ and $\eta_\ell < \ell$}{
Sample $\Mmell[\ell+\eta_\ell+1] - \Mmell[\ell + \eta_\ell]$ new, independent samples from $Y$ given $X_{t_0}$\;
Refine $\p{\Emell[\ell+\eta_\ell]\p{X_{t_0}},\sigma_{\ell+\eta_\ell}\p{X_{t_0}}}$ to $\p{\Emell[\ell+\eta_\ell+1]\p{X_{t_0}},\sigma_{\ell+\eta_\ell+1}\p{X_{t_0}}}$ using all existing samples\;
Set $\eta_\ell = \eta_\ell+1$\;
}
\KwOut{$\Emell[\ell+\eta_\ell]\p{X_{t_0}}$}
\caption{Adaptive sampling at level $\ell$ \cite[Algorithm 1]{hajiali2021adaptive}}
\label{alg:ad-samp}
\end{algorithm}
\end{center}
\subsection{Quantile Estimation and Value-at-Risk}
\label{sec:quantile}
As mentioned in the introduction, it is often of interest to solve \eqref{eqn:var}, \eqref{eqn:varf} for \(L_\eta\) with a pre-defined value of \(\eta\). This equates to finding the unique root of
\[
F(\lambda) = \E[\bigg]{\Hbig{\E[\bigg]{\Lambda\p[\big]{\E*{g\p{X_{{T}}}\given X_{t_1}},\ t_1} \given X_{t_0}} - \lambda}} - \eta.
\]
This problem has been studied using stochastic approximation methods in \cite{Robbins1951,Kushner2003, Kushner2010} and recently applied to the MLMC framework in \cite{Frikha2016,Dereich2019,Dereich2021}. An alternative algorithm is proposed in \cite[Algorithm 2]{GilesHajiAli:2018}, although a {complete} analysis of this method {was} left to future work. The estimator of \(F\p{\lambda}\) constructed in previous sections can be included in such algorithms to efficiently approximate \(L_\eta\).
\section{CVA Capital Charge}
\label{sec:cva}
This section describes the model CVA problem \cite{Gregory2020, BrigoDamiano2013Ccrc, Pykhtin2012} used for the numerical experiments in \Cref{sec:num}. For ease of notation, the main discussion is based around a single counterparty. However, the method is intended for use in the more practical case of many counterparties. In \Cref{sec:multcp}, we discuss the extension from a single counterparty to an arbitrary number of counterparties.
\subsection{Market Setup}
\label{sec:market}
Consider a market based on a stock process $S_t \in \R^d$ given by
\[
\text{d}S_t = a(t, S_t)\text{d}t + b(t, S_t)\text{d}W_t; \quad S_0 \ge 0,
\]
up to the contract maturity $T>0$, where $W_t$ is a Wiener process under the physical measure $\mathbb{P}$ {and \(a, b\) represent the drift and volatility of the market, respectively.} We denote the (possibly stochastic) interest rate by $r_t$ so that the value of an initial investment $B_0$ in a risk-free bond at time $t$ is $B_t = B_0e^{\int_0^t r_t\text{d}t}$. {We consider a contract between a bank and a counterparty with discounted future cash-flow to the bank given by $\pi_t(S_T) = B_tB_T^{-1} \Pi\p{S_T}$, where \(\Pi\p{S_T}\) denotes the payoff at maturity \(T\) given stock value \(S_T\)}. We assume the existence of a risk-neutral measure $\mathbb{Q}$ under which the discounted stock process $B_t^{-1}S_t$ is a \(\mathbb{Q}\)-martingale. The arbitrage-free value of the contract at time $t$ is then given by
\begin{equation}
\label{eqn:valproc}
V_t(S_t) \coloneqq \E^\mathbb{Q}{\pi_t(S_T) | S_t}.
\end{equation}
\subsection{Credit Valuation Adjustment}
Suppose that the counterparty will default at a random time $\tau > 0$. In practice, the default time \(\tau\) is modelled according to the counterparties credit spread. In the event of default, the bank stands to lose a proportion $\textnormal{LGD}$ (Loss Given Default) of any positive cash flow. The impact this has on the valuation of the banks position at time $0\le t\le T$ is the CVA. {After discounting to time \(0\), the CVA is given by \cite{BrigoDamiano2013Ccrc}}
\[
\textnormal{CVA}_t \coloneqq \Ern*{\I{\tau_t \le T}\frac{1}{B_{\tau_t}}\textnormal{LGD} \p*{\Ern{\pi_{\tau_t}(S_T) | S_{\tau_t}, \tau_t}}^+\bigg | S_t, c_t},
\]
where $c_t$ represents the counterparties credit spread at time $t$ and $\tau_t = \tau_t\p{c_t}$ defines the random default time based on the market conditions at time $t$. {The factor \(B_t / B_{\tau_t}\) discounts the CVA to time \(t\), and the positive part of the value process, \(\p*{\Ern{\pi_{\tau_t}(S_T) | S_{\tau_t}, \tau_t}}^+\), is known as the exposure at default. {For convenience, we express \(\textnormal{CVA}_t = \E^\mathbb{Q}{U_{\tau_t}\p{S_{\tau_t}}\given S_t, c_t}\), where}
\begin{equation}
\label{eqn:u-def}
\begin{aligned}
U_{\tau_t}\p{x} &\coloneqq \I{t\le T}\frac{\textnormal{LGD}}{B_{\tau_t}}\p*{\Ern{\pi_{\tau_t}(S_T) | S_{\tau_t} = x, {\tau_t}}}^+.
\end{aligned}
\end{equation}
}
By an abuse of notation, we will consider \(U_{\tau_t}\) both as a function of \(x\) and of \(\Ern{\pi_{\tau_t}(S_T) | S_{\tau_t} = x, {\tau_t}} \) depending on the context. \\
We {introduce the following terms, which are used frequently in what follows}:
\begin{itemize}
\item $S_\rh^\mathbb{P}$ denotes the risky value of the stock price simulated under the physical measure $\mathbb{P}$ {at a short risk horizon \(\rh\)}.
\item $S_t^{t_0, s}$ denotes the stock process with \(S_{t_0}^{t_0, s} = s \).
\end{itemize}
The CVA-VaR is defined as the threshold $L_\eta$ above which losses due to changes in CVA between time $0$ and a short risk horizon $\rh\ll T$ occur with small probability $\eta$ (typically $\eta = 1\%$) under $\mathbb{P}$. Specifically,
\begin{equation}
\label{eqn:cva-var}
\begin{aligned}
\eta &= \prob{{\textnormal{CVA}_\rh - \textnormal{CVA}_0} > L_\eta} \\
&= \Ephys{\H{\Ern{\Lambda - L_\eta|S_\rh^\mathbb{P}, c_\rh}}},
\end{aligned}
\end{equation}
where the (random) loss due to a shift in the CVA is
\[
\Lambda \coloneqq U_{\tau_\rh}\p*{S_{\tau_\rh}^{\rh, S_\rh^\mathbb{P}}} - U_{\tau_0}\p*{S_{\tau_0}},
\]
where \(U_{\tau_t}\p*{x}\) given by \eqref{eqn:u-def} is a Lipschitz function of \(\Ern{\pi_{\tau_t}(S_T) | S_{\tau_t} = x, {\tau_t}}\) and hence \eqref{eqn:cva-var} may be written in the form \eqref{eqn:var}{. In the context of \eqref{eqn:var}, \(X_t\) represents both the stock process \(S_t\) and credit spread \(c_t\), \(f\) is equivalent to \(\Lambda - L_\eta\) whereas \(g\) represents the discounted payoff \(\pi_\tau\p{S_T}\) and \(t_1\) represents both \(\tau_0\) and \(\tau_H\). In \Cref{sec:var-reduct}, we construct an equivalent estimator which requires sampling only \(t_1 = \tau_0\). Letting \(t_1\) be a random variable and \(g\p{x} = \pi_\tau\p{x}\) depend on \(\tau\) does not change the results of \Cref{sec:mlmc}, and is discussed in \Cref{rem:tau}.} The risk horizon $\rh$ usually represents a short interval, for example 1 week or 10 days, whereas the contract maturity $T$ could represent several years.\\
\subsection{Modelling the Default Time}
\label{sec:default-time}
As in \cite{Pykhtin2012}, we assume a flat credit term structure so that $c_t$ is constant under \(\mathbb{Q}\). Following the model for default probabilities in \cite[Chapter 12]{Gregory2020}, it follows that
\[
\tau_t = t + \hat \tau_t,
\]
where $\hat \tau_t$ are exponentially distributed random variables with rate $c_t / \textnormal{LGD}$. The choice of a flat credit term structure simplifies the analysis in the following sections, however we emphasize that the {multilevel Monte Carlo framework used} is general {and would also work with a variable} credit structure.\\
In practice, the initial credit spread can be inferred from market data. The risky dynamics of \(c_t\) up to the risk horizon under \(\mathbb{P}\) are assumed to follow \cite{Pykhtin2012}
\[
\text{d}c_t = \sigma^{\text{cs}} c_t\text{d}W_t^{\text{cs}},
\]
where $\sigma^{\text{cs}}>0$ is a fixed volatility constant, and $W_t^{\text{cs}}$ is a Wiener process under $\mathbb{P}$, {independent of \(W_t\) which drives the underlying stock process in \Cref{sec:market}}. The product $\sigma^{\text{cs}}c_0$ is linked with the external credit rating of the counterparty in \cite[Table A]{Pykhtin2012}.
\subsection{General Formulation for Multiple Counterparties}
\label{sec:multcp}
Typically, the CVA capital charge considers large portfolios containing many options with \(K\ge1\) counterparties, opposed to the one-dimensional problem considered above. In this case, the {adjustment made to the valuation of the banks portfolio is}
\[
\textnormal{CVA}_t \coloneqq \sum_{k=1}^K\textnormal{CVA}_t^{(k)},
\]
where \(\textnormal{CVA}^{(k)} \) denotes the CVA owing to counterparty \(k\). In this approach, the credit spread and loss given default will differ for each counterparty. The VaR becomes the value \(L_\eta\) solving
\begin{equation}
\label{eqn:cvavar-multcp}
\eta = \Ephys*{\Hbig{\Ern*{\sum_{k=1}^K\Lambda^{(k)} - L_\eta\ \bigg|\ S_\rh^\mathbb{P}, \{c_\rh^{(i)}\}_{i=1}^K}}},
\end{equation}
where \(\Lambda^{(k)}\) represents the loss due to CVA for the counterparty \(k\). {While we focus on the case \(K=1\) in this paper, the discussion applies in the extended case where \(K>1\). Some additional points to consider when {\(K\gg1\)} are discussed in \Cref{sec:extension}.}
\subsection{Variance Reduction}
\label{sec:var-reduct}
We now turn our attention to the approximation of \eqref{eqn:cva-var} using the MLMC methods in \Cref{sec:mlmc}. As mentioned in the introduction this approach reduces the $\Order{\textnormal{TOL}^{-5}}$ cost for standard Monte Carlo approximation to achieve root mean square error \(\textnormal{TOL}\) to $\Order{\textnormal{TOL}^{-2}\p{\log\textnormal{TOL}}^2}$. {The performance of (multilevel) Monte Carlo approximation} depends strongly on the variance of \(\Lambda \) as defined in \Cref{sec:cva}. As such we first consider variance reduction techniques, which are able to significantly improve the performance of MLMC. \\
There are two key factors which decrease the correlation between $U_{\tau_\rh}\p*{S_{\tau_\rh}^{\rh, S_\rh^\mathbb{P}}}$ and $U_{\tau_0}\p*{S_{\tau_0}}$ significantly, giving $\Lambda$ a large variance:
\begin{itemize}
\item $\tau_\rh, \tau_0$ have different rates and thus are difficult to correlate. It is unlikely that both of these terms will default before $T$.
\item Due to simulation of the market risk factor under $\mathbb{P}$, the stocks $S_t^{\rh, S_\rh^\mathbb{P}}$ and $S_t$ follow different {stochastic processes up} to the risk horizon \(\rh\).
\end{itemize}
Note that for the exponentially distributed default times in \Cref{sec:default-time} and some function \(F\) we have
\[
\begin{aligned}
\Ern{F(\tau_\rh) |c_\rh} &= \int_\rh^\infty F(t)\frac{c_\rh}{\textnormal{LGD}}e^{-c_\rh\p{t-\rh}/\textnormal{LGD}}\text{d}t\\
&= \frac{c_\rh}{c_0}e^{c_\rh \rh/\textnormal{LGD}}\int_0^\infty \I{t\ge \rh} F\p{t} \frac{c_0}{\textnormal{LGD}} e^{-\p{c_\rh - c_0}t/\textnormal{LGD}} e^{-c_0 t/\textnormal{LGD}}\text{d}t\\
&= \Ern*{\I{\tau_0\ge \rh}F(\tau_0)g_{\tau_0}(c_\rh)e^{c_\rh \rh/\textnormal{LGD}}\big|c_\rh},
\end{aligned}
\]
where $g_t(x) \coloneqq c_0^{-1}xe^{-\p{x-c_0}t/\textnormal{LGD}}$. {For ease of notation, we will express \(\tau_0\equiv \tau\) for the remainder of this paper.}
The above computation then motivates replacing \(\Lambda \) with
\begin{equation}
\label{eqn:lambd1}
\begin{aligned}
\Lambda^{\textrm{I}} = \I{\tau\ge \rh}\bigg( g_\tau(c_\rh)e^{c_\rh \rh/\textnormal{LGD}}&\underbrace{\p*{U_\tau\p{S_\tau^{\rh, S_\rh^\mathbb{P}}} - U_\tau(S_\tau)}}_{\text{Market risk}} \\
& +\underbrace{\p*{g_\tau\p{c_\rh}e^{c_\rh \rh/\textnormal{LGD}} - 1}}_{\text{Credit risk}} U_\tau(S_\tau) \bigg),
\end{aligned}
\end{equation}
whenever \( \tau \ge \rh\). The component of $\Lambda$ reflecting default before $\rh$ is $-\I{\tau<\rh}U_\tau\p{S_\tau}$. Since this is independent of the risk factors {at the horizon \(\rh\)}, we can cancel the bias induced by considering only $\tau\ge \rh$ in $\Lambda^{\textrm{I}}$ by setting
\[
L_\eta^{\textrm{I}} \coloneqq L_\eta + \Ern{\I{\tau< \rh}U_\tau\p{S_\tau}},
\]
where $\Ern{\I{\tau< \rh}U_\tau\p{S_\tau}}$ can be pre-computed offline at insignificant cost compared with the cost of computing $\eta$.
By the change of measure from \(\tau_\rh\) to \(\tau\) above, it follows that $\Ern{\Lambda - L_\eta|S_\rh^{\mathbb{P}}, c_\rh} = \Ern{\Lambda^{\textrm{I}} - L_\eta^{\textrm{I}}|S_\rh^{\mathbb{P}}, c_\rh}$. Note that we have decomposed $\Lambda^{\textrm{I}}$ into two terms with variance driven primarily by the market or credit risk factors as discussed above. We consider each term in turn.
\subsubsection{Credit Risk}
\label{sec:var-credit}
In \cite{Pykhtin2012}, it is shown that the standardised CVA capital charge can be derived by linearising
\begin{equation}
\label{eqn:gtau_lin}
g_\tau\p{c_\rh}e^{c_\rh \rh/\textnormal{LGD}} - 1 = \underbrace{\p{c_\rh - c_0}\p*{\frac{1}{c_0} - \frac{\tau}{\textnormal{LGD}}}}_{\Order{\sqrt{\rh}}} +\Order{\rh}.
\end{equation}
To remove order \(\sqrt{\rh}\) terms, we should take
\[
\Lambda^{\textrm{II}} \coloneqq \Lambda^{\textrm{I}} - \I{\tau\ge \rh} \p{c_\rh - c_0}\chi_\tau(S_\tau),
\]
where we introduce the control variate
\[
\chi_\tau(S_\tau) \coloneqq \p*{\frac{1}{c_0} - \frac{\tau}{\textnormal{LGD}}}U_\tau\p{S_\tau}.
\]
To ensure this does not bias the expected value of $\Lambda - L_\eta$, we set
\[
L_\eta^{\textrm{II}} \coloneqq L_\eta^{\textrm{I}} - \p{c_\rh-c_0}\Ern{\chi_\tau\p{S_\tau}},
\]
so that again $\Ern{\Lambda^{\textrm{II}} - L_\eta^{\textrm{II}}|S_\rh^{\mathbb{P}}, c_\rh} = \Ern{\Lambda - L_\eta|S_\rh^{\mathbb{P}}, c_\rh}$.
\subsubsection{Market Risk}
\label{sec:var-reduct-market}
The use of antithetic sampling and delta control variates to reduce the variance owed to market risk in VaR calculations is well documented \cite{glasserman2004, GilesHajiAli:2019sampling}. Following this approach leads to the term
\[
\begin{aligned}
\Lambda^{\textrm{III}} \coloneqq \I{\tau\ge \rh}&\bigg(g_\tau\p{c_\rh}e^{c_\rh \rh/\textnormal{LGD}}U_\tau\p{S_\tau^{\rh,S_\rh^\mathbb{P}}} \\
&- \frac{1}{2}\p*{U_\tau\p{S_\tau^+} + U_\tau\p{S_\tau^-}}\\
&-\frac{1}{2} \p{c_\rh-c_0}\p*{\chi_\tau\p{S_\tau^+} + \chi_\tau\p{S_\tau^-}}\\
& - \frac{1}{2}\p{S_\rh^{\rh, S_\rh^\mathbb{P}} - S_0}\p*{\nabla_{S_0}\cdot U_\tau\p{S_\tau^+} + \nabla_{S_0}\cdot U_\tau\p{S_\tau^-}}\bigg),
\end{aligned}
\]
where the antithetic pair $S_t^{\pm}$ follow the Brownian paths {$\{\pm W_t\}_{0\le t\le \rh}$} respectively. After $\rh$, all of the processes $S_t^\pm, S_t^{\rh, S_\rh^\mathbb{P}}$ follow the same Brownian path. The delta control variates \(\nabla_{S_0}\cdot U_\tau\p{S_\tau^\pm} \) arise from removing the leading order term in a Taylor expansion of \(U_\tau\p{S_\tau^{\rh, S_\rh^\mathbb{P}}} - U_\tau(S_\tau) \). The factor \(g_\tau\p{c_\rh}e^{c_\rh \rh/\textnormal{LGD}}\) of the market risk in \eqref{eqn:lambd1} can be ignored in this expansion since we see in \eqref{eqn:gtau_lin} that this term is of the form \(1 + \Order{\sqrt{\rh}}\). As before, to compensate for the bias induced by the delta control variate, we replace $L_\eta^{\textrm{II}}$ with
\[
L_\eta^{\textrm{III}} \coloneqq L_\eta^{\textrm{II}} - \p{S_\rh^{\rh, S_\rh^\mathbb{P}} - S_0}\:\Ern{\nabla_{S_0}\cdot U_\tau\p{S_\tau}}
\]
so that $\Ern{\Lambda^{\textrm{III}} - L_\eta^{\textrm{III}}|S_\rh^{\mathbb{P}}, c_\rh} = \Ern{\Lambda - L_\eta|S_\rh^{\mathbb{P}}, c_\rh}$. Since all of the expectations appearing in the definition of $L_\eta^{\textrm{III}}$ are independent of $S_\rh^{\mathbb{P}}$ {and} $c_\rh$, they may be pre-computed offline at insignificant cost compared to {\(\E^{\mathbb{Q}}{\Lambda^\textrm{III}}\)}.
\subsection{Sampling Rare Default Times}
\label{sec:rare-sampling}
Observe that the CVA loss, $\Lambda^{\textrm{III}}$, is non-zero only when $\rh\le\tau\le T$. For credible counterparties this is a rare event, making it costly to simulate a single instance of default. To address this issue note that, defining $\tilde p \coloneqq \prob{\rh\le\tau\le T}$, we have
\begin{equation}
\label{equation:importsampling}
\Ern*{\Lambda^{\textrm{III}}- L_\eta^{\textrm{III}}\bigg|S_\rh^\mathbb{P}, c_\rh} = \Ern*{\Lambda^{\textrm{III}}\tilde p - L_\eta^{\textrm{III}}\bigg|S_\rh^\mathbb{P}, c_\rh, \rh\le\tau\le T}.
\end{equation}
For $\rh\le t \le T$ and the model for \(\tau\) in \Cref{sec:default-time}, we have
\[
\prob{\tau \le t|\rh\le\tau\le T} = \frac{\prob{\rh\le\tau\le t}}{\prob{\rh\le\tau\le T}} = \frac{e^{-c_0\rh/\textnormal{LGD}} - e^{-c_0t/\textnormal{LGD}}}{e^{-c_0\rh/\textnormal{LGD}} - e^{-c_0T/\textnormal{LGD}}}.
\]
By inverting this equation for \(t\), we can sample directly from the conditional distribution using inverse transform sampling.\\
This approach relies explicitly on the model for default times described in \Cref{sec:default-time}. For more complex models of the default time, more advanced importance sampling techniques are required to achieve a similar result.
\subsection{Unbiased Estimation of the Loss}
\label{sec:unbiased-cva}
As previously remarked, it is not typically possible to sample the CVA loss \(\Lambda \) directly due to complicated market dynamics. The approach used in \Cref{sec:umlmc} provides an unbiased estimator of the form \(\E*{\Lambda \given S_\rh^\mathbb{P}, c_\rh} = \E*{ \dl[{\hat{\ell}}]^{(\textnormal{ant})}{\Lambda}\:q_{{\hat{\ell}}}^{-1} }\) where the multilevel corrections \(\dl{\Lambda} \) rely on Milstein simulation of the market and Monte Carlo valuation of the underlying contract. It is also possible to implement this approach for the estimator \(\Lambda^{\textrm{III}} \) above with the exception that the delta control variate \(\nabla_{S_0}\cdot U\) is discontinuous, and the multilevel correction term involving this control variate does not satisfy \Cref{assumpt:umlmc}, precluding the application of unbiased MLMC. However, to retain the benefits of the delta control variate in variance reduction, one may include this term only for a fixed, deterministic range of levels as in \cite{GilesHajiAli:2019sampling}.
\section{Numerical Experiments}
\label{sec:num}
To illustrate the preceding methods, we consider the estimation of the probability of large loss, $\eta$, for a toy CVA problem \eqref{eqn:cva-var} consisting of a single stock process and counterparty outlined in \Cref{sec:toyprob}. In \Cref{sec:extension}, we discuss how the methods can be readily extended to larger scale problems.
\subsection{Simple Synthetic Portfolio}
\label{sec:toyprob}
The underlying contract between the bank and counterparty is given by a linear combination of two European call options with discounted future cash-flow
\[
\pi_t(S_T) = \frac{B_t}{B_T}\p*{{c}_0\p{S_T - K_0}^+ + {c}_1\p{S_T - K_1}^+},
\]
where {constants} ${c}_0, {c}_1$ are determined by $V_0$ and the delta-neutral condition $\partial_{S_0}V_0 = 0$.
We assume the stock process follows a Geometric Brownian Motion under the Black-Scholes market assumptions with fixed interest rate $r = 1\%$ so that $B_t = e^{t/100}$. Then, under $\mathbb{P}$ and $\mathbb{Q}$, respectively, we have
\[
\begin{aligned}
\text{d}S_t &= \mu S_t\text{d}t + \sigma S_t\text{d}W_t,\\
\text{d}S_t &= rS_t\text{d}t + \sigma S_t \text{d}W_t^\mathbb{Q}.
\end{aligned}
\]
We take $T = 1$ year, as well as $S_0 = 1$ and $\mu=\sigma = 0.1$. To test the methods in greater generality, we ignore the existing analytic formulae for the value and stock processes. Instead, we resort to Milstein simulation of \(S_t\) and Monte Carlo estimation of the exposure at default. Thus, we rely upon the methods of \Cref{sec:unbiased-cva} to construct an unbiased estimator of the CVA loss. \\
Following the model for default times in \Cref{sec:default-time} we take $c_0 = 5\%$ and $\sigma^{\text{cs}} = 0.8\%/c_0$, corresponding to an A-rated counterparty according to \cite[Table A]{Pykhtin2012}. The probability of large loss \eqref{eqn:cva-var} is computed with risk horizon \(\rh = 10\) days and a fixed loss threshold \(L_\eta = 5\times 10^{-4}\), {giving \(\eta \approx 2\%\)}.
\subsection{MLMC Setup}
\label{sec:mlmcset}
Simulation of the outer expectation in \eqref{eqn:cva-var} over \(\mathbb{P} \) is performed using the methods in \Cref{sec:discontmlmc}. We consider both adaptive sampling with \(\Mmell = \Mmell[0]2^\ell\) middle simulation samples per level and non-adaptive sampling with \(\Mmell = \Mmell[0]2^{\gamma\ell}\) for \(\gamma = 1, 2\). {At level \(\ell = 0\) we use \(\Mmell[0] = 8\) samples per level.} The expected value of the loss \(\Lambda\) over \(\mathbb{Q}\) is approximated using a hierarchy of unbiased MLMC methods as in \Cref{sec:unbiased-cva}, and using the control variates discussed therein.
\subsubsection*{Optimal Starting Level}
\label{sec:opt-ell0}
For an effective comparison, it is important all methods {are tested with ideal hyper parameters}. For nested simulation, it is often optimal to modify the MLMC estimator to start not from $\ell = 0$ but at level $\ell = \ell_0>0$ \cite{GilesHajiAli:2018, GilesHajiAli:2019sampling}. The optimal starting level may be approximated by iteratively estimating the ratio \(R_{\ell_0}\) between the cost of a two level estimator starting at level \(\ell_0\) and a single level estimator at level \(\ell_0+1\). If \(R_\ell > 1\), it is more efficient to start at level \(\ell_0 + 1 \).
\subsubsection*{Error Estimation}
Robust computation of the MLMC estimator \eqref{eqn:mlmc-telsum} requires accurate estimates of the error terms $V_\ell \coloneqq \var{\dl{\mathbb{H}}}$ and $E_\ell \coloneqq \abs{\E{\dl{\mathbb{H}}}}$ in order to approximate the optimal maximum level $L$ and number of samples $\{M_\ell \}_{\ell=\ell_0}^L$ to use per level \cite{giles:2015review}. Reliable estimates for $V_\ell$ need $\Order {\kappa_\ell}$ outer samples, where $\kappa_\ell$ is the Kurtosis of $\dl{\mathbb{H}}$. {For the Heaviside function}, it follows that $\kappa_\ell \approx V_\ell^{-1}$. This contradicts the {methodology of} MLMC to reduce the number of samples required at {fine} levels, and can severely impact the robustness of the estimator. \\
In \cite{xu2020}, the authors approximate $\H{\cdot}$ using a sequence of smooth logistic sigmoid functions which depend on $\ell$ and significantly reduce $V_\ell, \kappa_\ell$ by a constant factor. Alternatively, in \cite{Collier:CMLMC} the authors assume $V_\ell \approx a_02^{-\beta\ell}, E_\ell\approx a_12^{-\alpha\ell}$ and use Bayesian estimates for the parameters $a_0, a_1, \alpha, \beta$. \\
Here, we describe an alternative, heuristic method based on a similar approach in \cite{Elfverson:2016selectiverefinement}. The multilevel correction terms \(\dl{\mathbb{H}}\) take the form
\[
\dl{\mathbb{H}} =
\begin{cases}
1 & \text{probability } p_\ell\\
-1 & \text{probability } q_\ell\\
0 & \text{probability} 1 - p_\ell - q_\ell
\end{cases},
\]
with
\[
\begin{aligned}
V_\ell \le p_\ell + q_\ell, \quad E_\ell = \abs{p_\ell - q_\ell}.
\end{aligned}
\]
When \(\ell\) is large, both \(p_\ell, q_\ell \ll 1\). Heuristically, when the number of samples \(\Moell\) of \(\dl{\mathbb{H}}\) is small, it is extremely unlikely we will observe a non-zero sample, creating unstable approximations to \(V_\ell, E_\ell\). The idea is to impose a Bayesian prior on \(p_\ell, q_\ell\) which incorporates knowledge of the theoretical convergence rates of \(V_\ell\) and \(E_\ell\) such that when the number of samples \(\Moell\) is low, we rely more heavily on the theoretical rates than on the observations, and vice-versa as \(\Moell\) becomes large.\\
Assume that we know a relationship of the form
\begin{equation}
\label{eqn:pqrel}
V_\ell \approx a_0 2^{-\beta\ell},
\end{equation}
where the constant \(a_0<1\) may be reasonably estimated using information from coarse levels where samples of \(\dl{\mathbb{H}} \) are inexpensive. Assume a Dirichlet prior density on \(p_\ell, q_\ell\)
\[
\rho_\ell\p{p, q} = \p{1 - p - q}^{c_\ell - 1}p^kq^k,
\]
for constants \(c_\ell, k > 0\). For the robustness of the MLMC estimator, it is preferable to over-estimate opposed to under-estimate \(V_\ell\). Hence, it is reasonable to impose
\[
\E_{\rho_\ell}{p_\ell + q_\ell} \ge V_\ell,
\]
which is true provided
\[
c_\ell \le 2\p{k+1}\p*{V_\ell^{-1} - 1}.
\]
Hence, using \eqref{eqn:pqrel}, a reasonable conservative estimate is
\begin{equation}
\label{eqn:clheur}
c_\ell = 2k\p{a_0^{-1}2^{\beta\ell} - 1}.
\end{equation}
Given observations \(\{\dl{\mathbb{H}}^{(m)}\}_{m=1}^{\Moell}\), define
\[
\nu_\ell^{\pm} \coloneqq \sum_{m=1}^{\Moell} \I{\dl{\mathbb{H}}^{(m)} = \pm 1}.
\]
Then, the maximum-a-posteriori estimate of \(p_\ell + q_\ell\) is
\[
\hat V_\ell = \frac{\nu_\ell^+ + \nu_\ell^- + 2k}{\Moell + 2k + c_\ell - 1}.
\]
{The parameter \(k\) can be chosen to control how confident we {are} that \(V_\ell \le \hat V_\ell\) for a small number of samples \(\Moell\). Larger values of \(k\) produce more reliable{, conservative} bounds.} \\
{To estimate \(E_\ell = \abs{p_\ell-q_\ell} \) we construct biased estimates for \(p_\ell - q_\ell\) and \(q_\ell - p_\ell\) separately. First, in the case where \(p_\ell > q_\ell\), we wish to estimate of \(p_\ell - q_\ell\). In doing so, we impose a \(\text{Dirichlet}\p{d_\ell, j + 1, 1}\) prior on \(p_\ell, q_\ell\) with density
\[
\tilde \rho_\ell\p{p, q} = \p{1 - p - q}^{d_\ell - 1}p^j.
\]
Given the observations \(\{\dl{\mathbb{H}}^{m}\}_{m=1}^{\Moell}\) the posterior distribution becomes a \(\text{Dirichlet}\p{d_\ell, \nu_\ell^+ + j +1, \nu_\ell^- + 1}\) distribution with maximum a posteriori estimator given by
\[
\frac{\nu_\ell^+ - \nu_\ell^- + j}{\Moell + j + d_\ell - 1}.
\]
Under the prior distribution we have
\[
\E_{\tilde \rho_\ell}{p_\ell - q_\ell} = \frac{j}{j + d_\ell + 2},
\]
so enforcing the condition
\[
\E_{\tilde \rho_\ell}{p_\ell - q_\ell} \ge e_02^{-\alpha\ell} \approx \abs{p_\ell - q_\ell}
\]
requires
\[
d_\ell \le j\p{e_0^{-1}2^{\alpha\ell} - 1} - 2.
\]
In the case where \(q_\ell > p_\ell\) one can repeat this process by interchanging the roles of \(p\) and \(q\). Taking the maximum of the two resulting estimators gives
\[
\hat E_\ell \coloneqq \frac{\abs{\nu_\ell^+ - \nu_\ell^-} + j}{\Moell + j + d_\ell - 1}.
\]
For the experiments presented here we choose \(k = j = 1\).}
\subsection{Results}
\Cref{fig:level-plots} plots various statistics of the multilevel correction terms \(\dl{\mathbb{H}}\) for each method. \Cref{fig:complexity} further plots the total work required to compute \(\eta\) to accuracy \(\textnormal{TOL}\), against \(\textnormal{TOL}\) for each method. We discuss the results below:
\subsubsection*{Work \(C_\ell\)}
The top left plot in \Cref{fig:level-plots} shows the expected work \(C_\ell\) of generating a single sample of \(\dl{\mathbb{H}}\). By construction, the non-adaptive samplers have \(\Order{2^{\gamma\ell}}\) cost (with some variability owed to the random cost of the unbiased MLMC estimators). On the other hand, the adaptive sampler has expected cost \(\E{2^{\ell + \eta_\ell}} = \Order{2^\ell}\) as discussed in \Cref{sec:discontmlmc}.
\subsubsection*{Variance \(V_\ell \)}
The top right plot in \Cref{fig:level-plots} shows the variance \(V_\ell = \var{\dl{\mathbb{H}}}\) as a function of \(\ell\). Following the discussion in \Cref{sec:discontmlmc}, the non-adaptive samplers have \(V_\ell = \Order{N_\ell^{-1/2}} = \Order{2^{-\gamma\ell /2}} \). On the other hand, the adaptive sampler has \(V_\ell = \Order{N_\ell^{-1}} = \Order{2^{-\ell}} \). Combined with an expected sampling cost of order \(2^\ell\), this improves the convergence rate of the adaptive MLMC over the non-adaptive estimators.
\subsubsection*{Kurtosis \(\kappa_\ell \)}
The bottom left plot in \Cref{fig:level-plots} shows the kurtosis of \(\dl{\mathbb{H}}\), \(\kappa_\ell\), against \(\ell\). We observe \(\kappa_\ell = \Order{V_\ell^{-1}} \) for all methods. As mentioned above, this reduces the accuracy of standard estimates for the overall error made by MLMC, especially when \(L\) is large. This motivates the use of the methods discussed in \Cref{sec:mlmcset} to improve the robustness of MLMC.
\subsubsection*{Optimal Starting Point}
The bottom right plot in \Cref{fig:level-plots} shows the ratio \(R_{\ell_0}\) between a two-level and single-level estimator starting at level \(\ell_0\) as discussed in \Cref{sec:mlmcset}. To ensure MLMC performs optimally, one should start from a level \(\ell_0\) for which \(R_{\ell_0} < 1\).
\subsubsection*{Total Work}
In \Cref{fig:complexity}, we see a reduction in cost by a factor of around 7 times when using adaptive opposed non-adaptive MLMC when computing \(\eta\) to 2 significant figures. We also see that our results closely match the theoretical complexities \(\Order{\textnormal{TOL}^{-5/2}} \) and \(\Order{\textnormal{TOL}^{-2}\p{\log \textnormal{TOL}}^2} \) stated for non-adaptive and adaptive sampling, respectively in \Cref{sec:discontmlmc}. {A reference line shows the theoretical \(\Order{\textnormal{TOL}^{-5}}\) complexity discussed in \Cref{sec:intro} for traditional Monte Carlo without any of the multilevel techniques in \Cref{sec:mlmc}. This illustrates savings of several orders of magnitude at the smallest computed error tolerance by using the adaptive hierarchical MLMC approach compared to using standard Monte Carlo.}
\begin{figure}
\centering
\begin{tabular}{ccc}
\ref*{pl:ad} Adaptive (\(r = 1.95\)) & \ref*{pl:gam1} $\gamma = 1$ & \ref*{pl:gam2} $\gamma = 2$
\end{tabular}
\begin{tabular}{c@{\hskip 2cm}c}
\begin{tikzpicture}[trim axis left, trim axis right]
\begin{axis}[
xlabel = $\ell$,
ylabel = $C_\ell$,
grid = both,
ymode = log,
ymin = 10^(2), ymax = 10^(7),
]
\addplot+[mark = asterisk, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{cost}/\thisrow{M}, col sep = comma,]{data/st-gam1-exp-pi-levels.csv};
\label{pl:gam1}
\addplot+[mark = square, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{cost}/\thisrow{M}, col sep = comma,]{data/st-gam2-exp-pi-levels.csv};
\label{pl:gam2}
\addplot+[mark = o, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{cost}/\thisrow{M}, col sep = comma,]{data/ad-umlmc-exp-pi-levels.csv};
\label{pl:ad}
\end{axis}
\end{tikzpicture}
&
\begin{tikzpicture}[trim axis left, trim axis right]
\begin{axis}[
xlabel = $\ell$,
ylabel = $V_\ell$,
grid = both,
ymode = log,
ymax = 10^(-0.5), ymin = 10^(-5.5),
]
\addplot+[mark = asterisk, mark size = 3pt, black, opacity = 0.6]
table[x = level, y expr = \thisrow{V}, col sep = comma,]{data/st-gam1-exp-pi-levels.csv};
\addplot+[mark = square, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{V}, col sep = comma,]{data/st-gam2-exp-pi-levels.csv};
\addplot+[mark = o, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{V}, col sep = comma,]{data/ad-umlmc-exp-pi-levels.csv};
\draw(3.8, 0.00002) node[fill = white, shape = rectangle]{\textcolor{gray}{$\Order{2^{-\ell}}$}};
\addplot+[mark = none, gray, domain = 2:8]{2^(-x)*10^(-2.5)};
\draw(5, 0.04) node[fill = white, shape = rectangle]{\textcolor{gray}{$\Order{2^{-\ell/2}}$}};
\addplot+[mark = none, densely dashed, gray, domain = 2:8]{2^(-x/2)*10^(-1.5)};
\end{axis}
\end{tikzpicture}\\
\begin{tikzpicture}[trim axis left, trim axis right]
\begin{axis}[
xlabel = $\ell$,
ylabel = $\kappa_\ell$,
grid = both,
ymode = log,
]
\addplot+[mark = asterisk, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{kurtosis}, col sep = comma,]{data/st-gam1-exp-pi-levels.csv};
\addplot+[mark = square, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{kurtosis}, col sep = comma,]{data/st-gam2-exp-pi-levels.csv};
\addplot+[mark = o, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{kurtosis}, col sep = comma,]{data/ad-umlmc-exp-pi-levels.csv};
\end{axis}
\end{tikzpicture}
&
\begin{tikzpicture}[trim axis left, trim axis right]
\begin{axis}[
xlabel = $\ell_0$,
ylabel = $R_{\ell_0}$,
grid = both,
ymin = 0.4, ymax = 1.6,
]
\addplot+[mark = asterisk, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{R}, col sep = comma,]{data/st-gam1-exp-pi-ell0.csv};
\addplot+[mark = square, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{R}, col sep = comma,]{data/st-gam2-exp-pi-ell0.csv};
\addplot+[mark = o, mark size = 3pt, black, opacity = 0.6,]
table[x = level, y expr = \thisrow{R}, col sep = comma,]{data/ad-umlmc-exp-pi-ell0.csv};
\addplot+[mark = none, black, domain=0:7]{1};
\end{axis}
\end{tikzpicture}
\end{tabular}
\caption{Statistics of the correction terms \(\dl{\mathbb{H}}\) plotted against $\ell$ for the MLMC estimators using both adaptive and non-adaptive sampling.}
\label{fig:level-plots}
\end{figure} \textsl{}\begin{figure}
\centering
\begin{tabular}{ccc}
\ref*{pl:ad} Adaptive (\(r = 1.95\)) & \ref*{pl:gam1} $\gamma = 1$ & \ref*{pl:gam2} $\gamma = 2$
\end{tabular}\\
\begin{tabular}{ccc}
\ref*{plot:complexity-MC} \(\textnormal{TOL}^{-5}\) & \ref*{pl:complex-log} \(\textnormal{TOL}^{-2}\p{\log \textnormal{TOL}}^2\) & \ref*{plot:complexity-big} \(\textnormal{TOL}^{-5/2}\)
\end{tabular}\\
\begin{tikzpicture}[trim axis left]
\begin{axis}[
xlabel = Normalised $\textnormal{TOL}$,
ylabel = Total work $\times \textnormal{TOL}^2$,
grid = both,
ymode = log,
xmode = log,
ymax = 10^(7.5),
width = 11cm, height = 5cm
]
\addplot+[mark = asterisk, mark size = 3pt, black, opacity = 0.6,]
table[x expr = \thisrow{tol}/0.035, y expr = \thisrow{cost}*\thisrow{tol}^2, col sep = comma,]{data/st-gam1-exp-pi-mlmc.csv};
\addplot+[mark = square, mark size = 3pt, black, opacity = 0.6,]
table[x expr = \thisrow{tol}/0.035, y expr = \thisrow{cost}*\thisrow{tol}^2, col sep = comma,]{data/st-gam2-exp-pi-mlmc.csv};
\addplot+[mark = o, mark size = 3pt, black, opacity = 0.6,]
table[x expr = \thisrow{tol}/0.035, y expr = \thisrow{cost}*\thisrow{tol}^2, col sep = comma,]{data/ad-umlmc-exp-pi-mlmc.csv};
\addplot+[black, mark = none, loosely dashdotted, domain = 0.1:0.005]{2*x^(-3)};
\label{plot:complexity-MC}
\addplot+[black, mark = none, densely dotted, domain = 0.1:0.005]{3600*x^(-0.5)};
\label{plot:complexity-big}
\addplot+[black, mark = none, densely dashed, domain = 0.1:0.005]{300*ln(x)^2};
\label{pl:complex-log}
\end{axis}
\end{tikzpicture}
\caption{Total work required to compute the probability of large loss due to changes in the CVA multiplied by $\textnormal{TOL}^2$ against $\textnormal{TOL}$, normalised by the true value of $\eta$.}
\label{fig:complexity}
\end{figure}
\subsection{Extension to Larger Portfolios}
\label{sec:extension}
The experiment above is based on a toy problem consisting of a single stock and counterparty. This setup is enough to test the methods of the preceding sections in the context of risk estimation of the CVA. However, it is important to stress that these methods can be readily extended to more realistic problems, with many counterparties and a large market to simulate. There are several additional {considerations} than can be useful when considering such problems:
\begin{itemize}
\item (\textit{Randomized subsampling of counterparties.}) For a large number of counterparties (\(K\gg 1 \) in \eqref{eqn:cvavar-multcp}) it can be more efficient to write the sum
\[
\sum_{k=1}^K \Lambda^{(k)} = \E{\Lambda^{\p*{k^\star}}q_{k^\star}^{-1}},
\]
where \(k^\star\) is drawn from a discrete probability distribution on \(\{1, \dots, k \} \). The expectation \(\E{\Lambda^{k^\star}q_{k^\star}^{-1}}\) may then be approximated using a Monte Carlo average over \(k^\star\). As a result, we need only consider a single counterparty with each sample of the inner expectation rather than evaluate each of \(\{\Lambda^{(k)} \}_{k=1}^K\) for each sample. This approach, known as randomised sub-sampling \cite{GilesHajiAli:2019sampling}, can reduce the cost of the MLMC estimator when \(K\) is sufficiently large.
\item (\textit{Simulation of market at large number of intermediate points.}) When there are many counterparties with different default times and/or many options per counterparty with differing maturities, it may be necessary to evaluate the stock process at many intermediate time points. It is possible to modify the Milstein discretisation to ensure the underlying Brownian path is sampled at each intermediate point. However, when the number of intermediate points is large, this can add a significant cost to the computation of the unbiased MLMC. One possible alternative is to interpolate the intermediate points using an antithetic approach similar to that used in \Cref{sec:var-reduct-market}, by sampling an antithetic pair \(S^{f, \pm}_t \) at the fine level which follow \(\pm\) an underlying Brownian path between each pair of coarse time points. If the coarse path is \(S_t^c\) and the Brownian path is sampled at evenly space timepoints over \([0, T] \), numerical experiments indicate that using linear interpolation the multilevel correction term
\[{\dl{^\textnormal{ant} S_t}} \coloneqq \frac{1}{2}\p{S^{f,+}_t + S^{f,-}_t} - S^c_t \]
satisfies the necessary condition for unbiased MLMC in \Cref{assumpt:umlmc} for all values \(t \in [0,T]\). {Analysis of such an approach is left to future work.}
\end{itemize}
\section{Conclusion}
Due to large costs induced by the approximation of market factors combined with several nested Monte Carlo averages, accurate estimation of xVA risk measures is often {considered} infeasible using traditional Monte Carlo methods. By combining recent developments in MLMC for nested simulation with variance reduction techniques, this paper illustrates an efficient framework for accurate simulation of these risk measures at greatly reduced costs. In the context of the CVA-VaR problem, we are able to reduce the cost of achieving accuracy $\textnormal{TOL}$ from $\Order{\textnormal{TOL}^{-5}}$ for traditional Monte Carlo to $\Order{\textnormal{TOL}^{-2}\p{\log\textnormal{TOL}}^2}$ using MLMC with an adaptive number of inner samples of the loss. {This results in computational savings of several orders of magnitude for realistic problems.} Paired with the capabilities of high performance computing, this shows that Monte Carlo methods are a viable option for efficient and accurate simulation of xVA risk measures. \\
{Future aspects of this work involve the use of stochastic approximation techniques to compute the Value-at-Risk as in \Cref{sec:quantile}, along with an application to larger portfolios as discussed in \Cref{sec:extension}. An additional area of future research concerns multilevel Quasi-Monte Carlo methods for the approximation of \eqref{eqn:var}.} Such methods are used in \cite{xu2020} to construct an estimator with $\Order{\textnormal{TOL}^{-2}\p{\log\textnormal{TOL}}^2}$ cost without adaptive sampling techniques. When combined with the adaptive methods, we can potentially~ construct an estimator with even better costs. For problems with similar convergence properties \cite{gw09:MLQMC}, multilevel Quasi-Monte Carlo estimators have been shown to have $\Order{\textnormal{TOL}^{-1}}$ cost to achieve accuracy $\textnormal{TOL}$.
\subsubsection*{Acknowledgments}
J. Spence was supported by EPSRC grant EP/S023291/1.
\small
\bibliographystyle{abbrv}
|
\section{INTRODUCTION}
Quantum entanglement is one of the most important features in quantum mechanics. The quantum entangled states \cite{EPR,GHZ1,W1,QE} are significant ingredients in quantum information processing. Over past decades, various theoretical and experimental proposals have been presented for processing quantum information by using various systems such as atoms \cite{AP,A1,A2,A3,A4,A5,A6,A7,A8,A9}, spins \cite{S1,S2,S3,S4,S5,S6,S7}, ions \cite{I1,I2,I3,I4,I5,I6,I7,I8}, photons \cite{AP,PT1,PT2,PT3,PT4,PT5,PT6,PT7,PT8,PT9,PT10}, phonons \cite{PN1,PN2,PN3}, and so on. With the development of technologies, the quantum entanglement has been established not only in microscopic systems, but also in the macroscopic systems such as superconducting circuits \cite{SC1,SC2,SC3,SC4,SC5,SC6} and magnons system \cite{LJ1,LJ2,Kong,M1,M2,M3}.
Hybrid systems exploit the advantages of different quantum systems in achieving certain quantum tasks, such as creating quantum entanglement and carrying out quantum logic gates. Many works have been presented so far for quantum information processing in the hybrid systems \cite{HS1,HS2,HS3,HS4}. For instance, as an important quantum technology \cite{HSC}, the hybrid quantum circuits combine superconducting systems with other physical systems which can be fabricated on a chip. The superconducting (SC) qubit circuits \cite{SCQ1,SCQ2}, based on the Josephson junctions, can exhibit quantum behaviors even at macroscopic scale. Generally, the interaction between the SC qubits and the environment, e.g., systems in strong or even ultrastrong coupling regime via quantized electromagnetic fields, would result in short coherence time. Thus many researches on circuit quantum electrodynamics (QED) \cite{CQED} have been presented with respect to the SC qubits, superconducting coplanar waveguide resonators, \emph{LC} resonators and so on. This circuit QED focuses on studies of the light-matter interaction by using the microwave photons, and has become a relative independent research field originated from cavity QED.
The hybrid systems composed of collective spins (magnons) in ferrimagnetic systems and other systems are able to constitute the magnon-photon \cite{MT1,MT2}, magnon-phonon \cite{MN3,MN1,MN2}, magnon-photon-phonon \cite{LJ1,LJ2,MTN} systems and so on, giving rise to new interesting applications. Ferrimagnetic systems such as yttrium iron garnet (YIG) sphere have attracted considerable attention in recent years, which provide new platforms for investigating the macroscopic quantum phenomena particularly. Such systems are able to achieve strong and even ultrastrong couplings \cite{SU} between the magnons and the microwave photons, as a result of the high density of the collective spins in YIG and the lower dissipation. The YIG has the unique dielectric microwave properties with very lower microwave magnetic loss parameter. Meanwhile, some important works have been presented on magnon Kerr effect \cite{Kerr1,Kerr2}, quantum transduction \cite{MQT}, magnon squeezing \cite{MS1,MS2}, magnon Fock state \cite{MFS} and entanglement of magnons. For example, In 2018 Li \emph{et al}. \cite{LJ1} proposed a system consisted of magnons, microwave photons and phonons for establishing tripartite entangled states based on the magnetostrictive interaction and that the entangled state in magnon-photon-phonon system is robust. In 2019 Li \emph{et al.} \cite{LJ2} constructed the entangled state of two magnon modes in a cavity magnomechanical system by applying a strong red-detuned microwave field on a magnon mode to activate the nonlinear magnetostrictive interaction. In 2021 Kong \emph{et al.} \cite{Kong} used the indirect coherent interaction for accomplishing two magnons entanglement and squeezing via virtual photons in the ferromagnetic-superconducting system.
In this work, we first present a hybrid system composed of two YIG spheres, two identical microwave cavities and a SC qubit to establish quantum entanglement on two nonlocal magnons. In this system, two YIGs are coupled to respective microwave cavities that cross each other. And a SC qubit is placed at the center of the crossing of two identical cavities, namely, the SC qubit interacts with the two cavities simultaneously. The magnons in YIGs can be coupled to the microwave cavities in the resonant way, owing to that the frequencies of two magnons can be tuned by biased magnetic fields, respectively. Compared with other works, the SC qubit is coupled to the two microwave cavities in the far-detuning regime, meaning that the two identical cavities indirectly interact with each other by exchanging virtual photons. Then, we give the effective Hamiltonian of the subsystem composed of the SC qubit and two cavities, and present the protocol of entanglement establishment. In Sec. \ref{3M}, we consider the case of three magnons. In the hybrid system shown in Fig.\ref{fig3}, the three identical microwave cavities could indirectly interact via the virtual photons, and each magnon is resonant with the respective cavity by tuning the frequency of the magnon. At last, we get the isoprobability entanglement on three nonlocal magnons. Moreover, the hybrid system composed of \emph{N} magnons, \emph{N} identical microwave cavities and a SC qubit is derived in Sec. \ref{4M}. We summarize in Sec. \ref{5M}.
\section{QUANTUM ENTANGLEMENT ON TWO NONLOCAL MAGNONS}\label{2M}
\subsection{Hamiltonian of the hybrid system}
We consider a hybrid system, see Fig.\ref{fig1}, in which two microwave cavities cross each other, two yttrium iron garnet (YIG) spheres are coupled to the microwave cavities, respectively. A superconducting (SC) qubit, represented by black spot in the Fig.\ref{fig1}, is placed at the center of the crossing in order to interact with the two microwave cavities simultaneously. The YIG spheres are placed at the antinode of two microwave magnetic fields, respectively, and a static magnetic field is locally biased in each YIG sphere. In our model, the SC qubit is a two-level system with ground state $|g\rangle\!_q$ and excited state $|e\rangle\!_q$.
\begin{figure
\centering
\includegraphics[width=8.5cm,angle=0]{Fig1.eps}
\caption{(Color online) Schematic of the hybrid system composed of two yttrium iron garnet spheres coupled to respective microwave cavities. Two cavities cross each other, and a superconducting qubit (black spot) is placed at the center of the crossing. }\label{fig1}
\end{figure}
The magnetostatic modes in YIG can be excited when the magnetic component of the microwave cavity field is perpendicular to the biased magnetic field. We only consider the Kittel mode \cite{Kittel} in the hybrid system, namely, the magnon modes can be excited in YIG. The frequency of the magnon is in the gigahertz range. Thus the magnon generally interacts with the microwave photon via the magnetic dipole interaction. The frequency of the magnon is given by $\omega_m=\gamma H$, where $H$ is the biased magnetic field and $\gamma/2\pi=28$ GHz/T is the gyromagnetic ratio.
In recent years, some experiments have already realized the strong and ultrastrong magnon-magnon coupling \cite{mm1,mm2,mm3} as well as the magnon-qubit interaction \cite{MSC,PMS}, which means that in the hybrid system shown in Fig.\ref{fig1} the magnon is both coupled to the SC qubit and another magnon. However, we mainly consider that the magnons which frequencies are tuned by the locally biased static magnetic fields can be resonant with the cavities. In the meantime, the two cavities modes interact indirectly in the far-detuning regime for exchanging photons. The entanglement of two nonlocal magnons can be constructed by using two cavities and the SC qubit. Given that there are magnon-magnon and magnon-qubit interactions, the magnon can be detuned with the qubit and another magnon in order to neglect their interactions. In the rotating wave approximation the Hamiltonian of the hybrid system is ($\hbar=1$ hereafter) \cite{Hm}
\begin{eqnarray}
H^{(\mathrm{S})}\!=H_0+H_{\mathrm{int}} \nonumber
\end{eqnarray}
\begin{eqnarray}
H_0&\!\!\!=\!\!\!&\omega_{m_1}m_1^\dag m_1+\omega_{m_2}m_2^\dag m_2+\frac{1}{2}\omega_q\sigma_z \nonumber \\
&&+\omega_{a_1}a_1^\dag a_1+\omega_{a_2}a_2^\dag a_2 \nonumber
\end{eqnarray}
\begin{eqnarray}
H_{\mathrm{int}}&\!\!\!=\!\!\!&\lambda_{m_1}(a_1m^\dag_1+a_1^\dag m_1)+\lambda_{m_2}(a_2m^\dag_2+a_2^\dag m_2) \nonumber \\
&&+\lambda_{q_1}(a_1\sigma^++a_1^\dag\sigma)+\lambda_{q_2}(a_2\sigma^++a_2^\dag\sigma).
\end{eqnarray}
Here, $H_0$ is the free Hamiltonian of the two cavities, two magnons and the SC qubit. $H_{\mathrm{int}}$ is the interaction Hamiltonian among the cavities, magnons and SC qubit. $\omega_{m_1}$ and $\omega_{m_2}$ are the frequencies of the two magnons, which are tunable under biased magnetic fields, respectively. $\omega_{a_1}$ and $\omega_{a_2}$ are the frequencies of two cavities, and $\omega_q$ is the state transition frequency between $|g\rangle\!_q\leftrightarrow|e\rangle\!_q$ of the SC qubit. In the Kittel mode, the collective spins in YIGs can be expressed by the boson operators. $m_1$ ($m_2$) and $m_1^\dag$ ($m_2^\dag$) are the annihilation and creation operators of magnon mode 1 (2). $a_1$ ($a_2$) and $a_1^\dag$ ($a_2^\dag$) denote the annihilation and creation operators of cavity mode 1 (2), respectively. They satisfy commutation relations $[O,O^\dag]=1$ for $O=a_1, a_2, m_1, m_2$. $\sigma_z=|e\rangle_q\langle e|-|g\rangle_q\langle g|$. $\sigma=|g\rangle_q\langle e|$ and $\sigma^+=|e\rangle_q\langle g|$ are the lowing and raising operators of the SC qubit. $\lambda_{q_1}$ ($\lambda_{q_2}$) is the coupling strength between the SC qubit and the cavity mode 1 (2). $\lambda_{m_1}$ ($\lambda_{m_2}$) is the coupling between the magnon mode 1 (2) and the cavity mode 1 (2).
As mentioned above, the two microwave cavities are identical ones with the same frequency $\omega_{a_1}=\omega_{a_2}=\omega_a$. Meanwhile, one can assume that $\lambda_{q_1}=\lambda_{q_2}=\lambda_q$. In the interaction picture with respect to $e^{-\mathrm{i}H_0 t}$, the Hamiltonian is expressed as
\begin{eqnarray}
H^{(\mathrm{I})}&\!\!=\!\!&\lambda_{m_1}a_1m^\dag_1e^{\mathrm{i}\delta_1 t}+\lambda_{m_2}a_2m^\dag_2e^{\mathrm{i}\delta_2 t}+\lambda_qa_1\sigma^+e^{\mathrm{i}\Delta_1 t} \nonumber \\
&&+\lambda_qa_2\sigma^+e^{\mathrm{i}\Delta_2 t}+H.c.,
\end{eqnarray}
where $\delta_1=\omega_{m_1}-\omega_a$, $\delta_2=\omega_{m_2}-\omega_a$, $\Delta_1=\omega_q-\omega_a$ and $\Delta_2=\omega_q-\omega_a$. The SC qubit is coupled to the two cavities simultaneously. Owing to $\Delta_1=\Delta_2=\Delta_0\neq0$ and $\Delta_0\gg\lambda_q$, the two identical microwave cavities indirectly interact with each other in the far-detuning regime. Therefore, the effective Hamiltonian of the subsystem composed of the two microwave cavities and the SC qubit in the far-detuning regime is given by \cite{EH}
\begin{eqnarray}
H_{\mathrm{eff}}\!=\!\widetilde{\lambda}_q\!\left[\sigma_z(a_1^{\dag}a_1+a_2^{\dag}a_2+a_1^{\dag}a_2+a_1a_2^{\dag})+2|e\rangle_q\langle e|\right], \label{eff1}
\end{eqnarray}
where $\widetilde{\lambda}_q=\lambda_q^2/\Delta_0$.
\subsection{Entangled state generation on two nonlocal magnons}
We now give the protocol of quantum entanglement generation on two nonlocal magnons. Generally, the magnon can be excited by a drive magnetic field. For convenience the state of magnon 1 is prepared as $|1\rangle\!_{m\!_1}$ via the magnetic field. The initial state of the hybrid system is $|\varphi\rangle_0=|1\rangle\!_{m\!_1}|0\rangle\!_{m\!_2}|0
\rangle\!_{a\!_1}|0\rangle\!_{a\!_2}|g\rangle\!_q$, in which the two cavities are all in the vacuum state, magnon 2 is in the state $|0\rangle\!_{m\!_2}$, and the SC qubit is in state $|g\rangle\!_q$ which is unaltered all the time due to the indirect interaction between the two cavities.
\emph{step 1}: The frequency of magnon 1 is tuned to be $\omega_{m_1}\!=\!\omega_{a_1}$ so that the cavity 1 could be resonated with it. Therefore, the magnon 1 and cavity 1 are in a superposed state after time $T_1=\pi/4\lambda_{m_1}$. The local evolution is $|1\rangle\!_{m\!_1}|0\rangle\!_{a\!_1}\rightarrow\frac{1}{\sqrt{2}}(|1\rangle\!_{m\!_1}|0\rangle\!_{a\!_1}-\mathrm{i}|0\rangle\!_{m\!_1}|1\rangle\!_{a\!_1})$, which means that the states of SC qubit, magnon 2 and cavity 2 are unchanged due to decoupling between the SC and two cavities, and the magnon 2 is far-detuned with cavity 2. The state evolves to
\begin{eqnarray}
|\varphi\rangle_1&\!\!\!=\!\!\!&\frac{1}{\sqrt{2}}(|1\rangle\!_{m\!_1}|0\rangle\!_{a\!_1}-\mathrm{i}|0\rangle\!_{m\!_1}|1\rangle\!_{a\!_1}) \nonumber \\
&&\otimes|0\rangle\!_{m\!_2}\otimes|0\rangle\!_{a\!_2}\otimes|g\rangle\!_q.
\end{eqnarray}
\emph{step 2}: The magnons are tuned to far detune with respective cavities. From Eq. (\ref{eff1}), the evolution of subsystem composed of two microwave cavities and SC qubit is given by
\begin{eqnarray}
|\chi(t)\rangle\!_{\mathrm{sub}}&\!\!=\!\!&e^{\mathrm{i}\widetilde{\lambda}_qt}\big[\cos(\widetilde{\lambda}_qt)|1\rangle\!_{a\!_1}|0\rangle\!_{a\!_2}
+\mathrm{i}\sin(\widetilde{\lambda}_qt)|0\rangle\!_{a\!_1}|1\rangle\!_{a\!_2}\big] \nonumber \\
&&\otimes|g\rangle\!_q
\end{eqnarray}
under the condition $\Delta_0\gg\lambda_q$.
After time $T_2=\pi/2\widetilde{\lambda}_q$, the evolution between two cavities is $|1\rangle\!_{a\!_1}|0\rangle\!_{a\!_2}\rightarrow-|0\rangle\!_{a\!_1}|1\rangle\!_{a\!_2}$, which indicates that the photon can be indirectly transmitted between the two cavities, with the state of SC qubit unchanged. Therefore, the state after this step changes to
\begin{eqnarray}
|\varphi\rangle_2&\!\!\!=\!\!\!&\frac{1}{\sqrt{2}}(|1\rangle\!_{m\!_1}|0\rangle\!_{a\!_1}|0\rangle\!_{a\!_2}
+\mathrm{i}|0\rangle\!_{m\!_1}|0\rangle\!_{a\!_1}|1\rangle\!_{a\!_2}) \nonumber \\
&&\otimes|0\rangle\!_{m\!_2}\otimes|g\rangle\!_q.
\end{eqnarray}
\emph{step 3}: The frequency of magnon 2 is tuned with $\omega_{m_2}=\omega_{a_2}$ to resonate with the cavity 2. In the meantime the cavities are decoupled to the SC qubit and the magnon 1 is far detuned with the cavity 1. After time $T_3=\pi/2\lambda_{m_2}$, the local evolution $|0\rangle\!_{m\!_2}|1\rangle\!_{a\!_2}\rightarrow-\mathrm{i}|1\rangle\!_{m\!_2}|0\rangle\!_{a\!_2}$ is attained. The final state is
\begin{eqnarray}
|\varphi\rangle_3&\!\!\!=\!\!\!&\frac{1}{\sqrt{2}}(|1\rangle\!_{m\!_1}|0\rangle\!_{m\!_2}
+|0\rangle\!_{m\!_1}|1\rangle\!_{m\!_2}) \nonumber \\
&&\otimes|0\rangle\!_{a\!_1}\otimes|0\rangle\!_{a\!_2}\otimes|g\rangle\!_q,
\end{eqnarray}
which is just the single-excitation Bell state on two nonlocal magnons.
In the whole process, we mainly consider the interactions between the magnons and the cavities, and between the cavities and the SC qubit. However, the SC qubit can be coupled to the magnons. In terms of Ref.\cite{MSC}, the interactions between the magnons and the SC qubit are described as $H_{qm,1}=\lambda_{qm,1}(\sigma^+m_1+H.c.)$ and $H_{qm,2}=\lambda_{qm,2}(\sigma^+m_2+H.c.)$ where $\lambda_{qm,1}=\lambda_q\lambda_{m_1}/\Delta_0$ and $\lambda_{qm,2}=\lambda_q\lambda_{m_2}/\Delta_0$, while the conditions $\omega_q=\omega_{m_1}$ and $\omega_q=\omega_{m_2}$ are attained. In the meantime, the two magnons are interacts each other by using the SC qubit. Generally, the frequencies of two magnon modes are tuned by the locally biased magnetic fields. Therefore, the magnon can be detuned with the SC qubit and another magnon in order to neglect the interactions between the magnons and the SC qubit.
\subsection{Numerical result}
\begin{figure*
\centering
\begin{tabular}{cc}
\includegraphics[width=9cm,angle=0]{Fig2a.eps} & \includegraphics[width=9cm,angle=0]{Fig2b.eps}\\
\includegraphics[width=9cm,angle=0]{Fig2c.eps} & \includegraphics[width=9cm,angle=0]{Fig2d.eps}
\end{tabular}
\caption{(a) The fidelity of the Bell state of two nonlocal magnons with respect to the coupling strength $\lambda_q$. Since $\widetilde{\lambda}_q=\lambda_q^2/\Delta_0$ in Eq.(\ref{eff1}), the fidelity is similar to parabola. (b)-(d) The fidelity of the Bell state versus the dissipations of cavities, magnons, and SC qubit, respectively. }\label{fig2}
\end{figure*}
We here simulate \cite{SIM} the fidelity of the Bell state on two nonlocal magnons by considering the dissipations of all constituents of the hybrid system. The realistic evolution of the hybrid system composed of magnons, microwave cavities and SC qubit is governed by the master equation
\begin{eqnarray}
\dot{\rho}&\!\!=\!\!&-\mathrm{i}[H^{(I)},\rho]+\kappa_{m_1}D[m_1]\rho+\kappa_{m_2}D[m_2]\rho\nonumber \\
&&+\kappa_{a_1}D[a_1]\rho+\kappa_{a_2}D[a_2]\rho+\gamma_qD[\sigma]\rho.
\end{eqnarray}
Here, $\rho$ is the density operator of the hybrid system, $\kappa_{m_1}$ and $\kappa_{m_2}$ are the dissipation rates of magnon 1 and 2, $\kappa_{a_1}$ and $\kappa_{a_2}$ denote the dissipation rates for the two microwave cavities 1 and 2, $\gamma_q$ is the dissipation rate of the SC qubit, $D[X]\rho\!\!=\!\!(2X\rho X^\dag-X^\dag X\rho-\rho X^\dag X)/2$ for $X=m_1, m_2, a_1, a_2, \sigma$. The fidelity of the entangled state of two nonlocal magnons is defined by $F=_3\!\!\langle\varphi|\rho|\varphi\rangle_3$.
The related parameters are chosen as $\omega_q/2\pi=7.92$ GHz, $\omega_a/2\pi=6.98$ GHz, $\lambda_q/2\pi=83.2$ MHz, $\lambda_{m_1}/2\pi=15.3$ MHz, $\lambda_{m_2}/2\pi=15.3$ MHz \cite{PMS}, $\kappa_{m_1}/2\pi$=$\kappa_{m_2}/2\pi=\kappa_m/2\pi=1.06$ MHz, $\kappa_{a_1}/2\pi=\kappa_{a_2}/2\pi=\kappa_a/2\pi=1.35$ MHz \cite{SU}, $\gamma_q/2\pi=1.2$ MHz \cite{MSC}. The fidelity of the entanglement between two nonlocal magnons can reach 92.9$\%$.
The influences of the imperfect relationship among parameters is discussed next. The Fig.\ref{fig2}(a) shows the fidelity influenced by the coupling strength between the microwave cavities and the SC qubit. Since $\widetilde{\lambda}_q=\lambda_q^2/\Delta_0$ in Eq.(\ref{eff1}), the fidelity is similar to parabola. In Fig.\ref{fig2}(b)-(d), we give the fidelity varied by the dissipations of cavities, magnons, and SC qubit. As a result of the virtual photon, the fidelity is almost unaffected by the SC qubit, shown in Fig.\ref{fig2}(d).
\section{ENTANGLEMENT GENERATION FOR THREE NONLOCAL MAGNONS}\label{3M}
\subsection{Entangled state of three nonlocal magnons}
Similar to the protocol of entangled state generation for two nonlocal magnons in two microwave cavities, we consider the protocol for entanglement of three nonlocal magnons. As shown in Fig.\ref{fig3}, similar to the hybrid system composed of two magnons coupled to the respective microwave cavities and a SC qubit in Fig.\ref{fig1}, there are three magnons in three YIGs coupled to respective microwave cavities and a SC qubit placed at the center of the three identical cavities ($\omega_{a_1}\!=\!\omega_{a_2}\!=\!\omega_{a_3}\!=\!\omega_a$). Each magnon is in biased static magnetic field and is located at the antinode of the microwave magnetic field.
\begin{figure
\centering
\includegraphics[width=8.5cm,angle=0]{Fig3.eps}
\caption{(Color online) Schematic of the hybrid system composed of three yttrium iron garnet spheres coupled to respective microwave cavities. A superconducting qubit (black spot) is placed at the center of the three cavities. }\label{fig3}
\end{figure}
In the interaction picture, the Hamiltonian of the hybrid system depicted in Fig.\ref{fig3} is
\begin{eqnarray}
H^{(\mathrm{I})}_3&\!\!=\!\!&\lambda_{m_1}a_1m^\dag_1e^{\mathrm{i}\delta_1 t}+\lambda_{m_2}a_2m^\dag_2e^{\mathrm{i}\delta_2 t} \nonumber \\
&&+\lambda_{m_3}a_3m^\dag_3e^{\mathrm{i}\delta_3 t}+\lambda_qa_1\sigma^+e^{\mathrm{i}\Delta_1 t} \nonumber \\
&&+\lambda_qa_2\sigma^+e^{\mathrm{i}\Delta_2 t}+\lambda_qa_3\sigma^+e^{\mathrm{i}\Delta_3 t}+H.c.,
\end{eqnarray}
where $\lambda_{m_3}$ is the coupling strength between magnon 3 and microwave cavity 3, $a_3$ and $m^\dag_3$ are annihilation operator of the cavity 3 and creation operator of the magnon 3, respectively. $\lambda_q$ is the coupling between the SC qubit and three cavities, $\delta_3=\omega_{m_3}-\omega_a$. The frequency $\omega_{m_3}$ can be tuned by the biased magnetic field in microwave cavity 3. $\Delta_3=\omega_q-\omega_a=\Delta_0$.
At the beginning we have the initial state $|\psi\rangle_0^{(3)}\!\!\!=\!\!|\psi\rangle\!_m^{(3)}\otimes|\psi\rangle\!_a^{(3)}\otimes|g\rangle\!_q$ with
$|\psi\rangle\!_m^{(3)}\!\!\!=\!\!|1\rangle\!_{m\!_1}|0\rangle\!_{m\!_2}|0\rangle\!_{m\!_3}\!=\!|100\rangle\!_m$ and
$|\psi\rangle\!_a^{(3)}\!\!\!=\!\!|0\rangle\!_{a\!_1}|0\rangle\!_{a\!_2}|0\rangle\!_{a\!_3}\!=\!|000\rangle\!_a$.
The single-excitation is set in the magnon 1. The magnon 1 is resonant with the cavity 1 by tuning the frequency of magnon 1, and the SC qubit is decoupled to the cavities. After time $T_1^{(3)}=\pi/2\lambda_{m_1}$, the local evolution $|1\rangle\!_{m\!_1}|0\rangle\!_{a\!_1}\rightarrow-\mathrm{i}|0\rangle\!_{m\!_1}|1\rangle\!_{a\!_1}$ is attained. The state is evolved to
\begin{eqnarray}
|\psi\rangle_1^{(3)}=-\mathrm{i}|000\rangle\!_m\!|100\rangle\!_a|g\rangle\!_q.
\end{eqnarray}
The SC qubit is coupled to the three identical microwave cavities at the same time in far-detuning regime $\Delta_0\gg\lambda_q$. Therefore, the effective Hamiltonian of the subsystem composed of the SC qubit and the three identical cavities is of the form \cite{EH}
\begin{eqnarray}
H_{\mathrm{eff}}^{(3)}&\!\!\!=\!\!\!&\widetilde{\lambda}_q\bigg[\sigma_z(a_1^\dagger a_1+a_2^\dagger a_2+a_3^\dagger a_3)+3|e\rangle_q\langle e| \nonumber \\
&&+\sigma_z(a_1a_2^\dagger+a_1a_3^\dagger+a_2a_3^\dagger+H.c.)\bigg]. \label{eff2}
\end{eqnarray}
The magnons are then all detuned with the cavities. The local evolution $e^{-\mathrm{i}H_{\mathrm{eff}}^{(3)}t}|100\rangle\!_a|g\rangle$ of the subsystem is given by
\begin{eqnarray}
|\chi(t)\rangle\!_{\mathrm{sub}}^{(3)}&\!\!\!=\!\!\!&\bigg[C^{(3)}_{1,t}|100\rangle\!_a+C^{(3)}_{2,t}|010\rangle\!_a+C^{(3)}_{3,t}|001\rangle\!_a\bigg] \nonumber \\
&&\otimes|g\rangle\!_q,
\end{eqnarray}
where $C^{(3)}_{1,t}=\frac{e^{\mathrm{i}3\widetilde{\lambda}_qt}+2}{3}$ and $C^{(3)}_{2,t}=C^{(3)}_{3,t}=\frac{e^{\mathrm{i}3\widetilde{\lambda}_qt}-1}{3}$. It is easy to derive that
\begin{eqnarray}
|C^{(3)}_{1,t}|^2+|C^{(3)}_{2,t}|^2+|C^{(3)}_{3,t}|^2=1.
\end{eqnarray}
Fig.\ref{fig4} shows the probability related to the states $|100\rangle\!_a|000\rangle\!_m|g\rangle\!_q$, $|010\rangle\!_a|000\rangle\!_m|g\rangle\!_q$ and
$|001\rangle\!_a|000\rangle\!_m|g\rangle\!_q$. In particular, one has $|C^{(3)}_{1,t}|^2=|C^{(3)}_{2,t}|^2=|C^{(3)}_{3,t}|^2=\frac{1}{3}$, with
\begin{eqnarray}
C^{(3)}_1=\frac{\sqrt{3}+\mathrm{i}}{2\sqrt{3}}, C^{(3)}_2=C^{(3)}_3=\frac{-\sqrt{3}+\mathrm{i}}{2\sqrt{3}}
\end{eqnarray}
at time $T_2^{(3)}=2\pi/9\widetilde{\lambda}_q$. Correspondingly, the state evolves to
\begin{eqnarray}
|\psi\rangle_2^{(3)}&\!\!\!=\!\!\!&\bigg[\frac{\sqrt{3}+\mathrm{i}}{2\sqrt{3}}|100\rangle\!_a+\frac{-\sqrt{3}+\mathrm{i}}{2\sqrt{3}}|010\rangle\!_a
+\frac{-\sqrt{3}+\mathrm{i}}{2\sqrt{3}}|001\rangle\!_a\bigg] \nonumber \\
&&\otimes(-\mathrm{i})|000\rangle\!_m\otimes|g\rangle\!_q.
\end{eqnarray}
\begin{figure
\centering
\includegraphics[width=8.5cm,angle=0]{Fig4.eps}
\caption{(Color online) Evolution probabilities of the states: $P_1=|C^{(3)}_{1,t}|^2$ for $|100\rangle\!_a|000\rangle\!_m|g\rangle\!_q$ (red), $P_2=|C^{(3)}_{2,t}|^2$ for $|010\rangle\!_a|000\rangle\!_m|g\rangle\!_q$, $P_3=|C^{(3)}_{3,t}|^2$ for $|001\rangle\!_a|000\rangle\!_m|g\rangle\!_q$, and $P_2=P_3$ (blue). }\label{fig4}
\end{figure}
Finally, the magnons can be resonated with the respective cavities under the condition $\{\delta_1, \delta_2, \delta_3\}=0$. The local evolution and the time are
$|0\rangle\!_{m\!_k}|1\rangle\!_{a\!_k}\rightarrow-\mathrm{i}|1\rangle\!_{m\!_k}|0\rangle\!_{a\!_k}$ and $T_{3k}^{(3)}=\pi/2\lambda_{m_k}$ ($k=1, 2, 3$), respectively. Thus the final state is
\begin{eqnarray}
|\psi\rangle_3^{(3)}&\!\!\!=\!\!\!&-\bigg[\frac{\sqrt{3}+\mathrm{i}}{2\sqrt{3}}|100\rangle\!_m\!+\!\frac{-\sqrt{3}+\mathrm{i}}{2\sqrt{3}}|010\rangle\!_m\!
+\!\frac{-\sqrt{3}+\mathrm{i}}{2\sqrt{3}}|001\rangle\!_m\!\bigg] \nonumber \\
&&\otimes|000\rangle\!_a\otimes|g\rangle\!_q.
\end{eqnarray}
In the whole process, the state of the SC qubit is kept unchanged.
\subsection{Numerical result}
\begin{figure
\centering
\includegraphics[width=9.0cm,angle=0]{Fig5a.eps}
\includegraphics[width=9.0cm,angle=0]{Fig5b.eps}
\includegraphics[width=9.0cm,angle=0]{Fig5c.eps}
\caption{ (a)-(c) The fidelity of the entanglement on three nonlocal magnons versus the dissipations of cavities, magnons and SC qubit.}\label{fig5}
\end{figure}
The entanglement fidelity of three nonlocal magnons is given here by taking into account the dissipations of hybrid system. Firstly, the master equation which governs the realistic evolution of the hybrid system composed of three magnons, three microwave cavities and a SC qubit can be expressed as
\begin{eqnarray}
\dot{\rho}^{(3)}&\!\!=\!\!&-\mathrm{i}[H^{(I)}_3,\rho^{(3)}]+\kappa_{m_1}D[m_1]\rho^{(3)}+\kappa_{m_2}D[m_2]\rho^{(3)} \nonumber \\
&&+\kappa_{m_3}D[m_3]\rho^{(3)}+\kappa_{a_1}D[a_1]\rho^{(3)}+\kappa_{a_2}D[a_2]\rho^{(3)} \nonumber \\
&&+\kappa_{a_3}D[a_3]\rho^{(3)}+\gamma_qD[\sigma]\rho^{(3)},
\end{eqnarray}
where $\rho^{(3)}$ is the density operator of realistic evolution of the hybrid system, $\kappa_{m_3}$ is the dissipation rate of magnon 3 with $\kappa_{m_3}/2\pi=\kappa_m/2\pi=1.06$ MHz \cite{SU}, $\kappa_{a_3}$ denotes the dissipation rate for the microwave cavities 3 with $\kappa_{a_3}/2\pi=\kappa_a/2\pi=1.35$ MHz \cite{SU}, $D[X]\rho^{(3)}\!\!=\!\!(2X\rho^{(3)} X^\dag-X^\dag X\rho^{(3)}-\rho^{(3)} X^\dag X)/2$ for any $X=m_1, m_2, m_3, a_1, a_2, a_3, \sigma$.
The entanglement fidelity for three nonlocal magnons is defined by $F^{(3)}=_3^{(3)}\!\!\langle\psi|\rho^{(3)}|\psi\rangle_3^{(3)}$, which can reach 84.9$\%$. The fidelity with respect to the parameters is shown in Fig.\ref{fig5}.
\section{\emph{N} MAGNONS SITUATION}\label{4M}
In Sec. \ref{2M} and Sec. \ref{3M}, the entanglement of two and three nonlocal magnons have been established. In this section we consider the case of \emph{N} magnons. In the hybrid system shown in Fig.\ref{fig6}, the SC qubit is coupled to \emph{N} cavity modes that have the same frequencies $\omega_a$. A magnon is coupled to the cavity mode in each cavity. Each magnon is placed at the antinode of microwave magnetic field of the respective cavity and biased static magnetic field.
\begin{figure
\centering
\includegraphics[width=8.5cm,angle=0]{Fig6.eps}
\caption{(Color online) Schematic of the hybrid system composed of \emph{N} yttrium iron garnet spheres coupled to respective microwave cavities. A superconducting qubit is placed at the center of the \emph{N} identical microwave cavities. }\label{fig6}
\end{figure}
In the interaction picture the Hamiltonian of whole system shown in Fig.\ref{fig6} can be expressed as
\begin{eqnarray}
H^{(\mathrm{I})}_N\!\!&=\!\!&\sum_n \bigg[\lambda_{m_n}(a_nm^\dag_ne^{\mathrm{i}\delta_n t}+H.c.) \nonumber \\
&&\,\,\,\,\,\,\,\,\,\,\,\,\,+\lambda_q(a_n\sigma^+e^{\mathrm{i}\Delta_n t}+H.c.)\bigg],
\end{eqnarray}
where $a_n$ and $m^\dag_n$ ($n=1,2,3,\cdots,N$) are the annihilation operator of the \emph{n}th cavity mode and the creation operator of the \emph{n}th magnon, $\lambda_{m_n}$ is the coupling between the \emph{n}th magnon and the \emph{n}th cavity mode, $\lambda_q$ denotes the coupling strength between the SC qubit and the \emph{n}th cavity mode, $\delta_n=\omega_{m_n}-\omega_a$, $\omega_{m_n}$ is the frequency of the \emph{n}th magnon, $\Delta_n=\Delta_0=\omega_q-\omega_a$.
The initial state is prepared as
\begin{eqnarray}
|\psi\rangle^{(N)}_0\!\!\!&=&\!\!\!|\psi\rangle\!_m^{(N)}\otimes|\psi\rangle\!_a^{(N)}\otimes|g\rangle\!_q, \\
|\psi\rangle\!_m^{(N)}\!\!\!&=&\!\!\!|1\rangle\!_{m\!_1}|0\rangle\!_{m\!_2}|0\rangle\!_{m\!_3}\cdots|0\rangle\!_{m\!_N}\!=\!|100\cdots0\rangle\!_m, \nonumber \\
|\psi\rangle\!_a^{(N)}\!\!\!&=&\!\!\!|0\rangle\!_{a\!_1}|0\rangle\!_{a\!_2}|0\rangle\!_{a\!_3}\cdots|0\rangle\!_{a\!_N}\!=\!|000\cdots0\rangle\!_a. \nonumber
\end{eqnarray}
At first, we tune the frequency of magnon 1 under the condition $\delta_1=0$. The magnon 1 is resonant with the cavity 1, which means that the single photon is transmitted to cavity 1, and the SC qubit is decoupled to all the cavities. The state evolves to
\begin{eqnarray}
|\psi\rangle^{(N)}_1\!\!\!&=&\!\!\!-\mathrm{i}|000\cdots0\rangle\!_m|100\cdots0\rangle\!_a|g\rangle\!_q
\end{eqnarray}
after time $T_1^{(N)}=\pi/2\lambda_{m_1}$.
Next the magnons are tuned to detune with respective cavities. The SC qubit is coupled to the \emph{N} microwave cavities at the same time in far-detuning regime $\Delta_0\gg\lambda_q$. Under the condition $\Delta_n=\Delta_0$, the effective Hamiltonian of the subsystem composed of the SC qubit and \emph{N} microwave cavities is of the form \cite{EH}
\begin{eqnarray}
H_{\mathrm{eff}}^{(N)}&\!\!=\!\!&\sum_n \widetilde{\lambda}_q\bigg[\sigma_za_n^\dagger a_n+|e\rangle_q\langle e|\bigg] \nonumber \\
&&+\sum_{l<n}\widetilde{\lambda}_q\bigg[\sigma_z(a_la_n^\dagger+H.c.)\bigg].
\end{eqnarray}
Consequently, the evolution of the hybrid system is given by
\begin{eqnarray}
|\psi\rangle^{(N)}_2&\!\!\!=\!\!\!&\bigg[C^{(N)}_{1,t}|100\cdots0\rangle\!_a+C^{(N)}_{2,t}|010\cdots0\rangle\!_a \nonumber \\
&&+C^{(N)}_{3,t}|001\cdots0\rangle\!_a+\cdots+C^{(N)}_{N,t}|000\cdots1\rangle\!_a\bigg] \nonumber \\
&&\otimes(-\mathrm{i})|000\cdots0\rangle\!_m\otimes|g\rangle\!_q,
\end{eqnarray}
where
\begin{eqnarray}
C^{(N)}_{1,t}&\!\!\!=\!\!\!&\frac{e^{\mathrm{i}N\widetilde{\lambda}_qt}+(N-1)}{N}, \nonumber \\
C^{(N)}_{2,t}=C^{(N)}_{3,t}&\!\!\!=\!\!\!&\cdots=C^{(N)}_{N,t}=\frac{e^{\mathrm{i}N\widetilde{\lambda}_qt}-1}{N}. \label{coeff}
\end{eqnarray}
In addition, we have the following relation
\begin{eqnarray}
\sum_n|C^{(N)}_{n,t}|^2\!\!&\!\!\!=\!\!\!&\!\!|C^{(N)}_{1,t}|^2+|C^{(N)}_{2,t}|^2+|C^{(N)}_{3,t}|^2+\cdots+|C^{(N)}_{N,t}|^2 \nonumber \\
&=&\!\!1
\end{eqnarray}
by straightforward calculation.
At last, the SC qubit is decoupled to the cavities, and the magnons are resonant with the cavities, respectively. Thus, after the time $T_{3n}^{(N)}=\pi/2\lambda_{m_n}$, the final state is given by
\begin{eqnarray}
|\psi\rangle^{(N)}_3&\!\!\!=\!\!\!&-\bigg[C^{(N)}_{1,t}|100\cdots0\rangle\!_m+C^{(N)}_{2,t}|010\cdots0\rangle\!_m \nonumber \\
&&+C^{(N)}_{3,t}|001\cdots0\rangle\!_m+\cdots+C^{(N)}_{N,t}|000\cdots1\rangle\!_m\bigg] \nonumber \\
&&\otimes|000\cdots0\rangle\!_a\otimes|g\rangle\!_q.
\end{eqnarray}
In the whole process, the state of SC qubit is unchanged all the time.
\begin{figure*}[!htbp]
\centering
\begin{tabular}{ccc}
\includegraphics[width=5.5cm,angle=0]{Fig7a.eps} & \includegraphics[width=5.5cm,angle=0]{Fig7b.eps} & \includegraphics[width=5.5cm,angle=0]{Fig7c.eps}
\end{tabular}
\caption{(a)-(c) Evolution probabilities for $N=4$ (left), $N=5$ (middle), and $N=6$ (right). If $N\geqslant5$, $p_1^{(N)}\neq p_2^{(N)}$ implies that the isoprobability entanglement does not exist.}\label{fig7}
\end{figure*}
{\it [Remark]} Concerning the coefficients Eq. (\ref{coeff}), the probabilities with respect to the states $|100\cdots0\rangle\!_m|000\cdots0\rangle\!_a|g\rangle\!_q$, $|010\cdots0\rangle\!_m|000\cdots0\rangle\!_a|g\rangle\!_q$, $|001\cdots0\rangle\!_m|000\cdots0\rangle\!_a|g\rangle\!_q$, $\cdots$, $|000\cdots1\rangle\!_m|000\cdots0\rangle\!_a|g\rangle\!_q$ are $p^{(N)}_1=|C^{(N)}_{1,t}|^2$, $p^{(N)}_2=|C^{(N)}_{2,t}|^2$, $p^{(N)}_3=|C^{(N)}_{3,t}|^2$, $\cdots$, $p^{(N)}_N=|C^{(N)}_{N,t}|^2$, respectively, and $p^{(N)}_2=p^{(N)}_3=\cdots=p^{(N)}_N$. If the condition $p^{(N)}_1=p^{(N)}_2$ can be attained, the isoprobability entanglement can be obtained. For instance, for $N=4$, the entangled state of the four nonlocal magnons is given by
\begin{eqnarray}
|\psi\rangle^{(4)}_3&\!\!\!=\!\!\!&-\frac{1}{2}\!\bigg[|1000\rangle\!_m\!-\!|0100\rangle\!_m\!-\!|0010\rangle\!_m\!-\!|0001\rangle\!_m\bigg] \nonumber \\
&&\otimes|0000\rangle\!_a\otimes|g\rangle\!_q.
\end{eqnarray}
However, if $N\geqslant5$, the isoprobability entanglement does not exist as a result of $p^{(N)}_1\neq p^{(N)}_2$, see illustration in Fig.\ref{fig7}(b)(c).
\section{SUMMARY AND DISCUSSION}\label{5M}
We have presented protocols of establishing entanglement on magnons in hybrid systems composed of YIGs, microwave cavities and a SC qubit. By exploiting the virtual photon, the microwave cavities can indirectly interact in far-detuning regime, and the frequencies of magnons can be tuned by the biased magnetic field, which leads to the resonant interaction between the magnons and the respective microwave cavities. We have constructed single-excitation entangled state on two and three nonlocal magnons, respectively, and the entanglement for \emph{N} magnons has been also derived in term of the protocol for three magnons.
By analyzing the coefficients in Eq. (\ref{coeff}), the isoprobability entanglement has been also constructed for cases $N=2$, $N=3$ and $N=4$. In particular, such isoprobability entanglement no longer exists for $N\geqslant5$.
In the protocol for the case of two magnons discussed in Sec. \ref{2M}, we have firstly constructed the superposition of magnon 1 and microwave cavity 1. Then the photon could be transmitted $|1\rangle\!_{a\!_1}|0\rangle\!_{a\!_2}\rightarrow-|0\rangle\!_{a\!_1}|1\rangle\!_{a\!_2}$ between two cavities. At last, the single-excitation Bell state is finally constructed in resonant way. As for $N\geqslant3$, however, such method is no longer applicable because of $|100\cdots0\rangle\!_a\nrightarrow\alpha_2|010\cdots0\rangle\!_a
+\alpha_3|001\cdots0\rangle\!_a+\cdots+\alpha_N|000\cdots1\rangle\!_a$, namely, $p_1^{(N)}\neq0$.
\section*{Acknowledgements}
This work is supported by the National Natural Science Foundation of China (NSFC) under Grant Nos. 12075159 and 12171044, Beijing Natural Science Foundation (Grant No. Z190005), the Academician Innovation Platform of Hainan Province.
|
\section{Introduction}
Grunewald, Segal, and Smith introduced the subgroup zeta function of finitely-generated group~\cite{Grunewald-Segal-Smith}, and
Du Sautoy and Grunewald gave a general method to compute such zeta functions using $p$-adic intregration and resolution of singularities~\cite{duSautoy-Grunewald}.
This motivated Voll and the second author to examine the setting where the multivariate polynomials factor linearly and found that the $p$-adic integrals are specializations of multivariate rational functions depending only on the combinatorics of the corresponding hyperplane arrangement~\cite{Maglione-Voll}.
After a natural specialization, its denominator greatly simplifies, and they conjecture that the numerator polynomial has nonnegative coefficients.
\medskip
In this work, we prove their conjecture and bring new tools to aid in the understanding of the underlying combinatorics of these rational functions.
Specifically, we reinterpret these numerator polynomials by introducing and studying the \emph{(Poincaré-)extended $\mathbf{a}\mathbf{b}$-index}, a polynomial generalizing both the \emph{Poincaré polynomial} and $\mathbf{a}\mathbf{b}$-index of the \emph{intersection poset} of the arrangement.
Both these polynomials have been studied extensively in combinatorics, although from different perspectives.
The coefficients of the Poincaré polynomial have interpretations in terms of the combinatorics and the topology of the
arrangement~\cite[Section~2.5]{BjornerEtAl}.
The $\mathbf{a}\mathbf{b}$-index, on the other hand, carries information about the order complex of the poset and is
particularly well-understood in the case of face posets of oriented
matroids---or, more generally, Eulerian posets. In those settings, the
$\mathbf{a}\mathbf{b}$-index encodes topological data via the \emph{flag
$f$-vector}~\cite{bayer-survey}.
\medskip
We study the extended $\mathbf{a}\mathbf{b}$-index in the generality of graded posets admitting $R$-labelings.
This class of posets includes intersection posets of hyperplane arrangements and, more generally, geometric lattices and geometric semilattices.
We show that the extended $\mathbf{a}\mathbf{b}$-index has nonnegative coefficients by interpreting them in terms of a combinatorial statistic generalizing a statistic given for the $\mathbf{a}\mathbf{b}$-index by Billera, Ehrenborg, and Readdy~\cite{billera-ehrenborg-readdy}.
This interpretation proves the mentioned conjecture~\cite{Maglione-Voll}, as well as a related conjecture from Kühne and the second author~\cite{Kuhne-Maglione}.
It also recovers and generalizes additional results from~\cite{billera-ehrenborg-readdy} and explains a close relationship between the Poincaré polynomial and the $\mathbf{a}\mathbf{b}$-index.
Concretely, we show that the extended $\mathbf{a}\mathbf{b}$-index can be obtained from the $\mathbf{a}\mathbf{b}$-index by a
suitable substitution and that the Poincaré polynomial is already encoded in the $\mathbf{a}\mathbf{b}$-index.
This allows us to define the \emph{pullback $\mathbf{a}\mathbf{b}$-index}, a polynomial in noncommuting variables $\mathbf{c} = \mathbf{a} + \mathbf{b}$ and $2\mathbf{d} = 2(\mathbf{a}\mathbf{b} + \mathbf{b}\mathbf{a})$.
In the special case of oriented matroids, this recovers the relationship between the $\mathbf{a}\mathbf{b}$-index of the lattice of flats and the $\mathbf{c}\mathbf{d}$-index of the face poset.
Notably, our approach recovers several well-know results without using the \emph{coalgebra techniques} of~\cite{billera-ehrenborg-readdy}.
\medskip
The paper is organized as follows.
In \Cref{sec:main-stuff}, we introduce the main technical definitions (see \Cref{sec:main-defs}) and state all the main results of the paper (see \Cref{sec:mainresults}).
The proofs of the main results are given in \Cref{sec:proof-of-combinatorial-interp,,sec:refinement-of-ber}.
\section{The Poincaré-Extended $\mathbf{a}\mathbf{b}$-index}
\label{sec:main-stuff}
\subsection{Main definitions}
\label{sec:main-defs}
Throughout this paper,~$P$ is a finite \Dfn{graded poset} of rank~$n$. That is, $P$ is a
finite poset with unique minimum element~$\hat 0$ of rank~$0$ and unique maximum
element~$\hat 1$ of rank~$n$ such that ${\sf rank}(X)$ is equal to the length of all
maximal chains from $\hat 0$ to $X$. We follow~\cite[Chapter 3]{ec1} for
definitions and notation surrounding posets. The \Dfn{M\"obius function} $\mu$
of $P$ is given by $\mu(X, X) = 1$ for all $X\in P$ and $\mu(X, Y) =
-\sum_{X\leq Z < Y} \mu(X, Z)$ for all $X < Y$ in $P$. The \Dfn{Poincaré
polynomial} of $P$ is
\[
\POIN{P}{y} = \sum_{X\in P} |\mu(\hat{0}, X)| \cdot y^{{\sf rank}(X)} \ \in \mathbb{Z}[y].
\]
The \Dfn{chain Poincaré polynomial} of a chain $\mathcal{C} = \big\{\mathcal{C}_1 < \dots < \mathcal{C}_k \big\}$ in~$P$ is
\[
\POIN[\mathcal{C}]{P}{y} =\prod_{i = 1}^k \POIN{[\mathcal{C}_i,\mathcal{C}_{i+1}]}{y} \ \in \mathbb{Z}[y],
\]
where we set~$\mathcal{C}_{k+1} = \hat{1}$. By taking the singleton chain $\{\hat 0\}$,
we recover the usual Poincaré polynomial, $\POIN{P}{y} = \POIN[\{\hat 0\}]{P}{y}$. The ranks of a given chain~$\mathcal{C}$ is given by
\[
{\sf Rank}(\mathcal{C}) = \left\{ {\sf rank}(\mathcal{C}_i) ~\middle|~ 1 \leq i \leq k \right\}.
\]
The following is the main object of study of this paper.
\begin{definition}
The (\Dfn{Poincaré-})\Dfn{extended $\mathbf{a}\mathbf{b}$-index} of~$P$ is
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})
= \sum_{\mathcal{C}\text{ chain in } P\setminus\{\hat{1}\}} \POIN[\mathcal{C}]{P}{y} \cdot {\sf wt}_\mathcal{C}(\mathbf{a},\mathbf{b}) \ \in \mathbb{Z}[y]\langle\mathbf{a},\mathbf{b}\rangle ,
\]
where~$\mathbf{a}$ and~$\mathbf{b}$ are noncommuting variables and ${\sf wt}_\mathcal{C}(\mathbf{a},\mathbf{b}) = w_0 \cdots w_{n-1}$ is given by
\begin{equation}\label{eqn:ext-weights}
w_i = \begin{cases}
\mathbf{b} & \text{if }i\in {\sf Rank}(\mathcal{C}), \\
\mathbf{a} - \mathbf{b} & \text{else}.
\end{cases}
\end{equation}
\end{definition}
Since~$P$ is has a unique minimum, we always have $\POIN{P}{0} = 1$, implying
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;0,\mathbf{a},\mathbf{b})
= \sum_{\mathcal{C}\text{ chain in } P\setminus\{\hat{1}\}} {\sf wt}_\mathcal{C}(\mathbf{a},\mathbf{b})\,.
\]
This recovers\footnote{This is actually a mild variant of the usual definition of the $\mathbf{a}\mathbf{b}$-index, see~\Cref{rem:bottom-element-toss-out} below.} the \Dfn{$\mathbf{a}\mathbf{b}$-index}
\[
\Psi(P;\mathbf{a},\mathbf{b}) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;0,\mathbf{a},\mathbf{b})\,.
\]
\begin{example}
\label{ex:face-poset2}
We compute the extended $\mathbf{a}\mathbf{b}$-index of the poset $\mathcal{L}$ drawn below on the left.
On the right, we collect the relevant data.
\begin{center}
\begin{tikzpicture}[scale=1.2, baseline={([yshift=-.5ex]current bounding box.center)}]
\node[invisivertex] (1) at ( 0,2){$\hat 1$};
\node[invisivertex] (H1) at (-1,1){$\alpha_1$};
\node[invisivertex] (H2) at ( 0,1){$\alpha_2$};
\node[invisivertex] (H3) at ( 1,1){$\alpha_3$};
\node[invisivertex] (0) at ( 0,0){$\hat 0$};
\draw (0) -- (H1) -- (1);
\draw (0) -- (H2) -- (1);
\draw (0) -- (H3) -- (1);
\end{tikzpicture}
\hspace{2cm}
\begin{tabular}{|c|c|c|c|}
\hline
&&&\\[-10pt]
$\mathcal{C}$ & $\POIN[\mathcal{C}]{\mathcal{L}}{y}$ & ${\sf Rank}(\mathcal{C})$ & ${\sf wt}_\mathcal{C}(\mathbf{a},\mathbf{b})$ \\[5pt]
\hline
&&&\\[-10pt]
$\{\}$ & $1$ & $\{\}$ & $(\mathbf{a}-\mathbf{b})^2$ \\[5pt]
$\{\hat 0\}$ & $1 + 3y + 2y^2$ & $\{0\}$ & $\mathbf{b}(\mathbf{a}-\mathbf{b})$ \\[5pt]
$\{\alpha_i\}$ & $1+y$ & $\{1\}$ & $(\mathbf{a}-\mathbf{b})\mathbf{b}$ \\[5pt]
$\{\hat 0 < \alpha_i\}$ & $(1+y)^2$ & $\{0,1\}$ & $\mathbf{b}^2$\\
\hline
\end{tabular}
\end{center}
The extended $\mathbf{a}\mathbf{b}$-index and its specialization to the $\mathbf{a}\mathbf{b}$-index are thus
\begin{align*}
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};y,\mathbf{a},\mathbf{b})
&= (\mathbf{a}-\mathbf{b})^2\! +\! (1 + 3y + 2y^2)\mathbf{b}(\mathbf{a}-\mathbf{b})\! +\! 3\cdot (1+y)(\mathbf{a}-\mathbf{b})\mathbf{b}\! +\! 3\cdot(1+y)^2\mathbf{b}^2 \\
&= \mathbf{a}^2 + (3y+2y^2)\mathbf{b}\mathbf{a} + (2+3y)\mathbf{a}\mathbf{b} + y^2\mathbf{b}^2, \\
\Psi(\mathcal{L};\mathbf{a},\mathbf{b}) &= \mathbf{a}^2 + 2\mathbf{a}\mathbf{b}\,.
\end{align*}
\end{example}
\begin{remark}
\label{rem:top-element-toss-out}
Taking chains~$\mathcal{C}$ in $P \setminus \{\hat 1\}$, rather than in~$P$, is a
harmless reduction in the definition of the extended $\mathbf{a}\mathbf{b}$-index since
$\POIN[\mathcal{C}]{P}{y} = \POIN[\mathcal{C}\cup\{\hat 1\}]{P}{y}$. If we consider both~$\mathcal{C}$
and~$\mathcal{C} \cup\{\hat 1\}$ separately as summands of ${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})$,
we would need to consider weights $\wt^+_\mathcal{C}(\mathbf{a},\mathbf{b}) = w_0 \cdots w_n$
taking the $n$th position into account; see~\Cref{eqn:ext-weights}. We would,
thus, have the two terms $\POIN[\mathcal{C}]{P}{y}\cdot\wt^+_{\mathcal{C}}(\mathbf{a},\mathbf{b})$ and
$\POIN[\mathcal{C} \cup \{\hat 1\}]{P}{y}\cdot\wt^+_{\mathcal{C} \cup \{\hat 1\}}(\mathbf{a},\mathbf{b})$,
differing only in the last entry of the weight. So their sum is
$\POIN[\mathcal{C}]{P}{y}\cdot{\sf wt}_{\mathcal{C}}(\mathbf{a},\mathbf{b})\cdot \mathbf{a}$. This argument holds for
all chains, proving
\begin{equation}
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})\cdot\mathbf{a} = \sum_{\mathcal{C}\text{ chain in } P} \POIN[\mathcal{C}]{P}{y} \cdot \wt^+_\mathcal{C}(\mathbf{a},\mathbf{b}) .
\label{eq:enab01}
\end{equation}
\end{remark}
\begin{remark}
\label{rem:bottom-element-toss-out}
In the case of the $\mathbf{a}\mathbf{b}$-index, a similar argument to the one in the
\Cref{rem:top-element-toss-out} implies that we could further restrict to
chains in $P\setminus\{\hat 0, \hat 1\}$. Therefore, we have
\begin{equation}
\Psi(P;\mathbf{a},\mathbf{b}) = \mathbf{a}\cdot\left(\sum_{\mathcal{C}\text{ chain in } P\setminus\{\hat 0,\hat 1\}} \wt^-_\mathcal{C}(\mathbf{a},\mathbf{b})\right),
\label{eq:ab01}
\end{equation}
where $\wt^-_\mathcal{C}(\mathbf{a},\mathbf{b}) = w_1 \dots w_{n-1}$ as neither~$0$ nor~$n$ can appear
in ${\sf Rank}(\mathcal{C})$. The expression in the parentheses recovers the usual
definition of the $\mathbf{a}\mathbf{b}$-index as given, for example, in~\cite[Section
2]{bayer-survey}. Using additional information, the extended $\mathbf{a}\mathbf{b}$-index
admits a similar property to that in \Cref{eq:ab01}, which we prove in
\Cref{cor:enab0}.
\end{remark}
The fact that $\hat 1$ is included in every chain in the computation of the
chain Poincaré polynomial is inspired by the setting of hyperplane arrangements.
A (central, real) \Dfn{hyperplane arrangement} $\mathcal{A}$ is a finite collection of
hyperplanes in~$\mathbb{R}^d$, all of which have a
common intersection.
The \Dfn{lattice of flats} $\mathcal{L}$ of $\mathcal{A}$
is the poset of subspaces of $\mathbb{R}^d$ obtained from
intersections of subsets of the hyperplanes, ordered by reverse inclusion.
The open, connected components of the complement $\mathbb{R}^d\setminus \mathcal{A}$ are called
(open) \Dfn{chambers}.
The set of (closed) \Dfn{faces} $\Sigma$ is the set of \emph{closures} of chambers of $\mathcal{A}$, together with all possible intersections of closures of chambers (ignoring intersections which are empty).
This set comes equipped with a natural partial order by reverse inclusion, and for this reason we refer to $\Sigma$ as the \Dfn{face poset} of $\mathcal{A}$.
There is an order-preserving, rank-preserving
surjection ${\sf supp} : \Sigma \twoheadrightarrow \mathcal{L}$ sending a face to its affine
span~\cite[Proposition~4.1.13]{BjornerEtAl}. This map extends to chains, and the
fiber sizes are given, for $\mathcal{C} = \{ \mathcal{C}_1 < \dots < \mathcal{C}_k \}\subseteq \mathcal{L}$, by
\begin{equation}
\#{\sf supp}^{-1}(\mathcal{C}) = \prod_{i=1}^{k} \POIN{[\mathcal{C}_i,\mathcal{C}_{i+1}]}{1} = \POIN[\mathcal{C}]{P}{1},
\label{eq:supppre}
\end{equation}
with $\mathcal{C}_{k+1} = \hat 1$; see~\cite[Proposition~4.6.2]{BjornerEtAl}.
This is the key motivation for the next definition.
\begin{definition}
\label{def:pullbackab}
The \Dfn{pullback $\mathbf{a}\mathbf{b}$-index} of~$P$ is
\[
\Psi_{\sf pull}(P;\mathbf{a},\mathbf{b}) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;1,\mathbf{a},\mathbf{b}).
\]
\end{definition}
Let~$\Sigma$ be the face poset and $\mathcal{L}$ the lattices of flats of a real central
hyperplane arrangement. Since~$\Sigma$ may not have a unique minimum element, we
formally add a minimum element~$\hat{0}$ and let $\Sigma \cup \{\hat 0 \}$ be
the resulting poset. Now, \Cref{eq:supppre} relates the $\mathbf{a}\mathbf{b}$-index of the face poset and the pullback $\mathbf{a}\mathbf{b}$-index of the lattice of flats by
\begin{equation}
\Psi(\Sigma \cup\{\hat 0\};\mathbf{a},\mathbf{b}) = \mathbf{a}\cdot\Psi_{\sf pull}(\mathcal{L};\mathbf{a},\mathbf{b})\,.
\label{eq:pullbackab}
\end{equation}
Note that this corresponds to the evaluation of ${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\Sigma\cup\{\hat 0\};y,\mathbf{a},\mathbf{b})$ at $y=0$
to the evaluation of ${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};y,\mathbf{a},\mathbf{b})$ at $y=1$.
\Cref{eq:supppre} and thus also \Cref{eq:pullbackab} hold indeed in the more general context of oriented
matroids.
\begin{example}
\label{ex:pullbackab}
The pullback $\mathbf{a}\mathbf{b}$-index of the poset from \Cref{ex:face-poset2} is
\[
\Psi_{\sf pull}(\mathcal{L};\mathbf{a},\mathbf{b}) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};1,\mathbf{a},\mathbf{b}) = \mathbf{a}^2 + 5\mathbf{b}\mathbf{a} + 5\mathbf{a}\mathbf{b} + \mathbf{b}^2\,.
\]
Consider the arrangement of three lines in the plane through a common intersection as shown below on the left in a way that emphasizes its face structure.
Its lattice of flats is the poset $\mathcal{L}$ from \Cref{ex:face-poset2}.
To the right, we draw its face poset~$\Sigma$ with~$\hat 0$ included.
\begin{center}
\begin{tikzpicture}[scale=1.2]
\draw[very thick] ( 0:-2) -- ( 0:2);
\draw[thick] (120:-2) -- (120:2);
\draw[thick] (240:-2) -- (240:2);
\draw[white, fill=white] (0,0) circle [radius=4pt];
\draw[fill] (0,0) circle [radius=2pt];
\node[invisivertex] at (30:1){\tiny$+++$};
\node[invisivertex] at (90:1){\tiny$-++$};
\node[invisivertex] at (150:1){\tiny$--+$};
\node[invisivertex] at (210:1){\tiny$---$};
\node[invisivertex] at (270:1){\tiny$+--$};
\node[invisivertex] at (330:1){\tiny$++-$};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[scale=1.5]
\node[invisivertex] (O) at (2.5,2){\tiny $000$};
\node[invisivertex] (H1+) at (0,1){\tiny $0++$};
\node[invisivertex] (H2+) at (1,1){\tiny $+0+$};
\node[invisivertex] (H3-) at (2,1){\tiny $-+0$};
\node[invisivertex] (H2-) at (3,1){\tiny $-0-$};
\node[invisivertex] (H1-) at (4,1){\tiny $0--$};
\node[invisivertex] (H3+) at (5,1){\tiny $+-0$};
\node[invisivertex] (C1) at (0,0){\tiny $+++$};
\node[invisivertex] (C2) at (1,0){\tiny $-++$};
\node[invisivertex] (C3) at (2,0){\tiny $-+-$};
\node[invisivertex] (C4) at (3,0){\tiny $---$};
\node[invisivertex] (C5) at (4,0){\tiny $+--$};
\node[invisivertex] (C6) at (5,0){\tiny $+-+$};
\node[invisivertex] (1) at (2.5,-1){\tiny $\hat 0$};
\path[-] (O) edge [] node[above] {} (H1+);
\path[-] (O) edge [] node[above] {} (H2+);
\path[-] (O) edge [] node[above] {} (H3-);
\path[-] (O) edge [] node[above] {} (H2-);
\path[-] (O) edge [] node[above] {} (H1-);
\path[-] (O) edge [] node[above] {} (H3+);
\path[-] (C1) edge [] node[above] {} (H1+);
\path[-] (C1) edge [] node[above] {} (H2+);
\path[-] (C2) edge [] node[above] {} (H1+);
\path[-] (C2) edge [] node[above] {} (H3-);
\path[-] (C3) edge [] node[above] {} (H3-);
\path[-] (C3) edge [] node[above] {} (H2-);
\path[-] (C4) edge [] node[above] {} (H2-);
\path[-] (C4) edge [] node[above] {} (H1-);
\path[-] (C5) edge [] node[above] {} (H1-);
\path[-] (C5) edge [] node[above] {} (H3+);
\path[-] (C6) edge [] node[above] {} (H3+);
\path[-] (C6) edge [] node[above] {} (H2+);
\path[-] (C1) edge [dashed] node[above] {} (1);
\path[-] (C2) edge [dashed] node[above] {} (1);
\path[-] (C3) edge [dashed] node[above] {} (1);
\path[-] (C4) edge [dashed] node[above] {} (1);
\path[-] (C5) edge [dashed] node[above] {} (1);
\path[-] (C6) edge [dashed] node[above] {} (1);
\end{tikzpicture}
\end{center}
The $\mathbf{a}\mathbf{b}$-index of $\Sigma\cup\{\hat 0\}$ can be computed as
\[
\mathbf{a}^3 + 5\mathbf{a}\mathbf{b}\mathbf{a} + 5 \mathbf{a}^2\mathbf{b} + \mathbf{a}\mathbf{b}^2 = \mathbf{a}(\mathbf{a}^2 + 5\mathbf{b}\mathbf{a} + 5\mathbf{a}\mathbf{b} + \mathbf{b}^2).
\]
As seen above, this equals $\mathbf{a}\cdot\Psi_{\sf pull}(\mathcal{L};\mathbf{a},\mathbf{b})$.
\end{example}
\subsection{Main results}
\label{sec:mainresults}
All main results of this paper concern $R$-labeled graded posets.
A function~$\lambda$ from the set of cover relations $X\lessdot Y$ in~$P$ into the positive integers is an \Dfn{$R$-labeling} of $P$ if, for every interval $[X, Y]$ in~$P$, there is a unique
maximal chain $X=\mathcal{M}_i\lessdot \mathcal{M}_{i+1} \lessdot \cdots \lessdot \mathcal{M}_j=Y$ such that
\[
\lambda(\mathcal{M}_i,\mathcal{M}_{i+1}) \leq \lambda(\mathcal{M}_{i+1},\mathcal{M}_{i+2}) \leq \cdots \leq \lambda(\mathcal{M}_{j-1},\mathcal{M}_j).
\]
We say a poset $P$ is \Dfn{$R$-labeled} if it is finite, graded, and admits an
$R$-labeling.
Throughout this section, we consider $R$-labeled posets with a fixed $R$-labeling~$\lambda$.
Many interesting families of posets admit
$R$-labelings. Most importantly for the present context, geometric lattices,
{\textit{i.e.}}, lattices of flats for finite matroids, admit $R$-labelings;
see~\cite[Example 3.8]{bjorner-shellings}.
\medskip
The first is
a combinatorial statistic describing the coefficients of the extended
$\mathbf{a}\mathbf{b}$-index which witnesses their nonnegativity. This
generalizes~\cite[Corollary~7.2]{billera-ehrenborg-readdy} and also reproves it
using purely combinatorial arguments.
For a maximal chain~$\mathcal{M} =
\{\mathcal{M}_0\lessdot\mathcal{M}_1\lessdot\dots\lessdot\mathcal{M}_n\}$, define the monomial
${\sf u}(\mathcal{M}) = u_1\cdots u_n$ in $\mathbf{a},\mathbf{b}$ given by $u_1 = \mathbf{a}$
and for $i \in \{2,\dots,n\}$ by
\begin{equation}
\label{eq:statmon}
u_i = \begin{cases}
\mathbf{a} &\text{if } \lambda(\mathcal{M}_{i-2},\mathcal{M}_{i-1}) \leq \lambda(\mathcal{M}_{i-1},\mathcal{M}_{i})\,, \\
\mathbf{b} &\text{if } \lambda(\mathcal{M}_{i-2},\mathcal{M}_{i-1}) > \lambda(\mathcal{M}_{i-1},\mathcal{M}_{i})\,.
\end{cases}
\end{equation}
Now, let $E \subseteq \{1,\dots,n\}$, viewed as a subset of the cover relations
in the chain~$\mathcal{M}$. Define the monomial ${\sf u}(\mathcal{M},E) = v_1\dots v_n$ in
$\mathbf{a},\mathbf{b}$ to be obtained from ${\sf u}(\mathcal{M})$ by
\begin{itemize}
\item replacing all variables~$\mathbf{a}$ by $\mathbf{b}$ at positions~$i\in \{1,\dots,
n\}$ if $i \in E$ and
\item replacing all variables~$\mathbf{b}$ by $\mathbf{a}$ at positions~$i\in \{2,\dots,
n\}$ if $i-1 \in E$.
\end{itemize}
In symbols this means, for the given position~$i \in \{1,\dots,n\}$, that
\begin{equation*}
\begin{array}{ll}
v_i = \mathbf{a} &\text{ if } \begin{cases}
u_i = \mathbf{a}, \quad i\hspace*{21.5pt} \notin E\quad\text{or} \\
u_i = \mathbf{b}, \quad i-1 \in E\,,
\end{cases} \\ \\
v_i = \mathbf{b} & \text{ if } \begin{cases}
u_i = \mathbf{a}, \quad i\hspace*{21.5pt} \in E\quad\text{or} \\
u_i = \mathbf{b}, \quad i-1 \notin E\,.
\end{cases}
\end{array}
\end{equation*}
We have, in particular, ${\sf u}(\mathcal{M},\emptyset) = {\sf u}(\mathcal{M})$ and
\begin{equation}
\label{eq:firstindex}
v_1 = \begin{cases}
\mathbf{a} & \text{ if } 1 \notin E\,, \\
\mathbf{b} & \text{ if } 1 \in E\,.
\end{cases}
\end{equation}
\begin{thm}
\label{thm:combinatorial-interp}
Let~$P$ be an $R$-labeled poset of rank~$n$.
Then
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})
= \sum_{(\mathcal{M},E)} y^{\#E}\cdot{\sf u}(\mathcal{M},E)
\]
where the sum ranges over all maximal chains~$\mathcal{M}$ in~$P$ and all subsets~$E \subseteq \{1,\dots,n\}$.
\end{thm}
\begin{example}
\label{ex:face-poset3}
The poset from the previous examples admits the $R$-labeling given below on the left.
On the right, we collect the relevant data to compute the combinatorial description of the extended $\mathbf{a}\mathbf{b}$-index.
\begin{center}
\begin{tikzpicture}[scale=.7, baseline={([yshift=-.5ex]current bounding box.center)}]
\node[invisivertex] (1) at (2,4){$\hat 1$};
\node[invisivertex] (H1) at (0,2){$\alpha_1$};
\node[invisivertex] (H2) at (2,2){$\alpha_2$};
\node[invisivertex] (H3) at (4,2){$\alpha_3$};
\node[invisivertex] (0) at (2,0){$\hat 0$};
\node[edgelabel] at (.5,1){\tiny $1$};
\node[edgelabel] at (1.7,1){\tiny $2$};
\node[edgelabel] at (3.5,1){\tiny $3$};
\node[edgelabel] at (.5,3){\tiny $2$};
\node[edgelabel] at (1.7,3){\tiny $1$};
\node[edgelabel] at (3.5,3){\tiny $1$};
\draw (0) -- (H1) -- (1);
\draw (0) -- (H2) -- (1);
\draw (0) -- (H3) -- (1);
\end{tikzpicture}
\hspace*{2cm}
\begin{tabular}{|c|c|c|c|c|}
\hline
& & & &\\[-10pt]
$E$ & $y^{\#E}$
& $\hat 0 \lessdot \alpha_1 \lessdot \hat 1$
& $\hat 0 \lessdot \alpha_2 \lessdot \hat 1$ &
$\hat 0 \lessdot \alpha_3 \lessdot \hat 1$ \\[5pt]
\hline
& & & &\\[-10pt]
$\{\}$ & $1$ & $\mathbf{a}\aaa$ & $\mathbf{a}\mathbf{b}$ & $\mathbf{a}\mathbf{b}$ \\
$\{ 1 \}$ & $y$ & $\mathbf{b}\mathbf{a}$ & $\mathbf{b}\mathbf{a}$ & $\mathbf{b}\mathbf{a}$ \\
$\{ 2 \}$ & $y$ & $\mathbf{a}\mathbf{b}$ & $\mathbf{a}\mathbf{b}$ & $\mathbf{a}\mathbf{b}$ \\
$\{ 1,2 \}$ & $y^2$ & $\mathbf{b}\bbb$ & $\mathbf{b}\mathbf{a}$ & $\mathbf{b}\mathbf{a}$\\[5pt]
\hline
\end{tabular}
\end{center}
Then ${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};y,\mathbf{a},\mathbf{b})$ is obtained as
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};y,\mathbf{a},\mathbf{b}) = \mathbf{a}\aaa + (3y+2y^2)\mathbf{b}\mathbf{a} + (2 + 3y)\mathbf{a}\mathbf{b} + y^2\mathbf{b}\bbb\,,
\]
in agreement with our computation in \Cref{ex:face-poset2}.
\end{example}
\Cref{thm:combinatorial-interp} has many consequences, which we formulate in
six corollaries. The most important gives a substitution sending the
$\mathbf{a}\mathbf{b}$-index to the extended $\mathbf{a}\mathbf{b}$-index---meaning that the extended
$\mathbf{a}\mathbf{b}$-index is already encoded in the $\mathbf{a}\mathbf{b}$-index.
\begin{corollary}
\label{thm:refinement-of-ber}
For an $R$-labeled poset~$P$, we have
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) = \omega\big(\Psi(P;\mathbf{a},\mathbf{b})\big)
\]
where the substitution $\omega$ replaces all occurrences of $\mathbf{a}\mathbf{b}$ with
$\mathbf{a}\mathbf{b} + y\mathbf{b}\mathbf{a} + y\mathbf{a}\mathbf{b} + y^2\mathbf{b}\mathbf{a}$ and then simultaneously
replaces all remaining occurrences of $\mathbf{a}$ with $\mathbf{a}+y\mathbf{b}$ and $\mathbf{b}$ with
$\mathbf{b}+y\mathbf{a}$.
\end{corollary}
\begin{example}
We have seen that the poset~$\mathcal{L}$ from \Cref{ex:face-poset2} has $\mathbf{a}\mathbf{b}$-index
\[
\Psi(\mathcal{L};\mathbf{a},\mathbf{b}) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L},0,\mathbf{a},\mathbf{b}) = \mathbf{a}\aaa + 2\mathbf{a}\mathbf{b}\,.
\]
Applying $\omega$ gives
\[
\omega\big(\Psi(\mathcal{L};\mathbf{a},\mathbf{b})\big)
= (\mathbf{a}+y\mathbf{b})^2 + 2(\mathbf{a}\mathbf{b} + y\mathbf{b}\mathbf{a} + y\mathbf{a}\mathbf{b} + y^2\mathbf{b}\mathbf{a}) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};y,\mathbf{a},\mathbf{b})\,,
\]
which coincides with the extended $\mathbf{a}\mathbf{b}$-index we computed in \Cref{ex:face-poset3}.
\end{example}
One consequence of \Cref{thm:refinement-of-ber} is that the Poincaré polynomial
of $P$ is encoded in its $\mathbf{a}\mathbf{b}$-index. To see this, we define another
substitution~$\iota$, which simply deletes the first letter from every $\mathbf{a}\mathbf{b}$-monomial, so
$\iota(\mathbf{a}^3\mathbf{b}\mathbf{a} + (1+y)\mathbf{b}\mathbf{a}) = \mathbf{a}^2\mathbf{b}\mathbf{a} + (1+y)\mathbf{a}$ for example. Then
\begin{equation}
\mathcal{\sf Poin}(P;y) = [\mathbf{a}^{n-1}]\ \iota\big(\omega\big(\Psi(P;\mathbf{a},\mathbf{b})\big)\big)\,,
\label{eq:exabpoin}
\end{equation}
where $[{\sf m}]f \in \mathbb{Z}[y]$ is equal to the coefficient of the monomial ${\sf m}$ in the polynomial
$f \in \mathbb{Z}[y]\langle\mathbf{a},\mathbf{b}\rangle$. \Cref{thm:refinement-of-ber}
generalizes~\cite[Theorem~3.1]{billera-ehrenborg-readdy} relating the
$\mathbf{a}\mathbf{b}$-index of the lattice of flats of an oriented matroid with the
$\mathbf{a}\mathbf{b}$-index of its face poset. As a consequence, we see that
${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})$ is akin to a refinement of a $\mathbf{c}\mathbf{d}$-index. We
make this observation precise in the following corollary.
\begin{corollary}
\label{cor:cd-index}
For an $R$-labeled poset~$P$, there exists a polynomial $\Phi(P;
\mathbf{c}_1,\mathbf{c}_2,\mathbf{d})$ in noncommuting variables $\mathbf{c}_1,\mathbf{c}_2,\mathbf{d}$ such that
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P; y, \mathbf{a}, \mathbf{b}) = \Phi(P;\, \mathbf{a}+y\mathbf{b},\, \mathbf{b}+y\mathbf{a},\, \mathbf{a} \mathbf{b} + y \mathbf{b} \mathbf{a} + y \mathbf{a} \mathbf{b} + y^2 \mathbf{b} \mathbf{a}) .
\]
In particular, the pullback $\mathbf{a}\mathbf{b}$-index $\Psi_{\sf pull}(P;\mathbf{a},\mathbf{b})$ is a polynomial in noncommuting variables $\mathbf{c} = \mathbf{a}+\mathbf{b}$ and $2\mathbf{d} = 2(\mathbf{a}\mathbf{b}+\mathbf{b}\mathbf{a})$.
\end{corollary}
\Cref{cor:cd-index} gives a natural pullback $\mathbf{c}\mathbf{d}$-index for any $R$-labeled
poset which, in the case of the lattice of flats for an oriented matroids,
equals the $\mathbf{c}\mathbf{d}$-index of its face poset. This, in particular, gives an
analogue of the face poset $\mathbf{c}\mathbf{d}$-index for non-orientable matroids, which
do not have access to an actual face posets.
\begin{remark}[Oriented Interval Greedoids]
\label{rem:oig-version}
The same argument used for oriented matroids and their lattice of flats applies to \emph{oriented interval greedoids}, where the analogue of~\Cref{eq:supppre} is given in~\cite[Theorem~6.8]{saliola-thomas}.
Since the lattice of flats of an interval greedoid is a semimodular lattice~\cite[Theorem~8.8.7]{white}, it admits an $R$-labeling, see~\cite[Theorem~3.7]{bjorner-shellings}.
Applying~\Cref{thm:refinement-of-ber} and setting $y=1$ gives~\cite[Corollary~6.12]{saliola-thomas}.
\end{remark}
By examining the statistic from~\Cref{thm:combinatorial-interp}, we see that the
extended $\mathbf{a}\mathbf{b}$-index has symmetry among its coefficients. We encode this in
the following theorem, which generalizes the bivariate version
of~\cite[Theorem~A]{Maglione-Voll}.
\begin{corollary}
\label{thm:symmetry-of-coeffs}
Let~$P$ be an $R$-labeled poset of rank~$n$.
Let moreover~$\mathcal{M}$ be a maximal chain in~$P$ and let $E, E^c$ be complementary subsets of $\{1,\dots,n\} = E \cup E^c$.
Then ${\sf u}(\mathcal{M},E) = v_1\dots v_{n}$ and ${\sf u}(\mathcal{M},E^c) = v_1^c\dots v_{n}^c$ are complementary words, {\textit{i.e.}}, $\{v_i,v_i^c\} = \{\mathbf{a},\mathbf{b}\}$ for every $i \in \{1,\dots,n\}$.
In particular,
\[
[y^\ell {\sf m}]\ {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) = [y^{n-\ell} {\sf m}^c]\ {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})
\]
for complementary words~${\sf m}$ and~${\sf m}^c$ in $\mathbf{a},\mathbf{b}$ and for $\ell\in
\{0,\dots, n\}$.
\end{corollary}
We next turn toward the coarse flag Hilbert--Poincaré series introduced and
studied in~\cite{Maglione-Voll}. The numerator of this rational function is
defined in~\cite[Equation~(1.13)]{Maglione-Voll}, and we extend this definition
to graded posets via
\[
{\sf Num}(P;y,t) = \sum_{\mathcal C\text{ chain in } P\setminus\{\hat{0},\hat 1\}} \POIN[\{\hat 0\} \cup \mathcal C]{P}{y}\cdot t^{\#\mathcal C}(1-t)^{n-1-\#\mathcal C}\, \in \mathbb{Z}[y,t]\,.
\]
\begin{example}
Let~$\mathcal{L}$ be the poset from \Cref{ex:face-poset2}.
This poset has rank~$2$ and numerator polynomial
\[
{\sf Num}(P;y,t) = (1 + 3y + 2y^2)(1-t) + 3\cdot(1+y)^2t = 1 + 3y + 2y^2 + (2 + 3y + y^2)t\,.
\]
\end{example}
The following corollary generalizes the property in \Cref{eq:ab01} to extended
$\mathbf{a}\mathbf{b}$-indices and relates them to coarse flag Hilbert--Poincaré series.
Recall from \Cref{eq:exabpoin}, the function $\iota$ that deletes the first
letter from every $\mathbf{a}\mathbf{b}$-monomial.
\begin{corollary}
\label{cor:enab0}
For an $R$-labeled poset~$P$, we have
\[
\iota\big({\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})\big) = \sum_{\mathcal{C}\text{ chain in } P\setminus\{\hat 0,\hat 1\}} \mathcal{\sf Poin}_{\{\hat 0\} \cup \mathcal{C}}(P;y) \cdot \wt^-_\mathcal{C}(\mathbf{a},\mathbf{b})\ \in \mathbb{N}[y]\langle \mathbf{a},\mathbf{b}\rangle \,,
\]
where $\wt^-_\mathcal{C}(\mathbf{a},\mathbf{b}) = w_1\dots w_{n-1}$.
\end{corollary}
Specializing the equation from \Cref{cor:enab0}
via $\mathbf{a} \mapsto 1$ and $\mathbf{b} \mapsto t$
proves~\cite[Conjecture~E]{Maglione-Voll} and its generalization to $R$-labeled
posets. We collect this in the following corollary.
\begin{corollary}
\label{thm:maglione-voll-conjecture}
For an $R$-labeled poset~$P$, the coefficients of ${\sf Num}(P;y,t)$ are
nonnegative.
\end{corollary}
Together with \Cref{eq:exabpoin}, we obtain
\begin{equation}
\mathcal{\sf Poin}(P;y) = [t^0]\ {\sf Num}(P;y,t).
\label{eq:poinfromnum}
\end{equation}
The substitutions in the previous corollaries show that \Cref{thm:combinatorial-interp} also gives analogous combinatorial interpretations for the coefficients of $\iota\big({\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})\big)$ and of ${\sf Num}(P;y,t)$.
\begin{example}
Recall that $\iota$ is the function that deletes the first letter of every
$\mathbf{a}\mathbf{b}$-monomial. Thus, ignoring the first position in the computation in
\Cref{ex:face-poset3} yields
\[
\iota\big({\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};y,\mathbf{a},\mathbf{b})\big) = (1 + 3y + 2y^2)\mathbf{a} + (2 + 3y + y^2)\mathbf{b}\,.
\]
The substitution $\mathbf{a} \mapsto 1$ and $\mathbf{b} \mapsto t$ yields ${\sf Num}(\mathcal{L};y,t)$.
Since $n=2$, we have
\[
[\mathbf{a}^{n-1}]\ \iota\big({\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(\mathcal{L};y,\mathbf{a},\mathbf{b})\big) = [t^0]\ {\sf Num}(\mathcal{L};y,t) = \mathcal{\sf Poin}(\mathcal{L};y)\,,
\]
also in agreement with \Cref{eq:exabpoin}.
\end{example}
\begin{remark}[Geometric Semilattices]
Note that~\cite[Conjecture~E]{Maglione-Voll} concerns all hyperplane arrangements (central and affine).
While the intersection posets of central hyperplane arrangements are geometric lattices and, thus, admit $R$-labelings~\cite[Example 3.8]{bjorner-shellings}, the intersection posets of affine arrangements are part of a more general family called \emph{geometric semilattices}, first explicitly studied by Wachs and Walker in~\cite{wachs-walker}.
A theorem of Ziegler shows that if $\mathcal{L}$ is a geometric semilattice, then $\mathcal{L}\cup\{\hat{1}\}$ admits an $R$-labeling~\cite[Theorem 2.2]{ziegler-shellability}.
Thus~\Cref{thm:combinatorial-interp} holds for intersection posets of affine arrangements (now with a formal unique maximal element included, which is consistent with the formulation in~\cite[Conjecture~E]{Maglione-Voll}).
\end{remark}
The specialization of ${\sf Num}(P;y,t)$ at $y=1$ was studied further for matroids
and oriented matroids by the second author and Kühne in~\cite{Kuhne-Maglione},
who showed ${\sf Num}(P;1,t)$ is the sum of $h$-polynomials of simplicial complexes
related to the chambers if $P$ is the lattice of flats of a real central
hyperplane arrangement. The following corollary gives a lower bound for the
coefficients of ${\sf Num}(P;1,t)$, whose proof also addresses the conjectured lower
bound in~\cite[Conjecture~1.4]{Kuhne-Maglione} generalized to the setting of
$R$-labeled posets---originally stated for geometric lattices.
\begin{corollary}
\label{thm:lower-bound-on-coefficients}
Let~$P$ be an $R$-labeled poset of rank~$n$.
The coefficients of ${\sf Num}(P;1,t)$ are bounded below by
\[
[t^k]\ {\sf Num}(P;1,t) \geq \binom{n-1}{k}\cdot \POIN{P}{1}\,.
\]
\end{corollary}
\section{The combinatorial description of the extended $\mathbf{a}\mathbf{b}$-index}
\label{sec:proof-of-combinatorial-interp}
In this section we prove~\Cref{thm:combinatorial-interp} as well as~\Cref{thm:symmetry-of-coeffs,,thm:maglione-voll-conjecture,,thm:lower-bound-on-coefficients}.
First we use a property of $R$-labelings to rewrite the chain Poincaré polynomial as a sum over maximal chains (\Cref{lem:poincare-expansion}).
Then, in~\Cref{sec:inclusion-exclusion}, we use an inclusion-exclusion argument to give a combinatorial description of the coefficients of the extended $\mathbf{a}\mathbf{b}$-index.
The proof of~\Cref{thm:combinatorial-interp} and the corollaries are given in \Cref{sec:reinterp-coeffs} and come from reformulating this combinatorial description.
\medskip
Throughout this section, we fix an $R$-labeled poset~$P$ of rank~$n$ together
with an $R$-label~$\lambda$. We start with recalling the following rewriting of
the Möbius function in terms of maximal chains.
\begin{lemma}[\!\!{\cite[Corollary 2.3]{bjorner-garsia-stanley}}]
\label{lem:moebiuscount}
Let $X,Y \in P$ with $X \leq Y$.
Then
\[
\big|\mu(X,Y)\big|
=
\#
\big\{
X = \mathcal{C}_i \lessdot \mathcal{C}_{i+1} \lessdot \dots \lessdot \mathcal{C}_{j} = Y
\mid
\lambda(\mathcal{C}_{i},\mathcal{C}_{i+1}) > \dots > \lambda(\mathcal{C}_{j-1},\mathcal{C}_{j})
\big\}\,.
\]
\end{lemma}
We distinguish multisets from sets by using two curly brackets, and a multichain is a totally ordered multiset.
For $k\geq 0$ and a chain $\mathcal{C} = \{\mathcal{C}_1 < \cdots < \mathcal{C}_k\}$ in $P$, say that a multichain $\mathcal{D}=\multichain{\mathcal{D}_1 \leq \cdots \leq \mathcal{D}_k}$ in~$P$ \Dfn{interlaces}~$\mathcal{C}$ if
\begin{equation}\label{eqn:interlaced}
\mathcal{C}_1 \leq \mathcal{D}_1 \leq \mathcal{C}_2 \leq \mathcal{D}_2 \leq \dots \leq \mathcal{C}_k \leq \mathcal{D}_k\,.
\end{equation}
Note that we allow~$\mathcal{D}$ to be a multichain only because we could have $\mathcal{D}_{i-1} = \mathcal{C}_i = \mathcal{D}_i$ for some position~$i$.
We call such a pair $(\mathcal{C},\mathcal{D})$ an \Dfn{interlacing pair}.
For a chain $\mathcal{C}$ in~$P$, let
\[
\mathsf{IL}(\mathcal{C}) = \{ \mathcal{D} ~|~ (\mathcal{C}, \mathcal{D}) \text{ is an interlacing pair}\},
\]
the set of multichains interlacing~$\mathcal{C}$.
For $(\mathcal{C},\mathcal{D})$ interlacing, we write
\begin{align*}
{\sf IRank}(\mathcal{C},\mathcal{D})
&= \big\{{\sf rank}(\mathcal{C}_1)+1,\dots,{\sf rank}(\mathcal{D}_1) \big\} \cup \dots \cup \big\{{\sf rank}(\mathcal{C}_k)+1,\dots,{\sf rank}(\mathcal{D}_k) \big\} \\
&= \big\{ r \in \{1,\dots,n\} \ \big\vert\ {\sf rank}(\mathcal{C}_i) < r \leq {\sf rank}(\mathcal{D}_i) \text{ for some position } i \big\}\,,
\end{align*}
and we denote the cardinality of this set by
\[
{\sf irank}(\mathcal{C},\mathcal{D}) = \# {\sf IRank}(\mathcal{C},\mathcal{D}) = \sum_{i=1}^k{\sf rank}(\mathcal{D}_i) - \sum_{i=1}^k {\sf rank}(\mathcal{C}_i)\,.
\]
A maximal chain $\mathcal{M} = \{ \mathcal{M}_0 \lessdot \mathcal{M}_1 \lessdot \dots \lessdot \mathcal{M}_n\}$ in~$P$ \Dfn{decreases along} the interval $[\mathcal{M}_i,\mathcal{M}_j]$ if
\[
\lambda(\mathcal{M}_i, \mathcal{M}_{i+1}) > \dots > \lambda(\mathcal{M}_{j-1}, \mathcal{M}_{j})\,.
\]
Similarly, we say that $\mathcal{M}$ \Dfn{weakly increases along} $[\mathcal{M}_i,\mathcal{M}_j]$ if
\[
\lambda(\mathcal{M}_i, \mathcal{M}_{i+1}) \leq \dots \leq \lambda(\mathcal{M}_{j-1}, \mathcal{M}_{j})\,.
\]
We say $\mathcal{M}$ is \Dfn{decreasing} (resp.\ \Dfn{weakly increasing}) if $\mathcal{M}$ is decreasing (resp.\ weakly increasing) along $[\mathcal{M}_0, \mathcal{M}_n] = P$. Let ${\sf IncDec}(\mathcal{C},\mathcal{D})$ be the set of maximal chains~$\mathcal{M}$ in~$P$ that refine the chain in~\Cref{eqn:interlaced} which
\begin{itemize}
\item decrease along all the intervals of the form $[\mathcal{C}_i,\mathcal{D}_i]$, and
\item weakly increase along all the intervals of the forms $[\hat{0}, \mathcal{C}_1]$, $[\mathcal{D}_i,\mathcal{C}_{i+1}]$ and $[\mathcal{D}_k, \hat{1}]$.
\end{itemize}
We say that such a maximal chain $\mathcal{M} \in {\sf IncDec}(\mathcal{C},\mathcal{D})$ is \Dfn{increasing-decreasing} with respect to $(\mathcal{C},\mathcal{D})$.
\begin{proposition}
\label{lem:poincare-expansion}
Let~$P$ be an $R$-labeled poset, and let~$\mathcal{C}$ be a chain in~$P$.
Then
\[
\POIN[\mathcal{C}]{P}{y} = \sum_{\ell \geq 0} \left( y^\ell\cdot \sum_{\substack{\mathcal{D} \in \mathsf{IL}(\mathcal{C}) \\ {\sf irank}(\mathcal{C},\mathcal{D})=\ell}} \# {\sf IncDec}(\mathcal{C},\mathcal{D})\right).
\]
\end{proposition}
\begin{proof}
We first rewrite the chain Poincaré polynomial, and then invoke~\Cref{lem:moebiuscount}. For $\mathcal{C} = \{\mathcal{C}_1 < \cdots < \mathcal{C}_k\}$, we have
\begin{align*}
\POIN[\mathcal{C}]{P}{y} & = \prod_{i=1}^{k} \POIN{[\mathcal{C}_i,\mathcal{C}_{i+1}]}{y}
= \prod_{i=1}^{k} \left(\sum_{X \in [\mathcal{C}_i,\mathcal{C}_{i+1}]} |\mu(\mathcal{C}_i,X)| \cdot y^{{\sf rank}(X)-{\sf rank}(\mathcal{C}_i)}\right)\\
& = \sum_{\ell\geq 0} \left(y^\ell\cdot \sum_{\substack{\mathcal{D} \in \mathsf{IL}(\mathcal{C}) \\{\sf irank}(\mathcal{C},\mathcal{D}) = \ell}}\prod_{i=1}^{k} |\mu(\mathcal{C}_i,\mathcal{D}_i)|\right)\,.
\end{align*}
We may interpret $|\mu(\mathcal{C}_i,\mathcal{D}_i)|$ via maximal chains from $\mathcal{C}_i$ to $\mathcal{D}_i$ using~\Cref{lem:moebiuscount}.
From the definition of an $R$-label, each interval $[\mathcal{D}_i,\mathcal{C}_{i+1}]$ has a unique weakly increasing chain.
By gluing together weakly increasing and decreasing pieces, we finally obtain, for $\mathcal{D}\in \mathsf{IL}(\mathcal{C})$,
\[
\prod_{i=1}^{k} |\mu(\mathcal{C}_i,\mathcal{D}_i)| = \# {\sf IncDec}(\mathcal{C},\mathcal{D}) . \qedhere
\]
\end{proof}
\subsection{An inclusion-exclusion construction}
\label{sec:inclusion-exclusion}
The inclusion-exclusion construction starts with a set ${\sf A}_\ell(S)$ associated to a subset $S\subseteq \{0,\dots,n-1\}$ and a parameter~$\ell \geq 0$.
It is given by
\[
{\sf A}_\ell(S) =
\left\{
(\mathcal{C},\mathcal{D},\mathcal{M})\
\middle\vert\
\begin{matrix}
(\mathcal{C},\mathcal{D}) \text{ interlacing}\\
{\sf Rank}(\mathcal{C}) = S,\ {\sf irank}(\mathcal{C},\mathcal{D}) = \ell\\
\mathcal{M} \in {\sf IncDec}(\mathcal{C},\mathcal{D})
\end{matrix}\
\right\}.
\]
In other words, we consider triples $(\mathcal{C},\mathcal{D},\mathcal{M})$ carrying the following data:
a chain~$\mathcal{C}$ in~$P$ with the prescribed ranks~${\sf Rank}(\mathcal{C}) = S$,
a multichain $\mathcal{D}$ interlacing~$\mathcal{C}$ with ${\sf irank}(\mathcal{C},\mathcal{D}) = \ell$, and
a maximal chain $\mathcal{M}$ that is increasing-decreasing with respect to the pair $(\mathcal{C},\mathcal{D})$.
\begin{lemma}
\label{lem:AIncDec}
For all $S \subseteq \{0, \dots, n-1\}$ and $\ell\geq 0$, we have
\[
\# {\sf A}_\ell(S) = \sum_{(\mathcal{C},\mathcal{D})} \# {\sf IncDec}(\mathcal{C},\mathcal{D}),
\]
where the sum ranges over all interlacing pairs $(\mathcal{C},\mathcal{D})$ with ${\sf Rank}(\mathcal{C})\! =\! S$ and ${\sf irank}(\mathcal{C},\mathcal{D})\! =\! \ell$.
\end{lemma}
\begin{proof}
For each interlacing pair $(\mathcal{C}, \mathcal{D})$ with ${\sf Rank}(\mathcal{C}) = S$ and
${\sf irank}(\mathcal{C},\mathcal{D}) = \ell$, there is an embedding of ${\sf IncDec}(\mathcal{C},\mathcal{D})$ into
${\sf A}_\ell(S)$ via $\mathcal{M} \mapsto (\mathcal{C}, \mathcal{D}, \mathcal{M})$, and the images of
${\sf IncDec}(\mathcal{C},\mathcal{D})$ and ${\sf IncDec}(\mathcal{C}',\mathcal{D}')$ are disjoint if and only if
$(\mathcal{C},\mathcal{D})\neq (\mathcal{C}',\mathcal{D}')$. Moreover, every $(\mathcal{C}, \mathcal{D}, \mathcal{M})\in
{\sf A}_{\ell}(S)$ is an image of such an embedding.
\end{proof}
\begin{remark}
Given a graded poset~$P$ of rank~$n$ and a subset $S \subseteq \{0,\dots,n-1\}$, let $\alpha(S)$ denote the set of chains in~$P$ with prescribed ranks $S$, and let $\beta(T)$ be the signed sum $\sum_{S \subseteq T} (-1)^{\#(S\setminus T)} \alpha(S)$.
The $\mathbf{a}\mathbf{b}$-index of~$P$ can then be expressed as
\[
\Psi(P;\mathbf{a},\mathbf{b}) = \sum_{T\subseteq \{0,\dots,n-1\}} \beta(T) \cdot {\sf u}_T
= \sum_{T\subseteq \{0,\dots,n-1\}} \left(\sum_{S \subseteq T} (-1)^{\#(S\setminus T)} \alpha(S)\right)\cdot {\sf u}_T,
\]
where ${\sf u}_T = w_0\dots w_{n-1}$ with $w_i = \mathbf{b}$ if $i \in T$ and $w_i = \mathbf{a}$ if $i \notin T$.
In this section we define an analogue of~$\alpha(S)$ for the extended $\mathbf{a}\mathbf{b}$-index, and use an inclusion-exclusion construction to obtain the analogue of~$\beta(S)$.
This describes, in particular, how we use $R$-labelings since our analogue of~$\alpha(S)$ uses this machinery to interpret the Möbius function via chains as described in \Cref{lem:moebiuscount}.
\end{remark}
Define the embedding $\varphi_{S, T} : {\sf A}_\ell(S) \hookrightarrow {\sf A}_\ell(T)$ for $S \subseteq T \subseteq \{0,\dots,n-1\}$ given by
\[
\varphi_{S,T}(\mathcal{C},\mathcal{D},\mathcal{M}) = \big(\mathcal{C}\cup\{\mathcal{M}_r \mid r \in T \setminus S\},\ \mathcal{D}\cup\{\mathcal{M}_r \mid r\in T \setminus S\},\ \mathcal{M} \big)\,.
\]
It is immediate that $\mathcal{M} \in {\sf IncDec}(\mathcal{C},\mathcal{D})$ implies $\mathcal{M} \in {\sf IncDec}(\mathcal{C}\cup\{\mathcal{M}_r\}, \mathcal{D}\cup\{\mathcal{M}_r\})$ for $r \notin {\sf Rank}(S)$, so this embedding is well-defined.
These maps are compatible with one another in the sense that $\varphi_{T,U} \circ \varphi_{S,T} = \varphi_{S,U}$ for $S \subseteq T \subseteq U$.
For $T\subseteq \{0,\dots,n-1\}$, this allows us to define the set
\[
{\sf B}_\ell(T) = {\sf A}_\ell(T) \big\backslash \bigcup_{S\subsetneq T} \varphi_{S,T}\big({\sf A}_\ell(S)\big).
\]
\begin{proposition}
\label{lem:A-B-expansion}
For $T\subseteq \{0,\dots,n-1\}$ and $\ell\geq 0$, we have
\[
\# {\sf B}_\ell(T)
= \sum_{S \subseteq T} (-1)^{\#T - \#S}
\sum_{(\mathcal{C},\mathcal{D})} \# {\sf IncDec}(\mathcal{C},\mathcal{D})\,,
\]
where the inner sum ranges over all interlacing pairs $(\mathcal{C},\mathcal{D})$ in~$P$ for which ${\sf Rank}(\mathcal{C}) = S$ and ${\sf irank}(\mathcal{C},\mathcal{D}) = \ell$.
\end{proposition}
\begin{proof}
We have
\[
\#{\sf B}_\ell(T) = \# {\sf A}_\ell(T) - \# \bigcup_{S\subsetneq T} \varphi_{S,T}\big({\sf A}_\ell(S)\big)
= \sum_{S \subseteq T} (-1)^{\#T - \#S} \# {\sf A}_\ell(S)\,,
\]
where the second equality follows from the principle of inclusion-exclusion.
The statement follows with \Cref{lem:AIncDec}.
\end{proof}
We are now ready to give a first combinatorial description for the coefficients of the extended $\mathbf{a}\mathbf{b}$-index, which are given by the cardinalities of the sets ${\sf B}_\ell(T)$.
\begin{thm}
\label{thm:toppiecount}
For all $T\subseteq \{0, \dots, n-1\}$ and $\ell\geq 0$,
\[
[y^\ell\cdot {\sf u}_T]\ {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) = \# {\sf B}_\ell(T)\,,
\]
where ${\sf u}_T = w_0\dots w_{n-1}$ with $w_i = \mathbf{b}$ if $i \in T$ and $w_i = \mathbf{a}$ if $i \notin T$.
\end{thm}
\begin{proof}
Rewriting the weight function in the definition of ${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})$ gives
\begin{align*}
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) & = \sum_{\mathcal{C} \text{ chain in }P\setminus \{\hat{1}\}} \POIN[\mathcal{C}]{P}{y} \cdot {\sf wt}_\mathcal{C}(\mathbf{a},\mathbf{b})\\
= & \sum_{\mathcal{C} \text{ chain in }P\setminus \{\hat{1}\}}~\sum_{\substack{T \subseteq \{0,\dots,n-1\} \\ {\sf Rank}(\mathcal{C}) \subseteq T}} (-1)^{\#T\setminus{\sf Rank}(\mathcal{C})} \POIN[\mathcal{C}]{P}{y}\cdot {\sf u}_T.
\end{align*}
By specifying the ranks of these chains and rearranging the order of summation, we obtain
\begin{align*}
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})
= & \sum_{S\subseteq \{0,\dots,n-1\}}~\sum_{\substack{\mathcal{C} \text{ chain in }P\\{\sf Rank}(\mathcal{C}) = S}}~\sum_{\substack{T \subseteq \{0,\dots,n-1\} \\ S \subseteq T}} (-1)^{\#T\setminus S} \POIN[\mathcal{C}]{P}{y} \cdot {\sf u}_T\\
= & \sum_{S \subseteq T \subseteq \{0,\dots,n-1\}}~\sum_{\substack{\mathcal{C} \text{ chain in }P\\{\sf Rank}(\mathcal{C}) = S}} (-1)^{\# T\setminus S} \POIN[\mathcal{C}]{P}{y}\cdot {\sf u}_T\,.
\end{align*}
Applying \Cref{lem:poincare-expansion,,lem:A-B-expansion} yields
\begin{align*}
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) &= \sum_{\ell\geq 0} y^{\ell} \sum_{T\subseteq \{0,\dots,n-1\}} \left(\sum_{S \subseteq T} (-1)^{\#(T\setminus S)} \sum_{(\mathcal{C},\mathcal{D})} \#{\sf IncDec}(\mathcal{C},\mathcal{D}) \right) \cdot {\sf u}_T
\\ &=
\sum_{\ell\geq 0} y^{\ell} \sum_{T\subseteq \{0,\dots,n-1\}}\#{\sf B}_\ell(T) \cdot {\sf u}_T . \qedhere
\end{align*}
\end{proof}
\begin{example}
Below we compute ${\sf A}_\ell(T)$ and ${\sf B}_\ell(T)$ for the poset $\mathcal{L}$ from~\Cref{ex:face-poset2}.
We record a triple $(\mathcal{C},\mathcal{D},\mathcal{M}) \in {\sf A}_\ell(T)$ as follows:
a maximal chain~$\mathcal{M} = \{\hat 0 \lessdot \alpha_i \lessdot \hat 1\}$ is recorded by~$\alpha_i$;~$\mathcal{C}$ is given by~$\mathcal{M}$ and the property ${\sf Rank}(\mathcal{C}) = T$; and~$\mathcal{D}$ is finally given by ${\sf IRank}(\mathcal{C},\mathcal{D}) = E \subseteq \{1,2\}$.
\begin{center}
\begin{tabular}{|c|c||c|c|c|}
\hline
& & & & \\[-10pt]
T & $\ell$ & $E$ & ${\sf A}_\ell(T)$ & ${\sf B}_\ell(T)$\\[2pt]
\hline
& & & & \\[-10pt]
$\emptyset$ & $0$ & $\emptyset$ & $\alpha_1$ & $\alpha_1$\\
$\{0\}$ & $0$ & $\emptyset$ & $\alpha_1$ & $\emptyset$ \\
$\{1\}$ & $0$ & $\emptyset$ & $\alpha_1,\alpha_2,\alpha_3$ & $\alpha_2,\alpha_3$ \\
$\{ 0,1 \}$ & $0$ & $\emptyset$ & $\alpha_1,\alpha_2,\alpha_3$ & $\emptyset$ \\
\hline
& & & \\[-10pt]
$\emptyset$ & $1$ & - & $\emptyset$ & $\emptyset$\\
$\{0\}$ & $1$ & $\{1\}$ & $\alpha_1,\alpha_2,\alpha_3$ & $\alpha_1,\alpha_2,\alpha_3$ \\
$\{1\}$ & $1$ & $\{2\}$ & $\alpha_1,\alpha_2,\alpha_3$ & $\alpha_2,\alpha_3$ \\
$\{ 0,1 \}$ & $1$ & $\{1\},\{2\}$ & $\alpha_1,\alpha_2,\alpha_3$ & $\emptyset$ \\
\hline
& & & \\[-10pt]
$\emptyset$ & $2$ & - & $\emptyset$ & $\emptyset$\\
$\{0\}$ & $2$ & $\{1,2\}$ & $\alpha_2,\alpha_3$ & $\alpha_2,\alpha_3$ \\
$\{1\}$ & $2$ & - & $\emptyset$ & $\emptyset$ \\
$\{ 0,1 \}$ & $2$ & $\{1,2\}$ & $\alpha_1,\alpha_2,\alpha_3$ & $\alpha_1$ \\
\hline
\end{tabular}
\end{center}
\end{example}
\subsection{Reinterpreting the coefficients}
\label{sec:reinterp-coeffs}
Next we reinterpret ${\sf B}_\ell(T)$ in order to prove \Cref{thm:combinatorial-interp}.
We show, for a given maximal chain~$\mathcal{M}$ in~$P$ and a set~$E\subseteq \{1,\dots,n\}$, that there is a unique interlacing pair~$(\mathcal{C},\mathcal{D})$ and a unique set~$T\subseteq\{0,\dots,n-1\}$ such that
\begin{equation}
{\sf IRank}(\mathcal{C},\mathcal{D}) = E\quad \text{and}\quad (\mathcal{C},\mathcal{D},\mathcal{M}) \in {\sf B}_{\#E}(T)\,.
\label{eq:uniquecond}
\end{equation}
For a set $E\subseteq\{1,\dots,n\}$, let $I_E,J_E \subseteq \{1,\dots,n\}$ denote the set of ranks where intervals of consecutive elements outside and inside of~$E$ end, respectively.
In symbols,
\[
I_E = \big\{i \in \{1,\dots,n\} \mid i \not\in E, i+1\in E \big\}
\quad\text{and}\quad
J_E = \big\{i \in \{1,\dots,n\} \mid i \in E, i+1\not\in E \big\}\,.
\]
The next lemma now follows immediately from unpacking definitions.
\begin{lemma}\label{lem:IRank-characterization}
Let $E \subseteq \{1,\dots,n\}$.
An interlacing pair $(\mathcal{C},\mathcal{D})$ satisfies ${\sf IRank}(\mathcal{C},\mathcal{D}) = E$ if and only if
\[
I_E = {\sf Rank}(\mathcal{C}) \setminus \big({\sf Rank}(\mathcal{C})\cap{\sf Rank}(\mathcal{D})\big), \quad J_E = {\sf Rank}(\mathcal{D}) \setminus \big({\sf Rank}(\mathcal{C})\cap{\sf Rank}(\mathcal{D})\big)\,.
\]
\end{lemma}
For a set~$E\subseteq \{1,\dots,n\}$ and a subset $R\subseteq \{0,\dots,n\} \setminus I_E$, set
\begin{equation}\label{eqn:C_R-D_R}
\mathcal{C}_R = \{ \mathcal{M}_i \mid i \in I_E \cup R\}\qquad \text{and}\qquad
\mathcal{D}_R = \multichain{\mathcal{M}_i \mid i \in J_E \sqcup R}\,,
\end{equation}
where $J_E \sqcup R$ is a \emph{multiset} union. For a maximal chain $\mathcal{M}$ in
$P$, define
\[
\mathsf{IL}_{\mathcal{M},E} = \big\{ (\mathcal{C},\mathcal{D}) \mid \mathcal{D} \in \mathsf{IL}(\mathcal{C}),\ {\sf IRank}(\mathcal{C},\mathcal{D}) = E,\ \mathcal{M} \in {\sf IncDec}(\mathcal{C}, \mathcal{D}) \big\}\, .
\]
Recall from \Cref{sec:inclusion-exclusion} that the set ${\sf A}_\ell(S)$ consists of triples $(\mathcal{C},\mathcal{D},\mathcal{M})$ for interlacing pairs $(\mathcal{C},\mathcal{D})$ such that ${\sf Rank}(\mathcal{C}) = S$ and~$\mathcal{M}$ is increasing-decreasing with respect to $(\mathcal{C},\mathcal{D})$.
Our next step is to use $\mathcal{C}_R$ and $\mathcal{D}_R$ to rewrite ${\sf A}_\ell(S)$.
To that end, we identify the subsets $R \subseteq\{0,\dots,n\}$ for which a given maximal chain $\mathcal{M}$ is increasing-decreasing with respect to $(\mathcal{C}_R,\mathcal{D}_R)$.
\begin{lemma}
\label{prop:Tdescription}
Let~$\mathcal{M}$ be a maximal chain, and let $E \subseteq \{1,\dots,n\}$.
Given the monomial ${\sf u}(\mathcal{M}) = u_1 \cdots u_n$, let $\TT{\mathcal{M}}{E}$ be the set of
indices $i \in \{0,\dots,n-1\}$ given by
\[
\TT{\mathcal{M}}{E} = \big\{ i \mid i, i+1 \in E,\ u_{i+1} = \mathbf{a}\big\} \cup \big\{ i \mid i, i+1 \notin E,\ u_{i+1} = \mathbf{b}\big\}\,.
\]
Then we have
\[
\mathsf{IL}_{\mathcal{M},E} = \big\{ (\mathcal{C}_R,\mathcal{D}_R) \mid \TT{\mathcal{M}}{E} \subseteq R \subseteq \{0,\dots,n\} \setminus I_E \big\}\,.
\]
\end{lemma}
\begin{proof}
By \Cref{lem:IRank-characterization}, an interlacing pair $(\mathcal{C}, \mathcal{D})$
satisfies ${\sf IRank}(\mathcal{C}_R, \mathcal{D}_R) = E$ if and only if there exists $R\subseteq
\{0,\dots, n\}\setminus I_E$ such that $(\mathcal{C}, \mathcal{D})= (\mathcal{C}_R, \mathcal{D}_R)$. Thus, it
suffices to identify those sets~$R$ for which the given maximal chain~$\mathcal{M}$ is
increasing-decreasing with respect to~$(\mathcal{C}_R,\mathcal{D}_R)$. That is, $\lambda$
decreases along all intervals of $\mathcal{M}$ of the form $[\mathcal{C}_i,\mathcal{D}_i]$ and weakly
increases along intervals of the form $[\hat{0},\mathcal{C}_1]$, $[\mathcal{D}_i,\mathcal{C}_i]$ and
$[\mathcal{D}_i,\hat{1}]$, where $\mathcal{C}=\mathcal{C}_R$ and $\mathcal{D}=\mathcal{D}_R$. Recall that
$u_{i+1}=\mathbf{a}$ if $\lambda(\mathcal{M}_{i-1}, \mathcal{M}_{i})\leq \lambda(\mathcal{M}_{i},
\mathcal{M}_{i+1})$ and $u_{i+1}=\mathbf{b}$ otherwise. Let $i\in \TT{\mathcal{M}}{E}$, and assume
$i\notin R$.
If $i,i+1 \in E$, then $\lambda(\mathcal{M}_{i-1}, \mathcal{M}_{i})\leq
\lambda(\mathcal{M}_{i}, \mathcal{M}_{i+1})$, so $\mathcal{M}$ is not decreasing along
$[\mathcal{C}_i,\mathcal{D}_i]$.
The case where $i,i+1 \notin E$ is analogous.
Hence,~$\mathcal{M}\in {\sf IncDec}(\mathcal{C}_R,\mathcal{D}_R)$ if and only if $\TT{\mathcal{M}}{E} \subseteq R$,
so the lemma follows.
\end{proof}
\begin{example}
Recall the poset $\mathcal{L}$ from \Cref{ex:face-poset2}. We give, for each maximal
chain in $\mathcal{L}$ and each subset of $\{0,1,2\}$, the unique pair of interlacing
chains satisfying \Cref{eq:uniquecond}.
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
& & &\\[-10pt]
& $\hat 0 \underset{\color{magenta}1}{\lessdot} \alpha_1 \underset{\color{magenta}2}{\lessdot} \hat 1$
& $\hat 0 \underset{\color{magenta}2}{\lessdot} \alpha_2 \underset{\color{magenta}1}{\lessdot} \hat 1$ &
$\hat 0 \underset{\color{magenta}3}{\lessdot} \alpha_3 \underset{\color{magenta}1}{\lessdot} \hat 1$ \\[5pt]
\hline
& & & \\[-10pt]
$\{\}$ & $\{\},\{\}$ & $\{\alpha_2\},\{\alpha_2\}$ & $\{\alpha_3\},\{\alpha_3\}$ \\
$\{ 0 \}$ & $\{\hat{0}\},\{\alpha_1\}$ & $\{\hat{0}\},\{\alpha_2\}$ & $\{\hat{0}\},\{\alpha_3\}$ \\
$\{ 1 \}$ & $\{\alpha_1\},\{\hat{1}\}$ & $\{\alpha_2\},\{\hat{1}\}$ & $\{\alpha_3\},\{\hat{1}\}$ \\
$\{ 0,1 \}$ & $\{\hat{0}<\alpha_1\},\{\alpha_1<\hat{1}\}$ & $\{\hat{0}\},\{\hat{1}\}$ & $\{\hat{0}\},\{\hat{1}\}$ \\
\hline
\end{tabular}
\end{center}
\end{example}
The next step is to translate \Cref{prop:Tdescription} into a statement about ${\sf A}_\ell(S)$, which we then use to simplify the description of ${\sf B}_\ell(T)$.
\begin{proposition}
\label{prop:PIEdescription}
Let $S\subseteq\{0,\dots,n-1\}$ and $\ell\geq 0$.
We have the following decomposition of ${\sf A}_\ell(S)$ into disjoint subsets
\[
{\sf A}_\ell(S) = \bigcup_{(\mathcal{M},E)} \big\{ (\mathcal{C},\mathcal{D},\mathcal{M}) \mid (\mathcal{C},\mathcal{D}) \in \mathsf{IL}_{\mathcal{M},E} \text{ and } {\sf Rank}(\mathcal{C}) = S \big\}\,,
\]
where the disjoint union ranges over all maximal chains~$\mathcal{M}$ and all subsets $E \subseteq \{1,\dots,n\}$ of cardinality $\#E = \ell$.
Moreover,
\[
{\sf B}_{\ell}(T) = \left\{ \left(\mathcal{C}_{\TT{\mathcal{M}}{E}},\mathcal{D}_{\TT{\mathcal{M}}{E}},\mathcal{M}\right) ~\middle|~ \begin{array}{c} \mathcal{M} \text{ maximal chain},\ E \subseteq \{1,\dots,n\}, \\ \#E = \ell,\ {\sf Rank}(\mathcal{C}_{\TT{\mathcal{M}}{E}}) = T \end{array} \right\}\,.
\]
\end{proposition}
\begin{proof}
Let~$\mathcal{M}$ be a maximal chain, $E\subseteq\{1,\dots,n\}$ of cardinality $\# E = \ell$, and $S \subseteq \{0,\dots,n-1\}$.
We then write
\[
\mathsf{IL}^{\ast}_{\mathcal{M},E}(S) = \big\{ (\mathcal{C},\mathcal{D},\mathcal{M}) \mid (\mathcal{C},\mathcal{D}) \in \mathsf{IL}_{\mathcal{M},E} \text{ and } {\sf Rank}(\mathcal{C}) = S \big\}\,,
\]
and observe that both $\mathsf{IL}^{\ast}_{\mathcal{M},E}(S) \subseteq {\sf A}_\ell(S)$ and $\mathsf{IL}^{\ast}_{\mathcal{M},E}(S) \cap \mathsf{IL}^{\ast}_{\mathcal{M},E'}(S) = \emptyset$ for different subsets $E$ and $E'$ of $\{1,\dots,n\}$.
Moreover, for $(\mathcal{C},\mathcal{D},\mathcal{M})\in {\sf A}_\ell(S)$ with $E = {\sf IRank}(\mathcal{C},\mathcal{D})$, we have that $(\mathcal{C},\mathcal{D},\mathcal{M}) \in \mathsf{IL}^{\ast}_{\mathcal{M},E}$.
This proves the first claim.
The set ${\sf B}_\ell(T)$ comprises the elements in ${\sf A}_\ell(T)$ that do not appear in any ${\sf A}_\ell(S)$ for $S\subsetneq T$.
From \Cref{prop:Tdescription}, it follows that
\[
\mathsf{IL}^{\ast}_{\mathcal{M},E}(S) = \big\{ (\mathcal{C}_R,\mathcal{D}_R,\mathcal{M}) \mid \TT{\mathcal{M}}{E} \subseteq R \subseteq \{0,\dots,n\} \setminus I_E \text{ and } {\sf Rank}(\mathcal{C}_R) = S\big\}\,.
\]
The elements in ${\sf A}_\ell(T)$ not in ${\sf A}_\ell(S)$ for $S\subsetneq T$ are, thus, exactly the elements of the form $(\mathcal{C}_R,\mathcal{D}_R,\mathcal{M})$ for some maximal chain~$\mathcal{M}$ and some subset $E \subseteq \{1,\dots,n\}$ such that ${\sf Rank}(\mathcal{C}_R) = T$ and $R = \TT{\mathcal{M}}{E}$.
This yields the proposed description of~${\sf B}_{\ell}(T)$.
\end{proof}
We finally describe the cardinality of the set ${\sf B}_\ell(T)$ in terms of the statistic ${\sf u}(\mathcal{M},E)$ given in \Cref{sec:mainresults}.
Recall for a subset $T \subseteq \{0,\dots,n-1\}$, the monomial ${\sf u}_T = w_0\dots w_{n-1}$ in $\mathbf{a},\mathbf{b}$ is given by $w_i = \mathbf{b}$ if $i \in T$ and $w_i = \mathbf{a}$ if $i \notin T$.
\begin{corollary}
\label{prop:topPIEStat}
Let $\ell \geq 0$, and let $T \subseteq \{0,1,\dots,n-1\}$.
Then $\#{\sf B}_\ell(T)$ is the number of pairs $(\mathcal{M},E)$ of a maximal chain~$\mathcal{M}$ and a subset $E \subseteq \{1,\dots,n\}$ of cardinality $\#E = \ell$ such that ${\sf u}(\mathcal{M},E) = {\sf u}_T$.
\end{corollary}
\begin{proof}
\Cref{prop:PIEdescription} shows that for every pair $(\mathcal{M},E)$, the triple $(\mathcal{C}_{R},\mathcal{D}_{R},\mathcal{M})$ is contained in ${\sf B}_{\#E}(T)$ for $T = {\sf Rank}(\mathcal{C}_{R})$ and $R=\TT{\mathcal{M}}{E}$.
The definition of $\mathcal{C}_{R}$ yields that~$T$ is the set of positions in $\{0,\dots,n-1\}$ given by
\[
\big\{ i \mid i \notin E,\ i+1 \in E \big\}
\cup \big\{ i \mid i, i+1 \in E,\ u_{i+1} = \mathbf{a}\big\}
\cup \big\{ i \mid i, i+1 \notin E,\ u_{i+1} = \mathbf{b}\big\}\,,
\]
where ${\sf u}(\mathcal{M}) = {\sf u}(\mathcal{M},\emptyset) = u_1\cdots u_n$.
The description of ${\sf u}(\mathcal{M},E)$ in \Cref{sec:mainresults} can be easily seen to be equivalent to ${\sf u}(\mathcal{M},E) = v_1\dots v_n$ with
\begin{equation*}
\begin{array}{ll}
v_{i+1} = \mathbf{a} &\text{ if } \begin{cases}
i\in E,\quad i+1\notin E, \quad \text{or} \\
i \notin E, \quad i+1\notin E, \quad u_{i+1} = \mathbf{a} \quad \text{ or} \\
i\in E, \quad i+1\in E, \quad u_{i+1} = \mathbf{b}\,,
\end{cases} \\ \\
v_{i+1} = \mathbf{b} &\text{ if } \begin{cases}
i\notin E,\quad i+1\in E, \quad \text{or} \\
i \in E, \quad i+1\in E, \quad u_{i+1} = \mathbf{a} \quad \text{ or} \\
i\notin E, \quad i+1\not\in E, \quad u_{i+1} = \mathbf{b}
\end{cases}
\end{array}
\end{equation*}
for~$i \in \{0,\dots,n-1\}$.
Comparing these two descriptions of~$T$ and of ${\sf u}(\mathcal{M},E)$ proves that ${\sf u}_T = {\sf u}(\mathcal{M},E)$.
\end{proof}
\begin{proof}[Proof of~\Cref{thm:combinatorial-interp}]
Combining \Cref{prop:topPIEStat} and \Cref{thm:toppiecount} shows that the coefficient of $y^\ell {\sf u}_T$ in ${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})$ equals
\[
\#\big\{ (\mathcal{M},E) \mid \mathcal{M} \text{ maximal chain},\ E \subseteq \{1,\dots,n\},\ \#E = \ell,\ {\sf u}(\mathcal{M},E) = {\sf u}_T \big\}\,.
\]
In other words, we have
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) = \sum_{(\mathcal{M},E)} y^{\#E}\cdot{\sf u}(\mathcal{M},E)\,,
\]
where the sum ranges over all maximal chains~$\mathcal{M}$ and all subsets $E \subseteq \{1,\dots,n\}$.
\end{proof}
\begin{proof}[Proof of~\Cref{thm:symmetry-of-coeffs}]
For a given chain $\mathcal{M}$, exchanging $E\subseteq \{1,\dots,n\}$ with its complement $E^c = \{1,\dots,n\}\setminus E$ also exchanges roles of $\mathbf{a}$ and $\mathbf{b}$ in ${\sf u}(\mathcal{M},E) = w_1\dots w_n$ to ${\sf u}(\mathcal{M},E^c) = w^c_1\dots w^c_n$, meaning that $\{w_i,w_i^c\} = \{\mathbf{a},\mathbf{b}\}$ for every position~$i$.
The result now follows with \Cref{thm:combinatorial-interp}.
\end{proof}
\begin{proof}[Proof of~\Cref{cor:enab0}]
Setting
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi^\circ(P;y,\mathbf{a},\mathbf{b}) = \sum_{\mathcal{C}\text{ chain in } P\setminus\{\hat 0,\hat 1\}} \POIN[\{\hat 0\} \cup \mathcal{C}]{P}{y} \cdot \wt^-_\mathcal{C}(\mathbf{a},\mathbf{b})\ \in \mathbb{N}[y]\langle \mathbf{a},\mathbf{b}\rangle \,,
\]
we show that
\begin{equation}
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi^\circ(P;y,\mathbf{a},\mathbf{b}) = \iota\big({\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})\big)\,.
\label{eq:iotaextpsi}
\end{equation}
All previously given arguments remain valid when only considering those sets ${\sf A}_\ell(S)$ and ${\sf B}_\ell(T)$ that contain~$0$.
For $T \subseteq \{1,\dots,n-1\}$, we set
\begin{equation}
{\sf B}^\circ_\ell(T) = {\sf A}_\ell(T\cup\{0\}) \big\backslash \bigcup_{S\subsetneq T} \varphi_{S,T}\big({\sf A}_\ell(S\cup\{0\})\big)\,.
\label{eq:toppiecirc}
\end{equation}
It is immediate that the analogue of \Cref{thm:toppiecount} holds, meaning that
\begin{equation}
[y^\ell {\sf u}_T]\ {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi^\circ(P;y,\mathbf{a},\mathbf{b}) = \# {\sf B}^\circ_\ell(T)
\label{eq:extpsicirccoeff}
\end{equation}
for ${\sf u}_T = w_1\cdots w_{n-1}$ with $w_i = \mathbf{b}$ if $i \in T$ and $w_i = \mathbf{a}$ if $i \notin T$.
Also the argument for \Cref{prop:topPIEStat} remains valid when considering ${\sf u}_T = w_1\cdots w_{n-1}$ instead of $w_0\cdots w_{n-1}$.
We get that $\#{\sf B}^\circ_\ell(T)$ equals the number of pairs $(\mathcal{M},E)$ of a maximal chain~$\mathcal{M}$ and a subset $E \subseteq \{1,\dots,n\}$ of cardinality $\#E = \ell$ such that $\iota({\sf u}(\mathcal{M},E)) = {\sf u}_T$.
This implies \Cref{eq:iotaextpsi}, finishing the proof of \Cref{cor:enab0}.
\end{proof}
We finish this section with a proof of \Cref{thm:lower-bound-on-coefficients}.
That is, we show that
\begin{equation}\label{eqn:LB}
[t^k]\ {\sf Num}(P;1,t) \geq \binom{n-1}{k}\cdot \POIN{P}{1}\,.
\end{equation}
By \Cref{thm:maglione-voll-conjecture} and \Cref{eq:poinfromnum}, we have
\[
{\sf Num}(P;1,t) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi^\circ(P;1,1,t)\quad\text{and}\quad
\mathcal{\sf Poin}(P;1) = {\sf Num}(P;1,0) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi^\circ(P;1,1,0)
\]
for ${\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi^\circ(P;y,\mathbf{a},\mathbf{b}) = \iota\big({\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b})\big)$ as given in \Cref{eq:iotaextpsi}.
Together with \Cref{eq:extpsicirccoeff}, \Cref{eqn:LB} is equivalent to
\begin{equation}
\sum_T \sum_{\ell\geq 0} \#{\sf B}^{\circ}_\ell(T)
\geq
\binom{n-1}{k}\cdot\sum_{\ell \geq 0} \#{\sf B}^{\circ}_\ell(\emptyset)\,,
\label{eq:boundingsets}
\end{equation}
where the leftmost sum runs over all subsets $T \subseteq \{1,\dots,n-1\}$ of cardinality~$k$.
\begin{lemma}\label{lem:injective}
Let $T \subseteq \{1,\dots,n-1\}$. Then
\[
\sum_{\ell\geq 0} \#{\sf B}^{\circ}_\ell(T) \geq \sum_{\ell\geq 0}\# {\sf B}^{\circ}_\ell(\emptyset).
\]
\end{lemma}
\begin{proof}
It is clearly sufficient to construct an injection
\[
\bigcup_{\ell\geq0}{\sf B}^{\circ}_\ell(\emptyset) \hookrightarrow
\bigcup_{\ell\geq0}{\sf B}^{\circ}_\ell(T)\,.
\]
Fix $\ell\geq0$.
Then $(\mathcal{C},\mathcal{D},\mathcal{M}) \in {\sf B}^{\circ}_\ell(\emptyset)$ if and only if $\mathcal{C} = \{\mathcal{M}_0\}$, $\mathcal{D} = \multichain{\mathcal{M}_\ell}$, and~$\mathcal{M}$ is decreasing along
the interval $[\mathcal{M}_0,\mathcal{M}_{\ell}]$ and weakly increasing along the interval
$[\mathcal{M}_{\ell},\mathcal{M}_n]$.
Now map $(\mathcal{C},\mathcal{D},\mathcal{M})$ to $(\mathcal{C}_{\mathcal{M},T},\mathcal{D}_{\mathcal{M},T,\ell},\mathcal{M})$.
Here, $\mathcal{C}_{\mathcal{M},T} = \{ \mathcal{M}_0 \} \cup \{ \mathcal{M}_i \mid i \in T \}$,
\[
\mathcal{D}'_{\mathcal{M},T,\ell} = \multichain{ \mathcal{M}_\ell} \sqcup \multichain{ \mathcal{M}_{i-1} \mid i \in T, i < \ell} \sqcup \multichain{ \mathcal{M}_{i+1} \mid i \in T, i > \ell}
\]
and $\mathcal{D}_{\mathcal{M},T,\ell} = \mathcal{D}'_{\mathcal{M},T,\ell}$ if $\ell \notin T$.
If $\ell \in T$,
\[
\mathcal{D}_{\mathcal{M},T,\ell} = \begin{cases}
\mathcal{D}'_{\mathcal{M},T,\ell} \sqcup \multichain{\mathcal{M}_{\ell-1}} &\text{if } \lambda(\mathcal{M}_{\ell-1},\mathcal{M}_\ell) > \lambda(\mathcal{M}_{\ell},\mathcal{M}_{\ell+1})\,, \\
\mathcal{D}'_{\mathcal{M},T,\ell} \sqcup \multichain{\mathcal{M}_{\ell+1}} &\text{if }\lambda(\mathcal{M}_{\ell-1},\mathcal{M}_\ell) \leq \lambda(\mathcal{M}_{\ell},\mathcal{M}_{\ell+1})\,. \\
\end{cases}
\]
By construction, we have that ${\sf Rank}(\mathcal{C}_{\mathcal{M}, T}) = \{0\} \cup T$ and moreover that
\[
(\mathcal{C}_{\mathcal{M},T},\mathcal{D}_{\mathcal{M},T,\ell},\mathcal{M}) \in {\sf B}^{\circ}_{\ell'}(T)
\]
for some parameter~$\ell'$.
This shows that the map is well-defined.
It remains to show it is also injective.
First note that if~$\mathcal{M}$ is a maximal chain which first decreases and then only weakly increases, then there are exactly two interlacing pairs $(\mathcal{C}_1,\mathcal{D}_1),(\mathcal{C}_2,\mathcal{D}_2)$ and exactly two values of $\ell_1,\ell_2\geq 0$ such that $(\mathcal{C}_i,\mathcal{D}_i,\mathcal{M}) \in {\sf B}^{\circ}_{\ell_i}(\emptyset)$.
Moreover,~$\ell_1$ and~$\ell_2$ differ by $1$ meaning that $\{\ell_1,\ell_2\} = \{\ell,\ell+1\}$ for some $\ell\geq 0$ and we have
\[
\big(\{\mathcal{M}_0\},\multichain{\mathcal{M}_\ell},\mathcal{M}\big) \in {\sf B}^{\circ}_\ell(\emptyset),\quad \big(\{\mathcal{M}_0\},\multichain{\mathcal{M}_{\ell+1}},\mathcal{M}\big) \in {\sf B}^{\circ}_{\ell+1}(\emptyset)\,.
\]
Injectivity follows with $\mathcal{D}_{\mathcal{M},T,\ell} \neq \mathcal{D}_{\mathcal{M},T,\ell+1}$ which is immediate from its definition.
\end{proof}
\begin{proof}[Proof of \Cref{thm:lower-bound-on-coefficients}]
This follows from \Cref{eq:boundingsets} and \Cref{lem:injective}.
\end{proof}
\section{Extending the $\mathbf{a}\mathbf{b}$-index}
\label{sec:refinement-of-ber}
In this section we prove~\Cref{thm:refinement-of-ber}.
The proof comes from first proving a special case of~\Cref{thm:refinement-of-ber} and then showing how to extend this special case to the general setting.
\medskip
As before, we fix an $R$-labeled poset~$P$ of rank~$n$ together with an $R$-label $\lambda$.
Recall that $\omega({\sf m})$ for a monomial~${\sf m}$ in $\mathbf{a},\mathbf{b}$ substitutes all occurrences of $\mathbf{a}\mathbf{b}$ with $\mathbf{a}\mathbf{b} + y\mathbf{b}\mathbf{a} + y\mathbf{a}\mathbf{b} + y^2\mathbf{b}\mathbf{a}$ and then simultaneously substitutes all remaining occurrences of $\mathbf{a}$ with $\mathbf{a}+y\mathbf{b}$ and all occurrences of $\mathbf{b}$ with $\mathbf{b}+y\mathbf{a}$.
The main ingredient in the proof of the corollary is the following proposition.
Also recall the definition of ${\sf u}(\mathcal{M}) = {\sf u}(\mathcal{M},\emptyset)$ given in \Cref{eq:statmon}.
\begin{proposition}
\label{eq:extending-equality}
Let~$\mathcal{M}$ be a maximal chain in~$P$.
Then
\[
\omega\big({\sf u}(\mathcal{M})\big) = \sum_{E\subseteq\{1,\dots,n\}} y^{\#E}\cdot {\sf u}(\mathcal{M},E)\,.
\]
\end{proposition}
Before proving the proposition, we show how to deduce the corollary.
\begin{proof}[Proof of \Cref{thm:refinement-of-ber}]
\Cref{thm:combinatorial-interp} gives that
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) = \sum_{\mathcal{M},E} y^{\#E}\cdot{\sf u}(\mathcal{M},E)\,,
\]
and in particular,
\[
\Psi(P;\mathbf{a},\mathbf{b}) = {\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;0,\mathbf{a},\mathbf{b}) = \sum_{\mathcal{M}} {\sf u}(\mathcal{M})\,.
\]
By \Cref{eq:extending-equality}, we have
\[
{\text{\scriptsize\sf ex}\hspace*{-1pt}}\Psi(P;y,\mathbf{a},\mathbf{b}) = \sum_{\mathcal{M}} \omega\big({\sf u}(\mathcal{M})\big) = \omega\left(\sum_{\mathcal{M}} {\sf u}(\mathcal{M})\right) = \omega\big(\Psi(P;\mathbf{a},\mathbf{b})\big)\,. \qedhere
\]
\end{proof}
For a maximal chain $\mathcal{M}$, we have that ${\sf u}(\mathcal{M}) = v_1\dots v_{n}$ starts with $v_1 = \mathbf{a}$, compare \Cref{eq:firstindex}.
For this reason, ${\sf u}(\mathcal{M})$ can be decomposed into a product of monomials of the form $\mathbf{a} \mathbf{b}^j$ for some $j \geq 0$.
This decomposition of ${\sf u}(\mathcal{M})$ induces a unique decomposition of $\mathcal{M}$ into intervals $[X_i,X_{i+1}]$ such that the restriction $\mathcal{M}_{[X_i,X_{i+1}]}$ of $\mathcal{M}$ to the interval $[X_i,X_{i+1}]$ satisfies
\[
{\sf u}(\mathcal{M}_{[X_i,X_{i+1}]}) = \mathbf{a}\mathbf{b}^j
\]
for some $j\geq 0$.
We call this the \Dfn{decomposing chain} $\{\hat{0} = X_1 < \cdots < X_k = \hat{1}\}$ of~$\mathcal{M}$.
The following lemma shows how we can use decomposing chains to reduce the proof of~\Cref{eq:extending-equality} to a special case.
\begin{lemma}
\label{lem:dijon-subsets}
For a maximal chain~$\mathcal{M}$ with decomposing chain $\{\hat{0} = X_1 < \cdots < X_k = \hat{1}\}$ as above, we have
\[
\omega\big({\sf u}(\mathcal{M})\big) = \prod_{i = 1}^{k - 1} \omega\big({\sf u}(\mathcal{M}_{[X_i,X_{i+1}]})\big).
\]
\end{lemma}
\begin{proof}
This follows from the definition of the substitution~$\omega$ and the fact that each factor~$\mathbf{a}\mathbf{b}$ in ${\sf u}(\mathcal{M})$ appears in one of the intervals of the decomposition.
\end{proof}
We are now ready to prove~\Cref{eq:extending-equality} on each of the invervals given by the decomposing chain.
From there, we then use \Cref{lem:dijon-subsets} to recover $\omega\big({\sf u}(\mathcal{M})\big)$.
\begin{lemma}\label{lem:omeggs-over-easy}
Suppose ${\sf u}(\mathcal{M}) = \mathbf{a} \mathbf{b}^j$ for some $j\geq 0$.
Then
\[
\omega\big({\sf u}(\mathcal{M})\big) = \sum_{E\subseteq\{1,\dots,j+1\}} y^{\#E}\cdot {\sf u}(\mathcal{M},E)\,.
\]
\end{lemma}
\begin{proof}
For $j = 0$, we have $\omega\big({\sf u}(\mathcal{M})\big) = \mathbf{a} + y\mathbf{b}$, as desired.
For $j\geq 1$, we have
\begin{equation}
\label{eqn:cracking-omeggs}
\omega\big({\sf u}(\mathcal{M})\big) = (\mathbf{a}\mathbf{b} + y\mathbf{b}\mathbf{a} + y\mathbf{a}\mathbf{b} + y^2\mathbf{b}\mathbf{a})(\mathbf{b} + y\mathbf{a})^{j-1}.
\end{equation}
There are $4\cdot 2^{j-1} = 2^{j+1}$ terms in the expansion.
For each $E \subseteq \{1,\dots, j+1\}$, we explicitly describe a unique summand $\mathsf{s}_E$ of the expansion of the right side of \Cref{eqn:cracking-omeggs}.
Fix $E\subseteq \{1,\dots, j+1\}$ and write $\mathsf{s}_E = s_1\cdots s_j$.
For $i\in \{2,\dots, j\}$, we set $s_i$ to $y\mathbf{b}$ if $i\in E$ and otherwise
to $\mathbf{a}$.
Lastly, we set
\[
s_1 = \begin{cases}
\mathbf{a}\mathbf{b} & \text{if } 1\notin E,\ j+1 \notin E, \\
y\mathbf{b}\mathbf{a} & \text{if } 1\in E,\ j+1 \notin E, \\
y\mathbf{a}\mathbf{b} & \text{if } 1\notin E,\ j+1 \in E, \\
y^2\mathbf{b}\mathbf{a} & \text{if } 1 \in E,\ j+1 \in E.
\end{cases}
\]
It is clear that $\mathsf{s}_E$ is a summand in (the expansion of)
\Cref{eqn:cracking-omeggs} since~$E$ encodes exactly how to select terms in the~$j$ different factors.
Moreover, $\mathsf{s}_E = y^{\# E} \cdot {\sf u}(\mathcal{M}, E)$.
\end{proof}
\begin{proof}[Proof of~\Cref{eq:extending-equality}]
Let $\{\hat{0} = X_1 < \cdots < X_k = \hat{1}\}$ be the decomposing chain for~$\mathcal{M}$.
For $i\in \{1,\dots, k-1\}$, set $\mathsf{IR}_i = \{{\sf rank}(X_i)+1,\dots,{\sf rank}(X_{i+1})\}$.
If $E_i\subseteq \mathsf{IR}_i$ for each $i\in \{1,\dots, k-1\}$ and $E=E_1\cup \cdots \cup E_{k-1}$, then
\[
{\sf u}(\mathcal{M},E) = {\sf u}(\mathcal{M}_{[X_1,X_{2}]},E_1)\cdots{\sf u}(\mathcal{M}_{[X_{k-1},X_{k}]},E_{j-1}).
\]
Together with \Cref{lem:omeggs-over-easy,,lem:dijon-subsets}, this gives
\begin{align*}
\omega\big({\sf u}(\mathcal{M})\big) & = \prod_{i = 1}^{j - 1} \omega\big({\sf u}(\mathcal{M}_{[X_i,X_{i+1}]})\big)\\
& = \prod_{i = 1}^{j - 1} \left(\sum_{E_i\subseteq\mathsf{IR}_i} y^{\#E_i}\cdot {\sf u}\big(\mathcal{M}_{[X_i,X_{i+1}]},E_i\big) \right)\\
& = \sum_{E\subseteq\{1,\dots,n\}} y^{\#E}\cdot {\sf u}(\mathcal{M},E)\,. \qedhere
\end{align*}
\end{proof}
\bibliographystyle{plain}
|
\section{The Interstellar Medium}
The Interstellar Medium (ISM) constitutes the reservoir of matter, that was and still is turned into stars and planets and gave also rise to the existence of our own solar system and our world.
If its study wasn't already interesting for just that reason, there are many complex processes that impact it chemically as well as energetically.
The ISM is being enriched with heavier elements by the more massive stars in their late evolutionary phases, but also diluted by the influx of extragalactic matter.
Feedback from the different phases of stellar life, but also cosmic rays and AGN inject energy, which is released by emission in many atomic and molecular lines ([C\,{\sc ii}], [O\,{\sc i}], C$_2$H$_2$, H$_2$O, PAHs, etc.) as well as thermal emission by different kinds of dust.
Even though the general scenario of star formation is reasonably well understood, the details of the complex interplay of stellar radiation, gravitation, turbulence and magnetic fields, that determine the timescales and the interstellar mass function, are not.
A large number of these lines as well as the peak of the thermal emission are located in the mid- to far-infrared (MIR 3-30$\mu$m, FIR 30-300$\mu$m) wavelength range as illustrated in Fig.~\ref{Meixner:fig:ismsed} (top), making this portion of the electromagnetic spectrum key to studying the ISM and a multitude of related scientifically interesting phenomena.
However, this is also a rather difficult spectral range to observe as shown in the lower part of Fig.~\ref{Meixner:fig:ismsed}, which illustrates the atmospheric transmission at the levels of the Atacama Large Millimeter/submillimeter Array (ALMA) and the Stratospheric Observatory for Infrared Astronomy (SOFIA). Telluric water vapor and ozone leave only certain windows in the MIR and sub-millimeter ranges, while the FIR is effectively unobservable from the ground and requires observatories in the stratosphere or in space.
\begin{figure}[htp]
\begin{minipage}{0.50\textwidth}
\includegraphics[width=\textwidth]{MeixnerIsmSED.pdf}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[t]{0.4\textwidth}
\caption{The Spectral Energy Distribution (SED) of the interstellar medium from mid- to far-infrared wavelengths (top) and the corresponding transmission spectra of the Earth atmosphere at the operating altitudes of ALMA and SOFIA (bottom).}
\label{Meixner:fig:ismsed}
\end{minipage}
\end{figure}
\section{JWST}
Launched at the end of 2021, the James Webb Space Telescope (JWST) provides access to the full MIR spectrum in space since its first science data were released in July 2022.
Its high spatial resolution, similar to that of Hubble in the visible spectrum, and its access to PAH emission as well as the ro-vibrational lines of molecular hydrogen and water, make JWST an excellent probe of star formation regions.
JWST images of 30 Doradus, aka the Tarantula Nebula, show the stars, molecular hydrogen and PAHs with NIRCam (0.6-5~$\mu$m) and warm dust and PAHs with MIRI (4.9 to 28.8~$\mu$m).
These kind of maps provide an unprecedented amount of detail at those wavelengths and will play an important role in further investigating the hot and warm ISM.
The spectroscopic capabilities of JWST are considerable as well, yet limited in terms of spectral resolution with R~$\approx 2700$ \citep{meixner:cite:Boeker2022} for NIRSpec and R~$\approx 1300$ to 3700 for MIRI \citep{meixner:cite:Wells2015}.
This is where SOFIA provided complementary high spectral resolution spectroscopy (R~$\approx10^5$) with EXES \citep{meixner:cite:Richter2018}, even though at lower spatial resolution and sensitivity.
\section{SOFIA}
\subsection{Importance and Successes}
With its five exchangeable scientific instruments (SIs), SOFIA filled nicely the large spectral gap in the FIR between JWST and ALMA and provided further complementary capabilities like high spectral resolution at JWST wavelengths with EXES and the ability to observe very bright sources with FORCAST, that filled in the overexposed areas in MIR maps of the Galactic Center region, made by Spitzer \citep{meixner:cite:Hankins2020}.
The recent discovery of water on the sunlit surface of the Moon by \cite{meixner:cite:Honniball2021} falls into that category as well.
The heterodyne instrument GREAT covers such important atomic fine structure lines as [C\,{\sc ii}]~and [O\,{\sc i}]~at the highest spectral resolutions of up to R~$\approx 10^6$ and fills in the spectral gaps that are inaccessible for ALMA due to atmospheric extinction.
This enabled not only the discovery of new molecules in the ISM like Helium Hydride \citep{meixner:cite:Guesten2019}, but also very detailed kinematic studies e.g. feedback processes in Orion by \cite{meixner:cite:Pabst2019} that triggered a very successful SOFIA legacy program by \cite{meixner:cite:Schneider2020}.
When sensitivity became an issue and could be gained by sacrificing spectral resolution, in particular for extragalactic work, the FIFI-LS spectrometer provided a good alternative for observations of fine structure lines as shown by \cite{meixner:cite:Fadda2021}, \cite{meixner:cite:Spinoglio2022} or \cite{meixner:cite:Pineda2018}.
Last but not least, where very high sensitivity was required to reveal the peak of the cold dust emission of high redshift objects, HAWC+ provided the FIR imaging capability.
HAWC+, however, also provided an entire new dimension for ISM studies, that had before only briefly been available with ISO in the FIR.
Polarization mapping revealed the vectors of magnetic fields in the ISM thanks to the FIR emission of aligned elongated dust particles. Many publications sparked a lot of new observational as also theoretical interest in this previously rather dormant field \citep{meixner:cite:Pillai2020, meixner:cite:Lopez-Rodriguez2021, meixner:cite:Zielinski2021}.
\subsection{Mission Success and Conclusion}
In the face of the tremendous scientific successes of this true ISM-Machine, the decision by NASA and DLR to end the SOFIA mission after only 9 observing cycles, is certainly very hard to understand.
Following the recommendations from the Flagship Mission Review from 2019, the project has transformed since then with a tremendous growth in science productivity as demonstrated in the SOFIA Status and Future Prospects Report \citep{meixner:cite:Rangwala2022}\footnote{This report was already prepared for NASA's Senior Review Process.}.
Annual publication rates for SOFIA have doubled over the past three years on topics ranging from the Earth to high-z galaxies \citep{meixner:cite:Schmelz2021}.
The Decadal Survey Astro~2020 recommended to NASA to terminate the SOFIA mission, which unfortunately was based on outdated ($>2$~years) and incorrect information\footnote{SOFIA science addresses 50\% of Astro~2020 key science questions, not 10\%.}.
NASA holds Astro 2020 recommendations as superior to Senior Review process results and hence removed SOFIA from the Senior Review.\footnote{This avoided potentially ending up with two contradicting recommendations.}
Arguments that SOFIA's science productivity was insufficient can be easily refuted by comparing the observing time that is spent on average per refereed publication to that of Herschel. Eight years after launch Herschel had provided about 23,500 hours of observing time and produced 2,145 publications, resulting in $\approx$11~h$/$paper. SOFIA with 3458 hours and 330 publications after 8 years since achieving full science operational capability in 2014 results in very similar 10.5~h$/$paper.
Fortunately the last year was particularly productive in terms of observations, so there is a considerable amount of science data in the IRSA archive. As there is only a minimal post-operational phase of one year planned by NASA at this point, we hope DLR will provide the means to conduct data reprocessing also for the time before Cycle~5, advanced water vapor and pointing analysis and more comprehensive corrections, which are currently not included in the plans. In the next section we'll lay out that the time to the next FIR mission might be rather long. Already collected FIR photons might thus be even more valuable for astronomy and funds for maximizing their scientific usability will be well spent.
\section{Future Far-Infrared Observatories}
\subsection{History and Guidance}
Fig.~\ref{Meixner:fig:firhistory} illustrates the history of FIR astronomy by showing the operational phases of all major observatories as green boxes, starting in the sixties until today and the current outlook towards 2045. Up to today, there was an almost continuous capability to supply astronomers with current FIR observations except for the few years between ISO and Spitzer. With the sudden cancellation of SOFIA, which was originally scheduled to continue until 2034, and the cancellation of SPICA by ESA in 2021, the opportunities for FIR data collection have become sparse.
In \cite{meixner:cite:Rangwala2022} Page~4, a traceability matrix can be found, that links Astro 2020 science questions to key measurements in the MIR and FIR, that could have been performed with SOFIA.
This list should still be useful as a collection of science requirements for the design of future stratospheric- and space-observatories.
\subsection{New Opportunities in Space}
Even though Astro~2020 recommended the cancellation of SOFIA, it acknowledged the importance of the FIR spectral region for astrophysics and recommended the launch of a Probe space mission for 2030 that will specialize either in FIR- or X-ray- astronomy.
NASA followed this up by issuing an announcement of opportunity and a proposal deadline of October 2023, a downselection end of 2025, a cost cap of 1B$\$$ excluding the launcher and a launch date not later than 2032 \citep{meixner:cite:NASA-AO2022}.
If history is a guide, such a schedule is highly optimistic. In reality a launch might rather be expected in the mid 2030s, not to mention that continuing the SOFIA mission until its planned end would have cost substantially less, especially when taking into account the launcher as well.
Given that the X-ray community is also competing for another opportunity, it is everything but a done deal that NASA's probe mission will be dedicated to the FIR.
If that doesn't happen, then also the dream of a more ambitious true observatory for the FIR such as ORIGINS in the 2040s \citep{meixner:cite:Meixner2019} may become unrealistic with observational FIR astronomy having lost a lot of its expertise by then.
Therefore at this point it is quite important for the FIR community to look towards the future which at least in space will be the Probe mission. There are four mission proposals for the FIR named PRIMA (PI, Jason Glenn)\footnote{\url{https://prima.ipac.caltech.edu}}, SPICE (PI Lee Mundy)\footnote{\url{https://asd.gsfc.nasa.gov/spice/index.html}}, FIRSST (PI Asantha Cooray) and SALTUS (PI Chris Walker), which were presented at the IR Astrophysics Workshop 2022 in Colorado. The concepts comprise more traditional space observatories with cold telescopes like PRIMA and FIRSST, and more unusual ones like the interferometer SPICE or the large inflatable telescope concept SALTUS. Details as presented at the workshop are available at the workshop website \citep{meixner:cite:Irstig2022}.
\subsection{Stratospheric Opportunities}
In the meantime the FIR community should also investigate other opportunities to reclaim a permanent capability in that part of the spectrum.
This will in particular enable more time dependent FIR astronomy, that we consider being still in its infancy.
The fairly short life spans of FIR missions so far have been a hindrance while time-domain astronomy has really taken off in other parts of the electromagnetic spectrum.
The astrophysical community should investigate the available potential in the FIR.
SOFIA was likely the last airplane observatory, and future stratospheric platforms will probably be of the lighter-than-air category.
Current balloon experiments are, however, rather short lived, extremely weather dependent with very few launch opportunities, can't stay in a particular region for long and have only a 50~\% survival rate upon landing.
Such missions are still seen rather as serving technology maturation and the training of instrumentalists than being able to support serious general observatory type projects for the astronomical community.
This school of thought needs to change as better technologies become available that could address many of the shortcomings mentioned above.
Longer lived robotic stratospheric platforms with propulsion may also be interesting to a wider community including UV- and FIR-astronomy but also climate research and general Earth observation \citep{meixner:cite:Miller2014}.
\section{Conclusion}
Even though the end of SOFIA is a blow to FIR astronomy, the mission and its team have performed excellently and are concluding at peak performance with much data in the archive that await analysis and publication.
JWST is the observatory now to study the ISM in warm/hot conditions, while there will be new opportunities for observatories that can study the cold ISM from space or from the stratosphere.
|
\section{Introduction}
The \textit{inner product} of two vectors $\bv{a}$ and $\bv{b}$, which is defined as $\langle \bv{a}, \bv{b} \rangle = \sum_{k=1}^n \bv{a}[k]\bv{b}[k]$, is a ubiquitous operation. Among many other applications, inner products can be used to compute document similarities~\cite{SaltonWongYang:1975}, to evaluate learned classification models, and to estimate join sizes%
~\cite{AlonMatiasSzegedy:1999,Achlioptas:2003,RusuDobra:2008}.
However, in modern applications involving very high-dimensional vectors, computing exact inner products can be intractable. The naive computational cost is $O(n)$. Moreover, computing $\langle \bv{a}, \bv{b}\rangle$ requires loading $O(n)$ numbers from memory, or communicating $O(n)$ numbers if $\bv{a}$ and $\bv{b}$ are stored on different machines.
A common approach for resolving this issue is to pre-compute a small space compression (a sketch) of each vector, which we will denote by $\mathcal{S}(\bv{a})$ and $\mathcal{S}(\bv{b})$, respectively. An estimation function $\mathcal{F}$ is then used to approximate the inner product as $\mathcal{F}\left(\mathcal{S}(\bv{a}), \mathcal{S}(\bv{b})\right) \approx \langle \bv{a}, \bv{b} \rangle$.
The beauty of sketching is that it simultaneously reduces storage, communication, and runtime complexity. Moreover, once computed, sketches can be reused again and again to estimate inner products with other vectors. For example, given another vector $\bv{c}$ we can estimate $\langle \bv{a}, \bv{c} \rangle \approx \mathcal{F}\left(\mathcal{S}(\bv{a}), \mathcal{S}(\bv{c})\right)$.
Sketching methods for approximating inner products are already widely used throughout computer science. In machine learning, they can be used to accelerate the training of large-scale linear models like support vector machines or logistic regression
\cite{ArriagaVempala:2006,LiShrivastavaMoore:2011}. In relational databases, inner product sketches are used in query optimizers to choose optimal query plans without having to execute expensive queries that involve large joins \cite{CormodeGarofalakisHaas:2011}. More recently, inner product sketches have found applications in dataset search and discovery. Here, they are used to discover joinable tables \cite{FernandezMinNava:2019} and to estimate other column statistics, such as correlation~\cite{SantosBessaChirigati:2021}, without explicitly performing a {join} operation between two tables. We discuss these applications and others in \Cref{sec:apps}.
\subsection{What was Previously Known?}
In all of the applications above, a primary concern is optimizing the trade-off between the sketch size (which governs storage, communication, and runtime efficiency) and how accurately $\mathcal{F}\left(\mathcal{S}(\bv{a}), \mathcal{S}(\bv{b})\right)$ approximates
$\langle \bv{a}, \bv{b} \rangle$. A large sketch size will in general lead to better approximation, but the question is by exactly how much.
Currently, the only methods with strong theoretical guarantees on this tradeoff
for \textit{general vectors} (i.e., vectors without any assumed value distribution or magnitude) are based on \emph{linear sketching} algorithms. Such algorithms include the famous ``tug-of-war'' sketch, a.k.a. the AMS sketch \cite{AlonMatiasSzegedy:1999,AlonGibbonsMatias:1999}, the CountSketch algorithm \cite{CharikarChenFarach-Colton:2002}, and methods based on Johnson-Lindenstrauss (JL) random projection~\cite{Achlioptas:2003,DasguptaGupta:2003}.
All of these approaches have a similar form. We choose a random matrix $\bv{\Pi} \in \R^{m\times n}$ ($\bv{\Pi}$ might have i.i.d. random entries or more complex structure) and sets $\mathcal{S}(\bv{a})= \bv{\Pi}\bv{a}$ and $\mathcal{S}(\bv{b}) = \bv{\Pi}\bv{b}$.
Each sketch is a length $m$ vector and is considered a \emph{linear sketch} since $\mathcal{S}$ is a linear function.
To estimate the inner product, the typical approach is to simply return the sketch inner product $\langle \mathcal{S}(\bv{a}), \mathcal{S}(\bv{b})\rangle$.\footnote{Other estimators involving e.g., the median of multiple approximate inner products, are also used \cite{LarsenPaghTetek:2021}. However, theoretical guarantees are similar, typically differing in the dependence on the failure probability $\delta$.}
\begin{table*}[t]
\vspace{-.75em}
\def1.0{1}
\centering
\begin{tabular}{ p{0.37\linewidth}|p{0.35\linewidth}|p{0.16\linewidth}}
\toprule
\textbf{Method} & \textbf{Error for sketches of size $O(1/\epsilon^2)$} & \textbf{Assumptions}\\
\midrule
Johnson-Lindenstrauss Projection \cite{ArriagaVempala:2006}, AMS \cite{AlonMatiasSzegedy:1999}, CountSketch \cite{CharikarChenFarach-Colton:2002} & $\epsilon \cdot\|\bv{a}\| \|\bv{b}\|$ & None \\
\midrule
MinHash (MH) Sampling \cite{BeyerHaasReinwald:2007} & $\epsilon \cdot \max\left(\|\bv{a}_{\mathcal{I}}\|\|\bv{b}\|, \|\bv{a}\|\|\bv{b}_{\mathcal{I}}\| \right)$ & $\bv{a}, \bv{b}$ are binary, i.e. with $\{0,1\}$ entries. \\
\midrule
\textbf{Our Method: Weighted MinHash (WMH) Sampling} & $\epsilon \cdot \max\left(\|\bv{a}_{\mathcal{I}}\|\|\bv{b}\|, \|\bv{a}\|\|\bv{b}_{\mathcal{I}}\| \right)$ & None\\
\bottomrule
\end{tabular}
\vspace{.25em}
\caption{Comparison of high-probability additive error guarantees for estimating $\langle \bv{a}, \bv{b}\rangle$ using various sketching methods.
We let $\mathcal{I} = \{i: \bv{a}[i] \neq 0\text{ and } \bv{b}[i] \neq 0\}$ denote the intersection of $\bv a$'s and $\bv{b}$'s supports. $\bv{a}_{\mathcal{I}}$ and $\bv{b}_{\mathcal{I}}$ are $\bv{a}$ and $\bv{b}$ restricted to indices in $\mathcal{I}$.
Since $\max\left(\|\bv{a}_{\mathcal{I}}\|\|\bv{b}\|, \|\bv{a}\|\|\bv{b}_{\mathcal{I}}\| \right) \leq \|\bv{a}\| \|\bv{b}\|$, the bound for our Weighted MinHash (WMH) method always beats the linear sketching methods. Our bound matches that of unweighted MinHash, but without the strong limiting assumption that $\bv{a}$ and $\bv{b}$ are binary: it holds for all vectors.}
\label{tab:error_guarantees}
\end{table*}
A textbook theoretical accuracy guarantee for inner product estimation based on linear sketching is:
\begin{fact}[Linear Sketching for Inner Products \cite{ArriagaVempala:2006}]
\label{fact:jl_result}
Let $\epsilon,\delta \in (0,1)$ be accuracy and failure probability parameters respectively and let $m = O(\log(1/\delta)/\epsilon^2)$.
Let $\bv{\Pi} \in \R^{m \times n}$ be a random matrix with each entry set independently to $+\sqrt{1/m}$ or $-\sqrt{1/m}$ with equal probability. For length $n$ vectors $\bv{a},\bv{b} \in \R^n$, let $\mathcal{S}(\bv{a}) = \bv{\Pi}\bv{a}$ and $\mathcal{S}(\bv{b}) = \bv{\Pi}\bv{b}$. With probability at least $1-\delta$,
\begin{align*}
\left | \langle \mathcal{S}(\bv{a}), \mathcal{S}(\bv{b})\rangle - \langle \bv{a}, \bv{b} \rangle \right | \leq \epsilon \|\bv{a}\| \|\bv{b}\|
\end{align*}
where $\norm{\bv x} = \left (\sum_{i=1}^n x[i]^2\right )^{1/2}$ denotes the standard Euclidean norm.
\end{fact}
In addition to dense random matrices, analgous results to \Cref{fact:jl_result} can be proven for sparse JL matrices, CountSketch matrices, and other linear sketches \cite{CormodeGarofalakisHaas:2011}.
The fact provides a powerful accuracy guarantee that improves with the sketch size $m$ and depends naturally on the norms of $\bv{a}$ and $\bv{b}$. To the best of our knowledge, linear sketching methods were previously the only known algorithms to obtain such a strong theoretical guarantee.
\subsection{Our New Result: Better Inner Product Sketches for Sparse Vectors}
In this paper we introduce a novel method for inner product sketching based on the Weighted MinHash sketch \cite{GollapudiPanigrahy:2006,ManasseMcSherryTalwar:2010,Ioffe:2010}, which is a variant of the classic MinHash method \cite{Broder:1997,BroderCharikarFrieze:1998}. We prove that our method obtains a more refined guarantee than \Cref{fact:jl_result}. In particular, it matches the result for linear sketches in the worst case when $\bv{a}$ and $\bv{b}$ are dense\footnote{For dense vectors, \Cref{fact:jl_result} is actually optimal up to constants: recent work implies that no sketch of size $m = o(\log(1/\delta)/\epsilon^2)$ can achieve error $\epsilon \|\bv{a}\|\|\bv{b}\|$ with probability $1-\delta$ for general inputs \cite{LarsenNelson:2017,AlonKlartag:2017}. Our result also matches this lower bound.}, but always obtains a \emph{better bound} when $\bv{a}$ and $\bv{b}$ are sparse vectors with limited overlap between non-zero entries. As discussed further in \Cref{sec:apps}, such pairs of vectors are the norm in many applications of inner product sketching to database problems and modern dataset search applications. Formally we prove the following:
\begin{theorem}[Main Result]
\label{thm:main}
Let $\epsilon,\delta \in (0,1)$ be accuracy and failure probability parameters and let $m = O(\log(1/\delta)/\epsilon^2)$. There is an algorithm $\mathcal{S}$
that produces size-$m$ sketches (\Cref{alg:weighted_sketch}),
along with an estimation procedure $\mathcal{F}$ (\Cref{alg:weight_est}), such that for any $\bv a,\bv b \in \R^n$, with probability at least $1-\delta$,
\begin{align*}
\left | \mathcal{F}\left(\mathcal{S}(\bv{a}), \mathcal{S}(\bv{b})\right) - \langle \bv{a}, \bv{b} \rangle\right | \leq \epsilon \max\left(\|\bv{a}_{\mathcal{I}}\|\|\bv{b}\|, \|\bv{a}\|\|\bv{b}_{\mathcal{I}}\| \right)
\end{align*}
Above, $\mathcal{I} = \{i: \bv{a}[i] \neq 0\text{ and } \bv{b}[i] \neq 0\}$ is the intersection of $\bv a$'s and $\bv{b}$'s supports. $\bv{a}_{\mathcal{I}}$ and $\bv{b}_{\mathcal{I}}$ denote $\bv{a}$ and $\bv{b}$ restricted to indices in $\mathcal{I}$.
\end{theorem}
We always have $\|\bv{a}_{\mathcal{I}}\| \leq \|\bv{a}\|$ and $\|\bv{b}_{\mathcal{I}}\| \leq \|\bv{b}\|$, so we can bound $ \max\left(\|\bv{a}_{\mathcal{I}}\|\|\bv{b}\|, \|\bv{a}\|\|\bv{b}_{\mathcal{I}}\| \right) \leq \|\bv{a}\|\|\bv{b}\|$.
That is, the guarantee of \Cref{thm:main} matches that of \Cref{fact:jl_result} in the worse-case, but can be significantly better.
For example, consider $\bv{a}$ and $\bv{b}$ that have roughly the same number of non-zero entries, but only a $\gamma < 1$ fraction of those entries are non-zero in \emph{both} $\bv{a}$ and $\bv{b}$. In this case we
expect that $\|\bv{a}_{\mathcal{I}}\|^2 \approx \gamma \|\bv{a}\|^2$ and $\|\bv{b}_{\mathcal{I}}\|^2 \approx \gamma\| \bv{b}\|^2$. As a result, the bound from \Cref{thm:main} will be better than \Cref{fact:jl_result} by a factor of $\sqrt{\gamma}$.
So, to obtain the same error as a linear sketch, our method can set
$m$ smaller by a factor of $\gamma$.
In many applications, $\gamma$ is very small. For example, in \Cref{sec:exp} we consider a document similarity problem where $\gamma \leq .05$ for $95\%$ of vector pairs sketched. This amounts to
$20x$ improvement in sketch size required to achieve a specified level of error.
Thanks to their strong theoretical guarantees, linear sketching algorithms have become the go-to approach for generic inner product estimation \cite{CormodeGarofalakisHaas:2011}. Our results show for the first time that an alternative method can provide stronger bounds. We hope that this paper will serve as a starting point for further investigation into hashing-based algorithms for inner product sketching.
\subsection{Motivating Application: Dataset Search}
\label{sec:apps}
Before presenting the technical details of our results and discussing related work, we detail one application that could benefit from the proposed sketches, and helps illustrate the importance of obtaining bounds for inner product estimation that are sensitive to the number of overlapping non-zero entries in $\bv{a}$ and $\bv{b}$.
Specifically, we consider the problem of \emph{dataset search} which has received increasing attention in recent years~\cite{LehmbergEtAl:2014,ZhuDengNargesianMiller:2019,YangZhangZhang:2019,ZhuNargesianPu:2016,FernandezMinNava:2019,SantosBessaChirigati:2021,SantosBessaMusco:2022}.
Suppose that a data scientist wants to understand the reasons for fluctuations in taxi ridership in New York City in 2022. The analyst only has a table containing two columns: a \textit{date} column and the \emph{number of taxi rides} taken on that day. In order to carry out the analysis, she needs to find other tables, either in her organization's data lake or in public repositories like NYC Open Data (which contain thousands of datasets \cite{New-York:2022}), that would bring in other relevant variables when
joined with the original~table. For example, the analyst might hope to find weather data or a list of holidays and local festivals, which can impact ridership. Moreover, she would like to find relevant factors that she \emph{might not think of on her own}, in an \emph{automatic way}.
To solve this problem, we would like methods to automatically discover tables that are both 1) joinable with the target table (i.e., also contain columns with dates from 2022) and 2) meaningfully correlated with the analyst's data. For example, a table of local festivals should only be returned to the analyst if taxi ridership on days with festivals is either significantly higher or significantly lower than the ridership on average. To find such tables, brute force search is computationally infeasible -- we typically cannot afford to join the analyst's table with all tables in the search set to look for correlations. Instead, we need to efficiently estimate statistics between disparate tables \emph{without materializing their join} \cite{SantosBessaChirigati:2021}.
\input{table-example}
\input{arxiv-table-2-example}
Sketching has become the most popular approach for performing this sort of estimation between unjoined data tables~\cite{ZhuNargesianPu:2016,FernandezMinNava:2019,YangZhangZhang:2019,SantosBessaChirigati:2021,SantosBessaMusco:2022}.
Specifically, a small-space sketch is \emph{precomputed} for all data tables in the search set. When the analyst issues a query to find relevant data, a sketch of her table is compared against these preexisting sketches using a fraction of the computational resources in comparison to explicitly materializing table joins~\cite{SantosBessaChirigati:2021}.
\paragraph{Inner product sketching for dataset search} Interestingly, in the framework discussed above, many problems of interest can be formulated precisely as inner product sketching problems.
To see why this is the case, consider the example tables $\mathcal{T}_{A}$ and $\mathcal{T}_{B}$ shown in Figure~\ref{fig:example-tables}: each contains a column of keys, $K_A$ and $K_B$, and a column of values, $V_A$ and $V_B$.
A \textit{join} operation between the tables
on their keys
generates the output table $\mathcal{T}_{A \bowtie B}$. We list in \Cref{fig:example-tables} a number of statistics that we might hope to estimate in $\mathcal{T}_{A \bowtie B}$ when searching for relevant datasets. We claim that all of these statistics can be estimated using inner products between vector representations of the tables, which we denote $\bv{x}^{\ind{K_A}}, \bv{x}^{{K_A}}$ and $\bv{x}^{\ind{K_B}}, \bv{x}^{{K_B}}$ respectively and show in \Cref{fig:ex-vectors}.
First, it is easy to see that the size of $\mathcal{T}_{A \bowtie B}$ is equal to the intersection between the keys in $K_A$ and $K_B$, i.e., $|K_A \cap K_B| = 4$. This is in turn equal to the \emph{inner product} between $\bv{x}^{\ind{K_A}}$ and $\bv{x}^{\ind{K_B}}$. Similarly, the \texttt{SUM} aggregate of the values in $V_A$ after join (i.e., \texttt{SUM($V_{A \bowtie}$)}) is equal to the inner product $\text{\texttt{SUM($V_{A \bowtie}$)}} = \langle \bv{x}^{V_A}, \bv{x}^{\ind{K_B}} \rangle$.
To estimate a post-join mean (i.e., \texttt{MEAN}($V_{A\bowtie}$)), we can combine the join-size estimate with the SUM estimate:
\vspace{-.25em}
\begin{align*}
\text{\texttt{MEAN}}(V_{A \bowtie}) = \frac{\langle \bv{x}^{V_A}, \bv{x}^{\ind{K_B}} \rangle}{\langle \bv{x}^{\ind{K_A}}, \bv{x}^{\ind{K_B}} \rangle}.
\end{align*}
\vspace{-.75em}
\noindent
In the application discussed above, such a statistic would allow us to determine, based on sketches alone, if for instance the average taxi ridership on festival days was higher than typical. Alternatively, for tables containing precipitation data and taxi ridership, an estimate of
$\langle \bv{x}^{V_A}, \bv{x}^{{V_B}} \rangle$
might signify if high precipitation days align with high ridership days (this would lead to a high inner product).
\paragraph{Comparison of different methods} Given the above reductions, both linear sketching methods like JL projection and CountSketch, and our Weighted MinHash method, can be directly applied to the dataset search problem. We simply need to precompute $\mathcal{S}(\bv{x}^{\ind{K_B}})$ and $\mathcal{S}(\bv{x}^{{V_B}})$ for all tables $\mathcal{T}_B$ in our search set. Sketching other vector transformations like $\mathcal{S}((\bv{x}^{{V_B}})^2)$
opens up the possibility of also estimating other quantities like post-join variance.
In search applications, we note that the vector length $n$ can be very large. However, computing sketches does not require fully materializing the vectors $\bv{x}^{\ind{K_A}}$ and $\bv{x}^{\ind{K_B}}$: all sketching methods discussed in this paper only need to process the vectors' non-zero entries.
Furthermore, it is not necessary to know the $n$ beforehand: we can simply set $n$ to be large enough to cover the whole domain of the keys being sketched (e.g., $n=2^{32}$ or $n=2^{64})$.
To compare methods, \Cref{fact:jl_result} and \Cref{thm:main} suggest that any asymptotic differences in performance between our WMH method and linear sketching will depend on the overlap in non-zero entries between the vectors being sketched. In dataset search, this exactly corresponds to the \emph{Jaccard similarity} of the key sets ${K}_A$ and ${K}_B$. Our method will perform better when the Jaccard similarity is small. For example, in \Cref{fig:example-tables}, only $4$ out of $14$ unique keys are shared in both tables, so the similarity is $\approx .29$. In the scenario discussed above, we could imagine a much smaller ratio: for example, our data analyst might only have a table containing taxi data from 2022, but compare it to a weather data table with dates from 1960 through the present day. The Jaccard similarity would be $1/63 \approx . 016$. Conversely, we might compare to a table with local festivals that only contains a small fraction of dates from 2022. In Section \ref{sec:exp} we consider a dataset search use case involving data from the World Bank \cite{WBF_2022} where more than $42\%$ percent of table pairs had Jaccard similarity $<.1$, and more than $35\%$ have Jaccard similarity $< .05$.
\subsection{Paper Roadmap}
In \Cref{sec:relate_work} we review related prior work. In \Cref{sec:unweighted-minhash} we outline an analysis of the standard \emph{unweighted} MinHash method for inner product estimation. While this analysis falls far short of our desired goal (it requires strong assumptions on $\bv{a}$ and $\bv{b}$) the section serves as a technical warm-up for our main result on Weighted MinHash. We present that main result (\Cref{thm:main}) in \Cref{sec:main_result}. Finally, in \Cref{sec:exp} we support \Cref{thm:main} with an empirical comparison of our method with linear sketching and unweighted hashing methods.
\section{Related Work}
\label{sec:relate_work}
\paragraph{Inner Product Estimation for Binary Vectors} Beyond linear sketching methods for estimating the inner product between general real-valued vectors $\bv{a}$ and $\bv{b}$, there has been a lot of prior work on the special case of \emph{binary} vectors with $\{0,1\}$ entries.
For such vectors, approximating the inner product amounts to approximating the size of the intersection of two sets. Concretely, any $\bv{a}, \bv{b} \in \{0,1\}^n$ can be associated with sets $\mathcal{A}$ and $\mathcal{B}$ that contain integers from $\{1, \ldots, n\}$. We define $\mathcal{A}$ to contain all $i$ for which $\bv{a}[i] = 1$, and similarly $\mathcal{B}$ to contain all $i$ for which $\bv{b}[i] = 1$. Note that $\langle \bv{a}, \bv{b}\rangle = |\mathcal{A} \cap \mathcal{B}|$.
Applying \Cref{fact:jl_result}, we know that a linear sketch of size $m = O(1/\epsilon^2)$ can estimate $\langle \bv{a}, \bv{b}\rangle$ up to additive error $\epsilon \|\bv{a}\| \|\bv{b}\| = \epsilon \sqrt{|\mathcal{A}||\mathcal{B}|}$.
However, it has been shown that a better bound can be obtained using non-linear sketching methods based on
the classic MinHash sketch \cite{Broder:1997,BroderCharikarFrieze:1998,Manber:1994,Heintze:1996}, the $k$-minimum value (KMV) sketch \cite{BeyerHaasReinwald:2007}, or related techniques \cite{LiKonig:2010,LiOwenZhang:2012}. With $m = O(1/\epsilon^2)$ space, such methods are known to achieve error $\epsilon \sqrt{\max(|\mathcal{A}|, |\mathcal{B}|)\cdot |\mathcal{A}\cap \mathcal{B}|}$, which is always smaller than $\epsilon \sqrt{|\mathcal{A}||\mathcal{B}|}$ \cite{BeyerHaasReinwald:2007,PaghStockelWoodruff:2014}. For binary vectors, this bound was proven optimal up to constant factors in \cite{PaghStockelWoodruff:2014}.
Our work was motivated by this pre-existing result for binary vectors. In fact, our main result, \Cref{thm:main}, can be viewed as a strict generalization of the bound to \emph{all real-valued vectors}. In particular, when $\bv{a}$ and $\bv{b}$ are binary, we have that $\|\bv{a}_{\mathcal{I}}\|^2 = \|\bv{b}_{\mathcal{I}}\|^2 = |\mathcal{A}\cap \mathcal{B}|$. So it is not hard to see that $\epsilon \sqrt{\max(|\mathcal{A}|, |\mathcal{B}|)\cdot |\mathcal{A}\cap \mathcal{B}|} = \epsilon \cdot \max\left(\|\bv{a}_{\mathcal{I}}\|\|\bv{b}\|, \|\bv{a}\|\|\bv{b}_{\mathcal{I}}\| \right)$, which is exactly our bound from \Cref{thm:main}. We summarize how all prior inner product sketching methods compare to our result in \Cref{tab:error_guarantees}.
\paragraph{Beyond Binary Vectors}
There has been less work on obtaining better results for estimating inner products of vectors with non-binary entries. One recent paper \cite{LarsenPaghTetek:2021} proves refined bounds for the CountSketch method that depend on the $\ell_1$ norm of $\bv{a}$ and $\bv{b}$ (instead of the Euclidean norm). These bounds can be tighter than \Cref{fact:jl_result} for some vectors, especially when the sketch size $m$ is large. We refer the reader to that work for more details, although note that the results are not directly comparable to ours.
We take a different approach, moving beyond linear sketching entirely. Our main result is based on a class of sketches that we collectively refer to as ``Weighted MinHash'' methods \cite{ChiZhu:2017,Shrivastava:2016}.
These methods include weighted versions of coordinated random sampling \cite{CohenKaplan:2007,CohenKaplan:2013}, as well as the ``Consistent Weighted Sampling'' algorithm \cite{ManasseMcSherryTalwar:2010,GollapudiPanigrahy:2006} and its descendants, which are essentially equivalent, but computationally cheaper to apply \cite{Ioffe:2010,WuLiChen:2019,HaeuplerManasseTalwar:2014}. As shown in \Cref{sec:main_result}, Weighted MinHash sketches allows us to handle vectors whose entries have \emph{highly varying magnitude} (in contrast to binary vectors, where all non-zero entries have the same magnitude of $1$).
Weighted MinHash sketches have been used in a number of applications, including for approximating weighted Jaccard similarity \cite{WuLiChen:2019}, for near-duplicate detection with weighted features \cite{ManasseMcSherryTalwar:2010}, for approximating the distance between two vectors \cite{Ioffe:2010}, and for sketching image histograms \cite{Shrivastava:2016}. In many of these applications, the weighted sketches empirically outperform unweighted sketches. Weighted MinHash sketches have also been used to compute general ``sum aggregate'' queries, for which the inner product is a special case \cite{CohenKaplan:2013}. However, we are not aware of strong worst-case error guarantees for the above applications, let alone for the problem of general inner product estimation.
\paragraph{Locality Sensitive Hashing} Finally, we note that the problem of estimating inner products from sketches is closely related to maximum inner product search (MIPS) and cosine similarity search. In these problems, the goal is to retrieve vectors from a database with the \emph{highest} inner product (respectively, cosine similarity) with a given query vector. A common approach for solving these search problems is locality sensitive hashing \cite{GionisIndykMotwani:1999}, and there exist methods based on both random projections, like SimHash \cite{Charikar:2002}, and methods based on MinHash. It has been observed that MinHash methods often outperform SimHash for binary data, which parallels what was previously known for binary inner product estimation \cite{ShrivastavaLi:2014a}.
\section{Technical Warmup: Analyzing Unweighted MinHash}
\label{sec:unweighted-minhash}
\label{sec:notation}
\paragraph{Notation}
We first review notation used throughout the paper.
We use bold letters to denote vectors. For a vector $\bv{a}$ we let $\bv{a}[k]$ denote the $k^\text{th}$ entry (indexing starts with $1$). For two length $n$ vectors, $\bv a, \bv b$, we let $\langle \bv{a}, \bv{b}\rangle = \sum_{k=1}^n \bv{a}[k]\bv{b}[k]$ denote the inner product. We let $\|\bv{a}\| =\ \sqrt{\langle \bv{a}, \bv{a}\rangle}$ denote the Euclidean norm of a vector and let $\|\bv{a}\|_{\infty} = \max_{k\in \{1,\ldots, n\}} |\bv{a}[k]|$ denote the infinity norm. We let $\|\bv{a}\|_{1} = \sum_{k=1}^n |\bv{a}[k]|$ denote the $\ell_1$ norm.
As is standard in the literature \cite{BeyerHaasReinwald:2007}, for ease of exposition we assume access to uniformly random hash functions that map to the real line. I.e., we assume that we can construct a random function $h$ such that for any input $j \in \{1, \ldots, n\}$, $h(j)$ is distributed uniformly and independently on the interval $[0,1]$. In practice, $h$ can be replaced with a low-randomness function that map to a sufficient large discrete set $\{1/U, 2/U \ldots, 1\}$. Typically $U$ is chosen to equal $n^c$ for constant $c$ (e.g. $c = 3$) \cite{CormodeGarofalakisHaas:2011}.
We let $\Pr[E]$ denote the probability that a random event $E$ occurs, and $\mathbbm{1}[E]$ denote the indicator random variable that evaluates to $1$ if $E$ occurs and to $0$ otherwise. $\E[X]$ and $\operatorname{Var}[X]$ denote the expectation and variance of a random variable $X$.
\paragraph{An unweighted method} Before introducing our Weighted MinHash sketching method, we review the unweighted MinHash algorithm and prove a inner product estimation bound that can be obtained from this method. The bound closely follows prior work on binary vectors \cite{BeyerHaasReinwald:2007,PaghStockelWoodruff:2014} and only holds under strong assumptions on the sketched vectors $\bv{a}$ and $\bv{b}$ -- specifically that their entries are uniformly bounded in magnitude. Nevertheless, it serves as a warmup for our main result, which is proven using a similar strategy, but eliminates the assumption by using weighted sampling.
Given a vector $\bv{a}$, we obtain an entry in the standard MinHash sketch (see e.g., \cite{Broder:1997}) by hashing the index of every non-zero entry in $\bv{a}$ to the interval $[0,1]$. We then store the smallest hash value. This process is repeated $m$ times with independently chosen random hash functions. For binary vectors $\bv{a}$ and $\bv{b}$ with non-zero index sets $\mathcal{A} = \{k: \bv{a}[k]\neq 0\}$ and $\mathcal{B} = \{k: \bv{b}[k]\neq 0\}$, the minimum hash value alone can be used to estimate the Jaccard similarity $|\mathcal{A}\cap \mathcal{B}|/|\mathcal{A}\cup \mathcal{B}|$ or the union size $|\mathcal{A}\cup \mathcal{B}|$ \cite{FlajoletMartin:1985,BeyerHaasReinwald:2007,KaneNelsonWoodruff:2010}.
\begin{algorithm}[h]\caption{Unweighted MinHash Sketch}\label{alg:minhash_sketch}
\begin{algorithmic}[1]
\Require Length $n$ vector $\bv{a}$, sample number $m$, random seed $s$.
\Ensure Sketch $H_{\bv{a}} = \{H_{\bv{a}}^{hash}, H_{\bv{a}}^{val}\}$, where $H_{\bv{a}}^{hash}$ and $H_{\bv{a}}^{val}$ have length $m$ and contain values in $[0,1]$ and from $\bv{a}$, respectively
\algrule
\State Initialize random number generator with seed $s$.
\For{i = 1, \ldots, m}
\State Select uniformly random hash func. $h^i: \{1,..., n\}\rightarrow [0,1]$.
\State Compute $j^* = \argmin_{j \in \{1, \ldots, n\},\, \bv{a}[j] \neq 0} h^i(j)$. \label{alg:j_star}
\State Set $H_{\bv{a}}^{hash}[i] = h^i(j^*)$ and $H_{\bv{a}}^{val}[i] = \bv{a}[j^*]$
\EndFor
\State \Return $\{H_{\bv{a}}^{hash}, H_{\bv{a}}^{val}\}$
\end{algorithmic}
\end{algorithm}
For non-binary vectors, is it common to \emph{augment} the standard MinHash sketch by also storing the \emph{value} of the index with minimum hash value. This idea is used in ``coordinated sampling'' or ``conditional random sampling'' sketches \cite{CohenKaplan:2013,LiChurchHastie:2006,Cohen:2016}, and was recently used to extend MinHash and the closely related $k$-minimum values (KMV) sketch to estimate vector correlations \cite{SantosBessaChirigati:2021}.
The basic augmented MinHash sketching method is shown in \Cref{alg:minhash_sketch}, which returns $H_{\bv{a}}^{hash}$ and $H_{\bv{a}}^{val}$ as vectors of minimum hashes and their corresponding vector values, respectively.
For any single vector $\bv{a}$, the augmented MinHash sketch $H_\bv{a}$ contains a uniform subsample (collected with replacement) of the non-zero values in $\bv{a}$. This is simply because for every $i \in \{1, \ldots, m\}$ the minimum value of the $i^\text{th}$ hash is equally likely to come from any of the indices with non-zero value. More importantly, the MinHash sketch can be used to obtain a uniform subsample from the \emph{intersection} of $\bv{a}$ and $\bv{b}$, i.e., from entries where both vectors are non-zero. This subsample can in turn be used to estimate the $n$-term sum $\langle \bv{a},\bv{b}\rangle = \sum_{k=1}^n \bv{a}[k]\bv{b}[k]$, since $\bv{a}[k]\bv{b}[k]$ only contributes to the sum if $\bv{a}[k]$ and $\bv{b}[k]$ are both non-zero.
Concretely, we have the following well-known fact:
\begin{fact}
\label{fact:mhash_sketch}
Consider vectors $\bv{a}$ and $\bv{b}$ sketched using \Cref{alg:minhash_sketch} to produce sketches $H_{\bv{a}}$ and $H_{\bv{b}}$. Define the sets $\mathcal{A} = \{i: \bv{a}[i]\neq 0\}$ and $\mathcal{B} = \{i: \bv{b}[i]\neq 0\}$. Then for all
$i\in\{1, \ldots, m\}$ we have:
\begin{enumerate}
\item $H_{\bv{a}}^{hash}[i] = H_{\bv{b}}^{hash}[i]$ with probability $\frac{|\mathcal{A} \cap \mathcal{B}|}{|\mathcal{A} \cup \mathcal{B}|}$.
\item If $H_{\bv{a}}^{hash}[i] = H_{\bv{b}}^{hash}[i]$, then $H_{\bv{a}}^{val}[i] = \bv{a}[j]$ and $H_{\bv{b}}^{val}[i] = \bv{b}[j]$ for $j$ chosen uniformly at random from $\mathcal{A}\cap \mathcal{B}$.
\end{enumerate}
\end{fact}
\Cref{fact:mhash_sketch} indicates that, to obtain a uniform subsample from the intersection of $\bv{a}$ and $\bv{b}$, we simply need to take all entries in $H_{\bv{a}}^{val}$ and $H_{\bv{b}}^{val}$ where the corresponding entries in $H_{\bv{a}}^{hash}$ and $H_{\bv{b}}^{hash}$ are equal -- and, as per (1), they \emph{will be equal} with good probabilty. All other entries of the sketch will be thrown away when estimating the inner product. We give a formal argument for \Cref{fact:mhash_sketch} for completeness.
\begin{proof}
First consider the following equivalent statement: \text{For all } $j \in \mathcal{A}\cap \mathcal{B}$, with probability $\frac{1}{|\mathcal{A}\cup\mathcal{B}|}$, $H_{\bv{a}}^{hash}[i] =H_{\bv{b}}^{hash}[i] = h^i(j)$.
To prove this statement, note that since we hash uniformly to the real interval $[0,1]$, then with probability $1$, if two hash values collide (i.e., if $H_{\bv{a}}^{hash}[i] =H_{\bv{b}}^{hash}[i]$), it must be that they were produced by the same input. In other words, the same $j^*$ (see Algorithm~\ref{alg:minhash_sketch}, line~\ref{alg:j_star}) was chosen when computing $H_{\bv{a}}^{hash}[i]$ and $H_{\bv{b}}^{hash}[i]$. Such a collision can only happen if $j^*$ is contained in $\mathcal{A} \cap \mathcal{B}$ and if $j^*$ minimizes $h^i(j)$ over all $j$ contained in either $\mathcal{A}$ or $\mathcal{B}$ -- i.e., if $j^*$ minimizes $h^i(j)$ over all $j$ in $|\mathcal{A}\cup \mathcal{B}|$. For each $j^*\in \mathcal{A} \cap \mathcal{B}$, the chance of that happening is $1/ |\mathcal{A}\cup \mathcal{B}|$.
\end{proof}
With \Cref{fact:mhash_sketch} in place, we describe an inner product estimator based on MinHash, which is given as \Cref{alg:minhash_est}. This estimator will serve as a template for our weighted MinHash estimator presented in the next section.
\begin{algorithm}[h]\caption{Unweighted MinHash Estimate}\label{alg:minhash_est}
\begin{algorithmic}[1]
\Require Sketches $H_{\bv{a}}= \{H_{\bv{a}}^{hash}, H_{\bv{a}}^{val}\}, H_{\bv{b}}= \{H_{\bv{b}}^{hash}, H_{\bv{b}}^{val}\}$ constructed using \Cref{alg:minhash_sketch} with the same inputs $m,s$.
\Ensure Estimate of $\langle \bv{a}, \bv{b}\rangle$.
\algrule
\State Set $\tilde{U}= \frac{m}{\sum_{i=1}^m\min\left( H_{\bv{a}}^{hash}[i],H_{\bv{b}}^{hash}[i]\right)} - 1$
\State \Return $\frac{\tilde{U}}{m} \sum_{i=1}^m \mathbbm{1}\left[ H_{\bv{a}}^{hash}[i] =H_{\bv{b}}^{hash}[i]\right] \cdot H_{\bv{a}}^{val}[i]\cdot H_{\bv{b}}^{val}[i]$ \label{alg:minhash_est_summation}
\end{algorithmic}
\vspace{-.2em}
\end{algorithm}
Consider the summation in line~\ref{alg:minhash_est_summation} of \Cref{alg:minhash_est}.
Using linearity of expectation and \cref{fact:mhash_sketch}, we can compute the expectation:
\begin{align*}
\E\left[\sum_{i=1}^m \mathbbm{1}\left[ H_{\bv{a}}^{hash}[i]=H_{\bv{b}}^{hash}[i]\right] \cdot H_{\bv{a}}^{val}[i]\cdot H_{\bv{b}}^{val}[i]\right] &= m\cdot \E\left[\mathbbm{1}\left[ H_{\bv{a}}^{hash}[1] =H_{\bv{b}}^{hash}[1]\right] \cdot H_{\bv{a}}^{val}[1]\cdot H_{\bv{b}}^{val}[1]\right]\\ &= m\cdot \sum_{j\in \mathcal{A}\cap \mathcal{B}}\frac{1}{|\mathcal{A}\cup \mathcal{B}|} \bv{a}[j]\bv{b}[j] = \frac{m}{|\mathcal{A}\cup \mathcal{B}|}\cdot \langle \bv{a}, \bv{b}\rangle.
\end{align*}
It follows from the above that, if we multiplied the summation $\sum_{i=1}^m \mathbbm{1}\left[ H_{\bv{a}}^{hash}[i] =H_{\bv{b}}^{hash}[i]\right] \cdot H_{\bv{a}}^{val}[i]\cdot H_{\bv{b}}^{val}[i]$ by $\frac{|\mathcal{A}\cup \mathcal{B}|}{m}$, then we would have an unbiased estimate for $\langle \bv{a}, \bv{b}\rangle$, as desired.
The only catch is that we do not \emph{know} $|\mathcal{A}\cup \mathcal{B}|$. This union size cannot be computed exactly from our sketches $H_\bv{a}$ and $H_{\bv{b}}$. However, it can be \emph{estimated} using the same information contained in our MinHash sketches. In particular, since $h^i$ hashes uniformly to $[0,1]$, $\frac{m}{\sum_{i=1}^m\min\left( H_{\bv{a}}^{hash}[i],H_{\bv{b}}^{hash}[i]\right)} - 1$ provides a good estimate for $|\mathcal{A}\cup \mathcal{B}|$. This is in fact a standard variant of the well-known Flajolet-Martin distinct elements estimator \cite{FlajoletMartin:1985,BeyerHaasReinwald:2007}. In Line 1 of \cref{alg:minhash_est}, we set $\tilde{U}$ equal to
this estimator
and we multiply by $\frac{\tilde{U}}{m}$ in Line 2 as a surrogate for $\frac{|\mathcal{A}\cup \mathcal{B}|}{m}$. This gives our final estimator for $\langle \bv{a}, \bv{b}\rangle$.
The accuracy of the union size estimator can be guarantee using the following result from prior work:
\begin{lemma}[Union Size Estimator \cite{BlumHopcroftKannan:2020}]
\label{lem:distinct_elem}
Let $\mathcal{A}$ and $\mathcal{B}$ be non-empty subsets of $\{1, \ldots, n\}$ and let $h^1, \ldots, h^m: \{1, \ldots, n\}\rightarrow [0,1]$ be independent, uniform random hash functions.
For any $\epsilon, \delta \in (0,1)$, if $m = O \left (\frac{1}{\delta\epsilon^2}\right )$, then with prob. at least $1-\delta$, the estimator $\tilde{U}= \frac{m}{\sum_{i=1}^m\min_{j\in \mathcal{A}\cup\mathcal{B}} h^i(j)} - 1$ satisfies:
\begin{align*}
(1-\epsilon)|\mathcal{A}\cup \mathcal{B}| \leq \tilde{U} \leq (1+\epsilon)|\mathcal{A}\cup \mathcal{B}|.
\end{align*}
\end{lemma}
Note that, while it is written in a slightly different way, the $\tilde{U}$ in \Cref{lem:distinct_elem} is exactly equivalent to the $\tilde{U}$ in \Cref{alg:minhash_est} (when $\mathcal{A}$ and $\mathcal{B}$ contain the non-zero indices of $\bv{a}$ and $\bv{b}$). To see why this is the case, note that $H_{\bv{a}}^{hash}[i] = \min_{j\in \mathcal{A}} h^i(j)$ and $H_{\bv{b}}^{hash}[i] = \min_{j\in \mathcal{B}} h^i(j)$. So $\min\left( H_{\bv{a}}^{hash}[i],H_{\bv{b}}^{hash}[i]\right) = \min_{j\in \mathcal{A}\cup\mathcal{B}} h^i(j)$.
With \Cref{lem:distinct_elem} in place, we are able to prove the following concentration bound for the overall inner product estimator in \Cref{alg:minhash_est} when applied to any pair of \emph{bounded vectors}:
\begin{theorem}[Intermediate Result: Inner Product Sketching with Unweighted MinHash]
\label{thm:bounded}
Let $\epsilon,\delta \in (0,1)$ be accuracy and failure probability parameters and let $m = O(\log(1/\delta)/\epsilon^2)$. There is an algorithm $\mathcal{S}$ that produces size-$m$ sketches (\Cref{alg:minhash_sketch}), along with an estimation procedure $\mathcal{F}$, such that for any $\bv a,\bv b \in \R^n$ with entries bounded in $[-c,c]$, with probability at least $1-\delta$,
\begin{align*}
\left | \mathcal{F}\left(\mathcal{S}(\bv{a}), \mathcal{S}(\bv{b})\right) - \langle \bv{a}, \bv{b} \rangle\right | \leq \epsilon \cdot c^2 \cdot \sqrt{\max(|\mathcal{A}|, |\mathcal{B}|)\cdot |\mathcal{A}\cap \mathcal{B}|}
\end{align*}
for $\mathcal{A} = \{i: \bv{a}[i] \neq 0\}$ and $\mathcal{B} = \{i: \bv{b}[i] \neq 0\}$.
\end{theorem}
Note that for binary vectors, the constant $c$ above equals $1$, so \Cref{thm:bounded} exactly recovers the bounds from prior work \cite{PaghStockelWoodruff:2014}.
\begin{proof}
Let $\overline{\mathcal{F}}(H_{\bv{a}}, H_{\bv{b}})$ denote the estimator from \Cref{alg:minhash_est}. Ultimately we will set $\mathcal{F}$ in \Cref{thm:bounded} to be $\overline{\mathcal{F}}$, but repeated $O(\log(1/\delta)$ times to obtain failure probability $1-\delta$.
We focus on showing first that $\overline{\mathcal{F}}(H_{\bv{a}}, H_{\bv{b}})$ achieves error $\epsilon \cdot c^2 \cdot \sqrt{\max(|\mathcal{A}|, |\mathcal{B}|)\cdot |\mathcal{A}\cap \mathcal{B}|}$ with probability $\geq 2/3$.
To prove this, let $\mathcal{F}^*(H_{\bv{a}}, H_{\bv{b}})$ be an alternative idealized estimator where we replace $\tilde U$ in line 1 of \Cref{alg:minhash_est} with the true union size $U = |\mathcal{A}\cup \mathcal{B}|$:
\begin{align*}
\mathcal{F}^*(H_{\bv{a}}, H_{\bv{b}}) = \frac{U}{m} \sum_{i=1}^m \mathbbm{1}\left[ H_{\bv{a}}^{hash}[i] =H_{\bv{b}}^{hash}[i]\right] \cdot H_{\bv{a}}^{val}[i]\cdot H_{\bv{b}}^{val}[i].
\end{align*}
We will first analyze $\mathcal{F}^*$, before showing that $\overline{\mathcal{F}}$ obtains essentially as good of an estimate.
As established in \Cref{sec:unweighted-minhash}, using the properties of \Cref{fact:mhash_sketch}, we have that:
\begin{align*}
&\E\left[\mathcal{F}^*(H_{\bv{a}}, H_{\bv{b}})\right] = U\cdot \frac{1}{|\mathcal{A}\cup \mathcal{B}|}\cdot \langle \bv{a}, \bv{b}\rangle = \langle \bv{a}, \bv{b}\rangle.
\end{align*}
So we turn to bounding the variance of the estimator. Define the random variable $Z_i = \mathbbm{1}\left[H_{\bv a}^{hash}[i] = H_{\bv b}^{hash}[i]\right]\cdot H_{\bv a}^{val}[i] \cdot H_{\bv b}^{val}[i]$ and note that $\mathcal{F}^*(H_{\bv{a}}, H_{\bv{b}}) = \frac{U}{m}\sum_{i=1}^m Z_i$.
From \Cref{fact:mhash_sketch} we have:
\begin{align*}
Z_i = \begin{cases}
0 &\text{with probability } 1 - \frac{|\mathcal{A}\cap \mathcal{B}|}{|\mathcal{A}\cup \mathcal{B}|}\\
\bv{a}[j] \bv{b}[j] &\text{with probability } \frac{1}{|\mathcal{A}\cup \mathcal{B}|} \text{ for all } j \in \mathcal{A}\cap \mathcal{B}.
\end{cases}
\end{align*}
Since each $Z_i$ is independent, we can bound:
\begin{align*}
\operatorname{Var}\left[\mathcal{F}^*(H_{\bv{a}}, H_{\bv{b}}) \right] = \frac{U^2}{m^2} \sum_{i=1}^m\operatorname{Var}\left[Z_i \right],
\end{align*}
and using our assumption that $\bv{a}[k],\bv{b}[k]\leq c$ for all $k$, for each $Z_i$ we have:
\begin{align*}
\operatorname{Var}\left[Z_i\right] \leq \E\left[Z_i^2\right] = \sum_{j\in \mathcal{A}\cap\mathcal{B}} \frac{1}{|\mathcal{A}\cup \mathcal{B}|} \cdot \bv{a}[j]^2 \bv{b}[j]^2 \leq c^4 \cdot \frac{|\mathcal{A}\cap \mathcal{B}|}{|\mathcal{A}\cup \mathcal{B}|}.
\end{align*}
So we conclude that
\begin{align*}
\operatorname{Var}\left[\mathcal{F}^*(H_{\bv{a}}, H_{\bv{b}}) \right] \leq \frac{1}{m} \cdot c^4 \cdot |\mathcal{A}\cap \mathcal{B}||\mathcal{A}\cup \mathcal{B}|.
\end{align*}
Plugging our expectation and variance bounds into Chebyhev's inequality, we then have that, if $m = O(1/\epsilon^2)$, with probability at least $5/6$,
\begin{align}
\label{eq:union1_unw}
\left |\mathcal{F}^*(H_{\bv{a}}, H_{\bv{b}}) - \langle \bv{a},\bv{b}\rangle\right | \leq \epsilon \cdot c^2 \sqrt{|\mathcal{A}\cap \mathcal{B}||\mathcal{A}\cup \mathcal{B}|}.
\end{align}
The proof is almost complete -- we just need to extend this bound to the non-idealized estimator $\overline{\mathcal{F}} = \frac{\tilde{U}}{U} \cdot \mathcal{F}^*$. We do so by observing that $\tilde{U}$ is a good approximation to $U$. Specifically, by \Cref{lem:distinct_elem} applied with $\delta = 1/6$, we have that, when $m = O(1/\epsilon^2)$, $(1-\epsilon) U \leq \tilde{U} \leq (1+\epsilon) U$,
with probability $\ge 5/6$.
It follows that
\begin{align}
\label{eq:union2_unw}
(1-\epsilon) \mathcal{F}^*(H_{\bv a},H_{\bv b})\leq{\overline{\mathcal{F}}(H_{\bv a},H_{\bv b})} \leq (1+\epsilon)\mathcal{F}^*(H_{\bv a},H_{\bv b}).
\end{align}
By a union bound, with probability at least $2/3$, both \eqref{eq:union1_unw} and \eqref{eq:union2_unw} hold simultaneously. Finally, by triangle inequality and the fact that $\langle a,b\rangle \leq c^2 |\mathcal{A}\cap \mathcal{B}| \leq c^2 \sqrt{|\mathcal{A}\cap \mathcal{B}||\mathcal{A}\cup \mathcal{B}|}$ it follows that:
\begin{align*}
|\overline{\mathcal{F}}(H_{\bv a},H_{\bv b})- \langle \bv{a},\bv{b}\rangle | \leq 3\epsilon\cdot c^2\cdot \sqrt{|\mathcal{A}\cap \mathcal{B}||\mathcal{A}\cup \mathcal{B}|}.
\end{align*}
Noting that $|\mathcal{A}\cap \mathcal{B}||\mathcal{A}\cup \mathcal{B}| \leq 2 \max(|\mathcal{A}|, |\mathcal{B}|)\cdot |\mathcal{A}\cap \mathcal{B}|$ and adjusting $\epsilon$ by a constant factor, we thus have that when $m = O(1/\epsilon^2)$, $\overline{\mathcal{F}}(H_{\bv a},H_{\bv b})$ satisfies the guarantee of \Cref{thm:bounded} with probability at least $2/3$.
To boost success probability to $1-\delta$, we can use the exact same median-trick used in the proof of \Cref{thm:main}: instead of computing a single pair of sketches $H_\bv{a}, H_{\bv b}$ for inputs $\bv{a},\bv{b}$, we concatenate $O(\log(1/\delta))$ sketches, each constructed using an independent random seed. If we apply $\overline{\mathcal{F}}$ to each pair of independent sketches and return the median estimate for $\langle \bv a,\bv b\rangle$, with probability at least $1-\delta$, it will satisfy our desired guarantee.
\end{proof}
\section{Our Main Result: Weighted MinHash Sampling}
\label{sec:main_result}
The main technical challenge in our work is extending the results of the previous section (\Cref{thm:bounded}) to vectors whose entries have \emph{highly varying magnitude}. It is not hard to see that the simple MinHash method fails for such vectors. For example, consider the extreme case when $\bv{a}$ and $\bv{b}$ both contain a very large values at some index $i$, so large that the term $\bv{a}[i]\bv{b}[i]$ dominates the inner product $\langle \bv{a},\bv{b}\rangle = \sum_{k=1}^n \bv{a}[k]\bv{b}[k]$. To correctly approximate the inner product, we \emph{need} to include $\bv{a}[i]$ and $\bv{b}[i]$ in our sketches for $\bv{a}$ and $\bv{b}$, respectively. A MinHash sketch will only do so with low probability, since it uniformly samples entries from the intersection of the vectors. Thus, it will obtain a poor estimate for $\langle \bv{a},\bv{b}\rangle$.
To avoid the issue with {heavy} entries, we modify the approach of \Cref{sec:unweighted-minhash} to incorporate \emph{non-uniform} sampling weights using a Weighted MinHash sketch \cite{ManasseMcSherryTalwar:2010}. This allows us to sample high-magnitude entries in the vectors with higher probability. Specifically, our goal is to sample the $i^\text{th}$ entry of $\bv{a}$ with probability proportional to the squared magnitude, $\bv{a}[i]^2$, so larger magnitude entries are sampled with higher probability.
The Weighted MinHash sketch achieves non-uniform sampling in a simple way: we construct an \emph{extended} vector $\bv{\bar{a}}$ which has the same entries as $\bv{a}$, but entries are repeated multiple times, with the exact number of repetitions proportional to their magnitude. We then apply the standard MinHash sketch to $\bv{\bar a}$. This approach is detailed in \Cref{alg:weighted_sketch}. \break\break
\begin{algorithm}[h]\caption{Weighted MinHash Sketch}\label{alg:weighted_sketch}
\begin{algorithmic}[1]
\Require Length $n$ vector $\bv{a}$, sample number $m$, random seed $s$, integer discretization parameter $L$.
\Ensure Sketch $W_{\bv{a}} = \{W_{\bv{a}}^{hash}, W_{\bv{a}}^{val}, \|\bv{a}\|\}$, where $W_{\bv{a}}^{hash}$ is a length $m$ vector of values in $[0,1]$, $W_{\bv{a}}^{val}$ is a length $m$ vector containing a subset of entries from $\bv{a}$, and $\|\bv{a}\|$ is a scalar, the Euclidean norm of $\bv a$.
\algrule
\State Initialize random number generator with seed $s$.
\State Set $\tilde{\bv{a}} = \textsc{Round}(\bv{a}/\|\bv{a}\|$, L) using \cref{alg:round}.
\State For each $i \in \{1, \ldots, n\}$, let $\bar{\bv{a}}^{(i)}$ be a length $L$ vector whose first $\tilde{\bv{a}}[i]^2 \cdot L$ entries are set to $\tilde{\bv a}[i]$. Set the remaining entries to $0$.
\State Let $\bv{\bar{a}} = [\bar{\bv{a}}^{(1)}, \ldots, \bar{\bv{a}}^{(n)}]$ be a length $n\cdot L$ vector obtained by concatenating the vectors defined above.
\For{i = 1, \ldots, m}
\State Select uniform random hash func. $h^i: \{1,..., nL\}\rightarrow [0,1]$.
\State Compute $j^* = \argmin_{j \in \{1, \ldots, n\cdot L\},\, \bar{\bv{a}}[j] \neq 0} h^i(j)$.
\State Set $W_{\bv{a}}^{hash}[i] = h^i(j^*)$ and $W_{\bv{a}}^{val}[i] = \bar{\bv{a}}[j^*]$.
\EndFor
\State \Return $\{W_{\bv{a}}^{hash}, W_{\bv{a}}^{val}, \|\bv{a}\|\}$.
\end{algorithmic}
\end{algorithm}
\vspace{-2em}
\begin{algorithm}[h]\caption{Vector Rounding for Weighted MinHash}\label{alg:round}
\begin{algorithmic}[1]
\Require Length $n$ unit vector $\bv{z}$, integer discretization parameter $L$.
\Ensure Length $n$ unit vector $\bv{\tilde z}$ with $\bv{\tilde z}[i]^2$ an integer multiple of $1/L$ for all $i$.
\algrule
\State For all $i \in \{1, \ldots, n \}$, $\bv{\tilde z}[i] = \sign(\bv{z}[i]) \cdot \sqrt{\frac{\lfloor {\bv{z}}[i]^2 \cdot L \rfloor}{L}}$.
\State Let $i^* = \argmax_{i \in 1,\ldots,n} |\bv{z}[i]|$.
\State Fix $\delta = 1-\norm{\bv{\tilde z}}^2$, then set $\bv{\tilde z}[i^*] = \sign(\bv z[i]) \cdot \sqrt{\bv{\tilde z}[i^*]^2 + \delta}$.
\State \Return $\bv{\tilde z}$.
\end{algorithmic}
\end{algorithm}
\paragraph{Rounding \& Normalization} While Weighted MinHash allows us to sample entries with non-uniform probability, another challenge arises: since sketches for $\bv{a}$ and $\bv{b}$ are computed independently, we no longer sample with the \emph{same probability} from both vectors. For $\bv{b}$, Weighted MinHash samples indices with probability proportional to $\bv{b}[i]^2$ instead of $\bv{a}[i]^2$. This mismatch can actually \emph{reduce} the probability that we select entries from $\bv{a}$ and $\bv{b}$ with the same index.
We are able to balance this issue with a normalization strategy. In particular, line 2 in \Cref{alg:weighted_sketch} performs a simple but important preprocessing step that \emph{scales} and \emph{rounds} $\bv{a}$ to a unit vector $\bv{\tilde{a}}$ whose squared entries are all integer multiples of $1/L$ for some large integer $L$ (to be chosen later). The rounding handles a minor issue: since we control the frequency with which each entry $\bv{a}[i]$ is sampled by \emph{repetition}, we need the squared value of all entries to be integer multiples of the same fixed constant in order to sample precisely with probability proportional to $\bv{a}[i]^2$. As will be proven, $L$ can be chosen so that the discretization has little impact on the accuracy of our final inner product estimate, and the parameter also has no impact on the size of the sketch returned by \Cref{alg:minhash_sketch}.\footnote{Note that our rounding method (\Cref{alg:round}) is non-standard: It rounds all entries of the input vector {down} to smaller magnitude values, except for the \emph{largest} magnitude entry in the vector, which gets rounded up.
This scheme allows us to achieve small \emph{relative error} when rounding
and to avoid additive error depending on $1/L$.}
The scaling is what deals with the bigger issue discussed above -- the mismatch in sampling probabilities between $\bv{a}$ and $\bv{b}$. Surprisingly, we can show that the impact of this mismatch can be controlled when $\|\bv{a}\| = \|\bv{b}\|$. So while it is possible to come up with examples where the algorithm fails if we directly sketch $\bv{a}$ and $\bv{b}$, we can obtain a worst-case bound by sketching $\bv{a}/\|\bv{a}\|$ and $\bv{b}/\|\bv{b}\|$, approximating $\langle \bv{a}/\|\bv{a}\|, \bv{b}/\|\bv{b}\| \rangle$, and then post-multiplying the result by $\|\bv{a}\|\|\bv{b}\|$ to get our final estimator.
\begin{algorithm}[h]\caption{Weighted MinHash Estimate}\label{alg:weight_est}
\begin{algorithmic}[1]
\Require Sketches $W_{\bv{a}} = \{W_{\bv{a}}^{hash}, W_{\bv{a}}^{val}, \|\bv{a}\|\}$ and $W_{\bv{b}} = \{W_{\bv{b}}^{hash}, W_{\bv{b}}^{val}, \|\bv{b}\|\}$ constructed using \Cref{alg:weighted_sketch} with the same inputs $m$, $s$, and $L$.
\Ensure Estimate of $\langle \bv{a}, \bv{b}\rangle$.
\algrule
\State For $i \in \{1, \ldots, m \}$, set $q_i = \min\left(W_{\bv{a}}^{val}[i]^2,W_{\bv{b}}^{val}[i]^2\right)$.
\State Set $\tilde{M} = \frac{1}{L}\cdot \left(\frac{m}{\sum_{i=1}^m\min\left(W_{\bv{a}}^{hash}[i],W_{\bv{b}}^{hash}[i]\right)} - 1\right)$.
\State Set $I = \frac{\tilde{M}}{m} \sum_{i=1}^m \mathbbm{1}\left[W_{\bv{a}}^{hash}[i] = W_{\bv{b}}^{hash}[i]\right]\cdot \frac{W_{\bv{a}}^{val}[i] \cdot W_{\bv{b}}^{val}[i] }{q_i}.$
\State \Return $\|\bv{a}\|\|\bv{b}\|\cdot I$ \label{alg_weight_est_return}
\end{algorithmic}
\end{algorithm}
\paragraph{Deriving the Inner Product Estimator}
With our sketching method described, we next motivate \Cref{alg:minhash_est}, which is the algorithm used to estimate $\langle \bv{a}, \bv{b}\rangle$ from the sketches.
Note that our Weighted MinHash Sketch (\cref{alg:weighted_sketch}) in fact returns an Unweighted MinHash Sketch (\cref{alg:minhash_sketch}) for the expanded vectors $\bv{\bar a}, \bv{\bar b}$. So, we can apply \Cref{fact:mhash_sketch} to obtain the following:
\begin{fact}
\label{fact:mhash_sketchWeighted}
Consider vectors $\bv{a}$ and $\bv{b}$ sketched using \Cref{alg:weighted_sketch} to produce $W_{\bv{a}}$ and $W_{\bv{b}}$. Define $\mathcal{A}$ and $\mathcal{B}$ as in \Cref{fact:mhash_sketch}. For all $i\in\{1, \ldots, m\}$ we have:
\begin{enumerate}
\item $W_{\bv{a}}^{hash}[i] = W_{\bv{b}}^{hash}[i]$ with probability $\bar J$ where $\bar J$ is the
weighted Jaccard similarity $\bar{J} = \frac{\sum_{j=1}^n \min(\tilde{\bv a}[j]^2,\tilde{\bv b}[J]^2)}{\sum_{j =1}^n \max(\tilde{\bv a}[j]^2,\tilde{\bv b}[j]^2)}$.
\item If $W_{\bv{a}}^{hash}[i] = W_{\bv{b}}^{hash}[i]$, then we have that $W_{\bv a}^{val}= \bv{\tilde a}[j]$ and $W_{\bv b}^{val}= \bv{\tilde b}[j]$ for $j$ chosen from $\mathcal{A} \cap \mathcal{B}$ with probability equal to
$\min(\tilde{\bv a}[j]^2,\tilde{\bv b}[j]^2)/\sum_{i\in \mathcal{A}\cap \mathcal B} \min(\tilde{\bv a}[j]^2,\tilde{\bv b}[j]^2)$.
\end{enumerate}
\end{fact}
A formal proof of \Cref{fact:mhash_sketchWeighted} is given in \cref{app:weighted_proof}. With the statement in place, in \Cref{alg:weight_est} we present our procedure for estimating $\langle \bv{a}, \bv{b}\rangle$ based the Weighted MinHash sketches computed by \Cref{alg:weighted_sketch}. This procedure is reminiscent of our estimator for unweighted sketches from the previous section.
The only difference is that, since we are sampling with non-uniform probabilities, we need to inversely weight samples in our sum to keep everything correct in expectation.
In particular, consider the sum
in line 3 of the algorithm. By \cref{fact:mhash_sketchWeighted} and linearity of expectation, we have that:
\begin{align*}
\E\left[\sum_{i=1}^m \mathbbm{1}\left[W_{\bv{a}}^{hash}[i] = W_{\bv{b}}^{hash}[i]\right]\cdot \frac{W_{\bv{a}}^{val}[i] \cdot W_{\bv{b}}^{val}[i] }{q_i}\right] &=
m\cdot \E\left[\mathbbm{1}\left[W_{\bv{a}}^{hash}[i] = W_{\bv{b}}^{hash}[i]\right]\right]\cdot \frac{W_{\bv{a}}^{val}[i] \cdot W_{\bv{b}}^{val}[i] }{q_i}\\
&= m\cdot \sum_{j\in \mathcal{A}\cap \mathcal{B}} \frac{q_i}{\sum_{i=1}^n \max(\tilde{\bv a}[j]^2,\tilde{\bv b}[j]^2)} \frac{\tilde{\bv a}[j]\tilde{\bv b}[j]}{q_i} \\
&= \frac{m}{\sum_{i=1}^n \max(\tilde{\bv a}[j]^2,\tilde{\bv b}[j]^2)}\cdot \langle \tilde{\bv a}, \tilde{\bv b}\rangle.
\end{align*}
So, we have obtained an estimator that in expectation is equal to $\langle \tilde{\bv a}, \tilde{\bv b}\rangle$, multiplied by $m$ over another term: $M =\sum_{i=1}^n \max(\tilde{\bv a}[j]^2,\tilde{\bv b}[j]^2)$. This term $M$ is referred to as the \emph{weighted} union size between the vectors. We can multiply by $\frac{M}{m}$ to obtain an unbiased estimator for $\langle \tilde{\bv a}, \tilde{\bv b}\rangle$. Since $\tilde{\bv{a}}$ and $\tilde{\bv{b}}$ were obtained by scaling $\bv{a}$ and $\bv{b}$ inversely by their Euclidean norms (ignoring the effect of rounding for now), our final estimator in Line~\ref{alg_weight_est_return} of \Cref{alg:weight_est} multiplies by $\|\bv{a}\|\|\bv{b}\|$. The values of $\|\bv{a}\|$ and $\|\bv{b}\|$ are stored explicitly in the sketches for $\bv{a}$ and $\bv{b}$, respectively (as just one extra number per sketch).
The formal analysis of \Cref{alg:weight_est}, which yields \Cref{thm:main}, is included in \Cref{app:weighted_proof}.
It contains three parts. First, when analyzing the unweighted estimator, we do not know $M$ exactly, so must estimate it. We can take advantage of the fact that $M$ is exactly equal to the \emph{unweighted} union size $|\bar{\mathcal{A}} \cup \bar{\mathcal{B}}|$ between the non-zero index sets $\bar{\mathcal{A}}$ and $\bar{\mathcal{B}}$ of the \emph{expanded vectors} $\bar{\bv{a}}$ and $\bar{\bv{b}}$ constructed in \Cref{alg:weighted_sketch}. We can apply \Cref{lem:distinct_elem} directly to obtain an estimator, which is denoted as $\tilde{M}$ in \Cref{alg:weight_est}.
Second, we need to analyze the variance of the sum $\sum_{i=1}^m \mathbbm{1}\left[W_{\bv{a}}^{hash}[i] = W_{\bv{b}}^{hash}[i]\right]\cdot \frac{W_{\bv{a}}^{val}[i] \cdot W_{\bv{b}}^{val}[i] }{q_i}$. This analysis
uses the fact that
$\tilde{\bv{a}}$ and $\tilde{\bv{b}}$ are unit vectors. Third, we need to rigorously analyze the impact of the rounding procedure performed in Line 2 of \Cref{alg:weighted_sketch} to establish that a good estimate for $\langle \tilde{\bv a}, \tilde{\bv b}\rangle$ actually yields a good estimate for $\langle \bv{a}/\|\bv{a}\|, \bv{b}/\|\bv{b}\|\rangle = \frac{1}{\|\bv{a}\|\|\bv{b}\|} \langle \bv{a}, \bv{b}\rangle$.
We conclude by noting that our final analysis of \Cref{alg:weight_est} requires setting $L$ to be on the order of $n^6/\epsilon^2$ when sketching using \Cref{alg:weighted_sketch}. This may sound large, but note that the parameter has \emph{no impact} on the size of the sketches returned by \Cref{alg:weighted_sketch}, or on the runtime of our estimation procedure \Cref{alg:weight_est}. $L$ does impact the {runtime} of \Cref{alg:weighted_sketch}, but as discussed in \cref{sec:exp}, prior work can be used to implement the Weighted MinHash sketching method so that it has a \emph{logarithmic} dependence on $L$ -- i.e., on $O(\log(n/\epsilon))$.
\section{Experiments}
\label{sec:exp}
To support the results presented in \cref{sec:main_result},
we performed an experimental evaluation using synthetic data and real-world datasets.
\paragraph{Baselines}
We compare our Weighted MinHash approach against 4 baseline methods, 2 linear sketching based and 2 sampling-based, with the goal of evaluating the trade-off between sketch size and accuracy in estimating inner products. Those methods are:
\smallskip \vspace{-.5em}
\begin{description}[style=unboxed,leftmargin=.5cm]
\item[\em Johnson-Lindenstrauss Projection (JL):] equivalent to the AMS sketch \cite{AlonMatiasSzegedy:1999,Achlioptas:2003}. Uses a random matrix $\bs{\Pi}$ with scaled $\pm 1$ entries (\Cref{fact:jl_result}).
\item[\em CountSketch (CS):] classic linear sketch introduced in \cite{CharikarChenFarach-Colton:2002}, and corresponds to multiplication with a $\bs{\Pi}$ that has sparse random entries. We follow the implementation in \cite{LarsenPaghTetek:2021}, using $5$ repetitions of the sketch and taking the median to improve performance.
\item[\em MinHash Sampling (MH):] method described in \Cref{alg:minhash_sketch}; we use a single sketch without any median estimate.
\item[\em $k$-Minimum Values Sampling (KMV):] sampling-based sketch closely related to MinHash, but it draws samples from the vector being sketched \emph{without replacement}. It can also be used to estimate union size. We follow the implementations from \cite{BeyerHaasReinwald:2007} and \cite{SantosBessaChirigati:2021}.
\item[\em Weighted MinHash Sampling (WMH):] our method described in \Cref{alg:weighted_sketch}; we use a single sketch without any median estimate.
\end{description}
\paragraph{Storage Size}
We use standard implementations for all sketches.
For linear sketches, we store the output of the matrix multiplication $\bs{\Pi}\bv{a}$ as 64-bit doubles. We also store $W_\bv{a}^{val}$ and $H_\bv{a}^{val}$ as 64-bit doubles. Since sampling-based sketches need to store hash values (which in our case are 32-bit ints), a sampling-based sketch with $m$ samples takes $1.5x$ as much space as a JL sketch with $m$ rows. In our experiments, we plot \emph{storage size} which denotes the total number of bits in the sketch divided by 64, i.e., the total number of 64-bit doubles (or equivalent) used in the sketch. Standard quantization tricks could possibly be used to reduce the bit complexity of store numbers in all sketches (linear and sampling) \cite{Jacques:2015, LiMitzenmacherSlawski:2016}, but evaluating such methods is beyond the scope of our work.
\paragraph{Estimation Error} For all plots, we report the absolute difference between $\langle \bv{a}, \bv{b}\rangle$ and the estimate, divided by $\|\bv{a}\|\|\bv{b}\|$. This is the term appearing on the right-hand side of the accuracy guarantee for linear sketches \Cref{fact:jl_result}, so this scaling roughly ensures that errors are between $0$ and $1$,
making it easier to compare across different datasets. We always report average error over $10$ independent trials.
\paragraph{Choice of $L$}
Note that the choice of $L$ in \Cref{alg:weighted_sketch} does not impact the size of our final sketch, so in general, it should be set as large as possible. While our bounds from \Cref{lem:round} that suggest $L$ should be set on the order of $n^6$ are likely to lose (we did not attempt to optimize polynomial factors), in practice, we found that it \emph{is necessary} to ensure at least that $L > n$, and ideally it should be larger by a multiplicative factor 100 or 1000. The reason for this is that when we normalize the input $\bv{a}$ of \Cref{alg:minhash_sketch} to have a unit norm, if $\bv{a}$ is dense, \emph{most} of its entries could have squared value $< 1/n$ (as the average value of a squared entry in a unit norm vector is always $1/n$). If we set $L < 1/n$, then any entries with value $<1/n$ would get rounded to $0$, which could negatively impact the accuracy of an inner product estimate. We fix the parameter $L = 10^7$ in all experiments.
\paragraph{Efficient Weighted Hashing}
When $L$ is large, a {naive} implementation of \Cref{alg:weighted_sketch} would be prohibitively slow. In particular, the ``extended'' vector $\bar{\bv{a}}$ has length $n\cdot L$, and we must apply a hash function to every non-zero entry in that vector. Let $\mathcal{A} = \{i: \bv{a}[i]\neq 0\}$ as before, so $|\mathcal{A}|$ is equal to the number of non-zero values in $\bv{a}$. If each hash computation is considered unit cost, this amounts to a runtime of $O(|\mathcal{A}|m\cdot L)$, which is too large since $L$ is chosen larger than $n$.
Fortunately, it is possible to improve this cost to $O(|\mathcal{A}|m\cdot \log L) = O(|\mathcal{A}|m\cdot \log n) $ using techniques for speeding up weighted MinHash sketches. Such techniques have been heavily studied in recent years \cite{Ioffe:2010,WuLiChen:2019,HaeuplerManasseTalwar:2014,Shrivastava:2016}. The savings are significant, reducing the computation cost of sketching to nearly linear in the input size for each of our $m$ samples.
Among faster methods, we suggest specifically employing the simple ``active index'' technique, which was first introduced in \cite{GollapudiPanigrahy:2006}. The rough idea is that when hashing non-zero entries in a particular length $L$ block of $\bv{\bar{a}}$, there is no need to hash all non-zero indices in that block. We can skip over large sections of indices by observing that if $z$ is the minimum hash value generated so far, the next index where a lower hash value will be seen is distributed as a \emph{geometric random variable} with parameter $z$. We can efficiently sample from the geometric distribution (e.g., using a built-in Python routine) and skip ahead to that index. It is possible to prove that the expected cost of this approach is just $O(\log L)$ per block. See the exposition in \cite{ManasseMcSherryTalwar:2010}
for further details.
\paragraph{Choice of Hash Function} In practice, we cannot obtain a truly uniform random hash function from $\{1,\ldots, n\}$ to the reals, so we must use an approximation. In our experiments, we employ a standard 2-wise independent hash function (linear function with random coefficients) that maps from $\{1, \ldots, n\}$ to $\{1, \ldots, p\}$ for a 31-bit prime $p$ \cite{CarterWegman:1979}. \footnote{Our choice to use a 2-wise independent hash function was based on prior implementations of the weighted MinHash method \cite{WuLiChen:2020} that do so.} We then use as our hash value $h(i)/p$, which is a number between $0$ and $1$. Since $p$ is chosen to have 31 bits, we can store the value of $h(i)$ in our sketch using a standard 32-bit int.
\begin{figure}[t]
\centering
\begin{subfigure}{.6\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/legend_ip.png}
\end{subfigure}
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/ip_diff_overlap_0.01_outlier_0.1_corr_0.98.png}
\vspace{-1.8em}
\caption{1\% overlap}
\end{subfigure}
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/ip_diff_overlap_0.05_outlier_0.1_corr_0.98.png}
\vspace{-1.8em}
\caption{5\% overlap}
\end{subfigure}%
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/ip_diff_overlap_0.1_outlier_0.1_corr_0.98.png}
\vspace{-1.8em}
\caption{10\% overlap}
\end{subfigure}
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/ip_diff_overlap_0.5_outlier_0.1_corr_0.98.png}
\vspace{-1.8em}
\caption{50\% overlap}
\end{subfigure}
\vspace{-.5em}
\caption{Inner product estimation (synthetic data).}
\vspace{-1em}
\label{fig:InnerProduct_Corr098_Est}
\vspace{-.2em}
\end{figure}
\vspace{-.2cm}
\subsection{Synthetic Data}
We begin with an evaluation of our approach using synthetic data, so that we can control different parameters.
We generate length $10000$ vectors $\bv{a}$ and $\bv{b}$, each with $2000$ non-zero entries. The ratio of non-zero entries that \emph{overlap}, i.e., are non-zero in both $\bv{a}$ and $\bv{b}$, is adjusted to simulate different practical settings with different levels of \emph{joinability} between tables (see \Cref{sec:apps}). The non-zero entries in $\bv{a}$ and $\bv{b}$ are normal random variables with values between $-1$ and $1$,
except $10\%$ of entries are chosen randomly as outliers and set to random values between $20$ and $30$.
Results for varying amounts of overlap are reported in \Cref{fig:InnerProduct_Corr098_Est}. They closely align with our theoretical findings: when the overlap is small, the bounds for Weighted MinHash are significantly better than those of linear sketching methods. WMH outperforms all other methods for overlap ratio $\leq 10\%$. Note that unweighted sampling based sketches also outperform linear sketches for very low overlap ($1\%$). But as the overlap increases, the advantage brought about by \Cref{thm:main} over \Cref{fact:jl_result} decreases. We can see this in \Cref{fig:InnerProduct_Corr098_Est}(d): at 50\% overlap, the performance of linear sketching
is comparable to that of Weighted MinHash.
\subsection{Real-World Data}
\subsubsection{World Bank Data: Assessing the Effect of Table Overlap and Outliers}
Using sketches of storage size 400,\footnote{The size was chosen empirically.
Our goal here is to simulate the real-world situation where a fixed parameter must be selected for a given application.} we estimate the inner product between 5000 random pairs of numerical columns from 53 datasets published by the World Bank Group~\cite{WBF_2022}.
We normalized columns to have norm $1$ so that all inner products have magnitude less than $1$.
\jf{should we add a ref for kurtosis?}
To visualize the results, we use
a winning table (shown in \Cref{fig:real}) that filters results based on different overlap ratios (column) and kurtosis values -- a measure of outliers (row). Each cell shows the average of error difference (WMH estimation error minus the error of other method) of vector pairs filtered by overlap and kurtosis values.\footnote{We provide detailed statistics for overlap ratio distributions in \cref{tab:ip_real_overlap_stats} in \cref{app:ip_real_overlap_stats}.}
The blue cells (negative difference) correspond to combinations in which
WMH outperforms the other methods, while the red cells (positive difference) represent combinations in which the other methods win. The darker the cells, the bigger the difference.
A high kurtosis often indicates the presence of outliers, which will, based on our theoretical results,
present a difficulty for unweighted sampling methods like MinHash in comparison to JL or our Weighted MinHash method. This is supported by the experiments, which show that WMH has a great improvement over MH when kurtosis is high (up to -.031 vs. at most -.020 when kurtosis is low).
As predicted by \Cref{thm:main} and shown in our synthetic experiments, WMH also has a great edge over JL for low overlap values. For large overlaps (greater than .75), JL leads to slightly smaller errors (from 0.003 to 0.006).
This suggests that \emph{WMH provides a good compromise for applications in which the distribution of data is unknown: it provides much better estimates for many cases, and when it does not, its estimates are comparable to
the best results from existing linear and sampling-based sketches}.
\begin{figure}[t]
\centering
\begin{subfigure}{0.46\textwidth}
\centering
\includegraphics[width=.9\textwidth]{figures/realdata_wmh_minus_jl_div.png}
\caption{WMH estimation error minus JL estimation error.}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=.9\textwidth]{figures/realdata_wmh_minus_mh_div.png}
\caption{WMH estimation error minus MH estimation error.}
\end{subfigure}
\caption{Inner product estimation (World Bank data). Different shades of blue highlight combinations for which WMH outperforms the other methods.}
\label{fig:real}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}{.6\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/legend_ip.png}
\end{subfigure}
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/20news_tfidf_all.pdf}
\vspace{-1.8em}
\caption{All Documents}
\end{subfigure}
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.0\linewidth]{figures/20news_tfidf_700len.pdf}
\vspace{-1.8em}
\caption{Documents > 700 words}
\end{subfigure}%
\caption{Text similarity estimation (20 Newsgroups dataset).}
\label{fig:tfidf}
\end{figure}
\subsubsection{20 Newsgroups Dataset: Document Similarity Estimation}
\as{WIP}
We also evaluated the performance of WMH sketches for text similarity estimation.
We represent each document as a vector in which each entry represents a term or a combination of 2 terms (bigrams), and is associated with a value that encodes term/bigram importance using TF-IDF weights~\cite{SaltonWongYang:1975}.
This setting is well-known for generating sparse vectors of very high dimension.
As a similarity measure, we use the \textit{cosine}, which is equal to an inner product when the vectors are normalized using their $l^2$-norm.
We sampled 700 documents from the \textit{20 newsgroups} dataset~\cite{20newsgroups}, and estimated the cosine similarity for
over 200,000 pairs of documents.
The results in \cref{fig:tfidf} show that, similar to previous experiments, \emph{in the worst case, WMH produces estimates that are comparable to the other methods, but it can outperform them -- sometimes by a large margin.} In this case, it does so for documents containing more than 700 words.
\jf{need to explain why!}
Note that linear projection sketches have poor performance for small sketches even when the documents are small, whereas our sampling-based methods are able to obtain significantly better accuracy for the same storage budget.
Finally, also note that the Unweighted MinHash (MH) performs poorly for long documents, whereas the weighted version still performs well.
\paragraph{Acknowledgements}
This work was partially supported by the DARPA D3M program and NSF awards ISS-2106888 and CCF-2046235. Cameron Musco was also supported by a Google Research Scholar Award. Aline Bessa was supported by a 2021 CRA/CCC CIFellows Award.
Any opinions, findings, conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of NSF, DARPA, or other funding organizations.
\bibliographystyle{apalike}
|
\section{Introduction}
In 2015 the Laser Interferometer Gravitational-Wave Observatory (LIGO) Scientific Collaboration and the Virgo Collaboration \cite{Abbott:2016blz,TheLIGOScientific:2016agk} directly observed the first gravitational wave (GW) event GW150914 coming from the coalescence of binary black holes (BHs),
and the discovery opened us a new window to understand the property of BHs and gravity in the nonlinear and strong field regimes.
Until now tens of GW events in the frequency range of tens to hundreds Hertz have been confirmed \cite{LIGOScientific:2018mvr,LIGOScientific:2020ibl,LIGOScientific:2021usb,LIGOScientific:2021djp,LIGOScientific:2017vwq,LIGOScientific:2020aai,LIGOScientific:2021qlt}.
However, due to the seismic noise and gravity gradient noise, the ground-based GW observatories can only measure transient GWs in the frequency range $10-10^3$ Hz, radiated by the coalescences of stellar-mass compact binaries.
Apart from transient GW sources, the future space-based GW detectors like the Laser Interferometer Space Antenna (LISA) \cite{Danzmann:1997hm,Audley:2017drz}, TianQin \cite{Luo:2015ght} and Taiji \cite{Hu:2017mde} will help us uncover unprecedented information about GW sources and fundamental physics \cite{LISA:2017pwj,TianQin:2020hid,Ruan:2018tsw,Amaro-Seoane:2022rxf,LISA:2022kgy}.
One of the most conspicuous sources for the future space-based GW detectors is the extreme mass-ratio inspirals (EMRIs) \cite{Amaro-Seoane:2007osp,Babak:2017tow}.
EMRIs, which consist of a stellar-mass compact object (secondary object) with mass $m_p\sim1-100~M_{\odot}$ such as BHs, neutron stars, white dwarfs, etc. orbiting around a supermassive black hole (SMBH) (primary object) with mass $M\sim10^5-10^7~M_{\odot}$,
with the mass ratio $m_p/M$ in the range of $10^{-7}-10^{-4}$,
radiate millihertz GWs expected to be observed by the future space-based GW detectors.
Future detections of EMRIs with space-based detectors can provide highly precise measurements on source parameters such as the BH masses, spins, etc.
In \cite{Barack:2003fp}, the authors introduced a family of approximate waveforms for EMRIs to make parameter estimation with LISA.
For a typical source of $m_p=10~M_{\odot}$ and $M=10^6~M_{\odot}$ with a signal-to-noise ratio (SNR) of $30$,
LISA can determine the masses of both primary and secondary objects to within a fractional error of $\sim 10^{-4}$, measure the spin of the primary object to within $\sim 10^{-4}$, and localize the source on the sky within $\sim 10^{-3}$ steradians.
The improved augmented analytic kludge model \cite{Chua:2017ujo} provides more accurate and efficient GW waveforms to improve the errors of parameters by one order of magnitude.
Thus, EMRIs can be used to precisely measure the slope of the black-hole mass function \cite{Gair:2010yu} or as standard sirens \cite{Holz:2005df} to constrain cosmological parameters and investigate the expansion history of the Universe \cite{MacLeod:2007jd,Laghi:2021pqk,LISACosmologyWorkingGroup:2022jok}.
The observations of EMRIs can also help us explore gravitational physics.
For example, they can be used to figure out the spacetime structure around the central SMBH to high precision,
allowing us to test if the spacetime geometry is described by general relativity or an alternative theory
and analyze the environments such as dark matter surrounding the central SMBH \cite{Amaro-Seoane:2007osp,eLISA:2013xep,Eda:2013gg,Eda:2014kra,Barausse:2014tra,Yue:2017iwc,Yue:2018vtk,Babak:2017tow,Berry:2019wgg,Hannuksela:2019vip,Destounis:2020kss,Burton:2020wnj,Torres:2020fye,Barausse:2020rsu,Cardoso:2019rou,Maselli:2020zgv,Maselli:2021men,Guo:2022euk,Zhang:2022rfr,Barsanti:2022ana,Cardoso:2021wlq,Dai:2021olt,Jiang:2021htl,Zhang:2022hbt,Gao:2022hho,Gao:2022hsn,Destounis:2022obl,Barsanti:2022vvl,Liang:2022gdk}.
In \cite{Cardoso:2020iji,Liu:2020vsy,Liu:2020bag},
the authors investigated the eccentricity and orbital evolution
of BH binaries under the influence of accretion in addition to the scalar/vector and gravitational radiations, and discussed the competition between radiative mechanisms and accretion effects on eccentricity evolution.
However, these discussions were mainly based on the Newtonian orbit and dipole emission.
A generic, fully-relativistic formalism to study EMRIs in spherically symmetric and non-vacuum BH spacetime was established in \cite{Cardoso:2022whc}.
Considering the secondary object of mass $m_p$ orbiting the galactic BHs (GBHs) immersed in an astrophysical environment,
like an accretion disk or a dark matter halo, the authors found that the relative flux difference at $\omega M=0.02$ between a vacuum and a GBH with the halo mass $M_{\text{halo}}=0.1 M$ and the lengthscale $a_0=10^2 M_{\text{halo}}$ and $10^3 M_{\text{halo}}$ is $\sim 10\%$ and $1\%$, respectively.
The results clearly indicate that EMRIs can constrain smaller-scale matter distributions around GBHs \cite{Cardoso:2022whc}.
According to the no-hair theorem, any BH can be described by three parameters: the mass, angular momentum, and electric charge.
Current observations have not yet been able to confirm the no-hair theorem or the existence of extra fields besides the gravitational fields in modified gravity theories.
The coupling between scalar fields and higher-order curvature invariants can invalidate the no-hair theorem so that BHs can carry scalar charge which depends on the mass of BH \cite{Sotiriou:2013qea,Silva:2017uqg,Doneva:2017bvd,Antoniou:2017acq}.
The possible detection of scalar fields with EMRIs was discussed in \cite{Maselli:2020zgv,Maselli:2021men,Barsanti:2022ana,Guo:2022euk,Zhang:2022rfr,Barsanti:2022vvl}.
SMBHs are usually neutral because of long-time charge dissipation through the presence of the plasma around them or the spontaneous production of electron-positron pairs \cite{Gibbons:1975kk,Eardley:1975kp,1982PhRvD..25.2509H,1982Prama..18..385J,Gong:2019aqa}.
However, the existence of stable charged astrophysical stellar-mass compact objects such as BHs, neutron stars, white dwarfs, etc. in nature remains controversial.
In Refs. \cite{Bekenstein:1971ej,de1995relativistic,deFelice:1999qp,Ivanov:2002jy,Majumdar:1947eu,zhang_influence_1982,Anninos:2001yb,Bonnor:1975gba},
it was shown that there can be a maximal huge amount of charge with the charge-to-mass ratio of the order one for highly compact stars, whose radius is on the verge of forming an event horizon.
The balance between the attractive gravitational force from the matter part and the repulsive force from the electrostatic part is unstable and charged compact stars will collapse to a charged BH due to a decrease in the electric field \cite{Ray:2003gt}.
Also, BHs can be charged through the Wald mechanism by selectively accreting charge in a magnetic field \cite{Wald:1974np}, or by accreting minicharged dark matter beyond the standard model \cite{Cardoso:2016olt}.
The charge carried by compact objects can affect the parameter estimation of the chirp mass and BH merger \cite{Christiansen:2020pnv,Jai-akson:2017ldo}
and the merger rate distribution of primordial BH binaries \cite{Liu:2022wtq,Liu:2020cds}.
The electromagnetic self-force acting on a charged particle in an equatorial circular orbit of Kerr BH was calculated in \cite{Torres:2020fye}.
It showed the dissipative self-force balances with the sum of the electromagnetic flux radiated to infinity and down the BH horizon,
and prograde orbits can stimulate BH superradiance although the superradiance is not sufficient to support floating orbits even at the innermost stable circular orbit (ISCO) \cite{Torres:2020fye}.
The observations of GW150914 can constrain the charge-to-mass ratio of charged BHs to be as high as 0.3 \cite{Bozzola:2020mjx}.
In \cite{Zhang:2022hbt}, the energy fluxes for both gravitational and electromagnetic waves induced by a charged particle orbiting around a Schwarzschild BH were studied.
It was demonstrated that the electric charge leaves a significant imprint on the phase of GWs and is observable with space-based GW detectors.
In this paper, based on the Teukolsky formalism for BH perturbations \cite{Teukolsky:1973ha,Press:1973zz,Teukolsky:1974yv},
we numerically calculate the energy fluxes for both tensor and vector perturbations induced by a charged particle moving in an equatorial circular orbit around a Kerr BH, the orbital evolution of EMRIs up to the ISCO,
and investigate the capabilities to detect vector charge carried by the secondary compact object with space-based GW detectors such as LISA, TianQin, and Taiji.
We apply the methods of faithfulness and Fisher information matrix (FIM) to assess the capability of space-based GW detectors to detect the vector charge carried by the secondary compact object.
The paper is organized as follows.
In Sec. \ref{sec2}, we introduce the model with vector charge and the Teukolsky perturbation formalism.
In Sec. \ref{sec3}, we give the source terms as well as procedures for solving the inhomogeneous Teukolsky equations.
Then we numerically calculate the energy fluxes for gravitational and vector fields using the Teukolsky and generalized Sasaki-Nakamura (SN) formalisms in the background of a Kerr BH.
In Sec. \ref{sec4}, we give the numerical results of energy fluxes falling onto the horizon and radiated to infinity for gravitational and vector fields,
then we use the dephasing of GWs to constrain the charge.
In Sec. \ref{sec5}, we calculate the faithfulness between GWs with and without vector charge and perform the FIM to estimate the errors of detecting vector charge with LISA, TianQin, and Taiji.
Sec. \ref{sec5} is devoted to conclusions and discussions.
In this paper, we set $c=G=M=1$.
\section{Einstein-Maxwell field equations}
\label{sec2}
The simplest model including vector charges is Einstein-Maxwell theory, which is modeled via the massless vector field
\begin{equation}
S=\int d^4x\frac{\sqrt{-g}}{16\pi}\left[R-\frac{1}{4}F^{\mu\nu}F_{\mu\nu}-A_{\mu}J^{\mu}\right]-S_{\text{matter}}(g_{\mu\nu},\Phi),
\end{equation}
where $R$ is the Ricci scalar, $A_\mu$ is a massless vector field, $F_{\mu\nu}=\nabla_{\mu}A_{\nu}-\nabla_{\nu}A_{\mu}$ is the field strength and $\Phi$ is the matter field, $J^\mu$ is the electric current density.
Varying the action with respect to the metric tensor and the vector field yields the Einstein-Maxwell field equations
\begin{equation}
G^{\mu\nu}=8\pi T^{\mu\nu}_p+8\pi T^{\mu\nu}_e,
\end{equation}
\begin{equation}
\nabla_\nu F^{\mu\nu}=4\pi J^\mu,
\end{equation}
where $G_{\mu\nu}$ is the Einstein tensor, $T_p^{\mu\nu}$ and $T_e^{\mu\nu}$ are the particle's material stress-energy tensor and vector stress-energy tensor, respectively.
Since the amplitude of the vector stress-energy $T_e^{\mu\nu}$ is quadratic in the vector field, the contribution to the background metric from the vector field is second order.
For an EMRI system $(m_p\ll M)$ composing of a small compact object with mass $m_p$ and charge-to-mass ratio $q$ orbiting around a Kerr BH with mass $M$ and spin $Ma$, we can ignore the contribution to the background metric from the vector field.
The perturbed Einstein and Maxwell equations for EMRIs are
\begin{equation}
G^{\mu\nu}=8\pi T^{\mu\nu}_p,
\end{equation}
\begin{equation}
\nabla_\nu F^{\mu\nu}=4\pi J^\mu,
\end{equation}
where
\begin{equation}
T^{\mu\nu}_p(x)=m_p\int d\tau~u^{\mu}u^\nu\frac{\delta^{(4)}\left[x-z(\tau)\right]}{\sqrt{-g}},
\end{equation}
\begin{equation}
J^{\mu}(x)=q m_p\int d\tau~ u^{\mu}\frac{\delta^{(4)}\left[x-z(\tau)\right]}{\sqrt{-g}},
\end{equation}
and $u^{\mu}$ is the velocity of the particle.
We use the Newman-Penrose formalism \cite{Newman:1966ub} to study perturbations around a Kerr BH induced by a charged particle with mass $m_p$ and charge $q$.
In Boyer-Lindquist coordinate, the metric of Kerr BHs is
\begin{equation}\label{SBH}
\begin{split}
ds^2=&(1-2r/\varSigma)dt^2+(4ar\sin(\theta)/\varSigma)dtd\varphi-(\varSigma/\Delta)dr^2-\varSigma d\theta^2\\
&-\sin^2{\theta}(r^2+a^2+2a^2r\sin^2{\theta}/\varSigma)d\varphi^2.
\end{split}\end{equation}
where $\varSigma=r^2+a^2\cos^2{\theta}$, and $\Delta=r^2-2r+a^2$. When $a=0$, the metric reduces to the Schwarzschild metric.
Based on the metric \eqref{SBH}, we construct the null tetrad,
\begin{equation}
\begin{split}
\begin{split}
l^\mu&=[(r^2+a^2)/\Delta,1,0,a/\Delta],\\
n^\mu&=[r^2+a^2,-\Delta,0,a]/(2\varSigma),\\
m^\mu&=[ia\sin{\theta},0,1,i/\sin{\theta}]/(2^{1/2}(r+ia\cos{\theta})),\\
\bar m^\mu&=[-ia\sin{\theta},0,1,-i/\sin{\theta}]/(2^{1/2}(r-ia\cos{\theta})).
\end{split}
\end{split}
\end{equation}
The propagating vector field is described by the two complex quantities,
\begin{equation}
\phi_0=F_{\mu\nu}l^\mu m^\nu,\qquad \phi_2=F_{\mu\nu}\bar{m}^\mu n^\nu.
\end{equation}
The propagating gravitational field is described by the two complex Newman-Penrose variables
\begin{equation}
\begin{split}
\psi_0&=-C_{\alpha\beta\gamma\delta}l^\alpha m^\beta l^\gamma m^\delta ,\\
\psi_4&=-C_{\alpha\beta\gamma\delta}n^\alpha \bar{m}^\beta n^\gamma \bar{m}^\delta,
\end{split}
\end{equation}
where $C_{\alpha\beta\gamma\delta}$ is the Weyl tensor.
A single master equation for tensor ($s=-2$) and vector ($s=-1$) perturbations was derived as \cite{Teukolsky:1973ha},
\begin{equation}
\label{TB}
\begin{split}
&\left[\frac{(r^2+a^2)^2}{\Delta}-a^2\sin^2{\theta}\right]\frac{\partial^2\psi}{\partial t^2}+\frac{4ar}{\Delta}\frac{\partial^2\psi}{\partial t\partial\varphi}+\left[\frac{a^2}{\Delta}-\frac{1}{\sin^2{\theta}}\right]\frac{\partial^2\psi}{\partial \varphi^2}\\
&\qquad-\Delta^{-s}\frac{\partial}{\partial r}\left(\Delta^{s+1}\frac{\partial\psi}{\partial r}\right)-\frac{1}{\sin\theta}\frac{\partial}{\partial \theta}\left(\sin\theta\frac{\partial\psi}{\partial \theta}\right)-2s\left[\frac{a(r-1)}{\Delta}+\frac{i\cos\theta}{\sin^2{\theta}}\right]\frac{\partial\psi}{\partial \varphi}\\
&\qquad\qquad\qquad\qquad-2s\left[\frac{(r^2-a^2)}{\Delta}-r-ia\cos\theta\right]\frac{\partial\psi}{\partial t}+(s^2\cot^2\theta-s)\psi=4\pi\varSigma T,
\end{split}
\end{equation}
the explicit field $\psi$ and the corresponding source $T$ are given in Table \ref{source} \cite{Teukolsky:1973ha}.
\begin{table}[h]
\centering
\begin{tabular}{|p{0.5cm}<{\centering}|p{3.4cm}<{\centering}|p{3.4cm}<{\centering}|}
\hline
$s$ & $\psi$ & $T$\\ \hline
-1 & $(r-i a\cos\theta)^{2}\phi_2$ & $(r-i a\cos\theta)^{2}J_2$ \\ \hline
-2 & $(r-i a\cos\theta)^{4}\psi_4$ &$2(r-i a\cos\theta)^{4}T_4$ \\ \hline
\end{tabular}
\caption{The explicit expressions for the the field $\psi$ and the corresponding source $T$.}
\label{source}
\end{table}
In terms of the eigenfunctions ${_{s}}S_{lm}(\theta)$ \cite{Teukolsky:1973ha,Goldberg:1966uu}, the field $\psi$ can be written as %
\begin{equation}
\psi=\int d\omega \sum_{l,m}R_{\omega lm}(r)~{_{s}}S_{lm}(\theta)e^{-i\omega t+im\varphi},
\end{equation}
where the radial function $R_{\omega lm}(r)$ satisfies the inhomogeneous Teukolsky equation
\begin{equation}
\label{Teukolsky}
\Delta^{-s}\frac{d}{d r}\left(\Delta^{s+1}\frac{d R_{\omega lm}}{d r}\right)-V_{T}(r)R_{\omega lm}=T_{\omega lm},
\end{equation}
the function
\begin{equation}
V_{T}=-\frac{K^2-2is(r-1)K}{\Delta}-4is\omega r+\lambda_{lm\omega},
\end{equation}
$K=(r^2+a^2)\omega-am$, $\lambda_{lm\omega}$ is the corresponding eigenvalue which can be computed by the BH Perturbation Toolkit \cite{BHPToolkit}, and the source $T_{\omega lm}(r)$ is
\begin{equation}
T_{\omega lm}(r)=\frac{1}{2\pi}\int dt d\Omega ~4\pi \Sigma T ~{_s}S_{lm}(\theta)e^{i\omega t-im\varphi}.
\end{equation}
For the equatorial circular trajectory at $r_0$ under consideration,
the sources are
\begin{equation}
\begin{split}
T^{\mu\nu}_p(x)&=\frac{m_p}{r_0^2}\frac{u^{\mu}u^{\nu}}{u^t}\delta(r-r_0)\delta(\cos\theta)\delta(\varphi-\hat{\omega} t),\\
J^{\mu}(x)&=q\frac{m_p}{r_0^2}\frac{u^{\mu}}{u^t}\delta(r-r_0)\delta(\cos\theta)\delta(\varphi-\hat{\omega} t),
\end{split}
\end{equation}
where $\hat{\omega}$ is the orbital angular frequency.
Geodesic motion in Kerr spacetime
admits three constants of motion: the specific energy $\hat{E}$, the angular momentum $\hat{L}$, and the Carter constant $\hat{Q}$,
and the geodesic equations are
\begin{eqnarray}
m_p\Sigma \frac{d t}{d \tau} &=&\hat{E} \frac{\varpi^{4}}{\Delta}+a \hat{L}\left(1-\frac{\varpi^{2}}{\Delta}\right)-a^{2} \hat{E}\sin^{2}\theta, \label{timequa}\\
m_p\Sigma \frac{d r}{d \tau} &=&\pm \sqrt{V_{r}\left(r_{0}\right)}, \\
m_p\Sigma \frac{d \theta}{d \tau} &=&\pm \sqrt{V_{\theta}\left(\theta\right)}, \\
m_p\Sigma \frac{d \varphi}{d \tau} &=&a \hat{E}\left(\frac{\varpi^{2}}{\Delta}-1\right)-\frac{a^{2} \hat{L}}{\Delta}+ \hat{L} \csc ^{2} \theta,\label{anglequa}
\end{eqnarray}
where $\varpi\equiv\sqrt{r^2+a^2}$, the radial and polar potentials are
\begin{eqnarray}
&V_{r}(r)& =\left(\hat{E} \varpi^{2}-a \hat{L}\right)^{2}-\Delta\left(r^{2}+\left(\hat{L}-a \hat{E}\right)^{2}+\hat{Q}\right), \\
&V_{\theta}(\theta)& = \hat{Q}-\hat{L}^{2} \cot ^{2} \theta-a^{2}\left(1-\hat{E}^{2}\right) \cos ^{2} \theta.
\end{eqnarray}
In the adiabatic approximation, for a quasi-circular orbit on the equatorial plane,
the coordinates $r$ and $\theta$ are considered as constants,
then Eqs. \eqref{timequa} and \eqref{anglequa} are the remaining equations.
The conserved constants are \cite{Detweiler:1978ge}
\begin{eqnarray}\label{orbitE}
\hat{E}&=& m_p\frac{r_0^{3 / 2}-2 r_{0}^{1 / 2} \pm a }{r_{0}^{3 / 4}\left(r_{0}^{3 / 2}-3 r_{0}^{1 / 2} \pm 2 a \right)^{1 / 2}}, \\
\hat{L}&=&m_p \frac{\pm (r_{0}^{2}\mp 2a r_{0}^{1 / 2} + a^2)}{r_{0}^{3 / 4}\left(r_{0}^{3 / 2}-3 r_{0}^{1 / 2} \pm 2 a \right)^{1 / 2}},\\
\hat{Q}&=&0.
\end{eqnarray}
The orbital angular frequency is
\begin{equation}\label{orbitF}
\hat{\omega} \equiv \frac{d\varphi}{dt}=\frac{\pm 1}{r_{0}^{3 / 2} \pm a},
\end{equation}
where $\pm$ corresponds to co-rotating and counter-rotating, respectively.
In the following discussions, we use positive $a$ for co-rotating cases and negative $a$ for counter-rotating cases.
\section{Numerical calculation for the energy flux}
\label{sec3}
The homogeneous Teukolsky equation \eqref{Teukolsky} admits two linearly independent solutions $R^{\text{in}}_{\omega lm}$ and $R^{\text{up}}_{\omega lm}$, with the following asymptotic values at the horizon $r_+$ and at infinity,
\begin{equation}
R^{\text{in}}_{\omega lm}=
\begin{cases}
B^{\text{tran}}\Delta^{-s}e^{-i\kappa r^*},&\quad (r\to r_+)\\
B^{\text{out}}\frac{e^{i\omega r^*}}{r^{2s+1}}+B^{\text{in}}\frac{e^{-i\omega r^*}}{r}, &\quad (r\to+\infty)
\end{cases}
\end{equation}
\begin{equation}
R^{\text{up}}_{\omega lm}=
\begin{cases}
D^{\text{out}}e^{i\kappa r^*}+\frac{D^{\text{in}}}{\Delta^{s}}e^{-i\kappa r^*},&\quad (r\to r_+)\\
D^{\text{tran}}\frac{e^{i\omega r^*}}{r^{2s+1}},&\quad (r\to+\infty)\\
\end{cases}
\end{equation}
where $\kappa=\omega-m a/(2r_+)$, $r_\pm=1\pm\sqrt{1-a^2}$, and the tortoise radius of the Kerr metric
\begin{equation}
r^*=r+\frac{2r_+}{r_+-r_-}\ln \frac{r-r_+}{2}-\frac{2r_-}{r_+-r_-}\ln \frac{r-r_-}{2}.
\end{equation}
The solutions $R^{\text{in}}_{\omega lm}$ and $R^{\text{up}}_{\omega lm}$ are purely outgoing at infinity and purely ingoing at the horizon.
With the help of these homogeneous solutions, the solution to Eq.~\eqref{Teukolsky} is
\begin{equation}
\begin{split}
R_{\omega lm}(r)=\frac{1}{W}
\left(R^{\text{in}}_{\omega lm}\int_{r}^{+\infty}\Delta^{s}R^{\text{up}}_{\omega lm}T_{\omega lm}dr+R^{\text{up}}_{\omega lm}\int_{r_+}^{r}\Delta^{s}R^{\text{in}}_{\omega lm}T_{\omega lm}dr\right).
\end{split}
\end{equation}
with the constant Wronskian given by
\begin{equation}
W= \Delta^{s+1} \left(R^{\text{in}}_{\omega lm}\frac{d R^{\text{up}}_{\omega lm}}{dr}-R^{\text{up}}_{\omega lm}\frac{d R^{\text{in}}_{\omega lm}}{dr}\right)=2i\omega B^{\text{in}} D^{\text{tran}}.
\end{equation}
The solution is purely outgoing at infinity and purely ingoing at the horizon,
\begin{equation}
\begin{split}
R_{\omega lm}(r\to r_+)=Z^{\infty}_{\omega lm}\Delta^{-s}e^{-i\kappa r^*},\\
R_{\omega lm}(r\to \infty)=Z^{H}_{\omega lm}r^{-2s-1}e^{i\omega r^*},
\end{split}
\end{equation}
with
\begin{equation}
\begin{split}
Z^{\infty}_{\omega lm}&=\frac{B^{\text{tran}}}{W}\int_{r_+}^{+\infty}\Delta^{s}R^{\text{up}}_{\omega lm}T_{\omega lm}dr,\\
Z^{H}_{\omega lm}&=\frac{D^{\text{tran}}}{W}\int_{r_+}^{+\infty}\Delta^{s}R^{\text{in}}_{\omega lm}T_{\omega lm}dr.
\label{amplitudes}
\end{split}
\end{equation}
For a circular equatorial orbit with orbital angular frequency $\hat{\omega} $, we get
\begin{equation}
Z^{H,\infty}_{\omega lm}=\delta(\omega-m \hat{\omega})\mathcal{A}^{H,\infty}_{\omega lm}.
\end{equation}
For $s=-1$, the energy fluxes at infinity and the horizon read
\begin{equation}
\begin{split}
\dot{E}_q^{\infty}=\left(\frac{d E}{dt}\right)_{EM}^\infty&=\sum_{l=1}^{\infty}\sum_{m=1}^{l}\frac{|\mathcal{A}^{H}_{\omega lm}|^2}{\pi}, \\
\dot{E}_q^H=\left(\frac{d E}{dt}\right)_{EM}^H&=\sum_{l=1}^{\infty}\sum_{m=1}^{l}\alpha^E_{lm}\frac{|\mathcal{A}^{\infty}_{\omega lm}|^2}{\pi},
\end{split}
\end{equation}
where the coefficient $\alpha^E_{lm}$ is \cite{Teukolsky:1974yv}
\begin{equation}\label{energyformula}
\alpha^E_{lm}=\frac{128\omega\kappa r_+^3(\kappa^2+4\epsilon^2)}{|B_E|^2}
\end{equation}
with $\epsilon=\sqrt{1-a^2}/(4r_+)$ and
\begin{equation}
|B_E|^2=\lambda_{lm\omega}^2+4ma\omega-4a^2\omega^2.
\end{equation}
For $s=-2$, the gravitational energy fluxes at infinity and the horizon are given by
\begin{equation}
\begin{split}
\dot{E}_{\text{grav}}^{\infty}=\left(\frac{d E}{dt}\right)_{GW}^\infty&=\sum_{l=2}^{\infty}\sum_{m=1}^{l}\frac{|\mathcal{A}^{H}_{\omega lm}|^2}{2\pi\omega^2}, \\
\dot{E}_{\text{grav}}^H=\left(\frac{d E}{dt}\right)_{GW}^H&=\sum_{l=2}^{\infty}\sum_{m=1}^{l}\alpha^G_{lm}\frac{|\mathcal{A}^{\infty}_{\omega lm}|^2}{2\pi\omega^2},
\end{split}
\end{equation}
where the coefficient $\alpha^G_{l m}$ is \cite{Hughes:1999bq}
\begin{equation}\label{energyformula2}
\alpha^G_{l m}=\frac{256\left(2 r_{+}\right)^5 \kappa\left(\kappa^2+4 \epsilon^2\right)\left(\kappa^2+16 \epsilon^2\right)\omega^3}{\left|B_G\right|^2},
\end{equation}
and
\begin{equation}
\begin{aligned}
\left|B_G\right|^2 &=\left[\left(\lambda_{l m \omega}+2\right)^2+4 a\omega-4 a^2\omega^2\right]\times\left[\lambda_{l m \omega}^2+36 m a\omega-36 a^2\omega^2\right] \\
&+\left(2 \lambda_{l m \omega}+3\right)\left[96 a^2\omega^2-48 m a\omega\right]+144\omega^2\left(1-a^2\right) .
\end{aligned}
\end{equation}
Therefore, the total energy fluxes emitted from the EMRIs read
\begin{equation}
\dot{E}= \dot{E}_q+\dot{E}_{\text{grav}},
\end{equation}
where
\begin{equation}
\dot{E}_q=\dot{E}_q^\infty+\dot{E}_q^H,\ \ \dot{E}_{\text{grav}}=\dot{E}_{\text{grav}}^\infty+\dot{E}_{\text{grav}}^H.
\end{equation}
The detailed derivation of the above results is given in Appendix \ref{gsne}.
The energy flux emitted by tensor fields can be computed with the BH Perturbation Toolkit \cite{BHPToolkit}.
\section{Results}\label{sec4}
The top panel of Fig. \ref{energyd} shows the normalized vector energy flux $m_p^{-2}M^2\dot{E}_q$ for a charged particle with different charge values of $q$ on a circular orbit about a Kerr BH with the spin $a=0.9$, as a function of orbital radius.
The vector energy flux is proportional to the square of the vector charge $q^2$.
The vector energy flux increases as the charged particle inspirals into the central Kerr BH.
The ratio between the vector and gravitational energy flux is shown in the bottom panel of Fig. \ref{energyd}.
Both the vector and gravitational fluxes are in the same order of $(m_p/M)^2$,
the ratio of fluxes is independent of the mass ratio and increases as the orbital radius because the gravitational contribution falls off faster than the vector energy flux at a large orbital radius.
Figure \ref{energya} shows the normalized vector energy flux $m_p^{-2}M^2\dot{E}_q$ and the ratio of energy fluxes $\dot{E}_q/\dot{E}_{\textrm{grav}}$ as a function of the orbital radius for different values of $a$.
For larger $a$, the ISCO is smaller, resulting in higher GW frequency at the coalescence.
For the same orbital radius, the vector energy flux is slightly larger for smaller $a$.
However, the total energy flux increases with $a$ for one-year observations before the merger due to the smaller ISCO.
Figure \ref{energyHI} shows the ratio of energy flux falling onto the horizon to the energy flux radiated away to infinity,
as a function of the orbital radius, for various spin $a$, and for the vector and gravitational fields.
It is interesting to note that the sign of ratio becomes negative for Kerr BH with positive $a$ (co-rotating orbit) at a small orbital radius.
In these cases, the vector and gravitational fields generate superradiance, leading to extraction of energy from the horizon.
The superradiance only happens when the coefficient $\kappa$ in Eqs. \eqref{energyformula} and \eqref{energyformula2} becomes negative, which means that the orbital frequency slows down the rotation of the Kerr BH.
Our results are consistent with those found in Ref. \cite{Torres:2020fye}.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{energyd.pdf}
\caption{The energy fluxes versus the orbital distance.
The top panel shows the vector flux normalized with the mass ratio from a charged particle orbiting around a Kerr BH with the spin of $a/M=0.9$ for different values of the vector charge $q$.
The bottom panel shows the ratio between vector and gravitational energy fluxes for different values of the vector charge $q$.}
\label{energyd}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{energya.pdf}
\caption{Same as Fig. \ref{energyd}, but for different values of the primary spin $a$.
The vector charge $q$ is set to 1.}
\label{energya}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{energyHI.pdf}
\caption{The ratio of the energy flux falling onto the horizon to the energy flux radiated to infinity, as a function of the orbital radius, for various spin $a$, and for the vector (the top panel) and gravitational cases (the bottom panel).}
\label{energyHI}
\end{figure}
The extra energy leakage due to the vector field accelerates the coalescence of binaries.
Therefore, we expect the vector charge to leave a significant imprint on the GW phase over one-year evolution before the merger for EMRIs.
To detect the vector charge carried by the small compact object in EMRIs,
we study the dephasing of GWs caused by the additional energy loss during inspirals.
The observation time is one year before the merger,
\begin{equation}\label{sol-fre}
T_{\text{obs}}=\int^{f_{\text{max}}}_{f_{\text{min}}}\frac{1}{\dot{f}}df=1\ \text{year},
\end{equation}
where
\begin{equation}
f_{\text{max}}=\text{min}(f_{\text{ISCO}},f_{\text{up}}),~~~~~~f_{\text{min}}=\text{max}(f_{\text{low}},f_{\text{start}}),
\end{equation}
$f=\hat{\omega}/\pi$ is the GW frequency, $f_{\text{ISCO}}$ is the frequency at the ISCO \cite{Jefremov:2015gza},
$f_{\text{start}}$ is the initial frequency at $t=0$,
the cutoff frequencies $f_{\text{low}}=10^{-4}$ Hz and $f_{\text{up}}=1$ Hz.
The orbit evolution is determined by
\begin{equation}\label{orbittime}
\frac{d r}{dt}=-\dot{E}\left(\frac{d \hat{E}}{dr}\right)^{-1},\qquad \frac{d \varphi_{\text{orb}}}{d t}=\pi f,
\end{equation}
where $\dot{E}=\dot{E}_q+\dot{E}_{\text{grav}}$.
The total number of GW cycles accumulated over one year before the merger is \cite{Berti:2004bd}
\begin{equation}\label{phase-end}
\mathcal{N}=\int_{f_{\text{min}}}^{f_{\max }} \frac{f}{\dot{f}} d f.
\end{equation}
Considering EMRIs with the mass of the second compact object being fixed to be $m_p=10~M_{\odot}$,
we calculate the dephasing $\Delta\mathcal{N}=\mathcal{N}(q=0)-\mathcal{N}(q)$ for different vector charge $q$, spin $a$ and mass $M$,
and the results are shown in Fig. \ref{phase}.
For one-year observations before the merger,
the charged particle starts further away from ISCO due to extra radiation of the vector field and the difference $\Delta\mathcal{N}$ is always positive.
As shown in Fig. \ref{phase}, $\Delta \mathcal{N}$ increases monotonically with the spin $a$ and the charge-to-mass ratio $q$, and it strongly depends on the mass of the central BH such that lighter BHs have larger $\Delta \mathcal{N}$.
This means that the observations of EMRIs with a lighter and larger-spin Kerr BH can detect the vector charge easier.
For the same EMRI configuration in the Kerr background, the co-rotating orbit can detect the vector charge easier than the counter-rotating orbit.
Following Refs. \cite{Berti:2004bd,Maselli:2020zgv},
we take the threshold for a detectable dephasing that two signals are distinguishable by space-based GW detectors as $\Delta \mathcal{N}=1$.
Observations of EMRIs over one year before the merger may be able to reveal the presence of a vector charge as small as $q\sim 0.007$ for Kerr BHs with $a=0.9$ and $M=10^6~M_\odot$, and $q\sim 0.01$ for Schwarzschild BHs with $M=10^6~M_\odot$.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{dephasing.pdf}
\caption{The difference between the number of GW cycles accumulated by EMRIs with and without the vector charge in circular orbits. The left panel shows the dephasing as a function of the mass of the Kerr BH in the range $M\in\left[2\times 10^5,2\times 10^7\right]M_\odot$ for different spin values, the charge $q=0.01$.
The right panel shows the dephasing as a function of the charge-to-mass ratio $q$ for different $M$, the spin $a=0.9$. The red dashed line corresponds to the threshold above which two signals are distinguishable with space-based GW detectors. All observational time is one year before the merger.}
\label{phase}
\end{figure}
\section{Parameter Estimation}
To make the analysis more accurate and account for the degeneracy among parameters,
we calculate the faithfulness between two GW waveforms and carry out parameter estimation with the FIM method.
\label{sec5}
\subsection{Signals}
We can obtain the inspiral trajectory from adiabatic evolution in Eq. \eqref{orbittime}, then compute GWs in the quadrupole approximation.
The metric perturbation in the transverse-traceless (TT) gauge is
\begin{equation}
h_{i j}^{\mathrm{TT}}=\frac{2}{d_L}\left(P_{i l} P_{j m}-\frac{1}{2} P_{i j} P_{l m}\right) \ddot{I}_{l m},
\end{equation}
where $d_L$ is the luminosity distance of the source,
$P_{ij}=\delta_{ij}-n_i n_j$ is the projection operator acting onto GWs with the unit propagating direction $n_j$,
$\delta_{ij}$ is the Kronecker delta function,
and $\ddot{I}_{ij}$ is the second time derivative of the mass quadrupole moment.
The GW strain measured by the detector is
\begin{equation}\label{signal}
h(t)=h_{+}(t) F^{+}(t)+h_{\times}(t) F^{\times}(t),
\end{equation}
where $h_+(t)=\mathcal{A}\cos\left[2\varphi_{\rm orb}+2\varphi_0\right]\left(1+\cos^2\iota\right)$, $h_\times(t)=-2\mathcal{A}\sin\left[2\varphi_{\rm orb}+2\varphi_0\right]\cos\iota$, $\iota$ is the inclination angle between the binary orbital angular momentum and the line of sight,
the GW amplitude $\mathcal{A}=2m_{\rm p}\left[M\hat{\omega}(t)\right]^{2/3}/d_L$ and $\varphi_0$ is the initial phase.
The interferometer pattern functions $F^{+,\times}(t)$ and $\iota$ can be expressed in terms of four angles which specify the source orientation, $(\theta_s,\phi_s)$,
and the orbital angular direction $(\theta_1,\phi_1)$.
The faithfulness between two signals is defined as
\begin{equation}\label{eq:def_F}
\mathcal{F}_n[h_1,h_2]=\max_{\{t_c,\phi_c\}}\frac{\langle h_1\vert
h_2\rangle}{\sqrt{\langle h_1\vert h_1\rangle\langle h_2\vert h_2\rangle}}\ ,
\end{equation}
where $(t_c,\phi_c)$ are time and phase offsets \cite{Lindblom:2008cm},
the noise-weighted inner product between two templates $h_1$ and $h_2$ is
\begin{equation}\label{product}
\left\langle h_{1} \mid h_{2}\right\rangle=4 \Re \int_{f_{\min }}^{f_{\max }} \frac{\tilde{h}_{1}(f) \tilde{h}_{2}^{*}(f)}{S_{n}(f)} df,
\end{equation}
$\tilde{h}_{1}(f)$ is the Fourier transform of the time-domain signal $h(t)$,
its complex conjugate is $\tilde{h}_{1}^{*}(f)$,
and $S_n(f)$ is the noise spectral density for space-based GW detectors.
The signal-to-noise ratio (SNR) can be obtained by calculating $\rho=\left\langle h|h \right\rangle^{1/2}$.
The sensitivity curves of LISA, TianQin, and Taiji are shown in Fig. \ref{sensitivity}.
As pointed out in \cite{Chatziioannou:2017tdw},
two signals can be distinguished by LISA if $\mathcal{F}_n\leq0.988$.
Here we choose the source masses $m_p=10~M_{\odot}$, $M=10^6~M_{\odot}$,
the source angles $\theta_s=\pi/3,~\phi_s=\pi/2$ and $\theta_1=\phi_1=\pi/4$,
the luminosity distance is scaled to ensure SNR $\rho=30$,
the initial phase is set as $\varphi_0=0$ and the initial orbital separation is adjusted to experience one-year adiabatic evolution before the plunge $r_{\text{end}}=r_{\text{ISCO}}+0.1~M$.
In Fig. \ref{faithfulness}, we show
the faithfulness between GW signals with and without the vector charge for LISA as a function of the vector charge.
The results show that one-year observations of EMRIs with LISA may be able to reveal the presence of a vector charge as small as $q\sim 0.002$ for Kerr BHs with $a=0.9$ and $M=10^6~M_\odot$ (co-rotating orbit),
$q\sim 0.003$ for Schwarzschild BHs with $a=0$ and $M=10^6~M_\odot$, and $q\sim 0.004$ for Kerr BHs with $a=-0.9$ and $M=10^6~M_\odot$ (counter-rotating orbit).
Larger positive spin of the Kerr BH (co-rotating orbit) can help us detect the vector charge easier, which is consistent with the results obtained from the dephasing in the previous section.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{sensitivity.pdf}
\caption{The sensitivity curves for LISA, TianQin, and Taiji.
The horizontal solid lines represent the frequency band $f_{\textrm{start}}$ to $f_{\textrm{ISCO}}$ for EMRIs with $a=0.9$, and $M=2\times10^5~M_{\odot}$, $M=10^6~M_{\odot}$ and $M=10^7~M_{\odot}$ over one-year evolution before the merger.
}
\label{sensitivity}
\end{figure}
\begin{figure}[thbp]
\center{
\includegraphics[scale=0.74]{Faithfulness.pdf}
\caption{Faithfulness between GW signals with and without the vector charge for LISA as a function of the charge $q$. The spin of the Kerr BH is $a=0$, $a=0.9$, and $a=-0.9$. The horizontal dashed line represents the detection limit with LISA, $\mathcal{F}_n= 0.988$.
}\label{faithfulness}}
\end{figure}
\subsection{Fisher information matrix}
The signals \eqref{signal} measured by the detector are determined by the following eleven parameters
\begin{equation}
\xi=(\ln M, \ln m_p, a, q, r_0, \varphi_0, \theta_s, \phi_s, \theta_1, \phi_1, d_L).
\end{equation}
In the large SNR limit,
the posterior probability distribution of the source parameters $\xi$ can be approximated by a multivariate Gaussian distribution centered around the true values $\hat{\xi}$.
Assuming flat or Gaussian priors on the source parameters $\xi$,
their covariances are given by the inverse of the FIM
\begin{equation}
\Gamma_{i j}=\left\langle\left.\frac{\partial h}{\partial \xi_{i}}\right| \frac{\partial h}{\partial \xi_{j}}\right\rangle_{\xi=\hat{\xi}}.
\end{equation}
The statistical error on $\xi$ and the correlation coefficients between the parameters are provided by the diagonal and non-diagonal parts of ${\bf \Sigma}={\bf \Gamma}^{-1}$, i.e.
\begin{equation}
\sigma_{i}=\Sigma_{i i}^{1 / 2} \quad, \quad c_{\xi_{i} \xi_{j}}=\Sigma_{i j} /\left(\sigma_{\xi_{i}} \sigma_{\xi_{j}}\right).
\end{equation}
Because of the triangle configuration of the space-based GW detector, the total SNR is defined by $\rho=\sqrt{\rho_1^2+\rho_2^2}$, so the total covariance matrix of the binary parameters is obtained by inverting the sum of the Fisher matrices $\sigma_{\xi_i}^2=(\Gamma_1+\Gamma_2)^{-1}_{ii}$.
Here we fix the source angles $\theta_s=\pi/3,~\phi_s=\pi/2$ and $\theta_1=\phi_1=\pi/4$,
the initial phase is set as $\varphi_0=0$ and the initial orbital separation is adjusted to experience one-year adiabatic evolution before the plunge $r_{\text{end}}=r_{\text{ISCO}}+0.1~M$.
The luminosity distance $d_L$ is set to be $1$ Gpc.
We apply the FIM method for LISA, TianQin, and Taiji to estimate the errors of the vector charge.
The relative errors of the vector charge $q$ as a function of the vector charge with LISA, TianQin and Taiji are shown in Fig. \ref{sigmaq}.
For one-year observations before the merger,
the charged particle starts further away from ISCO due to extra radiation of the vector field,
so the $1\sigma$ error for the charge decreases with the charge $q$.
For EMRIs with $M=2\times 10^5~M_{\odot}$ and $a=0.9$, as shown in the top panel,
the relative errors of the charge $q$ with TianQin are better than LISA and Taiji.
For $M=10^6~M_{\odot}$ and $a=0.9$ as shown in the middle panel,
the relative errors of the charge $q$ with TianQin and LISA are almost the same.
For $M=10^7~M_{\odot}$ and $a=0.9$ as shown in the bottom panel,
the relative errors of the charge $q$ with TianQin are worse than LISA and Taiji.
In all the cases, the relative errors of the charge $q$ with Taiji are better than LISA, for the reason that the sensitivity of Taiji is always better than LISA.
For $M=2\times 10^5~M_{\odot}$ and $a=0.9$,
the relative errors with TianQin are better than LISA and Taiji since the sensitivity of TianQin is better than LISA and Taiji in the high-frequency band, but worse than LISA and Taiji in the low-frequency band as shown in Fig. \ref{sensitivity}.
For EMRIs with $m_p=10~M_{\odot}$, $M=10^6~M_{\odot}$ and $a=0.9$,
the vector charge can be constrained for LISA as small as $q\sim0.021$,
for TianQin as small as $q\sim0.028$ and for Taiji as small as $q\sim0.016$.
For EMRIs with $m_p=10~M_{\odot}$, $M=2\times10^5~M_{\odot}$ and $a=0.9$,
the vector charge can be constrained for LISA as small as $q\sim0.0082$,
for TianQin as small as $q\sim0.0049$ and for Taiji as small as $q\sim0.0057$.
Figure \ref{sigmaa} shows the relative errors of the vector charge $q$ versus the spin $a$ with LISA, TianQin, and Taiji.
In general, the relative errors of the charge for the co-rotating orbit are better than those for the counter-rotating orbit.
We only consider the co-rotating orbit for simplicity.
The $1\sigma$ error for the charge decreases with the spin $a$.
Comparing the relative errors for Kerr BHs with spin $a=0.9$ and spin $a=0$,
we find that the spin of Kerr BHs can decrease the charge uncertainty by about one or two orders of magnitude, depending on the mass of the Kerr BH.
For EMRIs with $M=10^6~M_\odot$, $m_p=10~M_\odot$, $q=0.05$, and different $a$,
the corner plots for source parameters with LISA are shown in Figs. \ref{corner09}, \ref{corner0} and \ref{cornerm09}.
For comparison, we also show the corner plot for charged EMRIs in the Schwarzschild BH background in Fig. \ref{corner0s}.
For $a=(0.9,0,-0.9)$, the corresponding errors of charge $\sigma_q$ are $(0.0031,0.086,0.65)$, respectively.
As expected, $\sigma_q$ is smaller for co-rotating orbits and bigger for counter-rotating orbits.
For EMRIs in the Kerr background, the co-rotating orbit can better detect the vector charge.
It is interesting to note that the charge $q$ are anti-correlated with the mass $M$ and the spin $a$ of Kerr BHs,
and the correlations between $q$ and $M$, and $q$ and $m_p$ in the Kerr BH background
are opposite to those in the Schwarzschild BH background.
\begin{figure}
\centering
\includegraphics[width=0.6\columnwidth]{M2e5q.pdf}\\
\includegraphics[width=0.6\columnwidth]{M1e6q.pdf}\\
\includegraphics[width=0.6\columnwidth]{M1e7q.pdf}\\
\caption{
The $1\sigma$ interval for the charge $q$ as a function of the charge $q$, inferred after one-year observations of EMRI with $a=0.9$ and different $M$ with LISA, Taiji, and TianQin. The horizontal dashed lines represent the $3\sigma$ limit $33.3\%$.
}
\label{sigmaq}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\columnwidth]{M2e5a.pdf}\\
\includegraphics[width=0.6\columnwidth]{M1e6a.pdf}\\
\includegraphics[width=0.58\columnwidth]{M1e7a.pdf}\\
\caption{
The $1\sigma$ interval for the charge $q$ as a function of the spin $a$, inferred after one-year observations of EMRI with $q=0.05$ and different $M$ with LISA, Taiji, and TianQin.
}
\label{sigmaa}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{LISA1e6A0.9.pdf}
\caption{Corner plot for the probability distribution of the source parameters $(\ln M,\ln m_p, a, q)$ with LISA, inferred after one-year observations of EMRIs with $q=0.05$ and $a=0.9$.
Vertical lines show the $1\sigma$ interval for the source parameter.
The contours correspond to the $68\%$, $95\%$, and $99\%$ probability confidence intervals.}
\label{corner09}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{LISA1e6A0.0.pdf}
\caption{Corner plot for the probability distribution of the source parameters $(\ln M,\ln m_p, a, q)$ with LISA, inferred after one-year observations of EMRIs with $q=0.05$ and $a=0$.
Vertical lines show the $1\sigma$ interval for the source parameter.
The contours correspond to the $68\%$, $95\%$, and $99\%$ probability confidence intervals.}
\label{corner0}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{LISA1e6A-0.9.pdf}
\caption{Corner plot for the probability distribution of the source parameters $(\ln M,\ln m_p, a, q)$ with LISA, inferred after one-year observations of EMRIs with $q=0.05$ and $a=-0.9$.
Vertical lines show the $1\sigma$ interval for the source parameter.
The contours correspond to the $68\%$, $95\%$, and $99\%$ probability confidence intervals.}
\label{cornerm09}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{LISA1e6A0.0s.pdf}
\caption{Corner plot for the probability distribution of the source parameters $(\ln M,\ln m_p, q)$ with LISA, inferred after one-year observations of EMRIs with $q=0.05$ in the Schwarzschild BH background.
Vertical lines show the $1\sigma$ interval for the source parameter.
The contours correspond to the $68\%$, $95\%$, and $99\%$ probability confidence intervals.}
\label{corner0s}
\end{figure}
\section{Conclusions}
\label{sec6}
We study the energy emissions and GWs from EMRIs consisting of a small charged compact object with mass $m_p$ and the charge to mass ratio $q$ inspiraling into a Kerr BH with spin $a$.
We derive the formula for solving the inhomogeneous Teukolsky equation with a vector field and calculate the power emission due to the vector field in the Kerr background.
By using the difference between the number of GW cycles $\Delta\mathcal{N}$ accumulated by EMRIs with and without the vector charge in circular orbits over one year before the merger,
we may reveal the presence of a vector charge as small as $q\sim 0.007$ for Kerr BHs with $a=0.9$ and $M=10^6~M_\odot$, and $q\sim 0.01$ for Schwarzschild BHs with $M=10^6~M_\odot$.
The dephasing increases monotonically with the charge-to-mass ratio $q$, and it strongly depends on the mass of the Kerr BH such that lighter BHs have larger dephasing.
We also apply the faithfulness between GW signals with and without the vector charge
to discuss the detection of the vector charge $q$.
We find that positive larger spin of the Kerr BH can help us detect the vector charge easier.
We show that one-year observations of EMRIs with LISA may be able to reveal the presence of a vector charge as small as $q\sim 0.002$ for Kerr BHs with $a=0.9$ and $M=10^6~M_\odot$,
$q\sim 0.003$ for Schwarzschild BHs with $a=0$ and $M=10^6~M_\odot$, and $q\sim 0.004$ for Kerr BHs with $a=-0.9$ and $M=10^6~M_\odot$.
To determine vector charge more accurately and account for the degeneracy among parameters,
we calculate the FIM to estimate the errors of the vector charge $q$.
For EMRIs with $M=2\times10^5~M_{\odot}$ and $a=0.9$,
the vector charge $q$ can be constrained as small as $q\sim0.0049$ with TianQin, $q\sim0.0057$ with Taiji, and $q\sim 0.0082$ with LISA.
For EMRIs with $M=10^6~M_{\odot}$ and $a=0.9$,
the vector charge can be constrained as small as $q\sim 0.016$ with Taiji, $q\sim 0.021$ with LISA, and $q\sim 0.028$ with TianQin.
Since the sensitivity of TianQin is better than LISA and Taiji in the high-frequency bands,
the ability to detect the vector charge for TianQin is better than LISA and Taiji when the mass $M$ of the Kerr BH with $a=0.9$ is lighter than $\sim 2\times 10^5~M_{\odot}$.
For the mass $M$ of the Kerr BH with $a=0.9$ above $10^6~M_{\odot}$, LISA and Taiji are more likely to detect smaller vector charges.
The relative errors of the charge $q$ with Taiji are always smaller than LISA because the sensitivity of Taiji is always better than LISA.
Due to the extra radiation of the vector field,
the charged particle starts further away from ISCO,
so the $1\sigma$ error for the charge decreases with the charge $q$.
As the spin $a$ of Kerr BHs increases, the ISCO becomes smaller,
the positive spin of Kerr BHs (co-rotating) can decrease the charge uncertainty by about one or two orders of magnitude, depending on the mass of the Kerr BH.
For EMRIs with $M=10^6~M_\odot$, $m_p=10~M_\odot$, $q=0.05$, and different $a=(0.9,0,-0.9)$, the corresponding errors of charge $\sigma_q$ with LISA are $(0.0031,0.086,0.65)$, respectively,
so co-rotating orbits can better detect the vector charge.
It is interesting to note that the charge $q$ are anti-correlated with the mass $M$ and the spin $a$ of Kerr BHs,
and the correlations between $q$ and $M$, and $q$ and $m_p$ in the Kerr BH background
are opposite to those in the Schwarzschild BH background.
In summary, the observations of EMRIs with a lighter and larger-spin Kerr BH can detect the vector charge easier.
\begin{acknowledgments}
This work makes use of the Black Hole Perturbation Toolkit package.
This research is supported in part by the National Key Research and Development Program of China under Grant No. 2020YFC2201504
\end{acknowledgments}
|
\section{Introduction} \label{sec: Introduction}
\subsection{Classical multiple zeta values} ${}$\par
Let $\mathbb N=\{1,2,\dots\}$ be the set of positive integers and $\mathbb Z^{\geq 0}=\{0,1,2,\dots\}$ be the set of non-negative integers. The multiple zeta values (MZV's for short) are real positive numbers that were studied notably by Euler in the eighteenth century. They are given by the following convergent series
\[ \zeta(n_1,\dots,n_r)=\sum_{0<k_1<\dots<k_r} \frac{1}{k_1^{n_1} \dots k_r^{n_r}} \]
where $n_i$ are positive integers with $n_i \geq 1$ and $n_r \geq 2$. Here $r$ is called the depth and $w := n_1+\dots+n_r$ is called the weight of the presentation $\zeta(n_1,\dots,n_r)$.
These numbers generalize the zeta values
\[ \zeta(n)=\sum_{k>0} \frac{1}{k^n}, \quad \text{where } n \in \mathbb{N} \text{ and } n \geq 2, \]
which were studied well before Riemann studied them as a function $\zeta(s)$ of a complex variable $s$, and its links with the distribution of primes. The simple zeta values, and more generally the multiple zeta values, still keep many secrets and play the role of fundamental constants. They are ubiquitous in many fields of mathematics and physics, in particular through the Feynman integrals which govern the interactions between elementary particles or through the Drinfeld associators coming from quantum groups and knot theory.
The even zeta values are well understood. In fact, Euler proved in 1735 that, when $n$ is even, $\zeta(n)$ is a rational multiple of $\pi^n$. Thanks to Lindemann’s proof of the transcendence of $\pi$, it follows that all these numbers are transcendental. However, the odd zeta values are much more mysterious. Indeed, a folklore conjecture states
\begin{conjecture}
The numbers $\pi, \zeta(3), \zeta(5), \dots$ are all algebraically independent over $\mathbb Q$.
\end{conjecture}
To our knowledge, we know nothing about the transcendence of odd zeta values. Concerning the irrationality of these numbers, Apéry \cite{Ape79} showed that $\zeta(3)$ is irrational, and Ball-Rivoal \cite{BR01} proved that there are infinitely many irrational numbers among the remaining odd zeta values (see \cite{Riv00,Riv02,Zud01} for related works).
The product of two multiple zeta values is a linear combination, with integral coefficients, of multiple zeta values. For instance, Euler proved the following identity
\[ \zeta(m) \zeta(n)=\zeta(m,n)+\zeta(n,m)+\zeta(m+n) \]
for all integers $m,n \geq 2$. It follows that the $\mathbb Q$-vector space $\mathcal Z$ spanned by all MZV's has an algebra structure. One can argue that the main goal of this theory is to understand all $\mathbb Q$-linear relations among MZV's. Unlike the algebra generated by simple zeta values, there are lots of linear relations among MZV's that endow $\mathcal Z$ with a rich combinatorial structure. One systematic way to produce linear relations among MZV's is to use the so-called extended double shuffle relations introduced by Ihara-Kaneko-Zagier \cite{IKZ06}. To do so, one first defines Hoffman's algebra~$\frak h$ and its subalgebras $\frak h^0 \subset \frak h^1 \subset \frak h$ with respect to the following algebra structures. Next, one constructs two algebra structures as particular cases of quasi-product algebras introduced by Hoffman \cite{Hof00}: the stuffle algebra $(\frak h^1,*)$ and the shuffle algebra $(\frak h,\shuffle)$. By regularization \cite[\S 2]{IKZ06}, there exist zeta maps which are $\mathbb Q$-algebra homomorphisms
\[ \zeta_*:(\frak h^1,*) \to \mathcal Z, \]
and
\[ \zeta_\shuffle:(\frak h,\shuffle) \to \mathcal Z, \]
which give rise to a generalization of the stuffle product and the shuffle product. The extended double shuffle relations are obtained by ``comparing'' the stuffle and shuffle products on $\frak h^1$ (see \cite[Theorem 2]{IKZ06} for a precise statement). Further, Ihara, Kaneko and Zagier formulated the following influential conjecture (see \cite[Conjecture 1]{IKZ06}):
\begin{conjecture}[Ihara-Kaneko-Zagier's conjecture] \label{conj: IKZ}
The extended double shuffle relations exhaust all $\mathbb Q$-linear relations among MZV's.
\end{conjecture}
In particular, it implies Goncharov's conjecture which states that all $\mathbb Q$-linear relations among MZV's can be derived from those among MZV's of the same weight.
Surprisingly, if $\mathcal{Z}_k$ denotes the $\mathbb Q$-vector space spanned by MZV's of weight $k$ for $k \in \mathbb{N}$, Zagier \cite{Zag94} and Hoffman \cite{Hof97} were able to predict the dimension and an explicit basis for $\mathcal{Z}_k$.
\begin{conjecture}[Zagier's conjecture] \label{conj: Zagier}
We define a Fibonacci-like sequence of integers $d_k$ as follows. Letting $d_0=1, d_1=0$ and $d_2=1$ we define $d_k=d_{k-2}+d_{k-3}$ for $k \geq 3$. Then for $k \in \ensuremath \mathbb{N}$ we have
\[ \dim_{\mathbb Q} \mathcal Z_k = d_k. \]
\end{conjecture}
\begin{conjecture}[Hoffman's conjecture] \label{conj: Hoffman}
The $\mathbb Q$-vector space $\mathcal Z_k$ is generated by the basis consisting of MZV's of weight $k$ of the form $\zeta(n_1,\dots,n_r)$ with $n_i \in \{2,3\}$.
\end{conjecture}
The algebraic part of these conjectures which concerns upper bounds for $\dim_{\mathbb Q} \mathcal Z_k$ was solved by Terasoma \cite{Ter02}, Deligne-Goncharov \cite{DG05} and Brown \cite{Bro12} using the theory of mixed Tate motives.
\begin{theorem}[Deligne-Goncharov, Terasoma] \label{thm: Zagier}
For $k \in \ensuremath \mathbb{N}$ we have $\dim_{\mathbb Q} \mathcal Z_k \leq d_k$.
\end{theorem}
\begin{theorem}[Brown] \label{thm: Hoffman}
The $\mathbb Q$-vector space $\mathcal Z_k$ is generated by MZV's of weight $k$ of the form $\zeta(n_1,\dots,n_r)$ with $n_i \in \{2,3\}$.
\end{theorem}
The proofs of this theorem use by a crucial manner different Hopf algebra structures of Hoffman's algebra $\frak h$ as described above. We mention that the transcendental part which concerns lower bounds for $\dim_{\mathbb Q} \mathcal Z_k$ is completely open. We refer the reader to \cite{BGF,Del13,Zag94} for more details and more exhaustive references.
\subsection{Multiple zeta values in positive characteristic} ${}$\par
There is a well-known analogy between number fields and function fields (see \cite{Iwa69,MW83,Wei39}). Inspired by Euler’s work on multiple zeta values and that of Carlitz \cite{Car35} on zeta values in positive characteristic, Thakur \cite{Tha17} introduced multiple zeta values attached to the affine line over a finite field. We now need to introduce some notations. Let $A=\mathbb F_q[\theta]$ be the polynomial ring in the variable $\theta$ over a finite field $\mathbb F_q$ of $q$ elements of characteristic $p>0$. We denote by $A_+$ the set of monic polynomials in $A$. Let $K=\mathbb F_q(\theta)$ be the fraction field of $A$ equipped with the rational point $\infty$. Let $K_\infty$ be the completion of $K$ at $\infty$. We denote by $v_\infty$ the discrete valuation on $K$ corresponding to the place $\infty$ normalized such that $v_\infty(\theta)=-1$, and by $\lvert\cdot\rvert_\infty= q^{-v_\infty}$ the associated absolute value on $K$.
In \cite{Car35} Carlitz introduced the Carlitz zeta values $\zeta_A(n)$ for $n \in \ensuremath \mathbb{N}$ given by
\[ \zeta_A(n) := \sum_{a \in A_+} \frac{1}{a^n} \in K_\infty \]
which are analogues of classical special zeta values in the function field setting. For any tuple of positive integers $\mathfrak s=(s_1,\ldots,s_r) \in \ensuremath \mathbb{N}^r$, Thakur \cite{Tha04} defined the characteristic $p$ multiple zeta value (MZV for short) $\zeta_A(\mathfrak{s})$ or $\zeta_A(s_1,\ldots,s_r)$ by
\begin{equation*}
\zeta_A(\mathfrak{s}):=\sum \frac{1}{a_1^{s_1} \ldots a_r^{s_r}} \in K_\infty
\end{equation*}
where the sum runs through the set of tuples $(a_1,\ldots,a_r) \in A_+^r$ with $\deg a_1>\cdots>\deg a_r$. We call $r$ the depth of $\zeta_A(\mathfrak{s})$ and $w(\mathfrak{s}):
=s_1+\dots+s_r$ the weight of $\zeta_A(\mathfrak{s})$. We note that Carlitz zeta values are exactly depth one MZV's. Thakur \cite{Tha09a} showed that all the MZV's do not vanish. We refer the reader to \cite{AT90,AT09,GP21,Gos96,LRT14,LRT21,Pel12,Tha04,Tha09,Tha10,Tha17,Tha20,Yu91} for more details about these objects.
Thakur proved that the product of two MZV's is a $K$-linear combination of MZV's and we call it the shuffle product in positive characteristic. As in the classical setting, the main goal of the theory is to understand all linear relations over $K$ among MZV's. Analogues of Zagier-Hoffman's conjectures in positive characteristic were formulated by Thakur in \cite[\S 8]{Tha17} and by Todd in \cite{Tod18}. In 2021, the fourth author \cite{ND21} solved these conjectures in the case of small weights. While the algebraic part uses tools introduced by Chen \cite{Che15}, Thakur \cite{Tha10,Tha17} and Todd \cite{Tod18}, the transcendental part uses the theory of $t$-motives and dual motives of Anderson \cite{And86,BP20,HJ20} and a powerful transcendental tool called the Anderson-Brownawell-Papanikolas criterion in \cite{ABP04} (see \cite{Pap08,Cha14,CPY19} for further development). Then the authors \cite{IKLNDP22} have developed a completely new approach and been able to solve these conjectures for all weights. More precisely, we prove (see \cite[Theorem B]{IKLNDP22}):
\begin{theorem}[Zagier's conjecture in positive characteristic]
For $w \in \ensuremath \mathbb{N}$ we denote by $\mathcal Z_w$ the $K$-vector space spanned by the MZV's of weight $w$. Letting
\begin{align*}
d(w)=\begin{cases}
1 & \text{ if } w=0, \\
2^{w-1} & \text{ if } 1 \leq w \leq q-1, \\
2^{w-1}-1 & \text{ if } w=q,
\end{cases}
\end{align*}
we put $d(w)=\sum_{i=1}^q d(w-i)$ for $w>q$. Then for any $w \in \ensuremath \mathbb{N}$, we have
\[ \dim_K \mathcal Z_w = d(w). \]
\end{theorem}
\begin{theorem}[Hoffman's conjecture in positive characteristic]
We keep the above notation. A $K$-basis for $\mathcal Z_w$ is given by $\mathcal T_w$ consisting of $\zeta_A(s_1,\ldots,s_r)$ of weight $w$ with $s_i \leq q$ for $1 \leq i <r$, and $s_r<q$.
\end{theorem}
We mention that in {\it loc. cit.} we also extended these results to the setting of alternating multiple zeta values introduced by Harada \cite{Har21}. We note that the classical alternating multiple zeta values have been studied by Broadhurst, Deligne–Goncharov, Hoffman, Kaneko–Tsumura and many others due to many connections in different contexts. We refer the reader to\cite{Cha21,Del10,Har21,Hof19,Zha16} for some references.
As pointed out by one of the referees of \cite{ND21}, we do not know any algebraic structures of MZV's in positive characteristic (see \cite[Remark 2.2, Part 1]{ND21}). As mentioned above, unlike the classical setting, the proofs of the above theorems are based on new ingredients: some operations introduced by Todd \cite{Tod18} and the fourth author \cite{ND21} as well as a transcendence criterion of Anderson-Brownawell-Papanikolas \cite{ABP04}.
\subsection{Main results} ${}$\par
In this manuscript we present a systematic study of algebraic structures of MZV's in positive characteristic. This paper grew from an attempt to answer the question in \cite[Remark 2.2, Part 1]{ND21} raised by one of the referees of {\it loc. cit.}. It turned out that in a private letter to Thakur in 2017, Deligne \cite{Del17} went further and suggested the existence of a Hopf algebra structure of MZV's in positive characteristic. Subsequently, the composition space which plays the role of Hoffman's algebra in our context was suggested by Shuji Yamamoto \cite{Tha17}, and Shi in \cite{Shi18} formulated a conjectural Hopf algebra structure for this composition space (see \cite[Conjectures 3.2.2 and 3.2.11]{Shi18}).
In this paper we succeed in constructing both the Hopf stuffle algebra and the Hopf shuffle algebra in positive characteristic. In particular, we completely solve all the aforementioned questions and conjectures of Deligne, Thakur and Shi in the previous paragraph. Our approach is based on various tools of analytic, algebraic and combinatorial nature.
Let us give now more precise statements of our results.
\subsubsection{Composition space} \label{section: Composition space} ${}$\par
We introduce the composition space $\frak C$ suggested by Shuji Yamamoto (see \cite[\S 5.2]{Tha17}) which plays the role of the Hoffman algebra $\frak h$ in our context. Let $\Sigma = \{x_n\}_{n \in \mathbb{N}}$ be a countable set equipped with the weight $w(x_n)=n$. The set $\Sigma$ will be called an alphabet and its elements will be called letters. A word over the alphabet $\Sigma$ is a finite string of letters. In particular, the empty word will be denoted by $1$. The depth $\depth(\mathfrak{a})$ of a word $\mathfrak{a}$ is the number of letters in the string of $\mathfrak{a}$, so that $\depth(1) = 0$. The weight of a word is the sum of the weights of its letter and we put $w(1)=0$. Let $\langle \Sigma \rangle$ denote the set of all words over $\Sigma$. We endow $\langle \Sigma \rangle$ with the concatenation product defined by the following formula:
\begin{equation*}
x_{i_1} \dotsc x_{i_n} \cdot x_{j_1} \dotsc x_{j_m} = x_{i_1} \dotsc x_{i_n} x_{j_1} \dotsc x_{j_m}.
\end{equation*}
Let $\mathfrak{C}=\mathbb F_q \langle \Sigma \rangle$ be the free $\mathbb F_q$-vector space with basis $\langle \Sigma \rangle$. The concatenation product extends to $\mathfrak{C}$ by linearity. For a letter $x_a \in \Sigma$ and an element $\mathfrak{a} \in \mathfrak{C}$, we write simply $x_a\mathfrak{a}$ instead of $x_a \cdot \mathfrak{a}$. For each nonempty word $\mathfrak{a} \in \langle \Sigma \rangle$, we can write $\mathfrak{a} = x_a \mathfrak{a}_-$ where $x_a$ is the first letter of $\mathfrak{a}$ and $\mathfrak{a}_-$ is the word obtained from $\mathfrak{a}$ by removing $x_a$.
\subsubsection{Shuffle algebra and shuffle map} ${}$\par
We define the unit $u:\mathbb F_q \to \frak C$ by sending $1$ to the empty word $1$. Next we define recursively two products on $\mathfrak{C}$ as $\mathbb F_q$-bilinear maps
\begin{align*}
\diamond \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C} \quad \text{and} \quad
\shuffle \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C}
\end{align*}
by setting $1 \diamond \mathfrak{a} = \mathfrak{a} \diamond 1 = \mathfrak{a}, 1 \shuffle \mathfrak{a} = \mathfrak{a} \shuffle 1 = \mathfrak{a}$ and
\begin{align*}
\mathfrak{a} \diamond \fb &= \ x_{a + b}(\mathfrak{a}_- \shuffle \fb_-) + \sum\limits_{i+j = a + b} \Delta^j_{a,b} x_i(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)),\\
\mathfrak{a} \shuffle \fb &= \ x_{a}(\mathfrak{a}_- \shuffle \fb) + x_{b}(\mathfrak{a} \shuffle \fb_-) + \mathfrak{a} \diamond \fb,
\end{align*}
for any words $\mathfrak{a},\fb \in \langle \Sigma \rangle$,
Here the coefficients $\Delta^i_{a,b}$ are given by
\begin{align*}
\Delta^i_{a,b}=\begin{cases} (-1)^{a-1} {i-1 \choose a-1}+(-1)^{b-1} {i-1 \choose b-1} & \quad \text{if } (q-1) \mid i \text{ and } 0<i<a+b, \\
0 & \quad \text{otherwise}.
\end{cases}
\end{align*}
We call $\diamond$ the diamond product and $\shuffle $ the shuffle product.
Our first result gives an affirmative answer to both questions in \cite[Remark 2.2, Part 1]{ND21} and \cite[Conjectures 3.2.2 and 3.2.11]{Shi18}. It reads as follows (see Theorems \ref{theorem: algebras} and \ref{thm: shuffle map}):
\begin{theoremx} \label{thm: A}
The spaces $(\mathfrak{C}, \diamond)$ and $(\mathfrak{C}, \shuffle )$ are commutative $\mathbb F_q$-algebras. Further, for all words $\mathfrak{a}, \fb \in \frak C$ we have
\begin{align*}
\zeta_A(\mathfrak{a}\shuffle \fb) =\zeta_A(\mathfrak{a}) \, \zeta_A(\fb).
\end{align*}
If we denote by $\mathcal Z$ the $K$-vector space spanned by MZV's, then the homomorphism of $K$-algebras
\begin{align*}
Z_\shuffle:\frak C \otimes_{\mathbb F_q} K &\to \mathcal Z \\
\mathfrak{a} &\mapsto \zeta_A(\mathfrak{a})
\end{align*}
is called the shuffle map in positive characteristic.
\end{theoremx}
To prove Theorem \ref{thm: A}, we are reduced to prove that the diamond and shuffle products are associative. It turns out that this claim is very hard to prove since a direct check involves complicated combinatorics as already noticed by Shi \cite[\S 3.1]{Shi18}. Our method uses analytic tools and consists of unpacking the coefficients $\Delta^i_{a,b}$ involved in the definition of the diamond product. Then we use the uniqueness of partial fraction decomposition to prove the desired associativity.
\subsubsection{Shuffle Hopf algebra} ${}$\par
We also define recursively a product on $\mathfrak{C}$ as a $\mathbb F_q$-bilinear map
\begin{align*}
\triangleright \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C}
\end{align*}
by setting $1 \triangleright \mathfrak{a} = \mathfrak{a} \triangleright 1 = \mathfrak{a}$ and
\begin{align*}
\mathfrak{a} \triangleright \fb &= \ x_{a}(\mathfrak{a}_- \shuffle \fb)
\end{align*}
for any words $\mathfrak{a},\fb \in \langle \Sigma \rangle$. We call $\triangleright$ the triangle product. We stress that the triangle product is neither commutative nor associative. Inspired by the work of Shi \cite[\S 3.2.3]{Shi18} we define a coproduct
\[ \Delta: \frak C \to \frak C \otimes \frak C. \]
using $\triangleright$ rather than the concatenation on recursive steps for words with depth $>1$ (see \S \ref{sec: coproduct}). The counit $\epsilon:\frak C \to \mathbb F_q$ is defined as follows: $\epsilon(1)=1$ and $\epsilon(\fu)=0$ otherwise.
We note that for quasi-shuffle algebras introduced by Hoffman \cite{Hof00} and their generalization, the coproduct is roughly speaking the deconcatenation. The coproduct~$\Delta$ defined as above is completely different from the deconcatenation and involves complicated combinatorics. We refer the reader to Proposition \ref{proposition: Hopf deltas depth 1 <= q^2} and Appendix~\ref{sec: numerical experiments} for numerical calculations of $\Delta$.
Our second result shows that this construction gives rise to a Hopf algebra structure of the shuffle algebra (see Theorem \ref{thm: Hopf algebra for shuffle product}).
\begin{theoremx} \label{thm: B}
The connected graded bialgebra $(\frak C,\shuffle,u,\Delta,\epsilon)$ is a connected graded Hopf algebra of finite type over $\mathbb F_q$.
\end{theoremx}
The proof of Theorem \ref{thm: B} is of algebraic nature by exploiting key properties among the diamond, shuffle and triangle products.
\subsubsection{Comparison with Shi's construction} ${}$\par
Next we study the coproduct $\Delta$ for letters in detail and prove some key properties in Proposition \ref{prop: formula delta xn}. As an immediate consequence, we deduce that the coproduct~$\Delta$ coincides with the coproduct introduced by Shi in \cite[\S 3.2.3]{Shi18}. Thus we settle Conjecture 3.2.11 of \cite{Shi18} (see Theorem \ref{thm: comparison with Shi's coproduct}).
\begin{theoremx} \label{thm: C}
Conjecture 3.2.11 in \cite{Shi18} holds.
\end{theoremx}
We note that to convince ourselves its validity, we made intensive numerical calculations (see Proposition \ref{proposition: Hopf deltas depth 1 <= q^2} and the Appendix \ref{sec: numerical experiments}). The proof of Theorem~\ref{thm: C} is based on combinatorial techniques.
\subsubsection{Stuffle algebra and stuffle Hopf algebra} ${}$\par
The stuffle algebra is easier to define. We introduce the stuffle product in the same way as that of $(\frak h^1,*)$ as above. The $*$ product
\begin{align*}
* \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C}
\end{align*}
is given by setting $1 * \mathfrak{a} = \mathfrak{a} * 1 = \mathfrak{a}$ and
\begin{align*}
\mathfrak{a} * \fb &= \ x_{a}(\mathfrak{a}_- * \fb) + x_{b}(\mathfrak{a} * \fb_-) + x_{a+b} (\mathfrak{a}_- * \fb_-)
\end{align*}
for any words $\mathfrak{a},\fb \in \langle \Sigma \rangle$. We call $*$ the \textit{stuffle product} and see that $(\frak C,*)$ is a commutative $\mathbb F_q$-algebra.
We now define a coproduct $\Delta_*:\frak C \to \frak C \otimes \frak C$ and a counit $\epsilon:\frak C \to \mathbb F_q$ by
\[ \Delta_*(w)= \sum_{uv=w} u \otimes v \]
and
\begin{align*}
\epsilon(w)=\begin{cases} 1 \quad \text{if } w=0, \\ 0 \quad \text{otherwise}, \end{cases}
\end{align*}
for any words $w \in \langle \Sigma \rangle$.
We deduce from the work of Hoffman \cite{Hof00} that
\begin{theoremx} \label{thm: D}
The stuffle algebra $(\frak C,*,u,\Delta_*,\epsilon)$ is a connected graded Hopf algebra of finite type over $\mathbb F_q$.
\end{theoremx}
\subsubsection{Stuffle map} ${}$\par
Finally, using our previous works \cite{IKLNDP22,ND21} we know that there is a connection between MZV's of Thakur and Carlitz's multiple polylogarithms. Thus we are able to construct a homomorphism of $K$-algebras called the stuffle map (see \S \ref{sec: stuffle map}).
\begin{theoremx} \label{thm: E}
Recall that $\mathcal Z$ is the $K$-vector space spanned by MZV's. Then there exists a homomorphism of $K$-algebras
\[ Z_*:\frak C \otimes_{\mathbb F_q} K \to \mathcal Z \]
called the stuffle map in positive characteristic.
\end{theoremx}
\subsection{Organization of the paper} ${}$\par
We briefly explain the plan of the manuscript.
\begin{itemize}
\item In \S \ref{sec: Hopf algebras} we recall the definition and basic facts of different notions of algebras (algebras, coalgebras, bialgebras and Hopf algebras) that will be used in this paper.
\item In \S \ref{sec: classical MZV's} we present the stuffle algebra and the shuffle algebra of MZV's in the classical setting. We recall different zeta maps and Hopf algebra structures associated to these algebras as well as important conjectures and theorems concerning these objects.
\item In \S \ref{sec: Thakur MZV's} we introduce the notion of MZV's in positive characteristic and define the composition space that is an analogue of Hoffman's algebra in this context.
\item In \S \ref{sec: shuffle algebra 0} and \S \ref{sec: algebra structure} we define the shuffle product $\shuffle$ and the shuffle map (see Theorem \ref{thm: shuffle map}). Using partial fraction decompositions we prove that the composition space equipped with the shuffle map is a commutative $\mathbb F_q$-algebra (see Theorem \ref{theorem: algebras}).
\item In \S \ref{sec: Hopf algebra structure} we define the coproduct $\Delta$ and the counit on the shuffle algebra. We prove that these give a Hopf algebra structure of the shuffle algebra (see Theorem \ref{thm: Hopf algebra for shuffle product}). In \S \ref{sec: depth one} we study the coproduct for words of depth one in detail and deduce that the coproduct $\Delta$ coincide with that introduced by Shi (see Theorem \ref{thm: comparison with Shi's coproduct}). Explicit formulas for the coproduct of such words are given in many cases in \S \ref{sec: explicit formula for small weights} and the Appendix \ref{sec: numerical experiments}.
\item Finally, in \S \ref{sec: stuffle algebra} we introduce the stuffle product $*$ and get the algebra and the Hopf algebra structures of the stuffle product as well as the stuffle map in our context (see Theorem \ref{thm: Hopf algebra for stuffle product}). The key ingredient is a connection between MZV's and Carlitz multiple polylogarithms as explained in \cite{ND21} and \cite{IKLNDP22}.
\end{itemize}
\subsection*{Acknowledgments}
The fourth author (T. ND.) would like to thank Dinesh Thakur for sharing a private letter of Deligne and for helpful discussions.
The first named author (B.-H. Im) was supported by the National Research Foundation of Korea (NRF) grant no.~2020R1A2B5B01001835 funded by the Korea government (MSIT). Two of the authors (KN. L. and T. ND.) were partially supported by the Excellence Research Chair ``$L$-functions in positive characteristic and applications'' financed by the Normandy Region. T. ND. and LH. P. were par-
tially supported by the Vietnam Academy of Science and Technology (VAST) under
grant no.~CTTH00.02/23-24 ``Arithmetic and Geometry of schemes over function fields and applications''.
\section{Review of Hopf algebras} \label{sec: Hopf algebras}
We briefly review the notion of Hopf algebras and follow closely the presentation of \cite[\S 3.2]{BGF}. Throughout this section, we let $k$ denote a ground field. Unless otherwise specified, all tensor products will be assumed to be over $k$.
\subsection{Hopf algebras} ${}$\par
For each $k$-vector space $H$, we denote by $\iota \colon H \otimes H \rightarrow H \otimes H $ the transposition map given by $x \otimes y \mapsto y \otimes x$.
\begin{definition}
An \textit{algebra} over $k$ is a triple $(H, m,u)$ consisting of a $k$-vector space $H$ together with $k$-linear maps $m \colon H \otimes H \rightarrow H$ called the \textit{multiplication} and $u \colon k \rightarrow H$ called the \textit{unit} such that the following diagrams are commutative:
\begin{enumerate}[$(1)$]
\item associativity
\begin{center}
\begin{tikzcd}
H\otimes H \otimes H \arrow[r, "m \otimes \text{id}"] \arrow[d, " \text{id} \otimes m "'] & H \otimes H \arrow[d, "m"] \\
H \otimes H \arrow[r, "m"] & H
\end{tikzcd}
\end{center}
\item unitary
\begin{center}
\begin{tikzcd}
H \otimes k \arrow[rd] \arrow[r, "\text{id} \otimes u"] & H \otimes H \arrow[d, "m"] & k \otimes H \arrow[ld] \arrow[l, "u \otimes \text{id}"'] \\
& H & \end{tikzcd}
\end{center}
where the diagonal arrows are canonical isomorphisms.
\end{enumerate}
The algebra is said to be \textit{commutative} if the following diagram is commutative:
\begin{center}
\begin{tikzcd}
H \otimes H \arrow[rd, "m"'] \arrow[rr, "\iota"] & & H \otimes H \arrow[ld, "m"] \\
& H &
\end{tikzcd}
\end{center}
\end{definition}
Turning all arrows around, one obtains the definition of coalgebras over $k$.
\begin{definition}
A \textit{coalgebra} over $k$ is a triple $(H, \Delta,\epsilon)$ consisting of a $k$-vector space $H$ together with $k$-linear maps $\Delta \colon H \rightarrow H \otimes H$ called the \textit{coproduct} and $\epsilon \colon H \rightarrow k$ called the \textit{counit} such that the following diagrams are commutative:
\begin{enumerate}[$(1)$]
\item coassociativity
\begin{center}
\begin{tikzcd}
H \otimes H\otimes H & \otimes H \arrow[l, "\Delta\otimes \text{id} "'] \\
H \otimes H \arrow[u, "\text{id} \otimes \Delta"] & {H} \arrow[u, "\Delta"'] \arrow[l, "\Delta"']
\end{tikzcd}
\end{center}
\item counitary
\begin{center}
\begin{tikzcd}
H \otimes k & H \otimes H \arrow[l, "\text{id} \otimes \epsilon "'] \arrow[r, "\epsilon \otimes \text{id}"] & k \otimes H \\
& H \arrow[u, "\Delta"'] \arrow[ru] \arrow[lu] &
\end{tikzcd}
\end{center}
where the diagonal arrows are canonical isomorphisms.
\end{enumerate}
The coalgebra is said to be \textit{cocommutative} if the following diagram is commutative:
\begin{center}
\begin{tikzcd}
H \otimes H \arrow[rr, "\iota"] & & H \otimes H \\
& H \arrow[ru, "\Delta"'] \arrow[lu, "\Delta"] &
\end{tikzcd}
\end{center}
\end{definition}
\begin{definition}
A \textit{bialgebra} over $k$ is a tuple $(H,m, u, \Delta,\epsilon)$ consisting of an algebra $(H,m, u)$ over $k$ and a coalgebra $(H,\Delta,\epsilon)$ over $k$ which are compatible, i.e., the following diagrams are commutative:
\begin{enumerate}[$(1)$]
\item product and coproduct
\begin{center}
\begin{tikzcd}
H \otimes H \arrow[d, "\Delta \otimes \Delta"'] \arrow[r, "m"] & H \arrow[r, "\Delta"] & H \otimes H \\
H \otimes H \otimes H \otimes H \arrow[rr, "\text{id} \otimes \iota \otimes \text{id}"] & & H \otimes H \otimes H \otimes H \arrow[u, "m \otimes m"']
\end{tikzcd}
\end{center}
\item unit and coproduct
\begin{center}
\begin{tikzcd}
H \arrow[r, "\Delta"] & H \otimes H \\
k \arrow[u, "u"] \arrow[r] & k \otimes k \arrow[u, "u \otimes u"']
\end{tikzcd}
\end{center}
\item counit and product
\begin{center}
\begin{tikzcd}
H \arrow[d, "\epsilon"'] & H \otimes H \arrow[d, "\epsilon \otimes \epsilon"] \arrow[l, "m"'] \\
k & k \otimes k \arrow[l]
\end{tikzcd}
\end{center}
\item unit and counit
\begin{center}
\begin{tikzcd}
k \arrow[rr, "\text{id}"] \arrow[rd, "u"'] & & k \\
& H \arrow[ru, "\epsilon"'] &
\end{tikzcd}
\end{center}
where the bottom arrows in the second diagram and the third diagram are canonical isomorphisms.
\end{enumerate}
\end{definition}
\begin{definition}
A \textit{Hopf algebra} over $k$ is a bialgebra $(H,m, u, \Delta,\epsilon)$ over $k$ together with a $k$-linear map $S\colon H \rightarrow H$ called \textit{antipode} such that the following diagram is commutative:
\begin{center}
\begin{tikzcd}
& H \otimes H \arrow[rr, "S \otimes \text{id}"] & & H \otimes H \arrow[rd, "m"] & \\
H \arrow[rr, "\epsilon"] \arrow[ru, "\Delta"] \arrow[rd, "\Delta"'] & & k \arrow[rr, "u"] & & H \\
& H \otimes H \arrow[rr, "\text{id}\otimes S "] & & H \otimes H \arrow[ru, "m"'] &
\end{tikzcd}
\end{center}
\end{definition}
We note that a bialgebra does not always admit an antipode (see \cite[Exercise 3.83]{BGF} for an example).
\subsection{Graded Hopf algebras} ${}$\par
In this section we introduce the notion of connected graded bialgebras which will be useful. We will see later that every connected graded bialgebra has an antipode, and hence a Hopf algebra structure.
\begin{definition} \label{defn: graded Hopf algebra}
\text{ }
\begin{enumerate}[$(1)$]
\item A bialgebra $(H,m, u, \Delta,\epsilon)$ over $k$ is said to be \textit{graded} if one can write $H$ as a direct sum of $k$-vectors subspaces
\begin{equation*}
H = \bigoplus \limits_{n = 0}^{\infty}H_n,
\end{equation*}
such that for all integers $r,s\geq 0$, we have
\begin{equation*}
m(H_r \otimes H_s) \subseteq H_{r + s} \quad \text{and} \quad \Delta(H_r) \subseteq \bigoplus \limits_{i + j = r} H_i \otimes H_j.
\end{equation*}
A graded bialgebra is said to be \textit{connected} if $H_0 = k$.
\item A \textit{graded Hopf algebra} is a Hopf algebra $H$ whose the underlying bialgebra is graded and the antipode $S$ satisfies $S(H_n) \subseteq H_n$.
\item A graded Hopf algebra is said to be \textit{connected} if $H_0 = k$.
\item A graded Hopf algebra is said to be {\it of finite type} if $H_n$ is a $k$-vector space of finite dimension.
\end{enumerate}
\end{definition}
The following proposition shows that a connected graded bialgebra automatically admits an antipode, thus it is always a Hopf algebra. It is given as an exercise in \cite[\S 11.2]{Swe69} (see also \cite[Exercise 3.84]{BGF}).
\begin{proposition} \label{prop: graded Hopf algebras}
Let $(H,m, u, \Delta,\epsilon)$ be a connected graded bialgebra over $k$.
\begin{enumerate}[$(1)$]
\item For each element $x \in H_n$ with $n \geq 1$, we have
\begin{equation*}
\Delta(x) = 1 \otimes x + x \otimes 1 + \sum x_{(1)} \otimes x_{(2)},
\end{equation*}
where $\sum x_{(1)} \otimes x_{(2)} \in \bigoplus \limits_{\substack{i,j > 0 \\ i + j = n}} H_i \otimes H_j$. Moreover, the counit $\epsilon$ vanishes on~$H_n$ for all $n \geq 1$.
\item We continue the notation as in (1) and define recursively a $k$-linear map $S \colon H \rightarrow H$ given by
\begin{align*}
S(x) = \begin{cases} x & \quad \text{if } x \in H_0, \\
- x - \sum m(S(x_{(1)}) \otimes x_{(2)}) & \quad \text{if } x \in H_n \text{ with } n \geq 1.
\end{cases}
\end{align*}
Then $H$ is a graded Hopf algebra whose antipode is $S$.
\end{enumerate}
\end{proposition}
\begin{proof}
See \cite[Lemma 2.1]{Ehr96}.
\end{proof}
\section{Classical multiple zeta values} \label{sec: classical MZV's}
In this section we review classical multiple zeta values studied by Euler in the late eighteenth century. In \S \ref{sec: quasi-shuffle algebras} we recall the theory of quasi-shuffle algebras introduced by Hoffman in 2000 and give basic facts such as the associated Hopf algebra structure. The stuffle algebra and the shuffle one defined in \S \ref{sec: Hoffman algebra} and \S \ref{sec: stuffle and shuffle algebras} are examples of this class. We explain their connection with MZV's and the regularization of Ihara-Kaneko-Zagier in \S \ref{sec: MZV's} and \S \ref{sec: regularization}.
\subsection{Quasi-shuffle algebras} ${}$\par \label{sec: quasi-shuffle algebras}
We review the notion of quasi-shuffle product introduced by Hoffman \cite{Hof00}.
Let $\Sigma = \{x_i\}_{i \in \mathbb{N}}$ be a countable set. To each letter $x_i$ we associate a weight $w(x_i) \in \mathbb N$ and we suppose that for any $n \in \mathbb N$ the set $\Sigma_n$ of letters of weight $n$ is finite. In this context, we follow the notations in Section \ref{section: Composition space}.
Let $\langle \Sigma \rangle$ denote the set of all words over $\Sigma$.
We denote by $\mathbb Q \langle \Sigma \rangle$ (resp. $\mathbb Q \Sigma$) the $\mathbb Q$-vector space with $\langle \Sigma \rangle$ (resp. $\Sigma$) as a basis. The concatenation product extends to $\mathbb Q \langle \Sigma \rangle$ by linearity so that $\mathbb Q \langle \Sigma \rangle$ is a graded algebra with respect to weight.
We set $\bar \Sigma=\Sigma \cup \{0\}$. Let $\diamond:\bar \Sigma \times \bar \Sigma \to \bar \Sigma$ be a commutative and associative product which preserves the grading. It means that this map satisfies the following properties: for all $a,b,c \in \Sigma$,
\begin{itemize}
\item $a \diamond 0=0$.
\item $a \diamond b=b \diamond a$.
\item $(a \diamond b) \diamond c=a \diamond (b \diamond c)$.
\item Either $a \diamond b=0$ or $w(a \diamond b)=w(a)+w(b)$.
\end{itemize}
We define a new product $*_\diamond$ on $\mathbb Q \langle \Sigma \rangle$ recursively by setting $1 *_\diamond \fu=\fu *_\diamond 1=\fu$, and
\[ a\fu *_\diamond b\fv=a(\fu *_\diamond b\fv)+b(a\fu *_\diamond \fv)+(a \diamond b)(\fu *_\diamond \fv)\]
for all letter $a,b \in \Sigma$ and all words $\fu,\fv \in \mathbb Q \langle \Sigma \rangle$. This product is called the quasi-shuffle product associated to $\diamond$. Hoffman \cite[Theorem 2.1]{Hof00} showed that the vector space $\mathbb Q \langle \Sigma \rangle$ equipped with the product $*_\diamond$ is a commutative $\mathbb Q$-algebra.
We now define a coproduct $\Delta:\mathbb Q \langle \Sigma \rangle \to \mathbb Q \langle \Sigma \rangle \otimes \mathbb Q \langle \Sigma \rangle$ and a counit $\epsilon:\mathbb Q \langle \Sigma \rangle \to \mathbb Q$ by
\[ \Delta(\fu)= \sum_{\mathfrak{a} \fb=\fu} \mathfrak{a} \otimes \fb \]
and
\begin{align*}
\epsilon(\fu)=\begin{cases} 1 \quad \text{if } \fu=1, \\ 0 \quad \text{otherwise}, \end{cases}
\end{align*}
for all words $\fu \in \langle \Sigma \rangle$. Hoffman \cite[Theorem 3.1]{Hof00} showed that $\mathbb Q \langle \Sigma \rangle$ equipped with the $*_{\diamond}$-multiplication and $\Delta$-comultiplication is a bialgebra. Since both $*_{\diamond}$ and $\Delta$ respect the grading, Proposition \ref{prop: graded Hopf algebras} implies
\begin{theorem} \label{theorem: Hopf algebras for qp}
The algebra $\mathbb Q \langle \Sigma \rangle$ with the $*_{\diamond}$-multiplication and $\Delta$-comultiplication is a graded Hopf algebra. Further, it is connected and of finite type.
\end{theorem}
Moreover, the antipode $S:\mathbb Q \langle \Sigma \rangle \to \mathbb Q \langle \Sigma \rangle$ is given explicitly in \cite[Theorem~3.2]{Hof00}: for any word $\fu=x_{i_1} \dots x_{i_n}$ we have
\[ S(\fu)= \sum_{(j_1,\dots,j_k)} (-1)^k x_{i_1} \dots x_{i_{j_1}}*_{\diamond}x_{i_{j_1+1}} \dots x_{i_{j_1+j_2}}*_{\diamond} \dots *_{\diamond} x_{i_{j_1+\dots+j_{k-1}+1}} \dots x_{i_{j_1+\dots+j_k}}\]
where the sum runs through the set of all partitions $(j_1,\dots,j_k)$ of $n$.
For recent developments on quasi-shuffle products, we refer the reader to \cite{Hof15,Hof20,HI17,IKOO11}.
\subsection{Multiple zeta values of Euler} ${}$\par \label{sec: MZV's}
We now illustrate two examples of quasi-product that are related to the classical MZVs. Recall that multiple zeta values of Euler (MZV's for short) are real positive numbers given by the following convergent series
\[ \zeta(n_1,\dots,n_r)=\sum_{0<k_1<\dots<k_r} \frac{1}{k_1^{n_1} \dots k_r^{n_r}}, \quad \text{where } n_i \geq 1, n_r \geq 2. \]
Here $r$ is called the depth and $w=n_1+\dots+n_r$ is called the weight of the presentation $\zeta(n_1,\dots,n_r)$. When $r=1$, we recover the special zeta values $\zeta(n)$ for $n \geq 2$ of the Riemann zeta function. It was studied by Euler in the eighteenth century and have been studied intensively especially in the last three decades by mathematicians and physicists. We refer the reader to \cite{BGF, Zag94} for more details. As mentioned in these references, the main goal of this theory is to understand all $\mathbb Q$-linear relations among MZV's. We note that precise conjectures formulated by Zagier \cite{Zag94} and Hoffman \cite{Hof97} give the dimension and an explicit basis for the $\mathbb Q$-vector space $\mathcal{Z}_k$ spanned by MZV's of weight $k$ for $k \in \mathbb{N}$. The algebraic part of these conjectures was completely settled by Brown \cite{Bro12}, Deligne-Goncharov \cite{DG05} and Terosoma \cite{Ter02}.
\subsection{The Hoffman algebra, stuffle product and shuffle product} ${}$\par \label{sec: Hoffman algebra}
In this section we take $k=\mathbb Q$. Let $X$ be the alphabet with two letters $x_0, x_1$ with weight 1, that means $w(x_0)=w(x_1)=1$. We denote $\frak h=\mathbb Q\langle X \rangle$ and call it the Hoffman algebra. A word in the alphabet $X$ is said to be positive if it is of the form $x_1 \fu$ and is said to be admissible if it is of the form $x_1 \fu x_0$. We denote by $\frak h^1$ (resp. $\frak h^0$) the subspace of $\frak h$ spanned by positive words (resp. admissible words).
For all $i \in \mathbb N$ we put $z_i=x_1x_0^{i-1}$. Then $w(z_i)=i$. Let $Z$ be the alphabet with letters $\{z_i\}_{i \in \mathbb{N}}$. Then $\frak h^1=\mathbb Q \langle Z \rangle$. We now equip the alphabet $Z$ with the commutative and associative product $\diamond:Z \times Z \to Z$ given by
\[ z_i \diamond z_j=z_{i+j} \]
for all $i, j \in \mathbb N$. The associated quasi-product on $\frak h^1=\mathbb Q \langle Z \rangle$ will be denoted by $*$ and called the stuffle product. A word in $\frak h^1$ is called admissible if it can be expressed as $z_{s_1} \dots z_{s_\ell}$ with $s_\ell>1$. We note that $\frak h^0$ is the subspace generated by admissible words in $\frak h^1$ and that $(\frak h^0,*)$ is a subalgebra of $(\frak h^1,*)$. Further, the harmonic product on MZV's gives rise to a homomorphism of $\mathbb Q$-algebras
\[ \zeta_*:\frak h^0 \to \mathbb R \]
which sends an admissible word $z_{s_1} \dots z_{s_\ell}$ to the associated zeta value $\zeta(s_1,\dots,s_r)$, that means
\[ \zeta_*(\fu*\fv)=\zeta_*(\fu) \zeta_*(\fv) \]
for all words $\fu,\fv \in \frak h^0$. This map is called the stuffle zeta map.
We now recall the shuffle algebra. We endow $X$ with the trivial product $\diamond:X \times X \to X$ given by
\[ a \diamond b=0 \]
for all $a,b \in X$. The associated quasi-product on $\frak h=\mathbb Q \langle X \rangle$ will be denoted by $\shuffle$ and called the shuffle product. We see that $(\frak h^0,\shuffle)$ and $(\frak h^1,\shuffle)$ are subalgebras of $(\frak h,\shuffle)$. The shuffle product on MZV's defines a homomorphism of $\mathbb Q$-algebras
\[ \zeta_\shuffle:\frak h^0 \to \mathbb R \]
which sends an admissible word $z_{s_1} \dots z_{s_\ell}$ to the associated zeta value $\zeta(s_1,\dots,s_r)$, that means
\[ \zeta_\shuffle(\fu \shuffle \fv)=\zeta_\shuffle(\fu) \zeta_\shuffle(\fv) \]
for all words $w,v \in \frak h^0$. This map is called the shuffle zeta map.
Using these zeta maps yield the so-called double shuffle relations in the convergent case: for all words $\fu,\fv \in \frak h^0$,
\[ \quad \zeta_*(\fu * \fv)=\zeta_\shuffle(\fu \shuffle \fv). \]
\subsection{Regularized zeta maps} ${}$\par \label{sec: regularization}
Following Ihara-Kaneko-Zagier \cite{IKZ06}, we note that the homomorphism of $(\frak h^0,*)$-algebras $\varphi_*:\frak h^0[T] \to \frak h^1$ which sends $T$ to $z_1$ is an isomorphism. Further, the following homomorphisms of $(\frak h^0,\shuffle)$-algebras
\begin{align*}
& \varphi_{\shuffle}:\frak h^0[T] \to \frak h^1, \quad T \mapsto x_1, \\
& \varphi_{\shuffle}:\frak h^0[T,U] \to \frak h, \quad T \mapsto x_1, \, U \mapsto x_0,
\end{align*}
are isomorphisms.
Now we define the stuffle regularized zeta map
\begin{equation} \label{eq: stuffle zeta map}
\zeta_*:\frak h^1 \to \mathbb R
\end{equation}
as the composition
\[ \frak h^1 \to \frak h^0[T] \to \mathbb R[T] \to \mathbb R \]
where the first map is $\varphi_*^{-1}$, the second map is induced by the stuffle zeta map and the last one is the evaluation at $T=0$. Similarly, we define the shuffle regularized zeta map
\begin{equation} \label{eq: shuffle zeta map}
\zeta_\shuffle:\frak h \to \mathbb R
\end{equation}
as the composition
\[ \frak h \to \frak h^0[T,U] \to \mathbb R[T,U] \to \mathbb R \]
where the first map is $\varphi_{\shuffle}^{-1}$, the second map is induced by the shuffle zeta map and the last one is the evaluation at $T=U=0$.
We mention that Ihara, Kaneko and Zagier \cite{IKZ06} use the restriction of these maps on $\frak h^1$ to extend the previous double shuffle relations among MZV's. Further, they conjectured that these extended double shuffle relations exhaust all linear relations among MZV's (see Conjecture \ref{conj: IKZ}). As mentioned in the Introduction, other important conjectures in this theory are those of Zagier and Hoffman which give precise dimension and a basis for the $\mathbb{Q}$-vector space spanned MZV's of fixed weight (see Conjectures \ref{conj: Zagier} and \ref{conj: Hoffman}).
Recall that the algebraic part of these conjectures was solved by Terasoma \cite{Ter02}, Deligne-Goncharov \cite{DG05} and Brown \cite{Bro12} using the theory of mixed Tate motives (see Theorems \ref{thm: Zagier} and \ref{thm: Hoffman}). The proofs of this theorem use by a crucial manner different Hopf algebra structures of Hoffman's algebra $\frak h$ as described above. We mention that the transcendental part which concerns lower bounds for $\dim_{\mathbb Q} \mathcal Z_k$ is completely open. We refer the reader to \cite{BGF,Del13,Zag94} for more details and more exhaustive references.
\subsection{Stuffle Hopf algebra and shuffle Hopf algebra} ${}$\par \label{sec: stuffle and shuffle algebras}
By the work of Hoffman \cite{Hof00} the above algebras can be endowed with a richer structure, i.e., that of Hopf algebras. In fact, as a direct consequence of Theorem~\ref{theorem: Hopf algebras for qp}, we get two Hopf algebras for classical MZV's.
\subsubsection{} The first graded Hopf algebra
\[ H_*=(\frak h^1,*) \]
comes from the stuffle product. We note that it is related to the algebra of quasi-symmetric functions over $k$ (see \cite{Ehr96,Hof15}). For some applications of Hopf algebra structure, we refer the reader to \cite{Hof15} (see also \cite{KXY21}).
\subsubsection{} The second graded Hopf algebra
\[ H_\shuffle=(\frak h,\shuffle) \]
is the shuffle algebra (see \cite{Reu93}). Explicitly,
\begin{itemize}
\item $\frak h=\mathbb Q\langle x_0,x_1 \rangle$.
\item The coproduct is given by the shuffle product $\shuffle$.
\item The unit is given by the empty word $1$.
\item The coproduct $\Delta:\frak h \to \frak h \otimes \frak h$ is given by the deconcatenation
\[ \Delta(\fu)= \sum_{\mathfrak{a} \fb=\fu} \mathfrak{a} \otimes \fb \]
for any words $\fu \in \frak h$.
\item The counit $\epsilon:\frak h \to \mathbb Q$ is given by
\begin{align*}
\epsilon(\fu)=\begin{cases} 1 \quad \text{if } \fu=1, \\ 0 \quad \text{otherwise}. \end{cases}
\end{align*}
\item The antipode $S:\frak h \to \frak h$ is given by
\[ S(x_{i_1} \dots x_{i_n})=(-1)^n x_{i_n} \dots x_{i_1}. \]
\end{itemize}
This Hopf algebra and its motivic version introduced by Goncharov \cite{Gon05} lie in the heart of the works of Brown \cite{Bro12}, Deligne-Goncharov \cite{DG05} and Terosoma \cite{Ter02} (see also \cite{BGF}).
\section{Multiple zeta values in positive characteristic} \label{sec: Thakur MZV's}
In \S \ref{sec: MZV's of Thakur} we recall the notion of multiple zeta values of Thakur and present the main goal and results of this theory. Then we define the composition space which plays the role of Hoffman's algebra in the function field setting (see \S \ref{sec: composition space}).
\subsection{Multiple zeta values in positive characteristic} ${}$\par \label{sec: MZV's of Thakur}
By analogy between number fields and function fields, Carlitz \cite{Car35} introduced zeta values in positive characteristic which are studied extensively in the last three decades. More recently, Thakur \cite{Tha04} generalized the work of Carlitz and defined analogues of multiple zeta values in positive characteristic. We recall some notations in the Introduction. The ring $A=\mathbb F_q[\theta]$ is the polynomial ring in the variable $\theta$ over a finite field $\mathbb F_q$ of $q$ elements of characteristic $p>0$. We recall that $A_+$ denotes the set of monic polynomials in $A$ and $K=\mathbb F_q(\theta)$ is the fraction field of $A$ equipped with the rational point $\infty$. Then $K_\infty$ is the completion of $K$ at $\infty$ and $\mathbb{C}_\infty$ is the completion of a fixed algebraic closure $\overline K$ of $K$ at $\infty$. We denote by $v_\infty$ the discrete valuation on $K$ corresponding to the place $\infty$ normalized such that $v_\infty(\theta)=-1$, and by $\lvert\cdot\rvert_\infty= q^{-v_\infty}$ the associated absolute value on $K$. The unique valuation of~$\mathbb C_\infty$ which extends $v_\infty$ will still be denoted by $v_\infty$.
For any tuple of positive integers $\mathfrak s=(s_1,\ldots,s_r) \in \ensuremath \mathbb{N}^r$, Thakur \cite{Tha04} defined the characteristic $p$ multiple zeta value (MZV for short) $\zeta_A(\mathfrak{s})$ or $\zeta_A(s_1,\ldots,s_r)$ by
\begin{equation*}
\zeta_A(\mathfrak{s}):=\sum \frac{1}{a_1^{s_1} \ldots a_r^{s_r}} \in K_\infty
\end{equation*}
where the sum runs through the set of tuples $(a_1,\ldots,a_r) \in A_+^r$ with $\deg a_1>\cdots>\deg a_r$. We call $r$ the depth of $\zeta_A(\mathfrak{s})$ and $w(\mathfrak{s})=s_1+\dots+s_r$ the weight of $\zeta_A(\mathfrak{s})$. We note that Carlitz zeta values are exactly depth one MZV's. Thakur \cite{Tha09a} showed that all the MZV's do not vanish. As in the classical setting, the main goal of the theory is to understand all linear relations over $K$ among MZV's. In fact, analogues of Zagier-Hoffman's conjectures in positive characteristic were formulated by Thakur in \cite[\S 8]{Tha17} and by Todd in \cite{Tod18} that we recall below.
For $w \in \ensuremath \mathbb{N}$ we denote by $\mathcal Z_w$ the $K$-vector space spanned by the MZV's of weight $w$. We denote by $\mathcal T_w$ the set of $\zeta_A(\mathfrak{s})$ where $\mathfrak{s}=(s_1,\ldots,s_r) \in \ensuremath \mathbb{N}^r$ of weight $w$ with $1\leq s_i\leq q$ for $1\leq i\leq r-1$ and $s_r<q$.
\begin{conjecture}[Zagier's conjecture in positive characteristic] \label{conj: dimension}
Letting
\begin{align*}
d(w)=\begin{cases}
1 & \text{ if } w=0, \\
2^{w-1} & \text{ if } 1 \leq w \leq q-1, \\
2^{w-1}-1 & \text{ if } w=q,
\end{cases}
\end{align*}
we put $d(w)=\sum_{i=1}^q d(w-i)$ for $w>q$. Then for any $w \in \ensuremath \mathbb{N}$, we have
\[ \dim_K \mathcal Z_w = d(w). \]
\end{conjecture}
\begin{conjecture}[Hoffman's conjecture in positive characteristic] \label{conj: basis}
A $K$-basis for $\mathcal Z_w$ is given by $\mathcal T_w$ consisting of $\zeta_A(s_1,\ldots,s_r)$ of weight $w$ with $s_i \leq q$ for $1 \leq i <r$, and $s_r<q$.
\end{conjecture}
These conjectures have been completely solved by the works of \cite{ND21} and \cite{IKLNDP22}.
\subsection{The composition space} ${}$\par \label{sec: composition space}
In this section, we recall the notion of the composition space $\frak C$ as mentioned in the Introduction. Let $\Sigma = \{x_n\}_{n \in \mathbb{N}}$ be a countable set equipped with the weight $w(x_n)=n$. The elements of $\Sigma$ will be called \textit{letters}. A \textit{word} over $\Sigma$ is a finite string of letters. In particular, the empty word will be denoted by $1$. Let $\langle \Sigma \rangle$ denote the set of all words over $\Sigma$. We endow $\langle \Sigma \rangle$ with the \textit{concatenation product} defined by the following formula:
\begin{equation*}
x_{i_1} \dotsc x_{i_n} \cdot x_{j_1} \dotsc x_{j_m} = x_{i_1} \dotsc x_{i_n} x_{j_1} \dotsc x_{j_m}.
\end{equation*}
The composition space $\mathfrak{C}$ is the free $\mathbb F_q$-vector space $\mathbb F_q \langle \Sigma \rangle$ with basis $\langle \Sigma \rangle$. The concatenation product extends to $\mathfrak{C}$ by linearity.
\section{Shuffle algebra and shuffle map in positive characteristic} \label{sec: shuffle algebra 0}
This section aims to introduce the notion of the shuffle product of MZV's in positive characteristic. Then we define different products related to the associated shuffle algebra (see \S \ref{sec: shuffle algebra pos char}) and also the shuffle map (see Theorem \ref{thm: shuffle map}).
\subsection{Shuffle product for power sums} ${}$\par
For $d \in \mathbb Z$ we introduce
\begin{equation*}
S_d(\mathfrak{s}):=\sum \frac{1}{a_1^{s_1} \ldots a_r^{s_r}} \in K_\infty
\end{equation*}
where the sum runs through the set of tuples $(a_1,\ldots,a_r) \in A_+^r$ with $d=\deg a_1>\ldots>\deg a_r$. Further, we define
\begin{equation*}
S_{<d}(\mathfrak{s}):=\sum \frac{1}{a_1^{s_1} \ldots a_r^{s_r}} \in K_\infty
\end{equation*}
where the sum is over $(a_1,\ldots,a_r) \in A_+^r$ with $d>\deg a_1>\ldots>\deg a_r$. Thus
\begin{align*}
S_{<d}(\mathfrak{s}) =\sum_{i=0}^{d-1} S_i(\mathfrak{s}), \quad S_{d}(\mathfrak{s}) =S_d(s_1) S_{<d}(\mathfrak{s}_-)=S_d(s_1) S_{<d}(s_2,\dots,s_r).
\end{align*}
Here by convention we define empty sums to be $0$ and empty products to be $1$. In particular, $S_{<d}$ of the empty tuple is equal to $1$.
We briefly recall some results of Thakur concerning the shuffle product for power sums in \cite{Tha10} (see also \cite[\S 5.2]{Tha17}). Thakur first proved (see \cite[Theorems 1 and 2]{Tha10}) that for all $a,b \in \mathbb N$, there exist $\Delta^i_{a,b} \in \mathbb F_p$ for $0 < i <a+b$ such that for all $d \in \mathbb Z$,
\begin{equation} \label{eq:product depth1}
S_d(a)S_d(b)=S_d(a+b)+\sum_{0 < i <a+b} \Delta^i_{a,b} S_d(a+b-i,i).
\end{equation}
Shortly after, Chen \cite{Che15} gave explicit formulas for the coefficients $\Delta^i_{a,b}$ and proved
\begin{align*}
\Delta^i_{a,b}=\begin{cases} (-1)^{a-1} {i-1 \choose a-1}+(-1)^{b-1} {i-1 \choose b-1} & \quad \text{if } (q-1) \mid i \text{ and } 0<i<a+b, \\
0 & \quad \text{otherwise}.
\end{cases}
\end{align*}
Here we recall that for integers $a, b$ with $b \geq 0
$,
\begin{equation*}
\binom{a}{b} = \dfrac{a(a-1) \dotsc (a-b+1)}{b!}.
\end{equation*}
It should be remarked that $\binom{a}{b} = 0$ if $b > a \geq 0$.
\begin{proposition}[Chen] \label{prop: chen1}
Let $r,s$ be positive integers. For all $d \in \mathbb{Z}$, we have
\begin{equation*}
S_d(r)S_d(s) = S_d(r+s) + \sum_{\substack{i,j \in \ensuremath \mathbb{N}, \\ i+j=r+s}} \Delta^{j}_{r,s}S_d(i,j).
\end{equation*}
\end{proposition}
\begin{proposition} \label{prop: chen2}
Let $r,s$ be positive integers. For all $d \in \mathbb{Z}$, we have
\begin{equation*}
S_{<d}(r)S_{<d}(s) = S_{<d}(r+s) + S_{<d}(r,s) + S_{<d}(s,r) + \sum_{\substack{i,j \in \ensuremath \mathbb{N}, \\ i+j=r+s}} \Delta^{j}_{r,s}S_{<d}(i,j).
\end{equation*}
\end{proposition}
\begin{proof}
Using Proposition \ref{prop: chen1}, we have
\begin{align*}
S_{<d}(r)S_{<d}(s) &= \sum \limits_{k<d} S_k(r)S_k(s) + \sum \limits_{k<d} S_k(r)S_{<k}(s) + \sum \limits_{k<d} S_k(s)S_{<k}(r)\\
&= \sum \limits_{k<d} \Bigg(S_k(r+s) + \sum_{\substack{i,j \in \ensuremath \mathbb{N}, \\ i+j=r+s}} \Delta^{j}_{r,s}S_k(i,j)\Bigg) + \sum \limits_{k<d} S_k(r,s) + \sum \limits_{k<d} S_k(s,r)\\
&= S_{<d}(r+s) + \sum_{\substack{i,j \in \ensuremath \mathbb{N}, \\ i+j=r+s}} \Delta^{j}_{r,s}S_{<d}(i,j) + S_{<d}(r,s) + S_{<d}(s,r).
\end{align*}
This proves the proposition.
\end{proof}
\subsection{The shuffle algebra in positive characteristic} ${}$\par \label{sec: shuffle algebra pos char}
Recall that the composition space is defined as in \S \ref{section: Composition space}. We define the unit $u:\mathbb F_q \to \frak C$ by sending $1$ to the empty word $1$. Next we define recursively two products on $\mathfrak{C}$ as $\mathbb F_q$-bilinear maps
\begin{align*}
\diamond \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C} \quad \text{and} \quad
\shuffle \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C}
\end{align*}
by setting $1 \diamond \mathfrak{a} = \mathfrak{a} \diamond 1 = \mathfrak{a}, 1 \shuffle \mathfrak{a} = \mathfrak{a} \shuffle 1 = \mathfrak{a}$ and
\begin{align*}
\mathfrak{a} \diamond \fb &= \ x_{a + b}(\mathfrak{a}_- \shuffle \fb_-) + \sum\limits_{i+j = a + b} \Delta^j_{a,b} x_i(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)),\\
\mathfrak{a} \shuffle \fb &= \ x_{a}(\mathfrak{a}_- \shuffle \fb) + x_{b}(\mathfrak{a} \shuffle \fb_-) + \mathfrak{a} \diamond \fb
\end{align*}
for any words $\mathfrak{a},\fb \in \langle \Sigma \rangle$. We call $\diamond$ the \textit{diamond product} and $\shuffle $ the \textit{shuffle product}.
\begin{proposition} \label{prop: commutative}
The diamond product and the shuffle product are commutative.
\end{proposition}
\begin{proof}
Let $\mathfrak{a}, \fb \in \langle \Sigma \rangle$ be two arbitrary words. It is suffices to show that
\begin{equation} \label{eq: commutative}
\mathfrak{a} \diamond \fb = \fb \diamond \mathfrak{a} \quad \text{and} \quad \mathfrak{a} \shuffle \fb = \fb \shuffle \mathfrak{a} .
\end{equation}
We proceed the proof by induction on $\depth(\mathfrak{a}) + \depth(\fb)$. If one of $\mathfrak{a}$ or $\fb$ is empty word, then \eqref{eq: commutative} holds trivially. We assume that \eqref{eq: commutative} holds when $\depth(\mathfrak{a}) + \depth(\fb) < n$ with $n \in \ensuremath \mathbb{N}$ and $n \geq 2$. We need to show that \eqref{eq: commutative} holds when $\depth(\mathfrak{a}) + \depth(\fb) = n$.
\par Indeed, we have
\begin{align*}
\mathfrak{a} \diamond \fb &= \ x_{a + b}(\mathfrak{a}_-\shuffle \fb_-) + \sum\limits_{i+j = a + b} \Delta^j_{a,b} x_i(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)),\\
\fb \diamond \mathfrak{a} &= \ x_{b + a}(\fb_- \shuffle \mathfrak{a}_-) + \sum\limits_{i+j = b + a} \Delta^j_{b,a} x_i(x_j \shuffle (\fb_- \shuffle \mathfrak{a}_-)).
\end{align*}
It follows from the induction hypothesis that $\mathfrak{a}_-\shuffle \fb_- = \fb_- \shuffle \mathfrak{a}_-$, hence $\mathfrak{a} \diamond \fb = \fb \diamond \mathfrak{a}$. On the other hand, we have
\begin{align*}
\mathfrak{a} \shuffle \fb &= \ x_{a}(\mathfrak{a}_- \shuffle \fb) + x_{b}(\mathfrak{a} \shuffle \fb_-) + \mathfrak{a} \diamond \fb,\\
\fb \shuffle \mathfrak{a} &= \ x_{b}(\fb_- \shuffle \mathfrak{a}) + x_{a}(\fb \shuffle \mathfrak{a}_-) + \fb \diamond \mathfrak{a}.
\end{align*}
It follows from the induction hypothesis and the above arguments that $\mathfrak{a}_- \shuffle \fb = \fb \shuffle \mathfrak{a}_-, \mathfrak{a} \shuffle \fb_- = \fb_- \shuffle \mathfrak{a}$ and $\mathfrak{a} \diamond \fb = \fb \diamond \mathfrak{a}$, hence $\mathfrak{a} \shuffle \fb = \fb \shuffle \mathfrak{a}$. This proves the proposition.
\end{proof}
We next define recursively a product on $\mathfrak{C}$ as a $\mathbb F_q$-bilinear map
\begin{align*}
\triangleright \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C}
\end{align*}
by setting $1 \triangleright \mathfrak{a} = \mathfrak{a} \triangleright 1 = \mathfrak{a}$ and
\begin{align*}
\mathfrak{a} \triangleright \fb &= \ x_{a}(\mathfrak{a}_- \shuffle \fb)
\end{align*}
for any words $\mathfrak{a},\fb \in \langle \Sigma \rangle$. We call $\triangleright$ the triangle product. We stress that the triangle product is neither commutative nor associative, as one verifies at once.
\begin{lemma} \label{lem: triangle formulas}
For all words $\mathfrak{a},\fb \in \langle \Sigma \rangle$, we have
\begin{enumerate}
\item $\mathfrak{a} \diamond \fb = (x_a \diamond x_b) \triangleright (\mathfrak{a}_- \shuffle \fb_-)$
\item $\mathfrak{a} \shuffle \fb = \mathfrak{a} \triangleright \fb + \fb \triangleright \mathfrak{a} + \mathfrak{a} \diamond \fb$.
\end{enumerate}
\end{lemma}
\begin{proof}
We have
\begin{align*}
\mathfrak{a} \diamond \fb
&= \ x_{a+b}(\mathfrak{a}_- \shuffle \fb_-) + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_i (x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-))\\
&= \ x_{a+b} \triangleright (\mathfrak{a}_- \shuffle \fb_-) + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_ix_j \triangleright (\mathfrak{a}_- \shuffle \fb_-)\\
&= \ (x_{a+b} + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_ix_j) \triangleright (\mathfrak{a}_- \shuffle \fb_-)\\
&= \ (x_a \diamond x_b) \triangleright (\mathfrak{a}_- \shuffle \fb_-).
\end{align*}
This proves part $(1)$. Part $(2)$ is straightforward from the commutativity of the shuffle product. We finish the proof.
\end{proof}
\subsection{The shuffle map in positive characteristic} ${}$\par
For all $d \in \mathbb Z$, we define two $\mathbb{F}_q$-linear maps
\begin{equation*}
S_{<d} \colon \frak C \rightarrow K_{\infty} \quad \text{and} \quad \zeta_A \colon \frak C \rightarrow K_{\infty},
\end{equation*}
which map the empty word to the element $1 \in K_{\infty}$, and map any word $x_{s_1} \dotsc x_{s_r}$ to $S_{<d}(s_1, \dotsc, s_r)$ and $\zeta_A(s_1, \dotsc, s_r)$, respectively. The main result of this section reads as follows:
\begin{theorem} \label{thm: shuffle map}
For all words $\mathfrak{a}, \fb \in \frak C$ and for all $d \in \mathbb Z$ we have
\begin{align*}
S_{<d}(\mathfrak{a}\shuffle \fb) &=S_{<d}(\mathfrak{a}) \, S_{<d}(\fb), \\
\zeta_A(\mathfrak{a}\shuffle \fb) &=\zeta_A(\mathfrak{a}) \, \zeta_A(\fb).
\end{align*}
\end{theorem}
\begin{proof}
See \cite[Theorem 3.1.4]{Shi18}.
\end{proof}
We denote by $\mathcal{Z}_w$ (resp. $\mathcal{Z}$) the $K$-vector space spanned by MZV's of weight $w$ (resp. by MZV's). Then the $K$-linear map \[ Z_\shuffle:\frak C \otimes_{\mathbb F_q} K \to \mathcal Z, \] which sends a word $\mathfrak{a} \in \frak C$ to $\zeta_A(\mathfrak{a})$, is a homomorphism of $K$-algebras, and is called the shuffle map in positive characteristic.
\section{Algebra structure of the shuffle algebra} \label{sec: algebra structure}
The main goal of this section is to prove that the composition space $\frak C$ equipped with the shuffle product $\shuffle $ given by Thakur is an algebra (see Theorem \ref{theorem: algebras}). The key point is to show the associativity of the shuffle product which consequently solves \cite[Conjecture 3.2.2]{Shi18}. In fact, the associativity property of $(\frak C,\shuffle )$ turned out to be very hard to prove (see \cite[Remark 2.2, Part 1]{ND21}) as pointed out one of the referees of \cite{ND21}. Our method is of algebraic nature. It consists of unpacking the nature of the coefficients $\Delta^i_{a,b}$ appearing in the shuffle product of Thakur by using partial fractional decompositions.
We mention that the associativity could follow from a transcendental approach. As mentioned in \cite[\S 3.2.1]{Shi18}, it follows from a conjecture of Thakur about the $\mathbb F_q$-linear independence of MZV's (see \cite[3.2.3]{Shi18}).
\subsection{Expansions for $S_d$ of depth one} ${}$\par
\subsubsection{} Let $r,s, t$ be positive integers. We first expand $(S_d(s) S_d(t))S_d(r)$. We have
\begin{align*}
&(S_d(r)S_d(s))S_d(t)\\
&= \left(S_d(r+s) + \sum_{i+j=r+s} \Delta^j_{r,s} S_d(i,j)\right)S_d(t)\\
&= S_d(r+s)S_d(t) + \sum_{i+j=r+s} \Delta^j_{r,s} S_d(i,j) S_d(t)\\
&= \left(S_d(r+s+t) + \sum_{i+j=r+s+t} \Delta^j_{r+s,t} S_d(i,j) \right) \\
&+ \sum_{i+j=r+s} \Delta^j_{r,s} \left(S_d(i+t,j)+\sum_{i_1+j_1=i+t} \Delta^{j_1}_{i,t} S_d(i_1) S_{<d}(j_1) S_{<d}(j)\right).
\end{align*}
The sum of the terms of depth $2$ in the above expansion is the following:
\begin{align} \label{eq: 1}
\sum_{i+j=r+s+t} \Delta^j_{r+s,t} S_d(i,j)+\sum_{i+j=r+s} \Delta^j_{r,s} \left(S_d(i+t,j)+\sum_{i_1+j_1=i+t} \Delta^{j_1}_{i,t} S_d(i_1,j_1+j)\right)
\end{align}
The sum of the terms of depth $3$ in the above expansion is the following:
\begin{align*}
\sum_{i+j=r+s} \Delta^j_{r,s} \sum_{i_1+j_1=i+t} \Delta^{j_1}_{i,t} \left(S_d(i_1,j_1,j)+S_d(i_1,j,j_1) + \sum \limits_{i_2 + j_2 = j+j_1 } \Delta_{j,j_1}^{j_2} S_d(i_1,i_2,j_2)\right).
\end{align*}
\subsubsection{} We next expand $S_d(r) (S_d(s)S_d(t))$.
\begin{align*}
&S_d(r)(S_d(s)S_d(t))\\
&= S_d(r)\left(S_d(s+t) + \sum_{i+j=s+t} \Delta^j_{s,t} S_d(i,j)\right) \\
&= S_d(r)S_d(s+t) + \sum_{i+j=s+t} \Delta^j_{s,t} S_d(r)S_d(i,j)\\
&= \left(S_d(r+s+t) + \sum_{i+j=r+s+t} \Delta^j_{r,s+t} S_d(i,j) \right) \\
&+ \sum_{i+j=s+t} \Delta^j_{s,t} \left(S_d(i+r,j)+\sum_{i_1+j_1=i+r} \Delta^{j_1}_{i,r} S_d(i_1) S_{<d}(j_1) S_{<d}(j)\right).
\end{align*}
The sum of the terms of depth $2$ in the above expansion is the following:
\begin{align} \label{eq: 2}
\sum_{i+j=r+s+t} \Delta^j_{r,s+t} S_d(i,j)+\sum_{i+j=s+t} \Delta^j_{s,t} \left(S_d(i+r,j)+\sum_{i_1+j_1=i+r} \Delta^{j_1}_{i,r} S_d(i_1,j_1+j)\right).
\end{align}
The sum of the terms of depth $3$ in the above expansion is the following:
\begin{align*}
\sum_{i+j=s+t} \Delta^j_{s,t} \sum_{i_1+j_1=i+r} \Delta^{j_1}_{i,r} \left(S_d(i_1,j_1,j)+S_d(i_1,j,j_1) + \sum \limits_{i_2 + j_2 = j+j_1 } \Delta_{j,j_1}^{j_2} S_d(i_1,i_2,j_2)\right).
\end{align*}
\subsubsection{Partial fraction decompositions} ${}$\par
In this section, we will give some partial fraction decompositions which will be used in the next section. To simplify the notations, for $j, r \in \mathbb{N}$, we set
\begin{align*}
\nabla^j_r :&= (-1)^{r} {j-1 \choose r-1}.
\end{align*}
\begin{lemma} \label{lem: basic indentity}
Let $r , s$ be positive integers. The following equality of rational function holds:
\begin{equation*}
\frac{1}{X^r Y^s}=\sum_{\substack{i,j \in \ensuremath \mathbb{N}, \\ i+j=r+s}} \left[\frac{{j-1 \choose s-1}}{(X+Y)^j X^i}+\frac{{j-1 \choose r-1}}{(X+Y)^j Y^i}\right].
\end{equation*}
\end{lemma}
\begin{proof}
See \cite[Lemma 1.49]{BGF}.
\end{proof}
\begin{lemma} \label{lem: decomposition}
Let $R$ be a ring, and let $a, b$ be elements in $R$ such that $a \ne b$. Then for all positive integers $r , s$, we have the following partial fraction decomposition:
\begin{equation*}
\frac{1}{(X+a)^r (X+b)^s}=\sum_{\substack{i,j \in \ensuremath \mathbb{N}, \\ i+j=r+s}} \left[\frac{\nabla^j_s}{(a-b)^j (X+a)^i}+\frac{\nabla^j_r}{(b-a)^j (X+b)^i}\right].
\end{equation*}
\end{lemma}
\begin{proof}
The result follows immediately from Lemma \ref{lem: basic indentity} by taking $X = X + a$ and $Y = - (X + b)$.
\end{proof}
It should be remarked that Chen's formula is based on the following identity.
\begin{corollary} \label{cor: identity}
Let $a, b$ be elements in $A$ such that $a \ne b$. Then for all positive integers $r , s$, we have
\begin{equation*}
\frac{1}{a^r b^s}=\sum_{\substack{i,j \in \ensuremath \mathbb{N}, \\ i+j=r+s}} \left[\frac{\nabla^j_s}{(a-b)^j a^i}+\frac{\nabla^j_r}{(b-a)^j b^i}\right].
\end{equation*}
\end{corollary}
\begin{proof}
The result is straightforward from Lemma \ref{lem: decomposition} when $X = 0$ and $R = A$.
\end{proof}
\par Let $r,s,t$ be positive integers. Let $R$ be a ring, and let $u, v$ be elements in $R$ such that $u \ne 0, v \ne 0 $ and $u \ne v$. Consider the following fractional function
\begin{equation*}
P(A) = \frac{1}{A^{r}(A+u)^s(A+v)^t}
\end{equation*}
Set $B= A + u$ and $C = A+v$. Using Lemma \ref{lem: decomposition}, we give the partial fraction decomposition of $P(A)$ in two ways. First, we expand $P(A)$ from the left to the right as follows:
\begin{align} \label{eq: PFD L-R}
\frac{1}{A^{r}B^s} \cdot \frac{1}{C^t} &=\sum \limits_{i+j = r+s} \left(\frac{\nabla^j_s}{(A-B)^jA^i} + \frac{\nabla^j_r}{(B-A)^jB^i}\right)\frac{1}{C^t}\\ \notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(A-B)^j} \frac{\nabla^{j_1}_t}{(A-C)^{j_1}}\frac{1}{A^{i_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(A-B)^j} \frac{\nabla^{j_1}_i}{(C-A)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(B-A)^j} \frac{\nabla^{j_1}_t}{(B-C)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(B-A)^j} \frac{\nabla^{j_1}_i}{(C-B)^{j_1}}\frac{1}{C^{i_1}}\notag.
\end{align}
Next we expand $P(A)$ from the right to the left as follows:
\begin{align} \label{eq: PFD R-L}
\frac{1}{A^r} \cdot \frac{1}{B^{s}C^t} &= \frac{1}{A^r} \sum \limits_{i+j = s + t} \left(\frac{\nabla^j_t}{(B-C)^jB^i} + \frac{\nabla^j_s}{(C-B)^jC^i}\right)\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(B-C)^j}\frac{\nabla^{j_1}_i}{(A-B)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(B-C)^j}\frac{\nabla^{j_1}_r}{(B-A)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(C-B)^j}\frac{\nabla^{j_1}_i}{(A-C)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(C-B)^j}\frac{\nabla^{j_1}_r}{(C-A)^{j_1}}\frac{1}{C^{i_1}}.
\end{align}
\subsection{Associativity for $S_d$ of depth one}
\subsubsection{Main results} ${}$\par
To simplify the notations, for $j,k \in \mathbb{N}$, we set
\begin{align*}
\delta_j :&= \sum \limits_{\lambda \in \mathbb F_q^\times} \frac{1}{\lambda^j} = \begin{cases} -1 & \quad \text{if } (q - 1) \mid j , \\
0 & \quad \text{otherwise,}
\end{cases}\\
\delta_{j,k} :&= \sum \limits_{\lambda,\mu \in \mathbb F_q^\times; \lambda \ne \mu} \frac{1}{\lambda^j \mu ^k}.
\end{align*}
The following formulas will be used frequently later:
\begin{align}
\Delta_{r,s}^j &= \delta_j(\nabla^j_r + \nabla^j_s)\label{eq: formula 1},\\
\delta_j &= \delta_j(-1)^j\label{eq: formula 2},\\
\delta_j\delta_k &= \delta_{j+k} + \delta_{j,k}.\label{eq: formula 3}
\end{align}
Consider all the cases of tuples $(a,b,c) \in A_+(d) \times A_+(d) \times A_+(d)$, we set
\begin{align*}
M_0 &= \{(a,b,c) \in A_+(d) \times A_+(d) \times A_+(d) : a = b = c \},\\
M_1 &= \{(a,b,c) \in A_+(d) \times A_+(d) \times A_+(d) : a = b \ne c \},\\
M_2 &= \{(a,b,c) \in A_+(d) \times A_+(d) \times A_+(d) : a \ne b = c \},\\
M_3 &= \{(a,b,c) \in A_+(d) \times A_+(d) \times A_+(d) : c = a \ne b \},\\
M_4 &= \{(a,b,c) \in A_+(d) \times A_+(d) \times A_+(d) : a \ne b, b \ne c, c \ne a \}.
\end{align*}
The last set $M_4$ can be decomposed by the following partition:
\begin{equation*}
M_4 = N_0 \sqcup N_1 \sqcup N_2 \sqcup N_3 \sqcup N_4.
\end{equation*}
Here
\begin{itemize}
\item the set $N_0$ consists of tuples $(a,b,c) \in M_4$ where $b = a + \lambda f, c = a + \mu f$ with $\lambda, \mu \in \mathbb F_q^\times$ such that $\lambda \ne \mu$ and $f \in A_+$ such that $d > \deg f$;
\item the set $N_1$ consists of tuples $(a,b,c) \in M_4$ where $b = a + \lambda f, c = a + \mu u$ with $\lambda, \mu \in \mathbb F_q^\times$ and $f,u \in A_+$ such that $d > \deg f > \deg u$;
\item the set $N_2$ consists of tuples $(a,b,c) \in M_4$ where $b = a + \lambda u, c = a + \mu f$ with $\lambda, \mu \in \mathbb F_q^\times$ and $f,u \in A_+$ such that $d > \deg f > \deg u$;
\item the set $N_3$ consists of tuples $(a,b,c) \in M_4$ where $b = a + \lambda f, c = a + \lambda f + \mu u$ with $\lambda, \mu \in \mathbb F_q^\times$ and $f,u \in A_+$ such that $d > \deg f > \deg u$;
\item the set $N_4$ consists of tuples $(a,b,c) \in M_4$ where $b = a + \lambda f, c = a + \mu f + \eta u$ with $\lambda, \mu, \eta \in \mathbb F_q^\times$ such that $\lambda \ne \mu$ and $f,u \in A_+$ such that $d > \deg f > \deg u$.
\end{itemize}
\begin{proposition} \label{prop: associativity}
Let $r,s , t$ be positive integers. The expansions using Corollary~\ref{cor: identity} of
\begin{equation*}
\sum \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums, where $(a,b,c)$ ranges over all tuples in the sets $M_0,M_1,M_2,M_3,N_0,N_1,N_2,N_3,N_4$. Moreover,
\begin{enumerate}[$(1)$]
\item the expansion of $\sum \frac{1}{a^rb^s} \cdot \frac{1}{c^t}$ (respectively, $\sum \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}$), where $(a,b,c)$ ranges over all tuples in $M_0$, yields the term of depth $1$ in the expression of $(S_d(r)S_d(s))S_d(t)$ (respectively, $S_d(r)(S_d(s)S_d(t))$);
\item the expansion of $\sum \frac{1}{a^rb^s} \cdot \frac{1}{c^t}$ (respectively, $\sum \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}$), where $(a,b,c)$ ranges over all tuples in $M_1 \sqcup M_2\sqcup M_3\sqcup N_0$, yields the terms of depth $2$ in the expression of $(S_d(r)S_d(s))S_d(t)$ (respectively, $S_d(r)(S_d(s)S_d(t))$);
\item the expansion of $\sum \frac{1}{a^rb^s} \cdot \frac{1}{c^t}$ (respectively, $\sum \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}$), where $(a,b,c)$ ranges over all tuples in $N_1 \sqcup N_2\sqcup N_3\sqcup N_4$, yields the terms of depth $3$ in the expression of $(S_d(r)S_d(s))S_d(t)$ (respectively, $S_d(r)(S_d(s)S_d(t))$).
\end{enumerate}
\end{proposition}
As a direct consequence, we obtain the following theorem.
\begin{theorem} \label{thm: assoc S_d depth 1}
Let $r,s , t$ be positive integers. For all $d \in \mathbb{N}$, the expansions using Chen's formula of $(S_d(s) S_d(t))S_d(r)$ and $S_d(s) (S_d(t)S_d(r))$ yield the same expression in terms of power sums.
\end{theorem}
The rest of this section is devoted to a proof of Proposition \ref{prop: associativity}.
\subsubsection{Depth 1 terms: proof of Proposition \ref{prop: associativity}, Part 1} ${}$\par
It is obvious that
\begin{equation*}
\sum \limits_{(a,b,c) \in M_0} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} = \sum \limits_{(a,b,c) \in M_0} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t} = S_d(r+s+t),
\end{equation*}
which yields the term of depth $1$ in the expression of $(S_d(r)S_d(s))S_d(t)$ and $S_d(r)(S_d(s)S_d(t))$.
\subsubsection{Depth 2 terms: proof of Proposition \ref{prop: associativity}, Part 2} ${}$\par
Consider the following cases:
\medskip
\noindent \textbf{Case 1:} $(a,b,c)$ ranges over all tuples in $M_1$.
\par For each $\lambda \in \mathbb F_q^\times$, consider the rational function
\begin{equation*}
P_{\lambda}(A,F) = \frac{1}{A^{r+s}(A+\lambda F)^t}.
\end{equation*}
We will deduce the partial fraction decomposition of $ P_{\lambda}(A,F) $ by the following process: we first give the partial fraction decomposition of $ P_{\lambda}(A,F) $ in variable $A$ with coefficients are rational functions of the form $Q(F) \in \mathbb F_q(F)$; then we continue to give the partial fraction decomposition of $Q(F)$ in variable $F$ whose coefficients are elements in $\mathbb F_q$.
\par Set $C= A + \lambda F$. Using Lemma \ref{lem: decomposition}, we proceed the process in two ways. First, we expand $P_{\lambda}(A,F)$ from the left to the right as follows:
\begin{align} \label{eq: L-R M1}
\frac{1}{A^{r+s}} \cdot \frac{1}{C^t} &= \sum \limits_{i+j = r + s+ t} \left(\frac{\nabla^j_t}{(A-C)^jA^i} + \frac{\nabla^j_{r+s}}{(C-A)^jC^i} \right)\\ \notag
&=\sum \limits_{i+j = r + s+ t} \left(\frac{\nabla^j_t}{(-\lambda F)^jA^i} + \frac{\nabla^j_{r+s}}{(\lambda F)^jC^i} \right)\\ \notag
&= \sum \limits_{i+j = r + s+ t} \frac{\nabla^j_t}{(-\lambda)^j}\frac{1}{ A^iF^j} + \sum \limits_{i+j = r + s+ t}\frac{\nabla^j_{r+s}}{\lambda^j}\frac{1}{C^iF^j}. \notag
\end{align}
Next we expand $P_{\lambda}(A,F)$ from the right to the left as follows:
\begin{align} \label{eq: R-L M1}
\frac{1}{A^r} \cdot\frac{1}{A^{s}C^t} &= \frac{1}{A^r} \sum \limits_{i+j = s+ t} \left(\frac{\nabla^j_t}{(A-C)^jA^{i}} + \frac{\nabla^j_{s}}{(C-A)^jC^i} \right)\\ \notag
&= \sum \limits_{i+j = s+ t} \frac{\nabla^j_t}{(A-C)^j}\frac{1}{A^{i+r}}\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_{s}}{(C-A)^j}\frac{\nabla^{j_1}_{i}}{(A-C)^{j_1}} \frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_{s}}{(C-A)^j}\frac{\nabla^{j_1}_{r}}{(C-A)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&= \sum \limits_{i+j = s+ t} \frac{\nabla^j_t}{(-\lambda F)^j}\frac{1}{A^{i+r}}\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_{s}}{(\lambda F)^j}\frac{\nabla^{j_1}_{i}}{(-\lambda F)^{j_1}} \frac{1}{A^{i_1}}\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_{s}}{(\lambda F)^j}\frac{\nabla^{j_1}_{r}}{(\lambda F)^{j_1}}\frac{1}{C^{i_1}}\\ \notag
&= \sum \limits_{i+j = s+ t} \frac{\nabla^j_t}{(-\lambda)^j}\frac{1}{A^{i+r}F^j}\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r}
\frac{\nabla^j_{s} (-1)^{j_1}\nabla^{j_1}_{i}}{\lambda^{j+j_1}}\frac{1}{A^{i_1}F^{j+j_1}}\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_{s} \nabla^{j_1}_{r}}{\lambda^{j+j_1}}\frac{1}{C^{i_1}F^{j+j_1}}. \notag
\end{align}
For each $(a,b,c)\in M_1$, there exist $\lambda \in \mathbb F_q^\times$ and $f \in A_+$ with $d > \deg f$ such that $c = a + \lambda f$. Replacing $A = a, F = f$, one deduces from \eqref{eq: L-R M1} that
\begin{align*}
\sum \limits_{(a,b,c) \in M_1}& \frac{1}{a^rb^s} \cdot \frac{1}{c^t}\\
&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda \in \mathbb F_q^\times} \frac{1}{a^{r+s}} \cdot \frac{1}{(a+\lambda f)^t}\\
&= \sum \limits_{i+j = r + s+ t} \delta_j\nabla^j_t \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{ a^if^j} + \sum \limits_{i+j = r + s+ t}\delta_j\nabla^j_{r+s}\sum\limits_{\substack{c,f \in A_+ \\ d = \deg c > \deg f}}\frac{1}{c^if^j}\\
&= \sum \limits_{i+j = r + s+ t} \delta_j\nabla^j_t S_d(i,j) + \sum \limits_{i+j = r + s+ t}\delta_j\nabla^j_{r+s}S_d(i,j)\\
&= \sum \limits_{i+j = r + s+ t}\Delta^j_{r+s,t}S_d(i,j).
\end{align*}
Similarly, one deduces from \eqref{eq: R-L M1} that
\begin{align*}
\sum \limits_{(a,b,c) \in M_1} &\frac{1}{a^r} \cdot \frac{1}{b^sc^t}
\\&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda \in \mathbb F_q^\times} \frac{1}{a^r} \cdot \frac{1}{a^s(a+\lambda f)^t}\\
&= \sum \limits_{i+j = s+ t} \delta_j\nabla^j_t \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i+r}f^j}\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r}\delta_{j+j_1}\nabla^j_{s} (-1)^{j_1}\nabla^{j_1}_{i} \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i_1}f^{j+j_1}}\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r} \delta_{j+j_1}\nabla^j_{s} \nabla^{j_1}_{r}\sum\limits_{\substack{c,f \in A_+ \\ d = \deg c > \deg f}}\frac{1}{c^{i_1}f^{j+j_1}}\\
&= \sum \limits_{i+j = s+ t} \delta_j\nabla^j_t S_d(i+r,j)\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r}\delta_{j+j_1}\nabla^j_{s} (-1)^{j_1}\nabla^{j_1}_{i} S_d(i_1,j+j_1)\\ \notag
&+ \sum \limits_{i+j = s+ t} \sum \limits_{i_1+j_1 = i+r} \delta_{j+j_1}\nabla^j_{s} \nabla^{j_1}_{r}S_d(i_1,j+j_1).
\end{align*}
Since the partial fraction decomposition of $P_{\lambda}(A,F)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in M_1} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in M_1} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.
\medskip
\noindent \textbf{Case 2:} $(a,b,c)$ ranges over all tuples in $M_2$.
\par For each $\lambda \in \mathbb F_q^\times$, consider the rational function
\begin{equation*}
P_{\lambda}(A,F) = \frac{1}{A^{r}(A+\lambda F)^{s+t}}.
\end{equation*}
Set $B = A + \lambda F$. From the same process as in the the case of $M_1$, we expand $P_{\lambda}(A,F)$ in two ways. First, we expand $P_{\lambda}(A,F)$ from the left to the right as follows:
\begin{align} \label{eq: L-R M2}
\frac{1}{A^rB^s} \cdot \frac{1}{B^t} &=
\sum \limits_{i+j = r + s} \left(\frac{\nabla^j_s}{(A-B)^jA^i} + \frac{\nabla^j_r}{(B-A)^jB^i}\right) \frac{1}{B^t}\\ \notag
&= \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \frac{\nabla^j_s}{(A-B)^j}
\frac{\nabla^{j_1}_t}{(A-B)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \frac{\nabla^j_s}{(A-B)^j}
\frac{\nabla^{j_1}_i}{(B-A)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = r + s} \frac{\nabla^j_r}{(B-A)^j}\frac{1}{B^{i+t}}\\\notag
&= \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \frac{\nabla^j_s}{(-\lambda F)^j}
\frac{\nabla^{j_1}_t}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \frac{\nabla^j_s}{(-\lambda F)^j}
\frac{\nabla^{j_1}_i}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = r + s} \frac{\nabla^j_r}{(\lambda F)^j}\frac{1}{B^{i+t}}\\\notag
&= \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \frac{\nabla^j_s
\nabla^{j_1}_t}{(-\lambda)^{j+j_1}}\frac{1}{A^{i_1}F^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \frac{\nabla^j_s
(-1)^{j_1}\nabla^{j_1}_i}{(-\lambda)^{j+j_1}}\frac{1}{B^{i_1}F^{j+j_1}}\\\notag
&+\sum \limits_{i+j = r + s} \frac{\nabla^j_r}{\lambda^j}\frac{1}{B^{i+t}F^j}.\notag
\end{align}
Next we expand $P_{\lambda}(A,F)$ from the right to the left as follows:
\begin{align}\label{eq: R-L M2}
\frac{1}{A^r} \cdot \frac{1}{B^{s+t}} &= \sum \limits_{i+j = r + s + t} \left(\frac{\nabla^j_{s+t}}{(A-B)^jA^i} + \frac{\nabla^j_{r}}{(B-A)^jB^i}\right)\\\notag
&=\sum \limits_{i+j = r + s + t} \left(\frac{\nabla^j_{s+t}}{(-\lambda F)^jA^i} + \frac{\nabla^j_{r}}{(\lambda F)^jB^i}\right)\\\notag
&=\sum \limits_{i+j = r + s + t} \frac{\nabla^j_{s+t}}{(-\lambda)^j} \frac{1}{A^iF^j} + \sum \limits_{i+j = r + s + t} \frac{\nabla^j_{r}}{\lambda^j}\frac{1}{B^iF^j}.\notag
\end{align}
For each $(a,b,c)\in M_2$, there exist $\lambda \in \mathbb F_q^\times$ and $f \in A_+$ with $d > \deg f$ such that $b = a + \lambda f$. Replacing $A = a, F = f$, one deduces from \eqref{eq: L-R M2} that
\begin{align*}
\sum \limits_{(a,b,c) \in M_2} & \frac{1}{a^rb^s} \cdot \frac{1}{c^t}\\
&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda \in \mathbb F_q^\times} \frac{1}{a^{r}(a+\lambda f)^s} \cdot \frac{1}{(a+\lambda f)^t}\\
&= \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \delta_{j+j_1}\nabla^j_s
\nabla^{j_1}_t
\sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i_1}f^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \delta_{j+j_1}\nabla^j_s
(-1)^{j_1}\nabla^{j_1}_i\sum\limits_{\substack{b,f \in A_+ \\ d = \deg b > \deg f}}\frac{1}{b^{i_1}f^{j+j_1}}\\\notag
&+\sum \limits_{i+j = r + s} \delta_j\nabla^j_r\sum\limits_{\substack{b,f \in A_+ \\ d = \deg b > \deg f}}\frac{1}{b^{i+t}f^j}\\\notag
&= \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \delta_{j+j_1}\nabla^j_s
\nabla^{j_1}_t
S_d(i_1,j+j_1)\\\notag
&+ \sum \limits_{i+j = r + s}\sum \limits_{i_1+j_1 = i + t} \delta_{j+j_1}\nabla^j_s
(-1)^{j_1}\nabla^{j_1}_i S_d(i_1,j+j_1)\\\notag
&+\sum \limits_{i+j = r + s} \delta_j\nabla^j_r S_d(i+t,j).
\end{align*}
Similarly, one deduces from \eqref{eq: R-L M2} that
\begin{align*}
\sum \limits_{(a,b,c) \in M_2} &\frac{1}{a^r} \cdot \frac{1}{b^sc^t}\\
&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda \in \mathbb F_q^\times} \frac{1}{a^r} \cdot \frac{1}{(a+\lambda f)^{s+t}}\\
&=\sum \limits_{i+j = r + s + t} \delta_j\nabla^j_{s+t} \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^if^j} + \sum \limits_{i+j = r + s + t} \delta_j\nabla^j_{r}\sum\limits_{\substack{b,f \in A_+ \\ d = \deg b > \deg f}}\frac{1}{b^if^j}\\
&=\sum \limits_{i+j = r + s + t} \delta_j\nabla^j_{s+t} S_d(i,j) + \sum \limits_{i+j = r + s + t} \delta_j\nabla^j_{r}S_d(i,j)\\
&=\sum \limits_{i+j = r + s + t} \Delta^j_{r,s+t} S_d(i,j).\notag
\end{align*}
Since the partial fraction decomposition of $P_{\lambda}(A,F)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in M_2} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in M_2} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.
\medskip
\noindent \textbf{Case 3:} $(a,b,c)$ ranges over all tuples in $M_3$.
\par For each $\lambda \in \mathbb F_q^\times$, consider the rational function
\begin{equation*}
P_{\lambda}(A,F) = \frac{1}{A^{r}(A+\lambda F)^{s}A^t}.
\end{equation*}
Set $B = A + \lambda F$. From the same process as in the the case of $M_1$, we expand $P_{\lambda}(A,F)$ in two ways. First, we expand $P_{\lambda}(A,F)$ from the left to the right as follows:
\begin{align} \label{eq: L-R M3}
\frac{1}{A^rB^s} \cdot \frac{1}{A^t} &= \sum \limits_{i+j = r+s} \left(\frac{\nabla^j_s}{(A-B)^jA^i} + \frac{\nabla^j_r}{(B-A)^jB^i}\right) \frac{1}{A^t}\\\notag
&= \sum \limits_{i+j = r+s}\frac{\nabla^j_s}{(A-B)^j}\frac{1}{A^{i+t}}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\frac{\nabla^j_r}{(B-A)^j}\frac{\nabla^{j_1}_t}{(B-A)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\frac{\nabla^j_r}{(B-A)^j}\frac{\nabla^{j_1}_i}{(A-B)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&= \sum \limits_{i+j = r+s}\frac{\nabla^j_s}{(-\lambda F)^j}\frac{1}{A^{i+t}}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\frac{\nabla^j_r}{(\lambda F)^j}\frac{\nabla^{j_1}_t}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\frac{\nabla^j_r}{(\lambda F)^j}\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&= \sum \limits_{i+j = r+s}\frac{\nabla^j_s}{(-\lambda)^j}\frac{1}{A^{i+t}F^j}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^{j+j_1}}\frac{1}{B^{i_1}F^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\frac{\nabla^j_r(-1)^{j_1}\nabla^{j_1}_i}{\lambda^{j+j_1}}\frac{1}{A^{i_1}F^{j+j_1}}.\notag
\end{align}
Next we expand $P_{\lambda}(A,F)$ from the right to the left as follows:
\begin{align}\label{eq: R-L M3}
\frac{1}{A^r} \cdot \frac{1}{B^{s}A^t} &= \frac{1}{A^r} \sum \limits_{i+j = s+ t} \left(\frac{\nabla^j_t}{(B-A)^jB^i} + \frac{\nabla^j_s}{(A-B)^jA^i}\right)\\\notag
&= \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_t}{(B-A)^j}\frac{\nabla^{j_1}_i}{(A-B)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_t}{(B-A)^j}\frac{\nabla^{j_1}_r}{(B-A)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s+ t}\frac{\nabla^j_s}{(A-B)^j}\frac{1}{A^{i+r}}\\\notag
&= \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_t}{(\lambda F)^j}\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_t}{(\lambda F)^j}\frac{\nabla^{j_1}_r}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s+ t}\frac{\nabla^j_s}{(-\lambda F)^j}\frac{1}{A^{i+r}}\\\notag
&= \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_t(-1)^{j_1}\nabla^{j_1}_i}{\lambda^{j+j_1}}\frac{1}{A^{i_1}F^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \frac{\nabla^j_t\nabla^{j_1}_r}{\lambda^{j+j_1}}\frac{1}{B^{i_1}F^{j+j_1}}\\\notag
&+\sum \limits_{i+j = s+ t}\frac{\nabla^j_s}{(-\lambda)^j}\frac{1}{A^{i+r}F^j}. \notag
\end{align}
For each $(a,b,c)\in M_3$, there exist $\lambda \in \mathbb F_q^\times$ and $f \in A_+$ with $d > \deg f$ such that $b = a + \lambda f$. Replacing $A = a, F = f$, one deduces from \eqref{eq: L-R M3} that
\begin{align*}
\sum \limits_{(a,b,c) \in M_3} &\frac{1}{a^rb^s} \cdot \frac{1}{c^t}
\\&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda \in \mathbb F_q^\times} \frac{1}{a^{r}(a+\lambda f)^s} \cdot \frac{1}{a^t}\\
&= \sum \limits_{i+j = r+s}\delta_j\nabla^j_s\sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i+t}f^j}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\delta_{j+j_1}\nabla^j_r\nabla^{j_1}_t\sum\limits_{\substack{b,f \in A_+ \\ d = \deg b > \deg f}}\frac{1}{b^{i_1}f^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\delta_{j+j_1}\nabla^j_r(-1)^{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i_1}f^{j+j_1}}.\notag\\
&= \sum \limits_{i+j = r+s}\delta_j\nabla^j_s S_d(i+t,j)\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\delta_{j+j_1}\nabla^j_r\nabla^{j_1}_tS_d(i_1,j+j_1)\\\notag
&+ \sum \limits_{i+j = r+s}\sum \limits_{i_1+j_1 = i+t}\delta_{j+j_1}\nabla^j_r(-1)^{j_1}\nabla^{j_1}_iS_d(i_1,j+j_1).\notag
\end{align*}
Similarly, one deduces from \eqref{eq: R-L M3} that
\begin{align*}
\sum \limits_{(a,b,c) \in M_3} &\frac{1}{a^r} \cdot \frac{1}{b^sc^t}\\
&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda \in \mathbb F_q^\times} \frac{1}{a^r} \cdot \frac{1}{(a+\lambda f)^{s}a^t}\\
&= \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \delta_{j+j_1}\nabla^j_t(-1)^{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i_1}f^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \delta_{j+j_1}\nabla^j_t\nabla^{j_1}_r\sum\limits_{\substack{b,f \in A_+ \\ d = \deg b > \deg f}}\frac{1}{b^{i_1}f^{j+j_1}}\\\notag
&+\sum \limits_{i+j = s+ t}\delta_j\nabla^j_s\sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i+r}f^j}. \notag\\
&= \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \delta_{j+j_1}\nabla^j_t(-1)^{j_1}\nabla^{j_1}_i S_d(i_1,j+j_1)\\\notag
&+ \sum \limits_{i+j = s+ t}\sum \limits_{i_1+j_1 = i+r} \delta_{j+j_1}\nabla^j_t\nabla^{j_1}_rS_d(i_1,j+j_1)\\\notag
&+\sum \limits_{i+j = s+ t}\delta_j\nabla^j_s S_d(i+r,j). \notag
\end{align*}
Since the partial fraction decomposition of $P_{\lambda}(A,F)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in M_3} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in M_3} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.
\medskip
\noindent \textbf{Case 4:} $(a,b,c)$ ranges over all tuples in $N_0$.
\par For each $\lambda, \mu \in \mathbb F_q^\times$ such that $\lambda \ne \mu$, consider the rational function
\begin{equation*}
P_{\lambda,\mu}(A,F) = \frac{1}{A^{r}(A+\lambda F)^{s}(A + \mu F)^t}.
\end{equation*}
Set $B = A + \lambda F$ and $C = A + \mu F$. From the same process as in the the case of $M_1$, we expand $P_{\lambda ,\mu}(A,F)$ in two ways. First, we expand $P_{\lambda, \mu}(A,F)$ from the left to the right. From \eqref{eq: PFD L-R}, we have;
\begin{align} \label{eq: L-R N0}
\frac{1}{A^rB^s} \cdot \frac{1}{C^t}
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\mu F)^{j_1}}\frac{1}{A^{i_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_i}{(\mu F)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_t}{[(\lambda - \mu)F]^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_i}{[( \mu-\lambda)F]^{j_1}}\frac{1}{C^{i_1}}\\\notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^j(-\mu)^{j_1}} \frac{1}{A^{i_1}F^{j+j_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s(-1 )^{j_1}\nabla^{j_1}_i}{(-\lambda)^j(-\mu)^{j_1}} \frac{1}{C^{i_1}F^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^j(\lambda - \mu)^{j_1}} \frac{1}{B^{i_1}F^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r(-1)^{j_1}\nabla^{j_1}_i}{\lambda^j(\lambda- \mu)^{j_1}} \frac{1}{C^{i_1}F^{j+j_1}}\notag.
\end{align}
Next we expand $P_{\lambda,\mu}(A,F)$ from the right to the left. From \eqref{eq: PFD R-L}, we have;
\begin{align}\label{eq: R-L N0}
\frac{1}{A^r} \cdot \frac{1}{B^{s}C^t}
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{[(\lambda - \mu)F]^j}\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{[(\lambda - \mu)F]^j}\frac{\nabla^{j_1}_r}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{[( \mu - \lambda)F]^j}\frac{\nabla^{j_1}_i}{(-\mu F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{[(\mu - \lambda)F]^j}\frac{\nabla^{j_1}_r}{(\mu F)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t(-1)^{j_1}\nabla^{j_1}_i}{(\lambda - \mu)^j\lambda^{j_1}}\frac{1}{A^{i_1}F^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t\nabla^{j_1}_r}{(\lambda - \mu)^j\lambda^{j_1}}\frac{1}{B^{i_1}F^{j+j_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s(-1)^{j_1}\nabla^{j_1}_i}{(\mu - \lambda)^j\mu ^{j_1}}\frac{1}{A^{i_1}F^{j+j_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s\nabla^{j_1}_r}{(\mu - \lambda)^j\mu ^{j_1}}\frac{1}{C^{i_1}F^{j+j_1}}.\notag
\end{align}
For each $(a,b,c)\in N_0$, we have $b = a + \lambda f, c = a + \mu f$ with $\lambda, \mu \in \mathbb F_q^\times$ such that $\lambda \ne \mu$ and $f \in A_+$ such that $d > \deg f$. Replacing $A = a, F = f$, one deduces from \eqref{eq: L-R N0} that
\begin{align*}
\sum \limits_{(a,b,c) \in N_0}& \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \\
&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda,\mu \in \mathbb F_q^\times; \lambda \ne \mu} \frac{1}{a^r(a+\lambda f)^s} \cdot \frac{1}{(a+\mu f)^t}\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_t \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i_1}f^{j+j_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s(-1 )^{j_1}\nabla^{j_1}_i \sum\limits_{\substack{c,f \in A_+ \\ d = \deg c > \deg f}}\frac{1}{c^{i_1}f^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r\nabla^{j_1}_t \sum\limits_{\substack{b,f \in A_+ \\ d = \deg b > \deg f}}\frac{1}{b^{i_1}f^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r(-1)^{j_1}\nabla^{j_1}_i \sum\limits_{\substack{c,f \in A_+ \\ d = \deg c > \deg f}}\frac{1}{c^{i_1}f^{j+j_1}}\\\notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_t S_d(i_1,j+j_1)\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s(-1 )^{j_1}\nabla^{j_1}_i S_d(i_1,j+j_1)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r\nabla^{j_1}_t S_d(i_1,j+j_1)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r(-1)^{j_1}\nabla^{j_1}_i S_d(i_1,j+j_1).\notag
\end{align*}
Similarly, one deduces from \eqref{eq: R-L N0} that
\begin{align*}
\sum \limits_{(a,b,c) \in N_0}& \frac{1}{a^r} \cdot \frac{1}{b^sc^t} \\&= \sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}} \sum \limits_{\lambda,\mu \in \mathbb F_q^\times; \lambda \ne \mu} \frac{1}{a^r} \cdot \frac{1}{(a+\lambda f)^s(a+\mu f)^t}\\
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}\nabla^j_t(-1)^{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i_1}f^{j+j_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}\nabla^j_t\nabla^{j_1}_r\sum\limits_{\substack{b,f \in A_+ \\ d = \deg b > \deg f}}\frac{1}{b^{i_1}f^{j+j_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s(-1)^{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,f \in A_+ \\ d = \deg a > \deg f}}\frac{1}{a^{i_1}f^{j+j_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_r\sum\limits_{\substack{c,f \in A_+ \\ d = \deg c > \deg f}}\frac{1}{c^{i_1}f^{j+j_1}}\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}\nabla^j_t(-1)^{j_1}\nabla^{j_1}_i S_d(i_1,j+j_1)\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}\nabla^j_t\nabla^{j_1}_r S_d(i_1,j+j_1)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s(-1)^{j_1}\nabla^{j_1}_i S_d(i_1,j+j_1)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_r S_d(i_1,j+j_1).\notag
\end{align*}
Since the partial fraction decomposition of $P_{\lambda,\mu}(A,F)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in N_0} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in N_0} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.\\
\par Using Formulas \eqref{eq: formula 1}, \eqref{eq: formula 2} and \eqref{eq: formula 3}, one verifies easily that
\begin{align*}
&\sum \limits_{(a,b,c)\in M_1} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} + \sum \limits_{(a,b,c)\in M_2} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} + \sum \limits_{(a,b,c)\in M_3} \frac{1}{a^rb^s} \cdot \frac{1}{c^t}+ \sum \limits_{(a,b,c)\in N_0} \frac{1}{a^rb^s} \cdot \frac{1}{c^t}\\
&= \sum_{i+j=r+s+t} \Delta^j_{r+s,t} S_d(i,j)+\sum_{i+j=r+s} \Delta^j_{r,s} \Bigg(S_d(i+t,j)+\sum_{i_1+j_1=i+t} \Delta^{j_1}_{i,t} S_d(i_1,j_1+j)\Bigg),
\end{align*}
which is the sum of the terms of depth $2$ in the expression of $(S_d(r)S_d(s))S_d(t)$. Similarly, one verifies easily that
\begin{align*}
&\sum \limits_{(a,b,c)\in M_1} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t} + \sum \limits_{(a,b,c)\in M_2} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t} + \sum \limits_{(a,b,c)\in M_3} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}+ \sum \limits_{(a,b,c)\in N_0} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}\\
&= \sum_{i+j=r+s+t} \Delta^j_{r,s+t} S_d(i,j)+\sum_{i+j=s+t} \Delta^j_{s,t} \Bigg(S_d(i+r,j)+\sum_{i_1+j_1=i+r} \Delta^{j_1}_{i,r} S_d(i_1,j_1+j)\Bigg),
\end{align*}
which is the sum the terms of depth $2$ in the expression of $S_d(r)(S_d(s)S_d(t))$.
\subsubsection{Depth 3 terms: proof of Proposition \ref{prop: associativity}, Part 3} ${}$\par
Consider the following cases:
\medskip
\noindent \textbf{Case 1:} $(a,b,c)$ ranges over all tuples in $N_1$.
\par For $\lambda,\mu \in \mathbb F_q^\times$, consider the rational function
\begin{equation*}
P_{\lambda, \mu }(A,F,U) = \frac{1}{A^{r}(A+\lambda F)^s(A+\mu U)^t}.
\end{equation*}
We will deduce the partial fraction decomposition of $ P_{\lambda, \mu }(A,F,U) $ by the following process: we first give the partial fraction decomposition of $ P_{\lambda, \mu }(A,F,U) $ in variable $A$ with coefficients are rational functions of the form $Q(F,U) \in \mathbb F_q(F,U)$; then we continue to give the partial fraction decomposition of $Q(F,U)$ in variable $F$ with coefficients are rational functions of the form $R(U) \in \mathbb F_q(U)$; finally, we give the partial fraction decomposition of $R(U)$ in variable $U$ with coefficients are elements in $\mathbb F_q$.
\par Set $B= A + \lambda F, C = A + \mu U$ and $G = F + \mu' U$ where $\mu' = - \frac{\mu}{\lambda}$, so that $B-C = \lambda G$. Using Lemma \ref{lem: decomposition}, we proceed the process in two ways. First, we expand $P_{\lambda, \mu }(A,F,U)$ from the left to the right. From \eqref{eq: PFD L-R}, we have;
\begin{align} \label{eq: L-R N1}
\frac{1}{A^{r}B^s} \cdot \frac{1}{C^t}
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\mu U)^{j_1}}\frac{1}{A^{i_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_i}{(\mu U)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_t}{(\lambda G)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_i}{(-\lambda G)^{j_1}}\frac{1}{C^{i_1}}.\notag
\end{align}
Here,
\begin{align*}
\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\mu U)^{j_1}}\frac{1}{A^{i_1}} &=\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^j(-\mu)^{j_1}} \frac{1}{A^{i_1}F^jU^{j_1}},\\
\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_i}{(\mu U)^{j_1}}\frac{1}{C^{i_1}}&=\frac{\nabla^j_s\nabla^{j_1}_i}{(-\lambda)^j\mu^{j_1}} \frac{1}{C^{i_1}F^jU^{j_1}}
\end{align*}
and
\begin{align*}
\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_t}{(\lambda G)^{j_1}}\frac{1}{B^{i_1}}
&= \frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-G)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-F)^{j_2}G^{i_2}}\Bigg)\frac{1}{B^{i_1}}\\
&= \frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\mu' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\mu' U)^{j_2}G^{i_2}}\Bigg)\frac{1}{B^{i_1}}\\
&= \frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j_1}}{(-\mu' )^{j_2}}\frac{1}{B^{i_1}F^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(\mu')^{j_2}}\frac{1}{B^{i_1}G^{i_2}U^{j_2}},\\
\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_i}{(-\lambda G)^{j_1}}\frac{1}{C^{i_1}} &= \frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^{j+j_1}}\sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-G)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-F)^{j_2}G^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\
&= \frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^{j+j_1}}\sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\mu' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\mu' U)^{j_2}G^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\
&= \frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^{j+j_1}}\sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j_1}}{(-\mu' )^{j_2}}\frac{1}{C^{i_1}F^{i_2}U^{j_2}}\\
&+
\frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(\mu')^{j_2}}\frac{1}{C^{i_1}G^{i_2}U^{j_2}}.
\end{align*}
Next we expand $P_{\lambda, \mu }(A,F,U)$ from the right to the left. From \eqref{eq: PFD R-L}, we have;
\begin{align} \label{eq: R-L N1}
\frac{1}{A^r} \cdot \frac{1}{B^{s}C^t}
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(\lambda G)^j}\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(\lambda G)^j}\frac{\nabla^{j_1}_r}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(-\lambda G)^j}\frac{\nabla^{j_1}_i}{(-\mu U)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(-\lambda G)^j}\frac{\nabla^{j_1}_r}{(\mu U)^{j_1}}\frac{1}{C^{i_1}}.\notag
\end{align}
Here,
\begin{align*}
\frac{\nabla^j_t}{(\lambda G)^j}\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}
&=\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(G-F)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(F-G)^{j_2}F^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\
&=\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(\mu' U)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(-\mu' U)^{j_2}F^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\
&=\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j_1}}{(\mu')^{j_2}} \frac{1}{A^{i_1}G^{i_2}U^{j_2}}\\
&+
\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(-\mu')^{j_2}}\frac{1}{A^{i_1}F^{i_2}U^{j_2}},\\
\frac{\nabla^j_t}{(\lambda G)^j}\frac{\nabla^{j_1}_r}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}} &=\frac{\nabla^j_t\nabla^{j_1}_r}{\lambda ^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(G-F)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(F-G)^{j_2}F^{i_2}}\Bigg)\frac{1}{B^{i_1}}\\
&=\frac{\nabla^j_t\nabla^{j_1}_r}{\lambda ^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\Bigg(\frac{\nabla^{j_2}_{j_1}}{(\mu' U)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(-\mu' U)^{j_2}F^{i_2}}\Bigg)\frac{1}{B^{i_1}}\\
&=\frac{\nabla^j_t\nabla^{j_1}_r}{\lambda ^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j_1}}{(\mu')^{j_2}} \frac{1}{B^{i_1}G^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_t\nabla^{j_1}_r}{\lambda ^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(-\mu')^{j_2}}\frac{1}{B^{i_1}F^{i_2}U^{j_2}}.
\end{align*}
and
\begin{align*}
\frac{\nabla^j_s}{(-\lambda G)^j}\frac{\nabla^{j_1}_i}{(-\mu U)^{j_1}}\frac{1}{A^{i_1}} &=\frac{\nabla^j_s\nabla^{j_1}_i}{(-\lambda )^j(-\mu)^{j_1}}\frac{1}{A^{i_1}G^jU^{j_1}},\\
\frac{\nabla^j_s}{(-\lambda G)^j}\frac{\nabla^{j_1}_r}{(\mu U)^{j_1}}\frac{1}{C^{i_1}}&=\frac{\nabla^j_s\nabla^{j_1}_r}{(-\lambda)^j\mu^{j_1}}\frac{1}{C^{i_1}G^jU^{j_1}}.
\end{align*}
For each $(a,b,c)\in N_1$, we have $b = a + \lambda f, c = a + \mu u$ with $\lambda, \mu \in \mathbb F_q^\times$ and $f,u \in A_+$ such that $d > \deg f > \deg u$. Set $g = f + \mu' u$ where $\mu' = - \frac{\mu}{\lambda}$. Replacing $A = a, F = f,U = u$, one deduces from \eqref{eq: L-R N1} that
\begin{align*}
\sum \limits_{(a,b,c) \in N_1}& \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\lambda, \mu \in \mathbb F_q^\times} \frac{1}{a^{r}(a+\lambda f)^s} \cdot \frac{1}{(a+\mu u)^t}\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_t \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^ju^{j_1}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_i \sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^ju^{j_1}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_r\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1}\sum\limits_{\substack{b,f,u \in A_+ \\ d = \deg b > \deg f> \deg u}}\frac{1}{b^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_r\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{b,g,u \in A_+ \\ d = \deg b > \deg g> \deg u}}\frac{1}{b^{i_1}g^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1}\sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{i_2}u^{j_2}}\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_t S_d(i_1,j,j_1)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_i S_d(i_1,j,j_1)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_r\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_r\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\Delta^{j_1}_{i,t} S_d(i_1,j,j_1)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_r\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1}\Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1}\Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2).
\end{align*}
Similarly, one deduces from \eqref{eq: R-L N1} that
\begin{align*}
\sum \limits_{(a,b,c) \in N_1}& \frac{1}{a^r} \cdot \frac{1}{b^sc^t}\\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\lambda, \mu \in \mathbb F_q^\times} \frac{1}{a^r} \cdot \frac{1}{(a+\lambda f)^s(a+\mu u)^t}\\
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}\nabla^j_t\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{b,g,u \in A_+ \\ d = \deg b > \deg g> \deg u}}\frac{1}{b^{i_1}g^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}\nabla^j_t\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{b,f,u \in A_+ \\ d = \deg b > \deg f> \deg u}}\frac{1}{b^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^ju^{j_1}}\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_r\sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^ju^{j_1}}\\
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}\nabla^j_t\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}\nabla^j_t\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_iS_d(i_1,j,j_1)\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_rS_d(i_1,j,j_1)\\
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j+j_1}\nabla^j_t\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1}\Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_s\Delta^{j_1}_{i,r}S_d(i_1,j,j_1).
\end{align*}
Since the partial fraction decomposition of $P_{\lambda,\mu}(A,F,U)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in N_1} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in N_1} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.
\medskip
\noindent \textbf{Case 2:} $(a,b,c)$ ranges over all tuples in $N_2$.
\par For $\lambda,\mu \in \mathbb F_q^\times$, consider the rational function
\begin{equation*}
P_{\lambda, \mu }(A,F,U) = \frac{1}{A^{r}(A+\lambda U)^s(A+\mu F)^t}.
\end{equation*}
\par Set $B= A + \lambda U, C = A + \mu F$ and $G = F + \lambda' U$ where $\lambda' = - \frac{\lambda}{\mu}$, so that $C-B = \mu G$. From the same process as in the the case of $N_1$, we expand $P_{\lambda, \mu }(A,F,U)$ in two ways. First, we expand $P_{\lambda, \mu }(A,F,U)$ from the left to the right. From \eqref{eq: PFD L-R}, we have;
\begin{align} \label{eq: L-R N2}
\frac{1}{A^{r}B^s} \cdot \frac{1}{C^t}
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda U)^j} \frac{\nabla^{j_1}_t}{(-\mu F)^{j_1}}\frac{1}{A^{i_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda U)^j} \frac{\nabla^{j_1}_i}{(\mu F)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda U)^j} \frac{\nabla^{j_1}_t}{(-\mu G)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda U)^j} \frac{\nabla^{j_1}_i}{(\mu G)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^j(-\mu)^{j_1}} \frac{1}{A^{i_1}F^{j_1}U^j}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s\nabla^{j_1}_i}{(-\lambda)^j\mu^{j_1}} \frac{1}{C^{i_1}F^{j_1}U^j}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^j(-\mu)^{j_1}} \frac{1}{B^{i_1}G^{j_1}U^j}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r\nabla^{j_1}_i}{\lambda^j\mu^{j_1}} \frac{1}{C^{i_1}G^{j_1}U^j}\\\notag
\end{align}
Next we expand $P_{\lambda, \mu }(A,F,U)$ from the right to the left. From \eqref{eq: PFD R-L}, we have;
\begin{align} \label{eq: R-L N2}
\frac{1}{A^r} \cdot \frac{1}{B^{s}C^t}
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(-\mu G)^j}\frac{\nabla^{j_1}_i}{(-\lambda U)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(-\mu G)^j}\frac{\nabla^{j_1}_r}{(\lambda U)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(\mu G)^j}\frac{\nabla^{j_1}_i}{(-\mu F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(\mu G)^j}\frac{\nabla^{j_1}_r}{(\mu F)^{j_1}}\frac{1}{C^{i_1}}.\notag
\end{align}
Here
\begin{align*}
\frac{\nabla^j_t}{(-\mu G)^j}\frac{\nabla^{j_1}_i}{(-\lambda U)^{j_1}}\frac{1}{A^{i_1}}
&= \frac{\nabla^j_t\nabla^{j_1}_i}{(-\mu)^j(-\lambda)^{j_1}}\frac{1}{A^{i_1}G^jU^{j_1}},\\
\frac{\nabla^j_t}{(-\mu G)^j}\frac{\nabla^{j_1}_r}{(\lambda U)^{j_1}}\frac{1}{B^{i_1}}
&= \frac{\nabla^j_t\nabla^{j_1}_r}{(-\mu)^j\lambda ^{j_1}}\frac{1}{B^{i_1}G^jU^{j_1}}
\end{align*}
and
\begin{align*}
\frac{\nabla^j_s}{(\mu G)^j}\frac{\nabla^{j_1}_i}{(-\mu F)^{j_1}}\frac{1}{A^{i_1}}
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{(-\mu)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(G-F)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(F-G)^{j_2}F^{i_2}}\Bigg)\frac{1}{A^{i_1}} \\
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{(-\mu)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(\lambda' U)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(-\lambda' U)^{j_2}F^{i_2}}\Bigg)\frac{1}{A^{i_1}} \\
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{(-\mu)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(\lambda' )^{j_2}}\frac{1}{A^{i_1}G^{i_2}U^{j_2}}\\
&+
\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{(-\mu)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(-\lambda')^{j_2}}\frac{1}{A^{i_1}F^{i_2}U^{j_2}},\\
\frac{\nabla^j_s}{(\mu G)^j}\frac{\nabla^{j_1}_r}{(\mu F)^{j_1}}\frac{1}{C^{i_1}}
&= \frac{\nabla^j_s\nabla^{j_1}_r}{\mu^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(G-F)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(F-G)^{j_2}F^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\
&= \frac{\nabla^j_s\nabla^{j_1}_r}{\mu^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(\lambda' U)^{j_2}G^{i_2}} + \frac{\nabla^{j_2}_{j}}{(-\lambda' U)^{j_2}F^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\
&= \frac{\nabla^j_s\nabla^{j_1}_r}{\mu^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(\lambda' )^{j_2}}\frac{1}{C^{i_1}G^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_s\nabla^{j_1}_r}{\mu^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(-\lambda')^{j_2}}\frac{1}{C^{i_1}F^{i_2}U^{j_2}}.
\end{align*}
For each $(a,b,c)\in N_2$, we have $b = a + \lambda u, c = a + \mu f$ with $\lambda, \mu \in \mathbb F_q^\times$ and $f,u \in A_+$ such that $d > \deg f > \deg u$. Set $g = f + \lambda' u$ where $\lambda' = - \frac{\lambda}{\mu}$. Replacing $A = a, F = f,U = u$, one deduces from \eqref{eq: L-R N2} that
\begin{align*}
\sum \limits_{(a,b,c) \in N_2}& \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\lambda, \mu \in \mathbb F_q^\times} \frac{1}{a^{r}(a+\lambda u)^s} \cdot \frac{1}{(a+\mu f)^t}\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_2}\nabla^{j_1}_t \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^{j_1}u^j}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_i \sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^{j_1}u^j}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_r\delta_{j_1}\nabla^{j_1}_t \sum\limits_{\substack{b,g,u \in A_+ \\ d = \deg b > \deg g> \deg u}}\frac{1}{b^{i_1}g^{j_1}u^j}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_r\delta_{j_1}\nabla^{j_1}_i \sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{j_1}u^j}\\\notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_2}\nabla^{j_1}_t S_d(i_1,j_1,j)\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_i S_d(i_1,j_1,j)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_r\delta_{j_1}\nabla^{j_1}_t S_d(i_1,j_1,j)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_r\delta_{j_1}\nabla^{j_1}_i S_d(i_1,j_1,j)\\\notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\Delta^{j}_{r,s} \Delta^{j_1}_{i,t}S_d(i_1,j_1,j).
\end{align*}
Similarly, one deduces from \eqref{eq: R-L N2} that
\begin{align*}
\sum \limits_{(a,b,c) \in N_2} &\frac{1}{a^r} \cdot \frac{1}{b^sc^t} \\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\lambda, \mu \in \mathbb F_q^\times} \frac{1}{a^r} \cdot \frac{1}{(a+\lambda u)^s(a+\mu f)^t}\\
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j}\nabla^j_t\delta_{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^ju^{j_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_t\delta_{j_1}\nabla^{j_1}_r\sum\limits_{\substack{b,g,u \in A_+ \\ d = \deg b > \deg g> \deg u}}\frac{1}{b^{i_1}g^ju^{j_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}\sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^{i_2}u^{j_2}} \\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^{i_2}u^{j_2}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}\sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{i_2}u^{j_2}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^{i_2}u^{j_2}}\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j}\nabla^j_t\delta_{j_1}\nabla^{j_1}_iS_d(i_1,j,j_1)\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_t\delta_{j_1}\nabla^{j_1}_rS_d(i_1,j,j_1)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}S_d(i_1,i_2,j_2) \\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2) \\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}S_d(i_1,i_2,j_2) \\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2) \\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j}\nabla^{j}_t\Delta^{j_1}_{i,r}S_d(i_1,j,j_1)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1}\Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2) \\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_r \sum \limits_{i_2 + j_2 = j + j_1} \Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2).\notag
\end{align*}
Since the partial fraction decomposition of $P_{\lambda,\mu}(A,F,U)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in N_2} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in N_2} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.
\medskip
\noindent \textbf{Case 3:} $(a,b,c)$ ranges over all tuples in $N_3$.
\par For $\lambda,\mu \in \mathbb F_q^\times$, consider the rational function
\begin{equation*}
P_{\lambda, \mu }(A,F,U) = \frac{1}{A^{r}(A+\lambda F)^s(A+ \lambda F + \mu U)^t}.
\end{equation*}
\par Set $B= A + \lambda F, C = A + \lambda F + \mu U$ and $G = F + \mu' U$ where $\mu' = \frac{\mu}{\lambda}$, so that $C-B = \mu U$ and $C-A = \lambda G$. From the same process as in the the case of $N_1$, we expand $P_{\lambda, \mu }(A,F,U)$ in two ways. First, we expand $P_{\lambda, \mu }(A,F,U)$ from the left to the right. From \eqref{eq: PFD L-R}, we have;
\begin{align} \label{eq: L-R N3}
\frac{1}{A^{r}B^s} \cdot \frac{1}{C^t}
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\lambda G)^{j_1}}\frac{1}{A^{i_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_i}{(\lambda G)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\mu U)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_i}{(\mu U)^{j_1}}\frac{1}{C^{i_1}}\notag.
\end{align}
Here,
\begin{align*}
\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\lambda G)^{j_1}}\frac{1}{A^{i_1}}
&= \frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-G)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-F)^{j_2}G^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\\notag
&= \frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\mu' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\mu' U)^{j_2}G^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\\notag
&= \frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(-\mu' )^{j_2}}\frac{1}{A^{i_1}F^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^{j+j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j}}{(\mu' )^{j_2}}\frac{1}{A^{i_1}G^{i_2}U^{j_2}},\\\notag
\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_i}{(\lambda G)^{j_1}}\frac{1}{C^{i_1}}
&= \frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^{j+j_1}}\sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-G)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-F)^{j_2}G^{i_2}}\Bigg) \frac{1}{C^{i_1}} \\\notag
&= \frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^{j+j_1}}\sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\mu' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\mu' U)^{j_2}G^{i_2}}\Bigg) \frac{1}{C^{i_1}} \\\notag
&= \frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^{j+j_1}}\sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(-\mu')^{j_2}} \frac{1}{C^{i_1}F^{i_2}U^{j_2}} \\
&+
\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^{j+j_1}}\sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(\mu')^{j_2}} \frac{1}{C^{i_1}G^{i_2}U^{j_2}}.\notag
\end{align*}
and
\begin{align*}
\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\mu U)^{j_1}}\frac{1}{B^{i_1}} &=\frac{\nabla^j_r\nabla^{j_1}_t}{(\lambda)^j(-\mu )^{j_1}} \frac{1}{B^{i_1}F^jU^{j_1}},\\\notag
\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_i}{(\mu U)^{j_1}}\frac{1}{C^{i_1}} &= \frac{\nabla^j_r\nabla^{j_1}_i}{\lambda^j\mu ^{j_1}} \frac{1}{C^{i_1}F^jU^{j_1}}\notag.
\end{align*}
Next we expand $P_{\lambda, \mu }(A,F,U)$ from the right to the left. From \eqref{eq: PFD R-L}, we have
\begin{align} \label{eq: R-L N3}
\frac{1}{A^r} \cdot \frac{1}{B^{s}C^t}
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(-\mu U)^j}\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{(-\mu U)^j}\frac{\nabla^{j_1}_r}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(\mu U)^j}\frac{\nabla^{j_1}_i}{(-\lambda G)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{(\mu U)^j}\frac{\nabla^{j_1}_r}{(\lambda G)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t\nabla^{j_1}_i}{(-\mu)^j(-\lambda)^{j_1}}\frac{1}{A^{i_1}F^{j_1}U^j}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t\nabla^{j_1}_r}{(-\mu )^j\lambda ^{j_1}}\frac{1}{B^{i_1}F^{j_1}U^j}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s\nabla^{j_1}_i}{\mu ^j(-\lambda)^{j_1}}\frac{1}{A^{i_1}G^{j_1}U^j}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s\nabla^{j_1}_r}{\mu ^j\lambda ^{j_1}}\frac{1}{C^{i_1}G^{j_1}U^j}.\notag
\end{align}
For each $(a,b,c)\in N_3$, we have $b = a + \lambda f, c = a + \lambda f + \mu u$ with $\lambda, \mu \in \mathbb F_q^\times$ and $f,u \in A_+$ such that $d > \deg f > \deg u$. Set $g = f + \mu'u$ where $\mu' = \frac{\mu}{\lambda} $. Replacing $A = a, F = f,U = u$, one deduces from \eqref{eq: L-R N3} that
\vspace{-0.2cm}
\begin{align*}
\sum \limits_{(a,b,c) \in N_3}& \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\lambda, \mu \in \mathbb F_q^\times} \frac{1}{a^{r}(a+\lambda f)^s} \cdot \frac{1}{(a+\lambda f + \mu u)^t}\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}\sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^{i_2}u^{j_2}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^{i_2}u^{j_2}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t} \delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^{i_2}u^{j_2}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j} \sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{i_2}u^{j_2}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_r\delta_{j_1}\nabla^{j_1}_t \sum\limits_{\substack{b,f,u \in A_+ \\ d = \deg b > \deg f> \deg u}}\frac{1}{b^{i_1}f^ju^{j_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j}\nabla^j_r\delta_{j_1}\nabla^{j_1}_i \sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^ju^{j_1}}\\\notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}S_d(i_1,i_2,j_2)\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t} \delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j} S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_j\nabla^j_r\delta_{j_1}\nabla^{j_1}_t S_d(i_1,j,j_1)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j}\nabla^j_r\delta_{j_1}\nabla^{j_1}_i S_d(i_1,j,j_1)\\\notag
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j+j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t} \delta_{j+j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1} \Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j}\nabla^j_r\Delta^{j_1}_{i,t} S_d(i_1,j,j_1).
\end{align*}
Similarly, one deduces from \eqref{eq: R-L N3} that
\vspace{-0.2cm}
\begin{align*}
\sum \limits_{(a,b,c) \in N_3}& \frac{1}{a^r} \cdot \frac{1}{b^sc^t} \\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\lambda, \mu \in \mathbb F_q^\times} \frac{1}{a^r} \cdot \frac{1}{(a+\lambda f)^s(a+\lambda f + \mu u)^t}\\
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_t\delta_{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^{j_1}u^j}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_t\delta_{j_1}\nabla^{j_1}_r\sum\limits_{\substack{b,f,u \in A_+ \\ d = \deg b > \deg f> \deg u}}\frac{1}{b^{i_1}f^{j_1}u^j}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_i\sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^{j_1}u^j}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_r\sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{j_1}u^j}\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_t\delta_{j_1}\nabla^{j_1}_iS_d(i_1,j_1,j)\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_j\nabla^j_t\delta_{j_1}\nabla^{j_1}_rS_d(i_1,j_1,j)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_iS_d(i_1,j_1,j)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_j\nabla^j_s\delta_{j_1}\nabla^{j_1}_rS_d(i_1,j_1,j)\\\notag
&=\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\Delta^{j}_{s,t} \Delta^{j_1}_{i,r}S_d(i_1,j_1,j).\notag
\end{align*}
Since the partial fraction decomposition of $P_{\lambda,\mu}(A,F,U)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in N_3} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in N_3} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.
\medskip
\noindent \textbf{Case 4:} $(a,b,c)$ ranges over all tuples in $N_4$.
\par For $\lambda,\mu, \eta \in \mathbb F_q^\times$ such that $\lambda \ne \mu$, consider the rational function
\begin{equation*}
P_{\lambda, \mu , \eta}(A,F,U) = \frac{1}{A^{r}(A+\lambda F)^s(A+\mu F + \eta U)^t}.
\end{equation*}
\par Set $B= A + \lambda F, C = A + \mu F + \eta U, G = F + \eta'U$ where $\eta' = \frac{\eta}{\mu}$ and $H = F + \eta'' U$ where $\eta'' = \frac{\eta}{\mu - \lambda}$, so that $C-A = \mu G$ and $C-B = (\mu - \lambda)H$. It should be remarked that $\eta' \ne \eta''$ since $\lambda \ne 0$, From the same process as in the the case of $N_1$, we expand $P_{\lambda, \mu , \eta}(A,F,U)$ in two ways. First, we expand $P_{\lambda, \mu , \eta}(A,F,U)$ from the left to the right. From \eqref{eq: PFD L-R}, we have
\begin{align} \label{eq: L-R N4}
\frac{1}{A^{r}B^s} \cdot \frac{1}{C^t}
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_t}{(-\mu G)^{j_1}}\frac{1}{A^{i_1}}\\ \notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_s}{(-\lambda F)^j} \frac{\nabla^{j_1}_i}{(\mu G)^{j_1}}\frac{1}{C^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_t}{[(\lambda - \mu )H]^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\frac{\nabla^j_r}{(\lambda F)^j} \frac{\nabla^{j_1}_i}{[(\mu - \lambda )H]^{j_1}}\frac{1}{C^{i_1}}\notag.
\end{align}
Here
\begin{align*}
\frac{\nabla^j_s}{(-\lambda F)^j} &\frac{\nabla^{j_1}_t}{(-\mu G)^{j_1}}\frac{1}{A^{i_1}} \\
&=\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^j(-\mu)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-G)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-F)^{j_2}G^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\ \notag
&=\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^j(-\mu)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\eta' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\eta' U)^{j_2}G^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\ \notag
&=\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^j(-\mu)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(-\eta')^{j_2}}\frac{1}{A^{i_1}F^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_s\nabla^{j_1}_t}{(-\lambda)^j(-\mu)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(\eta')^{j_2}}\frac{1}{A^{i_1}G^{i_2}U^{j_2}},\\ \notag
\frac{\nabla^j_s}{(-\lambda F)^j} &\frac{\nabla^{j_1}_i}{(\mu G)^{j_1}}\frac{1}{C^{i_1}} \\
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^j\mu^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-G)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-F)^{j_2}G^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^j\mu^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\eta' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\eta' U)^{j_2}G^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^j\mu^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(-\eta')^{j_2}} \frac{1}{C^{i_1}F^{i_2}U^{j_2}}\\
&+
\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{\lambda^j\mu^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j}}{(\eta' )^{j_2}}\frac{1}{C^{i_1}G^{i_2}U^{j_2}},\\\notag
\frac{\nabla^j_r}{(\lambda F)^j} &\frac{\nabla^{j_1}_t}{[(\lambda - \mu )H]^{j_1}}\frac{1}{B^{i_1}} \\
&=\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^j(\lambda - \mu )^{j_1}}\sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-H)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(H-F)^{j_2}H^{i_2}}\Bigg) \frac{1}{B^{i_1}}\\\notag
&=\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^j(\lambda - \mu )^{j_1}}\sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\eta'' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\eta'' U)^{j_2}G^{i_2}}\Bigg) \frac{1}{B^{i_1}}\\\notag
&=\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^j(\lambda - \mu )^{j_1}}\sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(-\eta'')^{j_2}} \frac{1}{B^{i_1}F^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_r\nabla^{j_1}_t}{\lambda^j(\lambda - \mu )^{j_1}}\sum \limits_{i_2 + j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(\eta'')^{j_2}} \frac{1}{B^{i_1}G^{i_2}U^{j_2}},\\\notag
\frac{\nabla^j_r}{(\lambda F)^j}& \frac{\nabla^{j_1}_i}{[(\mu - \lambda )H]^{j_1}}\frac{1}{C^{i_1}} \\
&=\frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^j(\mu - \lambda)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(F-H)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(H-F)^{j_2}H^{i_2}}\Bigg) \frac{1}{C^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^j(\mu - \lambda)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(-\eta'' U)^{j_2}F^{i_2}} + \frac{\nabla^{j_2}_{j}}{(\eta'' U)^{j_2}G^{i_2}}\Bigg) \frac{1}{C^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^j(\mu - \lambda)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(-\eta'' )^{j_2}} \frac{1}{C^{i_1}F^{i_2}U^{j_2}}\\
&+
\frac{(-1)^j\nabla^j_r\nabla^{j_1}_i}{(-\lambda )^j(\mu - \lambda)^{j_1}} \sum \limits_{i_2 + j_2 = j + j_1} \frac{\nabla^{j_2}_{j}}{(\eta'' )^{j_2}}\frac{1}{C^{i_1}G^{i_2}U^{j_2}}.\notag
\end{align*}
Next we expand $P_{\lambda, \mu , \eta}(A,F,U)$ from the right to the left. From \eqref{eq: PFD R-L}, we have;
\begin{align} \label{eq: R-L N4}
\frac{1}{A^r} \cdot \frac{1}{B^{s}C^t}
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{[(\lambda - \mu )H]^j}\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \frac{\nabla^j_t}{[(\lambda - \mu )H]^j}\frac{\nabla^{j_1}_r}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{[(\mu - \lambda )H]^j}\frac{\nabla^{j_1}_i}{(-\mu G)^{j_1}}\frac{1}{A^{i_1}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\frac{\nabla^j_s}{[(\mu - \lambda )H]^j}\frac{\nabla^{j_1}_r}{(\mu G)^{j_1}}\frac{1}{C^{i_1}}.\notag
\end{align}
Here
\begin{align*}
\frac{\nabla^j_t}{[(\lambda - \mu )H]^j}&\frac{\nabla^{j_1}_i}{(-\lambda F)^{j_1}}\frac{1}{A^{i_1}} \\
&=\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(\mu - \lambda)^j(-\lambda )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(H-F)^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{(F-H)^{j_2}F^{i_2}}\Bigg) \frac{1}{A^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(\mu - \lambda)^j(-\lambda )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(\eta''U)^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{(-\eta''U)^{j_2}F^{i_2}}\Bigg) \frac{1}{A^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(\mu - \lambda)^j(-\lambda )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(\eta'')^{j_2}} \frac{1}{A^{i_1}H^{i_2}U^{j_2}}\\
&+
\frac{(-1)^j\nabla^j_t\nabla^{j_1}_i}{(\mu - \lambda)^j(-\lambda )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(-\eta'')^{j_2}}\frac{1}{A^{i_1}F^{i_2}U^{j_2}},\\\notag
\frac{\nabla^j_t}{[(\lambda - \mu )H]^j}&\frac{\nabla^{j_1}_r}{(\lambda F)^{j_1}}\frac{1}{B^{i_1}} \\
&=\frac{\nabla^j_t\nabla^{j_1}_r}{(\lambda - \mu )^j\lambda^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(H-F)^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{(F-H)^{j_2}F^{i_2}}\Bigg)\frac{1}{B^{i_1}}\\\notag
&=\frac{\nabla^j_t\nabla^{j_1}_r}{(\lambda - \mu )^j\lambda^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(\eta''U)^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{(-\eta''U)^{j_2}F^{i_2}}\Bigg)\frac{1}{B^{i_1}}\\\notag
&=\frac{\nabla^j_t\nabla^{j_1}_r}{(\lambda - \mu )^j\lambda^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(\eta'')^{j_2}} \frac{1}{B^{i_1}H^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_t\nabla^{j_1}_r}{(\lambda - \mu )^j\lambda^{j_1}}\sum \limits_{i_2+j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(-\eta'')^{j_2}}\frac{1}{B^{i_1}F^{i_2}U^{j_2}},\\\notag
\frac{\nabla^j_s}{[(\mu - \lambda )H]^j}&\frac{\nabla^{j_1}_i}{(-\mu G)^{j_1}}\frac{1}{A^{i_1}} \\
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{( \lambda - \mu)^j(-\mu )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(H-G)^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-H)^{j_2}G^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{( \lambda - \mu)^j(-\mu )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{[(\eta''-\eta')U]^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{[(\eta'-\eta'')U]^{j_2}G^{i_2}}\Bigg)\frac{1}{A^{i_1}}\\\notag
&=\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{( \lambda - \mu)^j(-\mu )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(\eta''-\eta')^{j_2}}\frac{1}{A^{i_1}H^{i_2}U^{j_2}} \\
&+
\frac{(-1)^j\nabla^j_s\nabla^{j_1}_i}{( \lambda - \mu)^j(-\mu )^{j_1}}\sum \limits_{i_2+j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(\eta'-\eta'')^{j_2}}\frac{1}{A^{i_1}G^{i_2}U^{j_2}},\\\notag
\frac{\nabla^j_s}{[(\mu - \lambda )H]^j}&\frac{\nabla^{j_1}_r}{(\mu G)^{j_1}}\frac{1}{C^{i_1}} \\
&=\frac{\nabla^j_s\nabla^{j_1}_r}{(\mu - \lambda )^j\mu ^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{(H-G)^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{(G-H)^{j_2}G^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\\notag
&=\frac{\nabla^j_s\nabla^{j_1}_r}{(\mu - \lambda )^j\mu ^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \Bigg(\frac{\nabla^{j_2}_{j_1}}{[(\eta''-\eta')U]^{j_2}H^{i_2}} + \frac{\nabla^{j_2}_{j}}{[(\eta'-\eta'')U]^{j_2}G^{i_2}}\Bigg)\frac{1}{C^{i_1}}\\\notag
&=\frac{\nabla^j_s\nabla^{j_1}_r}{(\mu - \lambda )^j\mu ^{j_1}}\sum \limits_{i_2+j_2 = j + j_1} \frac{\nabla^{j_2}_{j_1}}{(\eta''-\eta')^{j_2}} \frac{1}{C^{i_1}H^{i_2}U^{j_2}}\\
&+
\frac{\nabla^j_s\nabla^{j_1}_r}{(\mu - \lambda )^j\mu ^{j_1}}\sum \limits_{i_2+j_2 = j + j_1}\frac{\nabla^{j_2}_{j}}{(\eta'-\eta'')^{j_2}}\frac{1}{C^{i_1}G^{i_2}U^{j_2}}.\notag
\end{align*}
For each $(a,b,c)\in N_4$, we have $b = a + \lambda f, c = a + \mu f + \eta u$ with $\lambda, \mu, \eta \in \mathbb F_q^\times$ such that $\lambda \ne \mu$ and $f,u \in A_+$ such that $d > \deg f > \deg u$. Set $g = f + \eta'u$ where $\eta' = \frac{\eta}{\mu}$ and $ h = f + \eta'' u$ where $\eta'' = \frac{\eta}{\mu - \lambda}$. Replacing $A = a, F = f,U = u$, one deduces from \eqref{eq: L-R N4} that
\vspace{-0.2cm}
\begin{align*}
\sum \limits_{(a,b,c) \in N_4} & \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\substack{\lambda, \mu, \eta \in \mathbb F_q^\times \\ \lambda \ne \mu}} \frac{1}{a^{r}(a+\lambda f)^s} \cdot \frac{1}{(a+\mu f + \eta u)^t}\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}\sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_t\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r\nabla^{j_1}_t\sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{b,f,u \in A_+ \\ d = \deg b > \deg f> \deg u}}\frac{1}{b^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r\nabla^{j_1}_t\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1}\nabla^{j_2}_{j} \sum\limits_{\substack{b,g,u \in A_+ \\ d = \deg b > \deg g> \deg u}}\frac{1}{b^{i_1}g^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{c,f,u \in A_+ \\ d = \deg c > \deg f> \deg u}}\frac{1}{c^{i_1}f^{i_2}u^{j_2}}\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{i_2}u^{j_2}}\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_t\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r\nabla^{j_1}_t\sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r\nabla^{j_1}_t\sum \limits_{i_2 + j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j_1}\nabla^{j_2}_{j} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\
&= \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_t \sum \limits_{i_2 + j_2 = j + j_1} \Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2 + j_2 = j + j_1} \Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}\nabla^j_r\nabla^{j_1}_t\sum \limits_{i_2 + j_2 = j + j_1} \Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2)\\
&+ \sum \limits_{i+j = r+s} \sum \limits_{i_1+j_1 = i + t}\delta_{j,j_1}(-1)^j\nabla^j_r\nabla^{j_1}_i \sum \limits_{i_2 + j_2 = j + j_1} \Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2).
\end{align*}
Similarly, one deduces from \eqref{eq: R-L N4} that
\vspace{-0.2cm}
\begin{align*}
\sum \limits_{(a,b,c) \in N_4}& \frac{1}{a^r} \cdot \frac{1}{b^sc^t} \\
&= \sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}} \sum \limits_{\substack{\lambda, \mu, \eta \in \mathbb F_q^\times \\ \lambda \ne \mu}} \frac{1}{a^r} \cdot \frac{1}{(a+\lambda f)^s(a+\mu f + \eta u)^t}\\
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{a,h,u \in A_+ \\ d = \deg a > \deg h> \deg u}}\frac{1}{a^{i_1}h^{i_2}u^{j_2}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{a,f,u \in A_+ \\ d = \deg a > \deg f> \deg u}}\frac{1}{a^{i_1}f^{i_2}u^{j_2}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_t\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{b,h,u \in A_+ \\ d = \deg b > \deg h> \deg u}}\frac{1}{b^{i_1}h^{i_2}u^{j_2}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_t\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{b,f,u \in A_+ \\ d = \deg b > \deg f> \deg u}}\frac{1}{b^{i_1}f^{i_2}u^{j_2}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}\sum\limits_{\substack{a,h,u \in A_+ \\ d = \deg a > \deg h> \deg u}}\frac{1}{a^{i_1}h^{i_2}u^{j_2}}\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{a,g,u \in A_+ \\ d = \deg a > \deg g> \deg u}}\frac{1}{a^{i_1}g^{i_2}u^{j_2}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} \sum\limits_{\substack{c,h,u \in A_+ \\ d = \deg c > \deg h> \deg u}}\frac{1}{c^{i_1}h^{i_2}u^{j_2}}\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}\sum\limits_{\substack{c,g,u \in A_+ \\ d = \deg c > \deg g> \deg u}}\frac{1}{c^{i_1}g^{i_2}u^{j_2}}\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_t\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_t\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1}S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1} \delta_{j_2}\nabla^{j_2}_{j_1} S_d(i_1,i_2,j_2)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1}\delta_{j_2}\nabla^{j_2}_{j}S_d(i_1,i_2,j_2)\\\notag
&= \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_t\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1} \Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_t\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1} \Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2)\\\notag
&+ \sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r} \delta_{j,j_1}(-1)^j\nabla^j_s\nabla^{j_1}_i\sum \limits_{i_2+j_2 = j + j_1} \Delta^{j_2}_{j,j_1}S_d(i_1,i_2,j_2)\\\notag
&+\sum \limits_{i+j = s + t}\sum \limits_{i_1+j_1 = i + r}\delta_{j,j_1}\nabla^j_s\nabla^{j_1}_r\sum \limits_{i_2+j_2 = j + j_1} \Delta^{j_2}_{j,j_1} S_d(i_1,i_2,j_2).\notag
\end{align*}
Since the partial fraction decomposition of $P_{\lambda, \mu , \eta}(A,F,U)$ obtained from the process is unique, it follows that the above expansions of
\begin{equation*}
\sum \limits_{(a,b,c)\in N_4} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} \quad \text{and} \quad \sum\limits_{(a,b,c)\in N_4} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}
\end{equation*}
yield the same expression in terms of power sums.\\
\par Using Formulas \eqref{eq: formula 1}, \eqref{eq: formula 2} and \eqref{eq: formula 3}, one verifies easily that
\begin{align*}
&\sum \limits_{(a,b,c)\in N_1} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} + \sum \limits_{(a,b,c)\in N_2} \frac{1}{a^rb^s} \cdot \frac{1}{c^t} + \sum \limits_{(a,b,c)\in N_3} \frac{1}{a^rb^s} \cdot \frac{1}{c^t}+ \sum \limits_{(a,b,c)\in N_4} \frac{1}{a^rb^s} \cdot \frac{1}{c^t}\\
&= \sum_{i+j=r+s} \Delta^j_{r,s} \sum_{i_1+j_1=i+t} \Delta^{j_1}_{i,t} \Bigg(S_d(i_1,j_1,j)+S_d(i_1,j,j_1) + \sum \limits_{i_2 + j_2 = j+j_1 } \Delta_{j,j_1}^{j_2} S_d(i_1,i_2,j_2)\Bigg),
\end{align*}
which is the sum of the terms of depth $3$ in the expression of $(S_d(r)S_d(s))S_d(t)$. Similarly, one verifies easily that
\begin{align*}
&\sum \limits_{(a,b,c)\in N_1} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t} + \sum \limits_{(a,b,c)\in N_2} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t} + \sum \limits_{(a,b,c)\in N_3} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}+ \sum \limits_{(a,b,c)\in N_4} \frac{1}{a^r} \cdot \frac{1}{b^s c^ t}\\
&= \sum_{i+j=s+t} \Delta^j_{s,t} \sum_{i_1+j_1=i+r} \Delta^{j_1}_{i,r} \Bigg(S_d(i_1,j_1,j)+S_d(i_1,j,j_1) + \sum \limits_{i_2 + j_2 = j+j_1 } \Delta_{j,j_1}^{j_2} S_d(i_1,i_2,j_2)\Bigg),
\end{align*}
which is the sum of the terms of depth $3$ in the expression of $S_d(r)(S_d(s)S_d(t))$. This completes the proof.
\subsection{Expansions for $S_{<d}$ of depth one} \label{sec: 4} ${}$\par
For two positive integers $r,s$ and for all $d \in \mathbb{N}$, we first recall the following formula:
\begin{equation} \label{eq: formula}
S_{<d}(r)S_{<d}(s) = \sum \limits_{i<d} S_i(r)S_i(s) + S_{<d}(r,s) + S_{<d}(s,r).
\end{equation}
Let $r,s,t$ be positive integers. In this section, we give the expansions of $(S_{<d}(r)S_{<d}(s))S_{<d}(t)$ and $S_{<d}(r)(S_{<d}(s)S_{<d}(t))$ by using Formula \eqref{eq: formula}.
\subsubsection{} We expand $(S_{<d}(r)S_{<d}(s))S_{<d}(t)$ as follows. We have;
\begin{align*}
&(S_{<d}(r)S_{<d}(s))S_{<d}(t)\\
&= \Bigg(\sum \limits_{i<d} S_i(r)S_i(s) + S_{<d}(r,s) + S_{<d}(s,r)\Bigg)S_{<d}(t)\\
&=\sum \limits_{i<d} S_i(r)S_i(s)S_{<d}(t) + S_{<d}(r,s)S_{<d}(t) + S_{<d}(s,r)S_{<d}(t)\\
&=\sum \limits_{i<d} S_i(r)S_i(s) \sum \limits_{i<d} S_i(t) + \sum \limits_{i<d} S_i(r,s)\sum \limits_{i<d} S_i(t)+ \sum \limits_{i<d} S_i(s,r)\sum \limits_{i<d} S_i(t)\\
&=\Bigg(\sum\limits_{i<d} (S_i(r)S_i(s))S_i(t) + \sum\limits_{i<d} (S_i(r)S_i(s))S_{<i}(t) + \sum\limits_{i<d} S_i(t) \sum \limits_{j<i} S_j(r)S_j(s)\Bigg) \\
&+ \Bigg(\sum\limits_{i<d}S_i(r,s)S_i(t) + \sum\limits_{i<d}S_i(r,s)S_{<i}(t) + \sum\limits_{i<d} S_i(t)S_{<i}(r,s)\Bigg)\\
&+ \Bigg(\sum\limits_{i<d}S_i(s,r)S_i(t) + \sum\limits_{i<d}S_i(s,r)S_{<i}(t) + \sum\limits_{i<d} S_i(t)S_{<i}(s,r)\Bigg)\\
&= \sum\limits_{i<d} (S_i(r)S_i(s))S_i(t) + \sum\limits_{i<d} (S_i(r)S_i(s))S_{<i}(t) + \sum\limits_{i<d}S_i(r,s)S_i(t) + \sum\limits_{i<d}S_i(r,s)S_{<i}(t)\\
&+ \sum\limits_{i<d}S_i(s,r)S_i(t) + \sum\limits_{i<d}S_i(s,r)S_{<i}(t) + \sum\limits_{i<d} S_i(t)\Bigg(\sum \limits_{j<i} S_j(r)S_j(s) + S_{<i}(r,s) + S_{<i}(s,r)\Bigg).
\end{align*}
\subsubsection{} We expand $S_{<d}(r)(S_{<d}(s)S_{<d}(t))$ as follows. We have;
\begin{align*}
&S_{<d}(r)(S_{<d}(s)S_{<d}(t))\\
&= S_{<d}(r) \Bigg(\sum \limits_{i<d} S_i(s)S_i(t) + S_{<d}(s,t) + S_{<d}(t,s)\Bigg)\\
&= S_{<d}(r) \sum \limits_{i<d} S_i(s)S_i(t) + S_{<d}(r)S_{<d}(s,t) + S_{<d}(r)S_{<d}(t,s)\\
&= \sum\limits_{i<d}S_{i}(r)\sum \limits_{i<d} S_i(s)S_i(t) + \sum\limits_{i<d}S_{i}(r)\sum \limits_{i<d} S_i(s,t) + \sum\limits_{i<d}S_{i}(r)\sum \limits_{i<d} S_i(t,s)\\
&= \Bigg(\sum\limits_{i<d}S_{i}(r)(S_i(s)S_i(t)) + \sum\limits_{i<d}S_{i}(r)\sum\limits_{j<i}S_j(s)S_j(t) + \sum\limits_{i<d}(S_i(s)S_i(t))S_{<i}(r)\Bigg)\\
&+ \Bigg(\sum\limits_{i<d}S_{i}(r)S_i(s,t) + \sum\limits_{i<d}S_{i}(r)S_{<i}(s,t) + \sum\limits_{i<d}S_i(s,t)S_{<i}(r)\Bigg)\\
&+ \Bigg(\sum\limits_{i<d}S_{i}(r)S_i(t,s) + \sum\limits_{i<d}S_{i}(r)S_{<i}(t,s) + \sum\limits_{i<d}S_i(t,s)S_{<i}(r)\Bigg)\\
&= \sum\limits_{i<d}S_{i}(r)(S_i(s)S_i(t)) + \sum\limits_{i<d}(S_i(s)S_i(t))S_{<i}(r) + \sum\limits_{i<d}S_{i}(r)S_i(s,t) + \sum\limits_{i<d}S_i(s,t)S_{<i}(r)\\
&+ \sum\limits_{i<d}S_{i}(r)S_i(t,s) + \sum\limits_{i<d}S_i(t,s)S_{<i}(r) + \sum\limits_{i<d}S_{i}(r) \Bigg(\sum\limits_{j<i}S_j(s)S_j(t) + S_{<i}(s,t) + S_{<i}(t,s)\Bigg).
\end{align*}
\subsection{Associativity for $S_{<d}$ of depth one} \label{sec: 5}
\begin{theorem} \label{thm: assoc S_<d depth one}
Let $r,s , t$ be positive integers. For all $d \in \mathbb{N}$, the expansions using Chen's formula and Formula \eqref{eq: formula} of $(S_{<d}(r)S_{<d}(s))S_{<d}(t)$ and $S_{<d}(r)(S_{<d}(s)S_{<d}(t))$ yield the same expression in terms of $S_{<d}$.
\end{theorem}
\begin{proof}
To prove the desired associativity, we compare the expansions of $(S_{<d}(r)S_{<d}(s))S_{<d}(t)$ and $S_{<d}(r)(S_{<d}(s)S_{<d}(t))$ in Section \ref{sec: 4}. From Theorem \ref{thm: assoc S_d depth 1}, it is obvious that the expansions of
\begin{equation*}
\sum\limits_{i<d} (S_i(r)S_i(s))S_i(t) \quad \text{and} \quad \sum\limits_{i<d}S_{i}(r)(S_i(s)S_i(t))
\end{equation*}
yield the same expression in terms of $S_{<d}$. From Formula \eqref{eq: formula}, one verifies easily that the expansions of
\begin{equation*}
\sum\limits_{i<d} S_i(t)\Bigg(\sum \limits_{j<i} S_j(r)S_j(s) + S_{<i}(r,s) + S_{<i}(s,r)\Bigg) \quad \text{and} \quad \sum\limits_{i<d}S_i(t,s)S_{<i}(r)
\end{equation*}
yield the same expression in terms of $S_{<d}$. Similarly, one deduces that the expansions of
\begin{equation*}
\sum\limits_{i<d}S_i(r,s)S_{<i}(t) \quad \text{and} \quad \sum\limits_{i<d}S_{i}(r) \Bigg(\sum\limits_{j<i}S_j(s)S_j(t) + S_{<i}(s,t) + S_{<i}(t,s)\Bigg)
\end{equation*}
yield the same expression in terms of $S_{<d}$. From a straightforward verification, one deduces that the expansions of
\begin{equation*}
\sum\limits_{i<d} (S_i(r)S_i(s))S_{<i}(t), \sum\limits_{i<d}S_i(r,s)S_i(t),\sum\limits_{i<d}S_i(s,r)S_i(t),\sum\limits_{i<d}S_i(s,r)S_{<i}(t)
\end{equation*}
yield respectively the same expressions in terms of $S_{<d}$ as those of
\begin{equation*}
\sum\limits_{i<d} S_i(r)S_i(s,t),\sum\limits_{i<d}S_i(r)S_i(t,s),\sum\limits_{i<d}(S_i(s)S_i(t))S_{<i}(r),\sum\limits_{i<d}S_i(s,t)S_{<i}(r).
\end{equation*}
This proves the theorem.
\end{proof}
\subsection{Expansions of arbitrary depth} \label{sec: 6} ${}$\par
We now extend our results for the case of arbitrary depth. Let $\mathfrak{a}=(a_1,\dotsc, a_m)$ and $\fb=(b_1,\dotsc, b_n)$ be two positive tuples. For simplicity, we set $\mathfrak{a}_- = (a_2,\dotsc, a_m)$ and $\fb_- = (b_2,\dotsc, b_n)$. For all $d \in \mathbb{N}$, we recall the following formulas (See \cite{ND21}):
\begin{align}
S_d(\mathfrak{a}) S_d(\fb) &= (S_d(a_1)S_d(b_1))(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))\label{eq: SdSd}\\
&=S_d(a_1+b_1)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-)) \notag\\
&+ \sum \limits_{i+j = a_1+b_1}\Delta_{a_1,b_1}^{j}S_d(i)[S_{<d}(j)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))], \notag\\
S_d(\mathfrak{a})S_{<d}(\fb) &= S_{<d}(\fb)S_d(\mathfrak{a}) = S_d(a_1)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb)),\label{eq: SdS<d}\\
S_{<d}(\mathfrak{a}) S_{<d}(\fb) &= \sum \limits_{i<d}S_i(\mathfrak{a})S_i(\fb) + \sum \limits_{i<d}S_i(\mathfrak{a})S_{<i}(\fb) + \sum \limits_{i<d}S_i(\fb)S_{<i}(\mathfrak{a}).\label{eq: S<dS<d}
\end{align}
Let $\mathfrak{a},\fb,\fc$ be positive tuples. We first give the expansions of $(S_{d}(\mathfrak{a})S_{d}(\fb))S_{d}(\fc)$ and $S_{d}(\mathfrak{a})(S_{d}(\fb)S_{d}(\fc))$ by using Formula \eqref{eq: SdSd}.
\subsubsection{} We expand $(S_{d}(\mathfrak{a})S_{d}(\fb))S_{d}(\fc)$ as follows. We have;
\begin{align} \label{eq: S_d L-R}
&(S_{d}(\mathfrak{a})S_{d}(\fb))S_{d}(\fc)\\
&= \Bigg(S_d(a_1+b_1)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))+ \sum \limits_{i+j = a_1+b_1}\Delta_{a_1,b_1}^{j}S_d(i)[S_{<d}(j)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))]\Bigg)S_{d}(\fc)\notag\\
&= S_d(a_1+b_1)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))S_{d}(\fc) + \sum \limits_{i+j = a_1+b_1}\Delta_{a_1,b_1}^{j}S_d(i)[S_{<d}(j)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))]S_{d}(\fc)\notag\\
&= S_d(a_1+b_1)S_d(c_1)[(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))S_{<d}(\fc_-)]\notag\\
&+ \sum \limits_{i+j = a_1+b_1}\Delta_{a_1,b_1}^{j}(S_d(i)S_d(c_1))[S_{<d}(j)(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))]S_{<d}(\fc_-).\notag
\end{align}
\subsubsection{} We expand $S_{d}(\mathfrak{a})(S_{d}(\fb)S_{d}(\fc))$ as follows. We have;
\begin{align}\label{eq: S_d R-L}
&S_{d}(\mathfrak{a})(S_{d}(\fb)S_{d}(\fc))\\
&= S_{d}(\mathfrak{a})\Bigg(S_d(b_1+c_1)(S_{<d}(\fb_-)S_{<d}(\fc_-))+ \sum \limits_{i+j = b_1+c_1}\Delta_{b_1,c_1}^{j}S_d(i)[S_{<d}(j)(S_{<d}(\fb_-)S_{<d}(\fc_-))]\Bigg)\notag\\
&= S_{d}(\mathfrak{a})S_d(b_1+c_1)(S_{<d}(\fb_-)S_{<d}(\fc_-))+ S_{d}(\mathfrak{a})\sum \limits_{i+j = b_1+c_1}\Delta_{b_1,c_1}^{j}S_d(i)[S_{<d}(j)(S_{<d}(\fb_-)S_{<d}(\fc_-))]\notag\\
&= (S_{d}(a_1)S_d(b_1+c_1))[S_{<d}(\mathfrak{a}_-)(S_{<d}(\fb_-)S_{<d}(\fc_-))]\notag\\
&+ \sum \limits_{i+j = b_1+c_1}\Delta_{b_1,c_1}^{j}(S_d(a_1)S_d(i))S_{<d}(\mathfrak{a}_-)[S_{<d}(j)(S_{<d}(\fb_-)S_{<d}(\fc_-))].\notag
\end{align}
We next give the expansions of $(S_{<d}(\mathfrak{a})S_{<d}(\fb))S_{<d}(\fc)$ and $S_{<d}(\mathfrak{a})(S_{<d}(\fb)S_{<d}(\fc))$ by using Formula \eqref{eq: S<dS<d}.
\subsubsection{} We expand $(S_{<d}(\mathfrak{a})S_{<d}(\fb))S_{<d}(\fc)$ as follows. We have;
\begin{align}\label{eq: S<d L-R}
&(S_{<d}(\mathfrak{a})S_{<d}(\fb))S_{<d}(\fc)\\
&= \Bigg(\sum \limits_{i<d}S_i(\mathfrak{a})S_i(\fb) + \sum \limits_{i<d}S_i(\mathfrak{a})S_{<i}(\fb) + \sum \limits_{i<d}S_i(\fb)S_{<i}(\mathfrak{a})\Bigg)S_{<d}(\fc)\notag\\
&= \sum \limits_{i<d}S_i(\mathfrak{a})S_i(\fb)S_{<d}(\fc) + \sum \limits_{i<d}S_i(\mathfrak{a})S_{<i}(\fb)S_{<d}(\fc) + \sum \limits_{i<d}S_i(\fb)S_{<i}(\mathfrak{a})S_{<d}(\fc)\notag\\
&= \sum \limits_{i<d}S_i(\mathfrak{a})S_i(\fb)\sum\limits_{i<d}S_{i}(\fc) + \sum \limits_{i<d}S_i(\mathfrak{a})S_{<i}(\fb)\sum\limits_{i<d}S_{i}(\fc) + \sum \limits_{i<d}S_i(\fb)S_{<i}(\mathfrak{a})\sum\limits_{i<d}S_{i}(\fc)\notag\\
&= \Bigg(\sum \limits_{i<d}(S_i(\mathfrak{a})S_i(\fb))S_i(\fc) + \sum \limits_{i<d}(S_i(\mathfrak{a})S_i(\fb))S_{<i}(\fc) + \sum \limits_{i<d}S_i(\fc)\sum \limits_{j<i}S_j(\mathfrak{a})S_j(\fb) \Bigg)\notag\\
&+ \Bigg(\sum \limits_{i<d}(S_i(\mathfrak{a})S_{<i}(\fb))S_i(\fc) + \sum \limits_{i<d}(S_i(\mathfrak{a})S_{<i}(\fb))S_{<i}(\fc) + \sum \limits_{i<d}S_i(\fc)\sum \limits_{j<i}S_j(\mathfrak{a})S_{<j}(\fb)\Bigg) \notag\\
&+ \Bigg(\sum \limits_{i<d}(S_i(\fb)S_{<i}(\mathfrak{a}))S_i(\fc) + \sum \limits_{i<d}(S_i(\fb)S_{<i}(\mathfrak{a}))S_{<i}(\fc) + \sum \limits_{i<d}S_i(\fc)\sum \limits_{j<i}S_j(\fb)S_{<j}(\mathfrak{a})\Bigg) \notag\\
&=\sum \limits_{i<d}(S_i(\mathfrak{a})S_i(\fb))S_i(\fc) + \sum \limits_{i<d}(S_i(\mathfrak{a})S_i(\fb))S_{<i}(\fc) + \sum \limits_{i<d}(S_i(\mathfrak{a})S_{<i}(\fb))S_i(\fc)\notag\\
& + \sum \limits_{i<d}(S_i(\mathfrak{a})S_{<i}(\fb))S_{<i}(\fc) + \sum \limits_{i<d}(S_i(\fb)S_{<i}(\mathfrak{a}))S_i(\fc) + \sum \limits_{i<d}(S_i(\fb)S_{<i}(\mathfrak{a}))S_{<i}(\fc) \notag\\
&+ \sum \limits_{i<d}S_i(\fc)\Bigg(\sum \limits_{j<i}S_j(\mathfrak{a})S_j(\fb) + \sum \limits_{j<i}S_j(\mathfrak{a})S_{<j}(\fb) + \sum \limits_{j<i}S_j(\fb)S_{<j}(\mathfrak{a}) \Bigg).\notag
\end{align}
\subsubsection{} We expand $S_{<d}(\mathfrak{a})(S_{<d}(\fb)S_{<d}(\fc))$ as follows. We have;
\begin{align}\label{eq: S<d R-L}
&S_{<d}(\mathfrak{a})(S_{<d}(\fb)S_{<d}(\fc))\\
&= S_{<d}(\mathfrak{a})\Bigg(\sum \limits_{i<d}S_i(\fb)S_i(\fc) + \sum \limits_{i<d}S_i(\fb)S_{<i}(\fc) + \sum \limits_{i<d}S_i(\fc)S_{<i}(\fb)\Bigg) \notag\\
&= S_{<d}(\mathfrak{a})\sum \limits_{i<d}S_i(\fb)S_i(\fc) + S_{<d}(\mathfrak{a})\sum \limits_{i<d}S_i(\fb)S_{<i}(\fc) + S_{<d}(\mathfrak{a})\sum \limits_{i<d}S_i(\fc)S_{<i}(\fb) \notag\\
&= \sum \limits_{i<d}S_{i}(\mathfrak{a})\sum \limits_{i<d}S_i(\fb)S_i(\fc) + \sum \limits_{i<d}S_{i}(\mathfrak{a})\sum \limits_{i<d}S_i(\fb)S_{<i}(\fc) + \sum \limits_{i<d}S_{i}(\mathfrak{a})\sum \limits_{i<d}S_i(\fc)S_{<i}(\fb) \notag\\
&= \Bigg(\sum \limits_{i<d}S_{i}(\mathfrak{a})(S_i(\fb)S_i(\fc)) + \sum \limits_{i<d}S_{i}(\mathfrak{a}) \sum \limits_{j<i}S_j(\fb)S_j(\fc) + \sum \limits_{i<d}(S_i(\fb)S_i(\fc))S_{<i}(\mathfrak{a})\Bigg) \notag\\
&+ \Bigg(\sum \limits_{i<d}S_{i}(\mathfrak{a})(S_i(\fb)S_{<i}(\fc)) + \sum \limits_{i<d}S_{i}(\mathfrak{a}) \sum \limits_{j<i}S_j(\fb)S_{<j}(\fc) + \sum \limits_{i<d}(S_i(\fb)S_{<i}(\fc))S_{<i}(\mathfrak{a})\Bigg)\notag\\
&+ \Bigg(\sum \limits_{i<d}S_{i}(\mathfrak{a})(S_i(\fc)S_{<i}(\fb)) + \sum \limits_{i<d}S_{i}(\mathfrak{a}) \sum \limits_{j<i}S_j(\fc)S_{<j}(\fb) + \sum \limits_{i<d}(S_i(\fc)S_{<i}(\fb))S_{<i}(\mathfrak{a})\Bigg)\notag\\
&=\sum \limits_{i<d}S_{i}(\mathfrak{a})(S_i(\fb)S_i(\fc)) + \sum \limits_{i<d}(S_i(\fb)S_i(\fc))S_{<i}(\mathfrak{a}) +
\sum \limits_{i<d}S_{i}(\mathfrak{a})(S_i(\fb)S_{<i}(\fc))\notag\\
&+ \sum \limits_{i<d}(S_i(\fb)S_{<i}(\fc))S_{<i}(\mathfrak{a}) + \sum \limits_{i<d}S_{i}(\mathfrak{a})(S_i(\fc)S_{<i}(\fb)) + \sum \limits_{i<d}(S_i(\fc)S_{<i}(\fb))S_{<i}(\mathfrak{a}) \notag\\
&+ \sum \limits_{i<d}S_{i}(\mathfrak{a}) \Bigg(\sum \limits_{j<i}S_j(\fb)S_j(\fc) + \sum \limits_{j<i}S_j(\fb)S_{<j}(\fc) + \sum \limits_{j<i}S_j(\fc)S_{<j}(\fb)\Bigg). \notag
\end{align}
\subsection{Associativity of arbitrary depth} \label{sec: 7} ${}$\par
\begin{theorem} \label{thm: assoc}
Let $\mathfrak{a},\fb,\fc$ be positive tuples.
\begin{enumerate}[$(1)$]
\item For all $d \in \mathbb{N}$, the expansions using \eqref{eq: SdSd}, \eqref{eq: SdS<d}, \eqref{eq: S<dS<d} of $(S_d(\mathfrak{a}) S_d(\fb))S_d(\fc)$ and $S_d(\mathfrak{a}) (S_d(\fb)S_d(\fc))$ yield the same expression in terms of power sums.
\item For all $d \in \mathbb{N}$, the expansions using \eqref{eq: SdSd}, \eqref{eq: SdS<d}, \eqref{eq: S<dS<d} of $(S_{<d}(\mathfrak{a}) S_{<d}(\fb))S_{<d}(\fc)$ and $S_{<d}(\mathfrak{a}) (S_{<d}(\fb)S_{<d}(\fc))$ yield the same expression in terms of $S_{<d}$.
\end{enumerate}
\end{theorem}
\begin{proof}
We proceed the proof by induction on $\depth(\mathfrak{a}) + \depth(\fb) + \depth(\fc)$. The base step $\depth(\mathfrak{a}) + \depth(\fb) + \depth(\fc)= 3$, i.e., $\depth(\mathfrak{a}) = \depth(\fb) = \depth(\fc)= 1$ of Theorem \ref{thm: assoc} follows from Theorem \ref{thm: assoc S_d depth 1} and Theorem \ref{thm: assoc S_<d depth one}.
Assume that Theorem \ref{thm: assoc} holds when $\depth(\mathfrak{a}) + \depth(\fb) + \depth(\fc) < n$ with $n \in \mathbb{N}$ and $n \geq 4$. We need to show that Theorem \ref{thm: assoc} holds when $\depth(\mathfrak{a}) + \depth(\fb) + \depth(\fc) = n$.
In order to prove Part (1), we apply the induction hypothesis on Expansions \eqref{eq: S_d L-R} and \eqref{eq: S_d R-L}. We deduce that the expansion of $(S_d(\mathfrak{a}) S_d(\fb))S_d(\fc)$ yields the same expression in terms of power sums as that of
\begin{align*}
&\Bigg[\Bigg(S_d(a_1 + b_1) + \sum \limits_{i+j = a_1 + b_1} \Delta^j_{a_1,b_1}S_d(i,j)\Bigg)S_d(c_1)\Bigg][(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))S_{<d}(\fc_-)]\\
&= [(S_d(a_1)S_d(b_1))S_d(c_1)][(S_{<d}(\mathfrak{a}_-)S_{<d}(\fb_-))S_{<d}(\fc_-)],
\end{align*}
and the expansion of $S_d(\mathfrak{a}) (S_d(\fb)S_d(\fc))$ yields the same expression in terms of power sums as that of
\begin{align*}
&\Bigg[S_d(a_1)\Bigg(S_d(b_1 + c_1) + \sum \limits_{i+j = b_1 + c_1} \Delta^j_{b_1,c_1}S_d(i,j)\Bigg)\Bigg][S_{<d}(\mathfrak{a}_-)(S_{<d}(\fb_-)S_{<d}(\fc_-))]\\
&=[S_d(a_1)(S_d(b_1)S_d(c_1))][S_{<d}(\mathfrak{a}_-)(S_{<d}(\fb_-)S_{<d}(\fc_-))].
\end{align*}
Using Theorem \ref{thm: assoc S_d depth 1} and the induction hypothesis again, we conclude that expansions of $(S_d(\mathfrak{a}) S_d(\fb))S_d(\fc)$ and $S_d(\mathfrak{a}) (S_d(\fb)S_d(\fc))$ yield the same expression in terms of power sums.
In order to prove Part (2), we compare Expansions \eqref{eq: S<d L-R} and \eqref{eq: S<d R-L}. From Part (1), it is obvious that the expansions of
\begin{equation*}
\sum \limits_{i<d}(S_i(\mathfrak{a})S_i(\fb))S_i(\fc)\quad \text{and} \quad \sum \limits_{i<d}S_i(\mathfrak{a})(S_i(\fb)S_i(\fc))
\end{equation*}
yield the same expression in terms of $S_{<d}$.
Note that from Formulas \eqref{eq: SdS<d} and \eqref{eq: S<dS<d}, we have;
\begin{align*}
&S_i(\fc)\Bigg(\sum \limits_{j<i}S_j(\mathfrak{a})S_j(\fb) + \sum \limits_{j<i}S_j(\mathfrak{a})S_{<j}(\fb) + \sum \limits_{j<i}S_j(\fb)S_{<j}(\mathfrak{a}) \Bigg) \\
&= S_i(\fc)(S_{<i}(\fb)S_{<i}(\mathfrak{a}))\\
&= S_i(c_1)[S_{<i}(\fc_-)(S_{<i}(\fb)S_{<i}(\mathfrak{a}))]
\end{align*}
and
\begin{align*}
(S_i(\fc)S_{<i}(\fb))S_{<i}(\mathfrak{a}) = S_i(c_1) [(S_{<i}(\fc_-)S_{<i}(\fb))S_{<i}(\mathfrak{a})].
\end{align*}
One then deduces from the induction hypothesis that the expansions of
\begin{align*}
\sum \limits_{i<d}S_i(\fc)\Bigg(\sum \limits_{j<i}S_j(\mathfrak{a})S_j(\fb) + \sum \limits_{j<i}S_j(\mathfrak{a})S_{<j}(\fb) + \sum \limits_{j<i}S_j(\fb)S_{<j}(\mathfrak{a}) \Bigg)
\end{align*}
and
\begin{align*}
\sum \limits_{i<d} (S_i(\fc)S_{<i}(\fb))S_{<i}(\mathfrak{a})
\end{align*}
yield the same expression in terms of $S_{<d}$. Similarly, one verifies easily that the expansions of
\begin{align*}
\sum \limits_{i<d}(S_i(\mathfrak{a})S_{<i}(\fb))S_{<i}(\fc)
\end{align*}
and
\begin{align*}
\sum \limits_{i<d}S_{i}(\mathfrak{a}) \Bigg(\sum \limits_{j<i}S_j(\fb)S_j(\fc) + \sum \limits_{j<i}S_j(\fb)S_{<j}(\fc) + \sum \limits_{j<i}S_j(\fc)S_{<j}(\fb)\Bigg)
\end{align*}
yield the same expression in terms of $S_{<d}$.
Note that from Formulas \eqref{eq: SdSd} and \eqref{eq: SdS<d}, we have;
\begin{align*}
(S_i(\mathfrak{a})S_i(\fb))S_{<i}(\fc) = (S_i(a_1)S_i(b_1))[(S_{<i}(\mathfrak{a}_-)S_{<i}(\fb_-))S_{<i}(\fc)]
\end{align*}
and
\begin{align*}
S_i(\mathfrak{a})(S_i(\fb)S_{<i}(\fc)) = (S_i(a_1)S_i(b_1))[S_{<i}(\mathfrak{a}_-)(S_{<i}(\fb_-)S_{<i}(\fc))].
\end{align*}
One then deduces from the induction hypothesis that the expansions of
\begin{align*}
\sum \limits_{i<d}(S_i(\mathfrak{a})S_i(\fb))S_{<i}(\fc) \quad \text{and} \quad \sum \limits_{i<d}S_i(\mathfrak{a})(S_i(\fb)S_{<i}(\fc))
\end{align*}
yield the same expression in terms of $S_{<d}$. Similarly, one verifies easily that the expansions of
\begin{align*}
\sum \limits_{i<d}(S_i(\mathfrak{a})S_{<i}(\fb))S_i(\fc), \sum \limits_{i<d}(S_i(\fb)S_{<i}(\mathfrak{a}))S_i(\fc), \sum \limits_{i<d}(S_i(\fb)S_{<i}(\mathfrak{a}))S_{<i}(\fc)
\end{align*}
yield respectively the same expressions in terms of $S_{<d}$ as those of
\begin{align*}
\sum \limits_{i<d}S_i(\mathfrak{a})(S_i(\fc)S_{<i}(\fb)), \sum \limits_{i<d}(S_i(\fb)S_i(\fc))S_{<i}(\mathfrak{a}), \sum \limits_{i<d}(S_i(\fb)S_{<i}(\fc))S_{<i}(\mathfrak{a}).
\end{align*}
From the above arguments, we conclude that the expansions of $(S_{<d}(\mathfrak{a}) S_{<d}(\fb))S_{<d}(\fc)$ and $S_{<d}(\mathfrak{a}) (S_{<d}(\fb)S_{<d}(\fc))$ yield the same expression in terms of $S_{<d}$. This completes the proof.
\end{proof}
\subsection{Associativity of the shuffle algebra}
\begin{proposition} \label{prop: associative}
The diamond product and the shuffle product are associative.
\end{proposition}
\begin{proof}
Let $\mathfrak{a}, \fb, \fc \in \langle \Sigma \rangle$ be arbitrary words. It is suffices to show that
\begin{equation} \label{eq: assocative}
(\mathfrak{a} \diamond \fb) \diamond \fc = \mathfrak{a} \diamond (\fb \diamond \fc) \quad \text{and} \quad (\mathfrak{a} \shuffle \fb) \shuffle \fc = \mathfrak{a} \shuffle (\fb \shuffle \fc) .
\end{equation}
We proceed the proof by induction on $\depth(\mathfrak{a}) + \depth(\fb) + \depth(\fc)$. If one of $\mathfrak{a}, \fb$ or $\fc$ is empty word, then \eqref{eq: assocative} holds trivially. We assume that \eqref{eq: assocative} holds when $\depth(\mathfrak{a}) + \depth(\fb) + \depth(\fc) < n$ with $n \in \ensuremath \mathbb{N}$ and $n \geq 3$. We need to show that \eqref{eq: assocative} holds when $\depth(\mathfrak{a}) + \depth(\fb) + \depth(\fc) = n$.
We first show that $(\mathfrak{a} \diamond \fb) \diamond \fc = \mathfrak{a} \diamond (\fb \diamond \fc)$. From Lemma \ref{lem: triangle formulas}, we have
\begin{align*}
& (\mathfrak{a} \diamond \fb) \diamond \fc \\
&= \Bigg[x_{a + b}(\mathfrak{a}_-\shuffle \fb_-) + \sum\limits_{i+j = a + b} \Delta^j_{a,b} x_i(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-))\Bigg] \diamond \fc\\
&= \ x_{a + b}(\mathfrak{a}_-\shuffle \fb_-) \diamond \fc + \sum\limits_{i+j = a + b} \Delta^j_{a,b} x_i(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)) \diamond \fc\\
&= \ (x_{a + b} \diamond x_c) \triangleright ((\mathfrak{a}_-\shuffle \fb_-) \shuffle \fc_-) + \sum\limits_{i+j = a + b} \Delta^j_{a,b} (x_i \diamond x_c) \triangleright [(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)) \shuffle \fc_-].
\end{align*}
For all $i,j \in \ensuremath \mathbb{N}$ with $i+j = a+b$, it follows from the induction hypothesis that
\begin{align*}
&(x_i \diamond x_c) \triangleright [(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)) \shuffle \fc_-]\\ &= \ (x_i \diamond x_c) \triangleright [x_j \shuffle ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)]\\
&= \Bigg[x_{i + c} + \sum\limits_{i_1+j_1 = i + c} \Delta^{j_1}_{i,c} x_{i_1}x_{j_1}\Bigg] \triangleright [x_j \shuffle ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)]\\
&= \ x_{i + c}\triangleright [x_j \shuffle ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)] + \sum\limits_{i_1+j_1 = i + c} \Delta^{j_1}_{i,c} x_{i_1}x_{j_1}\triangleright [x_j \shuffle ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)]\\
&= \ x_{i + c} [x_j \shuffle ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)] + \sum\limits_{i_1+j_1 = i + c} \Delta^{j_1}_{i,c} x_{i_1}[(x_{j_1} \shuffle x_j) \shuffle ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)]\\
&= \ x_{i + c}x_j \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-) + \sum\limits_{i_1+j_1 = i + c} \Delta^{j_1}_{i,c} x_{i_1}(x_{j_1} \shuffle x_j) \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)\\
&= \Bigg[ x_{i + c}x_j + \sum\limits_{i_1+j_1 = i + c} \Delta^{j_1}_{i,c} x_{i_1}(x_{j_1} \shuffle x_j)\Bigg] \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)\\
&= (x_ix_j \diamond x_c) \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-).
\end{align*}
Thus
\begin{align*}
(\mathfrak{a} \diamond \fb) \diamond \fc &= \ (x_{a + b} \diamond x_c) \triangleright ((\mathfrak{a}_-\shuffle \fb_-) \shuffle \fc_-) + \sum\limits_{i+j = a + b} \Delta^j_{a,b} (x_ix_j \diamond x_c) \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)\\
&= \Bigg[(x_{a + b} \diamond x_c) + \sum\limits_{i+j = a + b} \Delta^j_{a,b} (x_ix_j \diamond x_c) \Bigg]\triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-)\\
&= \ ((x_a \diamond x_b) \diamond x_c) \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_-).
\end{align*}
On the other hand, from Proposition \ref{prop: commutative} and the above arguments, we deduce that
\begin{align*}
\mathfrak{a} \diamond (\fb \diamond \fc) &= (\fb \diamond \fc) \diamond \mathfrak{a} \\
&= ((x_b \diamond x_c) \diamond x_a) \triangleright ((\fb_- \shuffle \fc_-) \shuffle \mathfrak{a}_-) \\
&= (x_a \diamond (x_b \diamond x_c) ) \triangleright (\mathfrak{a}_- \shuffle (\fb_- \shuffle \fc_-)).
\end{align*}
It is straightforward from Theorem \ref{thm: assoc S_d depth 1} that $(x_a \diamond x_b) \diamond x_c = x_a \diamond (x_b \diamond x_c)$. Moreover, it follows from the induction hypothesis that $(\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc_- = \mathfrak{a}_- \shuffle (\fb_- \shuffle \fc_-)$. We thus conclude that $(\mathfrak{a} \diamond \fb) \diamond \fc = \mathfrak{a} \diamond (\fb \diamond \fc)$.
We next show that $(\mathfrak{a} \shuffle \fb) \shuffle \fc = \mathfrak{a} \shuffle (\fb \shuffle \fc)$. From Lemma \ref{lem: triangle formulas}, we have;
\begin{align*}
(\mathfrak{a} \shuffle \fb) \shuffle \fc
&= \ (\mathfrak{a} \triangleright \fb + \fb \triangleright \mathfrak{a} + \mathfrak{a} \diamond \fb) \shuffle \fc\\
&= \ (\mathfrak{a} \triangleright \fb) \shuffle \fc + (\fb \triangleright \mathfrak{a}) \shuffle \fc + (\mathfrak{a} \diamond \fb) \shuffle \fc\\
&= \ ((\mathfrak{a} \triangleright \fb) \triangleright \fc + \fc \triangleright (\mathfrak{a} \triangleright \fb) + (\mathfrak{a} \triangleright \fb) \diamond \fc)\\
&+ ((\fb \triangleright \mathfrak{a}) \triangleright \fc + \fc \triangleright (\fb \triangleright \mathfrak{a}) + (\fb \triangleright \mathfrak{a}) \diamond \fc)\\
&+ ((\mathfrak{a} \diamond \fb) \triangleright \fc + \fc \triangleright (\mathfrak{a} \diamond \fb) + (\mathfrak{a} \diamond \fb) \diamond \fc)\\
&= \ (\mathfrak{a} \triangleright \fb) \triangleright \fc + (\mathfrak{a} \triangleright \fb) \diamond \fc + (\fb \triangleright \mathfrak{a}) \triangleright \fc + (\fb \triangleright \mathfrak{a}) \diamond \fc + (\mathfrak{a} \diamond \fb) \triangleright \fc\\
&+ (\fc \triangleright (\mathfrak{a} \triangleright \fb) + \fc \triangleright (\fb \triangleright \mathfrak{a}) + \fc \triangleright (\mathfrak{a} \diamond \fb)) + (\mathfrak{a} \diamond \fb) \diamond \fc\\
&= \ (\mathfrak{a} \triangleright \fb) \triangleright \fc + (\mathfrak{a} \triangleright \fb) \diamond \fc + (\fb \triangleright \mathfrak{a}) \triangleright \fc + (\fb \triangleright \mathfrak{a}) \diamond \fc + (\mathfrak{a} \diamond \fb) \triangleright \fc\\
&+ \fc \triangleright (\mathfrak{a} \shuffle \fb) + (\mathfrak{a} \diamond \fb) \diamond \fc
\end{align*}
and
\begin{align*}
\mathfrak{a} \shuffle (\fb \shuffle \fc)
&= \ \mathfrak{a} \shuffle (\fb \triangleright \fc + \fc \triangleright \fb + \fb \diamond \fc)\\
&= \ \mathfrak{a} \shuffle (\fb \triangleright \fc) + \mathfrak{a} \shuffle (\fc \triangleright \fb) + \mathfrak{a} \shuffle (\fb \diamond \fc)\\
&= \ (\mathfrak{a} \triangleright (\fb \triangleright \fc) + (\fb \triangleright \fc) \triangleright \mathfrak{a} + \mathfrak{a} \diamond (\fb \triangleright \fc))\\
&+ (\mathfrak{a} \triangleright (\fc \triangleright \fb) + (\fc \triangleright \fb) \triangleright \mathfrak{a} + \mathfrak{a} \diamond (\fc \triangleright \fb)) \\
&+ (\mathfrak{a} \triangleright (\fb \diamond \fc) + (\fb \diamond \fc) \triangleright \mathfrak{a} + \mathfrak{a} \diamond (\fb \diamond \fc))\\
&= \ (\fb \triangleright \fc) \triangleright \mathfrak{a} + \mathfrak{a} \diamond (\fb \triangleright \fc) + (\fc \triangleright \fb) \triangleright \mathfrak{a} + \mathfrak{a} \diamond (\fc \triangleright \fb) + (\fb \diamond \fc) \triangleright \mathfrak{a}\\
&+ (\mathfrak{a} \triangleright (\fb \triangleright \fc) + \mathfrak{a} \triangleright (\fc \triangleright \fb) + \mathfrak{a} \triangleright (\fb \diamond \fc)) + \mathfrak{a} \diamond (\fb \diamond \fc)\\
&= \ (\fb \triangleright \fc) \triangleright \mathfrak{a} + \mathfrak{a} \diamond (\fb \triangleright \fc) + (\fc \triangleright \fb) \triangleright \mathfrak{a} + \mathfrak{a} \diamond (\fc \triangleright \fb) + (\fb \diamond \fc) \triangleright \mathfrak{a}\\
&+ \mathfrak{a} \triangleright (\fb \shuffle \fc) + \mathfrak{a} \diamond (\fb \diamond \fc).
\end{align*}
We now compare the above expansions. We have showed that $(\mathfrak{a} \diamond \fb) \diamond \fc = \mathfrak{a} \diamond (\fb \diamond \fc)$. On the other hand, we have;
\begin{align*}
\fc \triangleright (\mathfrak{a} \shuffle \fb) = x_c(\fc_- \shuffle (\mathfrak{a} \shuffle \fb))
\end{align*}
and
\begin{align*}
(\fc \triangleright \fb) \triangleright \mathfrak{a} = x_c(\fc_- \shuffle \fb) \triangleright \mathfrak{a} = x_c((\fc_- \shuffle \fb) \shuffle \mathfrak{a}).
\end{align*}
From the induction hypothesis and commutativity of shuffle product, one deduces that $\fc \triangleright (\mathfrak{a} \shuffle \fb) = (\fc \triangleright \fb) \triangleright \mathfrak{a}$. Similarly, one deduces that $(\mathfrak{a} \triangleright \fb) \triangleright \fc= \mathfrak{a} \triangleright (\fb \shuffle \fc)$.
We have
\begin{align*}
(\mathfrak{a} \triangleright \fb) \diamond \fc &= x_a(\mathfrak{a}_- \shuffle \fb) \diamond \fc \\
&= (x_a \diamond x_c) \triangleright ((\mathfrak{a}_- \shuffle \fb) \shuffle \fc_-)
\end{align*}
and
\begin{align*}
\mathfrak{a} \diamond (\fc \triangleright \fb) &= \mathfrak{a} \diamond x_c(\fc_- \shuffle \fb)\\
&= (x_a \diamond x_c) \triangleright (\mathfrak{a}_- \shuffle (\fc_- \shuffle \fb)).
\end{align*}
From the induction hypothesis and commutativity of shuffle product, one deduces that $(\mathfrak{a} \triangleright \fb) \diamond \fc = \mathfrak{a} \diamond (\fc \triangleright \fb)$.
We have
\begin{align*}
(\fb \triangleright \mathfrak{a}) \triangleright \fc &= x_b(\fb_- \shuffle \mathfrak{a}) \triangleright \fc\\
&= x_b((\fb_- \shuffle \mathfrak{a}) \shuffle \fc)
\end{align*}
and
\begin{align*}
(\fb \triangleright \fc) \triangleright \mathfrak{a} &= x_b(\fb_- \shuffle \fc) \triangleright \mathfrak{a}\\
&= x_b((\fb_- \shuffle \fc) \shuffle \mathfrak{a}).
\end{align*}
From the induction hypothesis and commutativity of shuffle product, one deduces that $ (\fb \triangleright \mathfrak{a}) \triangleright \fc = (\fb \triangleright \fc) \triangleright \mathfrak{a}$.
It follows from the induction hypothesis that
\begin{align*}
(\mathfrak{a} \diamond \fb) \triangleright \fc
&= \Bigg[x_{a+b}(\mathfrak{a}_- \shuffle \fb_-) + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_i (x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-))\Bigg] \triangleright \fc\\
&= \ x_{a+b}(\mathfrak{a}_- \shuffle \fb_-) \triangleright \fc + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_i (x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)) \triangleright \fc\\
&= \ x_{a+b}((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc) + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_i [(x_j \shuffle (\mathfrak{a}_- \shuffle \fb_-)) \shuffle \fc]\\
&= \ x_{a+b}((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc) + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_i [x_j \shuffle ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc)]\\
&= \ x_{a+b} \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc) + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_ix_j \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc)\\
&= \ \Bigg[x_{a+b} + \sum \limits_{i+j = a + b} \Delta^j_{a,b} x_ix_j \Bigg] \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc)\\
&= \ (x_a \diamond x_b) \triangleright ((\mathfrak{a}_- \shuffle \fb_-) \shuffle \fc)\\
\end{align*}
and
\begin{align*}
\mathfrak{a} \diamond (\fb \triangleright \fc) = \mathfrak{a} \diamond x_b(\fb_- \shuffle \fc) = (x_a \diamond x_b) \triangleright (\mathfrak{a}_- \shuffle (\fb_- \shuffle \fc)).
\end{align*}
From the induction hypothesis, one deduces that $(\mathfrak{a} \diamond \fb) \triangleright \fc = \mathfrak{a} \diamond (\fb \triangleright \fc)$. Similarly, one deduces that $(\fb \triangleright \mathfrak{a}) \diamond \fc = (\fb \diamond \fc) \triangleright \mathfrak{a}$.
From the above arguments, we conclude that $(\mathfrak{a} \shuffle \fb) \shuffle \fc = \mathfrak{a} \shuffle (\fb \shuffle \fc)$. This completes the proof.
\end{proof}
As a direct consequence of Proposition \ref{prop: commutative} and Proposition \ref{prop: associative}, we obtain the following result.
\begin{theorem} \label{theorem: algebras}
The spaces $(\mathfrak{C}, \diamond)$ and $(\mathfrak{C}, \shuffle )$ are commutative $\mathbb F_q$-algebras. In particular, Conjecture 3.2.2 of \cite{Shi18} holds.
\end{theorem}
The following proposition summarizes several properties of different products $\triangleright$, $\diamond$ and $\shuffle $ that will be useful in the sequel.
\begin{proposition} \label{prop: key properties}
Let $\mathfrak{a}, \fb, \fc \in \langle \Sigma \rangle$ be arbitrary words. Then we have
\begin{enumerate}[$(1)$]
\item $\mathfrak{a} \diamond \fb=\fb \diamond \mathfrak{a}$.
\item $(\mathfrak{a} \diamond \fb) \diamond \fc=\mathfrak{a} \diamond (\fb \diamond \fc)$.
\item $\mathfrak{a} \shuffle \fb = \fb\shuffle \mathfrak{a}=\mathfrak{a} \triangleright \fb + \fb \triangleright \mathfrak{a} + \mathfrak{a} \diamond \fb$.
\item $(\mathfrak{a} \shuffle \fb) \shuffle \fc=\mathfrak{a} \shuffle (\fb \shuffle \fc)$.
\end{enumerate}
If we assume further that $\mathfrak{a}, \fb, \fc$ are nonempty words, then
\begin{enumerate}[$(1)$]
\item $(\mathfrak{a} \triangleright \fb) \triangleright \fc=(\mathfrak{a} \triangleright \fc) \triangleright \fb=\mathfrak{a} \triangleright (\fb\shuffle \fc)$.
\item $(\mathfrak{a} \diamond \fb) \triangleright \fc=\mathfrak{a} \diamond (\fb \triangleright \fc)=(\mathfrak{a} \triangleright \fc) \diamond \fb$.
\end{enumerate}
\end{proposition}
\section{Shuffle Hopf algebra in positive characteristic} \label{sec: Hopf algebra structure}
We first equip the shuffle algebra with a coproduct $\Delta$ and a counit map. The main theorem of this section states that these give a Hopf algebra structure of the shuffle algebra (see Theorem \ref{thm: Hopf algebra for shuffle product}).
Throughout this section we continue with the notation of the previous section.
\subsection{Coproduct} ${}$\par \label{sec: coproduct}
We first introduce the coproduct
\[ \Delta: \frak C \to \frak C \otimes \frak C. \]
We will define it on $\langle \Sigma \rangle$ by induction on weight and extend by $\mathbb F_q$-linearity to $\frak C$. We put
\begin{align*}
\Delta(1)&:=1 \otimes 1, \\
\Delta(x_1)&:=1 \otimes x_1 + x_1 \otimes 1.
\end{align*}
Let $w \in \mathbb N$ and we suppose that we have defined $\Delta(\fv)$ for all words $\fv$ of weight $w(\fv)<w$. We now give a formula for $\Delta(\fu)$ for all words $\fu$ with $w(\fu)=w$. For such a word $\fu$ with $\depth(\fu)>1$, we put $\fu=x_u \fv$ with $w(\fv)<w$. Since $x_u$ and $\fv$ are both of weight less than $w$, we have already defined
\begin{align*}
\Delta(x_u)&:=1 \otimes x_u + \sum a_u \otimes b_u, \\
\Delta(\fv)&:= \sum a_\fv \otimes b_\fv.
\end{align*}
Then we set
\begin{align*}
\Delta(\fu):=1 \otimes \fu + \sum (a_u \triangleright a_\fv) \otimes (b_u \shuffle b_\fv).
\end{align*}
Our last task is to define $\Delta(x_w)$. We know that
\[ x_1\shuffle x_{w-1}=x_w+x_1x_{w-1}+x_{w-1}x_1+\sum_{0<j<w} \Delta^j_{1,w-1} x_{w-j} x_j \]
where all the words $x_{w-j} x_j$ have weight $w$ and depth $2$ and all $\Delta^j_{1,w-1}$ belong to~$\mathbb F_q$. Therefore, we set
\begin{equation} \label{eq: coproduct}
\Delta(x_w):=\Delta(x_1) \shuffle \Delta(x_{w-1})-\Delta(x_1x_{w-1})-\Delta(x_{w-1}x_1)-\sum_{0<j<w} \Delta^j_{1,w-1} \Delta(x_{w-j} x_j).
\end{equation}
We note that this definition of coproduct is different from that given as in Shi's thesis (see \cite[\S 3.2.3]{Shi18}).
\begin{lemma} \label{lem: factor 1}
For all words $\fu$, we have
\[ \Delta(\fu)=1 \otimes \fu+\sum a_\fu \otimes b_\fu \]
where $a_\fu \neq 1$.
\end{lemma}
\begin{proof}
The proof is by induction on the weight $w=w(\fu)$. For $w=0$ and $w=1$ Lemma \ref{lem: factor 1} immediately holds as
\begin{align*}
\Delta(1)&:=1 \otimes 1, \\
\Delta(x_1)&:=1 \otimes x_1 + x_1 \otimes 1.
\end{align*}
Let $w \in \mathbb N$ with $w \geq 2$. We suppose that for all words $\fu$ with $w(\fu)<w$, we have
\[ \Delta(\fu)=1 \otimes \fu+\sum a_\fu \otimes b_\fu \]
where $a_\fu \neq 1$.
We have to prove that the statement holds for all words $\fu$ with $w(\fu)=w$. In fact, we first consider a word $\fu$ with $w(\fu)=w$ and $\depth(\fu)>1$. We put $\fu=x_u \fv$ with $\depth(\fv) \geq 1$. By the induction hypothesis, we write
\[ \Delta(x_u)=1 \otimes x_u+\sum a_u \otimes b_u \]
where $a_u \neq 1$. If we put $\Delta(x_\fv)=\sum a_\fv \otimes b_\fv$, then we know that
\begin{align*}
\Delta(\fu):=1 \otimes \fu + \sum (a_u \triangleright a_\fv) \otimes (b_u \shuffle b_\fv).
\end{align*}
Since $a_u \neq 1$, $a_u \triangleright a_\fv \neq 1$. Thus Lemma \ref{lem: factor 1} holds for $\fu$.
To conclude, it suffices to prove that Lemma \ref{lem: factor 1} holds for $x_w$. By the induction hypothesis, we deduce that
\begin{align*}
\Delta(x_1)\shuffle \Delta(x_{w-1})=1 \otimes (x_1\shuffle x_{w-1}) + \sum a \otimes b, \quad a \neq 1,
\end{align*}
and for all $0<j<w$,
\begin{align*}
\Delta(x_{w-j} x_j)=1 \otimes x_{w-j} x_j + \sum a_j \otimes b_j, \quad a_j \neq 1.
\end{align*}
Thus by \eqref{eq: coproduct},
\begin{align*}
& \Delta(x_w) \\
&=\Delta(x_1) \shuffle \Delta(x_{w-1})-\Delta(x_1x_{w-1})-\Delta(x_{w-1}x_1)-\sum_{0<j<w} \Delta^j_{j,w-j} \Delta(x_{w-j} x_j) \\
&= \left(1 \otimes (x_1\shuffle x_{w-1}) + \sum a \otimes b\right)-\left(1 \otimes x_1x_{w-1} + \sum a_{w-1} \otimes b_{w-1}\right) \\
&-\left(1 \otimes x_{w-1}x_1 + \sum a_1 \otimes b_1\right)-\sum_{0<j<w} \Delta^j_{j,w-j} \left(1 \otimes x_{w-j} x_j + \sum a_j \otimes b_j\right) \\
&= 1 \otimes \left(x_1\shuffle x_{w-1}-x_1x_{w-1}-x_{w-1}x_1-\sum_{0<j<w} \Delta^j_{j,w-j} x_{w-j} x_j \right) \\
&+ \sum a \otimes b- \sum a_{w-1} \otimes b_{w-1}- \sum a_1 \otimes b_1-\sum_{0<j<w} \Delta^j_{j,w-j} \sum a_j \otimes b_j \\
&= 1 \otimes x_w+ \sum a \otimes b-\sum a_{w-1} \otimes b_{w-1}-\sum a_1 \otimes b_1-\sum_{0<j<w} \Delta^j_{j,w-j} \sum a_j \otimes b_j.
\end{align*}
The proof is finished.
\end{proof}
\subsection{Compatibility of the coproduct} ${}$\par
In this section we prove the compatibility of the coproduct $\Delta$ given as in the previous section.
\begin{theorem} \label{thm: compatibility}
Let $\mathfrak{a}, \fb \in \langle \Sigma \rangle$. Then we have
\[ \Delta(\mathfrak{a}\shuffle \fb)=\Delta(\mathfrak{a})\shuffle \Delta(\fb). \]
\end{theorem}
The rest of this section is devoted to a proof of Theorem \ref{thm: compatibility}. The proof is by induction on the total weight $w=w(\mathfrak{a})+w(\fb)$.
For $w=0$ and $w=1$ we see that Theorem \ref{thm: compatibility} holds. Let $w \in \mathbb N$ with $w \geq 2$ and we suppose that for all $\mathfrak{a}, \fb \in \langle \Sigma \rangle$ such that $w(\mathfrak{a})+w(\fb)<w$, we have $\Delta(\mathfrak{a}\shuffle \fb)=\Delta(\mathfrak{a})\shuffle \Delta(\fb)$.
We now show that for all $\mathfrak{a}, \fb \in \langle \Sigma \rangle$ such that $w(\mathfrak{a})+w(\fb)=w$, we have
\[ \Delta(\mathfrak{a}\shuffle \fb)=\Delta(\mathfrak{a})\shuffle \Delta(\fb). \]
The proof will be divided into three steps.
\subsubsection{Step 1}
We first prove the following proposition.
\begin{proposition} \label{prop: compatibility step 1}
For all words $x_u \fu, x_v \in \langle \Sigma \rangle$ with $u,v \in \mathbb N$, $\depth(\fu) \geq 1$ and $w(x_u \fu)+w(x_v)=w$, we have
\[ \Delta(x_u \fu\shuffle x_v)=\Delta(x_u \fu)\shuffle \Delta(x_v). \]
\end{proposition}
\begin{proof}
By definition of the product $\shuffle $ and Lemma \ref{lem: triangle formulas}, we write as
\begin{align} \label{eq: expansion 1}
x_u \fu \shuffle x_v&=x_v \triangleright (x_u \fu)+(x_u \fu) \triangleright x_v+(x_u \fu) \diamond x_v \\
&=x_vx_u\fu+x_u(\fu \shuffle x_v) +(x_u \diamond x_v) \triangleright \fu \notag \\
&=x_v x_u \fu + x_u(\fu\shuffle x_v)+x_{u+v} \fu+ \sum_{0<j<u+v} \Delta^j_{u,v} x_{u+v-j} (x_j\shuffle \fu). \notag
\end{align}
where the coefficients $\Delta^j_{u,v}$ belong to $\mathbb F_q$. Therefore, we get
\begin{align} \label{eq:compatibility 1}
& \Delta(x_u \fu \shuffle x_v)-\Delta(x_u \fu)\shuffle \Delta(x_v) \\
&= \Delta(x_v x_u \fu) + \Delta(x_u(\fu\shuffle x_v)) \notag \\
&+\Delta(x_{u+v} \fu)+ \sum_{0<j<u+v} \Delta^j_{u,v} \Delta(x_{u+v-j} (x_j\shuffle \fu)) \notag \\
&-\Delta(x_u \fu)\shuffle \Delta(x_v). \notag
\end{align}
We now analyze each term of the RHS of the above expression. To do so we put
\begin{align*}
\Delta(\fu) &=1 \otimes \fu+\sum a_\fu \otimes b_\fu,
\end{align*}
and for all $j \in \mathbb N$, we simply put
\begin{align*}
\Delta(x_j) &=1 \otimes x_j+\sum a_j \otimes b_j.
\end{align*}
In particular,
\begin{align*}
\Delta(x_u) &=1 \otimes x_u+\sum a_u \otimes b_u, \\
\Delta(x_v) &=1 \otimes x_v+\sum a_v \otimes b_v.
\end{align*}
\noindent {\bf The first term $\Delta(x_v x_u \fu)$.} \par From the definition of the coproduct $\Delta$ we deduce
\begin{equation} \label{eq: x_u u}
\Delta(x_u \fu)=1 \otimes x_u \fu+\sum a_u \otimes (b_u\shuffle \fu)+\sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu).
\end{equation}
Thus
\begin{align} \label{eq:term 1}
\Delta(x_v x_u \fu)&=1 \otimes x_v x_u \fu+ \sum a_v \otimes ((x_u \fu)\shuffle b_v)+\sum (a_v \triangleright a_u) \otimes (b_u\shuffle \fu\shuffle b_v) \\
&+\sum (a_v \triangleright (a_u \triangleright a_\fu)) \otimes (b_u\shuffle b_\fu\shuffle b_v). \notag
\end{align}
\noindent {\bf The second term $\Delta(x_u(\fu\shuffle x_v))$.} \par Since $w(\fu)+w(x_v)<w$, the induction hypothesis implies
\begin{align*}
\Delta(\fu\shuffle x_v)&=\Delta(\fu)\shuffle \Delta(x_v) \\
&=\left(1 \otimes \fu+\sum a_\fu \otimes b_\fu \right)\shuffle \left(1 \otimes x_v+\sum a_v \otimes b_v \right) \\
&=1 \otimes (\fu\shuffle x_v)+\sum a_\fu \otimes (b_\fu\shuffle x_v) \\
&+\sum a_v \otimes (\fu\shuffle b_v)+\sum (a_\fu\shuffle a_v) \otimes (b_\fu\shuffle b_v).
\end{align*}
It follows that
\begin{align} \label{eq:term 2}
\Delta(x_u(\fu\shuffle x_v))&=1 \otimes (x_u(\fu\shuffle x_v))+\sum a_u \otimes (b_u\shuffle \fu\shuffle x_v) \\
&+\sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu\shuffle x_v)+\sum (a_u \triangleright a_v) \otimes (b_u\shuffle \fu\shuffle b_v) \notag \\
&+\sum (a_u \triangleright (a_\fu\shuffle a_v)) \otimes (b_u\shuffle b_\fu\shuffle b_v). \notag
\end{align}
\noindent {\bf The third term $\Delta(x_{u+v} \fu)$.} \par By definition,
\begin{align} \label{eq:term 3}
\Delta(x_{u+v} \fu) &=1 \otimes (x_{u+v} \fu)+ \sum a_{u+v} \otimes (b_{u+v}\shuffle \fu)+\sum (a_{u+v} \triangleright a_\fu) \otimes (b_{u+v}\shuffle b_\fu).
\end{align}
\noindent {\bf The fourth terms $\Delta(x_{u+v-j} (x_j\shuffle \fu))$ for all $0<j<u+v$.} \par
As $w(x_j)+w(\fu)<w$, by the induction hypothesis,
\begin{align*}
\Delta(x_j\shuffle \fu)&=\Delta(x_j)\shuffle \Delta(\fu) \\
&= \left(1 \otimes x_j+\sum a_j \otimes b_j \right)\shuffle \left(1 \otimes \fu+\sum a_\fu \otimes b_\fu \right) \\
&= 1 \otimes (x_j\shuffle \fu)+\sum a_j \otimes (b_j\shuffle \fu) \\
&+\sum a_\fu \otimes (x_j\shuffle b_\fu)+\sum (a_j\shuffle a_\fu) \otimes (b_j\shuffle b_\fu).
\end{align*}
We then get;
\begin{align} \label{eq:term 4}
\Delta(x_{u+v-j} (x_j\shuffle \fu))&=1 \otimes (x_{u+v-j}(x_j\shuffle \fu))+\sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j\shuffle \fu) \\
&+\sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j\shuffle \fu) \notag \\
&+\sum (a_{u+v-j} \triangleright a_\fu) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_\fu) \notag \\
&+\sum (a_{u+v-j} \triangleright (a_j\shuffle a_\fu)) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_\fu) \notag.
\end{align}
\noindent {\bf The last term $\Delta(x_u \fu)\shuffle \Delta(x_v)$.} \par
Recall that $\Delta(x_u \fu)$ is given by \eqref{eq: x_u u}. Thus,
\begin{align} \label{eq:term 5}
& \Delta(x_u \fu)\shuffle \Delta(x_v) \\
&= \left(1 \otimes x_u \fu+\sum a_u \otimes (b_u\shuffle \fu)+\sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu) \right)\shuffle \left(1 \otimes x_v+\sum a_v \otimes b_v\right) \notag \\
&= 1 \otimes ((x_u \fu)\shuffle x_v) + \sum a_v \otimes (b_v\shuffle (x_u \fu)) \notag \\
&+ \sum a_u \otimes (b_u\shuffle \fu\shuffle x_v)+\sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle \fu) \notag \\
&+ \sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu\shuffle x_v)+ \sum ((a_u \triangleright a_\fu)\shuffle a_v) \otimes (b_u\shuffle b_\fu\shuffle b_v). \notag
\end{align}
Plugging the equations \eqref{eq:term 1}, \eqref{eq:term 2}, \eqref{eq:term 3}, \eqref{eq:term 4}, \eqref{eq:term 5} into \eqref{eq:compatibility 1} yields
\begin{align*}
\Delta(x_u \fu \shuffle x_v)-\Delta(x_u \fu)\shuffle \Delta(x_v) =S_0-S_1-S_2+S_3+S_4.
\end{align*}
Here the sums $S_i$ with $0 \leq i \leq 4$ are given as follows:
\begin{align*}
S_0&= 1 \otimes x_v x_u \fu+1 \otimes (x_u(\fu\shuffle x_v))+1 \otimes (x_{u+v} \fu) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} 1 \otimes (x_{u+v-j}(x_j\shuffle \fu)) - 1 \otimes ((x_u \fu)\shuffle x_v). \\
S_1&=\sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle \fu)-\sum (a_u \triangleright a_v) \otimes (b_u\shuffle b_v\shuffle \fu) \\
&-\sum (a_v \triangleright a_u) \otimes (b_u\shuffle b_v\shuffle \fu). \\
S_2&=\sum ((a_u \triangleright a_\fu)\shuffle a_v) \otimes (b_u\shuffle b_\fu\shuffle b_v)-\sum (a_u \triangleright (a_\fu\shuffle a_v)) \otimes (b_u\shuffle b_\fu\shuffle b_v) \\
&-\sum (a_v \triangleright (a_u \triangleright a_\fu)) \otimes (b_u\shuffle b_\fu\shuffle b_v). \\
S_3&= \sum a_{u+v} \otimes (b_{u+v}\shuffle \fu)+\sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j\shuffle \fu) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j\shuffle \fu). \\
S_4&= \sum (a_{u+v} \triangleright a_\fu) \otimes (b_{u+v}\shuffle b_\fu) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_\fu) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_\fu) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright (a_j\shuffle a_\fu)) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_\fu).
\end{align*}
We claim that
\begin{enumerate}
\item $S_0=0$.
\item $S_1-S_3=0$.
\item $S_2-S_4=0$.
\end{enumerate}
We now prove the previous claim. For Part (1), we want to show that $S_0=0$. In fact, it follows immediately from \eqref{eq: expansion 1}, e.g.,
\begin{align*}
x_v x_u \fu+x_u(\fu\shuffle x_v)+x_{u+v} \fu +\sum_{0<j<u+v} \Delta^j_{u,v} x_{u+v-j}(x_j\shuffle \fu) - (x_u \fu)\shuffle x_v=0.
\end{align*}
For Part (2), we will show that
\begin{align*}
S_1=S_3=\sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v\shuffle \fu).
\end{align*}
In fact, we note that
\begin{align*}
S_1&=\sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle \fu)-\sum (a_u \triangleright a_v) \otimes (b_u\shuffle b_v\shuffle \fu) \\
&-\sum (a_v \triangleright a_u) \otimes (b_u\shuffle b_v\shuffle \fu) \\
&= \sum (a_u\shuffle a_v- a_u \triangleright a_v-a_v \triangleright a_u) \otimes (b_u\shuffle b_v\shuffle \fu) \\
&= \sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v\shuffle \fu).
\end{align*}
Here the last equality follows from Lemma \ref{lem: triangle formulas}, Part (2).
Next, since $w(x_u)+w(x_v)<w$, by the induction hypothesis we know that
\begin{equation} \label{eq: x_u x_v}
\Delta(x_u\shuffle x_v)=\Delta(x_u)\shuffle \Delta(x_v)
\end{equation}
The LHS equals
\begin{align*}
& \Delta(x_u\shuffle x_v) \\
&=\Delta(x_ux_v)+\Delta(x_vx_u)+\Delta(x_{u+v}) +\sum_{0<j<u+v} \Delta^j_{u,v} \Delta(x_{u+v-j} x_j) \\
&=1 \otimes x_ux_v + \sum a_u \otimes (b_u\shuffle x_v)+\sum (a_u \triangleright a_v) \otimes (b_u\shuffle b_v) \\
&+1 \otimes x_vx_u + \sum a_v \otimes (x_u\shuffle b_v)+\sum (a_v \triangleright a_u) \otimes (b_u\shuffle b_v) \\
&+ 1 \otimes x_{u+v} + \sum a_{u+v} \otimes b_{u+v} \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} 1 \otimes x_{u+v-j} x_j \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j).
\end{align*}
We see that
\begin{align*}
\Delta(x_u)\shuffle \Delta(x_v) &= \left(1 \otimes x_u+\sum a_u \otimes b_u \right)\shuffle \left(1 \otimes x_v+\sum a_v \otimes b_v \right) \\
&= 1 \otimes (x_u\shuffle x_v)+\sum a_u \otimes (b_u\shuffle x_v) \\
&+\sum a_v \otimes (x_u\shuffle b_v)+\sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v).
\end{align*}
Replacing these equalities into \eqref{eq: x_u x_v} gets
\begin{align*}
& \sum (a_u \triangleright a_v) \otimes (b_u\shuffle b_v) +\sum (a_v \triangleright a_u) \otimes (b_u\shuffle b_v) \\
&+ \sum a_{u+v} \otimes b_{u+v} + \sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j) \\
&= \sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v).
\end{align*}
Thus we deduce
\begin{align} \label{eq: x_u x_v 2}
& \sum a_{u+v} \otimes b_{u+v} + \sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j) \notag \\
&= \sum (a_u\shuffle a_v-a_u \triangleright a_v-a_v \triangleright a_u) \otimes (b_u\shuffle b_v) \notag \\
&= \sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v). \notag
\end{align}
As a direct consequence of \eqref{eq: x_u x_v 2}, we obtain
\begin{align*}
S_3&= \sum a_{u+v} \otimes (b_{u+v}\shuffle \fu)+\sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j\shuffle \fu) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j\shuffle \fu). \\
&= \sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v\shuffle u)
\end{align*}
as desired. We conclude that $S_1=S_3$ as claimed.
For Part (3), we will show that
\begin{align*}
S_2=S_4=\sum ((a_u \triangleright a_\fu) \diamond a_v) \otimes (b_u\shuffle b_\fu\shuffle b_v).
\end{align*}
In fact, to prove the equality for $S_2$, we note that
\begin{align*}
(a_u \triangleright a_\fu)\shuffle a_v&=(a_u \triangleright a_\fu) \triangleright a_v+a_v \triangleright (a_u \triangleright a_\fu)+(a_u \triangleright a_\fu) \diamond a_v \\
&=a_u \triangleright (a_\fu \shuffle a_v)+a_v \triangleright (a_u \triangleright a_\fu)+(a_u \triangleright a_\fu) \diamond a_v.
\end{align*}
The first equality follows from Lemma \ref{lem: triangle formulas} and the second one follows from Proposition \ref{prop: key properties}, Part 5. We then obtain
\begin{align*}
S_2&=\sum ((a_u \triangleright a_\fu)\shuffle a_v) \otimes (b_u\shuffle b_\fu\shuffle b_v)-\sum (a_u \triangleright (a_\fu\shuffle a_v)) \otimes (b_u\shuffle b_\fu\shuffle b_v) \\
&-\sum (a_v \triangleright (a_u \triangleright a_\fu)) \otimes (b_u\shuffle b_\fu\shuffle b_v) \\
&=\sum ((a_u \triangleright a_\fu)\shuffle a_v-a_u \triangleright (a_\fu\shuffle a_v)-a_v \triangleright (a_u \triangleright a_\fu)) \otimes (b_u\shuffle b_\fu\shuffle b_v) \\
&=\sum ((a_u \triangleright a_\fu) \diamond a_v) \otimes (b_u\shuffle b_\fu\shuffle b_v).
\end{align*}
We now consider the term $S_4$. We have
\begin{align*}
S_4&= \sum (a_{u+v} \triangleright a_\fu) \otimes (b_{u+v}\shuffle b_\fu) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_\fu) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_\fu) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright (a_j\shuffle a_\fu)) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_\fu) \\
&= \sum (a_{u+v} \triangleright a_\fu) \otimes (b_{u+v}\shuffle b_\fu) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_\fu) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_\fu) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum ((a_{u+v-j} \triangleright a_j) \triangleright a_\fu) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_\fu) \\
&= \sum ((a_u \diamond a_v) \triangleright a_\fu) \otimes (b_u\shuffle b_v\shuffle b_\fu) \\
&=\sum ((a_u \triangleright a_\fu) \diamond a_v) \otimes (b_u\shuffle b_\fu\shuffle b_v).
\end{align*}
Here the second and fourth equalities follow from Proposition \ref{prop: key properties}, Part 5 and 6. The third one is a direct consequence of \eqref{eq: x_u x_v 2}.
Thus we have proved that $S_2=S_4$ as claimed.
To conclude, we see immediately that
\begin{align*}
\Delta(x_u \fu \shuffle x_v)-\Delta(x_u \fu)\shuffle \Delta(x_v) =S_0-S_1-S_2+S_3+S_4=0.
\end{align*}
The proof of Proposition \ref{prop: compatibility step 1} is finished.
\end{proof}
\subsubsection{Step 2}
Next we generalize Proposition \ref{prop: compatibility step 1} for words of arbitrary depth.
\begin{proposition} \label{prop: compatibility step 2}
We work with the above assumption. Then for all words $x_u \fu, x_v \in \langle \Sigma \rangle$ with $u,v \in \mathbb N$, $\depth(\fu) \geq 1$, $\depth(\fv) \geq 1$ and $w(x_u \fu)+w(x_v \fv)=w$, we have
\[ \Delta(x_u \fu\shuffle x_v \fv)=\Delta(x_u \fu)\shuffle \Delta(x_v \fv). \]
\end{proposition}
\begin{proof}
We follow the same strategy as that of the proof of Proposition \ref{prop: compatibility step 1} but the proof is much more involved and complicated.
By the definition of the product $\shuffle $ and Lemma \ref{lem: triangle formulas}, we obtain
\begin{align} \label{eq: expansion 2}
& x_u \fu \shuffle x_v \fv \\
&=(x_u \fu) \triangleright (x_v \fv)+(x_v \fv) \triangleright (x_u \fu)+(x_u \fu) \diamond (x_v \fv) \notag \\
&=x_v(x_u \fu\shuffle \fv) + x_u(\fu\shuffle (x_v \fv))+(x_u \diamond x_v) \triangleright (\fu\shuffle \fv) \notag \\
&=x_v(x_u \fu\shuffle \fv) + x_u(\fu\shuffle (x_v \fv))+x_{u+v}(\fu\shuffle \fv)+ \sum_{0<j<u+v} \Delta^j_{u,v} x_{u+v-j} (x_j\shuffle \fu\shuffle \fv), \notag
\end{align}
where the coefficients $\Delta^j_{u,v}$ belong to $\mathbb F_q$. Therefore, we get
\begin{align} \label{eq:compatibility b}
& \Delta(x_u \fu \shuffle x_v \fv)-\Delta(x_u \fu)\shuffle \Delta(x_v \fv) \\
&= \Delta(x_v(x_u \fu\shuffle \fv)) + \Delta(x_u(\fu\shuffle (x_v \fv))+\Delta(x_{u+v}(\fu\shuffle \fv)) \notag \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \Delta(x_{u+v-j} (x_j\shuffle \fu\shuffle \fv)) -\Delta(x_u \fu)\shuffle \Delta(x_v \fv). \notag
\end{align}
We now analyze each term of the RHS of the above expression. To do so we put
\begin{align*}
\Delta(\fu) &=1 \otimes \fu+\sum a_\fu \otimes b_\fu, \\
\Delta(\fv) &=1 \otimes \fv+\sum a_\fv \otimes b_\fv,
\end{align*}
and for all $j \in \mathbb N$, we simply put
\begin{align*}
\Delta(x_j) &=1 \otimes x_j+\sum a_j \otimes b_j.
\end{align*}
In particular,
\begin{align*}
\Delta(x_u) &=1 \otimes x_u+\sum a_u \otimes b_u, \\
\Delta(x_v) &=1 \otimes x_v+\sum a_v \otimes b_v.
\end{align*}
\noindent {\bf The first term $\Delta(x_v(x_u \fu\shuffle \fv))$.} \par Recall that $\Delta(x_u \fu)$ is given as in \eqref{eq: x_u u}. As $w(x_u \fu)+w(\fv)<w$, we obtain
\begin{align*}
\Delta(x_u \fu\shuffle \fv)&= \Delta(x_u \fu)\shuffle \Delta(\fv) \\
&= 1 \otimes ((x_u \fu)\shuffle \fv)+\sum a_u \otimes (b_u\shuffle \fu\shuffle \fv)+\sum a_\fv \otimes ((x_u \fu)\shuffle b_\fv) \\
&+\sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu\shuffle \fv)+\sum (a_u \shuffle a_\fv) \otimes (b_u\shuffle \fu\shuffle b_\fv) \\
&+\sum ((a_u \triangleright a_\fu)\shuffle a_\fv) \otimes (b_u\shuffle b_\fu\shuffle b_\fv).
\end{align*}
Thus
\begin{align} \label{eq:term 1b}
\Delta(x_v(x_u \fu\shuffle \fv))&=1 \otimes x_v(x_u \fu\shuffle x_v)+ \sum a_v \otimes (b_v\shuffle (x_u \fu)\shuffle \fv) \\
&+\sum (a_v \triangleright a_u) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv) \notag \\
&+\sum (a_v \triangleright (a_u \triangleright a_\fu)) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv) \notag \\
&+\sum (a_v \triangleright a_\fv) \otimes (b_v\shuffle (x_u \fu)\shuffle b_\fv) \notag \\
&+\sum (a_v \triangleright (a_u\shuffle a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv) \notag \\
&+\sum (a_v \triangleright ((a_u \triangleright a_\fu)\shuffle a_\fv)) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle b_\fv). \notag
\end{align}
\noindent {\bf The second term $\Delta(x_u(\fu\shuffle (x_v \fv))$.} \par Similarly, we get
\begin{align} \label{eq:term 2b}
\Delta(x_u(\fu\shuffle (x_v \fv))&=1 \otimes x_u(\fu\shuffle x_v \fv)+ \sum a_u \otimes (b_u\shuffle \fu\shuffle (x_v \fv)) \\
&+\sum ((a_u \triangleright a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv) \notag \\
&+\sum (a_u \triangleright (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv) \notag \\
&+\sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu\shuffle (x_v \fv)) \notag \\
&+\sum (a_u \triangleright (a_\fu\shuffle a_v)) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle \fv) \notag \\
&+\sum (a_u \triangleright (a_\fu\shuffle (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle b_\fv). \notag
\end{align}
\noindent {\bf The third term $\Delta(x_{u+v}(\fu\shuffle \fv))$.} \par
We put
\[ \Delta(\fu\shuffle \fv)=1 \otimes (\fu\shuffle \fv)+\sum a_{\fu\shuffle \fv} \otimes b_{\fu\shuffle \fv}. \]
As $w(\fu)+w(\fv)<w$, the induction hypothesis implies that
\begin{align*}
\Delta(\fu\shuffle \fv)&=\Delta(\fu)\shuffle \Delta(\fv) \\
&= \left(1 \otimes \fu+\sum a_\fu \otimes b_\fu \right)\shuffle \left(1 \otimes \fv+\sum a_\fv \otimes b_\fv \right).
\end{align*}
Thus
\begin{align*}
& 1 \otimes (\fu\shuffle \fv)+\sum a_{\fu\shuffle \fv} \otimes b_{\fu\shuffle \fv} \\
&= \left(1 \otimes \fu+\sum a_\fu \otimes b_\fu \right)\shuffle \left(1 \otimes \fv+\sum a_\fv \otimes b_\fv \right) \\
&= 1 \otimes (\fu\shuffle \fv)+\sum a_\fu \otimes (b_\fu\shuffle \fv) \\
&+\sum a_\fv \otimes (\fu\shuffle b_\fv)+\sum (a_\fu\shuffle a_\fv) \otimes (b_\fu\shuffle b_\fv),
\end{align*}
which implies
\begin{equation} \label{eq: fu fv}
\sum a_{\fu\shuffle \fv} \otimes b_{\fu\shuffle \fv}=\sum a_\fu \otimes (b_\fu\shuffle \fv)+\sum a_\fv \otimes (\fu\shuffle b_\fv)+\sum (a_\fu\shuffle a_\fv) \otimes (b_\fu\shuffle b_\fv).
\end{equation}
Finally, we have
\begin{align} \label{eq:term 3b}
\Delta(x_{u+v}(\fu\shuffle \fv)) &=1 \otimes (x_{u+v}(\fu\shuffle \fv))+ \sum a_{u+v} \otimes (b_{u+v}\shuffle \fu\shuffle \fv) \\
&+\sum (a_{u+v} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v}\shuffle b_{\fu\shuffle \fv}). \notag
\end{align}
\noindent {\bf The fourth terms $\Delta(x_{u+v-j} (x_j\shuffle \fu\shuffle \fv))$ for all $0<j<u+v$.} \par
As $w(x_j)+w(\fu\shuffle \fv)<w$, by the induction hypothesis, we get
\begin{align*}
& \Delta(x_j\shuffle \fu\shuffle \fv) \\
&=\Delta(x_j)\shuffle \Delta(\fu\shuffle \fv) \\
&= \left(1 \otimes x_j+\sum a_j \otimes b_j \right)\shuffle \left(1 \otimes (\fu\shuffle \fv)+\sum a_{\fu\shuffle \fv} \otimes b_{\fu\shuffle \fv} \right) \\
&= 1 \otimes (x_j\shuffle \fu\shuffle \fv)+\sum a_j \otimes (b_j\shuffle \fu\shuffle \fv) \\
&+\sum a_{\fu\shuffle \fv} \otimes (x_j\shuffle b_{\fu\shuffle \fv})+\sum (a_j\shuffle a_{\fu\shuffle \fv}) \otimes (b_j\shuffle b_{\fu\shuffle \fv}).
\end{align*}
We then get
\begin{align} \label{eq:term 4b}
& \Delta(x_{u+v-j} (x_j\shuffle \fu\shuffle \fv)) \\
&=1 \otimes (x_{u+v-j}(x_j\shuffle \fu\shuffle \fv))+\sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j\shuffle \fu\shuffle \fv) \notag \\
&+\sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j\shuffle \fu\shuffle \fv) \notag \\
&+\sum (a_{u+v-j} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_{\fu\shuffle \fv}) \notag \\
&+\sum (a_{u+v-j} \triangleright (a_j\shuffle a_{\fu\shuffle \fv})) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_{\fu\shuffle \fv}) \notag.
\end{align}
\noindent {\bf The last term $\Delta(x_u \fu)\shuffle \Delta(x_v \fv)$.} \par
Recall that $\Delta(x_u \fu)$ is given by \eqref{eq: x_u u}. Similarly,
\begin{equation*}
\Delta(x_v \fv)=1 \otimes x_v \fv+\sum a_v \otimes (b_v\shuffle \fv)+\sum (a_v \triangleright a_\fv) \otimes (b_v\shuffle b_\fv).
\end{equation*}
Thus
\begin{align} \label{eq:term 5b}
& \Delta(x_u \fu)\shuffle \Delta(x_v \fv) \\
&= \left(1 \otimes x_u \fu+\sum a_u \otimes (b_u\shuffle \fu)+\sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu) \right) \notag \\
&\shuffle \left(1 \otimes x_v \fv+\sum a_v \otimes (b_v\shuffle \fv)+\sum (a_v \triangleright a_\fv) \otimes (b_v\shuffle b_\fv)\right) \notag \\
&= 1 \otimes (x_u \fu\shuffle x_v \fv) \notag \\
&+ \sum a_v \otimes ((x_u \fu)\shuffle b_v\shuffle \fv)+\sum (a_v \triangleright a_\fv) \otimes ((x_u \fu)\shuffle b_v\shuffle b_\fv) \notag \\
&+ \sum a_u \otimes (b_u\shuffle \fu\shuffle (x_v \fv))+\sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv) \notag \\
&+\sum (a_u\shuffle (a_v \triangleright a_\fv)) \otimes (b_u\shuffle \fu\shuffle b_v\shuffle b_\fv) \notag \\
&+ \sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu\shuffle (x_v \fv))+ \sum ((a_u \triangleright a_\fu)\shuffle a_v) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle \fv) \notag \\
&+\sum ((a_u \triangleright a_\fu)\shuffle (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv). \notag
\end{align}
Plugging the equations \eqref{eq:term 1b}, \eqref{eq:term 2b}, \eqref{eq:term 3b}, \eqref{eq:term 4b}, \eqref{eq:term 5b} into \eqref{eq:compatibility b} yields
\begin{align*}
\Delta(x_u \fu \shuffle x_v \fv)-\Delta(x_u \fu)\shuffle \Delta(x_v \fv) =S_0-S_1-S_2-S_\fu-S_\fv+S_3+S_4.
\end{align*}
Here the sums $S_i$ with $0 \leq i \leq 4$ and $S_\fu$, $S_\fv$ are given as follows:
\begin{align*}
S_0&= 1\otimes x_v(x_u \fu\shuffle \fv) + 1 \otimes x_u(\fu\shuffle (x_v \fv))+1 \otimes x_{u+v}(\fu\shuffle \fv) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} 1 \otimes x_{u+v-j} (x_j\shuffle \fu\shuffle \fv)-1 \otimes ((x_u \fu)\shuffle (x_v \fv)). \\
S_1&=\sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv)-\sum (a_u \triangleright a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv) \\
&-\sum (a_v \triangleright a_u) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv). \\
S_2&=\sum ((a_u \triangleright a_\fu)\shuffle (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv) \\
&-\sum (a_u \triangleright (a_\fu\shuffle (a_v \triangleright a_\fv))) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv) \\
&-\sum (a_v \triangleright ((a_u \triangleright a_\fu)\shuffle a_\fv) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv). \\
S_3&= \sum a_{u+v} \otimes (b_{u+v}\shuffle \fu\shuffle \fv) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j\shuffle \fu\shuffle \fv) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j\shuffle \fu\shuffle \fv). \\
S_4&= \sum (a_{u+v} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v}\shuffle b_{\fu\shuffle \fv}) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_{\fu\shuffle \fv}) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright (a_j\shuffle a_{\fu\shuffle \fv})) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_{\fu\shuffle \fv}),
\end{align*}
and
\begin{align*}
S_\fu&=\sum (a_u\shuffle (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv)-\sum (a_u \triangleright (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv) \\
&-\sum (a_v \triangleright (a_u\shuffle a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv), \\
S_\fv&=\sum ((a_u \triangleright a_\fu)\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv)-\sum (a_u \triangleright (a_\fu\shuffle a_v)) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv) \\
&-\sum (a_v \triangleright (a_u \triangleright a_\fu)) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv).
\end{align*}
We claim that
\begin{enumerate}
\item $S_0=0$.
\item $S_1-S_3=0$.
\item $S_2+S_\fu+S_\fv-S_4=0$.
\end{enumerate}
We now prove the previous claim. For Part (1), we want to show that $S_0=0$. In fact, it follows immediately from \eqref{eq: expansion 2}.
For Part (2), we will show that
\begin{align*}
S_1=S_3=\sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv).
\end{align*}
In fact, Lemma \ref{lem: triangle formulas} implies
\begin{align*}
S_1&=\sum (a_u\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv)-\sum (a_u \triangleright a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv) \\
&-\sum (a_v \triangleright a_u) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv) \\
&= \sum (a_u\shuffle a_v- a_u \triangleright a_v-a_v \triangleright a_u) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv) \\
&= \sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv).
\end{align*}
Recall that by \eqref{eq: x_u x_v 2},
\begin{align*}
& \sum a_{u+v} \otimes b_{u+v} + \sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j) \\
&= \sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v).
\end{align*}
As a direct consequence, we obtain
\begin{align*}
S_3&= \sum a_{u+v} \otimes (b_{u+v}\shuffle \fu\shuffle \fv) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum a_{u+v-j} \otimes (b_{u+v-j}\shuffle x_j\shuffle \fu\shuffle \fv) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_j) \otimes (b_{u+v-j}\shuffle b_j\shuffle \fu\shuffle \fv). \\
&= \sum (a_u \diamond a_v) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle \fv)
\end{align*}
as desired. We conclude that $S_1=S_3$ as claimed.
For Part (3), we will show that
\begin{align*}
S_2+S_\fu+S_\fv=S_4=\sum ((a_u \diamond a_v) \triangleright a_{\fu\shuffle \fv}) \otimes (b_u\shuffle b_v\shuffle b_{\fu\shuffle \fv}).
\end{align*}
In fact, we have
\begin{align*}
S_2&=\sum ((a_u \triangleright a_\fu)\shuffle (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv) \\
&-\sum (a_u \triangleright (a_\fu\shuffle (a_v \triangleright a_\fv))) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv) \\
&-\sum (a_v \triangleright ((a_u \triangleright a_\fu)\shuffle a_\fv) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv) \\
&=\sum ((a_u \triangleright a_\fu)\shuffle (a_v \triangleright a_\fv)-a_u \triangleright (a_\fu\shuffle (a_v \triangleright a_\fv))-a_v \triangleright ((a_u \triangleright a_\fu)\shuffle a_\fv)) \\
&\quad \quad \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv) \\
&=\sum ((a_u \triangleright a_\fu) \diamond (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_\fu\shuffle b_v\shuffle b_\fv).
\end{align*}
Here the last equality follows from Lemma \ref{lem: triangle formulas} and Proposition \ref{prop: key properties}. More precisely,
\begin{align*}
&(a_u \triangleright a_\fu)\shuffle (a_v \triangleright a_\fv) \\&=(a_u \triangleright a_\fu) \triangleright (a_v \triangleright a_\fv)+(a_v \triangleright a_\fv) \triangleright (a_u \triangleright a_\fu)+(a_u \triangleright a_\fu) \diamond (a_v \triangleright a_\fv) \\
&=a_u \triangleright (a_\fu\shuffle (a_v \triangleright a_\fv))+a_v \triangleright ((a_u \triangleright a_\fu)\shuffle a_\fv))+(a_u \triangleright a_\fu) \diamond (a_v \triangleright a_\fv).
\end{align*}
By Lemma \ref{lem: triangle formulas} and Proposition \ref{prop: key properties} again,
\begin{align*}
S_\fu&=\sum (a_u\shuffle (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv)-\sum (a_u \triangleright (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv) \\
&-\sum (a_v \triangleright (a_u\shuffle a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv) \\
&=\sum (a_u\shuffle (a_v \triangleright a_\fv)-a_u \triangleright (a_v \triangleright a_\fv)-a_v \triangleright (a_u\shuffle a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv) \\
&=\sum (a_u \diamond (a_v \triangleright a_\fv)) \otimes (b_u\shuffle b_v\shuffle \fu\shuffle b_\fv),
\end{align*}
and
\begin{align*}
S_\fv&=\sum ((a_u \triangleright a_\fu)\shuffle a_v) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv)-\sum (a_u \triangleright (a_\fu\shuffle a_v)) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv) \\
&-\sum (a_v \triangleright (a_u \triangleright a_\fu)) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv) \\
&=\sum ((a_u \triangleright a_\fu) \diamond a_v) \otimes (b_u\shuffle b_v\shuffle b_\fu\shuffle \fv).
\end{align*}
Combining the previous formulas for $S_2$, $S_\fu$, $S_\fv$ with \eqref{eq: fu fv} yields
\begin{align*}
S_2+S_\fu+S_\fv=\sum ((a_u \diamond a_v) \triangleright a_{\fu\shuffle \fv}) \otimes (b_u\shuffle b_v\shuffle b_{\fu\shuffle \fv}).
\end{align*}
We now consider the term $S_4$. We have
\begin{align*}
S_4&= \sum (a_{u+v} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v}\shuffle b_{\fu\shuffle \fv}) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_{\fu\shuffle \fv}) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright (a_j\shuffle a_{\fu\shuffle \fv})) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_{\fu\shuffle \fv}) \\
&= \sum (a_{u+v} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v}\shuffle b_{\fu\shuffle \fv}) \\
&+\sum_{0<j<u+v} \Delta^j_{u,v} \sum (a_{u+v-j} \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v-j}\shuffle x_j\shuffle b_{\fu\shuffle \fv}) \\
&+ \sum_{0<j<u+v} \Delta^j_{u,v} \sum ((a_{u+v-j} \triangleright a_j) \triangleright a_{\fu\shuffle \fv}) \otimes (b_{u+v-j}\shuffle b_j\shuffle b_{\fu\shuffle \fv}) \\
&= \sum ((a_u \diamond a_v) \triangleright a_{\fu\shuffle \fv}) \otimes (b_u\shuffle b_v\shuffle b_{\fu\shuffle \fv}).
\end{align*}
Here the second equality follows from Proposition \ref{prop: key properties}. The third one is a direct consequence of \eqref{eq: x_u x_v 2}.
We then have proved that $S_2+S_\fu+S_\fv=S_4$ as claimed.
To conclude, we see immediately that
\begin{align*}
\Delta(x_u \fu \shuffle x_v \fv)-\Delta(x_u \fu)\shuffle \Delta(x_v \fv)&=S_0-S_1-S_2-S_\fu-S_\fv+S_3+S_4 \\
&=S_0-(S_1-S_3)-(S_2+S_\fu+S_\fv-S_4) \\
&=0.
\end{align*}
The proof of Proposition \ref{prop: compatibility step 2} is finished.
\end{proof}
\subsubsection{Step 3}
By Propositions \ref{prop: compatibility step 1} and \ref{prop: compatibility step 2}, for all words $\fu,\fv$ such that $\depth(\fu)+\depth(\fv)>2$ and $w(\fu)+w(\fv)=w$, we have proved
\[ \Delta(\fu\shuffle \fv)=\Delta(\fu)\shuffle \Delta(\fv). \]
To finish the proof of Theorem \ref{thm: compatibility} we prove the remaining case where both $\fu$ and $\fv$ have depth 1.
\begin{proposition} \label{prop: compatibility step 3}
Let $u,v \in \mathbb N$ such that $u+v=w$. Then
\[ \Delta(x_u\shuffle x_v)=\Delta(x_u)\shuffle \Delta(x_v). \]
\end{proposition}
\begin{proof}
By the definition of the coproduct $\Delta$, for all $n \in \mathbb N$,
\begin{equation} \label{eq: x_1}
\Delta(x_1\shuffle x_n)=\Delta(x_1)\shuffle \Delta(x_n).
\end{equation}
It follows that Proposition \ref{prop: compatibility step 3} holds if either $u=1$ or $v=1$. We now suppose that $u,v \geq 2$.
We claim that
\begin{equation} \label{eq: x_1 x_u-1 x_v}
\Delta(x_1\shuffle x_{u-1}\shuffle x_v)=\Delta(x_1)\shuffle \Delta(x_{u-1})\shuffle \Delta(x_v).
\end{equation}
In fact, we write
\[ x_{u-1}\shuffle x_v=x_{u+v-1}+x_{u-1}x_v+x_vx_{u-1}+\sum_{0<j<u+v-1} \Delta^j_{u-1,v} x_{u+v-1-j} x_j. \]
Thus
\begin{align*}
& \Delta(x_1\shuffle x_{u-1}\shuffle x_v) \\
&= \Delta(x_1\shuffle (x_{u-1}\shuffle x_v)) \\
&=\Delta \left(x_1\shuffle \left(x_{u+v-1}+x_{u-1}x_v+x_vx_{u-1}+\sum_{0<j<u+v-1} \Delta^j_{u-1,v} x_{u+v-1-j} x_j\right)\right) \\
&=\Delta (x_1\shuffle x_{u+v-1})+\Delta(x_1\shuffle x_{u-1}x_v)+\Delta(x_1\shuffle x_vx_{u-1}) \\
&+\sum_{0<j<u+v-1} \Delta^j_{u-1,v} \Delta(x_1\shuffle x_{u+v-1-j} x_j) \\
&=\Delta (x_1)\shuffle \Delta(x_{u+v-1})+\Delta(x_1)\shuffle \Delta(x_{u-1}x_v) \\
&+\Delta(x_1)\shuffle \Delta(x_vx_{u-1})+\sum_{0<j<u+v-1} \Delta^j_{u-1,v} \Delta(x_1)\shuffle \Delta(x_{u+v-1-j} x_j).
\end{align*}
Here the last equality follows from \eqref{eq: x_1} and Proposition \ref{prop: compatibility step 1}. It implies
\begin{align*}
& \Delta(x_1\shuffle x_{u-1}\shuffle x_v) \\
&=\Delta (x_1)\shuffle \Delta(x_{u+v-1})+\Delta(x_1)\shuffle \Delta(x_{u-1}x_v) \\
&+\Delta(x_1)\shuffle \Delta(x_vx_{u-1})+\sum_{0<j<u+v-1} \Delta^j_{u-1,v} \Delta(x_1)\shuffle \Delta(x_{u+v-1-j} x_j) \\
&=\Delta (x_1)\shuffle \Delta\left(x_{u+v-1}+x_{u-1}x_v+x_vx_{u-1}+\sum_{0<j<u+v-1} \Delta^j_{u-1,v} x_{u+v-1-j} x_j \right) \\
&=\Delta (x_1)\shuffle \Delta(x_{u-1}\shuffle x_v).
\end{align*}
As $u+v-1=w-1<w$, $\Delta(x_{u-1}\shuffle x_v)=\Delta(x_{u-1})\shuffle \Delta(x_v)$, which implies the claim.
Now we express both sides of \eqref{eq: x_1 x_u-1 x_v} by a different way. First, the LHS of \eqref{eq: x_1 x_u-1 x_v} equals
\begin{align*}
&\Delta(x_1\shuffle x_{u-1}\shuffle x_v) \\
&=\Delta((x_1\shuffle x_{u-1})\shuffle x_v) \\
&=\Delta \left(\left(x_u+x_1x_{u-1}+x_{u-1}x_1+\sum_{0<j<u} \Delta^j_{1,u-1} x_{u-j} x_j \right)\shuffle x_v \right) \\
&=\Delta(x_u\shuffle x_v)+\Delta(x_1x_{u-1}\shuffle x_v)+\Delta(x_{u-1}x_1\shuffle x_v)+\sum_{0<j<u} \Delta^j_{1,u-1} \Delta(x_{u-j} x_j\shuffle x_v) \\
&=\Delta(x_u\shuffle x_v)+\Delta(x_1 x_{u-1})\shuffle \Delta(x_v)+\Delta(x_{u-1} x_1)\shuffle \Delta(x_v) \\
&+\sum_{0<j<u} \Delta^j_{1,u-1} \Delta(x_{u-j} x_j)\shuffle \Delta(x_v).
\end{align*}
The last equality holds by Proposition \ref{prop: compatibility step 1}.
Next, as $\Delta(x_1)\shuffle \Delta(x_{u-1})=\Delta(x_1\shuffle x_{u-1})$, the RHS of \eqref{eq: x_1 x_u-1 x_v} equals
\begin{align*}
& \Delta(x_1)\shuffle \Delta(x_{u-1})\shuffle \Delta(x_v) \\
&=(\Delta(x_1)\shuffle \Delta(x_{u-1}))\shuffle \Delta(x_v) \\
&=\Delta(x_1\shuffle x_{u-1})\shuffle \Delta(x_v) \\
&=\Delta \left(x_u+x_1x_{u-1}+x_{u-1}x_1+\sum_{0<j<u} \Delta^j_{1,u-1} x_{u-j} x_j \right)\shuffle \Delta(x_v) \\
&=\Delta(x_u)\shuffle \Delta(x_v)+\Delta(x_1 x_{u-1})\shuffle \Delta(x_v)+\Delta(x_{u-1} x_1)\shuffle \Delta(x_v) \\
&+\sum_{0<j<u} \Delta^j_{1,u-1} \Delta(x_{u-j} x_j)\shuffle \Delta(x_v).
\end{align*}
Putting all together, we deduce
\[ \Delta(x_u\shuffle x_v)=\Delta(x_u)\shuffle \Delta(x_v) \]
as desired.
\end{proof}
\subsection{Coassociativity of the coproduct} ${}$\par
In this section we prove the coassociativity of the coproduct $\Delta$.
\begin{theorem} \label{thm: coassociativity}
Let $\fu \in \langle \Sigma \rangle$. Then we have
\[ (\Id \otimes \Delta) \Delta(\fu)=(\Delta \otimes \Id)\Delta(\fu). \]
\end{theorem}
The rest of this section is devoted to a proof of Theorem \ref{thm: coassociativity}. The proof is by induction on the total weight $w=w(\mathfrak{a})+w(\fb)$.
For $w=0$ and $w=1$ we see that Theorem \ref{thm: coassociativity} holds. Let $w \in \mathbb N$ with $w \geq 2$ and we suppose that for all $\fu \in \langle \Sigma \rangle$ such that $w(\fu)<w$, we have
\[ (\Id \otimes \Delta) \Delta(\fu)=(\Delta \otimes \Id)\Delta(\fu). \]
We now show that for all $\fu \in \langle \Sigma \rangle$ with $w(\fu)=w$,
\[ (\Id \otimes \Delta) \Delta(\fu)=(\Delta \otimes \Id)\Delta(\fu). \]
The proof will be divided into several steps.
\subsubsection{Step 1}
We first prove the following proposition.
\begin{proposition} \label{prop: coassociativity step 1}
For all words $\fu, \fv \in \langle \Sigma \rangle$ with $w(\fu)+w(\fv)=w$, we have
\[ (\Id \otimes \Delta) \Delta(\fu\shuffle \fv)=(\Delta \otimes \Id)\Delta(\fu\shuffle \fv). \]
\end{proposition}
\begin{proof}
We see that if $\fu=1$ or $\fv=1$, then the proposition holds. We can suppose that $\fu \neq 1$ and $\fv \neq 1$. If we write
\begin{align*}
\Delta(\fu)&=\sum a_\fu \otimes b_\fu, \\
\Delta(\fv)&=\sum a_\fv \otimes b_\fv,
\end{align*}
then by the compatibility, we get
\begin{align*}
(\Id \otimes \Delta) \Delta(\fu\shuffle \fv)&=(\Id \otimes \Delta)(\Delta(\fu)\shuffle \Delta(\fv)) \\
&= (\Id \otimes \Delta)\left(\sum (a_\fu\shuffle a_\fv) \otimes (b_\fu\shuffle b_\fv) \right) \\
&= \sum (a_\fu\shuffle a_\fv) \otimes \Delta(b_\fu\shuffle b_\fv) \\
&= \sum (a_\fu\shuffle a_\fv) \otimes (\Delta(b_\fu)\shuffle \Delta(b_\fv)) \\
&= \left(\sum a_\fu \otimes \Delta(b_\fu)\right)\shuffle \left(\sum a_\fv \otimes \Delta(b_\fv)\right) \\
&=(\Id \otimes \Delta) \Delta(\fu) \shuffle (\Id \otimes \Delta) \Delta(\fv).
\end{align*}
Similarly, we get
\begin{align*}
(\Delta \otimes \Id) \Delta(\fu\shuffle \fv)&=(\Delta \otimes \Id) (\Delta(\fu)\shuffle \Delta(\fv)) \\
&=(\Delta \otimes \Id) \Delta(\fu) \shuffle (\Delta \otimes \Id) \Delta(\fv).
\end{align*}
Since $w(\fu)<w$ and $w(\fv)<w$, the induction hypothesis implies
\begin{align*}
(\Id \otimes \Delta) \Delta(\fu)&=(\Delta \otimes \Id)\Delta(\fu), \\
(\Id \otimes \Delta) \Delta(\fv)&=(\Delta \otimes \Id)\Delta(\fv).
\end{align*}
Putting all together, we deduce
\begin{align*}
(\Id \otimes \Delta) \Delta(\fu\shuffle \fv)&=(\Id \otimes \Delta) \Delta(\fu) \shuffle (\Id \otimes \Delta) \Delta(\fv) \\
&=(\Delta \otimes \Id)\Delta(\fu)\shuffle (\Delta \otimes \Id)\Delta(\fv) \\
&=(\Delta \otimes \Id) \Delta(\fu\shuffle \fv).
\end{align*}
The proof is finished.
\end{proof}
\subsubsection{Step 2}
We define the operator $\triangleright$ for tensors as follows:
\begin{align*}
\left(\sum \frak a_1 \otimes \frak b_1\right) \triangleright \left(\sum \frak a_2 \otimes \frak b_2\right):&=\sum (\frak a_1 \triangleright \frak a_2) \otimes (\frak b_1\shuffle \frak b_2), \\
\left(\sum \frak a_1 \otimes \frak b_1 \otimes \frak c_1\right) \triangleright \left(\sum \frak a_2 \otimes \frak b_2 \otimes \frak c_2 \right):&=\sum (\frak a_1 \triangleright \frak a_2) \otimes (\frak b_1\shuffle \frak b_2) \otimes (\frak c_1\shuffle \frak c_2).
\end{align*}
We prove the following lemma which will be useful in the sequel.
\begin{lemma} \label{lem: delta without factor 1}
Let $\fu,\fv \in \langle \Sigma \rangle$ with $\fu \neq 1$. Then
\[ (\Delta(\fu)-1 \otimes \fu) \triangleright \Delta(\fv)=\Delta(\fu \triangleright \fv)-1 \otimes (\fu \triangleright \fv). \]
\end{lemma}
\begin{proof}
Suppose that $\depth(\fu)=1$, says $\fu=x_u$. Then $\fu \triangleright \fv=x_u \fv$ and we have to show that $(\Delta(x_u)-1 \otimes x_u) \triangleright \Delta(\fv)=\Delta(x_u \fv)-1 \otimes (x_u \fv)$. We write
\begin{align*}
\Delta(x_u)=1 \otimes x_u + \sum a_u \otimes b_u, \quad \Delta(\fv)= \sum a_\fv \otimes b_\fv.
\end{align*}
Recall that $\Delta(x_u\fv)=1 \otimes (x_u \fv)+\sum (a_u \triangleright a_\fv) \otimes (b_u\shuffle b_\fv)$. Thus
\begin{align*}
\Delta(x_u\fv)-1 \otimes (x_u \fv)&=\sum (a_u \triangleright a_\fv) \otimes (b_u\shuffle b_\fv) \\
&=(\Delta(\fu)-1 \otimes \fu) \triangleright \Delta(\fv)
\end{align*}
as desired.
We now suppose that $\depth(\fu)>1$. We write $\fu=x_u \fu'$. Thus $\fu \triangleright \fv=x_u(\fu'\shuffle \fv)$ and we have to show that
\[ (\Delta(x_u \fu')-1 \otimes x_u \fu') \triangleright \Delta(\fv)=\Delta(x_u(\fu'\shuffle \fv))-1 \otimes x_u(\fu'\shuffle \fv). \]
We put
\begin{align*}
\Delta(x_u)&=1 \otimes x_u+ \sum a_u \otimes b_u, \\
\Delta(\fu')&=\sum a_{\fu'} \otimes b_{\fu'}, \\
\Delta(\fv)&=\sum a_\fv \otimes b_\fv.
\end{align*}
It follows that
\begin{equation} \label{eq: words}
\Delta(x_u \fu')-1 \otimes x_u \fu'=\sum (a_u \triangleright a_{\fu'}) \otimes (b_u\shuffle b_{\fu'}).
\end{equation}
By Theorem \ref{thm: compatibility}, we have
\[ \Delta(\fu'\shuffle \fv)=\Delta(\fu')\shuffle \Delta(\fv)=\sum (a_{\fu'}\shuffle a_\fv) \otimes (b_{\fu'}\shuffle b_\fv). \]
Thus
\begin{align*}
\Delta(x_u(\fu'\shuffle \fv))-1 \otimes x_u(\fu'\shuffle \fv) &=\sum (a_u \triangleright (a_{\fu'}\shuffle a_\fv)) \otimes (b_u\shuffle b_{\fu'}\shuffle b_\fv) \\
&=\sum ((a_u \triangleright a_{\fu'}) \triangleright a_\fv)) \otimes (b_u\shuffle b_{\fu'}\shuffle b_\fv) \\
&=(\Delta(x_u \fu')-1 \otimes x_u \fu') \triangleright \Delta(\fv).
\end{align*}
The second equality holds by Proposition \ref{prop: key properties} and the last one follows from \eqref{eq: words}. We finish the proof.
\end{proof}
Next we prove the following proposition.
\begin{proposition} \label{prop: coassociativity step 2}
For all words $x_u \fu \in \langle \Sigma \rangle$ with $u \in \mathbb N$, $\depth(\fu) \geq 1$ and $w(x_u \fu)=w$, we have
\[ (\Id \otimes \Delta) \Delta(x_u \fu)=(\Delta \otimes \Id)\Delta(x_u \fu). \]
\end{proposition}
\begin{proof}
We put
\begin{align*}
\Delta(x_u)&=1 \otimes x_u + \sum a_u \otimes b_u, \\
\Delta(\fu)&=\sum a_\fu \otimes b_\fu.
\end{align*}
In particular,
\[ \Delta(x_u \fu)=1 \otimes x_u\fu + \sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu). \]
It follows that
\begin{align*}
&(\Id \otimes \Delta)\Delta(x_u \fu) \\
&=(\Id \otimes \Delta) \left(1 \otimes x_u\fu + \sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu)\right) \\
&= 1 \otimes \Delta(x_u\fu) + \sum (a_u \triangleright a_\fu) \otimes \Delta(b_u\shuffle b_\fu) \\
&=1 \otimes \Delta(x_u\fu) + \sum (a_u \triangleright a_\fu) \otimes (\Delta(b_u)\shuffle \Delta(b_\fu)).
\end{align*}
The last equality follows from the compatibility proved in Theorem \ref{thm: compatibility}.
Next, we have
\begin{align*}
&(\Delta \otimes \Id)\Delta(x_u \fu) \\
&=(\Delta \otimes \Id) \left(1 \otimes x_u\fu + \sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu)\right) \\
&= 1 \otimes 1 \otimes x_u\fu + \sum \Delta(a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu).
\end{align*}
Thus we have to show that
\begin{align} \label{eq: coassociativity}
& 1 \otimes \Delta(x_u\fu) + \sum (a_u \triangleright a_\fu) \otimes (\Delta(b_u)\shuffle \Delta(b_\fu)) \\
&=1 \otimes 1 \otimes x_u\fu + \sum \Delta(a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu). \notag
\end{align}
As $w(x_u)<w$, the induction hypothesis implies
\[ (\Id \otimes \Delta) \Delta(x_u)=(\Delta \otimes \Id)\Delta(x_u). \]
We write down the expressions of both sides. The LHS equals
\begin{align*}
(\Id \otimes \Delta) \Delta(x_u)&=(\Id \otimes \Delta)\left(1 \otimes x_u+\sum a_u \otimes b_u\right) \\
&=1 \otimes \Delta(x_u)+\sum a_u \otimes \Delta(b_u),
\end{align*}
and the RHS equals
\begin{align*}
(\Delta \otimes \Id) \Delta(x_u)&=(\Delta \otimes \Id)\left(1 \otimes x_u+\sum a_u \otimes b_u\right) \\
&=1 \otimes 1 \otimes x_u+\sum \Delta(a_u) \otimes b_u.
\end{align*}
As $(\Id \otimes \Delta) \Delta(x_u)=(\Delta \otimes \Id)\Delta(x_u)$, we get;
\begin{equation*}
1 \otimes \Delta(x_u)+\sum a_u \otimes \Delta(b_u)=1 \otimes 1 \otimes x_u+\sum \Delta(a_u) \otimes b_u.
\end{equation*}
We use Lemma \ref{lem: factor 1} and cancel the terms of the form $1 \otimes \frak a \otimes \frak b$ on both sides to get
\begin{equation} \label{eq: coasso u}
\sum a_u \otimes \Delta(b_u)=\sum (\Delta(a_u)-1 \otimes a_u) \otimes b_u.
\end{equation}
Recall that $\Delta(\fu)=\sum a_\fu \otimes b_\fu$. By similar calculations, the equality $(\Id \otimes \Delta) \Delta(\fu)=(\Delta \otimes \Id)\Delta(\fu)$ implies
\begin{equation} \label{eq: coasso fu}
\sum a_\fu \otimes \Delta(b_\fu)=\sum \Delta(a_\fu) \otimes b_\fu.
\end{equation}
Combining \eqref{eq: coasso u} and \eqref{eq: coasso fu} yields
\begin{align*}
&\left(\sum a_u \otimes \Delta(b_u)\right) \triangleright \left(\sum a_\fu \otimes \Delta(b_\fu)\right) \\
&=\left(\sum (\Delta(a_u)-1 \otimes a_u) \otimes b_u\right) \triangleright \left(\sum \Delta(a_\fu) \otimes b_\fu\right).
\end{align*}
It follows that
\begin{align*}
&\sum (a_u \triangleright a_\fu) \otimes (\Delta(b_u)\shuffle \Delta(b_\fu)) \\
&=\sum \left((\Delta(a_u)-1 \otimes a_u) \triangleright \Delta(a_\fu)\right) \otimes (b_u\shuffle b_\fu) \\
&=\sum \left(\Delta(a_u \triangleright a_\fu)-1 \otimes (a_u \triangleright a_\fu)\right) \otimes (b_u\shuffle b_\fu) \\
&=\sum \Delta(a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu)- \sum 1 \otimes (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu).
\end{align*}
The second equality holds by Lemma \ref{lem: delta without factor 1}.
As a consequence, the equality \eqref{eq: coassociativity} follows immediately from the equality
\[ \Delta(x_u \fu)=1 \otimes x_u \fu+\sum (a_u \triangleright a_\fu) \otimes (b_u\shuffle b_\fu). \]
\end{proof}
\subsubsection{Step 3}
By Proposition \ref{prop: coassociativity step 2}, for all words $\fu$ such that $\depth(\fu)>1$ and $w(\fu)=w$, we have proved
\[ (\Id \otimes \Delta) \Delta(\fu)=(\Delta \otimes \Id)\Delta(\fu). \]
To finish the proof of Theorem \ref{thm: coassociativity} we prove the remaining case where $\fu=x_w$.
\begin{proposition}
We have
\[ (\Id \otimes \Delta) \Delta(x_w)=(\Delta \otimes \Id)\Delta(x_w). \]
\end{proposition}
\begin{proof}
By Proposition \ref{prop: coassociativity step 1},
\begin{equation*}
(\Id \otimes \Delta) \Delta(x_1\shuffle x_{w-1})=(\Id \otimes \Delta) \Delta(x_1\shuffle x_{w-1}).
\end{equation*}
Now we express both sides of the above equality by using
\[ x_1\shuffle x_{w-1}=x_w+x_1x_{w-1}+x_{w-1}x_1+\sum_{0<j<w} \Delta^j_{1,w-1} x_{w-j} x_j. \]
First, the LHS equals
\begin{align*}
&(\Id \otimes \Delta) \Delta(x_1\shuffle x_{w-1}) \\
&=(\Id \otimes \Delta) \Delta\left(x_w+x_1x_{w-1}+x_{w-1}x_1+\sum_{0<j<w} \Delta^j_{1,w-1} x_{w-j} x_j\right) \\
&=(\Id \otimes \Delta) \Delta(x_w)+(\Id \otimes \Delta) \Delta(x_1x_{w-1})+(\Id \otimes \Delta) \Delta(x_{w-1}x_1) \\
&+\sum_{0<j<w} \Delta^j_{1,w-1} (\Id \otimes \Delta) \Delta(x_{w-j} x_j),
\end{align*}
and the RHS equals
\begin{align*}
&(\Delta \otimes \Id) \Delta(x_1\shuffle x_{w-1}) \\
&=(\Delta \otimes \Id) \Delta\left(x_w+x_1x_{w-1}+x_{w-1}x_1+\sum_{0<j<w} \Delta^j_{1,w-1} x_{w-j} x_j\right) \\
&=(\Delta \otimes \Id) \Delta(x_w)+(\Delta \otimes \Id) \Delta(x_1x_{w-1})+(\Delta \otimes \Id)\Delta(x_{w-1}x_1) \\
&+\sum_{0<j<w} \Delta^j_{1,w-1} (\Delta \otimes \Id) \Delta(x_{w-j} x_j).
\end{align*}
Putting all together and using Propositions \ref{prop: coassociativity step 1} and \ref{prop: coassociativity step 2}, we deduce that
\[ (\Id \otimes \Delta) \Delta(x_w)=(\Delta \otimes \Id)\Delta(x_w). \]
\end{proof}
\subsection{Hopf algebra structure} ${}$\par
The counit $\epsilon:\frak C \to \mathbb F_q$ is defined as follows: $\epsilon(1)=1$ and $\epsilon(\fu)=0$ otherwise. By induction on weight, we can check that $\Delta$ preserves the grading. So $(\frak C,\shuffle,u,\Delta,\epsilon)$ is a connected graded bialgebra. By Proposition \ref{prop: graded Hopf algebras}, we get
\begin{theorem} \label{thm: Hopf algebra for shuffle product}
The connected graded bialgebra $(\frak C,\shuffle,u,\Delta,\epsilon)$ is a connected graded Hopf algebra over $\mathbb F_q$.
\end{theorem}
We also note that the Hopf shuffle algebra is of finite type (see Definition \ref{defn: graded Hopf algebra}).
\subsection{Numerical verification} ${}$\par
We end this section by presenting some numerical experiments for the shuffle algebra in positive characteristic. We mention that these calculations have been crucial for us during this project.
We have written codes which can calculate the operations on $\mathfrak{C}$ and verified Proposition~\ref{prop: associative}, Theorem~\ref{thm: coassociativity} and Theorem~\ref{thm: compatibility} in numerous cases as below, ``extending" the verification which were done in \cite[\S 3.2.3]{Shi18} The machine used for the computation is MacBook Pro (15-inch, 2018), with 2.6 GHz 6-core Intel Core i7 CPU and 16GB of memory.
\begin{description}
\item[Associativity] \par For $q=2,3,4, 5, 7$, $\left(S_d(a)S_d(b)\right) S_d(c) = S_d(a)\left(S_d(b) S_d(c)\right)$ for all depth 1 tuples $a, b, c$ with weight $<q^3$.
\begin{itemize}
\item The running time for each of the cases when $q=2, 3, 4, 5, 7$ are less than 1 second, 16 seconds, 170 seconds, 107 minutes, and 118 hours, respectively.
\end{itemize}
\item[Coassociativity] Coassociativity holds for some initial cases. Precisely, the coassociativity for depth one word $x_n$ for $1\le n < a_q$ and all words with weight $<w_q$ were verified within execution time $t_{q}$ and $t'_{q}$ seconds respectively, for following $q$'s:
\begin{table}[!htbp]
\begin{tabular}{c|cccccccc}
\hline
$q$ & $2$ & $3$ & $5$ & $7$ \\
$(a_q, t_q)$ & $(33,362)$ & $(62,432)$ & $(122,335)$ & $(182, 362)$ \\
$(w_q, t_q')$ & $(15, 2038)$ & $(14, 386)$ & $(18, 578)$ & $(18, 236)$ \\
\hline
$q$ & $8$ & $9$ & $11$ & $13$ \\
$(a_q, t_q)$ & $(300,222)$ & $(302, 183)$ & $(300, 191)$ & $(382, 283)$ \\
$(w_q, t_q')$ & $(18, 207)$ & $(18, 199)$ & $(18, 199)$ & $(18, 218)$ \\
\hline
\end{tabular}
\end{table}
\item[Compatibility] Compatibility holds for some initial cases. Precisely, we verified that $\Delta(w_1 \shuffle w_2) = \Delta(w_1)\Delta(w_2)$ holds for all words $w_1, w_2 \in \langle \Sigma \rangle$ with $\operatorname{weight}(w_1)+ \operatorname{weight}{w_2} \le 12$ within $t_q$ seconds of execution time respectively, where $t_2 = 1965$, $t_3 = 1414$, $t_4 = 450$, $t_5 = 825$, $t_7 = 800$, $t_8 = 365$, $t_9 = 506$, $t_{11} = 793$, $t_{13} = 789$.
Further, when $q=9$, we verified for all words $w_1, w_2 \in \langle \Sigma \rangle$ with $\operatorname{weight}(w_1)+ \operatorname{weight}(w_2) \le 13$ within 2450 seconds of execution time.
\end{description}
\section{Coproduct of depth one} \label{sec: depth one}
This section aims to prove Shi's conjecture on a Hopf algebra structure of the shuffle algebra (see Theorem \ref{thm: comparison with Shi's coproduct}). To do so we study coproduct for words of depth one and deduce that the coproduct $\Delta$ coincides with that introduced by Shi (see Proposition \ref{prop: comparison with Shi's coproduct}). Explicit formulas for coproduct of such words are given in many cases in \S \ref{sec: explicit formula for small weights} and the Appendix \ref{sec: numerical experiments}.
\subsection{Bracket operators} ${}$\par
In the next section we will give a formula for $\Delta(x_n)$ for all $n \in \ensuremath \mathbb{N}$ (see Prop \ref{prop: formula delta xn}). To do so, we need some preparatory results.
\begin{lemma} \label{lem: delta j<n}
Let $n$ be a natural number. We have
\begin{enumerate}
\item for all $j < n$,
\begin{equation*}
\Delta^j_{1,n} = \begin{cases} 1 & \quad \text{if } (q - 1) \mid j \\
0 & \quad \text{otherwise}.
\end{cases}
\end{equation*}
\item $\Delta^n_{1,n} = 0$.
\end{enumerate}
\end{lemma}
\begin{proof}
The result is straightforward from the definition of $\Delta^j_{1,n}$.
\end{proof}
\begin{remark} \label{rmk: delta j<n}
It follows from Lemma \ref{lem: delta j<n} that for all $j,n,m \in \ensuremath \mathbb{N}$ with $j < n < m$, we may identify $\Delta^j_{1,n} = \Delta^{j}_{1,m}$.
\end{remark}
We recall that $\langle \Sigma \rangle$ is the set of all words over $\Sigma = \{x_n\}_{n \in \ensuremath \mathbb{N}}$. Let $\mathfrak{a} = x_{i_1}\dotsb x_{i_r} $ be a non-empty word in $\langle \Sigma \rangle$. We define the \textit{bracket operator} by the following formula:
\begin{equation} \label{eq: bracket operator}
[\mathfrak{a}] := (-1)^{r}\Delta^{i_1}_{1,w(\mathfrak{a}) + 1}\dotsb\Delta^{i_r}_{1,w(\mathfrak{a}) + 1} x_{i_1} \shuffle \dotsb \shuffle x_{i_r}.
\end{equation}
As a matter of convention, we also agree that $[1] = 1$. The following lemmas will be useful.
\begin{lemma} \label{lem: prod of bracket}
Let $\fb$ and $\fc$ be two words in $\langle \Sigma \rangle$. Then we have
\begin{equation*}
[\fb] \shuffle [\fc] = [\fb\fc],
\end{equation*}
where $\fb\fc$ is the concatenation product of $\fb$ and $\fc$.
\end{lemma}
\begin{proof}
The result holds trivially if $\fb = 1$ or $\fc = 1$. We thus assume that $\fb = x_{i_1}\dotsb x_{i_r} $ and $\fc = x_{j_1}\dotsb x_{j_s} $. Then it follows from the definition of the bracket operator and Remark \ref{rmk: delta j<n} that
\begin{align*}
[\fb] &\shuffle [\fc] \\
&= (-1)^{r+s} \Delta^{i_1}_{1,w(\fb)+1} \dotsb \Delta^{i_r}_{1,w(\fb)+1}\Delta^{j_1}_{1,w(\fc)+1} \dotsb \Delta^{j_s}_{1,w(\fc)+1}(x_{i_1}\shuffle \cdots \shuffle x_{i_r}) \shuffle (x_{j_1} \shuffle \cdots \shuffle x_{j_s}) \\
&= (-1)^{r+s} \Delta^{i_1}_{1,w(\fb\fc)+1} \dotsb \Delta^{i_r}_{1,w(\fb\fc)+1}\Delta^{j_1}_{1,w(\fb\fc)+1} \dotsb \Delta^{j_s}_{1,w(\fb\fc)+1}x_{i_1}\shuffle \cdots \shuffle x_{i_r} \shuffle x_{j_1} \shuffle \cdots \shuffle x_{j_s} \\
&= [\fb\fc].
\end{align*}
This proves the lemma.
\end{proof}
\begin{lemma} \label{lem: bracket zero}
Let $\mathfrak{a}$ be a word in $\langle \Sigma \rangle$. If $(q-1) \nmid w(\mathfrak{a})$, then $[\mathfrak{a}] = 0$.
\end{lemma}
\begin{proof}
We may assume that $\mathfrak{a} = x_{i_1}\dotsb x_{i_r}$, so that $w(\mathfrak{a}) = i_1 + \cdots + i_r$. If $(q-1) \nmid w(\mathfrak{a})$, then there exists an index $i_k$ for $1 \leq k \leq r$ such that $(q-1) \nmid i_k$. It follows from Lemma \ref{lem: delta j<n} that $\Delta^{i_k}_{1,w(\mathfrak{a}) +1} = 0$, and hence $[\mathfrak{a}] =0$. This proves the lemma.
\end{proof}
For the convenience of computation, we introduce the following result.
\begin{lemma} \label{lem: delta diamond}
Let $\mathfrak{u}$ and $\mathfrak{v}$ be two non-empty words in $\langle \Sigma \rangle$. Suppose that
\begin{align*}
\Delta(\mathfrak{u}) &= 1 \otimes \mathfrak{u} + \sum \mathfrak{u}_{(1)} \otimes \mathfrak{u}_{(2)},\\
\Delta(\mathfrak{v}) &= 1 \otimes \mathfrak{v} + \sum \mathfrak{v}_{(1)} \otimes \mathfrak{v}_{(2)}.
\end{align*}
Then
\begin{equation*}
\Delta(\mathfrak{u} \diamond \mathfrak{v}) = 1 \otimes (\mathfrak{u} \diamond \mathfrak{v}) + \sum (\mathfrak{u}_{(1)} \diamond \mathfrak{v}_{(1)}) \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}_{(2)}).
\end{equation*}
\end{lemma}
\begin{proof}
From the compatibility, we have
\begin{align*}
\Delta(\mathfrak{u} \shuffle \mathfrak{v}) &= \Delta(\mathfrak{u}) \shuffle \Delta(\mathfrak{v})\\
&= 1 \otimes (\mathfrak{u} \shuffle \mathfrak{v}) + \sum \mathfrak{u}_{(1)} \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}) + \sum \mathfrak{v}_{(1)} \otimes (\mathfrak{u} \shuffle \mathfrak{v}_{(2)}) \\
& + \sum (\mathfrak{u}_{(1)} \shuffle \mathfrak{v}_{(1)}) \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}_{(2)}).
\end{align*}
On the other hand, it follows from Lemma \ref{lem: delta without factor 1} that
\begin{align*}
\Delta(\mathfrak{u} \triangleright \mathfrak{v}) &= 1 \otimes (\mathfrak{u} \triangleright \mathfrak{v}) + \sum \mathfrak{u}_{(1)} \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}) + \sum (\mathfrak{u}_{(1)} \triangleright \mathfrak{v}_{(1)}) \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}_{(2)}),\\
\Delta(\mathfrak{v} \triangleright \mathfrak{u}) &= 1 \otimes (\mathfrak{v} \triangleright \mathfrak{u}) + \sum \mathfrak{v}_{(1)} \otimes (\mathfrak{u} \shuffle \mathfrak{v}_{(2)}) + \sum (\mathfrak{v}_{(1)} \triangleright \mathfrak{u}_{(1)}) \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}_{(2)}).
\end{align*}
Thus
\begin{align*}
\Delta(\mathfrak{u} \diamond \mathfrak{v}) &= \Delta(\mathfrak{u} \shuffle \mathfrak{v} - \mathfrak{u} \triangleright \mathfrak{v} - \mathfrak{v} \triangleright \mathfrak{u})\\
&= \Delta(\mathfrak{u} \shuffle \mathfrak{v}) - \Delta(\mathfrak{u} \triangleright \mathfrak{v}) - \Delta(\mathfrak{v} \triangleright \mathfrak{u})\\
&= 1 \otimes (\mathfrak{u} \shuffle \mathfrak{v} - \mathfrak{u} \triangleright \mathfrak{v} - \mathfrak{v} \triangleright \mathfrak{u}) \\
&+ \sum (\mathfrak{u}_{(1)} \shuffle \mathfrak{v}_{(1)} - \mathfrak{u}_{(1)} \triangleright \mathfrak{v}_{(1)} - \mathfrak{v}_{(1)} \triangleright \mathfrak{u}_{(1)}) \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}_{(2)})\\
&= 1 \otimes (\mathfrak{u} \diamond \mathfrak{v}) + \sum (\mathfrak{u}_{(1)} \diamond \mathfrak{v}_{(1)}) \otimes (\mathfrak{u}_{(2)} \shuffle \mathfrak{v}_{(2)}).
\end{align*}
This proves the lemma.
\end{proof}
\subsection{A formula for the coproduct of depth one} ${}$\par
The first result of this section reads as follows.
\begin{proposition} \label{prop: formula delta xn}
For all $n \in \ensuremath \mathbb{N}$, we have
\begin{equation*}
\Delta(x_n) = 1 \otimes x_n + \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = n}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_r \otimes [\mathfrak{a}].
\end{equation*}
Here we recall that $[\mathfrak{a}]$ is given as in \eqref{eq: bracket operator}.
\end{proposition}
The rest of this section is devoted to a proof of Proposition \ref{prop: formula delta xn}. We proceed the proof by induction on $n$. For $n = 1$, we have
\begin{equation*}
\Delta(x_1) = 1 \otimes x_1 + \binom{-1}{0} x_1 \otimes [1] = 1 \otimes x_1 + x_1 \otimes 1,
\end{equation*}
which proves the base step. We assume that Proposition \ref{prop: formula delta xn} holds for all $n \leq m$ with $m \in \ensuremath \mathbb{N}$ and $m \geq 1$. We need to show that Proposition \ref{prop: formula delta xn} holds for $n = m + 1$. Indeed, from the induction hypothesis, we have
\begin{equation*}
\Delta(x_m) = 1 \otimes x_m + \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_r \otimes [\mathfrak{a}],
\end{equation*}
hence it follows from Lemma \ref{lem: delta diamond} that
\begin{align} \label{eq: delta x_m+1 (1)}
\Delta(x_1 \diamond x_m) &= 1 \otimes(x_1 \diamond x_m) + \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} (x_1 \diamond x_r) \otimes [\mathfrak{a}] \\ \notag
&= 1 \otimes x_{m+1} + \sum \limits_{i+j = m + 1} \Delta^j_{1,m} 1 \otimes x_jx_j\\ \notag
&+ \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_{r+1} \otimes [\mathfrak{a}]\\ \notag
&+ \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} \sum \limits_{h+k = r + 1} \Delta^k_{1,r} x_{h}x_k \otimes [\mathfrak{a}].
\end{align}
On the other hand, we have
\begin{equation} \label{eq: delta x_m+1 (2)}
\Delta(x_1 \diamond x_m) = \Delta(x_{m+1} + \sum \limits_{i+j = m + 1} \Delta^j_{1,m} x_ix_j) = \Delta(x_{m+1}) + \sum \limits_{i+j = m + 1} \Delta^j_{1,m} \Delta(x_ix_j).
\end{equation}
From the induction hypothesis, it follows that for all $i,j \in \ensuremath \mathbb{N}$ such that $i + j = m+1$,
\begin{align*}
\Delta(x_i) = 1 \otimes x_i + \sum \limits_{\substack{s \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ s + w(\fb) = i}} \binom{s + \depth(\fb) - 2}{\depth(\fb)} x_s \otimes [\fb],\\
\Delta(x_j) = 1 \otimes x_j + \sum \limits_{\substack{t \in \ensuremath \mathbb{N}, \fc \in \langle \Sigma \rangle\\ t + w(\fc) = j}} \binom{t + \depth(\fc) - 2}{\depth(\fc)} x_t \otimes [\fc],
\end{align*}
hence
\begin{align} \label{eq: delta x_m+1 (3)}
\Delta(x_ix_j) &= 1 \otimes x_ix_j + \sum \limits_{\substack{s \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ s + w(\fb) = i}} \binom{s + \depth(\fb) - 2}{\depth(\fb)} x_s \otimes ([\fb] \shuffle x_j)\\ \notag
&+ \sum \limits_{\substack{s,t \in \ensuremath \mathbb{N}; \fb,\fc \in \langle \Sigma \rangle\\ s + w(\fb) = i \\ t + w(\fc) = j}} \binom{s + \depth(\fb) - 2}{\depth(\fb)}\binom{t + \depth(\fc) - 2}{\depth(\fc)} x_sx_t \otimes ([\fb] \shuffle [\fc]).
\end{align}
From \eqref{eq: delta x_m+1 (1)}, \eqref{eq: delta x_m+1 (2)} and \eqref{eq: delta x_m+1 (3)}, we have
\begin{align*}
\Delta(x_{m+1}) = \Delta(x_1 \diamond x_m) - \sum \limits_{i+j = m + 1} \Delta^j_{1,m} \Delta(x_ix_j) = 1 \otimes x_{m+1} + S_1 + S_2,
\end{align*}
where
\begin{align*}
S_1 &= \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_{r+1} \otimes [\mathfrak{a}] \\
&- \sum \limits_{i+j = m + 1} \Delta^j_{1,m} \sum \limits_{\substack{s \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ s + w(\fb) = i}} \binom{s + \depth(\fb) - 2}{\depth(\fb)} x_s \otimes ([\fb] \shuffle x_j),\\
S_2 &= \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} \sum \limits_{h+k = r + 1} \Delta^k_{1,r} x_{h}x_k \otimes [\mathfrak{a}] \\
&- \sum \limits_{i+j = m + 1} \Delta^j_{1,m} \sum \limits_{\substack{s,t \in \ensuremath \mathbb{N}; \fb,\fc \in \langle \Sigma \rangle\\ s + w(\fb) = i \\ t + w(\fc) = j}} \binom{s + \depth(\fb) - 2}{\depth(\fb)}\binom{t + \depth(\fc) - 2}{\depth(\fc)} x_sx_t \otimes ([\fb] \shuffle [\fc]).
\end{align*}
We next compute the sums $S_1$ and $S_2$ as follows.
\subsubsection{The sum $S_1$} We first note that if $j <m$ then it follows from Remark \ref{rmk: delta j<n} and Lemma \ref{lem: prod of bracket} that $(-\Delta^j_{1,m})([\fb] \shuffle x_j) = [\fb] \shuffle [x_j] = [\fb x_j]$. Thus we have
\begin{align*}
S_1 &= \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_{r+1} \otimes [\mathfrak{a}] \\
&- \sum \limits_{i+j = m + 1} \Delta^j_{1,m} \sum \limits_{\substack{s \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ s + w(\fb) = i}} \binom{s + \depth(\fb) - 2}{\depth(\fb)} x_s \otimes ([\fb] \shuffle x_j)\\
&= \sum \limits_{r=2}^{m+1} \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \binom{r - 1 + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_{r} \otimes [\mathfrak{a}] \\
&+ \sum \limits_{s=1}^{m} \sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ s+ w(\fb) + j= m+1}} \binom{s + \depth(\fb) - 2}{\depth(\fb)} x_s \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j)\\
&= \binom{m-2}{0} x_{m+1} \otimes [1] + \sum \limits_{r=2}^{m} \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \binom{r - 1 + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_{r} \otimes [\mathfrak{a}]\\
&+ \sum \limits_{s=2}^{m} \sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ s+ w(\fb) + j= m+1}} \binom{s + \depth(\fb) - 2}{\depth(\fb)} x_s \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j) \\
&+ \sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ w(\fb) + j= m}} \binom{\depth(\fb) - 1}{\depth(\fb)} x_1 \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j)\\
&= x_{m+1} \otimes 1 + \Bigg[\sum \limits_{r=2}^{m} \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \binom{r + \depth(\mathfrak{a}) - 3}{\depth(\mathfrak{a})} x_{r} \otimes [\mathfrak{a}] \\
&+ \sum \limits_{s=2}^{m} \sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ s+ w(\fb) + j= m+1}} \binom{s + \depth(\fb) - 2}{\depth(\fb)} x_s \otimes [\fb x_j]\Bigg]
\\
&+ \sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ w(\fb) + j= m}} \binom{\depth(\fb) - 1}{\depth(\fb)} x_1 \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j)\\
&= x_{m+1} \otimes 1 + \sum \limits_{r=2}^{m} \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \Bigg[\binom{r + \depth(\mathfrak{a}) - 3}{\depth(\mathfrak{a})} + \binom{r + \depth(\mathfrak{a}) - 3}{\depth(\mathfrak{a})-1} \Bigg]x_{r} \otimes [\mathfrak{a}]
\\
&+ \sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ w(\fb) + j= m}} \binom{\depth(\fb) - 1}{\depth(\fb)} x_1 \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j)\\
&= x_{m+1} \otimes 1 + \sum \limits_{r=2}^{m} \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})}x_{r} \otimes [\mathfrak{a}] \\
&+ \sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ w(\fb) + j= m}} \binom{\depth(\fb) - 1}{\depth(\fb)} x_1 \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j).
\end{align*}
We claim that for all $j \in \ensuremath \mathbb{N}$ and for all $\fb \in \langle \Sigma \rangle$ with $j + w(\fb) = m$,
\begin{equation} \label{eq: S1 zero}
\binom{\depth(\fb) - 1}{\depth(\fb)} x_1 \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j) = 0.
\end{equation}
Indeed, if $j = m$, then it follows from Lemma \ref{lem: delta j<n} that $\Delta^m_{1,m} =0$, hence \eqref{eq: S1 zero} holds. If $j <m$, then $w(\fb) \geq 1$, i.e., $\depth(\fb) \geq 1$, hence $\binom{\depth(\fb) - 1}{\depth(\fb)} = 0$, showing that \eqref{eq: S1 zero} holds.
This proves the claim. As a consequence, one may identify
\begin{equation*}
\sum \limits_{\substack{j \in \ensuremath \mathbb{N}, \fb \in \langle \Sigma \rangle\\ w(\fb) + j= m}} \binom{\depth(\fb) - 1}{\depth(\fb)} x_1 \otimes (-\Delta^j_{1,m})([\fb] \shuffle x_j) = \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ w(\mathfrak{a}) = m}} \binom{\depth(\mathfrak{a}) - 1}{\depth(\mathfrak{a})}x_{1} \otimes [\mathfrak{a}] = 0.
\end{equation*}
Thus
\begin{align*}
S_1 &= x_{m+1} \otimes [1] + \sum \limits_{r=2}^{m} \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})}x_{r} \otimes [\mathfrak{a}] \\
&+ \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ w(\mathfrak{a}) = m}} \binom{\depth(\mathfrak{a}) - 1}{\depth(\mathfrak{a})}x_{1} \otimes [\mathfrak{a}]\\
&= \sum \limits_{r=1}^{m+1} \sum \limits_{\substack{\mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})}x_{r} \otimes [\mathfrak{a}]\\
&= \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m +1}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})}x_{r} \otimes [\mathfrak{a}].
\end{align*}
\subsubsection{The sum $S_2$} We claim that $S_2 = 0$. The following lemma will be useful.
\begin{lemma} \label{lem: comb indentity}
For positive integers $h,k,l$, we have;
\begin{equation*}
\sum \limits_{\substack{i , j \geq 0 \\i +j = l}} \binom{h + i}{i} \binom{k + j}{j} = \binom{h + k + l + 1}{l}.
\end{equation*}
\end{lemma}
\begin{proof}
For all $m ,n \in \ensuremath \mathbb{N}$, we have
\begin{equation*}
\sum \limits_{r = 0}^m \binom{n + r}{r} = \binom{n + m + 1}{m}.
\end{equation*}
We now proceed the proof by induction on $k$. The base step $k = 0$ follows from the above identity. We assume that Lemma \ref{lem: comb indentity} holds for $k \geq 0$. We need to show that Lemma \ref{lem: comb indentity} holds for $k + 1$. From the previous identity and the induction hypothesis, we have
\begin{align*}
\sum \limits_{\substack{i , j \geq 0 \\i +j = l}} \binom{h + i}{i} \binom{(k + 1) + j}{j} &= \sum \limits_{\substack{i , j \geq 0 \\i +j = l}} \binom{h + i}{i} \binom{ k + j + 1}{j}\\
&= \sum \limits_{\substack{i , j \geq 0 \\i +j = l}} \binom{h + i}{i} \sum \limits_{r =0}^j \binom{k + r}{r} \\
&= \sum \limits_{s = 0}^l \sum \limits_{\substack{i , r \geq 0 \\i +r = s}} \binom{h + i}{i}\binom{k + r}{r}\\
&= \sum \limits_{s = 0}^l \binom{h + k + s + 1}{s}\\
&= \binom{h + k + l + 2}{l}.
\end{align*}
This proves the lemma.
\end{proof}
It follows from Lemma \ref{lem: prod of bracket} that $[\fb] \shuffle [\fc] = [\fb\fc]$, hence
\begin{align*}
S_2&= \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} \sum \limits_{h+k = r + 1} \Delta^k_{1,r} x_{h}x_k \otimes [\mathfrak{a}] \\
&- \sum \limits_{i+j = m + 1} \Delta^j_{1,m} \sum \limits_{\substack{s,t \in \ensuremath \mathbb{N}; \fb,\fc \in \langle \Sigma \rangle\\ s + w(\fb) = i \\ t + w(\fc) = j}} \binom{s + \depth(\fb) - 2}{\depth(\fb)}\binom{t + \depth(\fc) - 2}{\depth(\fc)} x_sx_t \otimes [\fb \fc]\\
&= \sum \limits_{\substack{h,k \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ h + k + w(\mathfrak{a}) = m + 1}} \binom{h +k - 1 + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} \Delta^k_{1,h + k - 1} x_{h}x_k \otimes [\mathfrak{a}] \\
&- \sum \limits_{\substack{s,t \in \ensuremath \mathbb{N}; \fb,\fc \in \langle \Sigma \rangle\\ s + t + w(\fb) +w(\fc) = m + 1}} \binom{s + \depth(\fb) - 2}{\depth(\fb)}\binom{t + \depth(\fc) - 2}{\depth(\fc)} \Delta^{t + w(\fc)}_{1,m} x_sx_t \otimes [\fb \fc]\\
&= \sum \limits_{\substack{h,k \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ h + k + w(\mathfrak{a}) = m + 1}} \Bigg[ \binom{h +k + \depth(\mathfrak{a}) - 3}{\depth(\mathfrak{a})} \Delta^k_{1,h + k - 1} \\
&- \sum \limits_{\substack{ \fb,\fc \in \langle \Sigma \rangle\\ \fb \fc = \mathfrak{a}}} \binom{h + \depth(\fb) - 2}{\depth(\fb)}\binom{k + \depth(\fc) - 2}{\depth(\fc)}\Delta^{k + w(\fc)}_{1,m} \Bigg] x_{h}x_k \otimes [\mathfrak{a}]
\end{align*}
We will prove that for all $h, k \in \ensuremath \mathbb{N}$ and for all $\mathfrak{a} \in \langle \Sigma \rangle$ with $h + k + w(\mathfrak{a}) = m + 1$,
\begin{equation} \label{eq: S2 zero}
\Bigg[ \binom{h +k + \depth(\mathfrak{a}) - 3}{\depth(\mathfrak{a})} \Delta^k_{1,h + k - 1} - \sum \limits_{\substack{ \fb,\fc \in \langle \Sigma \rangle\\ \fb \fc = \mathfrak{a}}} \binom{h + \depth(\fb) - 2}{\depth(\fb)}\binom{k + \depth(\fc) - 2}{\depth(\fc)}\Delta^{k + w(\fc)}_{1,m} \Bigg] x_{h}x_k \otimes [\mathfrak{a}] = 0.
\end{equation}
If this is the case then $S_2 = 0$. We divide into three cases:\\
\noindent \textbf{Case 1: $ h = 1$}.
The LHS of \eqref{eq: S2 zero} equals
\begin{equation*}
\Bigg[ \binom{k + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} \Delta^k_{1,k} - \sum \limits_{\substack{ \fb,\fc \in \langle \Sigma \rangle\\ \fb \fc = \mathfrak{a}}} \binom{\depth(\fb) - 1}{\depth(\fb)}\binom{k + \depth(\fc) - 2}{\depth(\fc)}\Delta^{k + w(\fc)}_{1,m} \Bigg] x_{1}x_k \otimes [\mathfrak{a}].
\end{equation*}
It follows from Lemma \ref{lem: delta j<n} that $\Delta^k_{1,k} =0$. Moreover, if $\fb \ne 1$, then $\depth(\fb) \geq 1$, hence $\binom{\depth(\fb) - 1}{\depth(\fb)} = 0$. If $\fb = 1$, then $\fc = \mathfrak{a}$ hence $k + w(\fc) = k + w(\mathfrak{a}) = m$, showing that $\Delta^{k + w(\fc)}_{1,m} = \Delta^m_{1,m} = 0$. So \eqref{eq: S2 zero} holds in this case.\\
\noindent\textbf{Case 2: $ h \geq 2, k = 1$}. The LHS of \eqref{eq: S2 zero} equals
\begin{equation*}
\Bigg[ \binom{h + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} \Delta^1_{1,h} - \sum \limits_{\substack{ \fb,\fc \in \langle \Sigma \rangle\\ \fb \fc = \mathfrak{a}}} \binom{h +\depth(\fb) - 2}{\depth(\fb)}\binom{\depth(\fc) - 1}{\depth(\fc)}\Delta^{1 + w(\fc)}_{1,m} \Bigg] x_{h}x_1 \otimes [\mathfrak{a}].
\end{equation*}
Since $ 1 < h \leq m$, it follows from Remark \ref{rmk: delta j<n} that $\Delta^1_{1,h} = \Delta^1_{1,m}$. Moreover, if $\fc \ne 1$, then $\depth(\fc) \geq 1$, hence $\binom{\depth(\fc) - 1}{\depth(\fc)} = 0$. Then the LHS of \eqref{eq: S2 zero} equals
\begin{equation*}
\Bigg[ \binom{h + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} \Delta^1_{1,m} - \binom{h +\depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})}\binom{- 1}{0}\Delta^{1}_{1,m} \Bigg] x_{h}x_1 \otimes [\mathfrak{a}] = 0.
\end{equation*}
So \eqref{eq: S2 zero} holds in this case.\\
\noindent\textbf{Case 3: $ h \geq 2, k \geq 2$} Since $k < h + k - 1 \leq m$, it follows from Remark \ref{rmk: delta j<n} that $\Delta^k_{1,h+k - 1} = \Delta^k_{1,m}$. Moreover, we claim that
\begin{equation} \label{eq: S2 case 3}
\Delta^{k+w(\fc)}_{1,m} x_hx_k \otimes [\mathfrak{a}] = \Delta^{k}_{1,m} x_hx_k \otimes [\mathfrak{a}].
\end{equation}
Indeed, note that $k + w(\fc) < m$. If $(q-1) \mid w(\fc)$, then it follows from Lemma $\ref{lem: delta j<n}$ that $\Delta^{k+w(\fc)}_{1,m} = \Delta^{k}_{1,m}$, hence \eqref{eq: S2 case 3} holds. If $(q-1) \nmid w(\fc)$, then it follows from Lemma \ref{lem: bracket zero} that $[\fc] = 0$, hence $[\mathfrak{a}] = 0$, showing that \eqref{eq: S2 case 3} holds. Thus the LHS of \eqref{eq: S2 zero} becomes
\begin{equation*}
\Bigg[ \binom{h +k + \depth(\mathfrak{a}) - 3}{\depth(\mathfrak{a})} - \sum \limits_{\substack{ \fb,\fc \in \langle \Sigma \rangle\\ \fb \fc = \mathfrak{a}}} \binom{h + \depth(\fb) - 2}{\depth(\fb)}\binom{k + \depth(\fc) - 2}{\depth(\fc)} \Bigg] \Delta^{k}_{1,m} x_{h}x_k \otimes [\mathfrak{a}].
\end{equation*}
It follows from Lemma \ref{lem: comb indentity} that
\begin{equation*}
\binom{h +k + \depth(\mathfrak{a}) - 3}{\depth(\mathfrak{a})} = \sum \limits_{\substack{ \fb,\fc \in \langle \Sigma \rangle\\ \fb \fc = \mathfrak{a}}} \binom{h + \depth(\fb) - 2}{\depth(\fb)}\binom{k + \depth(\fc) - 2}{\depth(\fc)},
\end{equation*}
hence \eqref{eq: S2 zero} holds in this case.\\
From the above computations, we conclude that
\begin{equation*}
\Delta(x_{m+1}) = 1 \otimes x_{m+1} + S_1 + S_2 = 1 \otimes x_{m+1} + \sum \limits_{\substack{r \in \ensuremath \mathbb{N}, \mathfrak{a} \in \langle \Sigma \rangle\\ r + w(\mathfrak{a}) = m + 1}} \binom{r + \depth(\mathfrak{a}) - 2}{\depth(\mathfrak{a})} x_r \otimes [\mathfrak{a}].
\end{equation*}
This proves Proposition \ref{prop: formula delta xn}.
\subsection{Comparison with Shi's coproduct} ${}$\par
In \cite{Shi18}, Shi defined another coproduct
\[ \Delta_1: \frak C \to \frak C \otimes \frak C. \]
using the concatenation rather than $\triangleright$ on recursive steps for words with depth $>1$. More precisely, we define it on $\langle \Sigma \rangle$ by induction on weight and extend by $\mathbb F_q$-linearity to $\frak C$. First, we set
\begin{align*}
\Delta_1(1)&:=1 \otimes 1, \\
\Delta_1(x_1)&:=1 \otimes x_1 + x_1 \otimes 1.
\end{align*}
Let $w \in \mathbb N$ and we suppose that we have defined $\Delta_1(\fv)$ for all words $\fv$ of weight $w(\fv)<w$. We now give a formula for $\Delta_1(\fu)$ for all words $\fu$ with $w(\fu)=w$. For such a word $\fu$ with $\depth(\fu)>1$, we put $\fu=x_u \fv$ with $w(\fv)<w$. Since $x_u$ and $\fv$ are both of weight less than $w$, we have already defined
\begin{align*}
\Delta_1(x_u)&:=1 \otimes x_u + \sum a_u \otimes b_u, \\
\Delta_1(\fv)&:= \sum a_\fv \otimes b_\fv.
\end{align*}
Then we set
\begin{align*}
\Delta_1(\fu):=1 \otimes \fu + \sum (a_u a_\fv) \otimes (b_u \shuffle b_\fv).
\end{align*}
Our last task is to define $\Delta_1(x_w)$. We know that
\[ x_1\shuffle x_{w-1}=x_w+x_1x_{w-1}+x_{w-1}x_1+\sum_{0<j<w} \Delta^j_{1,w-1} x_{w-j} x_j \]
where all the words $x_{w-j} x_j$ have weight $w$ and depth $2$ and all $\Delta^j_{1,w-1}$ belong to $\mathbb F_q$. Therefore, we set
\begin{equation*}
\Delta_1(x_w):=\Delta_1(x_1) \shuffle \Delta_1(x_{w-1})-\Delta_1(x_1x_{w-1})-\Delta_1(x_{w-1}x_1)-\sum_{0<j<w} \Delta^j_{1,w-1} \Delta_1(x_{w-j} x_j).
\end{equation*}
As an application of Proposition \ref{prop: formula delta xn}, we prove:
\begin{proposition} \label{prop: comparison with Shi's coproduct}
For all words $\fu \in \langle \Sigma \rangle$, we have
\[ \Delta(\fu)=\Delta_1(\fu). \]
\end{proposition}
\begin{proof}
The proof is by induction on the weight $w$. We have to show that for all words $\fu$ of weight $w$,
\[ \Delta(\fu)=\Delta_1(\fu). \]
We denote this claim by $H_w$.
For $w=0$ and $w=1$, we are done as $\Delta(1)=\Delta_1(1)=1 \otimes 1$ and $\Delta(x_1)=\Delta_1(x_1)=1 \otimes x_1 + x_1 \otimes 1$. Suppose that for all words $\fu$ with $w(\fu)<w$, we have $\Delta(\fu)=\Delta_1(\fu)$. We will show that the claim $H_w$ holds.
Let $\fu$ be a word of weight $w$. Suppose that the depth of $\fu$ is at least 2. Then we put $\fu=x_u \fv$ with $w(\fv)<w$. By induction, we know that $\Delta(\fv)=\Delta_1(\fv)$. If $\Delta(x_u) = \sum \mathfrak{a}_u \otimes \mathfrak{b}_u$, then Proposition \ref{prop: formula delta xn} implies that $\depth(\mathfrak{a_n})\le1$. Thus $\Delta(\fu)=\Delta_1(\fu)$.
To conclude, we have to check the claim for $u=x_w$. By induction, for all $i<w$, $\Delta(x_i)=\Delta_1(x_i)$. It follows that
\begin{align*}
& \Delta(x_w) \\
&=\Delta(x_1) \shuffle \Delta(x_{w-1})-\Delta(x_1x_{w-1})-\Delta(x_{w-1}x_1)-\sum_{0<j<w} \Delta^j_{1,w-1} \Delta(x_{w-j} x_j) \\
&=\Delta_1(x_1) \shuffle \Delta_1(x_{w-1})-\Delta_1(x_1x_{w-1})-\Delta_1(x_{w-1}x_1)-\sum_{0<j<w} \Delta^j_{1,w-1} \Delta_1(x_{w-j} x_j) \\
&=\Delta_1(x_w).
\end{align*}
Thus we have proved $H_w$. The proposition follows.
\end{proof}
In particular, we get
\begin{theorem} \label{thm: comparison with Shi's coproduct}
Conjecture 3.2.11 in \cite{Shi18} holds.
\end{theorem}
\subsection{Auxiliary results} ${}$\par
In this section we prove some auxiliary results that will be useful in the sequel.
We define
\begin{align*}
x_{s_1}x_{s_2}\dots x_{s_r} \star q &:= x_{qs_1} x_{qs_2}\dots x_{q s_r}, \\
(\mathfrak {u}\otimes \mathfrak {v}) \star q &:= (\mathfrak {u}\star q) \otimes (\mathfrak {v}\star q)
\end{align*}
for $\mathfrak {u}, \mathfrak {v}\in \langle \Sigma \rangle$, and extend it to be $\mathbb{F}_p$-linear. We define $1 \star q = 1$ when $1$ is the empty word.
We recall the Lucas's theorem \cite{Gra97}: \[\binom{a_0+ a_1 p +a_2 p^2+ \dots +a_k p^k}{b_0 + b_1 p + b_2 p^2+ \dots +b_k p^k} \equiv \binom{a_0}{b_0} \binom{a_1}{b_1} \binom{a_2}{b_2}\dots \binom{a_k}{b_k} \pmod{p}, \] where $p$ is a prime and $0\le a_i, b_i<p$.
\begin{lemma}\label{lemma: modular properties on chen delta} The following modular equations hold.
\begin{enumerate}[$(1)$]
\item $\binom{i-1}{a-1}\equiv 0 \pmod{p}$ when $q\nmid i$, $q\mid a$. $\Delta_{a, b}^i \equiv 0\pmod p$ follows when $q\mid a, b$.
\item $\binom{pj-1}{pa-1} \equiv \binom{j-1}{a-1} \pmod{p}$ and $\Delta_{pa, pb}^{pj} \equiv \Delta_{a, b}^j \pmod{p}$ for $0<j<a+b$.
\end{enumerate}
\end{lemma}
\begin{proof} Let $q = p^k$.
To prove (1), let $i = Aq +B$ with $A, B \in \mathbb{N}$, $0<B<q$. We write \[i-1 = Aq + (B-1) =(\beta_0 + \beta_1 p + \dots + \beta_{k-1}p^{k-1}) + \sum_{s=k}^{r}\alpha_s p^s,\] and \[ a-1 = \left( (p-1)+ (p-1)p + \dots + (p-1)p^{k-1}\right) + \sum_{s=k}^{r}\alpha'_s p^s\]with $0 \le \alpha_s, \alpha'_s,\beta_s<p$. By Lucas's theorem,
\begin{align*}
\binom{i-1}{a-1} &\equiv \binom{\beta_0}{p-1}\binom{\beta_1}{p-1}\cdots \binom{\beta_{k-1}}{p-1} \cdot \binom{\alpha_1}{\alpha'_1}\binom{\alpha_2}{\alpha'_2}\cdots \equiv 0 \pmod{p},
\end{align*}
since at least one of $\beta_s < p-1$. (Note that $B<q$.)
For (2), $\binom{pj-1}{pa-1} \equiv \binom{j-1}{a-1} \pmod{p}$ is verified by similar routine calculations with Lucas's theorem.
\end{proof}
\begin{lemma} \label{lemma: star product raised by q}
For all words $\frak u, \frak v \in \frak C$, we have
$(\mathfrak {u}\shuffle \mathfrak{v})\star q = (\mathfrak{u}\star q)\shuffle (\mathfrak{v} \star q)$. Equivalently,
if $\mathfrak{u} \shuffle \mathfrak{v} = \sum \mathfrak{a}_i$, then $(\mathfrak{u}\star q) \shuffle (\mathfrak{v}\star q) = \sum (\mathfrak{a}_i\star q)$.
\end{lemma}
\begin{proof}
Let $\mathfrak{u} = x_u$, $\mathfrak{v} = x_v$ be depth one words. We have \begin{align*}
x_u\shuffle x_v &= x_{u+v}+ x_u+x_v + \sum_{0<i<u+v}\Delta^{i}_{u, v}x_{u+v-i}x_i, \\
x_{qu}\shuffle x_{qv} &= x_{q(u+v)}+ x_{qu}+x_{qv} + \sum_{0<i<q(u+v)}\Delta^{i}_{qu, qv}
x_{q(u+v)-i}x_i.
\end{align*}
By Lemma~\ref{lemma: modular properties on chen delta},
\begin{align*}
\sum_{0<i<q(u+v)}\Delta^{i}_{qu, qv}
x_{q(u+v)-i}x_i & = \sum_{q\mid i, \ 0<i<q(u+v)}\Delta^{i}_{qu, qv}x_{q(u+v)-i}x_i\\
& =
\sum_{0<j<u+v}\Delta^{qj}_{qu, qv}x_{q(u+v-j)}x_{qj}\\
&=\sum_{0<j<u+v}\Delta^{qj}_{u, v}x_{q(u+v-j)}x_{qj}.
\end{align*}
Thus the lemma is proved for depth 1 words $\mathfrak{u}, \mathfrak{v}$. By the induction on $\depth(\mathfrak u) + \depth(v)$, we can conclude the general result.
\end{proof}
We recall that
$\Delta(1) = 1\otimes 1$,
$\Delta(x_1) = 1\otimes x_1 + x_1 \otimes 1$ by definition, and
$\Delta(\fu) \in \frak C \otimes \frak C$ is defined for other all words $\fu$.
\begin{proposition} \label{proposition: Hopf deltas depth 1 <= q}
For all integers $1\le n \le q$,
\begin{equation} \label{eq: Hopf deltas depth 1 <= q}
\Delta(x_n) = 1\otimes x_n + x_n \otimes 1.
\end{equation}
\end{proposition}
\begin{proof}
\eqref{eq: Hopf deltas depth 1 <= q} holds when $n=1$. Let $2\le n \le q$, and assume \eqref{eq: Hopf deltas depth 1 <= q} holds for $1, 2,\dots, n-1$. Then,
\begin{align*}
x_1\shuffle x_{n-1} =x_1x_{n-1} +x_{n-1}x_{1} + x_n + \sum_{\substack{q-1 \mid j \\ 0<j< n}}\Delta^j_{1,n-1}x_{n-j}x_{j}.
\end{align*}
The $\sum$ term here is empty sum for $n<q$, and when $n=q$, $j=q-1$ term is the only possible index. $\Delta_{1, q-1}^{q-1} = \binom{q-2}{q-2} + (-1)^q \binom{q-2}{q-2}$ vanishes for all $q$, so
\begin{align*}
x_1\shuffle x_{n-1} =x_{1}x_{n-1} +x_{n-1}x_{1} + x_n,
\end{align*}
that is, $\Delta(x_1)\Delta(x_{n-1}) = \Delta(x_{1}x_{n-1}) + \Delta(x_{n-1}x_1) + \Delta(x_n)$.
By calculation and the induction hypothesis, we have
\begin{align*}
\Delta(x_1)\Delta(x_{n-1}) &= (1\otimes x_1 + x_1\otimes 1)(1\otimes x_{n-1} + x_{n-1}\otimes 1)\\
& =1\otimes \left(x_1\shuffle x_{n-1}\right) + x_{n-1}\otimes x_1 + x_1\otimes x_{n-1} + \left( x_1\shuffle x_{n-1}\right) \otimes 1,\\
\Delta(x_1x_{n-1}) & = e\otimes x_1x_{n-1} + x_1\otimes x_{n-1} + x_1 x_{n-1}\otimes 1, \\
\Delta(x_{n-1}x_1) & = 1\otimes x_{n-1}x_1 + x_{n-1}\otimes x_1 + x_{n-1}x_1\otimes 1
\end{align*}
Recall $x_1\shuffle x_{n-1} =x_{1}x_{n-1} +x_{n-1}x_{1} + x_n$ to have
\begin{align*}
\Delta(x_n) = 1\otimes x_n + x_n\otimes 1.
\end{align*}
\end{proof}
\begin{proposition}\label{proposition: Hopf coproduct raised by q}
For all $n \in \mathbb N$, we have
\[ \Delta (x_{qn}) = \Delta(x_n)\star q \] for all $n\in \mathbb{N}$. This can be stated as follows:
letting $\Delta(x_n) = e\otimes x_n + \sum a_n \otimes b_n$ we get $\Delta(x_{qn}) = e\otimes x_{qn} + \sum_i (a_n\star q) \otimes (b_n \star q)$.
\end{proposition}
\begin{proof}
The statement holds for $n=1$ by Proposition~\ref{proposition: Hopf deltas depth 1 <= q}.
Assume that the statement holds for all $n\le w$. Say $\Delta(x_w) = 1\otimes x_w + \sum a_w \otimes b_w$. Since
\begin{align*}
x_w\shuffle x_1 &= x_{w+1} + x_w x_1 + x_1x_w + \sum_{0 < j < w+1}\Delta_{1, w}^j x_{w+1-j}x_{j}, \\
\Delta(x_w x_1)& = 1\otimes x_w x_1 + \sum (a_{w}\triangleright x_1) \otimes b_w+ \sum a_{w}\otimes (b_{w}\shuffle x_1), \\
\Delta(x_1 x_w)& = 1\otimes x_1 x_w + x_1 \otimes x_w + \sum (x_1 \triangleright a_{w}) \otimes b_w,
\end{align*}
we have
\begin{align*}
& \Delta(x_{w+1}) \\
& = 1\otimes (x_1 \shuffle x_w) + x_1 \otimes x_w + \sum a_w \otimes (b_w \shuffle x_1) + \sum (a_w \shuffle x_1)\otimes b_w \\
& - 1 \otimes x_w x_1 - \sum (a_w \triangleright x_1)\otimes b_w - \sum a_{w}\otimes (b_{w}\shuffle x_1)\\
& - 1\otimes x_1 x_w - x_1 \otimes x_w - \sum (x_1 \triangleright a_{w}) \otimes b_w\\
& - \sum_{0 < j < w+1}\Delta_{1, w}^j \Delta(x_{w+1-j}x_{j})
\end{align*}
By the induction hypothesis, $\Delta (x_{qw}) = 1\otimes x_{qw} + \sum (a_w\star q) \otimes (b_w\star q)$. Note that $\Delta (x_q) = 1 \otimes x_q + x_q \otimes 1$ by Proposition~\ref{proposition: Hopf deltas depth 1 <= q} and since we proved the compatibility for $\Delta$,
\begin{align*}
x_{qw}\shuffle x_q &= x_{q(w+1)} + x_{qw} x_q + x_q x_{qw} + \sum_{0 < j < q(w+1)}\Delta_{q, qw}^j x_{qw+q-j}x_{j}, \\
\Delta(x_{qw} x_q)& = 1\otimes x_{qw} x_q + \sum ((a_{w}\star q)\triangleright x_q) \otimes (b_w\star q)+ \sum (a_{w}\star q)\otimes ((b_{w}\star q)\shuffle x_q), \\
\Delta(x_q x_{qw})& = 1\otimes x_q x_{qw} + x_q \otimes x_{qw} + \sum (x_q \triangleright (a_{w}\star q)) \otimes (b_w\star q),
\end{align*}
we have
\begin{align*}
& \Delta(x_{qw+q}) \\
& = 1\otimes (x_q \shuffle x_{qw}) + x_q \otimes x_{qw} + \sum (a_w\star q)\otimes ((b_w\star q) \shuffle x_q) + \sum ((a_w\star q) \shuffle x_q)\otimes (b_w\star q) \\
& -1\otimes x_{qw} x_q - \sum ((a_{w}\star q)\triangleright x_q) \otimes (b_w\star q)- \sum (a_{w}\star q)\otimes ((b_{w}\star q)\shuffle x_q) \\
& - 1\otimes x_q x_{qw} - x_q \otimes x_{qw} - \sum (x_q \triangleright (a_{w}\star q)) \otimes (b_w\star q) \\
& - \sum_{0 < j < q(w+1)}\Delta_{q, qw}^j \Delta(x_{qw+q-j}x_{j}).
\end{align*}
With Lemma~\ref{lemma: star product raised by q}, we are done if $\Delta((x_{w+1-i}x_i)\star q) = \Delta(x_{w+1-i}x_i)\star q$ for $i \le w$, which can be shown by applying the induction hypothesis.
\end{proof}
\subsection{Explicit formula for $\Delta(x_n)$ with $1\le n \le q^2$} ${}$\par \label{sec: explicit formula for small weights}
In this section we give an explicit formula for $\Delta(x_n)$ with $1\le n \le q^2$. It could be obtained as an application of Proposition \ref{prop: formula delta xn}. We will give below another way to do calculations.
For $1 \leq n \leq q$, an explicit formula for $\Delta(x_n)$ is given in Proposition \ref{proposition: Hopf deltas depth 1 <= q}:
\[ \Delta(x_n) = 1\otimes x_n + x_n \otimes 1. \]
We note that a direct consequence of Propositions \ref{proposition: Hopf deltas depth 1 <= q} and \ref{proposition: Hopf coproduct raised by q} implies
\[ \Delta(x_{aq^r}) = 1\otimes x_{a q^r} + x_{a q^r}\otimes 1 \]
for all $1\le a < q$. By Theorem \ref{thm: compatibility}, we have an algorithm to calculate $\Delta(x_{n})$:
\begin{itemize}
\item Write $n = n_0 + n_1q + \dots + n_r q^r$, with $0\le n_i < q$. Then $\Delta(x_{n_i q^i}) = 1\otimes x_{n_i q^i} + x_{n_i q^i} \otimes 1$ by Proposition~\ref{proposition: Hopf deltas depth 1 <= q}.
\item Use compatibility result to calculate $\Delta(x_n)$, by calculating $\Delta(x_{n_0})\shuffle \Delta(x_{n_1q})\shuffle \dots \shuffle \Delta(x_{n_r q^r}) = \Delta(x_{n_0}\shuffle x_{n_1 q}\shuffle \dots \shuffle x_{n_r q^r})$.
\end{itemize}
\begin{lemma}\label{lemma: Hopf delta for (q-1) multiples}
For $2\le k \le q$, we have
\[\Delta(x_{k(q-1)}) = 1 \otimes x_{k(q-1)} + x_{k(q-1)} \otimes 1 + \sum_{i=1}^{k-1} \binom{k}{i} x_{i(q-1)} \otimes x_{(k-i)(q-1)}.\]
\end{lemma}
\begin{proof}
The lemma follows from Lemma \ref{lemma: star product of q-1 multiples} and the compatibility equation (see Theorem \ref{thm: compatibility}) \[\Delta(x_{(k-1)(q-1)}\shuffle x_{q-1}) = \Delta(x_{(k-1)(q-1)})\shuffle \Delta(x_{q-1}).\]
\end{proof}
\begin{lemma} \label{lemma: star product of q-1 multiples}
For all $a,b \in \mathbb N$ with $a+b\le q$, we have
\begin{equation}x_{a(q-1)}\shuffle x_{b(q-1)} = \underbrace{x_{q-1}\shuffle \dots \shuffle x_{q-1}}_{a+b} = x_{(a+b)(q-1)}.
\label{eqn: star product of q-1 multiples}
\end{equation}
\end{lemma}
\begin{proof}
It suffices to prove the following claim: for $2\le k \le q$, we have $x_{(k-1)(q-1)}\shuffle x_{q-1} = x_{k(q-1)}.$ In fact, with Lucas's theorem, one can verify that $\Delta_{q-1,k(q-1)}^{k(q-1)} =-2$ when $k=1$, and $\Delta_{q-1,k(q-1)}^{q-1} \equiv \Delta_{q-1, k(q-1)}^{k(q-1)}\equiv -1 \pmod{p}$ and $\Delta_{q-1,k(q-1)}^{i(q-1)} \equiv 0\pmod p$ when $k>1$. Thus the claim follows from the definition of $\shuffle $.
\end{proof}
\begin{remark} \label{remark: q^2-1 Hopf delta}
Note that the Lemma~\ref{lemma: Hopf delta for (q-1) multiples} does not answer for $\Delta(x_{q^2-1})$. By similar calculation for $x_{q(q-1)}\shuffle x_{q-1}$, all lines except $\Delta_{q-1, q(q-1)}^{q(q-1)} =0$ are parallel, which yields
\[ x_{q(q-1)} \shuffle x_{q-1} = x_{q^2-1} + x_{q-1}x_{q(q-1)}.\]
Similar calculation gives \[ \Delta(x_{q^2-1}) = 1 \otimes x_{q^2-1} + x_{q^2-1}\otimes 1 + x_{q(q-1)}\otimes x_{q-1}.\]
\end{remark}
\begin{lemma}\label{lemma: Aq star B(q-1)}
Let $1\le A, B \le q-1$ with $A + B \le q$. Then $$x_{Aq}\shuffle x_{B(q-1)} = x_{Aq + B(q-1)}+ x_{B(q-1)}x_{Aq}.$$
\end{lemma}
\begin{proof}
Let $ n = Aq + B(q-1) = (A+B)(q-1) +A$. Since $1< A \le q-1$,
\begin{align*}x_{Aq}\shuffle x_{B(q-1)} & = x_{n} + x_{Aq}x_{B(q-1)} + x_{B(q-1)} x_{Aq}+ S_1 + S_2
\end{align*}
where
\begin{align*}
S_1 & =
(-1)^{A-1}\sum_{i = 1}^{A+B} \binom{i(q-1)-1 }{Aq-1} x_{n - i(q-1)} x_{i(q-1)}, \\
S_2 & =
-\sum_{i = 1}^{A+B} \binom{i(q-1)-1}{B(q-1)-1} x_{n - i(q-1)} x_{i(q-1)}.
\end{align*}
From Lemma~\ref{lemma: modular properties on chen delta}, $\binom{i(q-1)}{Aq}$ vanishes only when $q \mid i$, which is possible only when $A+B = q$. In this case, the only nonzero summand is
\begin{align*}
(-1)^{A-1}\binom{q(q-1)-1}{Aq-1} \equiv (-1)^{A-1}\binom{q-2}{A-1} \equiv A \pmod{p},
\end{align*}
thus
$
S_1 = \begin{cases}
0, & (A+B<q)\\
A x_{n-q(q-1)}x_{q(q-1)}. & (A+B=q)
\end{cases}
$
For $S_2$, when $i <q$,
\begin{align*}
\binom{(i-1)q + q - i - 1}{(B-1)q+(q-B-1)} \equiv \binom{i-1}{B-1}\binom{q-i-1}{q-B-1}, \pmod p
\end{align*}
is nonzero only when $i = B$. Thus, if $A+B \le q-1$, $S_2 = - x_{Aq}x_{B(q-1)}$. When $A+B=q$, the coefficient for the summand corresponding to $i=q$ is
\begin{align*}
\binom{(q-2)q+(q-1)}{(B-1)q + (q-B-1)}\equiv \binom{q-2}{B-1}\binom{q-1}{q-B-1}\equiv -B\pmod{p},
\end{align*}
so
$
S_2 = \begin{cases}
-x_{Aq}x_{B(q-1)}, & (A+B<q)\\
-x_{Aq}x_{B(q-1)} + B x_{n-q(q-1)}x_{q(q-1)}, & (A+B =q)
\end{cases}
$
which yields \[ S_1 + S_2 = -x_{Aq}x_{B(q-1)}\] in any case. Thus we have \[ x_{Aq}\shuffle x_{B(q-1)} = x_{Aq + B(q-1)} + x_{B(q-1)}x_{Aq}\] as we desired.
\end{proof}
\begin{proposition}\label{proposition: Hopf deltas depth 1 <= q^2} When $q+1 \le n \le q^2$, with some $1\le k < q$ such that $kq+1 \le n \le (k+1)q$, we have;
\begin{align*}
\Delta(x_n) & = 1\otimes x_n + x_n \otimes 1 -n x_{n-q+1}\otimes x_{q-1} + \frac{n(n+1)}{2} x_{n-2q+2}\otimes x_{2q-2} - \dots \\
& + (-1)^k \binom{n+k-1}{k} x_{n-kq+k}\otimes x_{kq-k} \\
& = 1\otimes x_n + x_n \otimes 1 +
\sum_{i =1}^k (-1)^i \binom{n-1 + i}{i} x_{n- i(q-1)}\otimes x_{i(q-1)}.
\end{align*}
\end{proposition}
\begin{proof}Note that $q\mid n$ and $(q-1)\mid n$ cases are already treated above. Note that Lemma~\ref{lemma: Hopf delta for (q-1) multiples} coincides with the statement as follows: when $n = r(q-1)$ with $2\le r \le q$, then
\begin{align*}
& (-1)^i\binom{rq-r-1+i}{i} = (-1)^i\binom{(r-1)q+(q-r-1+i)}{i}\\
& \equiv (-1)^i \binom{r-1}{0}\binom{q-r-1+i}{i} = \frac{(-1)^i}{i!}\cdot \frac{(q-(r-i+1))(q-(r-i+2))\dots 1}{(q-(r+1))(q-(r+2))\dots 1}\\
& \equiv \frac{(-1)^i}{i!}\cdot (-1)^{q-(r-i+1) + (q-r-1)}\frac{((r-i+1)-q)((r-i+2)-q)\dots ((q-1)-q)}{((r+1)-q)((r+2)-q)\dots ((q-1)-q)}\\
& = \frac{r!}{i!(r-i)!} = \binom{r}{i}\pmod{p},
\end{align*}
by Lucas's theorem and calculations. This kind of calculation will be used often.
We show the $k=1$ case for the initial step. The statement for $k=1$ is for $q+1\le n \le 2q$, but it also holds for $n=q$, since $-q x_1x_{q-1}$ term vanishes over $\mathbb{F}_p$. Based on this initial case, we let $q+1\le n < 2q$, and assume that the statement holds for all $q, q+1, \dots, n-1$. Then,
\begin{align*} x_1 \shuffle x_{n-1} & = x_1x_{n-1}+ x_{n-1}x_1+ x_n + \sum_{\substack{q-1 \mid j \\ j< n}}\Delta^j_{1,n-1}x_{n-j}x_j.
\end{align*}
Note that $j=q-1$ yields a summand. There is another summand for $j=2q-2$ only when $n=2q-1$.
Since $q+1 \le n$, $\Delta_{1, n-1}^{q-1} = 1$. Also since $\Delta_{1, (2q-1)-1}^{2q-2} = 1 + (-1)^{2q-1}=0$
we have
\begin{align*} & x_1 \shuffle x_{n-1} =
x_1 x_{n-1}+x_{n-1}x_1+ x_{n} + x_{n-q+1}x_{q-1},
\end{align*}
thus
\begin{align*}
\Delta(x_1)\shuffle \Delta(x_{n-1}) = \Delta(x_1 x_{n-1}) + \Delta(x_{n-1} x_1) + \Delta(x_{n-q+1}x_{q-1}) + \Delta(x_n),
\end{align*} and a routine calculation with the induction hypothesis yields the result.
We divide the remaining cases into two subcases: when $n=kq+b = k(q-1)+(k+b)$ with $1\le k, b \le q-1$, (1) $2\le k+b < q-1$ and (2) $q\le k+b$.
Define a temporal function $\Delta'$ to be $\Delta' (\mathfrak{u}) = \Delta(\mathfrak{u}) - 1 \otimes \mathfrak{u} - \mathfrak{u} \otimes 1$.
(1) Assume that the statement holds for all $kq + b$ with $k+b <q-1$ and $k < \kappa$, for some $\kappa\ge2$. Let $b\in \mathbb{N}$ with $ \kappa + b < q-1$, and let $n = \kappa q + b$. We have
\begin{align*}
x_{\kappa q} \shuffle x_b & = x_{\kappa q + b} + x_{\kappa q}x_b + x_bx_{\kappa q} + \sum_{i=1}^{\kappa } \Delta_{\kappa q, b}^{i(q-1)} x_{n - i(q-1)}x_{i(q-1)}, \\
\Delta (x_{\kappa q}) & = 1\otimes x_{\kappa q} + x_{\kappa q} \otimes 1,\\
\Delta( x_b) & = 1 \otimes x_b + x_b \otimes 1 ,\\
\Delta(x_{\kappa q} x_b) &= 1 \otimes x_{\kappa q} x_b + x_{\kappa q} \otimes x_b + x_{\kappa q} x_b \otimes 1 ,\\
\Delta (x_b x_{\kappa q})& = 1 \otimes x_b x_{\kappa q} + x_b \otimes x_{\kappa q} + x_b x_{\kappa q} \otimes 1,
\end{align*}
so
\begin{align*}
& \Delta(x_{\kappa q} \shuffle x_b) \\
&= \Delta (x_{\kappa q+b}) + 1 \otimes x_{\kappa q} x_b + x_{\kappa q} \otimes x_b + x_{\kappa q} x_b \otimes 1 + 1 \otimes x_b x_{\kappa q} + x_b \otimes x_{\kappa q} + x_b x_{\kappa q} \otimes 1 \\
& + \sum_{i=1}^{\kappa } \Delta_{\kappa q, b}^{i(q-1)} \Delta (x_{n - i(q-1)}x_{i(q-1)})\\
& = x_{\kappa q}\otimes x_b + x_b \otimes x_{\kappa q} \\
& + 1\otimes x_{\kappa q + b} + 1\otimes x_{\kappa q}x_b + 1\otimes x_bx_{\kappa q} + \sum_{i=1}^{\kappa } \Delta_{\kappa q, b}^{i(q-1)} 1\otimes x_{n - i(q-1)}x_{i(q-1)} \\
& + x_{\kappa q + b}\otimes 1 + x_{\kappa q}x_b\otimes 1 + x_bx_{\kappa q}\otimes 1 + \sum_{i=1}^{\kappa } \Delta_{\kappa q, b}^{i(q-1)} x_{n - i(q-1)}x_{i(q-1)}\otimes 1,
\end{align*}
thus
\begin{align*}
\Delta (x_{\kappa q+b}) = 1\otimes x_{\kappa q + b} + x_{\kappa q + b}\otimes 1 - \sum_{i=1}^{\kappa } \Delta_{\kappa q, b}^{i(q-1)} \Delta' (x_{n - i(q-1)}x_{i(q-1)}),
\end{align*}
for $1\le i \le \kappa$, $i(q-1)<\kappa q$. Also since $b<q$, Lucas's theorem and some calculations yield $\binom{iq-i-1}{b-1} \equiv \binom{q-1-i}{b-1} \equiv (-1)^{b-1} \binom{b+i-1}{b-1}
\pmod p$, so
\begin{align}
\Delta (x_{\kappa q+b}) = 1\otimes x_{\kappa q + b} + x_{\kappa q + b}\otimes 1
-\sum_{i=1}^{\kappa} \binom{b+i-1}{b-1}\Delta' (x_{n - i(q-1)}x_{i(q-1)}).
\label{eqn: Delta(n) up to q*q, case 1, delta formula}
\end{align}
Let $a_i = n-i(q-1) = (\kappa -i )q + (b+i)$ and $b_i = i(q-1)$. Note that $b+i \le b+ \kappa <q-1$, so by induction hypothesis we have
\begin{align*}
\Delta(x_{a_i}) = & \ 1\otimes x_{a_i} + x_{a_i} \otimes 1 + \sum_{j=1}^{\kappa -i } (-1)^j\binom{b+i+j-1}{j} x_{a_i -j(q-1)} \otimes x_{j(q-1)}, \\
\Delta(x_{b_i}) = & \ 1 \otimes x_{b_i} + x_{b_i} \otimes 1 + \sum_{s =1}^{i-1} \binom{i}{s} x_{s(q-1)}\otimes x_{(i-s)(q-1)},\\
\Delta(x_{a_i}x_{b_i}) = & \ 1\otimes x_{a_i}x_{b_i} + x_{a_i}\otimes x_{b_i} + x_{a_i}x_{b_i} \otimes 1 \\
& + \sum_{s =1}^{i-1} \binom{i}{s} x_{a_i}x_{s(q-1)}\otimes x_{(i-s)(q-1)} \\
& + \sum_{j=1}^{\kappa -i} (-1)^j\binom{b+i+j-1}{j} x_{a_i -j(q-1)} \otimes x_{j(q-1)}\shuffle x_{b_i}\\
& + \sum_{j=1}^{\kappa -i} (-1)^j\binom{b+i+j-1}{j} x_{a_i -j(q-1)}x_{b_i} \otimes x_{j(q-1)} \\
& + \sum_{j=1}^{\kappa -i} \sum_{s=1}^{i-1} (-1)^j\binom{b+i+j-1}{j}\binom{i}{s}
x_{a_i -j(q-1)} x_{s(q-1)}\otimes x_{j(q-1)} \shuffle x_{(i-s)(q-1)}\\
= & \ 1\otimes x_{a_i}x_{b_i} + x_{a_i}\otimes x_{b_i} + x_{a_i}x_{b_i} \otimes 1 \\
& + \sum_{s =1}^{i-1} \binom{i}{s} x_{a_i}x_{s(q-1)}\otimes x_{(i-s)(q-1)} \\
& + \sum_{j=1}^{\kappa -i} (-1)^j\binom{b+i+j-1}{j} x_{a_i -j(q-1)} \otimes x_{(i+j)(q-1)}\\
& + \sum_{j=1}^{\kappa -i} (-1)^j\binom{b+i+j-1}{j} x_{a_i -j(q-1)}x_{b_i} \otimes x_{j(q-1)} \\
& + \sum_{j=1}^{\kappa -i} \sum_{s=1}^{i-1} (-1)^j\binom{b+i+j-1}{j}\binom{i}{s}
x_{a_i -j(q-1)} x_{s(q-1)}\otimes x_{(i+j-s)(q-1)}
\end{align*}
by \eqref{eqn: star product of q-1 multiples}. By manipulating indices combining sums,
\begin{align*}
\Delta'(x_{a_i}x_{b_i}) & = x_{a_i}\otimes x_{b_i} + \sum_{s =1}^{i-1} \binom{i}{s} x_{a_i}x_{(i-s)(q-1)}\otimes x_{s(q-1)} + \\
& + \sum_{j=i+1}^{\kappa} (-1)^{j-i}\binom{b+j-1}{j-i} x_{a_i +(i-j)(q-1)} \otimes x_{j(q-1)}\\
& + \sum_{j=1}^{\kappa -i} \sum_{s=0}^{\infty} \mathbf{1}_{[0, i)}(s) (-1)^j \binom{b+i+j-1}{j}\binom{i}{s}
x_{a_i -j(q-1)} x_{(i-s)(q-1)}\otimes x_{(j+s)(q-1)}
\end{align*}
where $\mathbf{1}_{[0, i)}(s) =1$ when $0\le s <i$, otherwise 0.
So recalling the previous calculation \eqref{eqn: Delta(n) up to q*q, case 1, delta formula},
\begin{align*}
\Delta' (x_{\kappa q+b}) & = - \sum_{i=1}^{\kappa} \binom{b+i-1}{b-1}\Delta' (x_{a_i}x_{b_i})\\
& = S_1 + S_2 + S_3, \end{align*}
where
\begin{align*}
S_1 &= - \sum_{i=1}^{\kappa}\binom{b+i-1}{b-1} x_{a_i}\otimes x_{b_i}, \\
S_2 &= - \sum_{i=2}^{\kappa }(-1)^i \sum_{j=1}^{i-1}(-1)^{j} \binom{b+j-1}{b-1}\binom{b+i-1}{i-j} x_{a_i} \otimes x_{b_i},\\
S_3 & =
-\sum_{i=1}^{\kappa}
\sum_{j=0}^{\kappa -i} \sum_{s=0}^{\infty}
\mathbf{1}_{[0, i)}(s) (-1)^j \binom{b+i-1}{b-1}
\binom{b+i+j-1}{j}\binom{i}{s}
x_{a_i -b_j} x_{b_{i-s}}\otimes x_{b_{j+s}}
\end{align*}
We show that $S_1 + S_2$ yields the desired equation, and $S_3$ vanishes.
One can see that sum for $i=1$ in $S_1$ is $-b x_{n-q+1} \otimes x_{q-1}$.
For the remaining terms in $S_1+S_2$,
we can prove
\[- \binom{b+i-1}{b-1} - (-1)^i \sum_{j=1}^{i-1} (-1)^j \binom{b+j-1}{b-1} \binom{b+i-1}{i-j} = (-1)^i \binom{b+i-1}{i}
\]
for $i\ge2$ and $1\le b <q$ from the equation
\begin{align*}
\binom{b+j-1}{b-1}\binom{b+i-1}{i-j}& =
\binom{i}{j}\binom{b+i-1}{b-1}
\end{align*}
and the binomial equation $\sum_{j=1}^{i-1} (-1)^j \binom{i}{j} = -1 -(-1)^i$.
Thus it is enough to show that $S_3 =0$. We manipulate indices as $u=j+s$, and then $v = i+j$, to have
\begin{align*}
& -S_3 \\
& = \sum_{u=0}^\infty \sum_{i=1}^\kappa \sum_{j=0}^{\kappa-i}
\mathbf{1}_{[0, i)}(u-j) (-1)^j \binom{b+i-1}{b-1}
\binom{b+i+j-1}{j}\binom{i}{u-j}
x_{a_i -b_j} x_{b_{i+j-u}}\otimes x_{b_u}\\
& = \sum_{u=1}^{\kappa} \sum_{v=1}^\kappa \sum_{i=1}^{v}
\mathbf{1}_{[0, i)}(u-v+i) (-1)^{v-i} \binom{b+i-1}{b-1}
\binom{b+v-1}{v-i}\binom{i}{v-u}
x_{a_i -b_{v-i}} x_{b_{v-u}}\otimes x_{b_u}.
\end{align*}
The coefficient for $x_{a_i-(v-i)(q-1)} x_{(v-u)(q-1)}\otimes x_{u(q-1)}$ of $-S_3$ vanishes when $u\ge v$ due to $\mathbf{1}_{[0,i)}(u-v+i) =0$ factor.
Otherwise, it is
\begin{align*}
& \sum_{i=1}^{v}
\mathbf{1}_{[0, i)}(u-v+i) (-1)^{v-i} \binom{b+i-1}{b-1}
\binom{b+v-1}{v-i}\binom{i}{v-u}\\
& =\sum_{i=v-u}^{v}
(-1)^{v-i} \binom{b+i-1}{b-1}
\binom{b+v-1}{v-i}\binom{i}{v-u}\\
& =\sum_{i=0}^{u}
(-1)^{u-i} \binom{b+i+v-u-1}{b-1}
\binom{b+v-1}{u-i}\binom{i+v-u}{v-u}\\
& =\frac{(-1)^u(b+v-1)!}{u!(b-1)!(v-u)!}\sum_{i=0}^{u}
(-1)^{i}
\frac{u!}{i!(u-i)!}=0.
\end{align*}
Thus we are done.
\
(2)
Let $n = \kappa q + b$, and suppose that $q-1 <\kappa + b$. We have that $n = \kappa q + b = (\kappa +1) (q-1) + (\kappa + b - q + 1)$ with $0 < \kappa + b - q+ 1 < q$. We exclude the $n = q^2-1$ case, i.e. $\kappa=b=q-1$ case, since it is already treated earlier, so that $0<\kappa + b - q+1 \le q-2$.
Let $A = \kappa +b-q+1$, $B = q-b$. Then $n = Aq + B(q-1)$, $1 \le A \le q-2$, $1 \le B \le q-1$, and $A+B = \kappa +1$. What we need to prove turns out to be
\begin{equation} \Delta(x_n) = 1\otimes x_n + x_n \otimes 1 + \sum_{i=1}^{B}\binom{B}{i} x_{n - i(q-1)}\otimes x_{i(q-1)}
\label{eqn: Delta(n) up to q*q, case 2, in terms of A and B}
\end{equation} in terms of $A$ and $B$, by Lucas's theorem and direct calculations.
From Lemma~\ref{lemma: Aq star B(q-1)},
%
%
\begin{align*}
x_{Aq} \shuffle x_{B(q-1)} &= x_n + x_{B(q-1)}x_{Aq}
\end{align*}
for all $A$, $B$ with $A+B = \kappa+1 \le q$.
Now we can show that $\Delta(x_{Aq} \shuffle x_{B(q-1)}) = \Delta(x_{Aq}) \shuffle \Delta(x_{B(q-1)})$. From
\begin{align*}
\Delta(x_{B(q-1)})& = 1 \otimes x_{B(q-1)} + x_{B(q-1)}\otimes1 +
\sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}\otimes x_{(B-j)(q-1)}, \\
\Delta(x_{Aq}) & = 1\otimes x_{Aq} + x_{Aq} \otimes 1,
\end{align*}
we have
\begin{align*}
& \Delta(x_{B(q-1)}x_{Aq} ) \\
& = 1\otimes x_{B(q-1)}x_{Aq} +
x_{B(q-1)}\otimes x_{Aq} + x_{B(q-1)}x_{Aq} \otimes 1 \\
& + \sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}\otimes x_{Aq}\shuffle x_{(B-j)(q-1)} + \sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}x_{Aq} \otimes x_{(B-j)(q-1)}\\
&= 1\otimes x_{B(q-1)}x_{Aq} +
x_{B(q-1)}\otimes x_{Aq} + x_{B(q-1)}x_{Aq} \otimes 1 \\
& + \sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}\otimes \left( x_{Aq +(B-j)(q-1)} + x_{(B-j)(q-1)}x_{Aq}\right)\\
& + \sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}x_{Aq} \otimes x_{(B-j)(q-1)},
\end{align*}
thus
\begin{align*}
& \Delta(x_{Aq} \shuffle x_{B(q-1)}) \\
& = \Delta(n) + 1\otimes x_{B(q-1)}x_{Aq} +
x_{B(q-1)}\otimes x_{Aq} + x_{B(q-1)}x_{Aq} \otimes 1 \\
& + \sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}\otimes \left( x_{Aq +(B-j)(q-1)} + x_{(B-j)(q-1)}x_{Aq}\right)\\
& + \sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}x_{Aq} \otimes x_{(B-j)(q-1)}\\
& = \Delta(x_{Aq})\shuffle \Delta(x_{B(q-1)})\\
& = x_{Aq} \otimes x_{B(q-1)} + x_{Aq}\shuffle _{B(q-1)}\otimes1 +
\sum_{j=1}^{B-1} \binom{B}{j} x_{Aq}\shuffle x_{j(q-1)}\otimes x_{(B-j)(q-1)}\\
& + 1 \otimes x_{B(q-1)}\shuffle x_{Aq} + x_{B(q-1)}\otimes x_{Aq} +
\sum_{j=1}^{B-1} \binom{B}{j} x_{j(q-1)}\otimes x_{(B-j)(q-1)}\shuffle x_{Aq},
\end{align*}
so routine calculation yields
\begin{align*}
\Delta(x_n) & = 1 \otimes x_n + x_n\otimes 1 + x_{Aq} \otimes x_{B(q-1)}
+ \sum_{j=1}^{B-1} \binom{B}{j} x_{n - (B-j)(q-1)}\otimes x_{(B-j)(q-1)}\\
& = 1 \otimes x_n + x_n\otimes 1 + \sum_{j=1}^{B} \binom{B}{j} x_{n -j(q-1)}\otimes x_{j (q-1)},
\end{align*}
which is \eqref{eqn: Delta(n) up to q*q, case 2, in terms of A and B}, thus we are done.
\end{proof}
We mention that some explicit formulas for $\Delta(x_n)$ with $n>q^2$ can be found in the appendix (see Appendix \ref{sec: numerical experiments}).
\section{Stuffle Hopf algebra and stuffle map in positive characteristic} \label{sec: stuffle algebra}
In this section we define the stuffle algebra and the stuffle map in positive characteristic. The stuffle algebra is easy to define. However, to define the stuffle map we make use of a deep connection between the $K$-vector space spanned by the MZV's in positive characteristic and that spanned by the multiple polylogarithms in positive characteristic proved in \cite{IKLNDP22, ND21} (see Theorem \ref{thm:bridge}).
\subsection{The stuffle algebra in positive characteristic} ${}$\par
We recall that the composition space $\frak C$ is introduced in \S \ref{sec: composition space}. We define the stuffle product in the same way as that of $(\frak h^1,*)$ as given in \S \ref{sec: Hoffman algebra}. More precisely,
\begin{align*}
* \colon \mathfrak{C} \times \mathfrak{C} \longrightarrow \mathfrak{C}
\end{align*}
by setting $1 * \mathfrak{a} = \mathfrak{a} * 1 = \mathfrak{a}$ and
\begin{align*}
\mathfrak{a} * \fb &= \ x_{a}(\mathfrak{a}_- * \fb) + x_{b}(\mathfrak{a} * \fb_-) + x_{a+b} (\mathfrak{a}_- * \fb_-)
\end{align*}
for any words $\mathfrak{a},\fb \in \langle \Sigma \rangle$. We call $*$ the \textit{stuffle product}.
\begin{proposition}
With the above notation, $(\frak C,*)$ is a commutative $\mathbb F_q$-algebra.
\end{proposition}
\begin{proof}
The proof follows the same line as that of $(\frak h^1,*)$ (see \S \ref{sec: Hoffman algebra}).
\end{proof}
\subsection{Hopf algebra structure} ${}$\par
We now define a coproduct $\Delta_*:\frak C \otimes \frak C \to \frak C$ and a counit $\epsilon:\frak C \to \mathbb F_q$ by
\[ \Delta_*(\fu)= \sum_{\mathfrak{a}\fb=\fu} \mathfrak{a} \otimes \fb \]
and for any words $\fu \in \langle \Sigma \rangle$,
\begin{align*}
\epsilon(\fu)=\begin{cases} 1 \quad \text{if } \fu=1, \\ 0 \quad \text{otherwise}. \end{cases}
\end{align*}
By Theorem \ref{theorem: Hopf algebras for qp}, we get:
\begin{theorem} \label{thm: Hopf algebra for stuffle product}
The stuffle algebra $(\frak C,*,u,\Delta_*,\epsilon)$ is a connected graded Hopf algebra of finite type over $\mathbb F_q$.
\end{theorem}
\subsection{The stuffle map in positive characteristic} ${}$\par \label{sec: stuffle map}
We put $\ell_0 := 1$ and $\ell_d := \prod^d_{i=1}(\theta - \theta^{q^i})$ for all $d \in \mathbb{N}$. Letting $\mathfrak s = (s_1 , \dots, s_n) \in \mathbb{N}^n$, for $d \in \mathbb{Z}$, we define analogues of power sums by
\begin{equation*}
\Si_d(\mathfrak s) = \sum\limits_{d=d_1> \dots > d_n\geq 0} \dfrac{1}{\ell_{d_1}^{s_1} \dots \ell_{d_n}^{s_n}} \in K,
\end{equation*}
and
\begin{equation*}
\Si_{<d}(\mathfrak s) = \sum\limits_{d>d_1> \dots > d_n\geq 0} \dfrac{1 }{\ell_{d_1}^{s_1} \dots \ell_{d_n}^{s_n}} \in K.
\end{equation*}
Thus
\begin{align*}
\Si_{<d}(\mathfrak{s}) =\sum_{i=0}^{d-1} \Si_i(\mathfrak{s}), \quad \Si_{d}(\mathfrak{s}) =\Si_d(s_1) \Si_{<d}(\mathfrak{s}_-)=\Si_d(s_1) \Si_{<d}(s_2,\dots,s_n).
\end{align*}
Here by convention we define empty sums to be $0$ and empty products to be $1$. In particular, $\Si_{<d}$ of the empty tuple is equal to $1$.
Then we define the Carlitz multiple polygarithm (CMPL for short) as follows
\begin{equation*}
\Li(\mathfrak{s}) = \sum \limits_{d \geq 0} \Si_d (\mathfrak{s}) = \sum\limits_{d_1> \dots > d_n\geq 0} \dfrac{1}{\ell_{d_1}^{s_1} \dots \ell_{d_n}^{s_n}} \in K_{\infty}.
\end{equation*}
We agree also that $\Li(\emptyset) = 1$. We call $\depth(\mathfrak{s}) = n$ the depth, $w(\mathfrak{s}) = s_1 + \dots + s_n$ the weight of $\Li(\mathfrak{s})$.
\begin{lemma}
For all $\mathfrak{s}$ as above such that $s_i \leq q$ for all $i$, we have
\begin{equation*}
S_d(\mathfrak{s}) = \Si_d(\mathfrak{s}) \quad \text{for all } d \in \mathbb{Z}.
\end{equation*}
Therefore,
\begin{equation*}
\zeta_A (\mathfrak{s}) = \Li(\mathfrak{s}) .
\end{equation*}
\end{lemma}
\begin{proof}
See \cite[Lemma 1.1]{IKLNDP22}.
\end{proof}
Combining the above lemma and some extensions of results in \cite{ND21} we get a deep connection between the $K$-vector space spanned by the MZV's in positive characteristic and that spanned by the multiple polylogarithms in positive characteristic.
\begin{theorem} \label{thm:bridge}
The $K$-vector space $\mathcal{Z}_w$ of MZV's of weight $w$ and the $K$-vector space $\mathcal{L}_w$ of CMPL's of weight $w$ are the same.
\end{theorem}
\begin{proof}
See \cite[Theorem 4.3]{IKLNDP22}.
\end{proof}
For all $d \in \mathbb Z$, we define two $\mathbb{F}_q$-linear maps
\begin{equation*}
\Si_{<d} \colon \frak C \rightarrow K_{\infty} \quad \text{and} \quad \Li \colon \frak C \rightarrow K_{\infty},
\end{equation*}
which map the empty word to the element $1 \in K_{\infty}$, and map any word $x_{s_1} \dotsc x_{s_n}$ to $\Si_{<d}(s_1, \dotsc ,s_n)$ and $\Li(s_1, \dotsc, s_n)$, respectively. We have the following result:
\begin{proposition}\label{lastProp}
For all words $\mathfrak{a}, \fb \in \frak C$ and for all $d \in \mathbb Z$ we have
\begin{align*}
\Si_{<d}(\mathfrak{a} * \fb) &=\Si_{<d}(\mathfrak{a}) \, \Si_{<d}(\fb), \\
\Li(\mathfrak{a} *\fb) &=\Li(\mathfrak{a}) \, \Li(\fb).
\end{align*}
\end{proposition}
\begin{proof}
We leave the proof to the reader.
\end{proof}
Combining Theorem \ref{thm:bridge} and Proposition~\ref{lastProp} yields the $K$-linear map \[ Z_*:\frak C \otimes_{\mathbb F_q} K \to \mathcal Z, \] which sends a word $\mathfrak{a} \in \frak C$ to $\Li(\mathfrak{a})$, is a homomorphism of $K$-algebras, and is called the stuffle map in positive characteristic.
|
\section{Conclusion and Future Work}\label{conclusion}
In this paper, we defined the task of detecting stance of authorities towards rumors in tweets, and released the first dataset for the task targeting Arabic rumors. We studied the usefulness of existing Arabic datasets for stance detection for claim verification in our task. Based on our experiments and failure analysis, we found that although existing stance datasets showed to be somewhat useful for the task, they are obviously insufficient and there is a need to augment them with stance of authorities from Twitter data. In addition to expanding AuSTR{} to have sufficient training data for the task that can be used solely or to augment existing stance datasets, we plan to explore and contribute with stance models specific to the task.
\section*{Acknowledgments} The work of Fatima Haouari was supported by GSRA grant\# GSRA6-1-0611-19074 from the Qatar National Research Fund. The work of Tamer Elsayed was made possible by NPRP grant\# NPRP11S-1204-170060 from the Qatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsibility of the authors.
\section{Results and Discussion}\label{discuss}
The research question we address in this preliminary study is whether the existing stance detection datasets are useful or not in our task. To answer it, we use combinations of the existing datasets for training and AuSTR{} for testing. We also show how models trained on those combinations perform on their own corresponding in-domain test sets. While the results on the in-domain test sets are not comparable, since those test sets are different, they constitute an estimated upper bound performance. To evaluate the models, we report per-class $F_1$ and macro-$F_1$ scores.
Table~\ref{tab:results} presents the performance results of all experiments, which demonstrate several interesting observations.
First, we notice that almost all models (except a few) were able to achieve higher performance on their own in-domain test sets compared to AuSTR. This shows that domain adaptation was not very effective (thus in-domain data for our task is required for training the models).
Second, when using individual stance datasets for training, the model trained on AraStance clearly outperformed the others in all measures when tested on AuSTR. We note that ArabicFC is severely imbalanced, where the \textit{disagree} class represents only 3.3\% of the data, yielding a very poor performance on that class even when tested on its own in-domain test set. A similar conclusion was found by previous studies~\cite{baly-2018,Arastance}.
As for ANS, evidence is manually crafted, which is not as realistic as tweets from authorities. Alternatively, AraStance claims are extracted from three fact-checking websites,\footnote{Claims are collected from sources other than the ones we used to construct AuSTR.} covering multiple domains and Arab countries, similar to AuSTR, and the evidence is represented in articles written by journalists, not manually crafted.
Third, when tested on AuSTR, the model trained on all datasets combined exhibits the best performance on the \emph{disagree} class; however its performance was severely degraded compared to the AraStance model on the \emph{agree} class.
This indeed needs further investigation.
Furthermore, we observe that AraStance achieved the highest $F_1$(D) when used solely for training, and whenever combined with the other datasets. To investigate this, we manually examined a 10\% random sample of \textit{disagreeing} training articles. We found they have common words such as \textit{rumors}, \textit{not true}, \textit{denied}, and \textit{fake}; similar keywords appear in some \textit{disagreeing} tweets of AuSTR.
Finally, we observe that there is a clear discrepancy in the performance across different classes. Considering the model trained on all datasets for example, $F_1$(A) is 0.74 while $F_1$(D) is 0.65. Moreover, it is clear that detecting the \textit{disagree} stance is the most challenging subtask, which we expect to benefit from in-domain training. Overall, we believe training and testing on tweets is very different, as they are very short and informal, which needs special pre-processing.
\begin{table*}
\caption{Performance on both the in-domain test sets and AuSTR{}, measured in per-class $F_1$ (A: Agree, D: Disagree, U: Unrelated) and macro-$F_1$. On AuSTR, bold and underlined values indicate best and second-best performance respectively.}
\label{tab:results}
\centering
\begin{tabular}{l||ccc|c||ccc|c}
\toprule
& \multicolumn{4}{c||}{\textbf{Test on in-Domain Set}}
& \multicolumn{4}{c}{\textbf{Test on AuSTR}}
\\
\textbf{Training Set} & $F_1$(A) & $F_1$(D) & $F_1$(U)
& m-$F_1$ & $F_1$(A) & $F_1$(D) & $F_1$(U) &
m-$F_1$ \\
\hline
ANS & \,0.824\, & \,0.901\, & \,0.923\,
& \,0.882\, & \,0.653\, & \,0.578\, & \,0.709\,
& \,0.647\, \\
ArabicFC& \,0.770\, & \,0.090\, &\,0.915\,
& \,0.591\, & \,0.641 \, &\,0.434\, &\,0.799\,
&\,0.625\, \\
AraStance& \,0.898\, & \,0.833\, & \,0.95\,
& \,0.894\, & \,\textbf{0.837}\, & \,0.613\, & \,\underline{0.865}\,
& \,\textbf{0.772}\,
\\
\hline
ANS+ArabicFC &\,0.807\, & \,0.866\, & \,0.899\,
& \,0.857\, & \,0.678\, & \,0.587\, &\,0.862\,
&\,0.709\,
\\
ANS+AraStance& \,0.893\, & \,0.909\, & \,0.955\,
& \,0.919\, & \,0.743\, & \,0.629\, & \,0.847\,
& \,0.740\,
\\
ArabicFC+AraStance & \,0.765\, & \,0.555\, & \,0.897\,
& \,0.739\, & \,\underline{0.754}\, & \,\underline{0.635}\, & \,0.862\,
& \,0.750\,
\\
\hline
All Three Datasets\, & \,0.778\, & \,0.742\, &\,0.889\,
& \,0.803\, &\,0.741\, & \,\textbf{0.646}\, & \,\textbf{0.866}\,
& \,\,\underline{0.751}\,\, \\
\bottomrule
\end{tabular}
\end{table*}
\paragraph{\textbf{Failure Analysis}.}
We conducted a failure analysis on 17 examples from AuSTR{} that failed to be predicted correctly by \textit{all} of our 7 trained models. We found that we can attribute the failures to two main reasons: (1) \textit{Writing Style}, where the authority is denying a rumor about herself speaking in the first person. This constitutes 64.7\% of the examined failures. We believe this is due to the fact that none of the stance datasets we used for training have evidence written by authorities themselves, as the source was either news articles written by journalists, or paraphrased or contradicted news headlines manually crafted by annotators. (2) \textit{Indirect Disagreement/Agreement}, where the authority is indirectly denying/supporting the rumor. Examples of both types of failures are presented in Table~\ref{tab:failed_examples}. These findings motivate the need to augmenting existing stance datasets with rumor-evidence pairs from Twitter to further improve the performance of detecting the stance of authorities towards rumors from their tweets.
\begin{table}[h]
\caption{Sample examples failed to be predicted correctly by \underline{\textbf{all}} models. The golden label for the examples is either \color{ForestGreen} Agree \color{black} or \color{red}Disagree\color{black}. Failure types are writing style, indirect disagreement, and indirect agreement for the examples in order.}\label{tab:failed_examples}
\centering
\begin{tabular}{p{5.cm}@{\hskip 0.1in}p{7.cm}@{\hskip 0.1in}}
\hline
\textbf{Rumor tweet [posting date]} & \textbf{Evidence tweet [posting date]}\\
\hline
\texttt{Mortada Mansour passed away recently of a heart attack.[29-10-2021]}& \textbf{@Mortada5Mansour}:\color{red}\texttt{ I am having my dinner now, and after a few minutes I will share a voice and video to reassure you, and I will reply to those who disturbed my family members in my village and caused the anxiety to all my fans.[29-10-2021]}\\
\hline
\texttt{Egypt does not give a vaccine to its citizens, the Gulf countries sponsor them: Saudi Arabia / Sultanate of Oman / Qatar refuses their intervention, so there is no other than Kuwait, the country of humanity that receives them and feeds them. What is the mysterious secret? Kuwait treats Egypt with special treatment.[07-05-2021]} & \textbf{@mohpegypt}:\color{red}\texttt{ Information about the \#coronavirus vaccine. To book a vaccine, please visit the website http://egcovac.mohp.gov.eg
or go to the nearest health unit (for citizens who have difficulty registering online). For more information, please call the hotline: 15335 \#together\_rest\_assured.[10-05-2021] }
\\
\hline
\texttt{Urgent
The headquarters of the fourth channel was stormed by the militias of the Sadrist movement in the capital, Baghdad.[04-11-2022]} & \textbf{@MAKadhimi}:\color{ForestGreen}\texttt{ The attack on one of the Iraqi media outlets, and the threat to the lives of its employees, is a reprehensible act and represents the highest level of transgression against the law and freedom of the press and does not fall within the peaceful and legal practices and protests.
We directed that the perpetrators be held accountable, and that protection be tightened on press institutions.[04-11-2022] }
\\
\bottomrule
\end{tabular}
\end{table}
\section{Constructing AuSTR{} Dataset
}\label{ourData}
To construct AuSTR{}
where both the rumor and evidence are tweets, we exploit both fact-checking articles and variant authority Twitter accounts
\paragraph{\textbf{Exploiting Fact-checking Articles.}}
Fact-checkers who attempt to verify rumors usually provide in their fact-checking articles some examples of social media posts (e.g., tweets) propagating the specific rumors, and other posts from trusted authorities that constitute evidence to support their verification decisions. To construct AuSTR, we exploit both examples of tweets: stating rumors and showing evidence from authorities as provided by those fact-checkers. Specifically, we used AraFacts~\cite{ali2021arafacts}, a large dataset of Arabic rumors collected from 5 fact-checking websites. From those rumors, we selected only the ones that are expressed in tweets and have evidence in tweets as well.\footnote{We contacted the authors of AraFacts to get this information as it was not released.} We then extracted the rumor-evidence pairs as follows. For \textit{true} and \textit{false} rumors, we selected a single tweet example and all provided evidence tweets, which are then labeled as having \textit{agree} and \textit{disagree} stances respectively.\footnote{We only kept evidence expressed in \emph{text} rather than in image or video.}
If the fact-checkers provided the authority account but stated no evidence was found to support or deny the rumor, we selected one or two tweets from the authority timeline posted soon before the rumor time, and assigned the \textit{unrelated} label to the pairs.
\paragraph{\textbf{Exploiting Authority Accounts}.} Given that fact-checkers focus more on \textit{false} rumors than \textit{true} ones, we ended up with only 4 \textit{agree} pairs as opposed to 118 \textit{disagree} pairs following the above step. To further expand our \textit{agree} pairs, we did the reverse of the previous approach, where we collected the evidence first. Specifically, we started from a set of Twitter accounts of authorities (e.g., ministers, presidents, embassies, organization accounts, etc.) covering most of the Arab countries and multiple domains (e.g., politics, health, and sports), and selected recent tweets stating claims from their timelines. For each claim, we used Twitter search interface to look for tweets from regular users expressing it, but tried to avoid exact duplicates.
Finally, to get closer to the real scenario, where percentage of \textit{unrelated} tweets is usually higher than percentages of \textit{agree} and \textit{disagree} tweets in the authority timelines, we further expanded the \textit{unrelated} pairs by selecting one or two \textit{unrelated} recent tweets from the authority timeline posted before the rumor time for each \textit{agree} and \textit{disagree} pairs.
Overall, we end up with 409 pairs covering 171 unique claims, where 41 are \textit{true} and 130 are \textit{false}. Among those pairs, 118 are \emph{disagree} (29\%), 62 are \emph{agree} (15\%), and 229 are \emph{unrelated} (56\%).
\section{Experimental Setup}\label{setup}
\paragraph{\textbf{Datasets}.}
To study the usefulness of existing Arabic datasets that target stance for claim verification, we adopted the following ones for training:
\begin{enumerate}
\item \textbf{ANS~\cite{ANS}} of 3,786 \textbf{(claim, sentence)} pairs, where claims were extracted from news article titles from trusted sources, then annotators were asked to generate \textit{true} and \textit{false} sentences towards them by adopting paraphrasing and contradiction respectively. The sentences are annotated as either \textit{agree}, \textit{disagree}, or \textit{other} towards the claims.
\item \textbf{ArabicFC~\cite{baly-2018}} of 3,042 \textbf{(claim, article)} pairs, where claims are extracted from a single fact-checking website verifying political claims about war in Syria, and articles collected by searching Google using the claim. The articles are annotated as either \textit{agree}, \textit{disagree}, \textit{discuss}, or \textit{unrelated} to the claim.
\item \textbf{AraStance~\cite{Arastance}}: 3,676 \textbf{(claim, article)} pairs, where claims are extracted from 3 Arabic fact-checking websites covering multiple domains and Arab countries. The articles were collected and annotated similar to ArabicFC.
\end{enumerate}
To train our models, we considered only three labels, namely, \textit{agree}, \textit{disagree}, or \textit{unrelated}. For ANS and AraStance, we used the same data splits provided by the authors; however, we split the ArabicFC into 70\%, 10\%, and 20\% of the claims for training, development, and testing respectively\footnote{We release ArabicFC splits for reproducibility.}. When splitting data, we assigned all pairs having the same claim to the same split. Table~\ref{tab:train-data} shows the size of different data splits of the three datasets. Due to the limited size of AuSTR, in this work, we opt to utilize it only as a \emph{test set} while using the above datasets for training to show their usefulness in our task.
\begin{table}[h]
\caption{Data splits of the Arabic stance datasets used for training.}
\label{tab:train-data}
\centering
\begin{tabular}{c|ccc|ccc|ccc}
\multirow{2}{*}{\textbf{Label}} & \multicolumn{3}{c|}{\textbf{ANS}} &\multicolumn{3}{c|}{\textbf{ArabicFC}} & \multicolumn{3}{c}{\textbf{AraStance}} \\ & \textbf{Train} & \textbf{Dev} & \textbf{Test} & \textbf{Train} & \textbf{Dev} & \textbf{Test} & \textbf{Train} & \textbf{Dev} & \textbf{Test}
\\
\hline
\textbf{Agree} & 903 & 268 & 130 &323 & 32 & 119 & 739 & 129 & 154 \\
\textbf{Disagree}& 1686& 471& 242 & 66 & 8& 13 & 309 & 76 & 64 \\
\textbf{Unrelated} & 63 & 16 & 7 &1464&198&410 & 1553 & 294 & 358 \\
\hline
Total & 2652& 755 & 379 & 1853& 238&542& 2601 &499 & 576 \\
\bottomrule
\end{tabular}
\end{table}
\paragraph{\textbf{Stance Models}.}\label{stanceModels}
To train our stance models, we fine-tuned BERT~\cite{devlin2018bert} to classify whether the evidence sentence/article \textit{agrees} with, \textit{disagrees} with, or is \textit{unrelated} to the claim. We feed BERT the claim text as sentence \emph{A}, the evidence as sentence \emph{B} (truncated if needed) separated by the [SEP] token. Finally, we use the contextual representation of the [CLS] token as input to a single classification layer with three output nodes, added on top of the BERT architecture to compute the probability for each class of stance.
Various Arabic BERT-based models were released recently~\cite{antoun2020arabert,safaya-etal-2020-kuisail,lan-etal-2020-empirical,inoue2021interplay,abdul2021arbert}; we opted to choose ARBERT~\cite{abdul2021arbert} as it was shown to achieve better performance on the stance datasets adopted in our work~\cite{Arastance}. We adopted the authors' setup~\cite{Arastance} by training the models for a maximum of 25 epochs, where early stopping was set to 5 and sequence length to 512. We trained 7 different models in an ablation study using different combinations of the stance datasets mentioned earlier.
\section{Introduction}
Existing studies for rumor verification in social media exploited the propagation networks as a source of evidence, where they focused on the stance of replies~\cite{wu2019different,kumar-carley-2019-tree,chen2020modeling,yu-etal-2020-coupled,bai2022multi,roy2022gdart}, structure of replies~\cite{ma2018rumor,bian2020rumor,choi2021dynamic,song2021temporally,haouari2021arcov19,bai2022rumor}, and profile features of retweeters~\cite{liu2018early}. Recently, Dougrez-Lewis et al.~\cite{dougrez2022phemeplus} proposed augmenting the propagation networks with evidence from the Web. To our knowledge, no previous research has investigated exploiting evidence for rumor verification in social media from the timelines of trusted authorities, where an authority is \emph{an entity with the real knowledge or power to verify or deny a specific rumor}~\cite{2023clef}. We believe that detecting stance of relevant authorities towards rumors can be a great asset to augment the sources of evidence utilized by existing rumor verification systems. It can also serve as a valuable tool for fact-checkers to automate their process of checking authority tweets to verify certain rumors. It is worth mentioning that stance of authorities can be just one (\emph{but} important) source of evidence that compliment other sources and by itself may not (in some cases) be fully trusted to decide the veracity of rumors.
In this paper, we conduct a preliminary study for detecting stance of authorities towards rumors spreading in Twitter in the Arab world.
Exploiting sources of evidence for Arabic rumor verification in Twitter is still under-studied; existing studies exclusively focused on the tweet text for verification~\cite{elhadad2020covid,hasanain2020overview,mahlous2021fake,al2021arabic,9620509,alqurashi2021eating}. A notable exception is the work done by Haouari et al.~\cite{haouari2021arcov19} that utilized the replies, their structure, and repliers' profile features to verify Arabic COVID-19 rumors.
Several studies addressed Arabic stance detection in Twitter; however, the target was a specific topic not rumors~\cite{darwish2017improved,jaziriyan2021exaasc,alqurashi2022stance}. A few datasets for stance detection for Arabic claim verification were released recently, where the evidence is either news articles~\cite{baly-2018,Arastance} or manually-crafted sentences~\cite{ANS}. However, there is no dataset where the rumors are tweets and the evidence is retrieved from authority timelines, neither in Arabic nor in other languages. To fill this gap, the contribution of our work is four-fold: (1) we define the task of detecting stance of authorities towards rumors in tweets, (2) we construct and release the first Authority STance for Rumors (AuSTR{}) dataset,\footnote{\url{https://github.com/Fatima-Haouari/AuSTR}} (3) we present the first study on the usefulness of existing stance detection datasets for our task, and (4) we perform a failure analysis to gain insights for the future work on the task. The research question we aim to address in this work is whether the existing datasets of Arabic stance detection for claim verification are useful for detecting the stance of authorities in Arabic tweets.
The remainder of this paper is organized as follows. We outline the construction methodology of AuSTR{} in Section~\ref{ourData}. Our experimental setup is presented in Section~\ref{setup}. We discuss and analyze our results in Section~\ref{discuss}. Finally, we conclude and suggest some future directions in Section~\ref{conclusion}.
|
\section{Introduction}
A nitrogen-vacancy (NV) center in diamond is an atomic defect, in which magnetic sensitivity is in the range of several microtesla with the nanometer spatial resolution at room temperature ~\cite{tay08,maz08}. The magnetic field sensitivity even reaches below 1 pT/Hz$^{0.5}$ by adopting NV ensembles at the cost of the spatial resolution ~\cite{bar16,wol15,fes20,zha21c}. Various techniques such as high-density NV centers ~\cite{bar16}, improved readout fidelity ~\cite{wol15, cle15, dum13}, and flux concentrators~\cite{fes20, zha21c} have been implemented to improve sensitivity. These techniques have contributed to the detection of bio-magnetic fields from neurons, mammalian muscles, and a heart using NV ensembles ~\cite{bar16, web21a, ara21}. However, these measurements exhibit millimeter-scale spatial resolutions, which may limit the visualization of functional activity in neural networks ~\cite{hal12, kar18, kar21}.
A widefield NV microscope based on the frequency shift of optically detected magnetic resonance (ODMR) can be an alternative method for detecting magnetic field distributions at a micrometer-scale spatial resolution ~\cite{ste10, sim16, gle15, sch21}. Currently, density distribution in graphene, magnetic field arising from an integrated circuit, a 2-D magnet, geological samples, and living cells were imaged at sub- or several-micrometer spatial resolutions ~\cite{tet17, tur20, bro20, gle17, sag13}. However, the widefield NV microscopes based on the ODMR frequency shift may have difficulties in detecting several tens of nanotesla, which corresponds to several hundreds of Hz in the frequency shift. This can be addressed using lock-in-based NV magnetometry ~\cite{sch11}.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig1.pdf}
\caption{ODMR spectrum and decoherence times of the overgrown diamond layer. The ODMR spectrum(a), and decoherence times, $\mathrm{T_2^*}$(b) and $\mathrm{T_2}$(c), support the high quality of the crystal.}
\label{fig1}
\end{figure}
In this study, we adopt a lock-in camera to improve the magnetic field sensitivity in a wide-field NV microscope ~\cite{woj18a, har21, web22, par22b}. For a simple yet reliable operation, we continuously excite NV diamond using a 532nm laser and a microwave. Multiple hyperfine transitions are simultaneously excited for a higher ODMR contrast, and double resonance is implemented to suppress the influences of temperature drift and strain ~\cite{bar16, fes20, har21, fan13, mam15, shi22}. Additionally, all four NV axes are exploited by aligning an external magnetic field along the $<$001$>$ direction of the NV diamond. Sub-millisecond temporal resolution, 0.4 ms, is demonstrated by detecting a magnetic field from a short-pulsed current at tens of nanotesla per-pixel sensitivity.
\section{Experimental}
A thin nitrogen-doped ([$^{14}$N] $\sim$ 10 ppm) diamond layer ($^{12}$C $>$ 99.99$\%$, 40 $\mu$m thick) is grown by chemical vapor deposition (CVD) on top of an electronic grade diamond plate by Applied Diamond Inc.. The dimensions of the diamond plate are approximately 2 $\times$ 2 $\times$ 0.54 mm$^3$. The diamond is electron irradiated (1 MeV, 5 $\times$ 10$^8$/cm$^2$), and annealed in a vacuum at 800$^{\circ}$C for 4 hours and 1000 $^{\circ}$C for 2 hours, sequentially. An ODMR spectrum of the crystal is shown in Fig.~\ref{fig1}(a). The NV decoherence times, $\mathrm{T_2^*}$ and $\mathrm{T_2}$, are found to be 1.6 $\mathrm{\mu s}$ and 19.3 $\mathrm{\mu s}$ at B $\approx$ 3 mT, respectively. $\mathrm{T_2^*}$ is found by fitting the data to $C_0$exp[-$(\tau/T_2^*)$]$\sum_{i}$ cos[$\omega_i \tau]$, where $C_0$ is the maximal ODMR contrast, and $\omega_i$ is frequency due to the NV hyperfine splittings, Fig.~\ref{fig1}(b). Similarly, $\mathrm{T_2}$ is obtained by fitting the data to $C_0$exp[-$(2\tau/T_2)^p$], where $p$ is stretched exponential parameter, Fig.~\ref{fig1}(c)~\cite{par22,bau18}.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig2.pdf}
\caption{(a) Schematic of the widefield microscope and protocol for lock-in camera detection. Frequency modulated (square wave) microwave (MW) is delivered into an omega-shaped coil for lock-in detection. Two MW sources (MW1 and MW2) and one RF (2.16 MHz) generator are mixed and combined for double resonance and hyperfine driving. A continuous-wave (CW) double resonance lock-in protocol is described in the inset. Phases ($\Phi_1$ and $\Phi_2$) of the frequency modulation are synchronized to a camera trigger for signal integration (I$^+$, Q$^+$, I$^-$, and Q$^-$). For magnetic field/temperature detections, $|\Phi_1 -\Phi_2| $ were kept at $\pi / 0 $, respectively. The trigger, Cam$\mathrm{_{trig}}$, starts the acquisition of the fluorescence, and a single frame of in-phase (I = I$^+$ - I$^+$) and quadrature (Q = Q$^+$ - Q$^-$) images are found after repeating$N_{cycle}$ times.}
\label{fig2}
\end{figure}
A schematic of the experimental setup is shown in Fig.~\ref{fig2}. An omega-shaped coil is placed on the overgrown diamond layer and a 532 nm laser (Opus 3W, LaserQuantum) illuminates the diamond. The incident light power on the diamond was approximately 200 mW. An objective lens (MPLFN50x, Olympus) is used to collect the red fluorescence from the diamond. Long pass (BLP01-633R, Semrock) and dichroic (LM01-552, Semrock) filters are placed before the lock-in camera (heliCam C3, Heliotis) to separate the 532-nm pump laser. We use two microwave generators (SG394, SRS) and a single RF source (2.16 MHz, WX1282C, Tabor Elec.) for the double resonance with hyperfine driving. Each SG394 was mixed (ZX05-43MH+, mini-circuits) with WX1282C and then the outputs are combined (ZX10-2-42S+, mini-circuits) for the double resonance with hyperfine driving. The mixed and combined signals are sent to an amplifier (ZHL-16W43-S+, mini-circuits). A switch (ZASWA-2-50DRA+, mini-circuits) is placed before the amplifier to control the delivery of the microwave to the omega-shaped coil. A TTL pulse generator (PB24-100-4k-PCI, spincore) is used to control the switch.
A frequency modulation (square wave) is selected for the lock-in detection. The modulation depth is 300 kHz and the modulation frequencies (f$_{mod}$) are 2.5 or 10 kHz, depending on the temporal resolution ~\cite{ele17}. The phases ($\phi_1, \phi_2$) of the frequency modulation are controlled using an arbitrary waveform generator (AWG, 33522B, Keysight). For magnetic field and temperature detections, their phase differences are maintained at 180$^{\circ}$ and 0$^{\circ}$, respectively. The phases used for imaging magnetic fields are shown in Fig.~\ref{fig2}. An external trigger for the camera (Cam$_{trig}$) comes from the TTL pulse generator and is synchronized to the phases. The trigger internally initiates an integration of the fluorescence signals during four periods, (I$^+$, Q$^+$, I$^-$, and Q$^-$). The in-phase (I = I$^+$ - I$^+$) and quadrature (Q = Q$^+$ - Q$^-$) images are automatically calculated by the camera. A single cycle is composed of four periods, and a single frame is a repetition of a single-cycle N$_{cyc}$ times.
\begin{figure}[t]
\centering
\includegraphics[width=0.45 \textwidth]{fig3.pdf}
\caption{ODMR spectra at two aligned cases. (a) ODMR spectra, bias fields aligned along $<$111$>$ and $<$001$>$ of the crystal axes; (b) ODMR of the $<$001$>$ aligned case (SR) compared to the case of hyperfine driving (SR + HF). The contrast of the SR + HF case is enhanced by 2.4 times due to the HF driving. The contrast is further increased by adopting double resonance (DR + HF) with the same phase}
\label{fig3}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig4.pdf}
\caption{Volume-normalized magnetic field sensitivity map. (a) Two-dimensional map of the volume-normalized field sensitivity, $\eta_V$ and the scale bar represents 10 $\mu m$; (b) A histogram of the sensitivity map within the red circled area. Mean volume sensitivity is 43.9 $\mathrm{nT\cdot\mu m^{1.5}/Hz^{0.5}}$}
\label{fig4}
\end{figure}
The hyperfine(HF) interaction between the NV and N nuclear spin results in a reduced ODMR contrast, which is detrimental to the magnetic field sensitivity. A single frequency-sweeping ODMR spectrum due to HF interaction can be expressed as 1-$\sum_{p=-1}^{1}\frac{\delta \nu^2}{\delta \nu^2 +4(\omega-(\omega_0 + p ~HF))^2}$, where C,$\nu, \omega, \omega_0$, HF represent contrast, ODMR linewidth, applied MW frequency, resonant MW frequency, and hyperfine splitting (2.16 MHz in our case), respectively. The central contrast can be enhanced by up to three times if three equally spaced MW frequencies are swept simultaneously, 1-$\sum_{p,q =-1}^{1}\frac{\delta \nu^2}{\delta \nu^2 +4((\omega + q~HF)-(\omega_0 + p ~HF))^2}$.~\cite{bar16} In practice, the enhancement is approximately two because of power-broadened hyperfine features.
The negatively charged NV$^-$ has an electronic spin triplet (S = 1) state with a temperature-dependent zero-field splitting, D $\sim$ 2.87 GHz at room temperature, between the \ket{m_s=0} and degenerated \ket{m_s = \pm1}~\cite{doh12}. The degenerated states are splitted by the Zeeman effect as an external magnetic field is applied. The Hamiltonian for the NV$^-$ in a magnetic field ($|B| > 1 mT$) can be approximated as: ~\cite{keh19,har21}
\begin{equation}
\frac{H}{h} = D(T)S_z^2 + \frac{\gamma}{2\pi} B_{NV}S_{z},
\end{equation}
where z, Sz, and B$_{NV}$, denote the NV symmetry axis, the dimensionless spin-1 operator, the external magnetic field projected along the NV symmetry axis, and $\gamma/2\pi$ is the gyromagnetic ratio (28 GHz/T), respectively. The hyperfine interaction and spin-stress coupling parameters are ignored. The resonance frequencies can be expressed as: ${f_1 = D(t) -\gamma B_{NV}(t)}$ for \ket{m_s=0}$\leftrightarrow$ \ket{m_s=-1} and ${f_2 = D(t) + \gamma B_{NV}(t)}$ for \ket{m_s=0}$\leftrightarrow$\ket{m_s=1} where ${D(t) = D_0 + \Delta D(t)}$ and ${B_{NV}(t) = B_{NV0} + \Delta B_{NV}(t)}$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig5.pdf}
\caption{Sub-ms temporal resolution. (a) A pulsed (triangular shape) voltage (dashed blue line) is applied at a coil of which inductance and resistance are 1.8 mH and 2 $\Omega$, respectively. An expected current (dashed black line) in the coil is simulated by LTspice. The measured magnetic field (solid red line), parallel to the z-axis of the crystal, is found from a single pixel in the middle of the sensitivity map in Fig.~\ref{fig4} (a) and shown in the upper panel. A detailed view, (b), in the time domain shows that an inductive delay in the current, $\sim$ 0.9 ms, is qualitatively captured by the magnetic field measurement. }
\label{fig5}
\end{figure}
A double resonance (DR) simultaneously drives the resonance frequencies ($f_1$, $f_2$) using two microwave generators (MW1 and MW2) with the modulation frequency ($f_{mod}$) and phases ($\phi_1$, $\phi_2$). The output signals of the lock-in amplifier at $f_1$ and $f_2$ are $S_1(t) =\alpha [\Delta D(t) - \gamma B_{NV}(t)]$ and $S_2(t) =\alpha [\Delta D(t) + \gamma B_{NV}(t)]$, respectively, where $\alpha$ is the slope of the lock-in amplifier. If we apply DR with the same phase, i.e. $\phi_1$ = $\phi_2$, then the lock-in signal ($S_{LIA}$) is only sensitive to the temperature, $S_{LIA}$ = 2 $\alpha \Delta$ D(t). Alternatively, it depends only on the magnetic field ($\Delta B_{NV}$) if $ | \phi_1 - \phi_2 |$ = $\pi$ , $S_{LIA}$ = $2 \alpha\Delta B(t)$ ~\cite{fes20, woj18b, shi22, hat21}. Additionally, the sensitivity of the DR method is expected to be enhanced by $\sim$4/3 times compared to the sensitivity of the single-resonance method ~\cite{fes20}.
\section{Results}
The shot-noise-limited continuous wave (CW) magnetic field sensitivity,$\eta_{CW}$, is given by~\cite{bar16,dre11}:
\begin{equation}
\eta_{CW} = \frac{4}{3\sqrt3}\frac{h}{g_e\mu_B}\frac{\Delta\nu}{C\sqrt R},
\end{equation}
where R is the photon-detection rate, $\Delta\nu$ is the linewidth, and $C$ is the ODMR contrast. To minimize the $\eta_{CW}$, we adopt several methods to obtain a higher ODMR contrast in Eq. (2). The first method is projecting a magnetic field equally along all the NV axes. We compare the ODMR spectra where the external magnetic fields are aligned along the $<$111$>$ and $<$001$>$ directions of the crystal, as shown in Fig.~\ref{fig3}(a). When the magnetic field is along the $<$001$>$ direction of the crystal, the ODMR contrast can be maximized~\cite{bar16,shi22}. Hereafter, we fix the external field along the $<$001$>$ direction. The second method is to simultaneously excite the three HF features (SR + HF) instead of exciting a single frequency (SR). The ODMR contrast is improved by 2.4 times, compared to that in SR, as shown in Fig.~\ref{fig3}(b). The third method involves applying DR along with HF driving (DR + HF). This further enhances the contrast compared to that in SR + HF, as shown in Fig.~\ref{fig3} (b). DR is essential for minimizing errors in the magnetic field due to temperature drift in the system~\cite{fan13,mam15,shi22}.
The magnetic field sensitivity can be expressed as $\eta = \delta B\sqrt T$, where $\delta B$ is the minimum detectable magnetic field, and T is the measurement duration. The minimum magnetic field is given by the standard deviation of a series of measurements~\cite{har21,par22b} . To determine the minimum magnetic field, a test field is applied along the z-axis of the crystal. The test field is found to be 6.8 $\mu$T, and the projected magnetic field along the NV axes is 4 $\mu$T. The frame rate of the camera is 114 Hz (8.8 ms, $f_{mod}$ = 2.5 kHz, 22 cycles), and 110 frames are collected for the estimation.
A two-dimensional map of the volume-normalized magnetic field sensitivity, $\eta_V = \eta \sqrt V$, is shown in Fig. ~\ref{fig4}(a), where the field of view is approximately 46 $\times$ 46 $\mu m^2$, the pixel size is 0.54 $\times$ 0.54 $\mu m^2$, and the sensor volume, V, is 11.7 (0.54 $\times$ 0.54 $\times$ 40) $\mu m^3$. The mean $\eta_V$ within the red circled area is 43.9 $\mathrm{nT\cdot\mu m^{1.5}/Hz^{0.5}}$ and the mean per pixel $\eta$ is 12.8 $\mathrm{nT/Hz^{0.5}}$. The histogram of the $\eta_V$ within the area shows a Gaussian-like distribution due to the beam shape, as shown in Fig.~\ref{fig4} (b). Illuminating the NV diamond based on the total internal reflection is expected to result in a larger field of view and a more uniform sensitivity distribution than that in the current scheme~\cite{cle15,har21}.
To demonstrate a sub-millisecond temporal resolution, the frame rate is increased to 2500 Hz, and 200 frames are acquired, where the modulation frequency is set to 10 kHz. The increased modulation frequency decreases the signal-to-noise ratio due to the wider bandwidth. A series of pulsed voltages are applied to a coil with a diameter of 10 cm, an inductance of 1.8 mH, and a resistance of 2 $\Omega$. The voltage pulse has a triangular shape, and its polarity is changed within 2 ms and repeats every 10 ms, as shown in Fig.~\ref{fig5} (b). Because the dimensions of the coil are significantly larger than the field of view (Fig.~\ref{fig4} (a)), the magnetic field produced by the pulsed voltage is uniform.
A single acquisition of the magnetic field from the voltage pulse at a single central pixel is shown in Fig.~\ref{fig5}(a). The dashed blue line represents the applied voltage (scaled), and the dashed black line represents the expected current in the coil. The solid red line represents the magnetic field along the z-axis of the crystal. Even in a single measurement, we can distinguish $\pm$ 4 $\mu$T magnetic pulse trains from the noise. The expected delay between the current (magnetic field) and the voltage is approximately 0.9 (1.8/ 2) ms. A close-up in the time domain, Fig.~\ref{fig5}(b), indicates that our system can capture the transient behavior with sub-millisecond temporal resolution. The standard deviation of the noise level ($\approx$ 1 $\mu$T) during the acquisition duration (0.4 ms) leads to a per-pixel sensitivity of 20 $\mathrm{nT/Hz^{0.5}}$. These observations support the nanotesla sensitivity with sub-millisecond temporal resolution.
\section{Discussion}
In this study, we optimize the volume-normalized magnetic field sensitivity of NV center ensembles using a lock-in camera. The mean per pixel volume-normalized magnetic field sensitivity of 43.9 $\mathrm{nT\cdot\mu m^{1.5}/Hz^{0.5}}$ and the sub-ms temporal resolution are obtained at a relatively low optical power density of 0.12 mW/$\mu m^2$. However, we still need to improve the sensitivity to less than 1 $\mathrm{nT\cdot\mu m^{1.5}/Hz^{0.5}}$ to visualize neuronal networks~\cite{kar18,kar21}. In this section, we will discuss how we can further improve the magnetic field sensitivity.
Photonic structures such as diamond nano-pillars will enhance the volume-normalized sensitivity by improving readout fidelity~\cite{mom15, mcc20}. It has been reported that the sensitivity can be improved by more than four times owing to increased photoluminescence and spin coherence time by nano-pillar~\cite{mcc20}. An additional antireflective coating of 600 - 800 nm on the diamond further increases the photoluminescence further ~\cite{web19}.
The inhomogeneous spin dephasing time, $T_2^*$, can be extended by applying decoupling sequences. The dipolar coupling between NV centers and substitutional nitrogen (P1) can be suppressed by driving P1 spins~\cite{bau18, bal19}. Bauch $et$ $al.$ and Balasubramanian $et$ $al.$ reported that $T_2^*$ in a high P1 density increased more than four times using spin-bath driving~\cite{bau18, bal19}. Moreover, Balasubramanian $et$ $al.$ decoupled NV-NV interaction by adopting WAHUHA sequence and additionally extended $T_2^*$ by ten times~\cite{bal19}. These methods, combined with our technique, could reduce the volume-normalized sensitivity to less than 1 $\mathrm{nT\cdot\mu m^{1.5}/Hz^{0.5}}$, which is an essential tool for understanding neuronal connectivity~\cite{hal12, kar18, kar21}.
Illuminating the NV layer uniformly using total internal reflection geometry improves the magnetic sensitivity distribution and increases the field of view up to the millimeter scale~\cite{cle15, tur20}. This can be utilized to detect magnetic fields from an integrated circuit and 3-D current distribution in a multi-layer printed circuit board~\cite{tur20, oli21}. Combined with the sub-millisecond temporal resolution, the wide field of view could contribute to imaging transient events, which could be missed by scanning-based systems such as giant-magneto resistive (GMR) or superconducting quantum interference device (SQUID)-based current mapping equipment ~\cite{tur20, gau14}.
\section{Conclusion}
In conclusion, we have obtained a mean per pixel volume-normalized magnetic sensitivity of 43.9 $\mathrm{nT\cdot\mu m^{1.5}/Hz^{0.5}}$ and a sub-ms temporal resolution using NV center ensembles and a lock-in camera. The HF driving, DR, and exploitation of the four NV axes are adopted with CW lock-in detection to reach the sensitivity. These methods could be a step forward for visualizing microscopic distributions of sub-nanotesla changes due to neuronal currents in real-time, as well as defects in a packaged battery~\cite{zha21x}.
\section*{Declarations}
\subsection*{Acknowledgments}
The authors thank Kiwoong Kim for valuable discussions and Heloitis AG for experimental assistance in implementing the camera.
\subsection*{Funding}
This research was supported by Institute of Information $\&$ communications Technology Planning $\&$ Evaluation (IITP) grants funded by the Korea government (MSIT) (No.2019-000296, No.2021-0-00076) and a grant (GP2021-0010) from Korea Research Institute of Standards and Science.
\subsection*{Availability of data and materials}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
|
\section{Introduction}
Many real-world scenarios can be modeled as \textbf{team-competitive Markov games}~\citep{lowe2017multi} (also called \textit{mixed cooperative-competitive games}), where each agent receives an individual reward for its actions and a shared reward with other agents in the same team for a joint outcome. For example, while playing soccer, each player may receive a reward for their individual performance and a shared reward for the team's victory. In some social interactions, e.g., companies vying for customers, each employee may receive a reward for their individual performance and a shared reward for the team performance.
\textbf{Roles} emerge naturally in such reward scenarios, since they allow agents within a team to cooperate effectively. For soccer players, the roles could be ``striker'' and ``defender'', while for company employees, the roles could be ``production'' and ``sales''. In this work, we study the problem of learning policies with emergent roles in a reinforcement learning (RL) setting for team-competitive Markov games.
We use multi-agent reinforcement learning (MARL) to learn the optimal policies of individual agents in this setting. While many popular deep MARL techniques exist for the shared-reward setting
~\citep{gupta2017cooperative, mordatch2018emergence, rashid2018qmix,son2019qtran, wang2019learning}, relatively fewer techniques exist for the mixed-reward setting.
Multi-agent actor-critic algorithms~\citep{lowe2017multi,iqbal2019actor}, following the centralized training and decentralized executions concept, are the most popular MARL approaches in the mixed-reward setting, which is needed to model the team-competitive scenarios. MADDPG \citep{lowe2017multi}
is one of the earliest techniques for learning deterministic policies in mixed reward settings using the actor-critic framework. MAAC ~\citep{iqbal2019actor}
introduces attention mechanisms in critic modules to selectively pay attention to information collected from other agents, resulting in better scalability in complex multi-agent environments.
However, to the best of our knowledge, role-oriented policy learning in the actor-critic framework, particularly in competitive games, is yet unexplored.
On the other hand, the concept of predefined roles for agents has been explored to reduce design complexity in many MARL systems~\citep{spanoudakis2010using, bonjean2014adelfe}.
There are few works addressing dynamic role emergence based MARL in domain-specific problems~\citep{barrett2015cooperating,leottau2015study,urieli2011optimizing,roy2020promoting,zhang2021hierarchical,khaleefah2021exploring}.
Recently, Wang et al.~\citep{wang2020roma} introduce automatic role emergence in MARL for cooperative tasks in their system called \textsc{ROMA}. It proposes a latent role-based task decomposition among the agents built over the QMIX~\citep{rashid2018qmix} platform.
In a similar line, Wang et al.~\cite{wang2020rode} propose RODE, a bi-level role-based learning framework where joint action-space is decomposed into restricted role-action spaces by clustering actions by their effects. Finally, it employ a role-selector to search role at a smaller role space, while role policies learn at a reduced action-observation space, significantly improving learning efficiency. Liu et al.~\citep{liu2022rogc} present ROGC, a role-oriented graph convolution based MARL framework, where roles are generated by classifying agents, graph convolution module is emplyed for efficient intra-role communication and finally effective individual policies are generated by embedding role information into the algorithm.
However, all of the above frameworks are built over QMIX, where a global utility function is composed of per-agent local utility functions, and the composition relies upon the assumption that all agents share the same reward. Hence, this setting can not be immediately extended to a mixed-reward setting, prohibiting role-oriented policy learning in the presence of opponents through joint learning of any form.
Recently, opponent-aware actor-critic algorithms have been explored in many settings, including
learning time-dynamical opponent models TDOM-AC~\cite{wen2019probabilistic} , maximum entropy regularized opponent models ROMMEO~\cite{tian2019regularized}, opponent model employing variational Bayes PRR~\cite{tian2022multi}, etc. These works validate the necessity of opponent models for improving performance in competitive games. However, they are primarily limited to one-to-one competitions and do not consider role-aware opponent modeling. However, in a team competition setting, where each agent employs a role-aware policy within the team, considering roles in opponent modeling can benefit agents by reducing design complexity.
In this work, we propose \textsc{RAC}, an actor-critic algorithm that combines role learning and role-aware opponent modeling in mixed-reward setting.
\textsc{RAC}\ uses the actor-critic framework based on MAAC \cite{iqbal2019actor}, which allows the critic to use joint observations of all agents, hence incorporating the opponent's experience in the critic.
Our primary contributions are two-fold. First, we incorporate role-encoder in the actor-critic framework that learns dynamic and diverse emergent roles per team.
Moreover, for opponent modeling, we incorporate an opponent role predictor network for each team, which is accessible to the policy at the execution time.
The opponent role prediction network is trained jointly along with the role encoder, critic, and policy network. In particular, during centralized training, the opponent role predictor is trained to mimic the actual role encoding distribution of the opponent. During execution, that benefits agents to select their response.
The learned roles are dynamic depending on the agent's trajectory and diverse between agents within the team. Experiments using the proposed episodic board games, Touch-Mark\, and Market, show that \textsc{RAC}\ agents outperform MAAC\ agents in direct competitions. Furthermore, our experiments demonstrate that the dynamic roles learned by \textsc{RAC}\ conform to the intuitive role-specific behaviors for each game, thus confirming the effectiveness of emergent roles.
In summary, our contributions are as follows:
\begin{itemize}
\item We propose \textsc{RAC}{,} the first opponent-aware role-oriented policy learning in actor-critic framework, designed for mixed reward setting, as per our knowledge.
\item \textsc{RAC}\ encourages diverse roles within a team and ensures that the role encoding captures the agent trajectory.
\item Most importantly, \textsc{RAC}\ learns an opponent role predictor module, utilizing opponent role encodings during the centralized training that it employs in policy during the execution time. Eventually, the opponent role predictor guides agents in choosing strategic responses against adversaries.
\item Empirically, we validate improved policy learning of \textsc{RAC}{, } through increased reward as well as demonstrate the identifiability of learned roles with intuitive role-specific behaviors of agents in two episodic board games, Touch-Mark\ and Market{.}
\end{itemize}
\section{Related Literature}
Multi-agent Actor-critic methods, primarily following the concept of centralized training and decentralized executions, are one of the most dominant and popular techniques in deep Multi-agent Reinforcement Learning (MARL) approaches~\citep{sukhbaatar2016learning,gupta2017cooperative,lowe2017multi,foerster2017stabilising,mordatch2018emergence,rashid2018qmix,foerster2018counterfactual}.
Among these works, MADDPG~\citep{lowe2017multi}, a multi-agent adaptation of DDPG~\citep{lillicrap2016continuous}, considers continuous action space and deterministic policies in mixed-reward settings and successfully trains the agents complex coordination strategies in cooperative as well as competitive scenarios.
MAAC~\citep{iqbal2019actor}, an extension of SAC~\citep{haarnoja2018soft}, introduces attention mechanism in critic modules to selectively pay attention to information collected from other agents, resulting in massive scalability in complex multi-agent environments. Also, there exists a series of value-function-factorization-based methods that train decentralized policies in a centralized end-to-end fashion, employing a joint action value network~\citep{rashid2018qmix}.
There are many follow-ups on actor-critic-based MARL algorithms, addressing a variety of issues, namely SAC~\citep{qu2020scalable} for improving scalability, SEAC~\citep{christianos2020shared} for sharing experience, LICA~\citep{zhou2020learning} for the credit assignment problem, TESSERACT~\citep{mahajan2021tesseract} for tensorizing the critics, Bilevel Actor Critic~ \cite{zhang2020bi} for multi-agent coordination problem with unequal agents,
DAC-TD~\cite{figura2022cooperative} for training agents in a privacy-aware framework, VDAC~\cite{su2021value} for combining value-decomposition framework with actor-critic, Scalable Actor Critic~\cite{lin2021multi} for scalable learning in stochastic network of agents, etc.
However, none of these works consider the role emergence paradigm in their model.
On the other hand, role division is quite common and efficient for accomplishing any complex collective task in human society~\citep{butler2012condensed}. Following that, many multi-agent systems also decompose the task into pre-defined roles, resulting in reduced design complexity and faster learning ~\citep{wooldridge2000gaia,padgham2002prometheus,cossentino2005passi,spanoudakis2010using,bonjean2014adelfe}.
In recent times, Majumdar et al.~\citep{majumdar2020evolutionary} present a line of work where MADDPG agents learn different strategies by de-coupling and automatically weighting individual and global goals in a population-based training paradigm.
Liu et al.~\citep{liu2021coach} consider the problem of coordination with dynamic composition, where a coach agent with global view distribute individual strategies to dynamically varying player agents. Jiang et al. ~\citep{jiang2021emergence} promote sub-task specialization via emergence of individuality through generating dynamic intrinsic rewards. Su et al.~\cite{su2022divergence} propose another actor-critic algorithm where agents learn diverse policies in cooperative team games.
Also, there is a series of works for role-oriented MARL for specialized domains like Robo-soccer ~\citep{barrett2015cooperating,leottau2015study,urieli2011optimizing,ossmy2018variety}, football environment~\citep{roy2020promoting}, swarn systems~\citep{zhang2021hierarchical}, image feature extraction tasks~\citep{khaleefah2021exploring} showing how complex policies can be learned by decomposing them into simpler sub-policies.
In contrast, Wang et al. propose \textsc{ROMA}~\citep{wang2020roma} and RODE~\citep{wang2020rode}, two dynamic and adaptive role-oriented MARL frameworks built over QMIX~\citep{rashid2018qmix} in a quite general setting and show dynamic role emergence within the team on SMAC benchmark tasks.
Recently, Liu et al. present ROGC~\citep{liu2022rogc}, a graph convolution based role-oriented MARL framework. However, all of them are inherently restricted to shared reward settings, i.e., only cooperative tasks.
Our work closely resembles~\citep{wang2020roma}, with the main distinction being that our work incorporates opponent modeling along with adopting team-based role learning to the actor-critic framework to exploit the advantage of training in the presence of opponents.
There are few recent works exploring opponent modeling in actor-critic framework, namely, TDOM-AC~\cite{tian2022multi} for learning time dynamical opponent model, ROMMEO~\cite{tian2019regularized} for learning maximum entropy regularized opponent model, PRR~\cite{wen2019probabilistic} for learning opponent model employing variational Bayes, to name a few. They primarily consider the setting of two-player competitive games, where each agent maintains a prototype of opponent policy. However, they do not consider team competition settings. Extending their framework to our role-aware competitive team setting is not immediate.
\subsection{Preliminaries}
\begin{table}[t]
\begin{tabular}{c l}
\toprule
Symbol & Meaning \\\midrule
$\Scal,\Acal$ & State and action Space \\
$\Pcal,\Rcal$ & Transition and reward functions \\
$N, K$ & Total Number of agents and teams \\
$\Ncal(k)$ & Set of agents in team $k$\\
$\obar=(o_1,\dots,o_N)\in \Ocal $ & observation vector\\
$\abar=(a_1,\dots,a_N)\in \Acal $ & action vector\\
$\pi=(\pi_1,\dots,\pi_N)$ & Agent policies \\
$\tau^t=(\tau_1^t,\dots,\tau_N^t)$ & Trajectory till time $t$\\
$Q=(Q_1,\dots,Q_N)$ & Value-action function \\
$h_{S}=(h_S^1,\dots,h_S^K)$ & Self-role encoder \\
$h_{O}=(h_O^1,\dots,h_O^K)$ & Opponent-role encoder \\
$\rho=(\rho_1,\dots,\rho_N)$ & Self-role \\
$\hrho=(\hrho_1^o,\dots,\hrho_N^o)$ & Predicted opponent-role \\
$\lambda$ & Decay factor \\
$\Lcal_{Q}$ & Critic loss \\
$\Lcal_{MI}$ & Mutual information loss\\
$\Lcal_{D}$ & Divergence loss \\
$\Lcal_{Opp}$ & Opponent loss \\
\bottomrule
\end{tabular}
\caption{List of important notations.}
\label{tab:notations}
\end{table}
A Markov game is characterized by the tuple $\langle \Scal,\Acal,\Pcal,\Rcal \rangle$, where $\Scal$ denotes the set of states, $\Acal$ denotes the set of actions for each of the $N$ agents. Hence the joint space of actions becomes $\Acal^N$.
The state transition function maps every state and joint-action combination to a probability over future states. $\Pcal \colon \Scal \times \Acal^N \rightarrow \PP(\Scal)$.
In case of team-competitive games (also called mixed cooperative-competitive
games \cite{lowe2017multi}), let $\Kcal$ denote the set of $K$ teams, and $\Ncal(k)$ be the set of agents of team $k$,$k \in \Kcal$.
The reward function $\Rcal = (r_1,\dots,r_N) \colon \Scal \times \Acal^N \rightarrow \mathcal{R}^N $ specifies the reward scheme ($r_i$) for each agent $i$.
The per-agent reward function allows modeling of team competitive games~\citep{lowe2017multi}, which are the focus of this paper.
Also, in the decentralized execution setting, let $o_i \in \Ocal$ denote the observation for agent $i$.
Each agent $i$ learns a policy $\pi_i \colon \Ocal \rightarrow \PP( \Acal )$, which is a mapping from it's own observation to a distribution over it's action set $\Acal$.
We also define the trajectory $\tau_i(t) \in T = (\Ocal\times\Acal)^*$ for each agent $i$ till time $t$ as a sequence of observation-action pairs $\{ (o_i^j,a_i^j), j=1,\dots,t \}$. We also denote the joint action and observations vectors as: $\bar{a}=(a_1,\dots,a_N)$ and $\bar{o}=(o_1,\dots,o_N)$. Agents aim to learn their policy $\pi_i$ by optimizing the agent-specific cumulative discounted reward function $\Jcal_i$:
\begin{align}
\Jcal_i(\pi_{i}) &= E_{\abar \sim \pi, s \sim \Pcal} \left[ \sum_{u=0}^{\infty} \gamma^{u} r_{i} (s^u, \abar^u) \right]
\label{eq:J_pol_grad}
\end{align}
\iffalse
Corresponding policy gradient update for agent $i$ takes following form:
\begin{align}
\nabla_{\pi_i} \Jcal_i(\pi_{i})
&= \nabla_{\pi_i} \log (\pi_{i} (a_{i}^t | o_{i}^t) ) \left[ \sum_{\hat{t}=t}^{\infty} \gamma^{\hat{t}-t} r_i(s^{\hat{t}}, \bar{a}^{\hat{t}}) \right] \\
&= \nabla_{\pi_i} \log (\pi_{i} (a_{i}^t|o_{i}^t ) Q_i(\bar{o}^t,\bar{a}^t)
\end{align}
where $Q_i$ is the Value-action function of agent $i$
\fi
\sloppy
\iffalse
We follow a state-of-the-art multi-agent actor-critic method, MAAC~\citep{iqbal2019actor}, which supports centralized training of agents' policies with individual rewards. MAAC learns policy $\pi_i$ for agent $i$ using the policy gradient updates for the agent-specific discounted reward function $\Jcal_i$:
\begin{align}
\nabla_{\pi_i} \Jcal_i(\pi_{i}) &= \nabla_{\pi_i} \log (\pi_{i} (a_i | o_i) ) \sum_{\hat{t}=t}^{\infty} \gamma^{\hat{t}-t} r_{\hat{t}} (s_{\hat{t}}, a_{\hat{t}}) \\
&= \nabla_{\pi_i} \log (\pi_{i} (a_t|s_t) ) Q_i(\bar{o},\bar{a})
\end{align}
In MAAC, the critic function for agent $i$ depends on the embedding of agent $i$, $g_i(o_i,a_i)$, and the attention-weighted average reward of other agents $x_i$ as:
\begin{align}
& Q_i(\bar{o},\bar{a}) = f_i ( g_i(o_i,a_i), x_i)\\
& x_i = \sum_{j \neq i} \alpha_j \Vcal g_j( o_j,a_j)
\end{align}
where $f_i$ and $\Vcal$ are neural networks.
$\alpha_j$ is the dynamic-attention function given by agent $i$ to agent $j$, and depends on embeddings $g_i(o_i,a_i)$ and $g_j(o_j,a_j)$. The critic function is learned using a temporal difference loss calculated on trajectories collected using current and past policies, which are stored in a replay buffer.
\fi
\subsection{Multi-Agent Actor Critic}
Actor-critic methods and their extensions are widely used in multi-agent reinforcement learning \cite{haarnoja2018soft,iqbal2019actor} due to their flexibility and low variance reward estimates, which lead to faster learning of policy parameters. A popular extension is the soft actor-critic \cite{haarnoja2018soft}(SAC), where the reward of policy gradient (Eqn \ref{eq:J_pol_grad}) is augmented with the expected entropy of the policy distribution in each state. Hence their reward function becomes:
\begin{align}
\Jcal_i(\pi_{i}) &= E_{\abar \sim \pi, s \sim \Pcal} \left[ \sum_{u=0}^{\infty} \gamma^{u} (r_{i} (s^u, \abar^u) + \alpha \mathcal{H}(\pi_i(s^u)) \right]
\end{align}
where $\alpha$ is the user defined regularization coefficient. Note that they do not explicitly discuss multi-agent RL in team-competitive setting.
Subsequently, multi-attention actor critic (MAAC) \cite{iqbal2019actor} extended SAC to support centralized training of agents' policies with individual rewards. The learning algorithm in MAAC has two broad components: (1) SAC-style update of policy $\pi_i$ for the $i^{th}$ agent and (2) Attentive learning of the critic function $Q_i$ for the $i^{th}$ agent. The policy update is carried out using the policy gradient given by:
\begin{multline}
\nabla_{\pi_i} \Jcal_i(\pi_{i}) = E_{\obar \sim D,\abar \sim \pi} [\nabla_{\pi_i} \log (\pi_{i} (a_{i}|o_{i})
( Q_i(\bar{o},\bar{a}) \\
- b(\obar,\abar_{-i})
- \alpha \log (\pi_i (a_i | o_i ))) ]
\label{eq:pol_grad_maac}
\end{multline}
where $\alpha$ is the regularization constant for the entropy term. $D$ is the replay buffer that contains the tuple $(\obar, \abar, \rbar, \obar^{'})$ for each step of the episodes generated by the policies till the current RL training time $t$. Here action $\abar$ at observation $\obar$ generates reward of all agents $\rbar$ and next observation $\obar^{'}$. $b(o,a_{-i})$ is the multi-agent baseline calculated as follows.
\begin{align}
b(\obar,\abar_{-i}) &= E_{a_i \sim \pi_i(o_i) } \left[ Q_i(\obar,(a_i,\abar_{-i}))\right] \\
&= \sum_{a_i \in \mathcal{A}} \pi_i(a_i | o_i) Q_i(\obar,(a_i,\abar_{-i}))
\end{align}
The baseline is used calculate the advantage function, which helps in solving the multi-agent credit assignment problem \cite{iqbal2019actor}.
The critic function $Q_i$ for agent $i$ depends on the embedding $g_i(o_i,a_i)$ of agent $i$ and the attention-weighted average reward of other agents $x_i$ as:
\begin{align}
& Q_i(\bar{o},\bar{a}) = f_i ( g_i(o_i,a_i), x_i)\\
& x_i = \sum_{j \neq i} \alpha_{ij} \Vcal ( g_j( o_j,a_j) ) \\
&\alpha_{ij} \propto \exp( v_i^T W_i^T W_j v_j)
\end{align}
where $f_i$ and $\Vcal$ are feedforward neural networks.
$\alpha_{ij}$ is the dynamic-attention function given by agent $i$ to agent $j$, and depends on embeddings $v_i = g_i(o_i,a_i)$ and $v_j = g_j(o_j,a_j)$. $W_i$ and $W_j$ are agent specific ``key'' transformation matrices for the attention function. The critic function is learned using a temporal difference loss calculated on trajectories collected using current and past policies, which are stored in a replay buffer $D$. The loss function for updating the critic module for agent $i$ takes following form:
\begin{multline}
\Lcal_Q = \sum_{i=1}^{N} E_{(\obar,\abar,r,\obar^{'}) \sim D } \left[ (Q_i(\obar,\abar) - y_i )^2\right] \text{ , where}\\
y_i = r_i + \gamma E_{\abar^{'} \sim \bar{\pi}(\obar^{'}) } \left[ \bar{Q}_i(\obar^{'}, \abar^{'}) - \alpha \log ( \bar{\pi}_i (a_i | o_i) ) \right]
\label{eq:critic_loss_maac}
\end{multline}
where $\bar{Q}_i$ and $\bar{\pi}_i$ are target critic and target policy for agent $i$. In this work, we build on the multi-agent actor-critic framework described above to incorporate learning of emergent roles in each agent. Next, we describe a popular role learning framework in the cooperative setting.
\subsection{Role-aware policy learning}
Roles have been used in many MARL applications for promoting cooperation and coordination among agents \cite{barrett2015cooperating,roy2020promoting,urieli2011optimizing,zhang2021hierarchical}.
Recently, Wang et al. proposed \textsc{ROMA}{,}\cite{wang2020roma}, a MARL framework for learning emergent roles in a cooperative game setting.
In \textsc{ROMA}, the roles are designed to exhibit two crucial properties: (1) \textbf{Identifiability: } role should be identifiable with its behavior and diverse enough to fulfill the task, and (2) \textbf{Specialization: } agents with similar roles should specialize in similar responsibilities. For identifiability, it maximizes the mutual information $I( \rho_i ; \tau_i )$ between the role distribution $p(\rho_i|o_i)$, and the trajectory-conditional role distribution
$p(\rho_i| \tau_i,o_i)$
For ensuring specialization and diversity, it maintains a dissimilarity model $d_{\phi}$ to measure the distance of trajectories of two agents and seek to maximize $I( \rho_i ; \tau_j ) + d_{\phi} (\tau_i, \tau_j) $. With these regularizers, \textsc{ROMA}\ ensures agents adopt role-aware policies within a team in each trajectory.
\textsc{ROMA}\ builds on the QMIX~\citep{rashid2018qmix}, which is designed for settings where the agents receive a shared reward for joint observation and actions. \textsc{ROMA}\ also works within the Q-learning mode without any additional policy network. It learns per-agent local utility networks, $\Qcal_i$ output of which are passed to a mixing network to compute global TD-loss.
During execution, each agent acts according to optimum policy derived from its' local utility function.
In \textsc{ROMA}\, the local utility function $\Qcal_i(o_i,a_i,\rho_i)$ takes the agent's role $\rho_i$ also as input. The role is $\rho_i$ defined as a latent (emergent) variable with a distribution, conditioned on
the trajectory of the agent $\tau_i$.
However, the Q-learning framework used in \textsc{ROMA}\ cannot be trivially extended to the competitive setting since the composition of local utility functions $\Qcal_i$ into one global utility function $\Qcal$ strongly relies upon the assumption of the common reward of all agents which does not hold in a competitive setting with mixed-rewards.
Moreover, the local utility function $\Qcal_i$ takes only observation and action of $i$-th agent ($o_i$ and $a_i$) as input, and therefore agents cannot learn from the experience of opponents (other agents with decoupled rewards).
In the next section, we develop a novel formulation for learning emergent roles in competitive setting in the actor-critic framework, which is better suited towards incorporating mixed-reward.
In the actor-critic framework, the critic function $Q_i(\bar{o},\bar{a})$ takes observations and actions of all agents as input, hence facilitating the sharing of information among agents with decoupled reward and agents benefits from learning opponent-aware policies by exploiting the shared opponent information available during the training.
\subsection{Opponent-aware Role learning}
\begin{figure*}[!t]
\centering
\subfloat{\includegraphics[width=.8\textwidth]{Results/RAC_new.pdf}}
\vspace{-5mm}
\caption{ A schematic diagram of our approach.
The self-role encoder $h_S$ generate the role $\rho_i$ for the $i$-th agent and opponent role predictor $h_O$ predicts roles $\hat{\rho}_{i}^o$ for the opponents. Policy $\pi_i$ takes own observation $o_i$ as well as generated role encodings ($\rho_i, \hat{\rho}_{i}^o$) to generate action $a_i$.
Critic $Q_i$ takes both $((o_i, \rho_i, \hat{\rho}_{i}^o), a_i)$ to compute the state-action value.
Critic loss trains the critic module, whereas mutual information loss, diversity loss, and opponent loss train the role encoder modules.
The framework can be trained in an end-to-end manner. }
\label{fig:diagram}
\end{figure*}
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$eps = 1 \cdots $ max episodes}{
$\tau_i^{0} \gets 0$\\
$D_i \gets \emptyset, \forall i $\\
Reset the environment, and get initial $o_i^{1}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
Sample roles $\rho^{t}, \hrho^{t}$ using~(\ref{eq:roles}) for all agents\\
Sample action from respective policies $a_i^{t} \sim \pi_i(o_i^{t},\rho_i^{t},\hrho^{o,t}_i)$ for each agent $i$\\
Each agent receives $o_i^{t+1}, r_i^{t}$ from environment\\
$\tau_i^{t} \gets \tau_i^{t-1} + (o_i^{t},a_i^{t})$\\
$D_i \gets D_i + (o_i^{t},\tau_i^{t-1},a_i^{t},r_i^{t},o_i^{t+1})$\\
}
$T_{update}=T_{update}+1$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B$)}
\label{update}
Select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
$\Lcal_{MI} \gets 0,\Lcal_{D} \gets 0, \Lcal_{Opp} \gets 0$\\
\For{$j = 1 \cdots \text{episode length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S_k \gets (\obar^{t_k+j,k}, \bar{\tau}^{t_k-1+j,k}, \abar^{t_k+j,k}, \rbar^{t_k+j,k}, \obar^{t_k+1+j,k})$\\
$S \gets S + S_k$\\
$\rho^{j,k},\hat{\rho}^{j,k},\Lcal_{MI}^{j,k},\Lcal_{D}^{j,k}, \Lcal_{Opp}^{j,k} \gets Role(S_k)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^{j,k}$\\
$\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^{j,k}$\\
$\Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^{j,k}$\\
}
Update critics, policies and targets using $S, \rho^{j,\cdot}, \hat{\rho}^{j,\cdot}$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho^{t}, \hrho^{t}$ from eq. (\ref{eq:roles}) \\
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using (\ref{eq:miloss}), (\ref{eq:dloss}), and (\ref{eq:opp_loss})\\
return $\rho^t,\hat{\rho}^{t},\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
We introduce \textsc{RAC}, the proposed a role-based actor-critic algorithm, designed for opponent-aware learning in the team-competitive games. In this setting, the agents in the team receive individual as well as shared rewards while the agents in different teams do not receive any shared reward. In this section, we describe the algorithm for simplicity the setting with two teams, each with any number of cooperating agents, which compete with each other according to the rules of individual games (described in section \ref{sec:experiments}). An interesting question in this setting is to determine whether learning opponent-aware role-based policies is beneficial for the teams.
\textsc{RAC}\ is built on the MAAC\ framework described above, and uses both agent-specific policy network $\pi_i$ and critic network $Q_i$. Additionally \textsc{RAC}\ employs two other networks:
\begin{enumerate}
\item $h_S$: self role encoding network
\item $h_O$: opponent role prediction network
\end{enumerate}
Another recurrent neural network (GRU) is used to maintain a running encoding of the trajectory $\tau_i^t$ at time $t$. For simplicity, we overload the notation $\tau_i^t$ to denote the embedding of the trajectory till time $t$, for the rest of the paper. Intuitively, the policy of an agent $i$ depends on its observation $o_i$, latent role $\rho_i$, and predicted roles of opponent agents $\hrho_i^{o} = [\hrho_j]_{j \in Opp(i)}$, where $Opp(i)$ is the set of opponent agents of agent $i$. Hence, $\pi_i(o_i,\rho_i,\hrho_i^o)$ computes the probability of action $a_i$. The self role encoding network $h_S$ and opponent role prediction network $h_O$ take the current observation $o_i^t$ and past action $a_i^{t-1}$ for $h_S$, or trajectory encoding $\tau_i^{t-1}$ for $h_O$ as input; and output a Gaussian distribution over the respective roles:
\begin{align}
&( \mu_{\rho_i^t}, \sigma_{\rho_i^t})= h_S( o_i^t, a_i^{t-1} )\ ;\ \ \PP(\rho_i^t | o_i^t, a_i^{t-1}) = \Ncal ( \mu_{\rho_i^t}, \sigma_{\rho_i^t}) \\
&\hat{\mu}_{\rho_j^t},\hat{\sigma}_{\rho_j^t} = h_O( o_{i}^t, \tau_{i}^{t-1} ) \forall j\neq i\ ;\ \ \hat{\PP}( \hat{\rho}_j^t | o_{i}^t, \tau_{i}^{t-1}) = \Ncal ( \hat{\mu}_{\rho_i^t}, \hat{\sigma}_{\rho_i^t})
\label{eq:roles}
\end{align}
Finally, the critic network $Q_i$ for agent $i$ takes as input the joint observation $\bar{o}$, joint action $\bar{a}$, and the role embeddings of current agent $\rho_i$ and role predictions of its opponents $\hrho_i^o$, $Q_i(\bar{o},\rho_i,\hrho_i^o,\bar{a})$, and outputs an estimate of the value-to-go. Figure \ref{fig:diagram} shows the overall architecture of the \textsc{RAC}\ system.
In the network architecture described above, the per-agent policies $\pi_i$ are trained using standard policy gradient steps, whereas the parameters of critic network $Q_i$, self-role encoder network $h_S$, and opponent-role predictor network $h_O$ are updated by centralized training of critics.
To update policy network of agent $i$, the policy gradient in~(\ref{eq:pol_grad_maac}) is further modified as follows:
\begin{multline}
\nabla_{\pi_i} \Jcal_i(\pi_{i}) = E_{\obar \sim D,\abar \sim \pi,\rho \sim h_S, \hrho \sim h_O} \left[ \nabla_{\pi_i} \log (\pi_{i} (a_{i}|o_{i},\rho_i,\hrho_i^o) \right.\\
\left. ( Q_i(\bar{o},\rho_i, \hrho_i^o, \bar{a})
- b(\obar,\rho_i, \hrho_i^o, \abar_{-i})
- \alpha \log (\pi_i (a_i | \obar_i, \rho_i, \hrho_i^o))) \right]
\end{multline}
Correspondingly, baseline computation is also modified as:
\begin{multline}
b(\obar,\rho_i, \hrho_i^o, \abar_{-i})
= \EE_{a_i \sim \pi_i(o_i, \rho_i, \hrho_i^o)} \left[ Q_i(\obar,\rho_i, \hrho_i^o,(a_i,\abar_{-i}))\right] \\
= \sum_{a_i \in A_i} \pi_i(a_i | o_i, \rho_i, \hrho_i^o ) Q_i(\obar,\rho_i, \hrho_i^o,(a_i,\abar_{-i}))
\end{multline}
While parameters of $\pi_i$ are trained using policy-gradient updates, parameters of the other networks are learned jointly by minimizing an aggregate loss function described next.
Firstly, the critic loss $\Lcal_Q$ defined in~(\ref{eq:critic_loss_maac}) is modified as follows:
\begin{multline}
\Lcal_{Q}=\sum_{i=1}^N \EE_{(\obar,\abar,\rbar,\obar^{\prime}) \sim D,\rho_i \sim h_S, \hrho_i^{o} \sim h_O } \left[ (Q_i(\obar ,\rho_i,\hrho_i^o, \abar ) - y_i )^2 \right] \text{ , where} \\
y_i = r_i + \gamma \EE_{\abar^{\prime} \sim \bar{ \pi }, \rho_i^{'} \sim h_S, \hrho_i^{o,'} \sim h_O }
\left[ \bar{Q}_i(\obar^{'}, \rho_i^{'}, \hrho_i^{o,'},\abar^{'}) \right. \\
\left. - \alpha \log ( \bar{\pi}_i (a_i^{'} | o_i^{'}, \rho_i^{'}, \hrho_i^{o,'}) ) \right]
\label{eq:critic_loss_role}
\end{multline}
\begin{figure*}[!t]
\centering
\subfloat[Touch-Mark]{\includegraphics[ width=.35\textwidth]{Results/snap_catch_goal_moving.pdf}}\hspace{3mm}
\subfloat[Market]{\includegraphics[width=.35\textwidth]{Results/snap_market_moving.pdf}}
\caption{Touch-Mark~(left) is a two-dimensional board game where members of each team try to reach one of the landmarks before any agent from the opponent team reaches any landmark. In Market~(right) teams pick resources scattered on the board and drop those to respective consumers before the opponent does the same.}
\label{fig:evironments}
\end{figure*}
Now, we describe the losses employed for training the self and opponent role encoder modules. For maintaining identifiability of roles with trajectories, we like to minimize conditional entropy $H(\rho_i^t \vert o_i^t, \tau_i^{t-1} )$, which is often intractable. Therefore, we maintain a variational approximation of distribution over roles, conditioned upon trajectories: $q_{\xi}(\rho_i^t | o_i^t, a_i^{t-1}, \tau_i^{t-1})$ and maximize the mutual information between $q_{\xi}(\rho_i^t | o_i^t, a_i^{t-1}, \tau_i^{t-1})$ and $\PP(\rho_i^t |o_i^t, a_i^{t-1})$. Simultaneously we minimize the entropy of role distribution $\PP(\rho_i^t |o_i^t, a_i^{t-1})$ to encode the history in the role encoder network.
Hence, corresponding mutual information loss, $\Lcal_{MI}$, is defined as:
\begin{multline} \label{eq:miloss}
\mathcal{L}_{MI} =
\sum_{i=1}^N \EE_{\tau_i^{t-1},o_i^t \sim D} \left[ D_{KL}(\PP(\rho_i^t|o_i^t,a_i^{t-1})|| \right. \\
\left. q_{\xi}(\rho_i^t|o_i^t,a_i^{t-1},\tau_i^{t-1})) \right]
+ H ( \PP(\rho_i^t|o_i^t,a_i^{t-1}) )
\end{multline}
Here, $D_{KL}$ denotes KL-divergence,
and $H(\cdot)$ denotes the entropy function.
In order to promote diversity between the behaviors of agents within a team, we introduce a new diversity loss $\mathcal{L}_D$, defined as:
\begin{equation}\label{eq:dloss}
\Lcal_D = \sum_{k \in \Kcal } \sum_{(i,j) \in \Ncal(k)}
\EE_{(\bar{\tau}^{t-1},\obar^t) \sim D}
\left[ q_{\xi}(\rho_i^t|\tau_j^{t-1},o_j^t,a_i^{t-1}) \right]
\end{equation}
Here, $\Kcal$ is the set of teams and $\Ncal(k)$ are agents in team $k$. The loss penalizes the probability of $\rho_i$ being identified with trajectory $\tau_j$ of another agent in the same team.
Finally, to train the opponent role prediction network $h_O$, we define the opponent loss $\Lcal_{Opp}$ as:
\begin{equation}\label{eq:opp_loss}
\Lcal_{Opp} = \sum_{\substack{i,j:team(i) \\ \neq team(j)}}
\EE_{(\bar{\tau}^{t-1},\obar^t) \sim D}
\left[ D_{KL}(\hat{\PP}(\rho_j^t|o_{i}^t,\tau_i^{t}) || \PP(\rho_j^t|o_j^t,a_j^{t-1})) \right]
\end{equation}
This loss penalizes the divergence of $\hat{\PP}(\rho_j^t|o_{i}^t,\tau_{i}^t)$, the predicted opponent role distribution for agent $j$ from the point of view of agent $i$, from $\PP(\rho_j^t|o_j^t,a_j^{t-1})$, the self-role distribution of agent $j$.
We define the composite loss function for training the role networks as:
\begin{equation}
\Lcal_{Role}=\Lcal_{D}+\Lcal_{MI}+\Lcal_{Opp}
\end{equation}
Hence, $\Lcal_Q + \Lcal_{Role}$ is the total loss used to learn the combined critic module of \textsc{RAC}\ model. However, it is observed that the evolution of the role networks in the later phase of training causes instability in learning policies, which result in reduced average reward. This phenomenon is mitigated by using an exponentially decaying weightage on role loss $\Lcal_{Role}$. The total critic loss becomes:
\begin{equation}
\Lcal_{tot}=\Lcal_Q + \lambda^{\frac{u}{C}} \times \Lcal_{Role}
\end{equation}
where $\lambda\in [0,1]$ is the decay factor, $ u $ is the training episode count, and $ C $ is a constant affecting the rate of decay over episodes. Algorithm~\ref{modifiedAlgo} includes a pseudo-code for our algorithm. Table ~\ref{tab:notations} contains list of all variables.
\iffalse
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize $E$ parallel environments with $N$ agents\\
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$i_{ep} = 1 \cdots $ max episodes}{
$\tau_i^{0,e} \gets 0$\\
$H_i^{0,e} \gets \emptyset, \forall i $\\
Reset all environments, and get initial $o_i^{1,e}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
In each environment $e$, sample roles for each agent $\rho_i^{t,e}, [\hat{\rho}^{t,e}_j]$ using eq.~(\ref{eq:roles})\\
In each environment $e$, sample action from respective policies $a_i^{t,e} \sim \pi_i(o_i^{t,e},\rho_i^{t,e},[\hat{\rho}^{t,e}_j])$ \\
Each agent receives $o_i^{t+1,e}, r_i^{t,e}$ from respective environments\\
$\tau_i^{t,e} \gets \tau_i^{t-1,e} + (o_i^{t,e},a_i^{t,e})$\\
$H_i^{e} \gets H_i^{e} + (o_i^{t,e},\tau_i^{t-1,e},a_i^{t,e},r_i^{t,e},o_i^{t+1,e})$\\
}
$T_{update}=T_{update}+E$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$,$\text{window length}$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B,\text{window length}$)}
\label{update}
Randomly select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
\For{$j = 1 \cdots \text{window length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S \gets S + (o_{1 \cdots N}^{t_k+j,k}, \tau_{1 \cdots N}^{t_k-1+j,k}, a_{1 \cdots N}^{t_k+j,k}, r_{1 \cdots N}^{t_k+j,k}, o_{1 \cdots N}^{t_k+1+j,k})$
}
$\rho^j,\hat{\rho}^j,\Lcal_{MI}^j,\Lcal_{D}^j, \Lcal_{Opp}^j \gets Role(S)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^j,\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^j, \Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^j$\\
Update critics, policies and targets using $S, \rho^j, \hat{\rho}^j$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho_i^{t,e}, [\hat{\rho}_j^{t,e}]$ from eq. \ref{eq:roles} using for each agent $i$, environment $e$ from $S$\\
$\rho^t \gets \rho_1^{t,\cdot} \cdots \rho_N^{t,\cdot},\hat{\rho}^{t} \gets \hat{\rho}_1^{t,\cdot} \cdots \hat{\rho}_N^{t,\cdot}$\\
\For{$team = 1 \cdots \text{number of teams}$ }{
\For{$agent=1 \cdots \text{number of agents in a team} $}{
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using eq. ~\ref{eq:miloss}, eq.~\ref{eq:dloss}, and eq.~\ref{eq:opp_loss}
}
}
return $\rho^t,\hat{\rho}^t,\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
\fi
\section{Experimental Results}
\label{sec:experiments}
\begin{figure*}[!t]
\centering
\subfloat{\includegraphics[width=.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Rewards.pdf}}
\subfloat{\includegraphics[width=.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Landmark.pdf}}
\subfloat{\includegraphics[width=.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Rewards.pdf}}
\subfloat{\includegraphics[width=.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Dropped.pdf}}
\caption{Performance comparison. We play \textsc{RAC}\ agents against MAAC\ agents, who are separately trained using self-play for $300000$ training steps. We play them against each other after each $1000$ step and report performance, averaging over $4$ pairs of tournament-play. For Touch-Mark, we report the average team reward and the fraction of times each team reaches the landmark. For Market, we report the average team reward and the fraction of times each team drops a resource to its respective consumer, after picking it from its source. }
\label{fig:tournament}
\end{figure*}
\begin{figure}[!t]
\centering
\subfloat{\includegraphics[width=0.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_vary_manual_decay_update_rsubseq_v3_coll_avg_0_Roles.pdf}\label{fig:role-catch-goal}
\subfloat{\includegraphics[width=0.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_coll_avg_0_Roles.pdf}\label{fig:role-market}}
\caption{Role emergence. For Touch-Mark, we observe, \textsc{RAC}\ agents result in more collisions than MAAC\ agents. In Market, we observe that \textsc{RAC}\ agents choose to pick more diverse items than MAAC\ agents. }
\label{fig:role}
\end{figure}
\begin{figure}[!t]
\centering
\subfloat{\includegraphics[width=0.23\textwidth]{Results/CatchGoalFixed_cooperative_avg_Landmark.pdf}\label{fig:cooperative-catch-goal}
\subfloat{\includegraphics[width=0.23\textwidth]{Results/Market3_cooperative_avg_Dropped.pdf}\label{fig:cooperative-market-2}}
\caption{We compare \textsc{RAC}\ with RAC$_{Team}$, an extension of \textsc{ROMA}\ with opponent modeling removed. However, we observe that performance-wise \textsc{RAC}\ is still outperforming RAC$_{Team}$ . }
\label{fig:cooperative}
\end{figure}
\begin{figure*}[!t]
\centering
\subfloat[]{\includegraphics[width=.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_opponent_update_rsubseq_v3_vary_decay_avg_Landmark.pdf}\label{fig:vary_decay_catch_goal}}
\subfloat[]{\includegraphics[width=.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_vary_decay_avg_Dropped.pdf}\label{fig:vary_decay_market}}
\subfloat[]{\includegraphics[width=0.23\textwidth]{Results/CatchGoalFixed_ablations_avg_Landmark.pdf}\label{fig:ablations_catch_goal}}
\subfloat[]{\includegraphics[width=0.23\textwidth]{Results/Market3_ablations_avg_Dropped.pdf}\label{fig:ablations_market}}
\caption{(Left) Hyperparameter tuning. Here the decay parameter $\lambda$ is varied in $[0,1]$ and the performance of \textsc{RAC}\ teams, played against a random set of MAAC\ models, is reported. (Right) Ablations. Performance of various ablations of \textsc{RAC}\ are reported, which show that the three distinct objectives are necessary for best performance in both the environments.}
\end{figure*}
In this section, we investigate two primary objectives. First, we check how employing the role network influences the performance of our algorithm in direct competition with baseline algorithm. Second, we check the effectiveness of the role network in emerging roles dynamically within the team.
Towards this end, we employ two games, namely Touch-Mark\ and Market, designed in MPE~\citep{lowe2017multi,mordatch2018emergence}.
\noindent \textbf{Touch-Mark~\citep{koley2022offsetting}{:}}
It is an episodic board game that consists of two teams, each comprising of multiple agents (2 in our case) and two static landmarks, all randomly positioned at the beginning of each episode and each team tries to reach at least one of those landmarks earlier than any member of the other team.
The agents start with velocity zero and slowly accelerate until they reach maximum permissible velocity.
The episode ends when an agent reaches a landmark. The winning agent's team (i.e. both the team members) receives a large reward $r_l$, incurring $-r_l$ penalty to the opposite team.
Additionally, each agent receives a small penalty,
proportional to its distance from
the nearest target at every time step to guide it towards the nearest target.
Moreover, an agent can collide with an agent of the opponent team to divert it from its path. The collision makes both agents temporarily stationary. This mechanism is introduced so that an agent
has the option to stop an agent of the opponent team from reaching a landmark, thus facilitating the fellow teammate to reach a landmark first.
\noindent \textbf{Market{:}}
This is a novel episodic board game with two teams, each consisting of multiple players ($2$ in our case).
At the beginning of each episode, four resources, two for each type of resource are randomly placed on the board, along with one consumer for each type of resource.
Agents need to acquire the resources and drop those to the respective consumers.
After acquiring a resource, the agent continues to receive a penalty proportional to its distance from the appropriate consumer, to guide him towards the corresponding consumer. Both acquiring and dropping resources result in high rewards for corresponding team.
The termination condition for the episode is timeout of 50 steps.
The consumer remains alive only until one resource of its type is dropped to it. Therefore for highest scores, one team must acquire both types of resources and drop them to respective consumers before their opponents.
\noindent \textbf{Experimental setup}
We use a one-layer feed-forward neural network followed by GRU unit for getting initial embedding of the trajectories. For each team, a double-layer MLP is employed for each of $\PP(\rho_i^t | o_i^t, a_i^{t-1})$,$\hat{\PP}(\rho_j^t | o_i^t, a_i^{t-1})$ and $q_{\xi}(\rho_i^t | \tau_i^{t-1},o_i^t)$. Critic and policy network architecture follows MAAC~\citep{iqbal2019actor} except inputs of both critic and policy modules for each agent $i$ are modified to include a role tuple for that agent, containing self-role $\rho_i^t$ and predicted opponent-roles $\hrho_i^{o,t}$.
\noindent \textbf{Baselines}
Our primary baseline is MAAC~\citep{iqbal2019actor}.
We also compare our method with RAC$_{Team}$\ which follows \textsc{RAC}\ except does not share information across teams during the training phase.
This mimics the setting where competing teams separately employ a cooperative role-oriented algorithm like \textsc{ROMA}\ and only sacrifices the advantage of training in presence of opponents.
\noindent \textbf{Metrics}
We report cumulative team rewards per episode to measure the performance of the algorithms. As a qualitative measure, we report the rate of touching landmarks per episode for each team in Touch-Mark, and the average number of dropped resources per episode for each team in Market{.} The average resources dropped per episode can vary between $0$ and $2$ as there are $2$ resources in Market. All measures are reported after taking an average across all seeds used.
\subsection{Results}
\noindent \textbf{Direct tournament:} Here, we compare \textsc{RAC}\ with MAAC\ by directly playing them against each other. Here, we use the \textsc{RAC}\ and MAAC\ teams, trained in self-play for $3 \times 10^5$ training episodes starting with random seed. At fixed interval during the training we play them among themselves for $1000$ episodes. This experiment is repeated for $4$ pairs of \textsc{RAC}\ and MAAC\ teams, and the average metrics are reported in fig. \ref{fig:tournament} for Touch-Mark~(left) and Market~(right){.}
We observe that, as the training progresses, \textsc{RAC}\ continues to outperform MAAC. The performance gap is quite significant for Market, whereas, in Touch-Mark\ the gain is marginal.
We explain it as follows. Market\ deals with a comparatively complex setting, making it difficult for the agents to play strategically without the explicit assistance for sub-task specialization. The role-driven policy learning framework of \textsc{RAC}\ seems to be effective here to reduce the observation-action space to explore and quickly guide them towards improved policy learning by devising opponent-aware strategic behaviors.
Touch-Mark\ deals with a simpler setting, where agents learn effective strategies comparatively faster even without external asistance on implicit role division or sub-task specialization, resulting in only marginal improvement of \textsc{RAC}\ over MAAC{.}
Further, we observe a performance drop of \textsc{RAC}\ agents as the training approaches for Touch-Mark{.} We explain this phenomenon as follows. \textsc{RAC}\ accelerates learning strategic behavior in Touch-Mark, more specifically, splits the responsibility of collision and chasing the landmark among team members. However, as MAAC\ agents quickly master the strategy in the next phase of training, the advantage of \textsc{RAC}\ almost disappear. Therefore, with the improvement of the expertise of MAAC\ team, the performance of \textsc{RAC}\ team drops, finally resulting in marginal gain over MAAC\ agents.
\noindent \textbf{Role learning:}
Further, we investigate whether \textsc{RAC}\ is actually helping teammates to choose diverse roles. Also, we study the strategic behaviors of \textsc{RAC}\ agents.
In Touch-Mark, as we have already mentioned, though collision incurs a small penalty, one can strategically employ collision with an opponent to delay the opponent agent from reaching the landmark.
Similarly, in Market, if the teammates choose to go for different resources, they can acquire higher rewards in future.
To check whether algorithms are able to train their agents such strategic moves, we summarize role-discrimination statistics in fig \ref{fig:role} for both the games.
For Touch-Mark\ we report the fraction of winning episodes where the agent other than the one who touches the landmark, collides with the opponents. Similarly, for Market, we report the fraction of episodes where members within the team drop different types of resources.
In both cases, we find that \textsc{RAC}\ agents tend to employ strategic moves more than MAAC\ agents, which can possibly explain better performance of \textsc{RAC}\ in direct competitions.
The key factor behind such diverse behavior is employment of $\Lcal_D$ and $\Lcal_{Opp}$, which incentivizes the role network to assign diverse as well as opponent-aware role distributions to teammates respectively.
\noindent \textbf{Comparison with RAC$_{Team}$\ extending~\citep{wang2020roma}:} Fig \ref{fig:cooperative} compares \textsc{RAC}\ with RAC$_{Team}$, which is a trivial extension of \textsc{ROMA}~\citep{wang2020roma} to actor-critic framework, where we use the two regularizers used in ~\citep{wang2020roma} for role learning, but remove the opponent modeling. Instead, agents are trained using the simplified-role-module enriched MAAC\ for both of the teams separately, whereas a team it enjoys shared reward setting. Finally, agents trained under \textsc{RAC}\ and RAC$_{Team}$\ are separately played against random MAAC\ teams selected from a common set of MAAC\ teams and average comparative results are plotted
in fig \ref{fig:cooperative}.
The superiority of \textsc{RAC}\ validates that our opponent-aware role-oriented framework provides an additional layer of guidance to agents in competitive team games over a straight-forward extension of \textsc{ROMA}~\citep{wang2020roma} to actor-critics, thus emphasizing the motivation of our approach.
\noindent \textbf{Hyperparameter ($\lambda$) tuning:} Fig~[\ref{fig:vary_decay_catch_goal}-\ref{fig:vary_decay_market}] presents performance of \textsc{RAC}\ teams against randomly chosen MAAC\ teams for different value of decay parameter $\lambda$.
Both in Touch-Mark\ and Market, performance significantly varies based on chosen $\lambda$, with $\lambda=0.5$ giving the best results. $\lambda=0.9$ results in a slower decay of role loss, hence resulting in greater instability in policy learning, while $\lambda=0.1$ results in too fast a decay. This experiment confirms our intuitive hypothesis from real-world experience, that sustained role evolution beyond a point is detrimental to team performance.
\noindent \textbf{Comparison with ablations:}
Fig. [\ref{fig:ablations_catch_goal}-\ref{fig:ablations_market}] compare \textsc{RAC}\ with three ablations of \textsc{RAC}\, which minimize either $\Lcal_D$ or $\Lcal_{MI}$ or $\Lcal_D+\Lcal_{MI}$, whereas \textsc{RAC}\ minimizes $\Lcal_D + \Lcal_{MI}+\Lcal_{Opp}$. In both games, \textsc{RAC}\ is observed to outperform the rest.
In Market\, we also observe $\Lcal_{MI}$ learns roles more identifiable with trajectory and therefore training better dropping skill, in comparison with the other two baselines.
\if{0}
For Touch-Mark, now we take a deeper look by observing the event of touching landmark for each agent individually in Fig. \ref{fig:tournament-landmark-agent-wise}. It is found that in case of \textsc{RAC}\ one agent learns faster than the rest, resulting in initial advantages of team \textsc{RAC}\ over team MAAC. But the other \textsc{RAC}\ agent barely shows any improvement in touching the landmark. MAAC\ agents learn slower, but, the variation in fellow teammates' learning rate is smaller. MAAC\ slowly trains both of the agents, simultaneously, eventually resulting in cumulative superior performance.
This observation indicates employing role network fastens the learning process and useful in shorter term.
However, if both skills are not explicitly maximizing rewards, we may find that eventually this role distinction within teammates may result in higher variance in the team members' skill-specific expertise, resulting in slowing down the overall team performance, as in Touch-Mark.
On the other hand, this did not happen in Market. As a reason, we can observe that, in Market, agents within a team try to pick different resources as a result of role emergence. As both of these actions implicitly increase the rewards, learning specific roles does not hamper retaining the higher rewards throughout, unlike Touch-Mark.
\fi
\section{Conclusion and Future work}
\vspace{2mm}
We propose an algorithm for opponent-aware role-based learning in actor-critic framework targeted towards team competitions. Our algorithm combines a self-role encoder and an opponent-role predictor in actor-critic framework for learning an optimal policy. We analyze our approach in two scenarios where we show how our method improves the quality of learning. As a future direction, we intend to extend our algorithm to more general settings where multiple teams can compete among themselves.
\iffalse
\PK{I am keeping algorithm here temporarily for some modification. I am not editing the method section. }
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$i = 1 \cdots $ max episodes}{
$\tau_i^{0} \gets 0$\\
$H_i^{0} \gets \emptyset, \forall i $\\
Reset all environments, and get initial $o_i^{1}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
Sample roles for each agent $\rho_i^{t}, \hrho^{o,t}_i$ using~(\ref{eq:roles})\\
Sample action from respective policies $a_i^{t} \sim \pi_i(o_i^{t},\rho_i^{t},\hrho^{o,t}_i)$ \\
Each agent receives $o_i^{t+1}, r_i^{t}$ from environment\\
$\tau_i^{t} \gets \tau_i^{t-1} + (o_i^{t},a_i^{t})$\\
$H_i \gets H_i + (o_i^{t},\tau_i^{t-1},a_i^{t},r_i^{t},o_i^{t+1})$\\
}
$T_{update}=T_{update}+1$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$,$\text{window length}$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B,\text{window length}$)}
\label{update}
Randomly select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
$\Lcal_{MI} \gets 0,\Lcal_{D} \gets 0, \Lcal_{Opp} \gets 0$\\
\For{$j = 1 \cdots \text{window length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S \gets S + (o_{1 \cdots N}^{t_k+j,k}, \tau_{1 \cdots N}^{t_k-1+j,k}, a_{1 \cdots N}^{t_k+j,k}, r_{1 \cdots N}^{t_k+j,k}, o_{1 \cdots N}^{t_k+1+j,k})$
}
$\rho^j,\hat{\rho}^j,\Lcal_{MI}^j,\Lcal_{D}^j, \Lcal_{Opp}^j \gets Role(S)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^j,\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^j, \Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^j$\\
Update critics, policies and targets using $S, \rho^j, \hat{\rho}^j$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho_i^{t}, \hrho_i^{o,t}$ from eq. (\ref{eq:roles}) using for each agent $i$\\
$\rho^t \gets \rho_1^{t} \cdots \rho_N^{t},\hat{\rho}^{t} \gets \hat{\rho}_1^{t} \cdots \hat{\rho}_N^{t}$\\
\For{$agent=1 \cdots N$}{
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using (\ref{eq:miloss}), (\ref{eq:dloss}), and (\ref{eq:opp_loss})
}
return $\rho^t,\hat{\rho}^t,\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
\fi
\section{Pseudo Code of our algorithm}
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize $E$ parallel environments with $N$ agents\\
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$i_{ep} = 1 \cdots $ max episodes}{
$\tau_i^{0,e} \gets 0$\\
$H_i^{0,e} \gets \emptyset, \forall i $\\
Reset all environments, and get initial $o_i^{1,e}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
In each environment $e$, sample roles for each agent $\rho_i^{t,e}, [\hat{\rho}^{t,e}_j]$ using eq.~(\ref{eq:roles})\\
In each environment $e$, sample action from respective policies $a_i^{t,e} \sim \pi_i(o_i^{t,e},\rho_i^{t,e},[\hat{\rho}^{t,e}_j])$ \\
Each agent receives $o_i^{t+1,e}, r_i^{t,e}$ from respective environments\\
$\tau_i^{t,e} \gets \tau_i^{t-1,e} + (o_i^{t,e},a_i^{t,e})$\\
$H_i^{e} \gets H_i^{e} + (o_i^{t,e},\tau_i^{t-1,e},a_i^{t,e},r_i^{t,e},o_i^{t+1,e})$\\
}
$T_{update}=T_{update}+E$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$,$\text{window length}$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B,\text{window length}$)}
\label{update}
Randomly select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
\For{$j = 1 \cdots \text{window length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S \gets S + (o_{1 \cdots N}^{t_k+j,k}, \tau_{1 \cdots N}^{t_k-1+j,k}, a_{1 \cdots N}^{t_k+j,k}, r_{1 \cdots N}^{t_k+j,k}, o_{1 \cdots N}^{t_k+1+j,k})$
}
$\rho^j,\hat{\rho}^j,\Lcal_{MI}^j,\Lcal_{D}^j, \Lcal_{Opp}^j \gets Role(S)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^j,\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^j, \Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^j$\\
Update critics, policies and targets using $S, \rho^j, \hat{\rho}^j$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho_i^{t,e}, [\hat{\rho}_j^{t,e}]$ from eq. \ref{eq:roles} using for each agent $i$, environment $e$ from $S$\\
$\rho^t \gets \rho_1^{t,\cdot} \cdots \rho_N^{t,\cdot},\hat{\rho}^{t} \gets \hat{\rho}_1^{t,\cdot} \cdots \hat{\rho}_N^{t,\cdot}$\\
\For{$team = 1 \cdots \text{number of teams}$ }{
\For{$agent=1 \cdots \text{number of agents in a team} $}{
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using eq. ~\ref{eq:miloss}, eq.~\ref{eq:dloss}, and eq.~\ref{eq:opp_loss}
}
}
return $\rho^t,\hat{\rho}^t,\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
\section{Opponent-aware Role-based MARL}
\label{sec:problem_formulation}
In this section, we develop the proposed algorithm \textsc{RAC}\ for Multi-agent Policy learning with emergent roles in the team-competitive Markov games setting \cite{lowe2017multi}.
\input{100prelim}
\input{101maac}
\input{102roma}
\input{103rac}
\input{200exp}
\input{300conclusion}
\bibliographystyle{ACM-Reference-Format}
\subsection{Role-aware policy learning}
Roles have been used in many MARL applications for promoting cooperation and coordination among agents \cite{barrett2015cooperating,roy2020promoting,urieli2011optimizing,zhang2021hierarchical}.
Recently, Wang et al. proposed \textsc{ROMA}{,}\cite{wang2020roma}, a MARL framework for learning emergent roles in a cooperative game setting.
In \textsc{ROMA}, the roles are designed to exhibit two crucial properties: (1) \textbf{Identifiability: } role should be identifiable with its behavior and diverse enough to fulfill the task, and (2) \textbf{Specialization: } agents with similar roles should specialize in similar responsibilities. For identifiability, it maximizes the mutual information $I( \rho_i ; \tau_i )$ between the role distribution $p(\rho_i|o_i)$, and the trajectory-conditional role distribution
$p(\rho_i| \tau_i,o_i)$
For ensuring specialization and diversity, it maintains a dissimilarity model $d_{\phi}$ to measure the distance of trajectories of two agents and seek to maximize $I( \rho_i ; \tau_j ) + d_{\phi} (\tau_i, \tau_j) $. With these regularizers, \textsc{ROMA}\ ensures agents adopt role-aware policies within a team in each trajectory.
\textsc{ROMA}\ builds on the QMIX~\citep{rashid2018qmix}, which is designed for settings where the agents receive a shared reward for joint observation and actions. \textsc{ROMA}\ also works within the Q-learning mode without any additional policy network. It learns per-agent local utility networks, $\Qcal_i$ output of which are passed to a mixing network to compute global TD-loss.
During execution, each agent acts according to optimum policy derived from its' local utility function.
In \textsc{ROMA}\, the local utility function $\Qcal_i(o_i,a_i,\rho_i)$ takes the agent's role $\rho_i$ also as input. The role is $\rho_i$ defined as a latent (emergent) variable with a distribution, conditioned on
the trajectory of the agent $\tau_i$.
However, the Q-learning framework used in \textsc{ROMA}\ cannot be trivially extended to the competitive setting since the composition of local utility functions $\Qcal_i$ into one global utility function $\Qcal$ strongly relies upon the assumption of the common reward of all agents which does not hold in a competitive setting with mixed-rewards.
Moreover, the local utility function $\Qcal_i$ takes only observation and action of $i$-th agent ($o_i$ and $a_i$) as input, and therefore agents cannot learn from the experience of opponents (other agents with decoupled rewards).
In the next section, we develop a novel formulation for learning emergent roles in competitive setting in the actor-critic framework, which is better suited towards incorporating mixed-reward.
In the actor-critic framework, the critic function $Q_i(\bar{o},\bar{a})$ takes observations and actions of all agents as input, hence facilitating the sharing of information among agents with decoupled reward and agents benefits from learning opponent-aware policies by exploiting the shared opponent information available during the training.
\section{Pseudo Code of our algorithm}
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize $E$ parallel environments with $N$ agents\\
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$i_{ep} = 1 \cdots $ max episodes}{
$\tau_i^{0,e} \gets 0$\\
$H_i^{0,e} \gets \emptyset, \forall i $\\
Reset all environments, and get initial $o_i^{1,e}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
In each environment $e$, sample roles for each agent $\rho_i^{t,e}, [\hat{\rho}^{t,e}_j]$ using eq.~(\ref{eq:roles})\\
In each environment $e$, sample action from respective policies $a_i^{t,e} \sim \pi_i(o_i^{t,e},\rho_i^{t,e},[\hat{\rho}^{t,e}_j])$ \\
Each agent receives $o_i^{t+1,e}, r_i^{t,e}$ from respective environments\\
$\tau_i^{t,e} \gets \tau_i^{t-1,e} + (o_i^{t,e},a_i^{t,e})$\\
$H_i^{e} \gets H_i^{e} + (o_i^{t,e},\tau_i^{t-1,e},a_i^{t,e},r_i^{t,e},o_i^{t+1,e})$\\
}
$T_{update}=T_{update}+E$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$,$\text{window length}$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B,\text{window length}$)}
\label{update}
Randomly select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
\For{$j = 1 \cdots \text{window length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S \gets S + (o_{1 \cdots N}^{t_k+j,k}, \tau_{1 \cdots N}^{t_k-1+j,k}, a_{1 \cdots N}^{t_k+j,k}, r_{1 \cdots N}^{t_k+j,k}, o_{1 \cdots N}^{t_k+1+j,k})$
}
$\rho^j,\hat{\rho}^j,\Lcal_{MI}^j,\Lcal_{D}^j, \Lcal_{Opp}^j \gets Role(S)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^j,\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^j, \Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^j$\\
Update critics, policies and targets using $S, \rho^j, \hat{\rho}^j$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho_i^{t,e}, [\hat{\rho}_j^{t,e}]$ from eq. \ref{eq:roles} using for each agent $i$, environment $e$ from $S$\\
$\rho^t \gets \rho_1^{t,\cdot} \cdots \rho_N^{t,\cdot},\hat{\rho}^{t} \gets \hat{\rho}_1^{t,\cdot} \cdots \hat{\rho}_N^{t,\cdot}$\\
\For{$team = 1 \cdots \text{number of teams}$ }{
\For{$agent=1 \cdots \text{number of agents in a team} $}{
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using eq. ~\ref{eq:miloss}, eq.~\ref{eq:dloss}, and eq.~\ref{eq:opp_loss}
}
}
return $\rho^t,\hat{\rho}^t,\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
\subsection{Preliminaries}
\begin{table}[t]
\begin{tabular}{c l}
\toprule
Symbol & Meaning \\\midrule
$\Scal,\Acal$ & State and action Space \\
$\Pcal,\Rcal$ & Transition and reward functions \\
$N, K$ & Total Number of agents and teams \\
$\Ncal(k)$ & Set of agents in team $k$\\
$\obar=(o_1,\dots,o_N)\in \Ocal $ & observation vector\\
$\abar=(a_1,\dots,a_N)\in \Acal $ & action vector\\
$\pi=(\pi_1,\dots,\pi_N)$ & Agent policies \\
$\tau^t=(\tau_1^t,\dots,\tau_N^t)$ & Trajectory till time $t$\\
$Q=(Q_1,\dots,Q_N)$ & Value-action function \\
$h_{S}=(h_S^1,\dots,h_S^K)$ & Self-role encoder \\
$h_{O}=(h_O^1,\dots,h_O^K)$ & Opponent-role encoder \\
$\rho=(\rho_1,\dots,\rho_N)$ & Self-role \\
$\hrho=(\hrho_1^o,\dots,\hrho_N^o)$ & Predicted opponent-role \\
$\lambda$ & Decay factor \\
$\Lcal_{Q}$ & Critic loss \\
$\Lcal_{MI}$ & Mutual information loss\\
$\Lcal_{D}$ & Divergence loss \\
$\Lcal_{Opp}$ & Opponent loss \\
\bottomrule
\end{tabular}
\caption{List of important notations.}
\label{tab:notations}
\end{table}
A Markov game is characterized by the tuple $\langle \Scal,\Acal,\Pcal,\Rcal \rangle$, where $\Scal$ denotes the set of states, $\Acal$ denotes the set of actions for each of the $N$ agents. Hence the joint space of actions becomes $\Acal^N$.
The state transition function maps every state and joint-action combination to a probability over future states. $\Pcal \colon \Scal \times \Acal^N \rightarrow \PP(\Scal)$.
In case of team-competitive games (also called mixed cooperative-competitive
games \cite{lowe2017multi}), let $\Kcal$ denote the set of $K$ teams, and $\Ncal(k)$ be the set of agents of team $k$,$k \in \Kcal$.
The reward function $\Rcal = (r_1,\dots,r_N) \colon \Scal \times \Acal^N \rightarrow \mathcal{R}^N $ specifies the reward scheme ($r_i$) for each agent $i$.
The per-agent reward function allows modeling of team competitive games~\citep{lowe2017multi}, which are the focus of this paper.
Also, in the decentralized execution setting, let $o_i \in \Ocal$ denote the observation for agent $i$.
Each agent $i$ learns a policy $\pi_i \colon \Ocal \rightarrow \PP( \Acal )$, which is a mapping from it's own observation to a distribution over it's action set $\Acal$.
We also define the trajectory $\tau_i(t) \in T = (\Ocal\times\Acal)^*$ for each agent $i$ till time $t$ as a sequence of observation-action pairs $\{ (o_i^j,a_i^j), j=1,\dots,t \}$. We also denote the joint action and observations vectors as: $\bar{a}=(a_1,\dots,a_N)$ and $\bar{o}=(o_1,\dots,o_N)$. Agents aim to learn their policy $\pi_i$ by optimizing the agent-specific cumulative discounted reward function $\Jcal_i$:
\begin{align}
\Jcal_i(\pi_{i}) &= E_{\abar \sim \pi, s \sim \Pcal} \left[ \sum_{u=0}^{\infty} \gamma^{u} r_{i} (s^u, \abar^u) \right]
\label{eq:J_pol_grad}
\end{align}
\iffalse
Corresponding policy gradient update for agent $i$ takes following form:
\begin{align}
\nabla_{\pi_i} \Jcal_i(\pi_{i})
&= \nabla_{\pi_i} \log (\pi_{i} (a_{i}^t | o_{i}^t) ) \left[ \sum_{\hat{t}=t}^{\infty} \gamma^{\hat{t}-t} r_i(s^{\hat{t}}, \bar{a}^{\hat{t}}) \right] \\
&= \nabla_{\pi_i} \log (\pi_{i} (a_{i}^t|o_{i}^t ) Q_i(\bar{o}^t,\bar{a}^t)
\end{align}
where $Q_i$ is the Value-action function of agent $i$
\fi
\sloppy
\iffalse
We follow a state-of-the-art multi-agent actor-critic method, MAAC~\citep{iqbal2019actor}, which supports centralized training of agents' policies with individual rewards. MAAC learns policy $\pi_i$ for agent $i$ using the policy gradient updates for the agent-specific discounted reward function $\Jcal_i$:
\begin{align}
\nabla_{\pi_i} \Jcal_i(\pi_{i}) &= \nabla_{\pi_i} \log (\pi_{i} (a_i | o_i) ) \sum_{\hat{t}=t}^{\infty} \gamma^{\hat{t}-t} r_{\hat{t}} (s_{\hat{t}}, a_{\hat{t}}) \\
&= \nabla_{\pi_i} \log (\pi_{i} (a_t|s_t) ) Q_i(\bar{o},\bar{a})
\end{align}
In MAAC, the critic function for agent $i$ depends on the embedding of agent $i$, $g_i(o_i,a_i)$, and the attention-weighted average reward of other agents $x_i$ as:
\begin{align}
& Q_i(\bar{o},\bar{a}) = f_i ( g_i(o_i,a_i), x_i)\\
& x_i = \sum_{j \neq i} \alpha_j \Vcal g_j( o_j,a_j)
\end{align}
where $f_i$ and $\Vcal$ are neural networks.
$\alpha_j$ is the dynamic-attention function given by agent $i$ to agent $j$, and depends on embeddings $g_i(o_i,a_i)$ and $g_j(o_j,a_j)$. The critic function is learned using a temporal difference loss calculated on trajectories collected using current and past policies, which are stored in a replay buffer.
\fi
\subsection{Multi-Agent Actor Critic}
Actor-critic methods and their extensions are widely used in multi-agent reinforcement learning \cite{haarnoja2018soft,iqbal2019actor} due to their flexibility and low variance reward estimates, which lead to faster learning of policy parameters. A popular extension is the soft actor-critic \cite{haarnoja2018soft}(SAC), where the reward of policy gradient (Eqn \ref{eq:J_pol_grad}) is augmented with the expected entropy of the policy distribution in each state. Hence their reward function becomes:
\begin{align}
\Jcal_i(\pi_{i}) &= E_{\abar \sim \pi, s \sim \Pcal} \left[ \sum_{u=0}^{\infty} \gamma^{u} (r_{i} (s^u, \abar^u) + \alpha \mathcal{H}(\pi_i(s^u)) \right]
\end{align}
where $\alpha$ is the user defined regularization coefficient. Note that they do not explicitly discuss multi-agent RL in team-competitive setting.
Subsequently, multi-attention actor critic (MAAC) \cite{iqbal2019actor} extended SAC to support centralized training of agents' policies with individual rewards. The learning algorithm in MAAC has two broad components: (1) SAC-style update of policy $\pi_i$ for the $i^{th}$ agent and (2) Attentive learning of the critic function $Q_i$ for the $i^{th}$ agent. The policy update is carried out using the policy gradient given by:
\begin{multline}
\nabla_{\pi_i} \Jcal_i(\pi_{i}) = E_{\obar \sim D,\abar \sim \pi} [\nabla_{\pi_i} \log (\pi_{i} (a_{i}|o_{i})
( Q_i(\bar{o},\bar{a}) \\
- b(\obar,\abar_{-i})
- \alpha \log (\pi_i (a_i | o_i ))) ]
\label{eq:pol_grad_maac}
\end{multline}
where $\alpha$ is the regularization constant for the entropy term. $D$ is the replay buffer that contains the tuple $(\obar, \abar, \rbar, \obar^{'})$ for each step of the episodes generated by the policies till the current RL training time $t$. Here action $\abar$ at observation $\obar$ generates reward of all agents $\rbar$ and next observation $\obar^{'}$. $b(o,a_{-i})$ is the multi-agent baseline calculated as follows.
\begin{align}
b(\obar,\abar_{-i}) &= E_{a_i \sim \pi_i(o_i) } \left[ Q_i(\obar,(a_i,\abar_{-i}))\right] \\
&= \sum_{a_i \in \mathcal{A}} \pi_i(a_i | o_i) Q_i(\obar,(a_i,\abar_{-i}))
\end{align}
The baseline is used calculate the advantage function, which helps in solving the multi-agent credit assignment problem \cite{iqbal2019actor}.
The critic function $Q_i$ for agent $i$ depends on the embedding $g_i(o_i,a_i)$ of agent $i$ and the attention-weighted average reward of other agents $x_i$ as:
\begin{align}
& Q_i(\bar{o},\bar{a}) = f_i ( g_i(o_i,a_i), x_i)\\
& x_i = \sum_{j \neq i} \alpha_{ij} \Vcal ( g_j( o_j,a_j) ) \\
&\alpha_{ij} \propto \exp( v_i^T W_i^T W_j v_j)
\end{align}
where $f_i$ and $\Vcal$ are feedforward neural networks.
$\alpha_{ij}$ is the dynamic-attention function given by agent $i$ to agent $j$, and depends on embeddings $v_i = g_i(o_i,a_i)$ and $v_j = g_j(o_j,a_j)$. $W_i$ and $W_j$ are agent specific ``key'' transformation matrices for the attention function. The critic function is learned using a temporal difference loss calculated on trajectories collected using current and past policies, which are stored in a replay buffer $D$. The loss function for updating the critic module for agent $i$ takes following form:
\begin{multline}
\Lcal_Q = \sum_{i=1}^{N} E_{(\obar,\abar,r,\obar^{'}) \sim D } \left[ (Q_i(\obar,\abar) - y_i )^2\right] \text{ , where}\\
y_i = r_i + \gamma E_{\abar^{'} \sim \bar{\pi}(\obar^{'}) } \left[ \bar{Q}_i(\obar^{'}, \abar^{'}) - \alpha \log ( \bar{\pi}_i (a_i | o_i) ) \right]
\label{eq:critic_loss_maac}
\end{multline}
where $\bar{Q}_i$ and $\bar{\pi}_i$ are target critic and target policy for agent $i$. In this work, we build on the multi-agent actor-critic framework described above to incorporate learning of emergent roles in each agent. Next, we describe a popular role learning framework in the cooperative setting.
\subsection{Opponent-aware Role learning}
\begin{figure*}[!t]
\centering
\subfloat{\includegraphics[width=.8\textwidth]{Results/RAC_new.pdf}}
\vspace{-5mm}
\caption{ A schematic diagram of our approach.
The self-role encoder $h_S$ generate the role $\rho_i$ for the $i$-th agent and opponent role predictor $h_O$ predicts roles $\hat{\rho}_{i}^o$ for the opponents. Policy $\pi_i$ takes own observation $o_i$ as well as generated role encodings ($\rho_i, \hat{\rho}_{i}^o$) to generate action $a_i$.
Critic $Q_i$ takes both $((o_i, \rho_i, \hat{\rho}_{i}^o), a_i)$ to compute the state-action value.
Critic loss trains the critic module, whereas mutual information loss, diversity loss, and opponent loss train the role encoder modules.
The framework can be trained in an end-to-end manner. }
\label{fig:diagram}
\end{figure*}
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$eps = 1 \cdots $ max episodes}{
$\tau_i^{0} \gets 0$\\
$D_i \gets \emptyset, \forall i $\\
Reset the environment, and get initial $o_i^{1}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
Sample roles $\rho^{t}, \hrho^{t}$ using~(\ref{eq:roles}) for all agents\\
Sample action from respective policies $a_i^{t} \sim \pi_i(o_i^{t},\rho_i^{t},\hrho^{o,t}_i)$ for each agent $i$\\
Each agent receives $o_i^{t+1}, r_i^{t}$ from environment\\
$\tau_i^{t} \gets \tau_i^{t-1} + (o_i^{t},a_i^{t})$\\
$D_i \gets D_i + (o_i^{t},\tau_i^{t-1},a_i^{t},r_i^{t},o_i^{t+1})$\\
}
$T_{update}=T_{update}+1$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B$)}
\label{update}
Select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
$\Lcal_{MI} \gets 0,\Lcal_{D} \gets 0, \Lcal_{Opp} \gets 0$\\
\For{$j = 1 \cdots \text{episode length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S_k \gets (\obar^{t_k+j,k}, \bar{\tau}^{t_k-1+j,k}, \abar^{t_k+j,k}, \rbar^{t_k+j,k}, \obar^{t_k+1+j,k})$\\
$S \gets S + S_k$\\
$\rho^{j,k},\hat{\rho}^{j,k},\Lcal_{MI}^{j,k},\Lcal_{D}^{j,k}, \Lcal_{Opp}^{j,k} \gets Role(S_k)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^{j,k}$\\
$\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^{j,k}$\\
$\Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^{j,k}$\\
}
Update critics, policies and targets using $S, \rho^{j,\cdot}, \hat{\rho}^{j,\cdot}$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho^{t}, \hrho^{t}$ from eq. (\ref{eq:roles}) \\
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using (\ref{eq:miloss}), (\ref{eq:dloss}), and (\ref{eq:opp_loss})\\
return $\rho^t,\hat{\rho}^{t},\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
We introduce \textsc{RAC}, the proposed a role-based actor-critic algorithm, designed for opponent-aware learning in the team-competitive games. In this setting, the agents in the team receive individual as well as shared rewards while the agents in different teams do not receive any shared reward. In this section, we describe the algorithm for simplicity the setting with two teams, each with any number of cooperating agents, which compete with each other according to the rules of individual games (described in section \ref{sec:experiments}). An interesting question in this setting is to determine whether learning opponent-aware role-based policies is beneficial for the teams.
\textsc{RAC}\ is built on the MAAC\ framework described above, and uses both agent-specific policy network $\pi_i$ and critic network $Q_i$. Additionally \textsc{RAC}\ employs two other networks:
\begin{enumerate}
\item $h_S$: self role encoding network
\item $h_O$: opponent role prediction network
\end{enumerate}
Another recurrent neural network (GRU) is used to maintain a running encoding of the trajectory $\tau_i^t$ at time $t$. For simplicity, we overload the notation $\tau_i^t$ to denote the embedding of the trajectory till time $t$, for the rest of the paper. Intuitively, the policy of an agent $i$ depends on its observation $o_i$, latent role $\rho_i$, and predicted roles of opponent agents $\hrho_i^{o} = [\hrho_j]_{j \in Opp(i)}$, where $Opp(i)$ is the set of opponent agents of agent $i$. Hence, $\pi_i(o_i,\rho_i,\hrho_i^o)$ computes the probability of action $a_i$. The self role encoding network $h_S$ and opponent role prediction network $h_O$ take the current observation $o_i^t$ and past action $a_i^{t-1}$ for $h_S$, or trajectory encoding $\tau_i^{t-1}$ for $h_O$ as input; and output a Gaussian distribution over the respective roles:
\begin{align}
&( \mu_{\rho_i^t}, \sigma_{\rho_i^t})= h_S( o_i^t, a_i^{t-1} )\ ;\ \ \PP(\rho_i^t | o_i^t, a_i^{t-1}) = \Ncal ( \mu_{\rho_i^t}, \sigma_{\rho_i^t}) \\
&\hat{\mu}_{\rho_j^t},\hat{\sigma}_{\rho_j^t} = h_O( o_{i}^t, \tau_{i}^{t-1} ) \forall j\neq i\ ;\ \ \hat{\PP}( \hat{\rho}_j^t | o_{i}^t, \tau_{i}^{t-1}) = \Ncal ( \hat{\mu}_{\rho_i^t}, \hat{\sigma}_{\rho_i^t})
\label{eq:roles}
\end{align}
Finally, the critic network $Q_i$ for agent $i$ takes as input the joint observation $\bar{o}$, joint action $\bar{a}$, and the role embeddings of current agent $\rho_i$ and role predictions of its opponents $\hrho_i^o$, $Q_i(\bar{o},\rho_i,\hrho_i^o,\bar{a})$, and outputs an estimate of the value-to-go. Figure \ref{fig:diagram} shows the overall architecture of the \textsc{RAC}\ system.
In the network architecture described above, the per-agent policies $\pi_i$ are trained using standard policy gradient steps, whereas the parameters of critic network $Q_i$, self-role encoder network $h_S$, and opponent-role predictor network $h_O$ are updated by centralized training of critics.
To update policy network of agent $i$, the policy gradient in~(\ref{eq:pol_grad_maac}) is further modified as follows:
\begin{multline}
\nabla_{\pi_i} \Jcal_i(\pi_{i}) = E_{\obar \sim D,\abar \sim \pi,\rho \sim h_S, \hrho \sim h_O} \left[ \nabla_{\pi_i} \log (\pi_{i} (a_{i}|o_{i},\rho_i,\hrho_i^o) \right.\\
\left. ( Q_i(\bar{o},\rho_i, \hrho_i^o, \bar{a})
- b(\obar,\rho_i, \hrho_i^o, \abar_{-i})
- \alpha \log (\pi_i (a_i | \obar_i, \rho_i, \hrho_i^o))) \right]
\end{multline}
Correspondingly, baseline computation is also modified as:
\begin{multline}
b(\obar,\rho_i, \hrho_i^o, \abar_{-i})
= \EE_{a_i \sim \pi_i(o_i, \rho_i, \hrho_i^o)} \left[ Q_i(\obar,\rho_i, \hrho_i^o,(a_i,\abar_{-i}))\right] \\
= \sum_{a_i \in A_i} \pi_i(a_i | o_i, \rho_i, \hrho_i^o ) Q_i(\obar,\rho_i, \hrho_i^o,(a_i,\abar_{-i}))
\end{multline}
While parameters of $\pi_i$ are trained using policy-gradient updates, parameters of the other networks are learned jointly by minimizing an aggregate loss function described next.
Firstly, the critic loss $\Lcal_Q$ defined in~(\ref{eq:critic_loss_maac}) is modified as follows:
\begin{multline}
\Lcal_{Q}=\sum_{i=1}^N \EE_{(\obar,\abar,\rbar,\obar^{\prime}) \sim D,\rho_i \sim h_S, \hrho_i^{o} \sim h_O } \left[ (Q_i(\obar ,\rho_i,\hrho_i^o, \abar ) - y_i )^2 \right] \text{ , where} \\
y_i = r_i + \gamma \EE_{\abar^{\prime} \sim \bar{ \pi }, \rho_i^{'} \sim h_S, \hrho_i^{o,'} \sim h_O }
\left[ \bar{Q}_i(\obar^{'}, \rho_i^{'}, \hrho_i^{o,'},\abar^{'}) \right. \\
\left. - \alpha \log ( \bar{\pi}_i (a_i^{'} | o_i^{'}, \rho_i^{'}, \hrho_i^{o,'}) ) \right]
\label{eq:critic_loss_role}
\end{multline}
\begin{figure*}[!t]
\centering
\subfloat[Touch-Mark]{\includegraphics[ width=.35\textwidth]{Results/snap_catch_goal_moving.pdf}}\hspace{3mm}
\subfloat[Market]{\includegraphics[width=.35\textwidth]{Results/snap_market_moving.pdf}}
\caption{Touch-Mark~(left) is a two-dimensional board game where members of each team try to reach one of the landmarks before any agent from the opponent team reaches any landmark. In Market~(right) teams pick resources scattered on the board and drop those to respective consumers before the opponent does the same.}
\label{fig:evironments}
\end{figure*}
Now, we describe the losses employed for training the self and opponent role encoder modules. For maintaining identifiability of roles with trajectories, we like to minimize conditional entropy $H(\rho_i^t \vert o_i^t, \tau_i^{t-1} )$, which is often intractable. Therefore, we maintain a variational approximation of distribution over roles, conditioned upon trajectories: $q_{\xi}(\rho_i^t | o_i^t, a_i^{t-1}, \tau_i^{t-1})$ and maximize the mutual information between $q_{\xi}(\rho_i^t | o_i^t, a_i^{t-1}, \tau_i^{t-1})$ and $\PP(\rho_i^t |o_i^t, a_i^{t-1})$. Simultaneously we minimize the entropy of role distribution $\PP(\rho_i^t |o_i^t, a_i^{t-1})$ to encode the history in the role encoder network.
Hence, corresponding mutual information loss, $\Lcal_{MI}$, is defined as:
\begin{multline} \label{eq:miloss}
\mathcal{L}_{MI} =
\sum_{i=1}^N \EE_{\tau_i^{t-1},o_i^t \sim D} \left[ D_{KL}(\PP(\rho_i^t|o_i^t,a_i^{t-1})|| \right. \\
\left. q_{\xi}(\rho_i^t|o_i^t,a_i^{t-1},\tau_i^{t-1})) \right]
+ H ( \PP(\rho_i^t|o_i^t,a_i^{t-1}) )
\end{multline}
Here, $D_{KL}$ denotes KL-divergence,
and $H(\cdot)$ denotes the entropy function.
In order to promote diversity between the behaviors of agents within a team, we introduce a new diversity loss $\mathcal{L}_D$, defined as:
\begin{equation}\label{eq:dloss}
\Lcal_D = \sum_{k \in \Kcal } \sum_{(i,j) \in \Ncal(k)}
\EE_{(\bar{\tau}^{t-1},\obar^t) \sim D}
\left[ q_{\xi}(\rho_i^t|\tau_j^{t-1},o_j^t,a_i^{t-1}) \right]
\end{equation}
Here, $\Kcal$ is the set of teams and $\Ncal(k)$ are agents in team $k$. The loss penalizes the probability of $\rho_i$ being identified with trajectory $\tau_j$ of another agent in the same team.
Finally, to train the opponent role prediction network $h_O$, we define the opponent loss $\Lcal_{Opp}$ as:
\begin{equation}\label{eq:opp_loss}
\Lcal_{Opp} = \sum_{\substack{i,j:team(i) \\ \neq team(j)}}
\EE_{(\bar{\tau}^{t-1},\obar^t) \sim D}
\left[ D_{KL}(\hat{\PP}(\rho_j^t|o_{i}^t,\tau_i^{t}) || \PP(\rho_j^t|o_j^t,a_j^{t-1})) \right]
\end{equation}
This loss penalizes the divergence of $\hat{\PP}(\rho_j^t|o_{i}^t,\tau_{i}^t)$, the predicted opponent role distribution for agent $j$ from the point of view of agent $i$, from $\PP(\rho_j^t|o_j^t,a_j^{t-1})$, the self-role distribution of agent $j$.
We define the composite loss function for training the role networks as:
\begin{equation}
\Lcal_{Role}=\Lcal_{D}+\Lcal_{MI}+\Lcal_{Opp}
\end{equation}
Hence, $\Lcal_Q + \Lcal_{Role}$ is the total loss used to learn the combined critic module of \textsc{RAC}\ model. However, it is observed that the evolution of the role networks in the later phase of training causes instability in learning policies, which result in reduced average reward. This phenomenon is mitigated by using an exponentially decaying weightage on role loss $\Lcal_{Role}$. The total critic loss becomes:
\begin{equation}
\Lcal_{tot}=\Lcal_Q + \lambda^{\frac{u}{C}} \times \Lcal_{Role}
\end{equation}
where $\lambda\in [0,1]$ is the decay factor, $ u $ is the training episode count, and $ C $ is a constant affecting the rate of decay over episodes. Algorithm~\ref{modifiedAlgo} includes a pseudo-code for our algorithm. Table ~\ref{tab:notations} contains list of all variables.
\iffalse
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize $E$ parallel environments with $N$ agents\\
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$i_{ep} = 1 \cdots $ max episodes}{
$\tau_i^{0,e} \gets 0$\\
$H_i^{0,e} \gets \emptyset, \forall i $\\
Reset all environments, and get initial $o_i^{1,e}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
In each environment $e$, sample roles for each agent $\rho_i^{t,e}, [\hat{\rho}^{t,e}_j]$ using eq.~(\ref{eq:roles})\\
In each environment $e$, sample action from respective policies $a_i^{t,e} \sim \pi_i(o_i^{t,e},\rho_i^{t,e},[\hat{\rho}^{t,e}_j])$ \\
Each agent receives $o_i^{t+1,e}, r_i^{t,e}$ from respective environments\\
$\tau_i^{t,e} \gets \tau_i^{t-1,e} + (o_i^{t,e},a_i^{t,e})$\\
$H_i^{e} \gets H_i^{e} + (o_i^{t,e},\tau_i^{t-1,e},a_i^{t,e},r_i^{t,e},o_i^{t+1,e})$\\
}
$T_{update}=T_{update}+E$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$,$\text{window length}$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B,\text{window length}$)}
\label{update}
Randomly select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
\For{$j = 1 \cdots \text{window length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S \gets S + (o_{1 \cdots N}^{t_k+j,k}, \tau_{1 \cdots N}^{t_k-1+j,k}, a_{1 \cdots N}^{t_k+j,k}, r_{1 \cdots N}^{t_k+j,k}, o_{1 \cdots N}^{t_k+1+j,k})$
}
$\rho^j,\hat{\rho}^j,\Lcal_{MI}^j,\Lcal_{D}^j, \Lcal_{Opp}^j \gets Role(S)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^j,\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^j, \Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^j$\\
Update critics, policies and targets using $S, \rho^j, \hat{\rho}^j$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho_i^{t,e}, [\hat{\rho}_j^{t,e}]$ from eq. \ref{eq:roles} using for each agent $i$, environment $e$ from $S$\\
$\rho^t \gets \rho_1^{t,\cdot} \cdots \rho_N^{t,\cdot},\hat{\rho}^{t} \gets \hat{\rho}_1^{t,\cdot} \cdots \hat{\rho}_N^{t,\cdot}$\\
\For{$team = 1 \cdots \text{number of teams}$ }{
\For{$agent=1 \cdots \text{number of agents in a team} $}{
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using eq. ~\ref{eq:miloss}, eq.~\ref{eq:dloss}, and eq.~\ref{eq:opp_loss}
}
}
return $\rho^t,\hat{\rho}^t,\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
\fi
\section{Introduction}
Many real-world scenarios can be modeled as \textbf{team-competitive Markov games}~\citep{lowe2017multi} (also called \textit{mixed cooperative-competitive games}), where each agent receives an individual reward for its actions and a shared reward with other agents in the same team for a joint outcome. For example, while playing soccer, each player may receive a reward for their individual performance and a shared reward for the team's victory. In some social interactions, e.g., companies vying for customers, each employee may receive a reward for their individual performance and a shared reward for the team performance.
\textbf{Roles} emerge naturally in such reward scenarios, since they allow agents within a team to cooperate effectively. For soccer players, the roles could be ``striker'' and ``defender'', while for company employees, the roles could be ``production'' and ``sales''. In this work, we study the problem of learning policies with emergent roles in a reinforcement learning (RL) setting for team-competitive Markov games.
We use multi-agent reinforcement learning (MARL) to learn the optimal policies of individual agents in this setting. While many popular deep MARL techniques exist for the shared-reward setting
~\citep{gupta2017cooperative, mordatch2018emergence, rashid2018qmix,son2019qtran, wang2019learning}, relatively fewer techniques exist for the mixed-reward setting.
Multi-agent actor-critic algorithms~\citep{lowe2017multi,iqbal2019actor}, following the centralized training and decentralized executions concept, are the most popular MARL approaches in the mixed-reward setting, which is needed to model the team-competitive scenarios. MADDPG \citep{lowe2017multi}
is one of the earliest techniques for learning deterministic policies in mixed reward settings using the actor-critic framework. MAAC ~\citep{iqbal2019actor}
introduces attention mechanisms in critic modules to selectively pay attention to information collected from other agents, resulting in better scalability in complex multi-agent environments.
However, to the best of our knowledge, role-oriented policy learning in the actor-critic framework, particularly in competitive games, is yet unexplored.
On the other hand, the concept of predefined roles for agents has been explored to reduce design complexity in many MARL systems~\citep{spanoudakis2010using, bonjean2014adelfe}.
There are few works addressing dynamic role emergence based MARL in domain-specific problems~\citep{barrett2015cooperating,leottau2015study,urieli2011optimizing,roy2020promoting,zhang2021hierarchical,khaleefah2021exploring}.
Recently, Wang et al.~\citep{wang2020roma} introduce automatic role emergence in MARL for cooperative tasks in their system called \textsc{ROMA}. It proposes a latent role-based task decomposition among the agents built over the QMIX~\citep{rashid2018qmix} platform.
In a similar line, Wang et al.~\cite{wang2020rode} propose RODE, a bi-level role-based learning framework where joint action-space is decomposed into restricted role-action spaces by clustering actions by their effects. Finally, it employ a role-selector to search role at a smaller role space, while role policies learn at a reduced action-observation space, significantly improving learning efficiency. Liu et al.~\citep{liu2022rogc} present ROGC, a role-oriented graph convolution based MARL framework, where roles are generated by classifying agents, graph convolution module is emplyed for efficient intra-role communication and finally effective individual policies are generated by embedding role information into the algorithm.
However, all of the above frameworks are built over QMIX, where a global utility function is composed of per-agent local utility functions, and the composition relies upon the assumption that all agents share the same reward. Hence, this setting can not be immediately extended to a mixed-reward setting, prohibiting role-oriented policy learning in the presence of opponents through joint learning of any form.
Recently, opponent-aware actor-critic algorithms have been explored in many settings, including
learning time-dynamical opponent models TDOM-AC~\cite{wen2019probabilistic} , maximum entropy regularized opponent models ROMMEO~\cite{tian2019regularized}, opponent model employing variational Bayes PRR~\cite{tian2022multi}, etc. These works validate the necessity of opponent models for improving performance in competitive games. However, they are primarily limited to one-to-one competitions and do not consider role-aware opponent modeling. However, in a team competition setting, where each agent employs a role-aware policy within the team, considering roles in opponent modeling can benefit agents by reducing design complexity.
In this work, we propose \textsc{RAC}, an actor-critic algorithm that combines role learning and role-aware opponent modeling in mixed-reward setting.
\textsc{RAC}\ uses the actor-critic framework based on MAAC \cite{iqbal2019actor}, which allows the critic to use joint observations of all agents, hence incorporating the opponent's experience in the critic.
Our primary contributions are two-fold. First, we incorporate role-encoder in the actor-critic framework that learns dynamic and diverse emergent roles per team.
Moreover, for opponent modeling, we incorporate an opponent role predictor network for each team, which is accessible to the policy at the execution time.
The opponent role prediction network is trained jointly along with the role encoder, critic, and policy network. In particular, during centralized training, the opponent role predictor is trained to mimic the actual role encoding distribution of the opponent. During execution, that benefits agents to select their response.
The learned roles are dynamic depending on the agent's trajectory and diverse between agents within the team. Experiments using the proposed episodic board games, Touch-Mark\, and Market, show that \textsc{RAC}\ agents outperform MAAC\ agents in direct competitions. Furthermore, our experiments demonstrate that the dynamic roles learned by \textsc{RAC}\ conform to the intuitive role-specific behaviors for each game, thus confirming the effectiveness of emergent roles.
In summary, our contributions are as follows:
\begin{itemize}
\item We propose \textsc{RAC}{,} the first opponent-aware role-oriented policy learning in actor-critic framework, designed for mixed reward setting, as per our knowledge.
\item \textsc{RAC}\ encourages diverse roles within a team and ensures that the role encoding captures the agent trajectory.
\item Most importantly, \textsc{RAC}\ learns an opponent role predictor module, utilizing opponent role encodings during the centralized training that it employs in policy during the execution time. Eventually, the opponent role predictor guides agents in choosing strategic responses against adversaries.
\item Empirically, we validate improved policy learning of \textsc{RAC}{, } through increased reward as well as demonstrate the identifiability of learned roles with intuitive role-specific behaviors of agents in two episodic board games, Touch-Mark\ and Market{.}
\end{itemize}
\section{Related Literature}
Multi-agent Actor-critic methods, primarily following the concept of centralized training and decentralized executions, are one of the most dominant and popular techniques in deep Multi-agent Reinforcement Learning (MARL) approaches~\citep{sukhbaatar2016learning,gupta2017cooperative,lowe2017multi,foerster2017stabilising,mordatch2018emergence,rashid2018qmix,foerster2018counterfactual}.
Among these works, MADDPG~\citep{lowe2017multi}, a multi-agent adaptation of DDPG~\citep{lillicrap2016continuous}, considers continuous action space and deterministic policies in mixed-reward settings and successfully trains the agents complex coordination strategies in cooperative as well as competitive scenarios.
MAAC~\citep{iqbal2019actor}, an extension of SAC~\citep{haarnoja2018soft}, introduces attention mechanism in critic modules to selectively pay attention to information collected from other agents, resulting in massive scalability in complex multi-agent environments. Also, there exists a series of value-function-factorization-based methods that train decentralized policies in a centralized end-to-end fashion, employing a joint action value network~\citep{rashid2018qmix}.
There are many follow-ups on actor-critic-based MARL algorithms, addressing a variety of issues, namely SAC~\citep{qu2020scalable} for improving scalability, SEAC~\citep{christianos2020shared} for sharing experience, LICA~\citep{zhou2020learning} for the credit assignment problem, TESSERACT~\citep{mahajan2021tesseract} for tensorizing the critics, Bilevel Actor Critic~ \cite{zhang2020bi} for multi-agent coordination problem with unequal agents,
DAC-TD~\cite{figura2022cooperative} for training agents in a privacy-aware framework, VDAC~\cite{su2021value} for combining value-decomposition framework with actor-critic, Scalable Actor Critic~\cite{lin2021multi} for scalable learning in stochastic network of agents, etc.
However, none of these works consider the role emergence paradigm in their model.
On the other hand, role division is quite common and efficient for accomplishing any complex collective task in human society~\citep{butler2012condensed}. Following that, many multi-agent systems also decompose the task into pre-defined roles, resulting in reduced design complexity and faster learning ~\citep{wooldridge2000gaia,padgham2002prometheus,cossentino2005passi,spanoudakis2010using,bonjean2014adelfe}.
In recent times, Majumdar et al.~\citep{majumdar2020evolutionary} present a line of work where MADDPG agents learn different strategies by de-coupling and automatically weighting individual and global goals in a population-based training paradigm.
Liu et al.~\citep{liu2021coach} consider the problem of coordination with dynamic composition, where a coach agent with global view distribute individual strategies to dynamically varying player agents. Jiang et al. ~\citep{jiang2021emergence} promote sub-task specialization via emergence of individuality through generating dynamic intrinsic rewards. Su et al.~\cite{su2022divergence} propose another actor-critic algorithm where agents learn diverse policies in cooperative team games.
Also, there is a series of works for role-oriented MARL for specialized domains like Robo-soccer ~\citep{barrett2015cooperating,leottau2015study,urieli2011optimizing,ossmy2018variety}, football environment~\citep{roy2020promoting}, swarn systems~\citep{zhang2021hierarchical}, image feature extraction tasks~\citep{khaleefah2021exploring} showing how complex policies can be learned by decomposing them into simpler sub-policies.
In contrast, Wang et al. propose \textsc{ROMA}~\citep{wang2020roma} and RODE~\citep{wang2020rode}, two dynamic and adaptive role-oriented MARL frameworks built over QMIX~\citep{rashid2018qmix} in a quite general setting and show dynamic role emergence within the team on SMAC benchmark tasks.
Recently, Liu et al. present ROGC~\citep{liu2022rogc}, a graph convolution based role-oriented MARL framework. However, all of them are inherently restricted to shared reward settings, i.e., only cooperative tasks.
Our work closely resembles~\citep{wang2020roma}, with the main distinction being that our work incorporates opponent modeling along with adopting team-based role learning to the actor-critic framework to exploit the advantage of training in the presence of opponents.
There are few recent works exploring opponent modeling in actor-critic framework, namely, TDOM-AC~\cite{tian2022multi} for learning time dynamical opponent model, ROMMEO~\cite{tian2019regularized} for learning maximum entropy regularized opponent model, PRR~\cite{wen2019probabilistic} for learning opponent model employing variational Bayes, to name a few. They primarily consider the setting of two-player competitive games, where each agent maintains a prototype of opponent policy. However, they do not consider team competition settings. Extending their framework to our role-aware competitive team setting is not immediate.
\section{Conclusion and Future work}
\vspace{2mm}
We propose an algorithm for opponent-aware role-based learning in actor-critic framework targeted towards team competitions. Our algorithm combines a self-role encoder and an opponent-role predictor in actor-critic framework for learning an optimal policy. We analyze our approach in two scenarios where we show how our method improves the quality of learning. As a future direction, we intend to extend our algorithm to more general settings where multiple teams can compete among themselves.
\iffalse
\PK{I am keeping algorithm here temporarily for some modification. I am not editing the method section. }
\begin{algorithm}
\caption{ \textsc{RAC}\ }
\label{modifiedAlgo}
Initialize replay buffer, $D$\\
$T_{update} \leftarrow 0$\\
\For{$i = 1 \cdots $ max episodes}{
$\tau_i^{0} \gets 0$\\
$H_i^{0} \gets \emptyset, \forall i $\\
Reset all environments, and get initial $o_i^{1}$ for each agent $i$\\
\For{$t=1 \cdots $ episode length }{
Sample roles for each agent $\rho_i^{t}, \hrho^{o,t}_i$ using~(\ref{eq:roles})\\
Sample action from respective policies $a_i^{t} \sim \pi_i(o_i^{t},\rho_i^{t},\hrho^{o,t}_i)$ \\
Each agent receives $o_i^{t+1}, r_i^{t}$ from environment\\
$\tau_i^{t} \gets \tau_i^{t-1} + (o_i^{t},a_i^{t})$\\
$H_i \gets H_i + (o_i^{t},\tau_i^{t-1},a_i^{t},r_i^{t},o_i^{t+1})$\\
}
$T_{update}=T_{update}+1$\\
\If{$T_{update} \geq $ min steps per update}{
\For{$j=1 \cdots $number of updates}{
$H_B \gets B$ episodes sampled from $D$\\
Update($H_B$,$\text{window length}$)\\
}
$T_{update} \leftarrow 0$\\
}
}
\end{algorithm}
\begin{algorithm}
\caption{Update($H_B,\text{window length}$)}
\label{update}
Randomly select $t_1,\cdots, t_B$, the starting time-step for each of the $B$ episodes\\
$\Lcal_{MI} \gets 0,\Lcal_{D} \gets 0, \Lcal_{Opp} \gets 0$\\
\For{$j = 1 \cdots \text{window length}$ }{
$S \gets \emptyset$ \\
\For{$k=1 \cdots B $}{
$S \gets S + (o_{1 \cdots N}^{t_k+j,k}, \tau_{1 \cdots N}^{t_k-1+j,k}, a_{1 \cdots N}^{t_k+j,k}, r_{1 \cdots N}^{t_k+j,k}, o_{1 \cdots N}^{t_k+1+j,k})$
}
$\rho^j,\hat{\rho}^j,\Lcal_{MI}^j,\Lcal_{D}^j, \Lcal_{Opp}^j \gets Role(S)$\\
$\Lcal_{MI} \gets \Lcal_{MI} + \Lcal_{MI}^j,\Lcal_{D} \gets \Lcal_{D} + \Lcal_{D}^j, \Lcal_{Opp} \gets \Lcal_{Opp} + \Lcal_{Opp}^j$\\
Update critics, policies and targets using $S, \rho^j, \hat{\rho}^j$\\
}
$\psi_{Role} \gets \argmin_{\psi}(\Lcal_{MI}+\Lcal_{D}+\Lcal_{Opp})$
\end{algorithm}
\begin{algorithm}
\caption{Role($S$)}
\label{role}
Obtain roles $\rho_i^{t}, \hrho_i^{o,t}$ from eq. (\ref{eq:roles}) using for each agent $i$\\
$\rho^t \gets \rho_1^{t} \cdots \rho_N^{t},\hat{\rho}^{t} \gets \hat{\rho}_1^{t} \cdots \hat{\rho}_N^{t}$\\
\For{$agent=1 \cdots N$}{
Calculate $\Lcal_{MI}$, $\Lcal_{D}$ and $\Lcal_{Opp}$ using (\ref{eq:miloss}), (\ref{eq:dloss}), and (\ref{eq:opp_loss})
}
return $\rho^t,\hat{\rho}^t,\Lcal_{MI},\Lcal_{D},\Lcal_{Opp}$
\end{algorithm}
\fi
\section{Experimental Results}
\label{sec:experiments}
\begin{figure*}[!t]
\centering
\subfloat{\includegraphics[width=.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Rewards.pdf}}
\subfloat{\includegraphics[width=.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Landmark.pdf}}
\subfloat{\includegraphics[width=.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Rewards.pdf}}
\subfloat{\includegraphics[width=.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_avg_0_Dropped.pdf}}
\caption{Performance comparison. We play \textsc{RAC}\ agents against MAAC\ agents, who are separately trained using self-play for $300000$ training steps. We play them against each other after each $1000$ step and report performance, averaging over $4$ pairs of tournament-play. For Touch-Mark, we report the average team reward and the fraction of times each team reaches the landmark. For Market, we report the average team reward and the fraction of times each team drops a resource to its respective consumer, after picking it from its source. }
\label{fig:tournament}
\end{figure*}
\begin{figure}[!t]
\centering
\subfloat{\includegraphics[width=0.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_vary_manual_decay_update_rsubseq_v3_coll_avg_0_Roles.pdf}\label{fig:role-catch-goal}
\subfloat{\includegraphics[width=0.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_coll_avg_0_Roles.pdf}\label{fig:role-market}}
\caption{Role emergence. For Touch-Mark, we observe, \textsc{RAC}\ agents result in more collisions than MAAC\ agents. In Market, we observe that \textsc{RAC}\ agents choose to pick more diverse items than MAAC\ agents. }
\label{fig:role}
\end{figure}
\begin{figure}[!t]
\centering
\subfloat{\includegraphics[width=0.23\textwidth]{Results/CatchGoalFixed_cooperative_avg_Landmark.pdf}\label{fig:cooperative-catch-goal}
\subfloat{\includegraphics[width=0.23\textwidth]{Results/Market3_cooperative_avg_Dropped.pdf}\label{fig:cooperative-market-2}}
\caption{We compare \textsc{RAC}\ with RAC$_{Team}$, an extension of \textsc{ROMA}\ with opponent modeling removed. However, we observe that performance-wise \textsc{RAC}\ is still outperforming RAC$_{Team}$ . }
\label{fig:cooperative}
\end{figure}
\begin{figure*}[!t]
\centering
\subfloat[]{\includegraphics[width=.23\textwidth]{Results/CatchGoalFixed_Role_vs_MAAC_opponent_update_rsubseq_v3_vary_decay_avg_Landmark.pdf}\label{fig:vary_decay_catch_goal}}
\subfloat[]{\includegraphics[width=.23\textwidth]{Results/Market3_Role_vs_MAAC_opponent_update_rsubseq_v3_vary_decay_avg_Dropped.pdf}\label{fig:vary_decay_market}}
\subfloat[]{\includegraphics[width=0.23\textwidth]{Results/CatchGoalFixed_ablations_avg_Landmark.pdf}\label{fig:ablations_catch_goal}}
\subfloat[]{\includegraphics[width=0.23\textwidth]{Results/Market3_ablations_avg_Dropped.pdf}\label{fig:ablations_market}}
\caption{(Left) Hyperparameter tuning. Here the decay parameter $\lambda$ is varied in $[0,1]$ and the performance of \textsc{RAC}\ teams, played against a random set of MAAC\ models, is reported. (Right) Ablations. Performance of various ablations of \textsc{RAC}\ are reported, which show that the three distinct objectives are necessary for best performance in both the environments.}
\end{figure*}
In this section, we investigate two primary objectives. First, we check how employing the role network influences the performance of our algorithm in direct competition with baseline algorithm. Second, we check the effectiveness of the role network in emerging roles dynamically within the team.
Towards this end, we employ two games, namely Touch-Mark\ and Market, designed in MPE~\citep{lowe2017multi,mordatch2018emergence}.
\noindent \textbf{Touch-Mark~\citep{koley2022offsetting}{:}}
It is an episodic board game that consists of two teams, each comprising of multiple agents (2 in our case) and two static landmarks, all randomly positioned at the beginning of each episode and each team tries to reach at least one of those landmarks earlier than any member of the other team.
The agents start with velocity zero and slowly accelerate until they reach maximum permissible velocity.
The episode ends when an agent reaches a landmark. The winning agent's team (i.e. both the team members) receives a large reward $r_l$, incurring $-r_l$ penalty to the opposite team.
Additionally, each agent receives a small penalty,
proportional to its distance from
the nearest target at every time step to guide it towards the nearest target.
Moreover, an agent can collide with an agent of the opponent team to divert it from its path. The collision makes both agents temporarily stationary. This mechanism is introduced so that an agent
has the option to stop an agent of the opponent team from reaching a landmark, thus facilitating the fellow teammate to reach a landmark first.
\noindent \textbf{Market{:}}
This is a novel episodic board game with two teams, each consisting of multiple players ($2$ in our case).
At the beginning of each episode, four resources, two for each type of resource are randomly placed on the board, along with one consumer for each type of resource.
Agents need to acquire the resources and drop those to the respective consumers.
After acquiring a resource, the agent continues to receive a penalty proportional to its distance from the appropriate consumer, to guide him towards the corresponding consumer. Both acquiring and dropping resources result in high rewards for corresponding team.
The termination condition for the episode is timeout of 50 steps.
The consumer remains alive only until one resource of its type is dropped to it. Therefore for highest scores, one team must acquire both types of resources and drop them to respective consumers before their opponents.
\noindent \textbf{Experimental setup}
We use a one-layer feed-forward neural network followed by GRU unit for getting initial embedding of the trajectories. For each team, a double-layer MLP is employed for each of $\PP(\rho_i^t | o_i^t, a_i^{t-1})$,$\hat{\PP}(\rho_j^t | o_i^t, a_i^{t-1})$ and $q_{\xi}(\rho_i^t | \tau_i^{t-1},o_i^t)$. Critic and policy network architecture follows MAAC~\citep{iqbal2019actor} except inputs of both critic and policy modules for each agent $i$ are modified to include a role tuple for that agent, containing self-role $\rho_i^t$ and predicted opponent-roles $\hrho_i^{o,t}$.
\noindent \textbf{Baselines}
Our primary baseline is MAAC~\citep{iqbal2019actor}.
We also compare our method with RAC$_{Team}$\ which follows \textsc{RAC}\ except does not share information across teams during the training phase.
This mimics the setting where competing teams separately employ a cooperative role-oriented algorithm like \textsc{ROMA}\ and only sacrifices the advantage of training in presence of opponents.
\noindent \textbf{Metrics}
We report cumulative team rewards per episode to measure the performance of the algorithms. As a qualitative measure, we report the rate of touching landmarks per episode for each team in Touch-Mark, and the average number of dropped resources per episode for each team in Market{.} The average resources dropped per episode can vary between $0$ and $2$ as there are $2$ resources in Market. All measures are reported after taking an average across all seeds used.
\subsection{Results}
\noindent \textbf{Direct tournament:} Here, we compare \textsc{RAC}\ with MAAC\ by directly playing them against each other. Here, we use the \textsc{RAC}\ and MAAC\ teams, trained in self-play for $3 \times 10^5$ training episodes starting with random seed. At fixed interval during the training we play them among themselves for $1000$ episodes. This experiment is repeated for $4$ pairs of \textsc{RAC}\ and MAAC\ teams, and the average metrics are reported in fig. \ref{fig:tournament} for Touch-Mark~(left) and Market~(right){.}
We observe that, as the training progresses, \textsc{RAC}\ continues to outperform MAAC. The performance gap is quite significant for Market, whereas, in Touch-Mark\ the gain is marginal.
We explain it as follows. Market\ deals with a comparatively complex setting, making it difficult for the agents to play strategically without the explicit assistance for sub-task specialization. The role-driven policy learning framework of \textsc{RAC}\ seems to be effective here to reduce the observation-action space to explore and quickly guide them towards improved policy learning by devising opponent-aware strategic behaviors.
Touch-Mark\ deals with a simpler setting, where agents learn effective strategies comparatively faster even without external asistance on implicit role division or sub-task specialization, resulting in only marginal improvement of \textsc{RAC}\ over MAAC{.}
Further, we observe a performance drop of \textsc{RAC}\ agents as the training approaches for Touch-Mark{.} We explain this phenomenon as follows. \textsc{RAC}\ accelerates learning strategic behavior in Touch-Mark, more specifically, splits the responsibility of collision and chasing the landmark among team members. However, as MAAC\ agents quickly master the strategy in the next phase of training, the advantage of \textsc{RAC}\ almost disappear. Therefore, with the improvement of the expertise of MAAC\ team, the performance of \textsc{RAC}\ team drops, finally resulting in marginal gain over MAAC\ agents.
\noindent \textbf{Role learning:}
Further, we investigate whether \textsc{RAC}\ is actually helping teammates to choose diverse roles. Also, we study the strategic behaviors of \textsc{RAC}\ agents.
In Touch-Mark, as we have already mentioned, though collision incurs a small penalty, one can strategically employ collision with an opponent to delay the opponent agent from reaching the landmark.
Similarly, in Market, if the teammates choose to go for different resources, they can acquire higher rewards in future.
To check whether algorithms are able to train their agents such strategic moves, we summarize role-discrimination statistics in fig \ref{fig:role} for both the games.
For Touch-Mark\ we report the fraction of winning episodes where the agent other than the one who touches the landmark, collides with the opponents. Similarly, for Market, we report the fraction of episodes where members within the team drop different types of resources.
In both cases, we find that \textsc{RAC}\ agents tend to employ strategic moves more than MAAC\ agents, which can possibly explain better performance of \textsc{RAC}\ in direct competitions.
The key factor behind such diverse behavior is employment of $\Lcal_D$ and $\Lcal_{Opp}$, which incentivizes the role network to assign diverse as well as opponent-aware role distributions to teammates respectively.
\noindent \textbf{Comparison with RAC$_{Team}$\ extending~\citep{wang2020roma}:} Fig \ref{fig:cooperative} compares \textsc{RAC}\ with RAC$_{Team}$, which is a trivial extension of \textsc{ROMA}~\citep{wang2020roma} to actor-critic framework, where we use the two regularizers used in ~\citep{wang2020roma} for role learning, but remove the opponent modeling. Instead, agents are trained using the simplified-role-module enriched MAAC\ for both of the teams separately, whereas a team it enjoys shared reward setting. Finally, agents trained under \textsc{RAC}\ and RAC$_{Team}$\ are separately played against random MAAC\ teams selected from a common set of MAAC\ teams and average comparative results are plotted
in fig \ref{fig:cooperative}.
The superiority of \textsc{RAC}\ validates that our opponent-aware role-oriented framework provides an additional layer of guidance to agents in competitive team games over a straight-forward extension of \textsc{ROMA}~\citep{wang2020roma} to actor-critics, thus emphasizing the motivation of our approach.
\noindent \textbf{Hyperparameter ($\lambda$) tuning:} Fig~[\ref{fig:vary_decay_catch_goal}-\ref{fig:vary_decay_market}] presents performance of \textsc{RAC}\ teams against randomly chosen MAAC\ teams for different value of decay parameter $\lambda$.
Both in Touch-Mark\ and Market, performance significantly varies based on chosen $\lambda$, with $\lambda=0.5$ giving the best results. $\lambda=0.9$ results in a slower decay of role loss, hence resulting in greater instability in policy learning, while $\lambda=0.1$ results in too fast a decay. This experiment confirms our intuitive hypothesis from real-world experience, that sustained role evolution beyond a point is detrimental to team performance.
\noindent \textbf{Comparison with ablations:}
Fig. [\ref{fig:ablations_catch_goal}-\ref{fig:ablations_market}] compare \textsc{RAC}\ with three ablations of \textsc{RAC}\, which minimize either $\Lcal_D$ or $\Lcal_{MI}$ or $\Lcal_D+\Lcal_{MI}$, whereas \textsc{RAC}\ minimizes $\Lcal_D + \Lcal_{MI}+\Lcal_{Opp}$. In both games, \textsc{RAC}\ is observed to outperform the rest.
In Market\, we also observe $\Lcal_{MI}$ learns roles more identifiable with trajectory and therefore training better dropping skill, in comparison with the other two baselines.
\if{0}
For Touch-Mark, now we take a deeper look by observing the event of touching landmark for each agent individually in Fig. \ref{fig:tournament-landmark-agent-wise}. It is found that in case of \textsc{RAC}\ one agent learns faster than the rest, resulting in initial advantages of team \textsc{RAC}\ over team MAAC. But the other \textsc{RAC}\ agent barely shows any improvement in touching the landmark. MAAC\ agents learn slower, but, the variation in fellow teammates' learning rate is smaller. MAAC\ slowly trains both of the agents, simultaneously, eventually resulting in cumulative superior performance.
This observation indicates employing role network fastens the learning process and useful in shorter term.
However, if both skills are not explicitly maximizing rewards, we may find that eventually this role distinction within teammates may result in higher variance in the team members' skill-specific expertise, resulting in slowing down the overall team performance, as in Touch-Mark.
On the other hand, this did not happen in Market. As a reason, we can observe that, in Market, agents within a team try to pick different resources as a result of role emergence. As both of these actions implicitly increase the rewards, learning specific roles does not hamper retaining the higher rewards throughout, unlike Touch-Mark.
\fi
\section{Opponent-aware Role-based MARL}
\label{sec:problem_formulation}
In this section, we develop the proposed algorithm \textsc{RAC}\ for Multi-agent Policy learning with emergent roles in the team-competitive Markov games setting \cite{lowe2017multi}.
\input{100prelim}
\input{101maac}
\input{102roma}
\input{103rac}
\input{200exp}
\input{300conclusion}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
Nowadays, Unmanned Aerial Vehicles (UAVs) or drones, or more generally Unmanned Aerial Systems (UAS), are increasingly used in many applications and are invoked to drive advanced air mobility (AAM) and urban air mobility~\cite{airbusUrbanMobility} (UAM).
UAM is a subset of AAM that uses small automated aircraft to work at low altitudes in urban and suburban areas.
So far, only manned/unmanned ground and manned aerial vehicles have been employed in these markets.
However, trucks are constrained by terrain and suffer from highly congested infrastructures, while the use of planes is costly and unaffordable to most people.
Therefore, the use of UAS is a viable solution in many applications~\cite{romaDroneCAPOMASI} also because they do not emit greenhouse gases.
The operation of UAS is subdivided into categories depending on the risk of the application.
Operations with low risk do not require a prior authorization before the flights, while operations with a higher risk require special authorization.
Currently, any operation that forces a drone to move Beyond Visual Line of Sight (BVLoS\xspace) requires special authorization.
To enable BVLoS\xspace missions, there are currently many ways~\cite{europaUnmannedAircraft}, e.g., by a NOtice To AirMen (NOTAM) authorization, by Extended Visual Line of Sight (EVLoS\xspace) flights, or by exploitation of ``corridors'' or ``polygons''.
A NOTAM isolates a certain area, and the risk is lowered because only the authorized UAV operation can use that area and all the facilities on the ground are aware of the drone's presence.
An EVLoS\xspace flight emulates the BVLoS\xspace operation by connecting different VLoS\xspace operators in series.
Finally, exploitation of ``corridors'' (e.g., high-voltage transmission lines) or ``polygons'' (e.g., fields), is the solution closest to BVLoS\xspace flights in a low-risk scenario.
However, these methods are artificial, unflexible, and unfeasible to allow daily BVLoS\xspace flights in the immediate future.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.35]{figures/big_picture.png}
\caption{Big picture of our envisioned BVLoS\xspace scenario. The connectivity is provided by the cellular infrastructure. The risk is present on the ground in some areas. The UAV operator has a limited view and flies the drone in BVLoS\xspace according to a certain trajectory.}
\label{fig:big_picture}
\end{figure}
To pursue flexible and scalable BVLoS\xspace flights, there are two big challenges: risk and connectivity of the drone (see Figure~\ref{fig:big_picture}).
The \textit{ground risk} depends on the concentration of people over places in case of some malfunctioning~\cite{primatesta2019risk}.
The \textit{air risk} depends on the presence of other drones in the area.
In this paper, we assume that we fly at the legal height limit of 120 \unit{m}, which reduces the risk of coming across other aircraft, which normally fly much higher than this limit.
The presence of buildings, towers, and trees can slightly mitigate the risk since they can act as shields or shelters~\cite{milano2022air}, but they also represent obstacles during the flight.
A previous risk assessment helps the UAVs to be aware in advance of the possible ground and air risks.
So, our goal is to build the UAVs' risk awareness by leveraging ground maps and consequently trying to limit the risks of UAV flight.
Maps that describe and constrain specific areas on the ground/air can be retrieved from many providers, such as OpenStreetMap~\cite{openstreetmap} or D-Flight~\cite{dflight}.
Regarding \textit{connectivity}, a UAV must communicate its position anytime so that its route can be monitored by the ground, and possibly modified by providing new tasks while receiving real-time updates from the UAV~\cite{aydin2022authentication}.
Connectivity is indeed the way to extend the drone operation with the same level of dependability guaranteed by the VLoS\xspace operations.
Since thinking of deploying an ad hoc infrastructure for communicating is not sustainable, it is realistic to rely on already available ones like those used for ground users (cellular).
There are several generations of networks already deployed (e.g., 4/5G) that can provide BVLoS\xspace links with different characteristics (e.g., bandwidth, latency)~\cite{vallero2021base} suitable for different types of UAVs, and types of applications~\cite{renga2022can}.
In this paper, we propose a novel multi-layer framework that includes information from different layers like no-fly zones, obstacles, ground risk, and communication infrastructure.
Such layers are then discretized in order to build a graph-based data structure which will be exploited by drone path planning algorithms.
The rest of the paper is organized as follows.
Section~\ref{sec:related} reviews the related work.
Section~\ref{sec:model} formally defines the multi-layer framework, while Section~\ref{sec:graph} proposes the graph-based data structure for path planning.
Finally, Section~\ref{sec:conclusion} offers conclusions and future works.
\section{Related Work}\label{sec:related}
In this section, we describe the existing works on risk aware path planning and handover management for UAVs flying BVLoS\xspace.
We remark that none of the works addresses the two factors simultaneously.
\paragraph{Risk Analysis and Path Planning}
BVLoS\xspace operation involves a number of risks, including the potential for collisions with other aircraft, loss of communication with the operator, and failure of onboard sensors or systems.
Primatesta et al.~\cite{primatesta2019risk,primatesta2020risk,primatesta2020ground}, propose the use of a two-dimensional location-based Risk Map to define the risk to the population caused by the crashing of the UAV.
The risk map is generated using a probabilistic approach and combines several layers, including population density, sheltering factor, no-fly zones, and obstacles.
The risk values are defined by a risk assessment process using different uncontrolled descent events, drone parameters, and environmental characteristics, as well as uncertainties in parameters.
By using the Risk Map, the authors propose an optimum risk path based on Rapidly-exploring Random Tree (RRT*) and A* algorithms.
Note that in our work we included the risk map by Primatesta et al. as one of the layers in the framework we presented.
The authors in~\cite{balachandran2017path} present a path planning algorithm that accommodates real-time traffic and geofence constraints in low-altitude airspace BVLoS\xspace.
Their proposed algorithm integrates an RRT technique with Detect and Avoid Alerting Logic.
Similarly in~\cite{kim2022risk}, authors propose a methodology to analyze the capacity of UAV corridors by linking the collision rate of the corridor and the failure rates of UAVs with the number of fatalities on the ground.
Finally, the authors in~\cite{savkin2016problem} propose an optimization model to navigate an aircraft to reach its destination by minimizing the maximum threat level and the length of the flight path using a geometric procedure.
\paragraph{Cellular Communications and UAVs}
One of the main technical challenges of BVLoS\xspace operations is maintaining a reliable link between the drone and the operator.
This is particularly important to ensure that the operator can monitor the drone's flight and provide real-time guidance or take control if necessary.
In fact, one of the major goals in BVLoS\xspace flight is to keep high-quality radio communications.
One issue that can arise when using cellular networks for BVLoS\xspace communication is the problem of handover.
Handover refers to the process of transferring the connection between the drone and the cellular network from one base station to another as the drone moves through the coverage area.
Handover is a critical aspect of cellular communication, as it enables the drone to maintain a continuous connection to the network when it flies.
However, handover can also be a challenge for BVLoS\xspace drones, as the drone may be flying at high speeds or in areas with limited coverage, which can make it difficult to maintain a stable connection.
Several studies have addressed the problem of handover for BVLoS\xspace drones.
In~\cite{amer2020performance}, authors study the performance of cellular-connected UAVs under 3D practical antenna configurations.
Their results reveal that vertically-mobile UAVs are susceptible to altitude handover due to consecutive crossings of the nulls and peaks of the antenna side lobes.
The authors in~\cite{chowdhury2020handover} propose an approximation of the probability mass function (PMF) of handover count (HOC) as a function of the UAV's velocity, HOC measurement time window, and ground base station (GBS) densities.
Furthermore, the authors in~\cite{fakhreddine2019handover} contributed an experimental study on cell association and handover rates for drones, connected to an LTE-A (Long Term Evolution Advanced) network in a suburban environment.
Their experiments show that the handover frequency increases with increasing flight altitude.
\section{System Model}\label{sec:model}
Let us consider a 3D environment bounded by a rectangular parallelepiped (briefly, \textit{box}, see Figure~\ref{fig:box}) denoted by $B$ characterized by a length $B_L$, width $B_W$, and height $B_H$, that lies on the ground plane at height $0$, with $B_L$, $B_W$, and $B_H \in \mathbb{R}^+$.
Moreover, we consider $B$ divided into a number of \textit{cubes} (or \textit{cells}) of the same side length $\ell$, thus making a discrete approximation of the environment.
We denote each cube with its relative position from the origin with a tuple $c = (x_c, y_c, z_c) \in B$, where $x_c$, $y_c$, and $z_c \in \mathbb{N}$ are the $x$-coordinate, $y$-coordinate, and $z$-coordinate, respectively, of $c$ in the discretized environment.
For simplicity, we assume $B$ to be a flat environment.
In $B$, the cell $c=(x_c, y_c, 1)$ lies at the lowest level at a height $\ell$ above the ground.
The sizes of $B$ are $n=\frac{B_L}{\ell}$, $m=\frac{B_W}{\ell}$, and $h=\frac{B_H}{\ell}$, respectively, and so there are $nmh$ cubes.
Since flying in BVLoS\xspace requires some essential and detailed information about the surrounding environment in $B$, we propose a \textit{multi-layer framework} in which each layer contains useful data.
Technically, a layer is a 2D matrix in which each element is associated with a geo-referenced location, and has a specific value depending on the type of layer.
In this way, we can associate each cube $c \in B$ with such information.
In this work, the framework is composed of the following layers.
\begin{itemize}
\item \textbf{Obstacle Layer}: defines the height of buildings, trees, or other solid obstacles on the terrain;
\item \textbf{No-Fly zone Layer}: defines the areas where drone flight is not allowed or permitted;
\item \textbf{Wireless infrastructure Layer}: defines wireless connectivity in certain areas with respect to the available cellular network infrastructure;
\item \textbf{Risk-map Layer}: defines the ground risk to people in the event that a drone crashes in case of malfunctioning.
\end{itemize}
In the following, we describe the above four layers.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.35]{figures/box_example.pdf}
\caption{Example of the box $B$. A cell is highlighted in green, while obstacles and no-fly zones are highlighted in yellow and red, respectively. Notice that the no-fly zone occupies cells up to the maximum height.}
\label{fig:box}
\end{figure}
\subsection{Obstacle Layer}
Obstacles such as buildings or trees occupy some cubes in $B$, where obviously drones cannot fly.
However, drones are free to fly over obstacles at a sufficient height, and hence this layer is defined at multiple heights.
In general, the higher the drone altitude, the fewer obstacles will be encountered.
The cubes that include obstacles are not considered in $B$, and therefore the number of cubes in $B$ is at most $nmh$.
\subsection{No-Fly Zone Layer}
No Fly-zones are areas where drone flight is forbidden, e.g., close to airports and military warehouses.
We assume that no-fly zones are effective at any height.
As before, the cubes that include no-fly zones are removed from $B$.
\subsection{Wireless infrastructure Layer}
We assume that in the environment $B$ there is already a deployed wireless network infrastructure that comprises a set of \textit{cellular towers}.
These towers provide Internet connectivity (e.g., 4/5G) to ground users.
Specifically, let $T$ be a set of towers with a given discretized position $t = (x_t, y_t, z_t) \in B$.
For simplicity, we assume that each tower $t \in T$ is at ground level, so $z_t = 1$.
We also assume that there is no more than one tower on each cube.
Drones can connect to these antennas if some physical constraints are met at the same time.
We have to distinguish between the \textit{transmitter} device, i.e., the tower, and the \textit{receiver} device, i.e., the drone.
In this paper, we use the Friis transmission equation (see Eq.~\eqref{eq:friis}) in order to determine the quality of the BVLoS\xspace communication link if a drone in cell $c$ establishes a connection to a tower $t$, i.e., by evaluating:
\begin{equation}\label{eq:friis}
P_W(c,t) = P_W(t)\ G(t)\ G(c) \left( \frac{\lambda}{4 \pi d_S} \right)^2,
\end{equation}
where $P_W(c,t)$ is the receiving power at the drone, $P_W(t)$ is the transmitting power of the tower, $G(t)$ and $G(c)$ are the antenna gains of the transmitting and receiving devices, respectively, $\lambda$ is the wavelength representing the effective aperture area of the receiving antenna, and $d_S=\|c-t\|_2$ is the Euclidean \textit{slant distance} that separates the antennas.
Having said that, the wireless infrastructure layer is defined at multiple heights, for multiple reasons.
First, at different altitudes, the distance $d_S$ in Eq.~\eqref{eq:friis} varies, which in turn affects the received power at the drone.
Another aspect to take into account is the probability of being on the Radio Line of Sight (briefly, LoS\xspace) which increases if the height of the drone increases.
Namely, the chances of being LoS\xspace increase with the relative elevation angle between the drone and the tower.
Not only the drone altitude but also the presence of obstacles impact the LoS\xspace.
So, when dealing with obstacles, the chances of being LoS\xspace decrease if the density of obstacles increases.
To wrap up, given two parameters $\alpha$ and $\beta$\footnote{The $\alpha$ and $\beta$ parameters are called here the S-curve parameters, i.e., a modified Sigmoid function.} that model the surrounding environment (e.g., urban, rural) by assuming that the drone is flying at the cell $c=(x_c, y_c, z_c)$, the probability to be LoS\xspace with the tower $t = (x_t, y_t, 1)$ is defined as~\cite{al2014optimal}:
\begin{equation}
\mathcal{P}_{\text{\los}}(c, t) = \frac{1}{1 + \alpha \ e^{-\beta(\arctan{\frac{z_c}{d_G}}-\alpha)}}
\end{equation}
So, the higher the altitude, the longer the distance $d_S$, the less will be the received power $P_W(c,t)$ by Eq.~\eqref{eq:friis}, and the less will be the quality of the BVLoS\xspace link.
But, at a higher altitude, a drone can potentially detect more towers due to its increased elevation.
Finally, given both $c$ and $t$, let
\begin{equation}\label{eq:connectivity}
\mathcal{L}(c, t) = \mathcal{P}_{\text{\los}}(c, t) \ P_W(c,t)
\end{equation}
be the \textit{link-reliability function} that returns the product among the probability of being LoS\xspace, and the normalized received power at the drone.
Potentially, there is an edge between $(c,t)$ as long as $\mathcal{L}(c, t)>0$.
For a given cell $c=(x_c, y_c, z_c) \in B$, let $s(z_c) = 1 + 2 z_c$.
To simplify our multi-layer model, from now on we assume that the probability of LoS\xspace, i.e., $\mathcal{P}_{\text{\los}}(c, t)$, is null or negligible for all towers outside a square of size $s(z_c) \times s(z_c)$.
Thus, when $z_c = 1$, $s(z_c)=3$; while when $z_c = 2$, $s(z_c)=5$.
In general, the square sides depend on the $\alpha$ and $\beta$ parameters, but currently, we omit the dependency.
Formally, let $f_T(c)$ be the subset of towers that the drone can see when it flies inside the cell $c$, i.e., the towers that belong to the square of size $s(z_c)$ centered in $c$, as represented in Figure~\ref{fig:graph}.
Precisely
\begin{align}
f_T(c = (x_c, y_c, z_c)) := \bigcup_{ t \in T} (x_t, y_t, 1),
\end{align}
where $T$ consists of cells $(x_t,y_t,1)$ such that $x_c - z_c \le x_t \le x_c + z_c$ and $y_c - z_c \le y_t \le y_c + z_c$.
Under the above simplification, we assume that a drone in cell $c$ can communicate with any tower $t \in f_T(c)$ with a receiving power $P_W(c,t)$ determined according to Eq.~\eqref{eq:friis}.
For any $t \not \in f_T(c)$, there is no connection between the drone and the tower.
Note that $f_T(c)$ increases with the cell height.
Moreover, we normalize the value $P_W(c,t)$ in the interval $[0,1]$ where $P_W(c,t)=1$ represents the maximum received power at the drone, i.e., when the drone and the tower share the same position $c=t$.
Notice that $0 \le \mathcal{L}(c, t) \le 1$.
\subsection{Risk-map Layer}
The risk-map layer models the potential risks that unexpected events can cause to ground people.
In this paper, we do not consider air risk, i.e., the collision with other flying vehicles (manned or unmanned).
The ground risk-map layer specifically evaluates the risk on the ground depending on a concatenation of different probabilities.
Given a cell $c \in B$, the risk~\cite{primatesta2020ground} is defined as follows:
\begin{equation}\label{eq:risk}
\mathcal{R}(c) = \mathcal{P}_{\text{event}}(c) \ \mathcal{P}_{\text{impact}}(c) \ \mathcal{P}_{\text{fatality}}(c)
\end{equation}
where $\mathcal{P}_{\text{event}}$ is the probability that the drone crashes on the ground; $\mathcal{P}_{\text{impact}}$ is the probability to impact a person when the drone crash on the ground; and $\mathcal{P}_{\text{fatality}}$ is the probability to produce fatal injuries after a person has been impacted.
In the next section, exploiting the presented multi-layer framework, we will create a suitable graph-based data structure used to solve our problem.
\section{The Graph Model}\label{sec:graph}
In this section, our aim is to provide a graph-based data structure that can be exploited by path planning algorithms.
\subsection{Path Dependability}
After we have presented the multi-layer framework, which is at the basis of our BVLoS\xspace flight of drones, and before constructing the graph, we need to introduce some fundamental prerequisites.
The graph we will create has vertices and edges.
The vertices represent ``locations'' or ``locations associated with a tower'', while the edges represent ``movements between adjacent cells'' or ``tower handovers''.
The graph is also weighted, where the weight of each edge represents its dependability.
In this paper, we combine the information from the multi-layer framework that forms the \textit{path dependability}.
More formally, the path dependability jointly takes into account the \textit{ground-safeness}, the \textit{link-reliability}, and the \textit{handover-success-rate}.
For us, the term dependability means the quality of being trustworthy and reliable, and therefore we measure it through a probability, i.e., the larger the probability, the more the path dependability.
So, ground-safeness, link-reliability, and handover-success-rate are individual probabilities to combine in order to obtain the path dependability.
Consider an edge that enters the cell $c'$ associated with the tower $t$ and exits the cell $c$ associated with the tower $t$.
Intuitively, the ground-safeness represents the dependability of the link with respect to the ground and the possible related safety risks.
It is defined as $\mathcal{P}_{\text{GS}}(c',t) = 1-\mathcal{R}(c')$, where $\mathcal{P}_{\text{GS}}(c',t)=0$ represents the highest risk, while $\mathcal{P}_{\text{GS}}(c',t)=1$ does not represent risk.
The link-reliability is simply $\mathcal{P}_{\text{LR}}(c', t) = \mathcal{L}(c', t)$, which combines the probability to be in LoS\xspace, and the normalized received power at the drone.
To summarize, the edge that represents a movement between the two cells has weight $\mathcal{P} = \mathcal{P}_{\text{GS}}(c',t)\ \mathcal{P}_{\text{LR}}(c',t)$, i.e., the product among the ground-safeness, and the link-reliability of the destination cell.
Finally, we assume that the Handover-Success-Rate (HSR) is a value $\mathcal{P}_{\text{HSR}} \in \{\frac{1}{2}, 1\}$, where $\mathcal{P}_{\text{HSR}} = 1$ represents the fact that the drone does not change the tower, continuing communication with the same tower $t$, while $\mathcal{P}_{\text{HSR}}\ = \frac{1}{2}$ represents the scenario when the drone changes the connection with another nearby tower $t'$.
In this latter case, we assume that something can go wrong, and hence the HRS is fixed to $\frac{1}{2}$.
A more complex HSR probability can be utilized in our model, which depends on the $\mathcal{P}_{\text{\los}}$ and $\mathcal{P}_{\text{LR}}$ parameters.
Therefore, the edge that represents a tower handover has weight $\mathcal{P} = \mathcal{P}_{\text{HSR}}$.
During a drone mission, the drone moves from a source to a destination and traverses a path consisting of several edges. The path dependability is given by the product of the weights of the traversed edges.
\subsection{Graph Construction}
Let $G=(V,E)$ be the \textit{weighted directed graph} which is defined by a set $V$ of vertices and a set $E$ of edges, which is built from the box $B$.
Recall that the box $B$ does not include either obstacle or no-fly zones.
We build the graph $G$ as follows.
For each cell $c \in B$, we create a vertex called $v_c$ that represents the cell $c$, plus $|f_T(c)|$ additional vertices to represent the towers to which it is possible to connect from $c$, denoted as $v_c^t$.
So, the set of vertices is $V = \{v_c : c \in B\} \cup \{v_c^t : t \in f_T(c)\ \forall c \in B\}$, while the set of edges $E$ is a bit more complicated.
In general, there are two types of edges: \textit{intra-edges} (i.e., tower handover) and \textit{inter-edges} (i.e., moves between adjacent cells).
An intra-edge is in the form of $(v_c, v_c^t) \in E$, while an extra-edge is in the form of $(v_c^t, v_{c'}^t) \in E$.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.75]{figures/adj.pdf}
\caption{Example of adjacent locations.}
\label{fig:adjacent}
\end{figure}
With regard to the intra-edges, these are used to connect a drone located at a given cell to all the possible visible towers in that cell.
For each $v_c^t \in V$ we add a directed edge $(v_c^t, v_c) \in E$ with cost $\mathcal{P}_{\text{HSR}}=1$, and a directed edge $(v_c, v_c^t) \in E$ with cost $\mathcal{P}_{\text{HSR}}=\frac{1}{2}$ (Algorithm~\ref{alg:Graph Construction}, Line~\ref{code:intra}).
Crossing this latter edge represents the drone changing the connection to a different tower.
Note that we could have used the opposite manner, but the graph construction guarantees that the handover is only paid once.
The pseudocode of the graph construction is reported in Algorithm~\ref{alg:Graph Construction}.
\begin{algorithm}[ht]
\caption{Graph Construction}
\label{alg:Graph Construction}
\DontPrintSemicolon
\For{$c \in B$}{
create vertex $v_c$\;
\For{$t \in f_T(c)$}{
create vertex $v_c^t$\;
\tcp{intra-edges}
add edge $(v_c, v_c^t)$ with cost $\mathcal{P}_{\text{HSR}}=\frac{1}{2}$\;\label{code:intra}
add edge $(v_c^t, v_c)$ with cost $\mathcal{P}_{\text{HSR}}=1$\;
}
}
\For{$c \in B$}{
\For{$c' \in Adj(c)$}{
\For{$t \in f_T(c) \cap f_T(c')$}{
\tcp{inter-edges}
add edge $(v_c^t, v_{c'}^t)$ with cost $\mathcal{P}_{\text{LR}}(c', t) \ \mathcal{P}_{\text{GS}}(c')$\;\label{code:extra}
add edge $(v_{c'}^t, v_c^t)$ with cost $\mathcal{P}_{\text{LR}}(c, t) \ \mathcal{P}_{\text{GS}}(c)$\;
}
}
}
\end{algorithm}
With regard to the inter-edges, these are used when the drone moves between adjacent locations.
Given $c=(x_c, y_c, z_c) \in B$, let $Adj(c)$ be the set of adjacent cells of $c$ where $Adj(c) := \{c'=(x_{c'}, y_{c'}, z_{c'}) \neq c : x_{c'} = x_c \pm \{0, 1\}, y_{c'} = y_c \pm \{0, 1\}, z_{c'} = z_c \pm \{0, 1\}\}$.
So, as illustrated in Figure~\ref{fig:adjacent}, $\max |Adj(c)| = 26$.
For each neighbor of $c \in B$, i.e., $\forall c' \in Adj(c)$, we connect pairs of vertices that represent the same tower $t$ in the two different cells $c$ and $c'$, i.e., the towers in $t \in f_T(c) \cap f_T(c')$.
For each of these pairs, we add a directed edge $(v_c^t, v_{c'}^t) \in E$ with cost $\mathcal{P}_{\text{LR}}(c', t) \ \mathcal{P}_{\text{GS}}(c')$, and a directed edge $(v_{c'}^t, v_c^t) \in E$ with cost $\mathcal{P}_{\text{LR}}(c, t) \ \mathcal{P}_{\text{GS}}(c)$ (Line~\ref{code:extra}).
Crossing these edges represents the drone moving, e.g., from the cell $c$ to $c'$, thus considering the ground-safeness and the link-reliability of the coming cell.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.8]{figures/graph_example.pdf}
\caption{Example of subgraph with $3$ cells, i.e., $c=(2,3)$ (yellow), $c=(3,3)$ (blue), and $c=(3,2)$ (red); and $h=1$.
The grid on top represents the cells along the two dimensions.
Solid edges and dashed edges represent the intra-edges and the inter-edges, respectively.}
\label{fig:graph}
\end{figure}
An example of the graph construction is reported in Figure~\ref{fig:graph}.
The yellow vertex in position $(2,3)$ has $8$ neighbors, i.e., the ones inside the green highlighted square.
Moreover, there are $3$ towers, i.e., $t_a$, $t_b$, and $t_c$.
Among these, only $t_a$ and $t_b$ are in range with position $(2,3)$, and therefore intra-edges from $v_{2,3}$ are only created between $v_{2,3}^a$ and $v_{2,3}^b$.
The blue cell is in position $(3,3)$ which is in range with all the $3$ towers.
As we can see, both $v_{2,3}$ and $v_{3,3}$ share the towers $t_a$ and $t_b$, and hence inter-edges between $v_{2,3}^a$ and $v_{3,3}^a$, and $v_{2,3}^b$ and $v_{3,3}^b$, are created.
The same construction can be repeated for the red cell in $(3,2)$.
Note that by decoupling the tower handover from the movements between adjacent cells, the number of edges decreases.
Recall that at level $1 \le i \le h$, the drone can see at most $|T_i| = (1+2i)^2$ towers and $\max |Adj(c)| = 26$.
Also, let $T_{\max} = \max |T_i|$.
In our construction, instead of preparing an edge ad hoc for each pair $v_c^t=(c,t)$ to $(v_{c'}^{t'}=(c',t')$, we first move from $(c,t)$ to $(c,t')$, and then from $(c,t')$ to $(c',t')$.
Without the decoupling, $(c,t)$ would have had $\mathcal{O}(26 \times T_{\max})$ outgoing edges.
With the proposed decoupling, $(c,t)$ has $\mathcal{O}(T_{\max})$ intra-edges plus $\mathcal{O}(26)$ inter-edges.
Thus, the graph construction has been strongly improved by decoupling the tower handover and the cell movement.
With our assumptions, the box $B$ has $\mathcal{O}(nmh)$ cells and $\mathcal{O}(nmh)$ overall inter-edges.
So, the number $|V|$ of vertices of the graph $G$ is upper-bounded by $|V| \le nmh + \sum_{i=1}^{h}{|T_i| nm}=\mathcal{O}(nmh^3)$.
With regard to the edges, we can upper bound $|E| \le 2\sum_{i=1}^{h}{T_i nm} + 2\sum_{i=1}^{h}{26 T_i nm}$, where the first summation refers to the intra-edges, while the second summation refers to the inter-edges (note the multiplicative factor $2$ that creates the two directed edges).
In conclusion, $|V| \in \mathcal{O}(nmh^3)$, and $|E| \in \mathcal{O}(nmh^3)$.
The constructed graph represents the multi-layer framework and models any drone mission by a path whose weight is its dependability.
\subsection{Problem Formulation and Solution}
Given a box $B$, the multi-layer framework, a starting and a destination cell $s, g \in B$ respectively, and built the weighted dependability graph $G=(V,E)$, we are in position for solving the \problong (\textsc{MPDP}\xspace) whose objective is to find a drone's path $\pi^*$ that starts in $s$ and finishes in $g$ such that the dependability of $\pi = \{(s,\cdot), \ldots, (\cdot,g)\}$ is maximized.
Formally,
\begin{align}
\pi^* &= \argmax_{\pi}{\prod_{e \in \pi} \mathcal{P}(e)} \label{eq:objective}
\intertext{To maximize Eq.~\eqref{eq:objective} is the same as switching to exponentiation and maximizing the argument of the $\log$ operation in Eq.~\eqref{eq:max_log_prod},}
\pi^* &= \argmax_{\pi}{\log \left( \prod_{e \in \pi} \mathcal{P}(e) \right)} \label{eq:max_log_prod} \\
&= \argmax_{\pi}{\sum_{e \in \pi} \log \left( \mathcal{P}(e) \right)} \label{eq:max_sum_log} \\
\intertext{which is the same as minimizing Eq.~\eqref{eq:min_sum_logabs} since all the addends are negative}
&= \argmin_{\pi}{\sum_{e \in \pi} | \log \left( \mathcal{P}(e) \right) |} \label{eq:min_sum_logabs}
\end{align}
Thus, the original objective to find the maximum probability path is equivalent to finding the minimum shortest path where each edge weight (probability) is replaced by the absolute value of its logarithm, and the path cost is the sum of the edge weights.
We remark that the above graph construction enables standard path planning algorithms, such as Dijkstra's algorithm, to retrieve the optimal path for the drone to traverse in the area prioritizing both safety and communication reliability, and thus enabling the BVLoS\xspace operation.
So our problem is polynomially solvable in $|E|$.
Moreover, our model is flexible because by changing the path dependability we can find paths that optimize different criteria.
For example, focusing on the reliability of the communication and posing all $\mathcal{P}_{\text{GS}}=1$, \textsc{MPDP}\xspace find the path that maximizes the connectivity issues.
Similarly, fixing $\mathcal{P}_{\text{LR}}=\mathcal{P}_{\text{HSR}}=1$, \textsc{MPDP}\xspace will find the path with minimum ground risk.
Or, just preserving $\mathcal{P}_{\text{HSR}}$ (and $\mathcal{P}_{\text{LR}}=\mathcal{P}_{\text{GS}}=1$), $\pi^*$ optimizes the number of handovers.
\section{Conclusion and Next Steps}\label{sec:conclusion}
In this paper, we presented a new framework for the operation of drones BVLoS\xspace based on real-world scenarios and challenges.
Our framework takes into account ground risk and obstacles while ensuring a constant communication link between the drone and the operator.
Furthermore, we provide a polynomial graph construction that enables simple path planning algorithms for BVLoS\xspace flight.
There are several directions for future work that could build on our framework for BVLoS\xspace drone operation. One potential area of research is the development of new efficient algorithms for specific sub-problems.
For example, algorithms that can ensure the communication link when the path is already given as input, or algorithms that can handle dynamic scenarios where the ground risk may change during the flight or a communication transfer between towers fails.
Also the exploration of specific real-world scenarios and applications for BVLoS\xspace operation is interesting.
This could include the use of BVLoS\xspace drones for agriculture, emergency response, or package delivery, and involve the development of case studies or pilot projects to demonstrate the feasibility and benefits of BVLoS\xspace operation in these contexts.
Overall, there is significant potential for BVLoS\xspace drone operation to revolutionize a variety of fields and applications, and our framework provides a starting point for further research and development in this area.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\subsection{A Fourier--Mukai/Decomposition correspondence}\label{Sec0.1}
The purpose of this paper is to formulate and explore a correspondence between two geometric structures associated with an integrable system --- the Fourier--Mukai transform \cite{Mukai} and the decomposition theorem \cite{BBD}.
Roughly speaking, a geometric model of an (algebraically completely) integrable system is a Lagrangian fibration $\pi_M: M \to B$ with $M$ a symplectic variety. Motivated by mirror symmetry and the geometric Langlands correspondence, in many interesting cases the fibration $M \to B$ is conjectured to admit a ``dual'' Lagrangian fibration $\check{M} \to B$ which extends the dual abelian scheme associated with the nonsingular fibers of $\pi_M$. Moreover, the two ambient spaces $M$ and $\check{M}$ are expected to share the same derived category of coherent sheaves via a Fourier--Mukai transform
\begin{equation}\label{FM_conj}
\phi_{\mathrm{FM}}: D\mathrm{Coh}(M) \to D\mathrm{Coh}(\check{M})
\end{equation}
extending the classical Fourier--Mukai transform \cite{Mukai} of dual abelian schemes. We are interested in the Fourier--Mukai images of the locally free sheaves $\Omega^k_M$ of K\"ahler differentials on~$M$:
\begin{equation}\label{FM0}
\phi_{\mathrm{FM}}(\Omega_M^k) \in D^b\mathrm{Coh}(\check{M});
\end{equation}
when $M$ is hyper-K\"ahler, $\Omega_M^k$ are natural \emph{hyper-holomorphic bundles} and (\ref{FM0}) are their homological mirrors.\footnote{See Section \ref{HMS} for further discussions on this aspect.} Our proposal is that the objects (\ref{FM0}) are closely related to the decomposition theorem \cite{BBD} for the original integrable system $\pi_M: M \to B$:
\[
R\pi_{M*} {\mathbb{Q}}_M[ \dim M /2] \simeq \bigoplus_{i=0}^{\dim M} P_k [-k], \quad P_k= {^\mathfrak{p}{\mathcal H}}^k( R\pi_* {\mathbb{Q}}_M[\dim M/2] ).
\]
Here $P_k$ can either be viewed as a perverse sheaf or a holonomic ${\mathcal D}_B$-module on $B$.
The rough version of the Fourier--Mukai/Decomposition correspondence is:
\begin{equation}\label{FM/D0}
\phi_{\mathrm{FM}}(\Omega_M^k) \textup{ ``$\simeq$'' } P_k, \quad \textup{for all } k \in {\mathbb{Z}}.
\end{equation}
Clearly (\ref{FM/D0}) does not make sense if we interpret it in a na\"ive way; the left-hand side is a coherent object on the symplectic variety $\check{M}$ while the right-hand side lies in the abelian category of ${\mathcal D}_B$-modules. Ultimately, we are able to modify both sides of (\ref{FM/D0}) to get a mathematically precise formulation; we refer to Conjecture \ref{main_conj} for the statement.
Before diving into more technical details, we briefly summarize some main ingredients and ideas. By Saito's theory of mixed Hodge modules \cite{S1,S2}, the holonomic ${\mathcal D}_B$-module $P_k$ admits a natural good filtration. Taking the associated graded object, we obtain a coherent sheaf $\mathrm{gr}(P_k)$ on the symplectic variety $T^*B$, which can be viewed as the ``classical limit'' of~$P_k$. On the other hand, in the case when $\check{M} \to B$ admits a section, the normal bundle of the section is identified with the cotangent bundle by the symplectic form of $\check{M}$. Then, the modified version of (\ref{FM/D0}) is an isomorphism which holds in the common formal neighborhood $\hat{B}$ of $B$ in both symplectic varieties $\check{M}$ and $T^*B$, between two coherent objects:
\[
\phi_{\mathrm{FM}}(\Omega_M^k)|_{\hat{B}} \quad \mathrm{and} \quad \mathrm{gr}(P_k)|_{\hat{B}}.
\]
In other words, the Hodge module $P_k$ ``quantizes'' the restriction of $\phi_{\mathrm{FM}}(\Omega_M^k)$ to $\hat{B}$.
\subsection{Fourier--Mukai}\label{Sec0.2}
Throughout, we work over the complex numbers ${\mathbb{C}}$. Modifying Definition $2.6$ of \cite{AF}, we say that \[
\pi_M: (M, \sigma) \to B, \quad \dim M = 2n
\]
is an \emph{integrable system} if $(M, \sigma)$ is a nonsingular holomorphic symplectic variety carrying a Lagrangian fibration
\begin{equation*}\label{Lag}
\pi_M: M \to B
\end{equation*}
such that $\pi_M$ is projective, the base $B$ is nonsingular, and it further satisfies that:
\begin{enumerate}
\item[(i)] every geometric fiber of $\pi_M$ is integral, and
\item[(ii)] the fibration $\pi_M$ admits a section $s_B: B \rightarrow M$.\footnote{In the setting of Arinkin--Fedorov, an integrable system is not required to have a section. Here we impose a stronger assumption.}
\end{enumerate}
Hitchin and Beauville--Mukai systems (Example \ref{ex1.1}) provide rich examples of integrable systems. By \cite[Corollary 2.7]{AF}, the identity component of the relative Picard space is a smooth group (algebraic) space $\pi_P: P(:=\mathrm{Pic}^0(M/B)) \to B$ over the base $B$ which admits a universal line bundle
\[
{\mathcal L} \rightarrow M\times_B P.
\]
We may assume that ${\mathcal L}$ is trivialized along the $0$-sections $0_B: B\hookrightarrow P$ and $s_B: B \hookrightarrow M$.
Our main character is Arinkin--Fedorov's (partial) Fourier--Mukai transform \cite{AF}:
\begin{equation}\label{PFM}
\phi_{\mathrm{FM}}: D^b\mathrm{Coh}(M) \rightarrow D^b\mathrm{Coh}(P),\quad {\mathcal E} \mapsto R{q_{P*}}(q_M^* {\mathcal E} \otimes {\mathcal L}^\vee) \otimes \pi_P^*\omega_B^\vee[n]
\end{equation}
where the $q_{(-)}$ are the natural projections from $M \times_BP$. In general, the existence of a dual Lagrangian fibration $\check{M} \to B$ and the Fourier--Mukai equivalence (\ref{FM_conj}) is conjectural and wide open for most cases. If they exist, $\check{M}$ is expected to be a holomorphic symplectic partial compactification of $P \to B$ and (\ref{FM_conj}) should extend (\ref{PFM}). Nevertheless, since we are only interested in the Fourier--Mukai image near the section
\[
B \subset P \,(\subset \check{M})
\]as we explained at the end of Section \ref{Sec0.1}, the rigorously defined partial transform (\ref{PFM}) is enough for our purpose.
Concerning the decomposition side, recall that by \cite{S1,S2}, the perverse sheaves $P_k$ are naturally Hodge modules on $B$, whose associated graded objects induce coherent sheaves
\[
\mathrm{gr}(P_k) \in \mathrm{Coh}(T^*B).
\]
As we will see in Proposition \ref{prop1.4}, a choice of $\pi_M$-relatively ample bundle $\Theta$ on $M$ induces an identification
$\widehat{\kappa}_{\Theta}$ between the formal neighborhoods of $B$ inside $P$ and $T^*B$ respectively, each of which we denote by $\hat{B}$:
\[
\hat{B} \subset P, \quad \hat{B} \subset T^*B.
\]
Using this identification, we formulate the Fourier--Mukai/Decomposition correspondence as follows; see Conjecture \ref{main_conj2} for a more precise formulation.
\begin{conj}[Fourier--Mukai/Decomposition correspondence]\label{main_conj}
For an integrable system $\pi_M: (M, \sigma) \to B$, we have
\begin{equation}\label{main}
\phi_{\mathrm{FM}}(\Omega_M^k)|_{\hat{B}} \simeq \mathrm{gr}(P_k)|_{\hat{B}} \in D^b\mathrm{Coh}(\hat{B}).
\end{equation}
\end{conj}
Conjecture \ref{main_conj} is hence a coherent--constructible correspondence for an integrable system $\pi_M: M \to B$, which connects the coherent object $\Omega_M^k$ on $M$ to the constructible object~$P_k$ on $B$.
\subsection{Lagrangian Cohen--Macaulay sheaves}
Our first result is a consistency check for Conjecture~\ref{main_conj}, in which we check both sides are given by Cohen-Macaulay sheaves and match the reduced (Lagrangian) supports under~(\ref{1223}).
For the Hodge module $P_k$, these are known already as follows.
We denote by
\[
\mathrm{supp}^{\mathrm{red}}( \mathrm{gr}(P_k)) \subset T^*B,
\]
the support of $\mathrm{gr}(P_k)$ endowed with the reduced scheme structure; it
is a conical Lagrangian depending only on the underlying perverse sheaf $P_k$. More precisely, it is the singular support~$\mathrm{SS}(P_k)$ of the perverse sheaf $P_k$, and is described by the \emph{higher discriminants} of the morphism $\pi_M: M \to B$ \cite{disc}. For any $i\geq 0$, Migliorini--Shende introduced a higher discriminant $\Delta^i(\pi_M) \subset B$ determined by the topology of $\pi_M: M \to B$, which satisfies
\[
\mathrm{codim}_{B} \Delta^i(\pi_M) \geq i.
\]
They further proved for any $k$ that
\begin{equation}\label{Lambda}
\mathrm{SS}(P_k) \subset \Lambda: = \bigcup_i \bigcup_{Z_i} \overline{T^*_{Z_i}B}.
\end{equation}
Here $Z_i$ runs through purely $i$-codimensional irreducible components of $\Delta^i(\pi_M)$, and $\overline{T^*_{Z_i}B}$ stands for the closure of the conormal bundle of the nonsingular locus of $Z_i$. Clearly the conical Lagrangian $\Lambda \subset T^*B$ only depends on the topology of $\pi_M: M \to B$.
\begin{thm}[Saito, Migliorini--Shende]\label{gr(P)}
For any $k$ the coherent sheaf $\mathrm{gr}(P_k)$ is a Cohen--Macaulay sheaf of dimension $n$ with
\[
\mathrm{supp}^{\mathrm{red}}(\mathrm{gr}(P_k)) \subset \Lambda.
\]
\end{thm}
The second part of the theorem follows from \cite{disc} as discussed above; the first part is a property of any mixed Hodge module on a nonsingular variety \cite{S1}.
We first establish the counterpart of Theorem \ref{gr(P)} on the Fourier--Mukai side. In Section~\ref{Sec2.4}, we define a closed subset $\Lambda' \subset P$ using the higher discriminants $\Delta^i(\pi_M)$, parallel to $\Lambda \subset T^*B$; by construction, the restriction of $\Lambda'$ to the formal neighborhood of $B$ recovers the conical Lagrangian $\Lambda \subset T^*B$ by Corollary \ref{cor2.4}:
\begin{equation}\label{1223}
\Lambda' |_{\hat{B}} \simeq \Lambda |_{\hat{B}} \subset \hat{B}.
\end{equation}
\begin{thm}\label{thm0.3}
The object
\[
\phi_{\mathrm{FM}}(\Omega_M^k) \in D^b\mathrm{Coh}(P)
\]
is a Cohen--Macaulay sheaf of dimension $n$ concentrated in degree 0 with
\begin{equation*}
\mathrm{supp}^{\mathrm{red}}(\phi_{\mathrm{FM}}(\Omega_M^k))|_{\hat{B}} \subset {\Lambda'}|_{\hat{B}}.
\end{equation*}
\end{thm}
In general the Fourier--Mukai image of a coherent sheaf is a complex concentrated in degrees $[-n,0
]$. Therefore the proof of Theorem \ref{thm0.3} relies on properties of the sheaves~$\Omega^k_M$. As a consequence of Theorem \ref{thm0.3}, Conjecture \ref{main_conj} is reduced to an isomorphism of two Cohen--Macaulay sheaves on $\hat{B}$.
\subsection{Smooth and elliptic fibrations}
As further evidence for Conjecture \ref{main_conj}, we prove it
in the following cases.
First, we verify the conjecture for smooth Lagrangian fibrations.
\begin{thm}\label{thm0.4}
Conjecture \ref{main_conj} holds if $\pi_M$ is smooth.
\end{thm}
Next, we prove the conjecture when $M$ is an elliptic surface over a non-proper curve $B$ which has at worst nodal fibers. This is the first nontrivial case where singular fibers appear.
\begin{thm}\label{thm0.5}
Conjecture \ref{main_conj} holds if $M$ is an elliptic surface over a non-proper curve $B$ with nodal singular fibers.
\end{thm}
The strategy for the proof is to view both sheaves as iterated extensions of certain sheaves and then match the terms as well as the extensions. This matching involves a delicate argument using Deligne's canonical extension.
In Section \ref{sec6.7}, we also sketch a proof of Conjecture \ref{main_conj} for non-proper 2-dimensional $M$ when cuspidal fibers appear. The argument in this setting follows the same strategy as for Theorem \ref{thm0.5} but is even more complicated due to the cuspidal fibers, so we leave the details to the interested reader.
\subsection{Relations to other work and directions}
\subsubsection{The perverse--Hodge symmetry}
The motivation for Conjecture \ref{main_conj} is an effort to understand and categorify the perverse--Hodge symmetry for Lagrangian fibrations \cite{SY}; see also \cite{HLSY, FSY, HM, SY2}.
For a Lagrangian fibration $\pi_M: M \to B$ with $M$ a $2n$-dimensional nonsingular compact irreducible symplectic variety, a perverse--Hodge symmetry was found in \cite{SY}:
\begin{equation}\label{P=F}
\dim H^{i-n}(B, P_{k}) = \dim H^{k,i}(M),
\end{equation}
whose proof relies heavily on the global geometry of compact holomorphic symlectic/hyper-K\"ahler manifolds.
A categorification of (\ref{P=F}) was proposed recently in \cite{SY2} using the natural lift of each $P_k$ to a Hodge module on $B$. By taking graded pieces of the de Rham complex of $P_k$, we obtain a bounded complex of coherent sheaves:
\[
\mathrm{gr}_{i}\mathrm{DR}(P_k) \in D^b\mathrm{Coh}(B).
\]
The main conjecture of \cite{SY2} is:
\begin{equation}\label{P=F2}
\mathrm{gr}_{-i}\mathrm{DR}(P_{k})[n-k] \simeq \mathrm{gr}_{-k}\mathrm{DR}(P_{i})[n-i] \in D^b\mathrm{Coh}(B).
\end{equation}
A mysterious feature of (\ref{P=F2}) is that, unlike (\ref{P=F}), it does not need the compactness assumption of $M$. When $M$ is indeed compact, it was explained in \cite[Section 4]{SY2} that (\ref{P=F2}) recovers (\ref{P=F}).
Our attempt here goes back to the original form of (\ref{P=F}) where we view the right-hand side as cohomology of the vector bundles $\Omega_M^k$. This suggests that there may exist a coherent--constructible correspondence connecting $\Omega^k_M$ and $P_k$ directly, which ideally should hold for not necessarily compact $M$ admitting a Lagrangian fibration. Our conjecture realizes this for integrable systems, as defined in Section \ref{Sec0.2}, using the Fourier--Mukai transform.
The following proposition proves the compatibility between Conjecture \ref{main_conj}, the perverse--Hodge symmetry (\ref{P=F}), and its categorification (\ref{P=F2}) on the base $B$; it gives further evidence for Conjecture \ref{main_conj}.
\begin{prop}\label{prop0.6}
Let $\pi_M: M \to B$ be an integrable system for which (\ref{P=F2}) holds. Then we have
\begin{equation}\label{P=F4}
L0_B^* \,\phi_{\mathrm{FM}}(\Omega_M^k) = L0_B^*\, \mathrm{gr}(P_k) \in D^b\mathrm{Coh}(B),
\end{equation}
where $0_B$ denotes both the closed embeddings of the $0$-sections in $P$ and $T^*B$. In particular, if $B$ is projective, then we have
\begin{equation}\label{P=F3}
H^i(B, L0_B^*\,\phi_{\mathrm{FM}}(\Omega_M^k) ) = H^i(B, L0_B^*\,\mathrm{gr}(P_k)).
\end{equation}
\end{prop}
In fact as we will see in its proof, (\ref{P=F3}) is equivalent to (\ref{P=F}). Hence Conjecture \ref{main_conj} can be viewed as extending the perverse--Hodge symmetry from the $0$-section $B$ to a neighborhood~$\hat{B}$. We note that even in the case where $\pi_M: M \to B$ is smooth, in general the sheaves~$\phi_{\mathrm{FM}}(\Omega_M^k)$ and $\mathrm{gr}(P_k)$ are not scheme-theoretically supported on the $0$-sections. When the integrable system has singular fibers, the formal neighborhood $\hat{B}$ contains much richer geometry than the $0$-section $B$.
One advantage of Conjecture \ref{main_conj} compared to (\ref{P=F2}) and (\ref{P=F4}) is that involves matching Cohen--Macaulay sheaves, instead of matching complexes in the derived category.
It would be interesting to find an enhancement of Conjecture \ref{main_conj} that would imply the full
statement of (\ref{P=F2}).
\subsubsection{The geometric Langlands correspondence}
A more speculative direction for studying the Fourier--Mukai transform of the sheaf of K\"ahler differentials comes via the geometric Langlands correspondence.
For a nonsingular curve $C$ and a reductive group $G$, the geometric Langlands correspondence predicts an equivalence of categories
\begin{equation}\label{GLC}
\mathrm{GLC}: \textup{``}D{\mathrm{QCoh}}\textup{''}({\mathcal M}_{\mathrm{dR},G}) \xrightarrow{\simeq} D(\mathrm{Bun}_{^LG}, {\mathcal D})
\end{equation}
respecting various extra structures. Here ${\mathcal M}_{\mathrm{dR},G}$ stands for the de Rham moduli stack parameterizing $G$-local systems on $C$, $\mathrm{Bun}_{^LG}$ is the moduli stack of principal $^LG$-bundles on $C$, $D(-, {\mathcal D})$ is the derived category of $D$-modules, and the quotation marks means that ``the category of quasi-coherent sheaves'' needs to be modified; we refer to \cite{AG} for the precise statement.
The ``classical limit'' of (\ref{GLC}) is expected to be induced by a Fourier--Mukai transform
\[
\phi_{\mathrm{FM}}: D\mathrm{QCoh}({\mathcal M}_{\mathrm{Higgs},G}) \xrightarrow{\simeq} D\mathrm{QCoh}({\mathcal M}_{\mathrm{Higgs},{^LG}})
\]
for the dual Lagrangian fibrations given by the Hitchin systems associated with $G$ and its Langlands dual $^LG$.\footnote{The Fourier--Mukai $\phi_{\mathrm{FM}}$ here should differ from the Fourier--Mukai we use in this paper by a shift, as we expect that a skyscraper sheaf is sent to a Hecke eigensheaf which is quantized by a holonomic $D$-module. However, for notational convenience we ignore the shift for the discussion here.}
In our setting, if we consider the (derived) exterior power
\[
\Omega^k: = \wedge^k {\mathbb{L}}
\]
of the cotangent complex ${\mathbb{L}}$ on ${\mathcal M}_{\mathrm{Higgs},G}$,
this object can be lifted to the left-hand side of~(\ref{GLC})). Therefore, we might expect to find an object ${\mathcal Q}_k$ in $D(\mathrm{Bun}_{^LG}, {\mathcal D})$ which quantizes $\phi_{\mathrm{FM}}(\Omega^k)$.
\begin{question}
Can we describe the object ${\mathcal Q}_k \in D(\mathrm{Bun}_{^LG}, {\mathcal D})$ quantizing $\phi_{\mathrm{FM}}(\Omega^k)$?
\end{question}
Our proposal provides a strange, partial answer to the question. To avoid stacky issues, we focus on the open subset $A^\circ \subset A$ of the Hitchin base over which both Hitchin fibrations
\[
h_G: {\mathcal M}_{\mathrm{Higgs}, G} \to A, \quad h_{^LG}: {\mathcal M}_{\mathrm{Higgs}, {^LG}} \to A
\]
are proper with integral fibers. Our main conjecture predicts that, the object $\phi_{\mathrm{FM}}(\Omega^k)$ is quantized by a single (shifted) $D$-module $P_k$ obtained from the decomposition theorem associated with \[
h_{^LG}|_{A^\circ}: {\mathcal M}_{\mathrm{Higgs},^LG}|_{A^\circ} \to A^\circ
\]
if we restrict to a formal neighborhood of the Kostant section.\footnote{According to the Ng\^o support theorem \cite{Ngo}, the decomposition theorem associated with the Hitchin system~$h_G$ coincides with that associated with $h_{^LG}$, since they are both governed by intermediate extensions of the local systems given by the smooth fibers.} In particular this suggests that ${\mathcal Q}_k$ is a (shifted) holonomic $D$-module.
\subsubsection{Hyper-K\"ahler mirror symmetry}\label{HMS}
The transform $\phi_{\mathrm{FM}}(\Omega_M^k)$ can also be studied from the perspective of hyper-K\"ahler mirror symmetry and $S$-duality \cite{KW}.
When $M$ is hyper-K\"ahler, homological mirror symmetry is expected to interchange two types of \emph{branes} on $M$ and its mirror $M^\vee$ respectively; it predicts that ``BBB'' branes on~$M$ is sent to ``BAA'' branes on $M^\vee$ via the Fourier--Mukai transform. Roughly, ``A'' stands for the Lagrangian condition and ``B'' stands for the holomorphic condition; ``BBB'' and ``BAA'' indicate the corresponding conditions with respect to the three complex structures. In particular, this suggests that
\[
\phi_{\mathrm{FM}}( \textup{hyper-holomorphic bundle} )
\]
should be a sheaf supported on a complex Lagrangian. The main theme of this paper concerns a particular class of hyper-holomorphic bundles $\Omega_M^k$. Our main proposal gives a local description of the mirror to $\Omega_M^k$ in terms of the decomposition theorem and Hodge modules.
We note that for Hitchin systems, recently Hausel and Hitchin \cite{HH, H_ICM} studied another interesting class of hyper-holomorphic bundles given by the universal family of the Higgs bundles, whose mirrors are supported on the upward flows of certain very stable Higgs bundles.
\subsection{Acknowledgement}
We would like to thank Conan Naichung Leung, Ivan Losev, Tony Pantev, and Christian Schnell for their interest and for very helpful discussions. J.S. was supported by the NSF grant DMS-2134315. Q.Y. was supported by the NSFC grants 11831013 and 11890661.
\section{Symmetries of integrable systems}
\subsection{Overview}
In this section, we introduce two groups $A$ and $P$ associated with an integrable system $\pi_M: M\to B$ which encode its symmetries. The group scheme $A$ acts on $M$ directly, and the group space $P$ induces the Fourier--Mukai transform. They are connected by a relative polarization (\ref{uni}), which further identifies the formal neighborhood of the $0$-sections $B \subset P$ and $B \subset T^*B$.
\subsection{Integrable systems}
We recall from Section \ref{Sec0.2} that an integrable system $\pi_M: M \to B$ is a Lagrangian fibration associated with a nonsingular holomorphic symplectic variety~$(M,\sigma)$ satisfying that the base $B$ is nonsingular, the map $\pi_M$ is projective and surjective, the geometric fibers are integral, and $\pi_M$ admits a section $s_B: B\to M$.
\begin{example}\label{ex1.1}
Compactified Jacobian fibrations associated with integral curves in holomorphic symplectic surfaces provide a large class of interesting examples of integrable systems.
If we take $(S,L)$ to be a polarized abelian or $K3$ surface, we consider the open subset $U \subset |L|$ of the linear system parameterizing integral curves with ${\mathcal C} \to U$ the universal family. Then the compactified Jacobian fibration
\begin{equation}\label{Jacobian}
\pi_J: \overline{J}_{\mathcal C} \to U
\end{equation}
parameterizing rank 1 torsion-free sheaves ${\mathcal W}$ on geometric fibers ${\mathcal C}_u$ of ${\mathcal C} \to U$ with Euler characteristic
\[
\chi({\mathcal C}_u, {\mathcal W}) = \chi ({\mathcal C}_u, {\mathcal O}_{{\mathcal C}_u}), \quad u \in U
\]
gives rise to an integrable system. The section $s_B: U \to \overline{J}_{\mathcal C}$ is given by the trivial line bundles:
\[
u \in U \mapsto [{\mathcal O}_{{\mathcal C}_u}] \in \overline{J}_{\mathcal C}.
\]
The symplectic form of $\overline{J}_{\mathcal C}$ is the restriction of the symplectic form on the regular locus of the moduli of semistable 1-dimensional coherent sheave on $S$ supported on $L$. Such a construction is referred to as the Beauville--Mukai system \cite{Be}.
The same construction replacing $S$ by a the non-compact symplectic surface $T^*C$ given by the total cotangent bundle of a curve $C$ of genus $\geq 2$ then recovers examples of Hitchin systems. \qed
\end{example}
Let $M^{\mathrm{sm}} \subset M$ be the smooth locus of $\pi_M$, \emph{i.e.}, the maximal open subscheme of $M$ such that the restriction of $\pi_M$ is smooth. Then the section
\[
s_B(B) \simeq B \subset M
\]
lies in $M^{\mathrm{sm}}$.
When the integrable system is given by the compactified Jacobian fibration (\ref{Jacobian}) as in Example \ref{ex1.1}, the smooth locus $\overline{J}_C^{\mathrm{sm}}$ is exactly the relative degree 0 Picard scheme
\[
\mathrm{Pic}^0({\mathcal C}/B) \to B
\]
parameterizing degree 0 line bundles on geometric fibers; it is a group scheme over $B$ which acts naturally on $\overline{J}_C$.
This is generalized to any integrable system by Arinkin--Fedorov \cite{AF}.
\begin{prop}[{\cite[Proposition 8.7]{AF}}]\label{prop1.2}
The smooth locus of an integrable system is a group scheme $A:= M^{\mathrm{sm}}$ over $B$. It admits a natural action on $M$
\[
\mu: A \times_B M \to M
\]
preserving the smooth locus.
\end{prop}
\begin{proof}
For an integrable system $\pi_M: M\to B$, Arinkin and Fedorov constructed in \cite[Section~8]{AF} a $B$-group scheme $A$ from the automorphism group scheme $\mathrm{Aut}_B(X)$, so that:
\begin{enumerate}
\item[(i)] $A$ acts naturally on $M$ preserving the smooth locus, and
\item[(ii)] $M^{\mathrm{sm}}$ is an $A$-torsor.
\end{enumerate}
In particular, when $\pi_M$ admits a section $s_B$, the $A$-torsor $M^{\mathrm{sm}}$ is canonically identified with~$A$ via
\[
\mu\circ(\mathrm{id}_A \times s_B): A \rightarrow A\times_B M \to M, \quad \mathrm{Im}( \mu\circ(\mathrm{id}_A \times s_B) ) = M^{\mathrm{sm}} \subset M;
\]
see the proof of \cite[Proposition 8.7]{AF}.
\end{proof}
We conclude this section by recalling Ng\^o's $\delta$-inequality \cite{Ngo, AF}. For every closed point $b \in B$, the group $A_b$ admits the Chevalley decomposition
\[
1 \to R_b \to A_b \to H_b \to 1
\]
with $R_b \subset A_b$ the maximal affine subgroup and $H_b$ an abelian variety. This defines a semicontinuous function
\[
\delta: B \to {\mathbb{N}}, \quad \delta(b): = \dim R_b
\]
which calculates the dimension of the affine part.
The following proposition is a consequence of the existence of a symplectic form.
\begin{prop}[Ng\^o; see {\cite[Proposition 8.9]{AF}}]\label{prop1.3}
For an integrable system $\pi_M: M \to B$ with~$A$ the associated group scheme as in Proposition \ref{prop1.2}, we have
\[
\mathrm{codim}_B\{b\in B\,|\, \delta(A_b) \geq i \} \geq i.
\]
\end{prop}
\subsection{Picard}
We introduce here another group over $B$ closely related to $A$, the Picard space, which will appear in the Fourier--Mukai transform.
The (relative) Picard stack represents the functor
\[
\mathcal{P}ic(M/B): B\textrm{-schemes} \to \mathrm{groupoids}
\]
sending a $B$-scheme $S$ to the groupoid of line bundles over $M \times_BS$. The section $s_B: B \to M$ trivializes the ${\mathbb{G}}_m$-gerbe:
\[
\mathcal{P}ic(M/B) = \mathrm{Pic}(M/B) \times B{\mathbb{G}}_m.
\]
Here $\mathrm{Pic}(M/B)$ is a $B$-group algebraic space whose restriction to a geometric point $b \in B$ recovers the Picard scheme $\mathrm{Pic}(M_s)$. In particular, the algebraic space $\mathrm{Pic}(M/B)$ carries a universal line bundle
\begin{equation}\label{uni}
{\mathcal L} \rightarrow M\times_B \mathrm{Pic}(M/B).
\end{equation}
The identity component $P \subset \mathrm{Pic}(M/B)$ is also a group space over $B$, and we obtain a universal line bundle ${\mathcal L} \to M\times_B P$ by the restriction of (\ref{uni}), which satisfies
\[
{\mathcal L}|_{M\times_B 0_B(B)} \simeq {\mathcal O}_M.
\]
By the choice of splitting, the universal bundle ${\mathcal L}$ on $M\times_BP$ is normalized to satisfy the condition
\[
{\mathcal L}|_{s_B(B)\times_BP} \simeq {\mathcal O}_P.
\]
We will always work with this normalized universal family ${\mathcal L}$.
The two groups $A$ and $P$ over $B$ are closely related. If we choose a $\pi_M$-relatively ample line bundle
\begin{equation}\label{choice}
\Theta \in \mathrm{Pic}(M),
\end{equation}
we obtain a Poincar\'e line bundle
\[
{\mathcal P}_\Theta:= \mu^*\Theta \otimes \mathrm{pr}_M^* \Theta
\]
over $A\times_BM$, which further induces a morphism
\[
\kappa_\Theta: A \rightarrow \mathrm{Pic}(M/B).
\]
By \cite[Corollary 7.7]{AF}, the morphism $\kappa_\Theta$ is \'etale. Since $A = M^{\mathrm{sm}} \subset M$ is irreducible, the image of $\kappa_\Theta$ has to lie in the identity component $P \subset \mathrm{Pic}(M/B)$. We thus obtain the \'etale homomorphism of $B$-groups
\[
\kappa_\Theta: A \to P.
\]
The following proposition serves as the foundation for Conjecture \ref{main_conj} which identifies the formal neighborhoods of $B$ inside $P$ and $T^*B$.
\begin{prop}\label{prop1.4}
The $\pi_M$-relatively ample line bundle $\Theta$ in \eqref{choice} induce an isomorphism
\[
\hat{\kappa}_\Theta: \hat{B}_P \xrightarrow{\simeq} \hat{B}_{T^*B}.
\]
Here $\hat{B}_{(-)}$ stands for the formal neighborhood of $B$ inside $(-)$.
\end{prop}
\begin{proof}
Since the formal neighborhood of the $0$-section in a commutative group scheme splits, \emph{i.e.}, it is isomorphic by the logarithm to the formal completion of the $0$-section in its normal bundle, we have
\[
\hat{B}_{P} = \hat{B}_{N_{B/P}}.
\]
So it suffices to show that $\kappa_\Theta$ (together with $\sigma$) identifies the normal bundle $N_{B/P}$ and the cotangent bundle $T^*B$ over $B$. This is achieved in two steps. First, the \'etale $B$-morphism $\kappa_\Theta$ induces an isomorphism $N_{B/P} \simeq N_{B/A}$. Then the restriction of the symplectic form $\sigma$ on $B$~to
\[
A = M^{\mathrm{sm}} \subset M
\]
induces an isomorphism $N_{B/A} \simeq T^*B$. The composition of these two isomorphism gives the desired one.
\end{proof}
Now we may state our main conjecture precisely using Proposition \ref{prop1.4}.
\begin{conj}\label{main_conj2}
For an integrable system $\pi_M: (M, \sigma) \to B$, let $\Theta$ be any $\pi_M$-relatively ample line bundle (\ref{choice}). Under the isomorphism $\hat{\kappa}_\Theta$ between the formal neighborhoods of $B$ in~$P$ and $T^*B$, there exists an isomorphism
\begin{equation*}
\phi_{\mathrm{FM}}(\Omega_M^k)|_{\hat{B}_P} \simeq \mathrm{gr}(P_k)|_{\hat{B}_{T^*B}} \in \mathrm{Coh}(\hat{B}).
\end{equation*}
\end{conj}
\begin{rmk}
The statements of Conjecture \ref{main_conj2} come in pairs. On one hand, the symplectic form $\sigma$ induces an isomorphism $\Omega_M^k \simeq \Omega_M^{2n - k}$, where $2n = \dim M$. On the other hand, we have $\mathrm{gr}(P_k) \simeq \mathrm{gr}(P_{2n - k})$ by the relative Hard Lefschetz theorem.
\end{rmk}
\section{Singular supports}
\subsection{Overview}
In this section, we review the decomposition theorem of $\pi_M: M \to B$ and describe the reduced support
\[
\mathrm{supp}^{\mathrm{red}}( \mathrm{gr}(P_k) ) \subset T^*B
\]
following Migliorini--Shende \cite{disc} using the group scheme $A$. As this coincides with the singular support of the underlying perverse sheaf, in this section we can view $P_k$ merely as a perverse sheaf; in particular the filtration $F_\bullet P_k$ provided by Saito's theory does not play a role.
We conclude this section by introducing $\Lambda' \subset P$ which serves as the counter-part at the Fourier--Mukai side of the singular support associated with the decomposition theorem. A comparison between $\Lambda$ and $\Lambda'$ is deduced in Corollary \ref{cor2.4}.
\subsection{The decomposition theorem}
As before, we assume that
\[
\dim M = 2 \dim B = 2n.
\]
Since a Lagrangian fibration has equidimesional fibers of dimension $n$ \cite{Equidim}, the decomposition theorem \cite{BBD} yields
\[
R\pi_{M*} {\mathbb{Q}}_M[n] \simeq \bigoplus_{i=0}^{2n} P_k [-k], \quad P_k= {^\mathfrak{p}{\mathcal H}}^k( R\pi_* {\mathbb{Q}}_M[n] ).
\]
Here each perverse sheaf $P_k$ is semisimple.
We note that in the cases of Example \ref{ex1.1} Ng\^o's support theorem \cite{Ngo} implies that all $P_k$ have full support. In particular, each $P_k$ is given by the intermediate extension of the local systems obtained from the smooth locus of the integrable system $\pi_M: M \to B$. More refined information regarding the topology of $\pi_M$ is encoded in the \emph{singular supports} which are certain conical Lagrangians subvarieties of $T^*B$. We describe them in the following section for integrable systems.
\subsection{Singular supports}
For any proper morphism $f: X\to Y$ between nonsingular varieties, Migliorini and Shende \cite{disc} provide a concrete description of the singular supports
\[
\mathrm{SS}(Rf_*{\mathbb{Q}}_X) \subset T^*Y
\]
associated with the decomposition theorem of $f$ using higher discriminants.
Recall that for each $i \geq 0$, the higher discriminant $\Delta^i(f)$ is formed by points $y \in Y$ such that no $(i-1)$-dimensional subspace of the tangent space $T_yY$ at $y$ is transverse to $f$. We obtain a stratification
\[
Y= \Delta^0(f) \supset \Delta^1(f) \supset \Delta^2(f) \supset \dots,
\]
where each $\Delta^i(f) \subset Y$ is closed with
\begin{equation}\label{codim}
\mathrm{codim}_Y(\Delta^i(f)) \geq i.
\end{equation}
The closed subset $\Delta^i(f)$ generalizes the discriminant $\Delta^1(f)$ --- the locus where the fiber is singular.
The following is the main result of \cite{disc}.
\begin{thm}[Migliorini--Shende]\label{MS2.1}
The singular support of $Rf_*{\mathbb{Q}}_X$ is contained in the union of the conormal varieties to $i$-codimensional components of $\Delta^i(f)$ for all $i$.
\end{thm}
In the case of an integrable system $\pi_M: M \to B$, the higher discriminants $\Delta^i(\pi_M)$ are more concretely given by the $\delta$-stratification associated with the group scheme $A$ \cite[Proposition~4.3]{disc}.
\begin{prop}\label{prop2.2}
Let $A$ be the $B$-group scheme associated with an integrable system $\pi_M: M \to B$. We have
\[
\Delta^i(\pi_M) = \{b\in B\,|\,\delta(b) \geq i\}.
\]
\end{prop}
Using Proposition \ref{prop2.2}, we may reformulate Ng\^o's $\delta$-regularity in Proposition \ref{prop1.3} as
\[
\mathrm{codim}_B \Delta^i(\pi_M) \geq i,
\]
which is an immediate consequence of (\ref{codim}).
\subsection{The subvarieties $\Lambda \subset T^*B$ and $\Lambda' \subset P$}\label{Sec2.4}
Recall from (\ref{Lambda}) the conical Lagrangian $\Lambda \subset T^*B$; the ``support part'' of Theorem \ref{gr(P)} follows from Theorem \ref{MS2.1}:
\[
\mathrm{supp}^{\mathrm{red}}(\mathrm{gr}(P_k)) = \mathrm{SS}(P_k) \subset \mathrm{SS}(R{\pi_M}_* {\mathbb{Q}}_M) \subset \Lambda.
\]
Furthermore, in view of Proposition \ref{prop2.2} the higher discriminants $\Delta^i(\pi_M)$ and the conical Lagrangian $\Lambda \subset T^*B$ are governed by the $\delta$-stratification on the base $B$. In the following, we construct a closed subvariety $\Lambda'\subset P$ using also $\Delta^i(\pi_M)$, which serves as
the counter-part of $\Lambda \subset T^*B$ for the Fourier--Mukai side.
For notational convenience, we denote by $G$ a commutative group space over $B$ which is either $A$ or $P$. Since $\kappa_\Theta$ preserves the affine parts, the $\delta$-functions calculating the dimensions of the maximal affine subgroups for $A$ and $P$ coincide.
Let
\[
1 \to R_b \to G_b \to H_b \to 1
\]
be the Chevalley decomposition for the group space $G$ over a closed point $b\in B$. We define the algebraic closed subset $\Delta_G^{\mathrm{aff}} \subset G$ to be the locus $g\in G$ such that $g$ lies in the affine part~$R_{\pi_M(g)}$ over $\pi_M(g) \in B$.
The lemma below follows from Proposition \ref{prop1.3}.
\begin{lem}
Any irreducible component of $\Delta_G^{\mathrm{aff}}$ has dimension $\leq n$.
\end{lem}
We denote by
\[
\Lambda'_G \subset \Delta_G^{\mathrm{aff}}
\]
the union of all $n$-dimensional irreducible components.
We now give a more concrete description of $\Lambda'_G$ in terms of $\Delta^i(\pi_M)$ parallel to that of $\Lambda$ in (\ref{Lambda}). For a locally closed subset $Z \subset B$, there exists an open dense $V\subset Z$ such that the Chevalley decomposition over $V$ is of the form
\[
1\to R_V \to G|_V \to H_V \to 1
\]
with $R_V$ affine and $H_V$ abelian; \emph{c.f.}~the paragraph before \cite[Section 2.3]{dCRS}. We use $\overline{R_Z}$ to denote the Zariski closure of $R_V$ in $G$ for any choice of $V$ as above.
The following proposition can be compared with the definition (\ref{Lambda}) of $\Lambda \subset T^*B$.
\begin{prop}\label{prop2.3}
We have
\[
\Lambda'_G = \bigcup_i \bigcup_{Z_i} \overline{R_{Z_i}} \subset G.
\]
Here $Z_i$ runs through purely $i$-codimensional irreducible components of $\Delta^i(\pi_M)$.
\end{prop}
\begin{proof}
Since specializations preserve affine parts of the groups, we have
\[
\Lambda'_G \supseteq \bigcup_i \bigcup_{Z_i} \overline{R_{Z_i}}.
\]
To prove the other inclusion, we take $W$ to be an irreducible component of $\Lambda_G'$; by definition it is purely of dimension $n$. We consider $Z = \pi_M(W) \subset B$. Assume that a general point $b \in Z$ satisfies $\delta(b) = i$. Hence a general fiber of $W \to Z$ has dimension $i$. This implies that $Z$ has codimension $i$ since
\[
\dim W = n.
\]
On the other hand, we have $Z \subset \Delta^i(\pi_M)$ where the latter has codimension at least $i$ by~(\ref{codim}). Therefore $Z$ is an $i$-dimensional irreducible component of $\Delta^i(\pi_M)$, and the irreducible component $W$ is of the form $\overline{R_Z}$. This proves that $W \subset \Lambda_G'$.
\end{proof}
In the following corollary, we relate $\Lambda \subset T^*B$ to
\[
\Lambda'_A \subset A, \quad \Lambda'_P \subset P.
\]
\begin{cor}\label{cor2.4}
\begin{enumerate}
\item[(i)] For the identification $\hat{B}_A = \hat{B}_{T^*B}$ of formal neighborhoods induced by the symplectic form $\sigma$, we have
\[
\Lambda'_A|_{\hat{B}_A} = \Lambda|_{\hat{B}_{T^*B}}.
\]
\item[(ii)] For any choice of $\Theta$ in (\ref{choice}) which identifies $\hat{B}_P$ and $\hat{B}_{T^*B}$ as in Proposition \ref{prop1.4}, we have
\[
\Lambda_P'|_{\hat{B}_P} = \Lambda|_{\hat{B}_{T^*B}}.
\]
\end{enumerate}
\end{cor}
\begin{proof}
The first part follows from Proposition \ref{prop2.3} and the proof of \cite[Proposition 4.3]{disc}, that
\[
\overline{R_{Z_i}}|_{\hat{B}_A} = \overline{T^*_{Z_i}B}|_{\hat{B}_{T^*B}}.
\]
The second part follows from the fact that the \'etale map between the groups
\[
{\kappa}_\Theta: A \to P
\]
preserves the affine parts.
\end{proof}
From now on we set
\[
\Lambda' : = \Lambda'_P \subset P.
\]
\section{Fourier--Mukai I: the perverse--Hodge symmetry}
\subsection{Overview}
We review Arinkin--Fedorov's Fourier--Mukai transform and deduce some of its basic properties. This recovers the border cases $k=0,2n$ of Conjecture \ref{main_conj2}; see Corollary~\ref{cor3.2}. Then we discuss a reformulation from the Fourier--Mukai viewpoint of Matsushita's theorem \cite{Ma3} concerning the higher direct image of ${\mathcal O}_M$ and its generalization --- the perverse--Hodge symmetry (Conjecture \ref{conj3.4}) \cite{SY,SY2}. We conclude this section by proving Proposition~\ref{prop0.6}.
\subsection{Fourier--Mukai functors} \label{Sec3.2}
We start with a brief review of the classical Fourier--Mukai transform \cite{Mukai}.
Let $A$ be an $n$-dimensional abelian variety with $P = \mathrm{Pic}^0(A)$ its dual. Then a universal family induces a canonical (normalized) Poincar\'e line bundle ${\mathcal P}$ on $A \times P$; it further induces two functors for the bounded derived categories of coherent sheaves:
\[
\Phi_{\mathrm{FM}}: D^b\mathrm{Coh}(P) \to D^b\mathrm{Coh}(A), \quad {\mathcal E} \mapsto R{q_{A*}}(q_P^* {\mathcal E} \otimes {\mathcal P})
\]
and
\[
\phi_{\mathrm{FM}}: D^b\mathrm{Coh}(A) \to D^b\mathrm{Coh}(P),\quad {\mathcal E} \mapsto R{q_{P*}}(q_A^* {\mathcal E} \otimes {\mathcal P}^\vee )[n].
\]
Here $q_A$ and $q_P$ are the natural projections from $A\times P$ to the corresponding factors. Both functors $\Phi_{\mathrm{FM}}$ and $\phi_{\mathrm{FM}}$ are equivalences of categories, and they are inverses of each other:
\[
\phi_{\mathrm{FM}} \circ \Phi_{\mathrm{FM}} \simeq \mathrm{id}_{D^b\mathrm{Coh}(P)}, \quad \Phi_{\mathrm{FM}} \circ \phi_{\mathrm{FM}} \simeq \mathrm{id}_{D^b\mathrm{Coh}(A)}.
\]
Arinkin--Fedorov \cite{AF} generalizes the picture above and their construction works for a large class of degenerate abelian schemes including integrable systems; see also \cite{A1, A2} for the case of compactified Jacobians. Here we focus on the case of the integrable system $\pi_M: M \to B$ with
\[
\dim M = 2 \dim B = 2n.
\]
Recall the relative Picard space $P$ which is smooth over $B$, and the normalized universal line bundle ${\mathcal L}$ over $M \times_B P$. Similarly, we have two functors
\[
\Phi_{\mathrm{FM}}: D\mathrm{QCoh}(P) \to D\mathrm{QCoh}(M), \quad {\mathcal E} \mapsto R{q_{M*}}(q_P^* {\mathcal E} \otimes {\mathcal L})
\]
and
\[
\phi_{\mathrm{FM}}: D\mathrm{QCoh}(M) \to D\mathrm{QCoh}(P),\quad {\mathcal E} \mapsto R{q_{P*}}(q_M^* {\mathcal E} \otimes {\mathcal L}^\vee ) \otimes \pi_P^* \omega_B^\vee[n].
\]
When $\pi_M$ has singular fibers, the above two functors are no longer equivalences. Nevertheless, we still have by \cite{AF} that
\begin{equation} \label{eq:phiPhi}
\phi_{\mathrm{FM}} \circ \Phi_{\mathrm{FM}} \simeq \mathrm{id}_{D\mathrm{QCoh}(P)}: D\mathrm{QCoh}(P) \to D\mathrm{QCoh}(P),
\end{equation}
and in particular $\Phi_{\mathrm{FM}}$ is fully-faithful. We note that since $q_M: M\times_BP \to P$ is proper, the functor $\phi_{\mathrm{FM}}$ preserves bounded coherent complexes.
\begin{prop}\label{prop3.1}
For ${\mathcal K} \in D^b\mathrm{Coh}(B)$, we have
\begin{equation*}\label{eqn3.1}
\Phi_{\mathrm{FM}}(0_{B*}{\mathcal K}) \simeq \pi_M^*{\mathcal K}, \quad \phi_{\mathrm{FM}}(\pi_M^*{\mathcal K}) \simeq 0_{B*} {\mathcal K}.
\end{equation*}
\end{prop}
\begin{proof}
By \eqref{eq:phiPhi} it suffices to prove the first statement. Since ${\mathcal L}$ is trivialized along the $0$-section~$0_B:B \rightarrow P$, we have
\begin{equation}\label{3.4_1}
0_B^* {\mathcal L} \simeq 0_B^* {\mathcal L}^\vee \simeq {\mathcal O}_M.
\end{equation}
We use $0_M: M \rightarrow M\times_BP$ to denote the base change of the $0$-section $0_B: B \rightarrow P$. We have
\begin{align*}
\Phi_{\mathrm{FM}}(0_{B*}{\mathcal K}) & \simeq Rq_{M*}(q_P^*0_{B*}{\mathcal K} \otimes {\mathcal L}) \\
& \simeq Rq_{M*}(0_{M*}\pi_M^*{\mathcal K} \otimes {\mathcal L}) \\
& \simeq Rq_{M*}0_{M*}(\pi_M^*{\mathcal K} \otimes 0_M^*{\mathcal L}) \\
& \simeq \pi_M^*{\mathcal K}.
\end{align*}
Here the second isomorphism is the base change $q_P^*0_{B*} \simeq 0_{M*}\pi_M^*$, the third isomorphism is the projection formula, and the fourth isomorphism follows from $q_M \circ 0_M = \mathrm{id}_M$ and \eqref{3.4_1}.
\end{proof}
Applying Proposition \ref{prop3.1} to ${\mathcal K} = {\mathcal O}_B$, we can verify the border cases of Conjecture \ref{main_conj2}.
\begin{cor}\label{cor3.2}
Conjecture \ref{main_conj2} holds for $k = 0, 2n$.
\end{cor}
\begin{proof}
We have
\[
\phi_{\mathrm{FM}}({\mathcal O}_M) \simeq 0_{B*}{\mathcal O}_B \in \mathrm{Coh}(P), \quad \mathrm{gr}(P_0) \simeq 0_{B*}{\mathcal O}_B \in \mathrm{Coh}(T^*B).
\]
Both are structure sheaves of the $0$-sections; in particular, they are isomorphic in the formal neighborhoods of $B$.
\end{proof}
As in the abelian variety case, there is a (derived) Pontryagin product on $P$. Consider the addition map
\[
m_P : P \times_B P \to P.
\]
We define
\[
\star^R : D\mathrm{QCoh}(P) \times D\mathrm{QCoh}(P) \to D\mathrm{QCoh}(P), \quad ({\mathcal E}_1, {\mathcal E}_2) \mapsto Rm_{P*}(q_1^*{\mathcal E}_1 \otimes^L q_2^*{\mathcal E}_2)
\]
where $q_1$ and $q_2$ are the two projections from $P \times_B P$.
\begin{prop} \label{prop3.3}
For ${\mathcal E}_1, {\mathcal E}_2 \in D\mathrm{QCoh}(P)$ and ${\mathcal K}_1, {\mathcal K}_2 \in D^b\mathrm{Coh}(M)$, we have
\[
\Phi_{\mathrm{FM}}({\mathcal E}_1 \star^R {\mathcal E}_2) \simeq \Phi_{\mathrm{FM}}({\mathcal E}_1) \otimes^L \Phi_{\mathrm{FM}}({\mathcal E}_2), \quad \phi_{\mathrm{FM}}({\mathcal K}_1 \otimes^L {\mathcal K}_2) \simeq \phi_{\mathrm{FM}}({\mathcal K}_1) \star^R \phi_{\mathrm{FM}}({\mathcal K}_2).
\]
\end{prop}
\begin{proof}
The two statements are similar; we only prove the first. By the theorem of the cube applied to $M \times_B P \times_B P$, we have
\begin{equation} \label{eq:cube}
(\mathrm{id} \times m_P)^*{\mathcal L} \simeq q_{12}^*{\mathcal L} \otimes q_{13}^*{\mathcal L}
\end{equation}
where the $q_{ij}$ are the natural projections from $M \times_B P \times_B P$ to the respective factors. Then
\begin{align*}
\Phi_{\mathrm{FM}}({\mathcal E}_1 \star^R {\mathcal E}_2) & \simeq Rq_{M*}(q_P^*Rm_{P*}(q_1^*{\mathcal E}_1 \otimes^L q_2^*{\mathcal E}_2) \otimes {\mathcal L}) \\
& \simeq Rq_{M*} (R(\mathrm{id} \times m_P)_*q_{23}^*(q_1^*{\mathcal E}_1 \otimes^L q_2^*{\mathcal E}_2) \otimes {\mathcal L}) \\
& \simeq Rq_{M*} R(\mathrm{id} \times m_P)_*(q_{12}^*q_P^*{\mathcal E}_1 \otimes^L q_{13}^*q_P^*{\mathcal E}_2 \otimes (\mathrm{id} \times m_P)^*{\mathcal L}) \\
& \simeq Rq_{M*} Rq_{13*}(q_{12}^*q_P^*{\mathcal E}_1 \otimes^L q_{13}^*q_P^*{\mathcal E}_2 \otimes q_{12}^*{\mathcal L} \otimes q_{13}^*{\mathcal L}) \\
& \simeq Rq_{M*} (Rq_{13*}q_{12}^*(q_P^*{\mathcal E}_1 \otimes {\mathcal L}) \otimes^L (q_P^*{\mathcal E}_2 \otimes {\mathcal L})) \\
& \simeq Rq_{M*} (q_M^*Rq_{M*}(q_P^*{\mathcal E}_1 \otimes {\mathcal L}) \otimes^L (q_P^*{\mathcal E}_2 \otimes {\mathcal L})) \\
& \simeq \Phi_{\mathrm{FM}}({\mathcal E}_1) \otimes^L \Phi_{\mathrm{FM}}({\mathcal E}_2).
\end{align*}
Here the second isomorphism is the base change $q_P^*Rm_{P*} \simeq R(\mathrm{id} \times m_P)_*q_{23}^*$, the third isomorphism is the projection formula, the fourth isomorphism follows from $q_M \circ (\mathrm{id} \times m_P) = q_M \circ q_{13}$ and \eqref{eq:cube}, the fifth isomorphism is again the projection formula, the sixth isomorphism is the base change $Rq_{13*}q_{12}^* \simeq q_M^*Rq_{M*}$, and the last isomorphism follows from the projection formula and the definition of $\Phi_{\mathrm{FM}}$.
\end{proof}
\subsection{Pushforward and Fourier--Mukai}
The Fourier--Mukai functors $\Phi_{\mathrm{FM}}$ and $\phi_{\mathrm{FM}}$ are also compatible with the pushforwards
\[
R\pi_{P*}: D\mathrm{QCoh}(P) \to D\mathrm{QCoh}(B), \quad R\pi_{M*}: D^b\mathrm{Coh}(M) \rightarrow D^b\mathrm{Coh}(B).
\]
\begin{prop}\label{prop3.2}
For ${\mathcal E} \in D\mathrm{QCoh}(P)$ and ${\mathcal K} \in D^b\mathrm{Coh}(M)$, we have
\[
Ls_B^*\,\Phi_{\mathrm{FM}} ({\mathcal E}) \simeq R\pi_{P*}{\mathcal E}, \quad L0_B^*\,\phi_{\mathrm{FM}} ({\mathcal K}) \simeq R\pi_{M*}{\mathcal K} \otimes \omega_B^\vee[n].
\]
\end{prop}
\begin{proof}
The two statements are similar; we only prove the second. We have
\begin{align*}
L0_B^*\,\phi_{\mathrm{FM}}({\mathcal K}) & \simeq L0_B^* Rq_{P*} ( q_{M}^*{\mathcal K} \otimes {\mathcal L}^\vee) \otimes \omega_B^\vee [n] \\
& \simeq R\pi_{M*} L0_M^* (q_M^*{\mathcal K}\otimes {\mathcal L}^\vee ) \otimes \omega_B^\vee [n] \\
& \simeq R\pi_{M*} ( {\mathcal K} \otimes 0_M^* {\mathcal L}^\vee) \otimes \omega_B^\vee [n] \\
& \simeq R\pi_{M*} {\mathcal K} \otimes \omega_B^\vee[n].
\end{align*}
Here the second isomorphism is the base change $L0_B^*Rq_{P*} \simeq R\pi_{M*}L0_M^*$, the third isomorphism is given by $q_M \circ 0_M = \mathrm{id}_M$, and the fourth isomorphism follows from (\ref{3.4_1}).
\end{proof}
The following example explains a connection between the Fourier--Mukai transform and a theorem of Matsushita \cite{Ma3}. It can be thought of as a first example where the decomposition theorem and the Fourier--Mukai transform are related.
\begin{example}\label{ex3.3}
Applying Proposition \ref{prop3.2} to ${\mathcal K} = {\mathcal O}_M$, we obtain from Proposition \ref{prop3.1} that
\[
R\pi_{M*}{\mathcal O}_M \simeq L0_B^*{\mathcal O}_B \otimes \omega_B[-n].
\]
The object $L0_B^*{\mathcal O}_B$ is the derived self-tensor of the structure sheaf ${\mathcal O}_B$ of the $0$-section $B \subset P$. Using either the Koszul resolution or the derived self-intersection formula \cite{AC}, we have
\begin{equation} \label{eq:self}
L0_B^* {\mathcal O}_B \simeq \bigoplus_{k=0}^n \wedge^k N^\vee_{B/P}[k] \simeq \bigoplus_{k=0}^n T_B^k[k].
\end{equation}
Therefore
\begin{equation}\label{Mat0}
R\pi_{M*} {\mathcal O}_M \simeq \bigoplus_{k=0}^nT_B^k \otimes \omega_B[k-n] \simeq \bigoplus_{k=0}^n \Omega_B^k[-k].
\end{equation}
The equation (\ref{Mat0}) implies simultaneously that
\begin{enumerate}
\item[(i)] the derived direct image admits a splitting into sheaves
\begin{equation} \label{eq:kollar}
R\pi_{M*} {\mathcal O}_M \simeq \bigoplus_{k=0}^n R^k\pi_{M*} {\mathcal O}_M[-k],
\end{equation}
and
\item[(ii)] each higher direct image is given by
\[
R^k\pi_{M*} {\mathcal O}_M \simeq \Omega_B^k.
\]
\end{enumerate}
In fact, both statements are known to hold for any Lagrangian fibration $\pi_M: M \to B$ beyond the case of integrable systems we consider here. The statement (i) was deduced from Saito's enhancement of the decomposition theorem by passing to graded pieces. The statement (ii) is a theorem of Matsushita \cite{Ma3} which was proven for any Lagrangian fibration; Matsushita's argument relies on the decomposition theorem (i) as well as Hodge-theoretic and birational-geometric techniques.
For an integrable system $\pi_M: M\to B$, the Fourier--Mukai transform provides an alternative argument, yielding both (i) and (ii) from a single calculation
\[
Rq_{P*}{\mathcal L} \simeq 0_{B*}\omega_B[-n],
\]
which is proven in \cite{AF} essentially from Ng\^o's $\delta$-inequality (Proposition \ref{prop1.3} in \cite{AF}). In particular, the Fourier--Mukai approach suggests that (i) and (ii) are both determined by the topology of the integrable system. \qed
\end{example}
\subsection{The perverse--Hodge symmetry}
As illustrated in the example above, the object $R\pi_{M*} {\mathcal O}_M \in D^b\mathrm{Coh}(B)$ is governed by the Fourier--Mukai identity
\begin{equation}\label{FM3.4}
\phi_{\mathrm{FM}}({\mathcal O}_M) \simeq 0_{B*}{\mathcal O}_B.
\end{equation}
In this section we discuss a generalization (Conjecture \ref{conj3.4} below) of Example \ref{ex3.3} concerning the objects
\[
R\pi_{M*} \Omega_M^k \in D^b\mathrm{Coh}(B), \quad k=0,1, \cdots, 2n,
\]
where Conjecture \ref{main_conj2} plays the role of (\ref{FM3.4}). This is a version of the proposal \cite{SY2} for the sheaf-theoretic perverse--Hodge symmetry which simultaneously generalizes Matsushita's theorem~(\ref{Mat0}) and categorifies the cohomological ``perverse = Hodge'' identity (\ref{P=F}).
For our purpose here, we view $P_k$ of the decomposition theorem a Hodge module on the base $B$; in particular, we consider its graded object
\[
\mathrm{gr}(P_k) \in \mathrm{Coh}(T^*B).
\]
\begin{conj}[Categorified perverse--Hodge symmetry]\label{conj3.4}
We have
\[
R\pi_{M*} \Omega_M^k \simeq ( L0_B^*\, \mathrm{gr}(P_k) )^\vee \in D^b\mathrm{Coh}(B).
\]
\end{conj}
\begin{rmk}\label{remark}
\begin{enumerate}
\item[(i)]Conjecture \ref{conj3.4} is expected to hold for not only integrable systems, but any Lagrangian fibration $\pi_M: M \to B$ with $M$ and $B$ nonsingular. As we show in the proof of Proposition \ref{prop0.6} below, Conjecture \ref{conj3.4} follows from Conjecture \ref{P=F2}. In particular, by \cite[Theorem 0.4]{SY2} we have verified Conjecture \ref{conj3.4} for Lagrangian fibrations induced by the Hilbert scheme of points on a surface that admits an elliptic fibration.
\item[(ii)] For an integrable system, Conjecture \ref{conj3.4} can be rewritten as
\[
L0_{B}^*\,\phi_{\mathrm{FM}}(\Omega^k_B) \simeq L0_B^*\, \mathrm{gr}(P_k)
\]
where we applied Proposition \ref{prop3.2} and the fact that the vector bundle $\Omega_M^k$ is self-dual. Consequently, Conjecture \ref{main_conj2} implies Conjecture \ref{conj3.4}.
\item[(iii)] When $B$ is projective, we recover the identity (\ref{P=F}) from Conjecture \ref{conj3.4} by taking cohomology $H^i(B, -)$. This follows from applying Laumon's formula (\emph{c.f.}~\cite[Theorem~2.4]{PS}) to the projective map $B \to \mathrm{pt}$.
\end{enumerate}
\end{rmk}
The following diagram illustrates the role of the Fourier--Mukai transform:
\begin{equation*}
\begin{tikzcd}
\textup{Conj.~\ref{main_conj2}: FM/Decomp.} \arrow[d, snake it, "\textup{Specialize:~} k=0"] \arrow[rr, dashed, "L0_B^*(-)"] & & \textup{Conj.~\ref{conj3.4}: Categorified Perv.~= Hodge} \arrow[d, snake it, "\textup{Specialize:~} k=0"] \\
\textup{FM (\ref{FM3.4}) for } {\mathcal O}_M \arrow[rr,dashed, "L0_B^*(-)"] & & \textup{Matsushita:~} R^i\pi_{M*}{\mathcal O}_M\simeq \Omega_B^i.
\end{tikzcd}
\end{equation*}
Next, we complete the proof of Proposition \ref{prop0.6}.
\begin{proof}[Proof of Proposition \ref{prop0.6}]
We relate both sides of Conjecture \ref{conj3.4} to the perverse--Hodge complexes
\[
{\mathcal G}_{k,i} : = \mathrm{gr}_{-i}\mathrm{DR}(P_{k})[n-k] \in D^b\mathrm{Coh}(B)
\]
of \cite{SY2} and reduce Conjecture \ref{conj3.4} to the symmetry (\ref{P=F2}):
\[
{\mathcal G}_{k,i} \simeq {\mathcal G}_{i,k}.
\]
Recall that if we view $P_k$ as a Hodge module on $B$, it carries the structure as a ${\mathcal D}_B$-module endowed with a good filtration $F_\bullet P_k$. The $i$-th graded piece gives a sheaf of ${\mathcal O}_B$-module $\mathrm{gr}_i P_k$. We consider the de Rham complex
\[
\mathrm{DR}(P_k) = [P_k \to P_k\otimes \Omega_B^1 \to \cdots \to P_k \otimes \Omega_B^n][n];
\]
with the induced filtration
\[
F_i\mathrm{DR}(P_k) = [F_iP_k \to F_{i+1}P_k\otimes \Omega_B^1 \to \cdots \to F_{i+n}P_k\otimes \Omega_B^n][n].
\]
The associated graded object of the de Rham complex induces objects
\begin{equation}\label{dR1}
\mathrm{gr}_i\mathrm{DR}(P_k) = [\mathrm{gr}_iP_k \to \mathrm{gr}_{i+1}P_k\otimes \Omega_B^1 \to \cdots \to \mathrm{gr}_{i+n}P_k\otimes \Omega_B^n][n]
\end{equation}
taking values in the bounded derived category of coherent sheaves on $B$. In particular, the perverse--Hodge complex ${\mathcal G}_{k,i}$ encodes the information of the $k$-th and the $i$-th graded pieces of the perverse and the Hodge filtrations respectively, and thus (\ref{P=F2}) is a perverse--Hodge symmetry.
To relate the perverse--Hodge complexes to the more classical object of the left-hand side of Conjecture~\ref{conj3.4}, we evoke here Saito's formula \cite[2.3.7]{S1} (see also \cite[Section 2.2]{SY2}):
\[
R\pi_{M*}\mathrm{gr}_{-k}\mathrm{DR}({\mathbb{Q}}_M[2n]) \simeq \mathrm{gr}_{-k} \mathrm{DR}(R\pi_{M*} {\mathbb{Q}}_M[2n]) \simeq \bigoplus_{i} \mathrm{gr}_{-k} \mathrm{DR}(P_i)[n-i].
\]
Here ${\mathbb{Q}}_M[2n]$ is the Hodge module associated with the perverse sheaf ${\mathbb{Q}}_M[2n]$ given by the trivial local system, whose de Rham complex has graded piece
\[
\mathrm{gr}_{-k}^F \mathrm{DR}({\mathbb{Q}}_M[2n]) = \Omega_M^k[2n-k].
\]
Therefore, we obtain that
\begin{equation*}
R\pi_{M*} \Omega_M^k \simeq \bigoplus_{i} {\mathcal G}_{i, k} [k-2n];
\end{equation*}
in particular, by the Grothendieck--Verdier duality,
\begin{equation}\label{LHS}
(R\pi_{M*} \Omega_M^k )^\vee \simeq \left(\bigoplus_{i} {\mathcal G}_{i,k} \right) \otimes \pi_M^*\omega_B^\vee [k-n].
\end{equation}
On the other hand, we have
\begin{equation}\label{RHS1}
L0_B^*\, \mathrm{gr}(P_k) \simeq Rp_*( \mathrm{gr}(P_k) \otimes^L_{{\mathcal O}_{T^*B}} {\mathcal O}_B )
\end{equation}
with $p: T^*B \to B$ the natural projection. By taking the Koszul resolution of the structure sheaf ${\mathcal O}_B$ of the $0$-section $B \subset T^*B$, we see from the expression (\ref{dR1}) that the right-hand side of (\ref{RHS1}) is
\[
Rp_*( \mathrm{gr}(P_k) \otimes^L_{{\mathcal O}_{T^*B}} {\mathcal O}_B ) \simeq \left(\bigoplus_{i} {\mathcal G}_{k,i}\right) \otimes \pi_M^* \omega_B^\vee[k-n];
\]
it is matched with (\ref{LHS}) through the perverse--Hodge symmetry ${\mathcal G}_{k,i} \simeq {\mathcal G}_{i,k}$.
So far we have proved that (\ref{P=F2}) implies Conjecture \ref{conj3.4}. The first part of Proposition \ref{prop0.6} then follows from Remark \ref{remark} (ii), and the second part follows from Remark \ref{remark} (iii).
\end{proof}
\section{Fourier--Mukai II: Lagrangian Cohen--Macaulay sheaves}
\subsection{Overview}
In this section, we prove Theorem \ref{thm0.3} which shows that the object
\[
\phi_{\mathrm{FM}}( \Omega_M^k )|_{\hat{B}_P}
\]
is a Cohen--Macaulay sheaf supported on the conical Lagrangian $\Lambda'|_{\hat{B}_P}$. This is matched with the singular support of $P_k$ in view of Theorem \ref{gr(P)} and Corollary \ref{cor2.4}.
Our method is built on the ideas of Arinkin \cite{A1} and Arinkin--Fedorov \cite{AF}.
\subsection{Cohen--Macaulay sheaves}
Let ${\mathcal F}$ be a coherent sheaf on $P$. We say that ${\mathcal F}$ is \emph{Cohen--Macaulay} if for any closed point $x \in P$, we have
\[
\mathrm{depth}_{{\mathcal O}_{P,x}}({\mathcal F}_x) = \dim \mathrm{supp}({\mathcal F}_x).
\]
In particular, ${\mathcal F}$ is Cohen--Macaulay of pure dimension $d$ if and only if its Verdier dual
\[
R{\mathcal H}\mathrm{om}({\mathcal F}, \omega_P) \in D^b\mathrm{Coh}(P)
\]
is concentrated in degree $d$.
The following theorem is a Cohen--Macaulay criterion for the Fourier--Mukai transform of a locally free sheaf.
\begin{thm}\label{thm4.1}
Let ${\mathcal K}$ be a locally free sheaf on $M$. Let $Z \subset P$ be a closed subset for which each irreducible component has dimension $\leq n$. If
\[
\mathrm{supp}^{\mathrm{red}}(\phi_{\mathrm{FM}}({\mathcal K})) \subset Z,
\]
then $\phi_{\mathrm{FM}}({\mathcal K})$ is a Cohen--Macaulay sheaf of pure dimension $n$ concentrated in degree 0.
\end{thm}
\begin{proof}
By definition the object
\[
\phi_{\mathrm{FM}}({\mathcal K}) = R{q_{P*}}(q_M^* {\mathcal K} \otimes {\mathcal L}^\vee) \otimes \pi_P^*\omega_B^\vee [n] \in D^b\mathrm{Coh}(P)
\]
is concentrated in degrees $[-n,0]$. Concerning its Verdier dual, we have
\begin{align*}
R{\mathcal H}\mathrm{om}(\phi_{\mathrm{FM}}({\mathcal K}), \omega_P) & \simeq R{\mathcal H}\mathrm{om}( Rq_{P*}(q_M^*{\mathcal K} \otimes {\mathcal L}^\vee) \otimes \pi_P^*\omega_B^\vee [n], \omega_P ) \\
& \simeq Rq_{P*} R{\mathcal H}\mathrm{om}(q_M^*{\mathcal K} \otimes {\mathcal L}^\vee, \omega_{q_P})\\
& \simeq Rq_{P*} (q_M^*{\mathcal K}^\vee \otimes {\mathcal L} \otimes \omega_{q_P} ).
\end{align*}
Here the second isomorphism follows from Grothendieck--Verdier duality. Since ${\mathcal K}$ is locally free on $M$, we have that $q_M^*{\mathcal K}^\vee \otimes {\mathcal L} \otimes \omega_{q_P}$ is a (locally free) sheaf on $M\times_BP$. Therefore the object
\[
R{\mathcal H}\mathrm{om}(\phi_{\mathrm{FM}}({\mathcal K}), \omega_P) \in D^b \mathrm{Coh}(P)
\]
is concentrated in degrees $[0,n]$.
In conclusion, we know that:
\begin{enumerate}
\item[(i)] $\phi_{\mathrm{FM}}({\mathcal K})$ is concentrated in $[-n,0]$,
\item[(ii)] its Verdier dual $R{\mathcal H}\mathrm{om}(\phi_{\mathrm{FM}}({\mathcal K}), \omega_P)$ is concentrated in $[0,n]$, and
\item[(iii)] both objects have supports of dimension $\leq n$.
\end{enumerate}
The only way for this to happen is that $\phi_{\mathrm{FM}}({\mathcal K})$ is a sheaf concentrated in degree 0 of pure dimension $n$, and its Verdier dual is concentrated in degree $n$ (see \cite[Lemma 7.7]{A2}); in other words, $\phi_{\mathrm{FM}}({\mathcal K})$ is a Cohen--Macaulay sheaf of pure dimension $n$ (concentrated in degree 0).
\end{proof}
\subsection{Proof of Theorem \ref{thm0.3}}
\begin{prop}\label{prop4.2}
Assume that ${\mathcal K}\in D^b\mathrm{Coh}(M)$ underlies an $A$-equivariant bounded complex on $M$. Then we have
\begin{enumerate}
\item[(i)] the support of $\phi_{\mathrm{FM}}({\mathcal K})$ has dimension $\leq n$, and
\item[(ii)] its restriction to $\hat{B} \subset P$ is contained in $\Delta_P^{\mathrm{aff}}$; \emph{i.e.},
\[
\mathrm{supp}^{\mathrm{red}}( \phi_{\mathrm{FM}}({\mathcal K}))|_{\hat{B}} \subset \Delta_P^{\mathrm{aff}}|_{\hat{B}}.
\]
\end{enumerate}
\end{prop}
\begin{proof}
The proof essentially follows from the discussion in \cite[Sections 4 and 5]{AF}; we present it here for the reader's convenience.
Let $l \in P$ be a line bundle on $M$ with
\[
l \in \mathrm{supp}^{\mathrm{red}}(\phi_{\mathrm{FM}}({\mathcal K})) \subset P.
\]
We write $b: = \pi_P(l) \in B$, and denote by $A_b, M_b, P_b$ the restrictions of $A, M, P$ over the closed point $b$ respectively. The (derived) restriction of the $A$-equivariant object ${\mathcal K}$ on $M$ gives an $A_b$-equivariant object ${\mathcal K}_b \in D^b\mathrm{Coh}(M_b)$.\\
\noindent {\bf Step 1.} By base change, we know that
\begin{equation}\label{Step1}
H^j(M_b, {\mathcal K}_b \otimes l_b^\vee) \neq 0
\end{equation}
for some $j \in {\mathbb{Z}}$. The restriction of the line bundle $l_b^\vee$ to $A_b \subset M_b$ corresponds to a ${\mathbb{C}}^*$-torsor over $A_b$. By \cite[Proposition 5.6]{AF} this ${\mathbb{C}}^*$-torsor, denoted by $\widetilde{A}_b$, is actually a commutative group scheme given by the extension
\begin{equation}\label{step1_1}
1 \to {\mathbb{C}}^* \to \widetilde{A}_b \to A_b \to 1.
\end{equation}
It acts on the total space of $l_b^\vee$ such that the subtorus ${\mathbb{C}}^* \subset \widetilde{A}$ acts tautologically by dilations. Since ${\mathcal K}_b$ is $A_b$-equivariant, this gives rise to a natural $\widetilde{A}_b$-action on the vector space (\ref{Step1}).\\
\noindent{\bf Step 2.} Since $\widetilde{A}_b$ is commutative, we may find a 1-dimensional sub-representation
\[
V \subset H^j(M_b, {\mathcal K}_b \otimes l_b^\vee),
\]
which induces a character
\[
\chi: \widetilde{A}_b \rightarrow {\mathbb{C}}^*, \quad \chi|_{{\mathbb{C}}^* \subset \widetilde{A}_b } = \mathrm{id}_{{\mathbb{C}}^*}.
\]
It splits the extension (\ref{step1_1}). In particular, the restriction of $l_b^\vee$ to $A_b \subset M_b$ is a trivial line bundle; equivalently, the restriction of $l_b$ to $A_b$ is trivial.\\
\noindent{\bf Step 3.} Now we focus on the \emph{cohomology support locus}
\[
K_b = \{l \in \mathrm{Pic}^0(M_b)\,|\, l|_{A_b} \simeq {\mathcal O}_{A_b}\} \subset \mathrm{Pic}^0(M_b);
\]
it is a quasi-subgroup in the sense of \cite[Definition 4.11]{AF}, whose identity component $K_b^0 \subset K_b$ is affine by \cite[Corollary 4.6]{AF}. In particular, we have
\[
\dim K_b \leq \delta(b).
\]
By Steps 1 and 2, we know that
\[
\mathrm{supp}^{\mathrm{red}}(\phi_{\mathrm{FM}}({\mathcal K})) \subset \bigcup_{b\in B} K_b
\]
where the latter $K:= \bigcup_{b\in B} K_b$ is a countable union of constructible sets satisfying:
\[
\dim K \leq \dim \Delta_P^{\mathrm{aff}} \leq n, \quad
K|_{\hat{B}} \subset \Delta_P^{\mathrm{aff}}|_{\hat{B}}.
\]
The proposition is concluded.
\end{proof}
We complete in the following the proof of Theorem \ref{thm0.3}. Note that in general the locally free sheaves $\Omega_M^k$ are not $A$-equivariant as illustrated in the example below.
\begin{example}
Let $\pi: M \to B$ be a smooth elliptic fibration with a section, and let $E \subset M$ be a fiber. In this case the group scheme $A$ is identical to $M$ itself. If the tangent bundle tangent bundle $T_M$ is $A$-equivariant, then it is obtained as the pullback of a vector bundle from the base $B$; in particular, the short exact sequence
\[
0 \to T_E \to T_M|_E \to N_{M/S} \to 0
\]
splits. This implies that the boundary map of the long exact sequence
\[
H^0(E, N_{E/M}) \to H^1(E, T_E)
\]
vanishes. On the other hand, this is exactly the Kodaira--Spencer map which is nontrivial when the fibration is not isotrivial.
\end{example}
Therefore, we are not allowed to apply Proposition \ref{prop4.2} to the locally free sheaf ${\mathcal K} = \Omega_M^k$ directly. To solve this issue, we need the following lemma.
\begin{lem}\label{lem4.4}
Let
\[
0 \to {\mathcal E}_1 \to {\mathcal E}_2 \to {\mathcal E}_3 \to 0
\]
be a short exact sequence in $\mathrm{Coh}(M)$. Then associated with the derived exterior power $\wedge^k{\mathcal E}_2 \in D^b\mathrm{Coh}(M)$, we have finitely many objects in $D^b\mathrm{Coh}(M)$
\[
{\mathcal G}_0, \cdots, {\mathcal G}_i,{\mathcal G}_{i+1}, \cdots, {\mathcal G}_k = \wedge^k{\mathcal E}_2
\]
with morphisms ${\mathcal G}_{i-1} \to {\mathcal G}_{i}$ fitting into the exact triangles
\[
{\mathcal G}_{i-1} \to {\mathcal G}_i \to \wedge^{k-i} {\mathcal E}_1 \otimes^L \wedge^i{\mathcal E}_3 \xrightarrow{+1} {\mathcal G}_{i-1}[1].
\]
\end{lem}
\begin{proof}
This follows from the proof of \cite[Corollary 2.2]{Lich} via taking compatible locally free resolutions of the short exact sequence $0\to {\mathcal E}_1 \to {\mathcal E}_2 \to {\mathcal E}_3 \to 0$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm0.3}]
We first note that it suffices to prove
\begin{equation}\label{main1}
\mathrm{supp}^{\mathrm{red}}(\phi_{\mathrm{FM}}(\Omega_M^k)) \subset K
\end{equation}
with $K = \bigcup_{b\in B}K_p \subset P$ given in the proof of Proposition \ref{prop4.2}. In fact, (\ref{main1}) implies that \[
\dim \mathrm{supp}^{\mathrm{red}}(\phi_{\mathrm{FM}}(\Omega_M^k)) \leq n.
\]
Therefore, by Theorem \ref{thm4.1} the object $\phi_{\mathrm{FM}}(\Omega_M^k)$ is a Cohen--Macaulay sheaf of dimension $n$ concentrated in degree 0. Furthermore, since
\[
\phi_{\mathrm{FM}}(\Omega_M^k)|_{\hat{B}} \subset K|_{\hat{B}} \subset \Delta^{\mathrm{aff}}_P|_{\hat{B}},
\]
the Cohen--Macaulay property ensures that the support is contributed by the purely codimension $i$ irreducible components of $\Delta^i(\pi_M)$ (see Proposition \ref{prop2.3}), \emph{i.e.},
\[
\phi_{\mathrm{FM}}(\Omega_M^k)|_{\hat{B}} \subset \Lambda'|_{\hat{B}}.
\]
Now it remains to prove (\ref{main1}). We consider the short exact sequence
\begin{equation*}
0 \to \pi_M^* \Omega_B^1 \to \Omega_M^1 \to \Omega^1_{M/B} \to 0.
\end{equation*}
Here both the first and the third terms are $A$-equivariant coherent sheaves. In view of Lemma~\ref{lem4.4}, the vector bundle $\Omega_M^k = \wedge^k \Omega_M^1$ can be expressed in terms of an iterated extension via exact triangles of objects of the form
\begin{equation}\label{main2}
\pi_M^* \Omega_B^{k-i} \otimes \wedge^i \Omega^1_{M/B} \in D^b\mathrm{Coh}(M),
\end{equation}
where $\wedge^i(-)$ stands for the derived exterior power as in Lemma \ref{lem4.4}. Since each term (\ref{main2}) is $A$-equivariant, its Fourier--Mukai transform has reduced support lying in $K$ by Proposition~\ref{prop4.2}. Hence we obtain that the reduced support of the Fourier--Mukai transform of $\Omega_M^k$ also lies in~$K$. This complete the proof.
\end{proof}
\section{Smooth fibrations}\label{smooth}
\subsection{Overview}
In this section, we verify Conjecture \ref{main_conj2} for smooth Lagrangian fibrations which proves Theorem \ref{thm0.4}. Let $(M, \sigma)$ be a nonsingular holomorphic symplectic variety carrying a \emph{smooth} Lagrangian fibration $\pi_M : M \to B$ with a section $s_B: B \to M$. Let $\pi_P: P \to B$ be the relative Picard scheme. As before, we assume that
\[
\dim M = 2 \dim B = 2n.
\]
\subsection{The $k = 1$ case}\label{Sec5.2}
We first treat the $k = 1$ case; our strategy is the following:
\begin{enumerate}
\item[(i)] We express both $\phi_{\mathrm{FM}}(\Omega^1_M)$ and $\mathrm{gr}(P_1)$ as an extension of sheaves that are (scheme-theoretically) supported on the $0$-sections of $P$ and $T^*B$. The individual terms are pairwise matched via the symplectic form $\sigma$ and the $\pi_M$-relatively ample line bundle~$\Theta$.
\item[(ii)] We match the extensions by means of the Gauss--Manin connection. An important property of smooth Lagrangian fibrations, known as the Donagi--Markman cubic condition (see Lemma \ref{dm}), is crucial in the matching.
\end{enumerate}
Consider the (co)tangent sequence for the smooth morphism $\pi_M : M \to B$. We have a commutative diagram
\begin{equation}\label{eq:cotan}
\begin{tikzcd}
0 \arrow{r}{} & T_{M/B} \arrow{r}{} \arrow{d}{\sigma}[swap]{\simeq} & T_M \arrow{r}{} \arrow{d}{\sigma}[swap]{\simeq} & \pi_M^*T_B \arrow{r}{} \arrow{d}{\sigma}[swap]{\simeq} & 0 \\
0 \arrow{r}{} & \pi_M^*\Omega^1_B \arrow{r}{} & \Omega^1_M \arrow{r}{} & \Omega^1_{M/B} \arrow{r}{} & 0
\end{tikzcd}
\end{equation}
where the vertical arrows are isomorphisms induced by the symplectic form $\sigma$. Since $\Omega^1_{M/B}$ is pulled back from $B$, the natural morphism
\[
\pi_M^*\pi_{M*}\Omega^1_{M/B} \to \Omega^1_{M/B}
\]
is an isomorphism. Then, substituting $\Omega^1_{M/B}$ by $\pi_M^*\pi_{M*}\Omega^1_{M/B}$ in \eqref{eq:cotan} and applying the Fourier--Mukai functor $\phi_{\mathrm{FM}}$ as in Proposition \ref{prop3.1}, we obtain a short exact sequence
\begin{equation} \label{eq:Omega1}
0 \to 0_{B*}\Omega^1_B \to \phi_{\mathrm{FM}}(\Omega^1_M) \to 0_{B*}\pi_{M*}\Omega^1_{M/B} \to 0
\end{equation}
in $\mathrm{Coh}(P)$.
On the other side, the Hodge module $P_1$ corresponds to the variation of Hodge structures on~$R^1\pi_{M*}\mathbb{Q}_M$. We set ${\mathcal V}_1 = R^1\pi_{M*}\mathbb{Q}_M \otimes_{\mathbb{Q}} {\mathcal O}_B$ with grades pieces ${\mathcal V}^{i, 1-i} = \mathrm{gr}^i{\mathcal V}_1$. The associated graded~$\mathrm{gr}(P_1)$ may be viewed either as a coherent sheaf on~$T^*B$ whose reduced support is the $0$-section, or as a Higgs bundle on $B$. The Higgs bundle is an extension of~\mbox{${\mathcal V}^{1, 0} \simeq \pi_{M*}\Omega^1_{M/B}$} by ${\mathcal V}^{0, 1} \simeq R^1\pi_{M*}{\mathcal O}_M$ (both with trivial Higgs field); in terms of coherent sheaves this gives a short exact sequence
\begin{equation} \label{eq:P1}
0 \to 0_{B*}R^1\pi_{M*}{\mathcal O}_M \to \mathrm{gr}(P_1) \to 0_{B*}\pi_{M*}\Omega^1_{M/B} \to 0
\end{equation}
in $\mathrm{Coh}(T^*B)$.
There is a duality between~$R^1\pi_{M*}{\mathcal O}_M$ and~$\pi_{M*}\Omega^1_{M/B}$ depending on a $\pi_M$-relatively ample line bundle $\Theta$. This, together with the isomorphism $T_B \simeq \pi_{M*}\Omega^1_{M/B}$ induced by~$\sigma$, yields an isomorphism
\begin{equation} \label{eq:r1pi}
R^1\pi_{M*}{\mathcal O}_M \simeq \Omega^1_B.
\end{equation}
Comparing the two sequences \eqref{eq:Omega1} and \eqref{eq:P1}, we find that both terms on the left (resp.~right) are (scheme-theoretically) supported on the $0$-sections and match each other. It then remains to match the extensions, which is key to the proof of the $k = 1$ case.
\subsection{Identifying extensions}\label{Sec5.3}
As the terms on the sides of \eqref{eq:Omega1} and \eqref{eq:P1} are supported on the $0$-sections of $P$ and $T^*B$, their extensions only concern the formal neighborhoods of $B$ in~$P$ and $T^*B$. In particular, the isomorphism $\hat{\kappa}_\Theta$ of Proposition \ref{prop1.4} together with \eqref{eq:r1pi} induces an isomorphism of the extension groups
\begin{equation} \label{eq:extgp}
\mathrm{Hom}_{P}(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}\Omega^1_B[1]) \simeq \mathrm{Hom}_{T^*B}(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}R^1\pi_{M*}{\mathcal O}_M[1])
\end{equation}
The identification \eqref{eq:extgp} can be expressed in concrete terms. In fact, we have by adjunction
\begin{align}
\phantom{\simeq{}} & \mathrm{Hom}_P(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}\Omega^1_B[1]) \nonumber \\
\simeq{} & \mathrm{Hom}_B(L0_B^*0_{B*}\pi_{M*}\Omega^1_{M/B}, \Omega^1_B[1]) \nonumber \\
\simeq{} & \mathrm{Hom}_B\left(\bigoplus_{i = 0}^n\wedge^iN_{B/P}^{\vee} \otimes \pi_{M*}\Omega^1_{M/B}[i], \Omega^1_B[1]\right) \nonumber \\
\simeq{} & \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, \Omega^1_B[1]) \oplus \mathrm{Hom}_B(N_{B/P}^{\vee}[1] \otimes \pi_{M*}\Omega^1_{M/B}, \Omega^1_B[1]) \nonumber \\
\simeq{} & \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, \Omega^1_B[1]) \oplus \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, N_{B/P} \otimes \Omega^1_B) \label{eq:split1}
\end{align}
and respectively
\begin{align}
\phantom{\simeq{}} & \mathrm{Hom}_{T^*B}(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}R^1\pi_{M*}{\mathcal O}_M[1]) \nonumber \\
\simeq{} & \mathrm{Hom}_B(L0_B^*0_{B*}\pi_{M*}\Omega^1_{M/B}, R^1\pi_{M*}{\mathcal O}_M[1]) \nonumber \\
\simeq{} & \mathrm{Hom}_B\left(\bigoplus_{i = 0}^n\wedge^iT_B \otimes \pi_{M*}\Omega^1_{M/B}[i], R^1\pi_{M*}{\mathcal O}_M[1]\right) \nonumber \\
\simeq{} & \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, R^1\pi_{M*}{\mathcal O}_M[1]) \oplus \mathrm{Hom}_B(T_B \otimes \pi_{M*}\Omega^1_{M/B}[1], R^1\pi_{M*}{\mathcal O}_M[1]) \nonumber \\
\simeq{} & \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, R^1\pi_{M*}{\mathcal O}_M[1]) \oplus \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, \Omega^1_B \otimes R^1\pi_{M*}{\mathcal O}_M). \label{eq:split2}
\end{align}
Here we have used the derived self-intersection \eqref{eq:self} for $B \hookrightarrow P$ and the Koszul resolution for~$B \hookrightarrow T^*B$. The symplectic form $\sigma$ and the $\pi_M$-relatively ample line bundle $\Theta$ induce an isomorphism of the normal bundles
\[N_{B/P} \simeq \Omega_B^1.\]
This, together with the isomorphism \eqref{eq:r1pi}, identifies the summands of \eqref{eq:split1} and \eqref{eq:split2}.
The decompositions \eqref{eq:split1} and \eqref{eq:split2} are a priori not canonical. Yet by the natural truncations~$\tau_{\geq j}$ of $L0_B^*{\mathcal O}_B$, we have canonical short exact sequences
\begin{multline*}
0 \to \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, \Omega^1_B[1]) \to \mathrm{Hom}_P(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}\Omega^1_B[1]) \\
\to \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, N_{B/P} \otimes \Omega^1_B) \to 0
\end{multline*}
and
\begin{multline*}
0 \to \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, R^1\pi_{M*}{\mathcal O}_M[1]) \to \mathrm{Hom}_{T^*B}(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}R^1\pi_{M*}{\mathcal O}_M[1]) \\
\to \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, \Omega^1_B \otimes R^1\pi_{M*}{\mathcal O}_M) \to 0.
\end{multline*}
Moreover, both sequences admit a canonical splitting by pushing forward via $\pi_P: P \to B$ and~$p: T^*B \to B$, \emph{i.e.},
\[
R\pi_{P*}: \mathrm{Hom}_P(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}\Omega^1_B[1]) \to \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, \Omega^1_B[1]),
\]
and
\[
Rp_*: \mathrm{Hom}_{T^*B}(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}R^1\pi_{M*}{\mathcal O}_M[1]) \to \mathrm{Hom}_B(\pi_{M*}\Omega^1_{M/B}, R^1\pi_{M*}{\mathcal O}_M[1]).
\]
We conclude that \eqref{eq:split1} and \eqref{eq:split2} are canonical decompositions with pairwise identifiable summands compatible with \eqref{eq:extgp}.
Let $\epsilon_{\Omega^1_M}$ and $\epsilon_{P_1}$ denote the respective extension classes of \eqref{eq:Omega1} and \eqref{eq:P1}.
\begin{lem}
We have $R\pi_{P*}\epsilon_{\Omega^1_M} = 0$ and $Rp_*\epsilon_{P_1} = 0$.
\end{lem}
\begin{proof}
The vanishing of $Rp_*\epsilon_{P_1}$ is immediate: this amounts to forgetting the Higgs field of of the Higgs bundle, so that it becomes the trivial extension of $\pi_{M*}\Omega^1_{M/B}$ by $R^1\pi_{M*}{\mathcal O}_M$. For~$R\pi_{P*}\epsilon_{\Omega^1_M}$, we consider the short exact sequence
\begin{equation} \label{eq:pipush}
0 \to \Omega^1_B \to \pi_{P*}\phi_{\mathrm{FM}}(\Omega_M^1) \to \pi_{M*}\Omega^1_{M/B} \to 0
\end{equation}
obtained by applying $R\pi_{P*}$ to \eqref{eq:Omega1}. By Propositions \ref{prop3.1} and \ref{prop3.2}, the sequence \eqref{eq:pipush} is isomorphic to
\begin{equation} \label{eq:zerosec}
0 \to \Omega^1_B \to s_B^*\Omega_M^1 \to \pi_{M*}\Omega_{M/B}^1 \to 0
\end{equation}
obtained by applying $Ls_B^*$ to the cotangent sequence \eqref{eq:cotan}. The sequence \eqref{eq:zerosec} splits because for an abelian variety $A$, the cotangent space $\Omega^1_{A, 0}$ is naturally identified with $H^0(A, \Omega^1_A)$.
\end{proof}
In view of \eqref{eq:split1} and \eqref{eq:split2}, the extension class $\epsilon_{\Omega^1_M}$ is uniquely determined by a morphism
\begin{equation} \label{eq:Omega1'}
\pi_{M*}\Omega^1_{M/B} \to N_{B/P} \otimes \Omega^1_B
\end{equation}
and $\epsilon_{P_1}$ by a morphism
\begin{equation} \label{eq:P1'}
\pi_{M*}\Omega^1_{M/B} \to \Omega^1_B \otimes R^1\pi_{M*}{\mathcal O}_M.
\end{equation}
\begin{prop} \label{prop5.2}
Under the isomorphisms
\[
N_{B/P} \simeq \Omega_B^1, \quad R^1\pi_{M*}{\mathcal O}_M \simeq \Omega^1_B
\]
induced by the symplectic form $\sigma$ and the $\pi_M$-relatively ample line bundle $\Theta$, the two morphisms~\eqref{eq:Omega1'} and \eqref{eq:P1'} coincide. In particular, Conjecture \ref{main_conj2} holds for $\pi_M: M \to B$ for~$k = 1$.
\end{prop}
\begin{proof}
By the construction of the Spencer/Koszul resolution, the morphism \eqref{eq:P1'} is precisely the associated graded of the Gauss--Manin connection
\begin{equation} \label{eq:gm}
\overline{\nabla}: {\mathcal V}^{1, 0} \to \Omega^1_B \otimes {\mathcal V}^{0, 1}.
\end{equation}
To see the relation of \eqref{eq:Omega1'} with the Gauss--Manin connection, we consider the following commutative diagram
\begin{equation*}
\begin{tikzcd}
\mathrm{Hom}_M(\pi_M^*\pi_{M*}\Omega^1_{M/B}, \pi_M^*\Omega_B^1[1]) \arrow{r}{\phi_{\mathrm{FM}}}[swap]{\simeq} \arrow{d}{R\pi_{M*}} & \mathrm{Hom}_P(0_{B*}\pi_{M*}\Omega^1_{M/B}, 0_{B*}\Omega_B^1[1]) \arrow{d}{L0_B^*}\\
\mathrm{Hom}_B(R\pi_{M*}\pi_M^*\pi_{M*}\Omega^1_{M/B}, R\pi_{M*}\pi_M^*\Omega^1_B[1]) \arrow{r}{\otimes \omega_B^\vee[n]}[swap]{\simeq} & \mathrm{Hom}_B(L0_B^*0_{B*}\pi_{M*}\Omega^1_{M/B}, L0_B^*0_{B*}\Omega_B^1[1]).
\end{tikzcd}
\end{equation*}
Here the diagram commutes by Proposition \ref{prop3.2}, and the top row is an isomorphism since $\phi_{\mathrm{FM}}$ is fully-faithful (in fact, an equivalence). Applying the decomposition \eqref{eq:kollar} and the derived self-intersection \eqref{eq:self} to the bottom terms, we find
\begin{multline}\label{eq:dec1}
\mathrm{Hom}_B(R\pi_{M*}\pi_M^*\pi_{M*}\Omega^1_{M/B}, R\pi_{M*}\pi_M^*\Omega^1_B[1]) \\
\simeq{} \mathrm{Hom}_B\left(\bigoplus_{i = 0}^nR^i\pi_{M*}{\mathcal O}_M \otimes \pi_{M*}\Omega^1_{M/B}[-i], \bigoplus_{i = 0}^nR^i\pi_{M*}{\mathcal O}_M \otimes \Omega^1_B[-i + 1]\right)
\end{multline}
and respectively
\begin{multline}\label{eq:dec2}
\mathrm{Hom}_B(L0_B^*0_{B*}\pi_{M*}\Omega^1_{M/B}, L0_B^*0_{B*}\Omega_B^1[1]) \\
\simeq{} \mathrm{Hom}_B\left(\bigoplus_{i = 0}^n\wedge^iN_{B/P}^{\vee} \otimes \pi_{M*}\Omega^1_{M/B}[i], \bigoplus_{i = 0}^n\wedge^iN_{B/P}^{\vee} \otimes \Omega^1_B[i + 1]\right).
\end{multline}
As is usual with decomposition-type theorems, the decompositions \eqref{eq:dec1} and \eqref{eq:dec2} themselves are not canonical. Yet one can associate canonical filtrations via the natural truncations $\tau_{\geq j}, \tau_{\leq j}$ of the arguments of $\mathrm{Hom}(-, -)$. In particular, the left-hand side of \eqref{eq:dec1} admits a canonical quotient
\begin{equation} \label{eq:ko}
\mathrm{Hom}_B(R^{n - 1}\pi_{M*}{\mathcal O}_M \otimes \pi_{M*}\Omega^1_{M/B}[-n + 1], R^n\pi_{M*}{\mathcal O}_M\otimes \Omega^1_B[-n + 1]).
\end{equation}
Similarly, the left-hand side of \eqref{eq:dec2} admits a canonical quotient
\begin{equation} \label{eq:canquot}
\mathrm{Hom}_B(N_{B/P}^\vee \otimes \pi_{M*}\Omega^1_{M/B}[1], \Omega^1_B[1]).
\end{equation}
By definition, the morphism \eqref{eq:Omega1'} is obtained by applying $L0_B^*$ to the extension class $\epsilon_{\Omega^1_M}$, and then projecting to the quotient space \eqref{eq:canquot}. This in turn corresponds (via $\otimes \omega_B^\vee[n]$) to the morphism obtained by applying~$R\pi_{M*}$ to the extension class of the cotangent sequence~\eqref{eq:cotan}, and then projecting to the quotient space \eqref{eq:ko}. In other words, we have a commutative diagram
\begin{equation} \label{eq:gmseq}
\begin{tikzcd}
\pi_{M*}\Omega^1_{M/B} \arrow{r}{} \arrow{d}{\simeq} & N_{B/P} \otimes \Omega_B^1 \arrow{d}{\simeq} \\
\pi_{M*}\Omega^1_{M/B} \arrow{r}{} \arrow{d}{\simeq} & (R^{n - 1}\pi_{M*}{\mathcal O}_M)^\vee \otimes R^n\pi_{M*}{\mathcal O}_M \otimes \Omega^1_B \arrow{d}{\simeq} \\
\pi_{M*}\Omega^1_{M/B} \arrow{r}{} & R^1\pi_{M*}{\mathcal O}_M \otimes \Omega_B^1.
\end{tikzcd}
\end{equation}
Here the first row is the morphism \eqref{eq:Omega1'}, the second row is obtained from the element in \eqref{eq:ko}, and the third row uses the (fiberwise) perfect pairing
\[
R^1\pi_{M*}{\mathcal O}_M \otimes R^{n - 1}\pi_{M*}{\mathcal O}_M \to R^n\pi_{M*}{\mathcal O}_M
\]
for a smooth abelian fibration $\pi_M: M \to B$.
We set ${\mathcal V}_k = R^k\pi_{M*}{\mathbb{Q}}_M \otimes_{\mathbb{Q}} {\mathcal O}_B$ with graded pieces ${\mathcal V}^{i, k - i} = \mathrm{gr}^i{\mathcal V}_k$. The associated graded of the Gauss--Manin connection takes the form
\[
\overline{\nabla}: {\mathcal V}^{i, k - i} \to {\mathcal V}^{i - 1, k - i + 1} \otimes \Omega_B^1
\]
and is linear with respect to the ${\mathcal V}^{0, k - i}$-factor in ${\mathcal V}^{i, k - i} \simeq {\mathcal V}^{0, k - i} \otimes {\mathcal V}^{i, 0}$ by virtue of Griffiths transversality. Finally, by the Katz--Oda description of the Gauss--Manin connection \cite{KO}, the extension class of the cotangent sequence \eqref{eq:cotan} projected to \eqref{eq:ko} is the associated graded
\begin{equation} \label{eq:gm2}
\overline{\nabla}: {\mathcal V}^{1, n - 1} \to {\mathcal V}^{0, n} \otimes \Omega_B^1.
\end{equation}
Since \eqref{eq:gm2} is linear with respect to the ${\mathcal V}^{0, n - 1}$-factor in ${\mathcal V}^{1, n - 1} \simeq {\mathcal V}^{0, n - 1} \otimes {\mathcal V}^{1, 0}$, we find that the bottom row of \eqref{eq:gmseq} is precisely the associated graded
\[
\overline{\nabla}: {\mathcal V}^{1, 0} \to {\mathcal V}^{0, 1} \otimes \Omega^1_B.
\]
Note that this time the factor $\Omega^1_B$ is placed on the right as opposed to \eqref{eq:gm}. Also note that we have yet to use the symplectic form $\sigma$ and the $\pi_M$-relatively ample line bundle $\Theta$. To deal with the side-change we evoke the following result of Donagi--Markman~\cite{DM}; see also \cite[Theorem 4.4]{Voisin} and \cite[Lemma 1.2]{SY2}.
\end{proof}
\begin{lem}[Donagi--Markman cubic condition] \label{dm}
Under the isomorphisms
\[
T_B \simeq \pi_{M*}\Omega^1_{M/B}, \quad R^1\pi_{M*}{\mathcal O}_M \simeq \Omega^1_B
\]
induced by the symplectic form $\sigma$ together with the $\pi_M$-relatively ample line bundle $\Theta$, the associated graded of the Gauss--Manin connection
\[
\overline{\nabla}: \pi_{M*}\Omega^1_{M/B} \to R^1\pi_{M*}{\mathcal O}_M \otimes \Omega^1_B
\]
comes from a cubic form in $H^0(B, \mathrm{Sym}^3\Omega^1_B)$.
\end{lem}
\subsection{General case}
We proceed to higher $k$. We show by exploiting functorial properties of smooth abelian fibrations that the higher $k$ cases follow from $k = 1$.
We first extend the Pontryagin product in Section \ref{Sec3.2} to more general (formal) group spaces $\pi_G : G \to B$. Later we will mainly consider $G = P, T^*B, \hat{B}$. Let
\[
m_G: G \times_B G \to G
\]
be the (formal) addition map. We define
\[
\star^R : D\mathrm{QCoh}(G) \times D\mathrm{QCoh}(G) \to D\mathrm{QCoh}(G), \quad ({\mathcal E}_1, {\mathcal E}_2) \mapsto Rm_{G*}(q_1^*{\mathcal E}_1 \otimes^L q_2^*{\mathcal E}_2)
\]
where $q_1$ and $q_2$ are the two projections from $G \times_B G$. The underived version $\star$ is defined accordingly.
\begin{lem} \label{lem5.4}
Let ${\mathcal E} = \phi_{\mathrm{FM}}(\Omega^1_M) \in \mathrm{Coh}(P)$ (resp.~${\mathcal E} = \mathrm{gr}(P_1) \in \mathrm{Coh}(T^*B)$). For $k \geq 1$, we~have
\[
{\mathcal E}^{\star^R k} \simeq {\mathcal E}^{\star k} \in \mathrm{Coh}(P) \quad \textrm{(resp.~$\in \mathrm{Coh}(T^*B)$)}
\]
and
\[
({\mathcal E}^{\star k})|_{\hat{B}} \simeq ({\mathcal E}|_{\hat{B}})^{\star k} \in \mathrm{Coh}(\hat{B}).
\]
\end{lem}
\begin{proof}
Take ${\mathcal E}_1, {\mathcal E}_2 \in \mathrm{Coh}(P)$ such that $\mathrm{supp}^{\mathrm{red}}({\mathcal E}_i) \subset B$, $i = 1, 2$. Since the reduced support~of
\[
q_1^*{\mathcal E}_1 \otimes^L q_2^*{\mathcal E}_2 \simeq q_1^*{\mathcal E}_1 \otimes q_2^*{\mathcal E}_2 \in \mathrm{Coh}(P \times_B P)
\]
is contained in the $0$-section $B \subset P \times_B P$, we have
\[
{\mathcal E}_1 \star^R {\mathcal E}_2 \simeq Rm_{P*}(q_1^*{\mathcal E}_1 \otimes q_2^*{\mathcal E}_2) \simeq m_{P*}(q_1^*{\mathcal E}_1 \otimes q_2^*{\mathcal E}_2) \simeq {\mathcal E}_1 \star {\mathcal E}_2 \in \mathrm{Coh}(P)
\]
with $\mathrm{supp}^{\mathrm{red}}({\mathcal E}_1 \star {\mathcal E}_2 ) \subset B$. We also have
\begin{align*}
({\mathcal E}_1 \star {\mathcal E}_2)|_{\hat{B}} & \simeq (m_{P*}(q_1^*{\mathcal E}_1 \otimes q_2^*{\mathcal E}_2))|_{\hat{P}} \\
& \simeq m_{\hat{B}*}((q_1^*{\mathcal E}_1 \otimes q_2^*{\mathcal E}_2)|_{\hat{B}}) \\
& \simeq m_{\hat{B}*}(q_1^*({\mathcal E}_1|_{\hat{B}}) \otimes q_2^*({\mathcal E}_2|_{\hat{B}})) \\
& \simeq {\mathcal E}_1|_{\hat{B}} \star {\mathcal E}_2|_{\hat{B}} \in \mathrm{Coh}(\hat{B}).
\end{align*}
Here the second isomorphism uses the base change with respect to $m_P: \mathrm{supp}(q_1^*{\mathcal E}_1 \otimes q_2^*{\mathcal E}_2) \to P$, and the third isomorphism uses the flatness of $|_{\hat{B}}$. The statement for ${\mathcal E} = \phi_{\mathrm{FM}}(\Omega^1_M)$ follows by induction on $k$ and the proof for ${\mathcal E} = \mathrm{gr}(P_1)$ is identical.
\end{proof}
Lemma \ref{lem5.4} together with Propositions \ref{prop3.3} and \ref{prop5.2} implies that
\begin{equation} \label{eq:five}
\phi_{\mathrm{FM}}((\Omega_M^1)^{\otimes k})|_{\hat{B}} \simeq (\phi_{\mathrm{FM}}(\Omega_M^1)^{\star k})|_{\hat{B}} \simeq (\phi_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}})^{\star k} \simeq (\mathrm{gr}(P_1)|_{\hat{B}})^{\star k} \simeq (\mathrm{gr}(P_1)^{\star k})|_{\hat{B}}.
\end{equation}
To understand $\mathrm{gr}(P_1)^{\star k}$, we recall the Higgs interpretation of $\mathrm{gr}(P_1)$. Let $p: T^*B \to B$ be the natural projection. The pushforward $p_*\mathrm{gr}(P_1)$ is a $p_*{\mathcal O}_{T^*B} \simeq \mathrm{Sym}(T_B)$-module, which is precisely given by the vector bundle
\[
\mathrm{gr}({\mathcal V}_1) := {\mathcal V}^{0, 1} \oplus {\mathcal V}^{1, 0}
\]
together with the Higgs field
\[
\overline{\nabla}|_{{\mathcal V}^{0, 1}} = 0, \quad \overline{\nabla}|_{{\mathcal V}^{1, 0}}: {\mathcal V}^{1,0} \to \Omega^1_B \otimes {\mathcal V}^{0, 1}.
\]
Conversely, from the Higgs bundle $(\mathrm{gr}({\mathcal V}_1), \overline{\nabla})$ one also recovers the ${\mathcal O}_{T^*B}$-module
\[
\mathrm{gr}(P_1) \simeq p^{-1} \mathrm{gr}({\mathcal V}_1) \otimes_{p^{-1}p_*{\mathcal O}_{T^*B}} {\mathcal O}_{T^*B}.
\]
\begin{lem} \label{lem5.5}
For $k \geq 1$, we have
\[
p_*(\mathrm{gr}(P_1)^{\star k}) \simeq (\mathrm{gr}({\mathcal V}_1)^{\otimes k}, \overline{\nabla}^{\otimes k}) \in \mathrm{Coh}(p_*{\mathcal O}_{T^*B}).
\]
Here the Higgs field $\overline{\nabla}^{\otimes k}$ on $\mathrm{gr}({\mathcal V}_1)^{\otimes k}$ is defined by the Leibniz rule.
\end{lem}
\begin{proof}
The addition map $m_{T^*B}: T^*B \times_B T^*B \to T^*B$ corresponds to the morphism of sheaves
\begin{equation} \label{eq:hopf}
p_*{\mathcal O}_{T^*B} \to p_*{\mathcal O}_{T^*B} \otimes_{{\mathcal O}_B} p_*{\mathcal O}_{T^*B}.
\end{equation}
In local coordinates ${\mathcal O}_U[\xi_1, \ldots, \xi_n]$ over $U \subset B$, the map \eqref{eq:hopf} sends $\xi_j$ to $\xi_j \otimes 1 + 1 \otimes \xi_j$.
Take ${\mathcal F}_1, {\mathcal F}_2 \in \mathrm{Coh}(T^*B)$ such that $p_*{\mathcal F}_i \simeq ({\mathcal E}_i, \theta_i) \in \mathrm{Coh}(p_*{\mathcal O}_{T^*B})$, $i = 1, 2$. Then the $p_*{\mathcal O}_{T^*B}$-module $p_*({\mathcal F}_1 \star {\mathcal F}_2)$ is given by the tensor product ${\mathcal E}_1 \otimes {\mathcal E}_2$ whose $p_*{\mathcal O}_{T^*B}$-module structure is obtained by composing with \eqref{eq:hopf}, hence the Leibniz rule for the Higgs fields $\theta_1$ and $\theta_2$. The statement of the lemma follows by induction on $k$.
\end{proof}
We are ready to prove Conjecture \ref{main_conj2} for the smooth Lagrangian fibration $\pi_M : M \to B$.
\begin{proof}[Proof of Theorem \ref{thm0.4}]
As in the proof of Proposition \ref{prop5.2} we set ${\mathcal V}_k = R^k\pi_{M*}{\mathbb{Q}}_M \otimes_{\mathbb{Q}} {\mathcal O}_B$ with graded pieces ${\mathcal V}^{i, k - i} = \mathrm{gr}^i{\mathcal V}_k$. The associated graded $\mathrm{gr}(P_k) \in \mathrm{Coh}(T^*B)$ corresponds via $p_*$ to the Higgs bundle
\[
\left(\mathrm{gr}({\mathcal V}_k) := \bigoplus_{i = 0}^k{\mathcal V}^{i, k - i}, \overline{\nabla}\right) \in \mathrm{Coh}(p_*{\mathcal O}_{T^*B})
\]
where
\[
\overline{\nabla}|_{{\mathcal V}^{i, k - i}}: {\mathcal V}^{i, k - i} \to \Omega_B^1 \otimes {\mathcal V}^{i - 1, k - i + 1}
\]
is the associated graded of the Gauss--Manin connection. Now since $\pi_M: M \to B$ is a smooth abelian fibration, we have canonical isomorphisms
\[
{\mathcal V}_k \simeq \wedge^k{\mathcal V}_1, \quad {\mathcal V}^{i, k - i} \simeq {\mathcal V}^{0, k - i} \otimes {\mathcal V}^{i, 0} \simeq \wedge^{k - i}{\mathcal V}^{0, 1} \otimes \wedge^i{\mathcal V}^{1, 0},
\]
and
\begin{equation} \label{eq:wedge}
(\mathrm{gr}({\mathcal V}_k), \overline{\nabla}) \simeq (\wedge^k\mathrm{gr}({\mathcal V}_1), \wedge^k\overline{\nabla})
\end{equation}
for all $k \geq 1$. Here the Higgs field $\wedge^k\overline{\nabla}$ on $\wedge^k\mathrm{gr}({\mathcal V}_1)$ is defined by the Leibniz rule.
By \eqref{eq:five} the restrictions of $\phi_{\mathrm{FM}}((\Omega_M^1)^{\otimes k})$ and $\mathrm{gr}(P_1)^{\star k}$ to the formal neighborhood(s) $\hat{B}$ are isomorphic. Moreover $\mathrm{gr}(P_1)^{\star k}$ corresponds via $p_*$ to the Higgs bundle $(\mathrm{gr}({\mathcal V}_1)^{\otimes k}, \overline{\nabla}^{\otimes k})$ by Lemma~\ref{lem5.5}. It is also clear that both Lemmas \ref{lem5.4} and~\ref{lem5.5} are $\mathfrak{S}_n$-equivariant. Taking antisymmetric tensors and in view of \eqref{eq:wedge}, we conclude that
\[
\phi_{\mathrm{FM}}(\Omega_M^k)|_{\hat{B}} \simeq \mathrm{gr}(P_k)|_{\hat{B}}. \qedhere
\]
\end{proof}
\section{Two-dimensional cases}
\subsection{Overview} \label{sec:6.1}
In this section, we prove Theorem \ref{thm0.5}. By our assumption, the morphism
\[
\pi_M: M \to B
\]
is an elliptic fibration over a non-proper curve $B$ with integral fibers and a section $s_B: B \to M$. Moreover, the closed fibers of $\pi_M$ are either a nonsingular elliptic curve, or a nodal rational curve. Assume that $p_1, \cdots, p_m$ are the closed points on $B$ such that the fiber $F_i = \pi_M^{-1}(p_i)$ is nodal, and the restriction
\[
\pi^\circ:= \pi_{M^\circ}: M^\circ \to B^\circ:= B \setminus \{p_1,\dots,p_m\}, \quad M^\circ:= \pi^{-1}(B^\circ)
\]
is a smooth elliptic fibration. Denote by $j: B^\circ \hookrightarrow B$ the natural open embedding, and denote by ${\mathcal V}$ the variation of Hodge structures given by $R^1\pi^\circ_* {\mathbb{Q}}_M$ over $B^\circ$ as in Section \ref{smooth}. Then a direct calculation (\emph{e.g.}~Proof of \cite[Proposition 4.17]{Z}) yields
\[
P_1 = {j_{!*}}{\mathcal V},
\]
where we use the same notation ${\mathcal V}$ to denote the (pure) Hodge module given by the flat bundle $({\mathcal V}, \nabla)$.
In view of Corollary \ref{cor3.2}, Theorem \ref{thm0.5} is reduced to the following theorem.
\begin{thm}\label{thm6.1}
Under the isomorphism $\hat{\kappa}_\Theta$ of Proposition \ref{prop1.4} identifying the formal neighborhoods of $B$ in $P$ and $T^*B$, we have
\[
\phi_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}} \simeq \mathrm{gr}({j_{!*}}{\mathcal V})|_{\hat{B}} \in \mathrm{Coh}(\hat{B}).
\]
\end{thm}
\subsection{Gauss--Manin with poles}
We choose as in Section \ref{Sec5.2} a symplectic form $\sigma$ on $M$ as well as a $\pi_M$-relative ample bundle $\Theta$. They induce isomorphisms
\begin{equation}\label{isom1}
R^1\pi_{M*} {\mathcal O}_M \simeq \Omega^1_B, \quad N_{B/P} \simeq \Omega^1_B
\end{equation}
where the first isomorphism is given by \cite{Ma3}. We denote by $D$ the effective divisor $\sum_{i=1}^m p_i \subset B$. The divisor
\[
{F} : = \sum_{i=1}^m F_i \subset M
\]
given by the pullback of $D$ is normal crossing on $M$. The symplectic form $\sigma$ also induces an isomorphism
\begin{equation}\label{6.2_1}
\Omega^1_{M/B}(\log {F}) \simeq \Omega_B^{1\vee}.
\end{equation}
By the discussion of Section \ref{Sec5.3}, if we restrict over $B^\circ$, the fixed isomorphisms (\ref{isom1}) and (\ref{6.2_1}) and the Donagi--Markman cubic form induce an isomorphism
\[
\phi_{\mathrm{FM}}(\Omega_M^1) |_{\hat{B}^\circ} \xrightarrow{\simeq} \mathrm{gr}({\mathcal V})|_{\hat{B}^\circ}.
\]
Note that the symmetry of the Donagi--Markman cubic form is automatic in this case since the base $B$ is 1-dimensional.
Now in order to extend the isomorphism above over $D$, we first need to extend the associated graded of the Gauss--Manin connection, as well as the Donagi--Markman cubic form. We consider the logarithmic cotangent sequence on $M$:
\begin{equation}\label{6.2_0}
0 \to \pi_M^*\Omega_B^1(D) \to \Omega^1_M(\log {F}) \to \Omega^1_{M/B}(\log {F}) \to 0.
\end{equation}
After pushing it forward to $B$, we have the connecting map
\[
\pi_{M*} \Omega^1_{M/B}(\log {F}) \to R^1\pi_{M*} {\mathcal O}_M \otimes \Omega^1_B(D).
\]
We review briefly the Hodge theoretic interpretation of this map via Deligne's canonical extension of ${\mathcal V}$; see \cite[Section 2]{K}. Recall that the canonical extension depends on a real interval $[a, a+1)$ or $(a, a + 1]$ where the eigenvalues of the residue endomorphism should lie. In our situation, the monodromy around each point of $D$ is unipotent (given by the matrix $(\begin{smallmatrix} 1&1\\0&1 \end{smallmatrix})$ in local coordinates), so the eigenvalues are necessarily integers. Let $\overline{{\mathcal V}}$ be the canonical extension of ${\mathcal V}$ with respect to either $[0, 1)$ or~$(-1, 0]$; it is locally free of rank $2$ on $B$. Schmid's theorem says that~$F_\bullet \overline{{\mathcal V}} := j_*F_\bullet {\mathcal V} \cap \overline{{\mathcal V}}$ is a filtration by locally free subsheaves; here we write the Hodge filtration as an increasing filtration $F_i{\mathcal V}: = F^i{\mathcal V}$ to be compatible with the convention of Hodge modules,
\[
(\overline{{\mathcal V}}, F^\bullet, \nabla), \quad \nabla: \overline{{\mathcal V}} \to \overline{{\mathcal V}}\otimes \Omega^1_B(D),\quad F_{-1}\overline{{\mathcal V}} \subset F_{0}\overline{{\mathcal V}} = \overline{{\mathcal V}}.
\]
Then by the logarithmic version of the Katz--Oda theorem \cite{Katz}, the connecting map above associated with (\ref{6.2_0}) recovers the associated graded of the meromorphic Gauss--Manin connection $\overline{\nabla}: \mathrm{gr}_{-1}\overline{{\mathcal V}} \to \mathrm{gr}_{0}\overline{{\mathcal V}} \otimes \Omega^1_B(D)$:
\begin{equation}\label{6.2_2}
\overline{\nabla}: \pi_{M*} \Omega^1_{M/B}(\log {F}) \to R^1\pi_{M*} {\mathcal O}_M \otimes \Omega^1_B(D).
\end{equation}
Further using the isomorphisms (\ref{isom1}) and (\ref{6.2_1}), we obtain that (\ref{6.2_2}) comes from a section
\begin{equation}\label{cubic2}
[\overline{\nabla}] \in H^0(B, (\Omega^1_B)^{\otimes 3}(D)).
\end{equation}
This is indeed the meromorphic extension of the Donagi--Markman cubic form in Lemma \ref{dm}.
\subsection{Admissible sheaves}
Our strategy is similar to the proof of Theorem \ref{thm0.4}. We express both sheaves
\[
\phi_{\mathrm{FM}}(\Omega^1_M)|_{\hat{B}} \quad \textup{and} \quad \mathrm{gr}(j_{!*}{\mathcal V})|_{\hat{B}}
\]
in terms certain ``building blocks'' supported scheme-theoretically on either the 0-section \mbox{$B \subset \hat{B}$} or a closed fiber $\hat{F}_i : = F_i|_{\hat{B}}$, and then we match their extension classes. Since the existence of the singular fibers $F_i$ further complicates the extensions, we introduce the notion of \emph{admissible sheaves} to treat such complexity.
For notational convenience, we will uniformly use $F_i$ to denote the fiber over $p_i$ for either $\pi_M: M \to B$ or the projection $T^*B \to B$. Therefore $F_i$ is either a nodal rational curve or the affine line ${\mathbb{C}}$.
We say that an object
\[
{\mathcal A} \in \mathrm{Coh}(T^*B) \quad (\textrm{resp.~}{\mathcal A} \in \mathrm{Coh}(\hat{B}))
\]
is an \emph{admissible sheaf}, if ${\mathcal A}$ admits an increasing filtration of coherent subsheaves
\[
{\mathcal W}^{\mathcal A}_{-1} \subset {\mathcal W}^{{\mathcal A}}_0 \subset {\mathcal A}
\]
satisfying that
\begin{enumerate}
\item[(a)] ${\mathcal W}^{\mathcal A}_{-1} \simeq \oplus_{i=1}^m {\mathcal O}_{F_i}$ (resp.~${\mathcal W}^{\mathcal A}_{-1} \simeq \oplus_{i=1}^m {\mathcal O}_{\hat{F}_i}$),
\item[(b)] ${\mathcal W}^{\mathcal A}_0/ {\mathcal W}^{\mathcal A}_{-1} \simeq 0_{B*}\Omega^1_B(D)$, and
\item[(c)] ${\mathcal A}/{\mathcal W}^{\mathcal A}_0 \simeq 0_{B*}\Omega_B^{1\vee}$.
\end{enumerate}
By (a, b), the subsheaf ${\mathcal W}_0^{\mathcal A}$ fits into a short exact sequence\footnote{For notational convenience, we only describe here the case where ${\mathcal A} \in \mathrm{Coh}(T^*B)$; the case for $\hat{B}$ is completely parallel.}
\begin{equation}\label{eqn56}
0 \to \bigoplus_{i=1}^m {\mathcal O}_{F_i} \to {\mathcal W}_0^{\mathcal A} \to 0_{B*} \Omega^1_B(D) \to 0,
\end{equation}
whose extension class induces
\begin{equation}\label{eqn57}
[{\mathcal W}_0^{{\mathcal A}}] \in \mathrm{Ext}_{T^*B}^1\left(0_{B*}\Omega^1_B(D), \bigoplus_{i=1}^m {\mathcal O}_{F_i}\right) = \bigoplus_{i=1}^m \mathrm{Ext}^1_{T^*B}(0_{B*} \Omega^1_B(D), {\mathcal O}_{F_i}).
\end{equation}
By adjunction, each extension group $\mathrm{Ext}^1_{T^*B}(0_{B*} \Omega^1_B(D), {\mathcal O}_{F_i})$ on the right-hand side of (\ref{eqn57}) is 1-dimensional:
\[
\mathrm{Ext}^1_{T^*B}(0_{B*} \Omega^1_B(D), {\mathcal O}_{F_i}) = \mathrm{Ext}^1_{F_i}({\mathbb{C}}_{p_i}, {\mathcal O}_{F_i})\simeq {\mathbb{C}}.
\]
Here $p_i$ is viewed as a point on $F_i$ lying in the intersection with the 0-section $B \subset T^*B$. We say that the admissible sheaf ${\mathcal A}$ is \emph{good} if each summand of the extension class $[{\mathcal W}_0^{\mathcal A}]$ in~$\mathrm{Ext}^1_{T^*B}(0_{B*} \Omega^1_B(D), {\mathcal O}_{F_i})$ is nonzero. In this case, up to scaling ${\mathcal O}_{F_i}$ we may express~${\mathcal W}^{{\mathcal A}}_0$ as an extension (\ref{eqn56}) whose extension class (\ref{eqn57}) is of the form
\[
(1,1, \dots, 1) \in \bigoplus_{i=1}^m \mathrm{Ext}^1_{T^*B}(0_{B*} \Omega^1_B(D), {\mathcal O}_{F_i}).
\]
Now assume that ${\mathcal A}$ is a good admissible sheaf. The condition (c) further implies that ${\mathcal A}$ fits into an extension
\[
0 \to {\mathcal W}^{\mathcal A}_0 \to {\mathcal A} \to 0_{B*} \Omega_B^{1\vee} \to 0
\]
which yields a class
\[
[{\mathcal A}] \in \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, {\mathcal W}_0^{\mathcal A}) \xrightarrow{\rho_{\mathcal A}} \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D)).
\]
Here the map $\rho_{\mathcal A}$ is induced by (\ref{eqn56}).
The following proposition provides a criterion for two good admissible sheaves to be isomorphic. This is also the only place where we need the assumption that $B$ is not proper.
\begin{prop}\label{prop6.2}
Assume $B$ non-proper. Let ${\mathcal A}$ and ${\mathcal A}'$ be two good admissible sheaves such that the classes $\rho_{{\mathcal A}}([{\mathcal A}])$ and $\rho_{{\mathcal A}'}([{\mathcal A}'])$ coincide:
\[
\rho_{\mathcal A}([{\mathcal A}]) = \rho_{{\mathcal A}'}([{\mathcal A}'])\in \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D)).
\]
Then we have
\[{\mathcal A} \simeq {\mathcal A}'.
\]
\end{prop}
\begin{proof}
Applying $\mathrm{Hom}_{T^*B}(0_{B*}\Omega^{1\vee}_B, - )$ to (\ref{eqn56}), we obtain the long exact sequence:
\begin{align*}
\cdots \rightarrow \mathrm{Hom}_{T^*B}(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D)) \xrightarrow{\textup{(i)}} \bigoplus_{i=1}^m \mathrm{Ext}_{T^*B}^1(0_{B*}\Omega^{1\vee}_B, {\mathcal O}_{F_i}) & \\ \xrightarrow{\textup{(ii)}} \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, {\mathcal W}_0^{\mathcal A}) \xrightarrow{\rho_{{\mathcal A}}} \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D)).&
\end{align*}
We would like to show in the following that, for a good admissible sheaf ${\mathcal A}$, the map (i) in the chain above is surjective, so that (ii) is 0. In particular, the extension class $[{\mathcal A}]$ is completely characterized by its image $\rho_{\mathcal A}([{\mathcal A}])$ and the proposition follows.
To prove the surjectivity of (i), we consider the residue exact sequence
\[
0 \to \Omega^1_B \to \Omega^1_B(D) \xrightarrow{\mathrm{res}} \bigoplus_{i=1}^m {\mathbb{C}}_{p_i} \to 0.
\]
Applying $\mathrm{Hom}_B(\Omega_B^{1\vee}, - )$, this induces the long exact sequence
\[
\cdots \to \mathrm{Hom}_B(\Omega_B^{1\vee}, \Omega^1_B(D)) \xrightarrow{\textup{(i)'}} {\mathbb{C}}^{\oplus m} \to \mathrm{Ext}_B^1(\Omega_B^{1\vee}, \Omega^1_B) \to \cdots
\]
Using the fact that $B$ is not proper, we obtain that
\[
\mathrm{Ext}_B^1(\Omega_B^{1\vee}, \Omega^1_B) = 0.
\]
Therefore the map (i)' above is surjective. On the other hand, the map (i)' recovers (i):
\[
\mathrm{Hom}_{T^*B}(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D)) = \mathrm{Hom}_B(\Omega_B^{1\vee}, \Omega^1_B(D)) \xrightarrow{\textup{(i)} = \textup{(i)'}} {\mathbb{C}}^{\oplus m} = \bigoplus_{i=1}^m \mathrm{Ext}_{T^*B}^1(0_{B*}\Omega^{1\vee}_B, {\mathcal O}_{F_i}).
\]
Hence we conclude the surjectivity of (i), which completes the proof of the proposition.
\end{proof}
In the following three sections, we prove Theorem \ref{thm6.1} by showing that both sheaves obtained from Fourier--Mukai and the Hodge module theory respectively are good admissible sheaves, and their classes in the group
\[
\mathrm{Ext}^1_{\hat{B}}(0_{B*}\Omega_B^{1\vee}, 0_{B*} \Omega^1_B(D))
\]
are both essentially governed by the cubic form (\ref{cubic2}).
\subsection{Hodge modules}\label{Sec6.4}
We first treat the Hodge module side. The Hodge module $j_{!*}{\mathcal V}$ on $B$ can be described concretely using the canonical extension $(\overline{{\mathcal V}}, F^\bullet, \nabla)$. More precisely, by \cite[3.10]{S2} we have
\[
j_{!*}{\mathcal V} = {\mathcal D}_B \cdot \overline{{\mathcal V}} \subset \overline{{\mathcal V}}(*D)
\]
where the ${\mathcal D}_B$-action is induced by Deligne's meromorphic connection on $\overline{{\mathcal V}}(*D)$, and
\begin{equation}\label{eq:ic}
F_kj_{!*}{\mathcal V} = \sum_{i \geq 0} F_i{\mathcal D}_B \cdot F_{k - i}\overline{{\mathcal V}}.
\end{equation}
Each associated graded object $\mathrm{gr}_i(j_{!*}{\mathcal V}) = F_ij_{!*}{\mathcal V} /F_{i-1} j_{!*}{\mathcal V}$ is a coherent sheaf on $B$. The coherent sheaf
\[
\mathrm{gr}(j_{!*}{\mathcal V}) \in \mathrm{Coh}(T^*B)
\]
obtained from the Hodge module $j_{!*}{\mathcal V}$ is then completely described by the quasi-coherent sheaf
\[
\bigoplus_{k\geq -1} \mathrm{gr}_k(j_{!*}{\mathcal V})\in \mathrm{QCoh}(B)
\]
together with the Higgs field
\begin{equation}\label{Higgs_fields}
\overline{\nabla}: \mathrm{gr}_k(j_{!*}{\mathcal V}) \to \mathrm{gr}_{k+1}(j_{!*}{\mathcal V})\otimes \Omega_B^1
\end{equation}
where by Griffiths transversality all the nontrivial Higgs fields only increase the index by 1. Following a direct calculation using the formulas above as in \cite[Section 2.4]{SY2}, we have that
\[
\mathrm{gr}_{-1}(j_{!*}{\mathcal V}) \simeq \Omega_B^{1\vee}, \quad \mathrm{gr}_{0}(j_{!*}{\mathcal V}) \simeq \Omega^1_B(D), \quad \mathrm{gr}_{k}(j_{!*}{\mathcal V}) \simeq \bigoplus_{i=1}^m {\mathbb{C}}_{p_i}, \quad k>0.
\]
Moreover the nontrivial Higgs fields (\ref{Higgs_fields}) are given by (\ref{cubic2}) for $k=-1$, the nontrivial residue map
\[
\Omega^1_B(D) \xrightarrow{\mathrm{res}} \bigoplus_{i=1}^m {\mathbb{C}}_{p_i}
\]
for $k=0$, and the identity maps
\[
\mathrm{id}: \bigoplus_{i=1}^m {\mathbb{C}}_{p_i} \xrightarrow{\simeq } \bigoplus_{i=1}^m {\mathbb{C}}_{p_i}.
\]
for all $k >0$. This allows us to express the object
\[
\mathrm{gr}(j_{!*}{\mathcal V}) \in \mathrm{Coh}(T^*B)
\]
as a good admissible sheaf
\[
{\mathcal W}_{-1}^{\mathrm{HM}} \subset {\mathcal W}_{0}^{\mathrm{HM}} \subset \mathrm{gr}(j_{!*}{\mathcal V}).
\]
Here the subsheaf ${\mathcal W}_i^{\mathrm{HM}} \in \mathrm{Coh}(T^*B)$ is given by
\[
\bigoplus_{k \geq -i} \mathrm{gr}_k(j_{!*}{\mathcal V})
\]
endowed with the restricted Higgs field
\[
\overline{\nabla}: \bigoplus_{k \geq -i} \mathrm{gr}_k(j_{!*}{\mathcal V}) \to \left(\bigoplus_{k \geq -i + 1} \mathrm{gr}_k(j_{!*}{\mathcal V}) \right) \otimes \Omega^1_B.
\]
Finally, an identical argument of Section \ref{Sec5.3} yields a splitting
\begin{equation}\label{splitting_HM}
\mathrm{Hom}_{T^*{B}}( 0_{B*} \Omega_B^{1\vee}, 0_{B*} \Omega^1_B(D)[1] ) = \mathrm{Hom}_B(\Omega_B^{1\vee}, \Omega^1_B(D)[1]) \oplus H^0(B, (\Omega^1_B)^{\otimes 3}(D)).
\end{equation}
The class
\[
\rho_{\mathrm{HM}}([\mathrm{gr}(j_{!*}{\mathcal V})]) \in \mathrm{Hom}_{T^*{B}}( 0_{B*} \Omega_B^{1\vee}, 0_{B*} \Omega^1_B(D)[1] )
\]
is then $(0, [\overline{\nabla}])$ via the splitting (\ref{splitting_HM}) with $[\overline{\nabla}]$ given by (\ref{cubic2}).
\subsection{Fourier--Mukai}\label{Sec6.5}
For our elliptic fibration $\pi_M: M \to B$ with a section $s_B: B \to M$, the partial Fourier--Mukai transform
\[
\phi_{\mathrm{FM}}: D^b\mathrm{Coh}(M) \to D^b\mathrm{Coh}(P)
\]
can actually be upgraded to a \emph{full} Fourier--Mukai transform
\[
\widetilde{\phi}_{\mathrm{FM}}: D^b\mathrm{Coh}(M) \to D^b\mathrm{Coh}(M)
\]
as we review in the following.
We denote by $M^\vee \to B$ the relative compactified Jacobian fibration parameterizing torsion-free, rank 1, degree 0 sheaves on the fibers of $\pi: M \to B$. Since $M^\vee$ is a fine moduli space with a $0$-section $0_B: B \rightarrow M^\vee$, there is a universal Poincar\'e sheaf on $M \times_B M^\vee$. We choose the universal Poincar\'e sheaf ${\mathcal P}$ such that the induced Fourier--Mukai transform
\[
\widetilde{\phi}_{\mathrm{FM}}: D^b\mathrm{Coh}(M) \xrightarrow{\simeq} D^b\mathrm{Coh}(M^\vee)
\]
satisfies $\widetilde{\phi}_{\mathrm{FM}}({\mathcal O}_M) \simeq 0_{B*} {\mathcal O}_B$. We have that $M^\vee$ is naturally isomorphic to $M$ by
\[
M \xrightarrow{\simeq} M^\vee, \quad x \mapsto \iota_{s*}(\mathfrak{m}^\vee_x \otimes {\mathcal O}_{M_s}(-s))
\]
where $M_s$ is the closed fiber containing $x$, $\iota_s: M_s \hookrightarrow M$ is the closed embedding, $\mathfrak{m}_x$ is the ideal sheaf of $x$ in $M_s$, and $s$ is viewed as a point on $M_s$ lying in the intersection with the section $B \subset M$. From now on we identify $M^\vee$ with $M$. The group scheme $P \to B$ is then obtained as an open surface $P\subset M$ removing the nodes of the singular fibers. The partial Fourier--Mukai transform $\phi_{\mathrm{FM}}$ is the composition of the full Fourier--Mukai $\widetilde{\phi}_{\mathrm{FM}}$ with the restriction map associated with $P \subset M$.
\begin{lem}\label{lem6.3} The following hold for the Fourier--Mukai transform $\widetilde{\phi}_{\mathrm{FM}}$.
\begin{enumerate}
\item[(i)] For ${\mathcal K} \in D^b\mathrm{Coh}(B)$, we have
\[
\widetilde{\phi}_{\mathrm{FM}}(\pi_M^* {\mathcal K}) \simeq 0_{B*} {\mathcal K}.
\]
\item[(ii)] For a point $x \in M_s$ in a closed fiber $\iota_s: M_s \hookrightarrow M$, we have
\[
\widetilde{\phi}_{\mathrm{FM}}({\mathbb{C}}_x) \simeq \iota_{s*} (\mathfrak{m}_x \otimes {\mathcal O}_{M_s}(s))[1].
\]
\item[(iii)] Let $F_i$ be the nodal fiber of $\pi: M \to B$ over $p_i$ with $x_i \in F_i$ the node. Let $\nu_i: \widetilde{F_i} \to F_i$ be the normalization. Then we have
\[
\widetilde{\phi}_{\mathrm{FM}}(\nu_{i*} {\mathcal O}_{\widetilde{F}_i}) \simeq \iota_{p_i*} \mathfrak{m}_{x_i}[1].
\]
\end{enumerate}
\end{lem}
\begin{proof}
(i, ii) are easy exercises deduced from the definition. We now prove (iii). We consider the short exact sequence
\begin{equation}\label{eqn60}
0 \to {\mathcal O}_{F_i} \to \nu_{i*} {\mathcal O}_{\widetilde{F}_i} \to {\mathbb{C}}_{x_i} \to 0.
\end{equation}
Applying the functor $\widetilde{\phi}_{\mathrm{FM}}$, we obtain from (i, ii) the exact triangle
\begin{equation}\label{eqn61}
{\mathbb{C}}_{p_i} \to \widetilde{\phi}_{\mathrm{FM}}(\nu_{i*}{\mathcal O}_{\widetilde{F}_i}) \to \iota_{p_i*} (\mathfrak{m}_{x_i} \otimes {\mathcal O}_{F_i}(p_i))[1] \xrightarrow{+1}.
\end{equation}
The associated long exact sequence reads:
\[
0 \to {\mathcal H}^{-1}\left( \widetilde{\phi}_{\mathrm{FM}}(\nu_{i*}{\mathcal O}_{\widetilde{F}_i}) \right) \to\iota_{p_i*} (\mathfrak{m}_{x_i} \otimes {\mathcal O}_{F_i}(p_i)) \to {\mathbb{C}}_{p_i} \xrightarrow{(*)} {\mathcal H}^{0}\left( \widetilde{\phi}_{\mathrm{FM}}(\nu_{i*}{\mathcal O}_{\widetilde{F}_i}) \right) \to 0.
\]
It suffices to show that the map $(*)$ in the above sequence is trivial. Assume it is not. Then the arrow $\iota_{p_i*} (\mathfrak{m}_{x_i} \otimes {\mathcal O}_{F_i}(p_i)) \to {\mathbb{C}}_{p_i}$ has to be trivial which forces (\ref{eqn61}) to split; equivalently~(\ref{eqn60}) splits which is a contradiction. This completes the proof of (iii).
\end{proof}
Now we show that $\phi_{\mathrm{FM}}(\Omega_M^1)$ is a good admissible sheaf after restricting to a formal neighborhood of the $0$-section $B\subset M$. Instead we work with the object
\begin{equation*}\label{FM_shf}
\widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1) \in D^b\mathrm{Coh}(M).
\end{equation*}
We consider the following triangle of morphisms
\begin{equation}\label{triangle1}
\begin{tikzcd}[column sep=small]
\Omega_M^1 \arrow[dr, ""] \arrow[rr, ""] & & \Omega_M^1(\log F) \arrow[dl, ""] \\
& \Omega^1_{M/B}(\log F) &
\end{tikzcd}
\end{equation}
where $\Omega_M^1 \rightarrow \Omega_M^1(\log F)$ and $\Omega_M^1(\log F) \to \Omega^1_{M/B}(\log F)$ are the natural maps, and the third arrow is the composition of these two maps. By the octahedral axiom of triangulated categories, the cones associated with the three maps of (\ref{triangle1}) form an exact triangle
\[
\bigoplus_{i=1}^m \nu_{i*} {\mathcal O}_{\widetilde{F}_i}[-1] \to {\mathcal K}_0 \to \pi_{M}^* \Omega^1_B(D) \xrightarrow{+1}.
\]
Here ${\mathcal K}_0:= \mathrm{cone}\left( \Omega_{M}^1 \to \Omega_{M/B}^1(\log F)\right)[-1]$. Applying the functor $\widetilde{\phi}_{\mathrm{FM}}$ to this exact triangle, we obtain from Lemma \ref{lem6.3} that the object $\widetilde{\phi}_{\mathrm{FM}}({\mathcal K}_0)$ is a sheaf concentrated in degree 0 which fits into the exact sequence
\begin{equation}\label{eq64}
0 \to \bigoplus_{i=1}^m \iota_{p_i*}\mathfrak{m}_{x_i} \to \widetilde{\phi}_{\mathrm{FM}}({\mathcal K}_0) \to 0_{B*} \Omega^1_B(D) \to 0.
\end{equation}
We set
\[
\widetilde{{\mathcal W}}_{-1}^{\mathrm{FM}}: = \bigoplus_{i=1}^m \iota_{p_i*}\mathfrak{m}_{x_i}, \quad \widetilde{{\mathcal W}}_0^{\mathrm{FM}}:= \widetilde{\phi}_{\mathrm{FM}}({\mathcal K}_0).
\]
\begin{prop}\label{prop6.4}
The object $\widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1)$ is a coherent sheaf on $M$ with an increasing filtration of subsheaves:
\[
\widetilde{{\mathcal W}}_{-1}^{\mathrm{FM}} \subset \widetilde{{\mathcal W}}_0^{\mathrm{FM}} \subset \widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1)
\]
which satisfies
\begin{enumerate}
\item[(a)] $\widetilde{{\mathcal W}}_{-1}^{\mathrm{FM}} = \bigoplus_{i=1}^m \iota_{p_i*}\mathfrak{m}_{x_i}$, \item[(b)] $\widetilde{{\mathcal W}}^\mathrm{FM}_0/ \widetilde{{\mathcal W}}^\mathrm{FM}_{-1} \simeq 0_{B*}\Omega^1_B(D)$, and
\item[(c)] $\widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1)/\widetilde{{\mathcal W}}^\mathrm{FM}_0 \simeq 0_{B*}\Omega_B^{1\vee}$.
\end{enumerate}
\end{prop}
\begin{proof}
By the definition of ${\mathcal K}_0$, the isomorphism (\ref{6.2_1}), and Lemma \ref{lem6.3} (i), we obtain an exact triangle
\[
\widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1) \to 0_{B*}\Omega_B^{1\vee} \to \widetilde{\phi}_{\mathrm{FM}}({\mathcal K}_0)[1] \xrightarrow{+1}.
\]
Since $\widetilde{{\mathcal W}}_0^{\mathrm{FM}} = \widetilde{\phi}_{\mathrm{FM}}({\mathcal K}_0)$ is a sheaf concentrated in degree 0, the above exact triangle yields a short exact sequence
\[
0 \to \widetilde{{\mathcal W}}_0^{\mathrm{FM}} \to \widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1) \to 0_{B*}\Omega_B^{1\vee} \to 0.
\]
This proves (c). For the remaining parts, we note that (a) is given by the definition of $\widetilde{{\mathcal W}}_{-1}^{\mathrm{FM}}$, and (b) follows from the short exact sequence (\ref{eq64}).
\end{proof}
Analogously to (\ref{eqn56}), the description of $\widetilde{{\mathcal W}}^\mathrm{FM}_0$ gives an extension (\ref{eq64}) which induces a class
\[
[\widetilde{{\mathcal W}}^\mathrm{FM}_0] \in \bigoplus_{i=1}^m \mathrm{Ext}_M^1(0_{B*} \Omega^1_B(D), \iota_{p_i*}\mathfrak{m}_{x_i}).
\]
Each extension group on the right-hand side is 1-dimensional.
\begin{prop}\label{prop6.5}
Each summand of the extension class $[\widetilde{{\mathcal W}}^\mathrm{FM}_0]$ in
\[
\mathrm{Ext}_M^1(0_{B*} \Omega^1_B(D), \iota_{p_i*}\mathfrak{m}_{x_i})
\]
is nonzero.
\end{prop}
\begin{proof}
Since this is a local question, we may assume that $\pi: M \to B$ only has one singular nodal fiber $F \subset M$ over $p \in B$ with $x\in F$ the node. We need to show that the extension
\[
0 \to \iota_{p*}\mathfrak{m}_{x}\to \widetilde{{\mathcal W}}^{\mathrm{FM}}_0 \to 0_{B*} \Omega^1_B(P) \to 0, \quad \widetilde{{\mathcal W}}^{\mathrm{FM}}_0 = \widetilde{\phi}_{\mathrm{FM}}({\mathcal K}_0)
\]
is nontrivial. Assume this is trivial. Then applying the inverse Fourier--Mukai $\widetilde{\phi}^{-1}_{\mathrm{FM}}$ to this exact sequence, we obtain that ${\mathcal K}_0$ has nontrivial cohomology in degrees 0 and 1 with
\begin{equation}\label{contra}
{\mathcal H}^1({\mathcal K}_0) \simeq \nu_* {\mathcal O}_{\widetilde{F}}.
\end{equation}
Here $\nu: \widetilde{F} \to F$ is the normalization.
To reach a contradiction, now we describe ${\mathcal K}_0$ by another triangle of morphisms
\begin{equation}\label{triangle2}
\begin{tikzcd}[column sep=small]
\Omega_M^1 \arrow[dr, ""] \arrow[rr, ""] & & \Omega_{M/B}^1 \arrow[dl, ""] \\
& \Omega^1_{M/B}(\log F). &
\end{tikzcd}
\end{equation}
Here $\Omega_M^1 \to \Omega^1_{M/B}$ and $\Omega^1_{M/B} \to \Omega^1_{M/B}(\log F)$ are the natural maps, and their composition recovers the map $\Omega_{M}^1 \to \Omega_{M/B}^1(\log F)$ of (\ref{triangle1}). The octahedral axiom yields that the cones associated with the three maps of (\ref{triangle2}) form an exact triangle
\[
\pi_M^*\Omega^1_B \to {\mathcal K}_0 \to {\mathbb{C}}_x[-1] \xrightarrow{+1}
\]
where ${\mathbb{C}}_x \simeq\mathrm{cone}\left( \Omega_{M/B}^1 \to \Omega_{M/B}^1(\log F)\right)$. Consequently, we have
\[
{\mathcal H}^1({\mathcal K}_0) \simeq {\mathbb{C}}_x
\]
which contradicts (\ref{contra}).
\end{proof}
The extension class
\[
[\widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1) ] \in \mathrm{Ext}_M^1(0_{B*}\Omega_B^{1\vee}, \widetilde{{\mathcal W}}_0^{\mathrm{FM}} )
\]
is sent naturally to a class
\begin{equation}\label{Ext1}
\widetilde{\rho}([\widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1) ] ) \in \mathrm{Ext}_M^1(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D) )
\end{equation}
via the natural morphism induced by Proposition \ref{prop6.4}:
\[
\widetilde{\rho}: \mathrm{Ext}_M^1(0_{B*}\Omega_B^{1\vee}, \widetilde{{\mathcal W}}_0^{\mathrm{FM}} ) \to \mathrm{Ext}_M^1(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D) ).
\]
Applying the functor $\widetilde{\phi}_{\mathrm{FM}}$ to the triangle (\ref{triangle1}), we see that the class (\ref{Ext1}) represents the extension
\[
\widetilde{\phi}_{\mathrm{FM}}(\pi_M^*\Omega^1_B(D)) \to \widetilde{\phi}_{\mathrm{FM}}(\Omega^1_{M}(\log F)) \to \widetilde{\phi}_{\mathrm{FM}}(\Omega^1_{M/B}(\log F)) \xrightarrow{+1},
\]
which is the exact triangle obtained by applying $\widetilde{\phi}_{\mathrm{FM}}$ to the exact sequence (\ref{6.2_0}).
We note that both sheaves $0_{B*}\Omega_B^{1\vee}$ and $0_{B*}\Omega^1_B(D)$ are supported on the $0$-section. Therefore we have a natural isomorphism
\[
\mathrm{Ext}_M^1(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D) ) = \mathrm{Ext}_{\hat{B}}^1(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D) ).
\]
An identical argument as in Section \ref{Sec5.3} yields a splitting
\[
\mathrm{Ext}_{\hat{B}}^1(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D) ) = \mathrm{Ext}_{{B}}^1(\Omega_B^{1\vee}, 0_{B*}\Omega_B(D)) \oplus H^0(B, (\Omega^1_B)^{\otimes 3}(D));
\]
this is the counter-part of (\ref{splitting_HM}) on the Fourier--Mukai side. Furthermore, the class (\ref{Ext1}) is~$(0, [\overline{\nabla}])$ with $[\overline{\nabla}]$ given by (\ref{cubic2}).
\subsection{Proof of Theorem \ref{thm6.1}}
We fix an identification of the formal neighborhoods of the $0$-sections in $T^*B$ and $M$ respectively using Proposition \ref{prop1.4}; we denote them by $\hat{B}$ uniformly.
On the Hodge module side, by Section \ref{Sec6.4} $\mathrm{gr}(j_{!*}{\mathcal V})$ is a good admissible sheaf on $T^*B$. Therefore its restriction
\[
\mathrm{gr}(j_{!*}{\mathcal V})|_{\hat{B}}
\]
is a good admissible sheaf on $\hat{B}$.
On the Fourier--Mukai side, Propositions \ref{prop6.4} and \ref{prop6.5} imply that
\[
\phi_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}} = \widetilde{\phi}_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}}
\]
is a good admissible sheaf on $\hat{B}$.
Furthermore, by the discussion at the ends of Sections \ref{Sec6.4} and \ref{Sec6.5} respectively, we have
\[
\rho_{\mathrm{HM}}([ \mathrm{gr}(j_{!*}{\mathcal V})|_{\hat{B}} ]) = \rho_{\mathrm{FM}}([ \phi_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}} ]) = (0,[\overline{\nabla}])\in \mathrm{Ext}_{\hat{B}}^1(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D)). \]
Finally we apply Proposition \ref{prop6.2} and conclude that the two good admissible sheaves obtained from the Hodge module and the Fourier--Mukai transform are isomorphic. This completes the proof. \qed
\subsection{Remarks on the cuspidal case}\label{sec6.7}
In fact the same strategy also applies to cuspidal fibers.\footnote{As we work with an elliptic fibration with integral fibers, nodes and cusps are the only possible singularities.} Here we briefly sketch the key steps and leave the details to the interested reader.
From now on we allow the elliptic fibration $\pi_M : M \to B$ to have nodal fibers $F_1, \ldots, F_m$ over $p_1, \ldots, p_m \in B$, and cuspidal fibers $G_1, \ldots, G_l$ over $q_1, \ldots, q_l \in B$. The rest of the assumptions on $\pi_M$ remains the same as in Section \ref{sec:6.1}, including $B$ non-proper. We set
\[
D := \sum_{i = 1}^mp_i \subset B, \quad C := \sum_{i = 1}^lq_i \subset B, \quad F := \sum_{i = 1}^mF_i \subset M, \quad G := \sum_{i = 1}^lG_i \subset M.
\]
In the presence of cusps, the notion of an \emph{admissible sheaf} ${\mathcal A} \in \mathrm{Coh}(T^*B)$ should be altered to a $4$-step filtration
\[
{\mathcal W}^{{\mathcal A}, \mathrm{red}}_{-1} \subset {\mathcal W}_{-1}^{\mathcal A} \subset {\mathcal W}_{0}^{\mathcal A} \subset {\mathcal A}
\]
which satisfies
\begin{enumerate}
\item[(a)] ${\mathcal W}^{{\mathcal A}, \mathrm{red}}_{-1} \simeq \oplus_{i=1}^m {\mathcal O}_{F_i} \oplus \oplus_{i = 1}^l {\mathcal O}_{G_i}$,
\item[(b)] ${\mathcal W}^{\mathcal A}_{-1}/{\mathcal W}^{{\mathcal A}, \mathrm{red}}_{-1} \simeq \oplus_{i = 1}^l {\mathcal O}_{G_i}$,
\item[(c)] ${\mathcal W}^{\mathcal A}_0/ {\mathcal W}^{\mathcal A}_{-1} \simeq 0_{B*}\Omega^1_B(D + 2C)$, and
\item[(d)] ${\mathcal A}/{\mathcal W}^{\mathcal A}_0 \simeq 0_{B*}\Omega_B^{1\vee}$.
\end{enumerate}
An admissible sheaf ${\mathcal A}$ is \emph{good} if the following hold.
\begin{enumerate}
\item[(i)] The extension class
\[
[{\mathcal W}^{\mathcal A}_{-1}] \in \mathrm{Ext}^1_{T^*B}\left(\bigoplus_{i = 1}^l {\mathcal O}_{G_i}, \bigoplus_{i=1}^m {\mathcal O}_{F_i} \oplus \bigoplus_{i = 1}^l {\mathcal O}_{G_i}\right) = \bigoplus_{i = 1}^{l}\mathrm{Ext}^1_{T^*B}({\mathcal O}_{G_i}, {\mathcal O}_{G_i})
\]
has a nonzero summand in each $\mathrm{Ext}^1_{T^*B}({\mathcal O}_{G_i}, {\mathcal O}_{G_i}) \simeq {\mathbb{C}}$. Up to scaling ${\mathcal O}_{G_i}$ we may assume that each summand is $1$. In particular, the sheaf ${\mathcal W}^{\mathcal A}_{-1}$ is of the form
\[
{\mathcal W}^{\mathcal A}_{-1} \simeq \bigoplus_{i=1}^m {\mathcal O}_{F_i} \oplus \bigoplus_{i = 1}^l {\mathcal O}_{\mathbf{G}_i}
\]
where $\mathbf{G}_i \subset T^*B$ is a thickening of $G_i$.
\item[(ii)] The extension class
\begin{multline*}
[{\mathcal W}^{\mathcal A}_{0}] \in \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega^1_B(D + 2C), {\mathcal W}^{\mathcal A}_{-1}) \\
= \bigoplus_{i = 1}^m \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega^1_B(D + 2C), {\mathcal O}_{F_i}) \oplus \bigoplus_{i = 1}^l\mathrm{Ext}^1_{T^*B}(0_{B*}\Omega^1_B(D + 2C), {\mathcal O}_{\mathbf{G}_i})
\end{multline*}
has a nonzero summand in each $\mathrm{Ext}^1_{T^*B}(0_{B*}\Omega^1_B(D + 2C), {\mathcal O}_{F_i}) \simeq {\mathbb{C}}$ as well as in each~$\mathrm{Ext}^1_{T^*B}(0_{B*}\Omega^1_B(D + 2C), {\mathcal O}_{\mathbf{G}_i}) \simeq {\mathbb{C}}^2$. Up to an automorphism of ${\mathcal O}_{F_i}$ and ${\mathcal O}_{\mathbf{G}_i}$ we may assume that the summands are either $1$ or $(1, 0)$.
\end{enumerate}
Further, the short exact sequence
\[
0 \to {\mathcal W}^{\mathcal A}_{-1} \to {\mathcal W}^{\mathcal A}_{0} \to 0_{B*}\Omega_B^1(D + 2C) \to 0
\]
induces a natural morphism
\[
\rho_{\mathcal A} : \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, {\mathcal W}^{\mathcal A}_{0}) \to \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega_B^1(D + 2C)),
\]
sending the extension class $[{\mathcal A}] \in \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, {\mathcal W}^{\mathcal A}_{0})$ to $\rho_{\mathcal A}([{\mathcal A}])$.
A version of Proposition \ref{prop6.2} states that over a non-proper curve $B$, two good admissible sheaves ${\mathcal A}, {\mathcal A}' \in \mathrm{Coh}(T^*B)$ are isomorphic if the classes $\rho_{\mathcal A}([{\mathcal A}])$ and $\rho_{{\mathcal A}'}([{\mathcal A}'])$ coincide:
\[
\rho_{\mathcal A}([{\mathcal A}]) = \rho_{{\mathcal A}'}([{\mathcal A}']) \in \mathrm{Ext}^1_{T^*B}(0_{B*}\Omega_B^{1\vee}, 0_{B*}\Omega^1_B(D + 2C)).
\]
On the Hodge module side we have $P_1 = j_{!*}{\mathcal V}$ as in Section \ref{sec:6.1}. The fact that $\mathrm{gr}(j_{!*}{\mathcal V})$ is a good admissible sheaf on $T^*B$ is again proven by an explicit calculation using the local monodromy $(\begin{smallmatrix} 1&1\\-1&0 \end{smallmatrix})$ around each $q_i \in B$ corresponding to the cuspidal fiber $G_i$. Note that however, this time one should distinguish the two canonical extensions with respect to $[0, 1)$ and $(-1, 0]$ (which differ by a twist by ${\mathcal O}_B(C)$) and use the latter for $\overline{{\mathcal V}}$ in \eqref{eq:ic}. The end results, deduced from formulas in \cite[Theorem 2.6]{K} and \cite{Ma3}, are
\begin{gather*}
\mathrm{gr}_{-1}(j_{!*}{\mathcal V}) \simeq \Omega_B^{1\vee}, \quad \mathrm{gr}_{0}(j_{!*}{\mathcal V}) \simeq \Omega^1_B(D + 2C),\\
\mathrm{gr}_{k}(j_{!*}{\mathcal V}) \simeq \bigoplus_{i=1}^m {\mathbb{C}}_{p_i} \oplus \bigoplus_{i = 1}^l {\mathcal O}_{\mathbf{q}_i}, \quad k>0,
\end{gather*}
where $\mathbf{q}_i$ is the length $2$ fat point supported on $q_i$. The nontrivial Higgs fields \eqref{Higgs_fields} are given by a cubic form
\begin{equation} \label{eq:cubiccusp}
[\overline{\nabla}] \in H^0(B, (\Omega^1_B)^{\otimes 3}(D + 2C))
\end{equation}
for $k = -1$, the residue map for $k = 0$, and the identity maps for all $k > 0$. As in the nodal case we set ${\mathcal W}_i^{\mathrm{HM}} \in \mathrm{Coh}(T^*B)$ to be
\[
\bigoplus_{k \geq -i}\mathrm{gr}_k(j_{!*}{\mathcal V})
\]
endowed with the restricted Higgs field, and ${\mathcal W}_{-1}^{\mathrm{HM}, \mathrm{red}}$ is the obvious subsheaf of ${\mathcal W}_{-1}^{\mathrm{HM}}$. It is straightforward to check that $\mathrm{gr}(j_{!*}{\mathcal V})$ is a good admissible sheaf on $T^*B$.
The Fourier--Mukai side is treated with the help of a log resolution. Let $f: \widetilde{M} \to M$ be the resolution obtained by blowing up each cusp $y_i \in G_i$ three times. We write $\pi_{\widetilde{M}}: \widetilde{M} \to B$ for the composition $\pi_M \circ f$. Let $E_{1, i}, E_{2, i}, E_{3, i} \subset \widetilde{M}$ be the three (strict transforms of) exceptional divisors associated with $y_i$. For $k = 1, 2, 3$, we set $E_k := \sum_{i = 1}^lE_{k, i} \subset \widetilde{M}$. We also set
\[
\mathbf{E} = E_1 + 2E_2 + 5E_3 \subset \widetilde{M}.
\]
We consider the following triangle of morphisms
\begin{equation} \label{eq:fancytriangle}
\begin{tikzcd}[column sep=small]
\Omega_M^1 \arrow[dr, ""] \arrow[rr, ""] & & Rf_*\left(\Omega^1_{\widetilde{M}}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E})\right) \arrow[dl, ""] \\
& Rf_*\left(\Omega^1_{\widetilde{M}/B}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E})\right). &
\end{tikzcd}
\end{equation}
We define
\begin{gather*}
{\mathcal K}_{-1}^{\mathrm{red}} := \mathrm{cone}\left(\Omega_M^1 \to Rf_*\left(\Omega^1_{\widetilde{M}}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E})\right)\right)[-1], \\
{\mathcal K}_0 := \mathrm{cone}\left(\Omega_M^1 \to Rf_*\left(\Omega^1_{\widetilde{M}/B}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E})\right)\right)[-1].
\end{gather*}
By the octahedral axiom, the cones associated with the three maps of \eqref{eq:fancytriangle} form an exact triangle
\begin{equation} \label{eq:k-1k0}
{\mathcal K}_{-1}^{\mathrm{red}} \to {\mathcal K}_0 \to \pi_M^*\Omega^1_B(D + C) \otimes Rf_*{\mathcal O}_{\widetilde{M}}(\mathbf{E}) \xrightarrow{+1}.
\end{equation}
Moreover, the natural inclusion $\mathbf{E} \subset \pi_{\widetilde{M}}^{-1}(C)$ induces a morphism
\begin{equation} \label{eq:inclusion}
\pi_M^*\Omega^1_B(D + C) \otimes Rf_*{\mathcal O}_{\widetilde{M}}(\mathbf{E}) \to \pi_M^*\Omega^1_B(D + 2C).
\end{equation}
We define
\[
{\mathcal K}_{-1} := \mathrm{cone}\left({\mathcal K}_0 \to \pi_M^*\Omega^1_B(D + 2C)\right)[-1]
\]
where the arrow is obtained by composing \eqref{eq:k-1k0} and \eqref{eq:inclusion}. Finally we set
\[
{\mathcal W}^{\mathrm{FM}, \mathrm{red}}_{-1} := \phi_{\mathrm{FM}}({\mathcal K}_{-1}^{\mathrm{red}}), \quad {\mathcal W}^{\mathrm{FM}}_{-1} := \phi_{\mathrm{FM}}({\mathcal K}_{-1}), \quad {\mathcal W}^{\mathrm{FM}}_{0} := \phi_{\mathrm{FM}}({\mathcal K}_{0}).
\]
It remains to check that all three terms above are sheaves concentrated in degree $0$ which are part of a filtration
\[
{\mathcal W}^{\mathrm{FM}, \mathrm{red}}_{-1} \subset {\mathcal W}^{\mathrm{FM}}_{-1} \subset {\mathcal W}^{\mathrm{FM}}_{0} \subset \phi_{\mathrm{FM}}(\Omega_M^1),
\]
and that $\phi_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}}$ is indeed a good admissible sheaf on $\hat{B}$. For example, by \cite[2.10]{K} we have
\[
\Omega^1_{\widetilde{M}/B}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E}) \simeq \omega_{\widetilde{M}/B}
\]
so that under the symplectic form $\sigma$ of $M$ there is an isomorphism
\[
Rf_*\left(\Omega^1_{\widetilde{M}/B}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E})\right) \simeq \pi_B^*\Omega_B^{1\vee}.
\]
Comparing with the definition of ${\mathcal K}_0$, we find an exact triangle
\[
{\mathcal K}_0 \to \Omega^1_M \to \pi_B^*\Omega_B^{1\vee} \xrightarrow{+1}
\]
whose Fourier--Mukai image is the expected short exact sequence
\[
0 \to {\mathcal W}^{\mathrm{FM}}_{0} \to \phi_{\mathrm{FM}}(\Omega_M^1) \to 0_{B*}\Omega_B^{1\vee} \to 0.
\]
One also uses the exact triangle
\begin{multline*}
\pi_M^*\Omega^1_B(D + C) \otimes Rf_*{\mathcal O}_{\widetilde{M}}(\mathbf{E}) \to Rf_*\left(\Omega^1_{\widetilde{M}}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E})\right) \\
\to Rf_*\left(\Omega^1_{\widetilde{M}/B}(\log f^{-1}(F + G))\otimes {\mathcal O}_{\widetilde{M}}(\mathbf{E})\right) \xrightarrow{+1}
\end{multline*}
and the log Katz--Oda theorem \cite{Katz} to relate the extension class $\rho_{\mathrm{FM}}([\phi_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}}])$ to the same cubic form $[\overline{\nabla}]$ as in \eqref{eq:cubiccusp}, and to conclude that
\[
\rho_{\mathrm{HM}}([ \mathrm{gr}(j_{!*}{\mathcal V})|_{\hat{B}} ]) = \rho_{\mathrm{FM}}([\phi_{\mathrm{FM}}(\Omega_M^1)|_{\hat{B}}])
\]
as in the nodal case.
|
\section{Introduction}
Two-dimensional (2D) materials have been emerging as a promsing functional materials with wide applications due to the novel fundamental physics with the reduced dimension\cite{glavin2020emerging} The systematic discovery and synthesis of functional 2D materials has been the focus of many studies\cite{2DMatPedia2,lyngby2022data,V2DB,C2DB2,MC2D,wyss2022large,ares2022recent}.
Having exceptional and tunable properties, 2D materials hold strong promise in semiconductor, energy, and health applications\cite{briggs2019roadmap,li2020new}. Since the 2010 Nobel prize-winning discovery of graphenes\cite{graphene} with a simple 2D structure of carbon atoms but with attractive and complex physics, only a few thousands of distinct 2D materials have been successfully synthesized\cite{C2DB2}. The isolation of single graphene sheets, which proves that 2D systems can exist, gives rise to the discovery of many 2D materials with unique superconducting\cite{cao1}, electronic\cite{manzeli20172d}, magnetic\cite{huang2021two}, and topological properties\cite{kou2017two}. In addition to being test beds for studying the behavior of systems in reduced dimensions, 2D materials hold great promise for various applications in optoelectronics\cite{wang2012electronics}, catalysis\cite{deng2016catalysis}, and the energy sector\cite{anasori20172d}. The research effort has been mainly concentrated on the systems which have bulk counterparts representing anisotropic crystals with layers held together by van der Waals (vdW) forces, with the most prominent example being the graphene and graphite. The weak interlayer interaction leads to a natural structural separation of the 2D subunits in the crystals, therefore making the mechanical or liquid-phase exfoliation possible.
Currently, there are three approaches for generating 2D materials: the top-down exfoliation method starts with a bulk material and exfoliates to make it thinner and peel the layers to obtain 2D materials; the bottom-up approach instead starts with existing 2D materials and uses element substitution to generate new materials. The third one is the de novo structure generation approach\cite{lyngby2022data} based on deep learning generative models such as CDVAE\cite{CDVAE}.
To get new 2D materials through the exfoliation method, we need to judge whether the 3D bulk material is layered so that it can be exfoliated. The layer screening process first checks the distance between atoms to identify whether these atom pairs are bonded. It then calculates the bonded atom clusters both in a 3x3x3 supercell and the unit cell. If the number of clusters in the supercell is three times that in the unit cell, the structure is tagged as layered\cite{larsen2019definition}. 2D materials are theoretically exfoliated by extracting one cluster in the standard conventional unit cell of the screened layered bulk structures. In the element substitution method, all the elements of the periodic table are categorized into different groups according to their column number. Elements with the same column (group) number share the same number of electrons in their outermost orbit, and elements with the same row (period) number share the same number of electronic layers, which means that elements in the same group or neighbor share some similar chemical properties. The substitution method starts with the structure of a known 2D material and replaces one or more element in this material with other elements either in the same group or its neighbor elements.
Both the element substitution method and the de novo generation method start with known 2D crystal structures. Currently, there are several open-source 2D material databases generated through exfoliation, substitution, or de novo generation methods. The Computational 2D Materials Database (C2DB)\cite{C2DB1,C2DB2} uses both exfoliation and substitution methods to organize a wealth of computed properties for 4038 (checked in October 2022) atomically thin 2D materials. The materials in the C2DB comprise both experimentally known and not previously synthesized structures. They have been generated in a systematic fashion by the combinatorial decoration of different 2D crystal lattices. Starting from 108,423 unique, experimentally known 3D compounds, MC2D\cite{MC2D} uses only the exfoliation method to identify a subset of 5,619 compounds that appear layered according to robust geometric and bonding criteria (checked in October 2022). High-throughput calculations using van der Waals density functional theory (DFT) validated against experimental structural data and calculating random phase approximation binding energies further allowed the identification of 1,825 compounds that can be either easily or potentially be exfoliated. 2D Materials Encyclopedia (2DMatPedia) database\cite{2DMatPedia,2DMatPedia2} screened all bulk materials in the database of Materials Project for layered structures by a topology-based detection algorithm and theoretically exfoliated them into monolayers. Then, new 2D materials are generated by applying chemical substitution of elements in known 2D materials by other elements from the same group in the periodic table. There are a total of 6351 materials in the current 2DMatPedia database (checked in December 2022), whereas 2940 were obtained by exfoliating existing layered materials (top-down approach), 3409 were obtained by the chemical substitution of 2D materials (bottom-up approach), and 2 were obtained from the literature via neither a top-down or bottom-up approach. The bottom-up approach starts from the 35 unary and 755 binary compounds obtained from the top-down approach. Only the same-column elements are used for substitution. By employing 22 different 2D crystal prototypes and 52 different chemical elements from the periodic table, Virtual 2D Materials Database (V2DB)\cite{V2DB} applied a brute-force substitution method to generate a systematic library of more than 72 million 2D compounds. Next, symmetry, neutrality, and stability sequential filtering layers are applied to identify 316,505 likely stable 2D materials.
Materials Cloud\cite{MatC} provides a practical yet straightforward approach to assessing whether any 3D compound can be exfoliated into 2D layers. The multistep procedure starts by pre-screening layered structures based on geometrical criteria requiring only the atomic positions of the atoms in the structure. The resulting filtered structures are featurized, and finally, an ML model based on a random forest classifier is applied to assess whether the material can be exfoliated or, instead, has high binding energy (HBE).
Friedrich et al.\cite{friedrich2022data} outlined a new set of non-vdW 2D materials by employing data-driven concepts and extensive calculations. By filtering the AFLOW-ICSD database according to the structural prototype of the two experimentally realized systems Fe$_2$O$_3$ and FeTiO$_3$, they have obtained 8 binary and 20 ternary 2D material candidates. The most recent approach for 2D material generation is based on the deep learning generative model. Lyngby et al.\cite{lyngby2022data} use a crystal diffusion variational
autoencoder (CDVAE)\cite{CDVAE} to generate new 2D structures of high chemical
and structural diversity and with formation energies mirroring the training structures. They also use the element substitution method to generate new possible 2D materials based on the newly generated 2D structures. In total, they find 11630 predicted new 2D materials, where 3073 are generated by CDVAE and 8,599 come from element substitution of these 3073 structures. They find that 2,004 of their generated 2D candidates are within 50 meV of the convex hull and could potentially be synthesised. In order to capture all these structural features of 2D and quasi-2D materials, Wang et al.\cite{wang2012effective} developed a new 2D structure search module in CALYPSO code that is based on 2D PSO algorithm but allows the relaxation of atomic coordinates in the perpendicular direction. They predicted a new family B$_x$N$_y$ with different chemical compositions that have layered structures.
Here we propose a computational pipeline, material transformer generator(MTG), for generative discovery of new 2D materials (and other crystal materials). Our method is based on combining a 2D composition generator trained with known 2D material compositions, two template based crystal structure predictors, two machine learning potential-based structure relaxers, and DFT relaxation. Extensive experiments show that our MTG pipeline can be used discover a large number of hypothetical 2D materials.
\section{Method}
\begin{figure}[ht]
\centering
\begin{minipage}[c]{1.0\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/architecture.png}
\end{minipage}\\
\caption{Architecture of our material transformer generator(MTG) pipeline. BLMM is a tranformer neural network based composition generator. TCSP and CSPML are template based crystal structure prediction algorithms; and BOWSR and M3GNET are machine learning potential based structure relaxing algorithms. DFT relaxation is a first-principles calculation method.}
\label{fig:flowchart}
\end{figure}
Figure\ref{fig:flowchart} (a) shows how the framework of our MTG pipeline for 2D material generation. We collected known 2D formulas and their structures from open datasets C2DB, MC2D, 2DMatPedia, and V2DB. We then train a set of BLMM (blank language models for materials) composition generators with known 2d formulas to generate new 2D formulas.
Next, we use known 2D structures as templates for structure prediction of these candidate 2D formulas using two crystal structure prediction algorithms TCSP and CSPML. TCSP is a template-based crystal structure prediction algorithm based on oxidation state patterns. CSPML is a machine learning-based crystal structure prediction method using a machine learning model to select templates. For a given new 2D formula such as SrTiO$_3$, both models will first select all template structures with prototype ABC3, but they are very different when sorting all these templates. TCSP calculates the element mover distance score and elements oxidation states, which focus on element distance. However, CSPML selects candidates using structural similarity. This structure similarity measure uses only the topological features of the atomic coordinates and does not use any information about the elemental composition.
After choosing the appropriate templates and generating new 2D structures, we use two machine learning potential-based relaxation algorithms to optimize the structures. The first one is BOWSR
which is a Bayesian optimization with symmetry relaxation. The second one is M3GNET, which uses materials graph neural networks with 3-body interactions as energy estimation model for structure relaxation. The BOWSR algorithm relaxes each structure by changing the independent lattice parameters and atomic coordinates to obtain lower potential energy. During relaxation, the M3GNET algorithm takes all atom coordinates and the 3X3 lattice matrix into consideration. The attributes of the bond, atom, and state are updated in order. For each attribute update, all previous attributes of these three parameters are considered.
After all these operations, for all generated 2D formulas, we obtain near-equilibrium relaxed structures. After the fast machine learning potential-based relaxation, we further apply the DFT-based relaxation procedure to optimize the structures. Finally, we calculate the formation energy and e-above-hull energy of top structures to evaluate the final performance.
\subsection{BLMM: Transformer based 2D material composition generation}
The material composition can be mapped into a sequence generation problem as a composition such as SiTiO$_3$ can be conveniently expanded into a specific sequence (e.g. Si Ti O O O) sorted by the electronegativities of the elements. The BLMM model is a composition generator built on the latest transformer deep neural network models, shown to be excellent on sequence learning and sequence generation.
By adopting the self-attention mechanism to weighting the significance of all tokens in the input sequence, the transformer model\cite{vaswani2017attention} has been proved as the state-of-the-art in the fields of natural language processing and computer vision. Based on the traditional transformer, Shen et.al.\cite{BLM} proposed a blank language model (BLM) which could generate sequences by dynamically creating and filling in blanks.
Our BLMM composition generator\cite{wei2022crystal} is developed based on the BLM blank-filling model.
All material formulas can be rewritten as sequences (e.g., SiTiO$_3$ to Si Ti O O O) composed of a vocabulary with 118 or fewer elements. We then train a BLMM based 2D composition generator using our 2D materials dataset. The architecture of the BLMM algorithm is shown in Figure\ref{fig:BLMM}. Generation starts with a single blank and ends when there is no blank. In each step, the model selects a blank, predicts a word w, and replaces the blank with the word w and possibly adjoining blanks. By repeating this blank selecting and filling process, a blank can be expanded to any number of words.
Then we use this well-trained BLMM model to generate new 2D compositions. After getting the generated compositions, we first remove duplicate compositions that are already included in known 2D datasets and then take the nonredundant formulas as our new 2D material candidates to be fed to the step of template-based 2D material structure prediction.
\begin{figure}[ht]
\centering
\begin{minipage}[c]{0.8\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/BLMM.png}
\end{minipage}\\
\caption{BLMM architecture \cite{wei2022crystal}.}
\label{fig:BLMM}
\end{figure}
\subsection{Template based 2D material structure prediction}
Currently, generic crystal structure prediction is still an unsolved problem despite that global optimization-based algorithms such as USPEX and CALYPSO can be applied to solve structures for small systems. On the other hand, we find that, similar to bulk materials\cite{mehl2017aflow,su2017construction,griesemer2021high}, most existing 2D material structures can be categorized into a very limited number of structure prototypes, which implies that their structures can be obtained using template-based elemental substitution.
After composition generation and duplicate checking, we obtain a 2D material composition candidates dataset. To gain the probable structures of all candidates, we use two different template-based element substitution methods to select the most similar structure template and then use element substitution to generate target structures. The crystal structure generated by these two methods has the same lattice parameters and atomic coordinates as the template structure and needs to be further relaxed.
\begin{figure}[ht]
\centering
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/tcsp.png}
\subcaption{}
\end{minipage}
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/cspml.png}
\subcaption{}
\end{minipage}\\
\caption{Template-based CSP algorithms. (a) TCSP architecture.\cite{wei2022tcsp}. (b) CSPML architecture\cite{CSPML}}
\label{fig:tsp}
\end{figure}
\paragraph{TCSP: is a template-based crystal structure prediction algorithm.}
The architecture of the TCSP algorithm is shown in Figure\ref{fig:tsp} (a). For a given candidate 2D material formula, the TCSP algorithm first searches all known 2D material structure templates that share the same composition prototype as this formula (e.g., SiTiO$_3$ has prototype ABC$_3$). The Element’s mover distance(ElMD)\cite{ElMD} is used to measure the compositional similarity between the query formula and compositions of all possible template structures. It then picks the top 5 structures with the smallest compositional distances as the candidate templates . For each of these candidate templates, we use the Pymatgen\cite{Pymatgen} package to check whether it has the same oxidation states as the query formula. Templates with identical oxidation states are then added to the final template list. If no such templates are found, all five of the top structures are taken as the final templates. Pymatgen’s StructureMatcher module is then used to reduce the redundant template structures. For each structure cluster, only one of these structures can be kept, which can significantly reduce the number of similar structure templates. After the templates of the query formula are designated, the algorithm then enumerates all of the possible element substitution pairs between the query and template formulas as Algorithm A2 in Ref\cite{wei2022tcsp}. It is possible that there are several element pair substitutions in one template to get the target formula. A replacement quality score is then calculated by summing the ElMD of all element pairs' substitution arrangements. This score represents how similar the substitution element pairs are. A lower score means higher similarity and thus higher quality.
\paragraph{CSPML is a machine learning-based crystal structure prediction algorithm.}
CSPML relies on metric learning\cite{Metric} for crystal structure prediction, which can select template structures from known structure databases with high similarity to the given composition. Metric learning uses a binary classifier to distinguish whether two given compositions have similar structures as defined by a similarity threshold of local structure order parameters (LoStOPs)\cite{LoStOPs}. The architecture of the CSPML algorithm is shown in Figure\ref{fig:tsp} (b). For a given 2D formula, CSPML first restricts the candidates to structures with the same compositional ratio (e.g., SiTiO$_3$ has a composition ratio of 1:1:3). The compositional descriptor of query formula and templates is then calculated by XenonPy\cite{XenonPy}. XenonPy provides 58 physicochemical features for each element. For a given composition, by calculating the weighted mean, weighted sum, weighted variance, min-pooling, and max-pooling of all elements, XenonPy generates a 290-dimensional (58x5) descriptor vector. A traditional multi-layer perceptron (MLP) is used to figure out how similar the template structure and the query formula are. The absolute difference between two compositional descriptors is used as the input. We pick the top five template structures with the biggest similarity scores with the query formula as the template candidates. The structure of the query formula is then generated by replacing the atoms in the templates with atoms in the query composition. When two or more elements have the same composition ratio, the substitution element pairs are not uniquely determined. In such cases, we substitute a pair of elements with the most similar physicochemical properties, as described in Ref\cite{CSPML}.
\subsection{Structure relaxation}
Accurately predicting novel stable crystal structures and their properties is a fundamental goal in computation-guided materials discovery. While ab initio approaches such as density functional theory (DFT) have been phenomenally successful in this regard, their high computational cost and poor scalability have limited their broad application across the vast chemical and structural spaces.
To circumvent this limitation, machine learning has emerged as a new paradigm for developing efficient surrogate models for predicting material properties at scale. In this paper, after gaining basic structures from template-based element substitution methods, we apply and compare two different machine-learning potential-based structure relaxation methods.
\paragraph{BOWSR: Bayesian optimization with symmetry relaxation algorithm}
BOWSR is a graph-based neural network-based structure relaxation algorithm that uses Bayesian optimization as an optimizer. Bayesian optimization is an adaptive strategy for the global optimization of functions. In the crystal structure relaxation problem, the target function that needs to be optimized is the potential energy surface, which describes the energy of the crystal structure. During the relaxation process of the BOWSR algorithm, the symmetry of the lattice and the Wyckoff positions of the atoms are limited. Only the independent lattice parameters and atomic coordinates are allowed to change. The BOWSR algorithm sets parameters for each structure based on these changeable, independent lattice parameters and atomic coordinates. The potential energy surface of each training observation is calculated by a graph neural network (MEGNet) energy model, which is trained with 12,277 stable structures with DFT-calculated formation energies. Bayesian optimization is then used to relax structures iteratively towards states with lower energies.
The geometry relaxation of a structure of N atoms requires optimizing 3N + 6 variables, 3 fractional coordinates for each atom, and 6 lattice parameters in total. By keeping the symmetry the same during the relaxation process, it can reduce the number of independent variables. New structures are generated by Bayesian optimization, which minimizes the formation energy, and the changed variables are then used as inputs to predict new energy. The previous step is then repeated multiple times until the formation energy reaches the lowest point or reaches the maximum number of iterations. The final structure is the BOWSR relaxation results.
\paragraph{M3GNET: materials graph neural networks with 3-body interactions}
M3GNet (M3GNET) is a new materials graph neural network architecture that incorporates 3-body interactions for formation energy prediction. It combines graph-based deep learning interatomic potential (IAP) and many-body features of traditional IAPs with those of flexible graph material representations. The inputs of the M3GNet model are position-included graphs. The atomic numbers of elements and the pair bond distance in the input graph are embedded as graph features. The three-body and many-body interaction atom indices and angles are calculated by the many-body computation module. The bond and atom information is then updated through a graph convolution module.
A key difference with prior materials graph implementations such as MEGNet is the addition of the coordinates for atoms and the 3x3 lattice matrix in crystals, which are necessary for obtaining tensorial quantities such as forces and stresses via auto-differentiation. The difference of M3GNet with BOWSR's GNN potential is that it is trained with both stable structures and unstable structures. The M3GNet-based relaxation algorithm \cite{M3GNET} is also different from BOWSR's Bayesian optimization. It uses an algorithm named FIRE, which is derived from molecular dynamics with additional velocity modifications and adaptive time steps and inertia to achieve fast inertial relaxation engine.
The ability of a M3GNet-based relaxing algorithm to accurately and rapidly relax arbitrary crystal structures and predict their energies makes it ideal for large-scale materials discovery.
\subsection{DFT calculations}
We carried out the first-principles calculations based on the density functional theory (DFT) using the Vienna \textit{ab initio} simulation package (VASP)\cite{Vasp1,Vasp2,Vasp3,Vasp4} to optimize the candidate structures suggested by the machine learning models. The projected augmented wave (PAW) pseudopotentials were used to treat the electron-ion interactions\cite{PAW1, PAW2} with 520 eV plane-wave cutoff energy. The generalized gradient approximation (GGA) based Perdew-Burke-Ernzerhof (PBE) method was considered for the exchange-correlation functions \cite{GGA1, GGA2}. The energy convergence criterion was 10$^{-5}$ eV and the force convergence criterion was 10$^{-2}$ eV/{\AA} for all the DFT calculations. The Brillouin zone integration for the unit cells was performed employing the $\Gamma$-centered Monkhorst-Pack $k$-meshes. The formation energies (in eV/atom) of the materials were determined employing the formula in Eq.~\ref{eq:form}, where $E[\mathrm{Material}]$ is the total energy per unit formula of the target structure, $E[\textrm{A}_i]$ is the energy of $i^\mathrm{th}$ element of the material, $x_i$ indicates the number of A$_i$ atoms in a unit formula, and $n$ is the total number of atoms in a unit formula($n=\sum_i x_i$). The Pymatgen code\cite{Pymatgen} was used to compute the energy above hull values of the materials with negative formation energies.
\begin{equation}
E_{\mathrm{form}} =\frac{1}{n}(E[\mathrm{Material}] - \sum_i x_i E[\textrm{A}_i])
\label{eq:form}
\end{equation}
\subsection{Evaluation criteria}
We use a series of performance metrics to evaluate our 2D material generation pipeline.
To evaluate the BLMM 2D material composition generator, we calculate the validity, uniqueness, recovery rate, and novelty. Formation energy is used as an indicator to evaluate the template-based 2D structure generator and relaxer. To further verify the structures, we use VASP to calculate the energy-above-the-hull.
\textbf{Validity.} For all formulas generated by the BLMM algorithm, we use Semiconducting Materials by Analogy and
Chemical Theory(SMACT)\cite{davies2019smact} to check whether they obey the charge neutrality and electronegativity (CNEN) rules.
\textbf{Uniqueness.} Uniqueness percentage is calculated by using the number of unique samples divided by the total generated samples. The uniqueness indicator shows the BLMM model's ability to generate diverse samples.
\textbf{Recovery Rate and Novelty.} To check the BLMM model's capability to generate novel materials, we calculate the recovery rate and novelty of generated formulas. The recovery rate shows the percentage of training samples that have been rediscovered. Novelty shows how many new samples have been generated.
\textbf{Formation energy.} The way to evaluate the structure generation models is to check the stability of generated structures. For structures generated and relaxed through our pipeline, we calculate their formation energy using M3GNET.
\textbf{Energy above convex hull.} The energy convex hull \cite{liu2015spinel} is generated based on existing stable structures. Structures with energy lying on the convex hull are thermodynamically stable, and the ones above it are either metastable or unstable. For all structures with negative formation energy, we use the energy above convex hull as a further filter to select more stable structures.
\subsection{Hyperparameters and training}
For 2D formula generation, each BLMM model trained on 2D datasets generates 100,000 samples. After generation, we use the TCSP and CSPML methods separately to generate structure candidates for these samples. BOWSR and M3GNET are then used to relax generated structures. Table\ref{table:hyperparameter} shows the hyperparameters used in the BLMM, TCSP, CSPML, BOWSR, and M3GNET models. In the BLMM model, we use an element vocabulary with size 130, and the generated formula sequence length is limited to 205. The maximum number of tokens per batch is set to 40,000, and the number of training steps is set to 200,000.
The candidate template structure number of both the TCSP and CSPML models is set to 10. And only the top 5 candidates, sorted by ElMD and XenenPy, respectively, can be used as real templates. Relax method BOWSR uses a Bayesian optimizer with initial points 1000, iteration steps 1000, and seed number 42. The M3GNET relax method uses FIRE\cite{FIRE} optimizer with a 0.1 total force tolerance for relaxation convergence and 500 relax steps.
\begin{table}[h]
\begin{center}
\caption{Hyperparameters used in models.}
\label{table:hyperparameter}
\begin{tabular}{|cc|cc|cc}
\hline
\multicolumn{2}{|c|}{TCSP} & \multicolumn{2}{c|}{CSPML} & \multicolumn{2}{c|}{BLMM} \\ \hline
\multicolumn{1}{|c|}{candidate} & 10 & \multicolumn{1}{c|}{candidate} & 10 & \multicolumn{1}{c|}{data workers} & \multicolumn{1}{c|}{32} \\ \hline
\multicolumn{1}{|c|}{top} & 5 & \multicolumn{1}{c|}{top} & 5 & \multicolumn{1}{c|}{max steps} & \multicolumn{1}{c|}{200000} \\ \hline
\multicolumn{1}{|c|}{sort} & ElMD & \multicolumn{1}{c|}{sort} & XenonPy & \multicolumn{1}{c|}{max token} & \multicolumn{1}{c|}{40000} \\ \hline
\multicolumn{1}{|c|}{filter} & ratio & \multicolumn{1}{c|}{filter} & ratio & \multicolumn{1}{c|}{vocab size} & \multicolumn{1}{c|}{130} \\ \hline
\multicolumn{1}{|c|}{filter} & oxidation & \multicolumn{1}{c|}{} & & \multicolumn{1}{c|}{max len} & \multicolumn{1}{c|}{205} \\ \hline
\multicolumn{2}{|c|}{BOWSR} & \multicolumn{2}{c|}{M3GNET} & & \\ \cline{1-4}
\multicolumn{1}{|c|}{optimizer} & Bayesian Optimizer & \multicolumn{1}{c|}{optimizer} & FIRE & & \\ \cline{1-4}
\multicolumn{1}{|c|}{initial points} & 100 & \multicolumn{1}{c|}{force tolerance} & 0.1 & & \\ \cline{1-4}
\multicolumn{1}{|c|}{iteration steps} & 100 & \multicolumn{1}{c|}{relax steps} & 500 & & \\ \cline{1-4}
\multicolumn{1}{|c|}{seed} & 42 & \multicolumn{1}{c|}{} & & & \\ \cline{1-4}
\end{tabular}
\end{center}
\end{table}
\section{Results}
\subsection{Datasets}
As shown in Table\ref{table:dataset}, our template-based 2D materials generation models are trained using the materials downloaded from the C2DB\cite{C2DB1,C2DB2}, MC2D\cite{MC2D}, 2DMatPedia\cite{2DMatPedia,2DMatPedia2}, and V2DB\cite{V2DB} databases with a total of 328,719 formula samples and 12,214 structures. The C2DB dataset was initially generated by decorating an experimentally known crystal structure prototype with atoms chosen from a (chemically reasonable) subset of the periodic table. The MC2D dataset starts from experimentally known 3D compounds and finds 1,825 compounds that are either easily or potentially exfoliate. The 2DMatPedia dataset is searched from the Materials Project database\cite{MP} and uses exfoliate first to find possible 2D structures, and new structures generated by exfoliation are then used as templates of element substitution. The generation of the V2DB dataset employs the brute-force element substitution method. This method generated 72,522,240 possible combinations of 2D materials and only 0.4\% of these passed the symmetry, neutrality, and stability validation.
In this work, we separated these 2D samples into two datasets. The first one is an experimental dataset (exp2d for short) with 4,023 formulas and corresponding structures from the C2DB and MC2D datasets. The second dataset (all2d for short) contains the all the samples in the above-mentioned our known 2D databases with a total of 302,174 unique formulas and 8,019 structures.
\begin{table}[h]
\begin{center}
\caption{Open source datasets used in 2D material discovery.}
\label{table:dataset}
\resizebox{\columnwidth}{!}{\begin{tabular}{|c|c|c|c|c|cc|}
\hline
Dataset & Formula & Structure & Type & From & \multicolumn{1}{c|}{Exfoliation} & Substitution \\ \hline
C2DB & 4038 & 4038 & experimental & known 2D crystal structure prototype & \multicolumn{1}{c|}{N/A} & all \\ \hline
MC2D & 1825 & 1825 & experimental & known 3D compounds & \multicolumn{1}{c|}{all} & N/A \\ \hline
2DMatPedia & 6351 & 6351 & calculate & Material Project & \multicolumn{1}{c|}{2940} & 3409 \\ \hline
V2DB & 316505 & N/A & calculate & 22 known 2D crystal prototypes & \multicolumn{1}{c|}{N/A} & all \\ \hline
\end{tabular}}
\end{center}
\end{table}
\subsection{Composition generation performance}
We use the BLMM algorithm to generate new 2D material compositions based on two different datasets, the experimental 2D dataset, exp2d, and an all 2D dataset, all2d. For each dataset, we train a generation model using these formulas and then use these well-trained models to generate new formulas. We also use transfer learning to train BLMM models on the materials project database and then fine-tune these pre-trained models using our two datasets, which are named all2d-transfer and exp2d-transfer, respectively. Four composition generation models are trained to generate 100,000 formulas separately. The generated results are shown in Table\ref{table:composition}. Furthermore, to check whether generated formulas are chemically valid, we employ two filters to check their charge neutrality (CN) and electronegativity (EN), this checking step is called CNEN for short. The results are shown in Table\ref{table:composition}, 93.3\%, 67.1\%, 93.2\%, and 67.0\% generated compositions passed the CNEN check. Besides, we remove duplicate composition in each model, and they achieve 65.8\%, 26.5\%, 69.4\%, and 21.0\% uniqueness respectively. We also calculate the recovery rate and novelty of these generational models. As we can see in the Table \ref{table:composition}, their recovery rates are 0.6\%, 9.2\%, 0.6\%, and 11.2\% while the novelties are 63.7\%, 24.6\%, 67.5\%, and 18.8\%. This evaluation demonstrates that our methods have the ability to generate stable and innovative compositions that form stable 2D structures.
Since the exp2d dataset is smaller than the all2d dataset, the BLMM model trained with the exp2d dataset has much fewer samples to learn from and use for interpolation. Thus, the BLMM model trained with the exp2d dataset has lower validity, uniqueness, and novelty percentages than the BLMM model trained with the all2d dataset. However, the recovery rate of the BLMM model trained with fewer samples is higher because the interpolation space is smaller and the interpolated results generated by the BLMM model are more likely to be the same as the training samples.
The composition generator pre-trained with the materials project database and fine-tuned with the all2d dataset achieves higher uniqueness and novelty compared with the generator only trained with the all2d dataset. However, due to the lack of sufficient transfer learning samples, the BLMM model fine-tuned with the exp2d dataset has lower uniqueness and novelty compared with the BLMM model trained with the exp2d dataset.
\begin{table}[h]
\begin{center}
\caption{Composition generation results.}
\label{table:composition}
\begin{tabular}{|c|c|c|c|c|}
\hline
& all2d & exp2d & all2d-transfer & exp2d-transfer \\ \hline
Generated results & 100000 & 100000 & 100000 & 100000 \\ \hline
Validity & 93.3\% & 67.1\% & 93.2\% & 67.0\% \\ \hline
Uniqueness & 65.8\% & 26.5\% & 69.4\% & 21.0\% \\ \hline
Recover rate & 0.6\% & 9.2\% & 0.6\% & 11.2\% \\ \hline
Novelty & 63.7\% & 24.6\% & 67.5\% & 18.8\% \\ \hline
\end{tabular}
\end{center}
\end{table}
\FloatBarrier
\subsection{Distribution of generated candidate 2D compositions}
To check the composition generation performance of BLMM, we plot the element distribution of compositions in the 2DMatPedia dataset and our generated samples in Figure\ref{fig:element distribution}, where (a) and (b) show the element frequency in compositions in the 2DMatPedia dataset and the BLMM model generated dataset, respectively. Here we take the BLMM model trained with the exp2d dataset as an example. The top 5 most frequent elements in the 2DMatpedia dataset are O, S, F, Te, and Cl. Out of the total 6,351 formulas, element O has appeared 1,642 times, or about 26\% of the 2Dmatpedia dataset. The occurrences of the elements S, F, Te, and Cl are 653, 598, 586, and 584, respectively. The top 5 most frequent elements in our generated dataset are Se, O, S, Te, and Cl. The element Se has shown 13,294 times out of the whole set of 67,103 formulas, which is about 20\% of the whole generated dataset. The elements O, S, Te, and Cl appear 12,591, 11,440, 9,546, and 8,829 times respectively. Overall, we find that four out of the top five most common elements in these two datasets are the same, and six in the top ten most common elements are also the same, indicating that the BLMM generator has learned the key composition preferences of 2D materials.
\begin{figure}[ht]
\centering
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/2dmatpedia_ele_bar.png}
\subcaption{}
\end{minipage}
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/exp2d_ele_bar.png}
\subcaption{}
\end{minipage}
\caption{Elements distribution in training and generating samples. (a)Element distribution in the 2DMatPedia dataset. (b)Element distribution in BLMM model generated samples.}
\label{fig:element distribution}
\end{figure}
We also analyze the distribution of element pairs in the known 2D dataset and our generation results. To count the frequencies of element pairs, we take each of the possible 2-element combinations from the element set and count the number of compositions that contain this pair (we ignore the order of the two elements in the pair). The distribution of the top 50 element pairs in the 2DMatPedia and our generation datasets are shown in Figure\ref{fig:element pair} (a) and (b), respectively.
The top 5 most frequent element pairs in the 2DMatpedia dataset are H-O, P-O, Li-O, V-O, and Bi-O. However, only the H-O element pair is shown in the top 5 of our generation results. The other 4 element pairs in our generation results are C-O, N-O, Cl-O, and H-C. These two datasets only share 2 common element pairs in the top 10 most frequent ones.
\begin{figure}[h]
\centering
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/top50_2dmatpedia_element_pairs.png}
\subcaption{}
\end{minipage}
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/top50_exp2d_element_pairs.png}
\subcaption{}
\end{minipage}
\caption{Top 50 element pairs distribution in training and generating samples. (a)Element pairs distribution in 2DMatPedia dataset. (b) Element pairs distribution in generated samples by BLMM model trained with exp2d dataset.}
\label{fig:element pair}
\end{figure}
To verify whether our newly generated compositions share a similar distribution with the known 2D compositions, we use the t-distributed Stochastic Neighbor Embedding(t-SNE)\cite{JMLR:v9:vandermaaten08a} technique to map the one-hot matrix of compositions to their corresponding formation energy. Each point in Figure\ref{fig:tsne} corresponds to one formula and the colors represent the formation energy levels. Figure\ref{fig:tsne} (a) shows the formation energy distribution of the 2DMatPedia samples. It can be found that most samples have formation energy between 0 and -3 eV/atom. Figure\ref{fig:tsne} (b) displays the formation energy distribution of our generated compositions by BLMM-exp2d, which are developed through the following pipeline: firstly, we train the BLMM model using the exp2d dataset and generate compositions; next, we generate candidate structures using the TCSP method and then relax these structures using the M3GNET model; thirdly, the formation energies of these structures are predicted by the M3GNET method. As the BLMM model is trained by adding and filling in blanks in existing materials, it has a strong interpolation capability when generating new samples. Therefore, the newly generated samples are always located around known samples.
\begin{figure}[h]
\centering
\begin{minipage}[c]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/2dmatpedia-tsne.png}
\subcaption{}
\end{minipage}
\begin{minipage}[c]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/mixed-tsne.png}
\subcaption{}
\end{minipage}
\caption{Formation energy distributions in the training (2DMatPedia) and generated samples(BLMM-exp2d). (a) Formation energy distribution in the 2DMatPedia dataset. (b) Formation energy distribution in generated samples.}
\label{fig:tsne}
\end{figure}
\subsection{Stability distribution of generated samples}
Another way to check the quality of samples generated by our pipeline is to measure their formation energies and compare their distribution to that of the training set. We use formation energies predicted by the ML potentials M3GNET of both training samples and generated results.
We first check the formation energy distribution of a special material family AB$_2$, which is the most frequent prototype in all existing 2D datasets: C2DB, MC2D, and 2DMatPedia. There are 1,288 AB$_2$ samples in the exp2d dataset, and 1,928 AB$_2$ samples in our generated structures. The distributions of energies of these two datasets are shown in Figure\ref{fig:formation energy distribution}(a).
Next, we check the structure-based formation energy distribution of the whole exp2d dataset samples and compare it with those of our generated samples. Figure\ref{fig:formation energy distribution}(b) shows that these two set of structures have similar formation energy, which means that new structures generated through our MTG pipeline are of high quality.
Figure\ref{fig:formation energy distribution}(c) compares the energy distribution of samples in the exp2d training set with samples generated through out MTG pipeline. These compositions are generated by the BLMM model trained with four different datasets, as introduced in section 3.1. The energy distributions of formulas generated by BLMM trained with the all2d dataset and trained with the MP dataset but finetuned using the all2d dataset are very similar to each other. Same situation in BLMM models trained and finetuned with the exp2d dataset.
\begin{figure}[h]
\centering
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/ab2_violinplot.png}
\subcaption{}
\end{minipage}
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/exp2d_violinplot.png}
\subcaption{}
\end{minipage}\\
\begin{minipage}[c]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/total_violinplot.png}
\subcaption{}
\end{minipage}\\
\caption{Formation energy per atom distribution. (a) Formation energy distribution of AB$_2$ type structures in the exp2d dataset and generated through our MTG pipeline (predicted by M3GNET). (b) Formation energy distribution of structures in the exp2d dataset and structures generated by our MTG-exp2d pipeline (formation energy predicted by M3GNET). (c) Formation energy distribution of compositions in the exp2d dataset and structures generated by our MTG pipelines (BLMM models trained by all four datasets, formation energies predicted by M3GNET).}
\label{fig:formation energy distribution}
\end{figure}
\FloatBarrier
\subsection{Discovery results}
Our MTG pipeline generates 148,563 candidate 2D formulas. For each formula, we generate 10 structures using TCSP and CSPML and then we do M3GNET based structure relaxation and we pick the top 1 structure with the lowest energy. Then we conduct DFT-based relaxation to generate final structures.
Figure\ref{fig:newstructure} shows how we generated new structures based on specific template structures and how to relax newly generated structures to make them more stable. For formula K$_4$Cr$_2$Ge$_4$Te$_2$ generated by the BLMM algorithm, we first select the structure templates for predicting its crystal structure. As shown in Figure\ref{fig:newstructure} (a), TCSP picked Na$_4$Ti$_2$S$_4$O$_2$, a layered material, as the template structure. Figure\ref{fig:newstructure} (b) is then created through the TCSP algorithm. After relaxing by M3GNET, we get a more stable structure as shown in Figure\ref{fig:newstructure} (c). This relaxed structure is then sent to VASP to do further DFT relaxation and energy calculations. We can find that Figure\ref{fig:newstructure} (a) and (b) are more similar as in (b) only elemental substitutions are applied in the structure (a) with no atomic coordinate fine-tuning. Figure\ref{fig:newstructure} (c) changed the coordinates based on atom sizes and bond types to make this structure more structurally stable.
A similar procedure is applied to discover the structure of K$_4$Cr$_2$Sn$_4$, via the three steps as shown in Figure \ref{fig:newstructure} (d,e,f).
\begin{figure}[ht]
\centering
\begin{minipage}[c]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/4755.png}
\subcaption{Template structure Na$_4$Ti$_2$S$_4$O$_2$}
\end{minipage}
\begin{minipage}[c]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/4755_generated.png}
\subcaption{Generated structure K$_4$Cr$_2$Ge$_4$Te$_2$}
\end{minipage}
\begin{minipage}[c]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/4755_relaxed.png}
\subcaption{Relaxed structure K$_4$Cr$_2$Ge$_4$Te$_2$}
\end{minipage}\\
\begin{minipage}[c]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/5132.png}
\subcaption{Template structure Na$_2$Cl$_2$O$_4$}
\end{minipage}
\begin{minipage}[c]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/5132_generated.png}
\subcaption{Generated structure K$_4$Cr$_2$Sn$_4$}
\end{minipage}
\begin{minipage}[c]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/5132_relaxed.png}
\subcaption{Relaxed structure K$_4$Cr$_2$Sn$_4$}
\end{minipage}\\
\caption{Illustration of the structure generation and relaxation process of our MTG pipeline. (a) to (c) shows the template structure, structure upon element substitution, and the fine-tuned structure after ML potential based relaxation for predicting the structure of K$_4$Cr$_2$Ge$_4$Te$_2$. (d) to (f) shows the similar process for K$_4$Cr$_2$Sn$_4$.}
\label{fig:newstructure}
\end{figure}
Figure\ref{fig:E-above-hull} shows four new 2D structures discovered through our MTG pipeline that have 0 e-above-hull energy (See cif information in Supplementary file). All the structures show a layered structure with each layer forming a compact 2D structure, demonstrating that they have passed the DFT stability check and the capability of our generative 2D materials design pipeline to find new 2D materials.
\begin{figure}[ht]
\centering
\begin{minipage}[c]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/NiCl4.png}
\subcaption{NiCl$_4$}
\end{minipage}
\begin{minipage}[c]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/IrSBr.png}
\subcaption{IrSBr}
\end{minipage}
\begin{minipage}[c]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/CuBr3.png}
\subcaption{CuBr$_3$}
\end{minipage}
\begin{minipage}[c]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/CoBrCl.png}
\subcaption{CoBrCl}
\end{minipage}\\
\caption{Four new 2D structures discovered by our MTG pipeline with 0 E-above-hull energy.}
\label{fig:E-above-hull}
\end{figure}
\FloatBarrier
\section{Conclusion}
Two-dimensional materials have wide applications due to their unique properties. Here we propose a generative design pipeline for 2D materials discovery by integrating a transformer-based 2D material composition generator, two template-based crystal structure predictors, and a graph neural network potential-based structure relaxation algorithm. It is found that the transformer composition generator can capture the composition preference which allows it to generate chemically valid potential 2D materials. We have applied our 2D generator pipeline to discover four hypothetical 2D materials with e-above-hull energy less than 0. Our pipeline is generic and can be used to train other types of materials' generative design models.
\section{Code Availability}
The BLMM material composition generator is accessible at \href{https://github.com/usccolumbia/blmm}{https://github.com/usccolumbia/blmm}. The TCSP template-based structure predictor can be accessed at \href{http://materialsatlas.org/crystalstructure}{http://materialsatlas.org/crystalstructure}. The CSPML template-based structure predictor can be found at \href{https://github.com/Minoru938/CSPML}{https://github.com/Minoru938/CSPML}. The BOWSR structure relaxation algorithm can be found at \href{https://github.com/materialsvirtuallab/maml}{https://github.com/materialsvirtuallab/maml}. M3GNet relaxation module can be downloaded from \href{https://github.com/materialsvirtuallab/m3gnet}{https://github.com/materialsvirtuallab/m3gnet}.
\section{Data Availability}
The V2DB database is available at Harvard Dataverse \href{https://doi.org/10.7910/DVN/SNCZF4}{https://doi.org/10.7910/DVN/SNCZF4}. The C2DB database is available at \href{https://cmrdb.fysik.dtu.dk/c2db/}{https://cmrdb.fysik.dtu.dk/c2db/}. The MC2D database is available at \href{https://www.materialscloud.org/discover/mc2d/dashboard/ptable}{https://www.materialscloud.org/discover/mc2d/dashboard/ptable}. The 2DMatPedia database is available at \href{http://www.2dmatpedia.org/}{http://www.2dmatpedia.org/}.
\section{Contribution}
Conceptualization, J.H.; methodology,R.D., J.H. Y.S.,E.S.; software, R.D., Y.S. ; resources, J.H.; writing--original draft preparation, R.D., E.S., J.H.; writing--review and editing, J.H; visualization, R.D. and E.S.; supervision, J.H.; funding acquisition, J.H.
\section*{Acknowledgement}
The research reported in this work was supported in part by National Science Foundation under the grant and 1940099 and 1905775. The views, perspectives, and content do not necessarily represent the official views of the NSF.
\bibliographystyle{unsrt}
|
\section{Introduction}
Causal models are conceptual constructs we use in our everyday understanding of the world and in scientific modelling. Structural causal models (SCM) provide a mathematical formalism to express causal assumptions, encode quantities of interest, and reason about relationships of cause and effect. For instance, a research lab $L$ may investigate lung cancer, and decide to model this scenario considering causal connections between a set of relevant variables, such as the smoking habits of patients, the presence of tar deposits in their lungs, and whether they ended up developing lung cancer (see Figure \ref{fig:SCMs}(left)).
Another common feature of reasoning and scientific modelling is reliance on multiple levels of abstraction, whereby an identical system can be studied at multiple levels of granularity. For instance, in studying lung cancer, another research lab $L'$ may decide to record only two variables, ignoring the contribution of tar deposits (see Figure \ref{fig:SCMs}(right)).
While SCMs allow us to deal with causal relationships internal to a given model, an abstraction focuses on \textit{external} relations between different models. The idea of abstraction is widespread in artificial intelligence: it underlies the very notion of intelligence \citep{mitchell2021abstraction}, it has been invoked to explain the success of deep learning \citep{lecun2015deep}, and it has a central role in causal representation learning \citep{scholkopfLBKKGB21}. However, rigorous formalisms for abstractions have only been recently proposed \citep{rubenstein2017causal,beckers2019abstracting,rischel2020category}.
Our work starts from the abstraction framework of \cite{rischel2020category}, which provides a grounded way to express an abstraction between two SCMs and to quantitatively assess its consistency. Evaluation of consistency requires, beyond the definition of the SCMs, a full specification of an abstraction, which, in reality, may rarely be available. In this paper, we consider the problem of \textit{learning an abstraction} when only partial information about it is available. This would correspond, for instance, to the case in which two labs are aware of their own SCMs but do not have an exact mapping between them. Successfully learning an abstraction would enable them to automatically transfer data and results across the models; in a low-data regime, where collecting new samples may be costly, having a proper abstraction would allow them to integrate their evidence and improve inferences.
Our contributions are introducing a new learning problem with specific semantics (abstraction learning) but a very generic syntax (commutativity learning); proposing a relaxation of the ensuing combinatorial problem and a solution based on differentiable programming which jointly solves a number of combinatorial sub-problems at once; analyzing empirically the performance of this approach on synthetic settings; demonstrating the benefits of our approach on the important problem of learning coating models for the batteries of electric vehicles (EV) by learning an abstraction that allows us to relate small-scale datasets collected through expensive real-world experiments performed in labs across France and the UK. To the best of our knowledge, this work also constitutes the first real-world application of an abstraction learning framework between SCMs.
The rest of the paper is organized as follows. Section \ref{sec:Background} reviews important background definitions, and Section \ref{sec:RelatedWork} presents related work. Section \ref{sec:ProblemStatement} discusses the learning problem. Section \ref{sec:Methodology} introduces our proposed methodology,
and Section \ref{sec:Experiments} presents our empirical results. Section \ref{sec:Discussion} summarizes our work and offers considerations on our approach and results. Appendix \ref{app:Notation} offers a summary of the notation used throughout this paper.
\begin{figure}[H]
\begin{adjustbox}{center}
\begin{tikzpicture}[shorten >=1pt, auto, node distance=1cm, thick, scale=0.8, every node/.style={scale=0.8}]
\tikzstyle{node_style} = [circle,draw=black]
\node[node_style] (S) at (0,0) {S};
\node[node_style] (T) at (2,0) {T};
\node[node_style] (C) at (4,0) {C};
\draw[->] (S) to (T);
\draw[->] (T) to (C);
\end{tikzpicture}
\hspace{3cm}
\begin{tikzpicture}[shorten >=1pt, auto, node distance=1cm, thick, scale=0.8, every node/.style={scale=0.8}]
\tikzstyle{node_style} = [circle,draw=black]
\node[node_style] (S) at (0,0) {S'};
\node[node_style] (C) at (2,0) {C'};
\draw[->] (S) to (C);
\end{tikzpicture}
\end{adjustbox}
\caption{Lung cancer SCM designed by lab $L$ (left) and lab $L'$ (right).} \label{fig:SCMs}
\end{figure}
\section{Background} \label{sec:Background}
In this section we provide definitions for the main concepts related to causality and abstraction.
\subsection{Causality}\label{ssec:Causality}
In order to express causal models, we rely on Pearl's formalism of structural causal models \citep{pearl2009causality,peters2017elements}. See \cite{pearl2009causality} for the analogies and differences to the potential outcomes framework by Rubin \citep{rubin2005causal}.
\begin{definition}[Structural Causal Model (SCM) \citep{pearl2009causality}]
A structural causal model $\scm{}$ consists of a directed acyclic graph (DAG) $\mathcal{G}_{\scm{}}$, and a tuple $\langle \mathcal{U}, \mathcal{X}, \mathcal{F}, P(\mathcal{U})\rangle$, where:
\begin{itemize}
\item $\mathcal{U}$ is a finite set of exogenous (unobserved or latent) variables;
\item $\mathcal{X}$ is a finite set of endogenous (observed) variables, each one with a discrete domain $\mathcal{M}[X_i]$;
\item $\mathcal{F} = \{f_{1}, . . . , f_{|\mathcal{X}|}\}$ is a set of modular functions (mechanisms) such that $x_i = f_i (pa(X_i))$, where $x_i$ is the value of an endogenous variable $X_i \in \mathcal{X}$ and $pa(X_i) \subseteq (\mathcal{U} \cup \mathcal{X}) \setminus X_i$.
\item $P(\mathcal{U})$ is a joint distribution over the exogenous variables.
\end{itemize}
\end{definition}
Notice that the assumptions of acyclicity and joint distribution over the exogenous variables imply a \emph{semi-Markovian SCM}; in this setting, the set $pa(X_i)$ of variables that determines the value of an endogenous variable $X_i$ can be given the graph-theoretic interpretation of \emph{parents} of $X_i$, and edges can be interpreted as causal relations. The additional assumption of discrete domains for each endogenous variable is introduced as a requirement of the abstraction framework of \cite{rischel2020category}.
Models presented in Figure \ref{fig:SCMs} can be given a rigorous reading as DAGs $\mathcal{G}_{\scm{}}$ and $\mathcal{G}_{\scm{'}}$.
A SCM allows us to study a system not only in the observational domain, but also under interventions. Modularity of the mechanisms allows us to formally define interventions.
\begin{definition}[Intervention \citep{pearl2009causality}]
Given a SCM $\scm{}=\langle \mathcal{U}, \mathcal{X}, \mathcal{F}, P(\mathcal{U})\rangle$, an intervention $\iota \coloneqq do(X_i = x_i)$ is an operator that generates a new post-interventional SCM $\scm{_\iota}=\langle \mathcal{U}_\iota, \mathcal{X}_\iota, \mathcal{F}_\iota, P_\iota(\mathcal{U}_\iota)\rangle$ where $\mathcal{U}_\iota = \mathcal{U}$,
$\mathcal{X}_\iota = \mathcal{X}$,
$P_\iota(\mathcal{U}_\iota)=P(\mathcal{U})$, while
$f_{\iota_i} = x_i$ and $f_{\iota_j} = f_j$ for all $j \neq i$.
\end{definition}
Thus, an intervention $\iota \coloneqq do(X_i = x_i)$ creates a new model $\scm{_\iota}$, identical to the original one except for the structural function $f_i$ which is replaced with the constant $x_i$; the DAG of the post-interventional model $\scm{_\iota}$ is similarly identical, with the exception of the node associated with $X_i$ which has all the incoming edges removed. It is immediate to extend the definition of intervention to multiple endogenous variables: $do(\mathbf{X} = \mathbf{x})$, where $\mathbf{X} = [X_1, X_2,..., X_m]$ is a vector of variables in $\mathcal{X}$ and $\mathbf{x} = [x_1, x_2,..., x_m]$ is a vector of values associated with each variable.
\subsection{Abstraction}\label{ssec:Abstraction}
We introduce a notion of abstraction meant to relate two SCMs representing an identical system. This definition originates from category theory, and it assumes SCMs with a finite set $\mathcal{X}$ of variables, each one defined on a finite and discrete domain $\mathcal{M}[X_i]$ \citep{rischel2020category}.
\begin{definition}[Abstraction \citep{rischel2020category}]
Given two SCMs $\scm{}$ and $\scm{'}$, an abstraction $\boldsymbol{\alpha}$ is a tuple $\langle {R}, {a}, \alpha \rangle$ where:
\begin{itemize}
\item $R \subseteq \mathcal{X}$ is a subset of relevant variables in the model $\mathcal{M}$;
\item ${a} : R \rightarrow \mathcal{X'}$ is a surjective map between variables, from nodes in $\mathcal{M}$ to node in $\mathcal{M'}$;
\item $\alpha$ is a collection of surjective maps $\alpha_{X'}: \mathcal{M}[{a}^{-1}(X')] \rightarrow \mathcal{M'}[X']$ from outcomes of variables in $\mathcal{M}$ to outcomes of variables in $\mathcal{M'}$.
\end{itemize}
\end{definition}
An abstraction establishes an asymmetric relation from a base or low-level model $\scm{}$ to an abstracted or high-level model $\scm{'}$. This definition encodes a mapping on two layers: on a structural or graphical level between the nodes of the DAGs via $a$, and on a distributional level via the maps $\alpha_{X'}$ \citep{zennaro2022abstraction}.
Since we are dealing with causal models, we require the SCMs to behave consistently wrt interventions.
\begin{definition}[Consistency wrt an interventional distribution]\label{def:AbstractionConsistency}
Let $\boldsymbol{\alpha}$ be an abstraction from $\scm{}$ to $\scm{'}$. Let $\mathbf{X'}$ and $\mathbf{Y'}$ be two disjoint subsets of variables in $\mathcal{X'}$. The abstraction $\boldsymbol{\alpha}{}$ is consistent wrt the interventional distribution $P'(\mathbf{Y'} \vert do(\mathbf{X'}))$ if the following diagram commutes:
\begin{adjustbox}{center}
\begin{tikzpicture}[thin,,scale=0.8, every node/.style={scale=0.8}]
\node[] (v1){$\mathcal{M}[\mathrm{a}^{-1}(\mathbf{X'})]$};
\node[right= 1.5cm of v1] (v2){$\mathcal{M}[\mathrm{a}^{-1}(\mathbf{Y'})]$};
\node[below= 1.5cm of v1] (v3){$\mathcal{M'}[\mathbf{X'}]$};
\node[below= 1.5cm of v2] (v4){$\mathcal{M'}[\mathbf{Y'}]$};
\draw[arr1] (v1) -- (v2);
\draw[arr1] (v1) -- (v3);
\draw[arr1] (v2) -- (v4);
\draw[arr1] (v3) -- (v4);
\path [] (v1) -- node [midway,above,align=center ] {$\mu$} (v2);
\path [] (v1) -- node [midway,left,align=center ] {$\alpha_{\mathbf{X'}}$} (v3);
\path [] (v2) -- node [midway,right,align=center ] {$\alpha_{\mathbf{Y'}}$} (v4);
\path [] (v3) -- node [midway,below,align=center ] {$\nu$} (v4);
\end{tikzpicture}
\end{adjustbox}
that is:
\begin{equation}\label{eqn:commutativity}
\alpha_{\mathbf{Y'}} \circ \mu = \nu \circ \alpha_{\mathbf{X'}},
\end{equation}
where $\mu$ and $\nu$ are the interventional distributions $P(a^{-1}(\mathbf{Y'}) \vert do(a^{-1}(\mathbf{X'})))$ and $P'((\mathbf{Y'}) \vert do((\mathbf{X'})))$.
\end{definition}
Intuitively, commutativity means that we would obtain equivalent interventional results in two different ways: (i) by intervening on the base model, obtaining the interventional distribution $\mu$ on the base model and then abstracting via $\alpha_{\mathbf{Y'}}$; or, (ii) by intervening on the base model, abstracting via $\alpha_{\mathbf{X'}}$ and then obtaining the interventional distribution $\nu$ on the abstracted model.
Formally, commutativity has a category-theoretic meaning in the category of $\mathtt{FinStoch}$ where objects are sets and arrows are stochastic matrices \citep{fritz2020synthetic}. A rigorous explanation is provided in \cite{rischel2020category}, but here it is worth remarking that, when working with SCMs with finite variables and finite domains, every discrete variable (e.g.: ${X'}$) is associated with its domain set (e.g.: $\mathcal{M'}[{X'}]$), discrete distributions (e.g.: $P'(Y' \vert do(X'))$) can be encoded in stochastic matrices (e.g.: $\mu$), and deterministic abstractions (e.g.: $\alpha_{X'}$) can also be represented as binary stochastic matrices.
Given this interpretation, the commutativity equality in Equation \ref{eqn:commutativity} boils down to an equality between matrix products. More details on this algebraic encoding are offered in Appendix \ref{app:AlgebraicEncoding}.
As abstractions normally introduce approximations and rarely guarantee perfect commutativity,
it is useful to define an abstraction error wrt to interventions.
\begin{definition}[Abstraction error wrt an interventional distribution \citep{rischel2020category}]\label{def:AbstractionError}
Let $\boldsymbol{\alpha}$ be an abstraction from SCM $\scm{}$ to SCM $\scm{'}$. Let $\mathbf{X'}$ and $\mathbf{Y'}$ be two disjoint subsets of variables in $\mathcal{X'}$. The abstraction error $E_{\boldsymbol{\alpha}}(\mathbf{X'},\mathbf{Y'})$ wrt the interventional distribution $P'(\mathbf{Y'} \vert do(\mathbf{X'}))$ is the Jensen-Shannon distance (JSD) between the upper and the lower path in the diagram in Definition \ref{def:AbstractionConsistency}:
\begin{equation}\label{eqn:abstractionerror}
D_{JSD}(\alpha_{\mathbf{Y'}} \circ \mu; \nu \circ \alpha_{\mathbf{X'}}).
\end{equation}
\end{definition}
A definition of JSD is recalled in Appendix \ref{app:JSD}. Intuitively, the abstraction error quantifies the discrepancy between the upper and lower path in the abstraction diagram: how different are the results when (i) we first work with the low-level model and then abstract, and (ii) we first abstract and then work with the high-level model. This measure and the choice of JSD have a category-theoretic justification in the category $\mathtt{FinStoch}$ enriched in the category $\mathtt{Met}$ \citep{rischel2020category}.
From the above definitions, we can derive an overall notion of error.
\begin{definition}[Abstraction error] \label{def:OverallAbstractionError}
Let $\boldsymbol{\alpha}$ be an abstraction from $\scm{}$ to $\scm{'}$. Let $\mathcal{J}$ be the set of all disjoint pair sets $(\mathbf{X'},\mathbf{Y'}) \in \mathscr{P}(\mathcal{X'})\times\mathscr{P}(\mathcal{X'}), \mathbf{X'} \cap \mathbf{Y'} = \emptyset$, where $\mathscr{P}()$ is the powerset. The abstraction error is:
$$
e(\boldsymbol{\alpha}) = \sup_{(\mathbf{X'},\mathbf{Y'}) \in \mathcal{J}} E_{\boldsymbol{\alpha}}(\mathbf{X'},\mathbf{Y'}).
$$
\end{definition}
Thus, the overall abstraction error is simply the worst-case abstraction error when considering all possible interventional distributions. While $\mathcal{J}$ is formally defined as the set of all disjoint pair sets, it is possible to reduce $\mathcal{J}$ only to those pair sets $(\mathbf{X'},\mathbf{Y'})$ representing meaningful or relevant interventions $P'(\mathbf{Y'}\vert do(\mathbf{X'}))$. A consistent abstraction is then simply a zero-error abstraction:
\begin{definition}[Consistent abstraction]
Let $\boldsymbol{\alpha}$ be an abstraction from $\scm{}$ to $\scm{'}$. The abstraction is consistent if, for all pairs $(\mathbf{X'},\mathbf{Y'})$ in $\mathcal{J}$, the abstraction $\boldsymbol{\alpha}{}$ is consistent wrt $P'(\mathbf{Y'} \vert do(\mathbf{X'}))$.
\end{definition}
\section{Related Work} \label{sec:RelatedWork}
Alternative accounts of abstraction have been offered in the literature. A seminal definition proposed by \cite{rubenstein2017causal}, and refined by \cite{beckers2019abstracting}, characterized abstraction only on a distributional level; measures of abstraction approximation in this context are discussed in \cite{beckers2020approximate}.
More detailed definitions that consider both the structural and the distributional levels, and that are grounded in category theory, have been presented in \cite{rischel2020category,rischel2021compositional,otsuka2022equivalence}. A review and a comparison of these definitions is offered in \cite{zennaro2022abstraction}.
An attempt at defining a hierarchy of abstraction learning problems has been put forward in \cite{zennaro2022towards}.
Causal representation learning (CRL) \citep{chalupka2017causal} shares with this work a similar objective, but a different setup. Instead of learning a mapping between two SCMs, CRL normally starts from unstructured data and aims at learning causally-relevant representations. While abstraction learning as we defined it deals with mappings between SCMs, CRL may be seen as a preliminary or complementary step to abstraction learning, dealing with a mapping from an unstructured data space onto the space of causal variables potentially belonging to a SCM.
\section{Problem Statement}\label{sec:ProblemStatement}
The abstraction framework in Section \ref{sec:Background} provides a rigorous way to estimate an abstraction error once we are given a fully defined abstraction $\boldsymbol{\alpha} = \langle R,a,\alpha \rangle$ from $\scm{}$ to $\scm{'}$.
Instead, we consider the scenario in which two SCMs $\scm{}$ and $\scm{'}$ are known, but only a partially defined abstraction in the form $\boldsymbol{\alpha} = \langle R,a \rangle$ is available. This represents the common situation where structural knowledge ($R,a$) is readily available, but detailed distributional knowledge ($\alpha$) lacking;
in our lab example, this corresponds to the case where researchers from lab $L$ and $L'$ can exchange their models, agree on which variables are relevant, but they have no immediate knowledge on how to relate the results of their interventional experiments.
Given two SMCs $\scm{}$ to $\scm{'}$, and a partial abstraction $\boldsymbol{\alpha} = \langle R,a \rangle$, abstraction learning is the problem of learning the values for the maps $\alpha_{X'}$ that minimize the abstraction error, achieving, if possible, a consistent abstraction. We can then cast the abstraction learning problem as an optimization problem:
\begin{equation}\label{eqn:optimizationProblem}
\min_{\alpha\in\mathcal{A}}e(\boldsymbol{\alpha}),
\end{equation}
where $\mathcal{A}$ is the space of all feasible solutions for the collection of surjective maps $\alpha_{X'}$.
This optimization problem is challenging for three reasons: (i) it implies multiple sub-problems; (ii) these sub-problems are related; (iii) the solution space is combinatorial.
\paragraph{Multiple sub-problems.}
Let us consider and unpack the optimization in Equation \ref{eqn:optimizationProblem}:
\begin{eqnarray}\label{eqn:SubProblems}
&\min_{\alpha\in\mathcal{A}} & e(\boldsymbol{\alpha}) = \\ &\min_{\alpha\in\mathcal{A}} & \sup_{(\mathbf{X'},\mathbf{Y'}) \in \mathcal{J}} E_{\boldsymbol{\alpha}}(\mathbf{X'},\mathbf{Y'}) = \\
&\min_{\alpha\in\mathcal{A}} & \sup \left\{ E_{\boldsymbol{\alpha}}(\mathbf{X'},\mathbf{Y'}), E_{\boldsymbol{\alpha}}(\mathbf{Y'},\mathbf{Z'}),
E_{\boldsymbol{\alpha}}(\mathbf{X'},\mathbf{Z'}) ...\right\} = \\
&\min_{\alpha\in\mathcal{A}} & \sup \left\{
D_{JSD}(\alpha_{\mathbf{Y'}} \mu; \nu \alpha_{\mathbf{X'}}),
D_{JSD}(\alpha_{\mathbf{Z'}} \mu'; \nu' \alpha_{\mathbf{Y'}}),
D_{JSD}(\alpha_{\mathbf{Z'}} \mu''; \nu'' \alpha_{\mathbf{X'}}) \label{eqn:SubProblems_lastline}
...\right\}.
\end{eqnarray}
The minimization of a supremum implies a minimization over multiple \emph{sub-problems}. Each sub-problem is defined by a pair set of endogenous variables $(\mathbf{X'},\mathbf{Y'}) \in \mathcal{J}$ representing the interventional distribution $P'(\mathbf{Y'} \vert do(\mathbf{X'}))$. For each interventional distribution we set up a diagram as in Definition \ref{def:AbstractionConsistency}, and we solve it in $\alpha_{\mathbf{X'}}$ and $\alpha_{\mathbf{Y'}}$ with the objective of minimizing the error $E_{\boldsymbol{\alpha}}(\mathbf{X'},\mathbf{Y'})$. Thus, we have a number of sub-problems equal to the cardinality $|\mathcal{J}|$, each one requiring the minimization of a JSD, as shown in Equation \ref{eqn:SubProblems_lastline}. Notice the direct correspondence between \emph{one} sub-problem, \emph{one} diagram, and \emph{one} minimization of a JSD.
\paragraph{Related sub-problems.}
The sub-problems identified above are not necessarily independent. As soon as we consider two interventional distributions involving one identical subset of variables, we will obtain two diagrams sharing an abstraction variable, For instance, if we consider $P'({Y'} \vert do({X'}))$ and $P'({Z'} \vert do({X'}))$, then the induced diagrams will share the abstraction map $\alpha_{{X'}}$;
this implies that minimizing the abstraction error for $E_{\boldsymbol{\alpha}}({X'},{Y'})$ by changing the value of $\alpha_{{X'}}$ will affect the abstraction error of $E_{\boldsymbol{\alpha}}({X'},{Z'})$ too.
\paragraph{Combinatorial optimization.}
The domain of each $\alpha_{X'}$, encoded as a binary stochastic matrix with shape $N_i \times M_i$, is $\{0,1\}^{N_{i}\times M_{i}}$, together with the constraint of stochasticity and surjectivity. There exists a finite number of solutions, equal to all possible surjective functions from a discrete $M_i$-dimensional space to a discrete $N_i$-dimensional space, which can be computed as
$
N! \inlineiiks{M}{N},
$
where $\inlineiiks{M}{N} = \frac{1}{N!} \sum_{i=0}^{N} (-1)^{N-i} \binom{N}{i} i^M$ is the second-kind Stirling number \citep{graham1994concrete}. Consequently, since $\alpha$ is a collection of $|\mathcal{X}'|$ surjective functions, the space $\mathcal{A}$ of feasible solutions is the Cartesian product
$ \mathcal{A} = \prod_{i=1}^{|\mathcal{X}'|}\{0,1\}^{N_{i}\times M_{i}}$,
with the constraint of stochasticity and surjectivity. The number of solutions, given by all possible combination of $\alpha_{X'}$ matrices, is:
\begin{equation}\label{eqn:all_n_surj_functions}
|\mathcal{A}| = \prod_{i=1}^{|\mathcal{X}'|} N_i! \iiks{M_i}{N_i},
\end{equation}
The finiteness of the space $\mathcal{A}$ allows, theoretically, for searching an optimal solution by \emph{enumeration}. However, this quickly becomes unfeasible as the dimensionality of the input models grows
\section{Methodology} \label{sec:Methodology}
As the enumeration strategy is not generally feasible, some form of heuristic
becomes necessary.
In this section we first describe our solution approach to the abstraction learning problem
discussing: (i) a joint approach to solve all the abstraction sub-problems; (ii) a relaxation of the optimization problem; (iii) a parametrization of the relaxed problem; (iv) the enforcement of stochasticity on the parameters; (v) the enforcement of surjectivity on the parameters; (vi) solution by gradient descent.
We then present two immediate extensions of the algorithm: (vii) use of weighting on the loss function; and (viii) ensembling of models to better explore the solution space.
\paragraph{(i) Joint approach.} The abstraction learning problem consists of a set of related sub-problems. Simplistic approaches could ignore the structure connecting these problems. An \emph{independent approach}, for instance, would consider each problem separately, solve it, and, at the end, apply some form of aggregation to decide the value of those matrices appearing in multiple sub-problems. Similarly, a \emph{sequential approach} would establish an order among the sub-problems, and solve them accordingly, freezing the value of previously learned matrices. However, we aim at devising a \emph{joint approach} which, from the beginning, would take into considerations the existing structure and compute a jointly optimal solution over all the sub-problems at once. We will compare our joint approach against these simpler approaches and showcase the benefits of our algorithm.
\paragraph{(ii) Problem relaxation.}
Traversing the discrete solution space implied by a combinatorial optimization problem is notoriously challenging \citep{papadimitriou1998combinatorial} and the subject of current research \citep{titsias2017hamming,jaini2021sampling}.
We introduce a relaxation of the original problem, by redefining the original solution space $\mathcal{A}$ as
$\mathcal{A'} = \prod_{i=1}^{|\mathcal{X}|}[0,1]^{N_{i}\times M_{i}}$.
Solution matrices are now allowed to assume continuous values in $[0,1]$, which can be interpreted as allowing for stochastic or uncertain mappings.
\paragraph{(iii) Parametrization of the problem.}
Having allowed for a continuous solution space, we now parametrize the problem by defining a set of parameters $\mathbf{W}$ containing a weight matrix $W_{X'} \in \mathbb{R}^{N_{i}\times M_{i}}$ for each abstraction $\alpha_{X'}$. We then redefine our learning problem as:
\begin{equation}\label{eqn:parametrizedProblem}
\min_{\mathbf{W}\in\mathcal{A''}}e(\boldsymbol{\alpha}(\mathbf{W})),
\end{equation}
where the abstraction tuple $\boldsymbol{\alpha}(\mathbf{W}) = \langle R,a,\mathbf{W}\rangle$ is now parametrized by $\mathbf{W}$, and the solution space is
$\mathcal{A''} = \prod_{i=1}^{k}\mathbb{R}^{N_{i}\times M_{i}}$.
\paragraph{(iv) Enforcement of stochasticity.}
The solution to the parametrized problem in Equation \ref{eqn:parametrizedProblem} returns a set of weights matrices in $\mathbb{R}^{N_{i}\times M_{i}}$. To force the solution towards a binary form, we apply a tempering operator column-wise
$\tmp{W}=\tfrac{e^{\frac{W_{i,j}}{T}}}{\sum_{i}e^{\frac{W_{i,j}}{T}}}$,
where $T \in \mathbb{R}_+$ is the temperature hyperparameter. This operator projects the solution into $[0,1]^{N_{i}\times M_{i}}$, approaching a binary solution in $\{0,1\}^{N_{i}\times M_{i}}$ as $T \rightarrow 0$. The problem can be re-expressed as:
\begin{equation} \label{eqn:firstlossterm}
\min_{\mathbf{W}\in\mathcal{A''}} \mathcal{L}_1 (\mathbf{W}) = \min_{\mathbf{W}\in\mathcal{A''}} e(\boldsymbol{\alpha}(\tmp{\mathbf{W}})).
\end{equation}
Notice how the relaxation and the parametrization allowed us to move from the solution space $\mathcal{A}$ to $\mathcal{A'}$ to $\mathcal{A''}$, and how the tempering operator has allowed us to approximately project back to $\mathcal{A}$.
\paragraph{(v) Enforcement of surjectivity.} While the tempering operator returns a solution with the desired binary form, nothing guarantees that the solution matrices $\tmp{W_{X'}}$ will encode surjective functions.
To enforce this property we introduce a second term in our loss function which penalizes every row in the learned matrices which do not contain at least a single one:
\begin{equation}\label{eqn:secondlossterm}
\min_{\mathbf{W}\in\mathcal{A''}} \mathcal{L}_2 (\mathbf{W}) = \min_{\mathbf{W}\in\mathcal{A''}} \sum_{W\in\mathbf{W}}\sum_{i=1}^{N_{i}}\left(1-\max_{j}\tmp{W}_{ij}\right).
\end{equation}
\paragraph{(vi) Solution by gradient descent.}
Let our loss function be:
\begin{equation} \label{eqn:lossfunction}
\min_{\mathbf{W}\in\mathcal{A''}} \mathcal{L}(\mathbf{W}) =
\min_{\mathbf{W}\in\mathcal{A''}} \lambda \mathcal{L}_1(\mathbf{W}) + \mathcal{L}_2(\mathbf{W}),
\end{equation}
where $\lambda \in \mathbb{R}_{+}$ is a trade-off hyperparameter. $\mathcal{L}(\mathbf{W})$ is given by the sum of piecewise continuous functions: the first term is the supremum of JSDs, while the second term is related to the sum of maxima in the rows of the parameters.
Given a random starting set of candidate solutions $\mathbf{W}$, it is possible to move in the solution space towards a locally optimal solution via gradient descent algorithms, iteratively computing $\mathbf{W} = \mathbf{W} - \eta \frac{\partial \mathcal{L}(\mathbf{W})}{\partial\mathbf{W}}$, with $\eta \in \mathbb{R}_+$ being a learning rate.
Algorithm \ref{alg:jointlearning} in Appendix \ref{app:JointApproach} summarizes our joint approach. By relying on automatic differentiation we can see the algorithm as a form of neural network encoding the solution in the weights and having weight sharing defined by $\mathcal{J}$, as shown in Figure \ref{fig:nn}.
\begin{figure}
\centering
\begin{center}
\begin{tikzpicture}[node distance={15mm}, thick, main/.style = {draw, circle}]
\node[dashed,thick,draw=black!75] (1) at (0,0) {$\lambda\cdot\mathcal{L}_1 + \mathcal{L}_2$};
\node[main] (2) at (-2,-1) {$\nu'$};
\node[rectangle,thick,draw=black!75] (3) at (-4,-1) {$W_{Z'}$};
\node[main] (4) at (-2,1) {$\nu$};
\node[rectangle,thick,draw=black!75] (5) at (-4,0) {$W_{Y'}$};
\node[rectangle,thick,draw=black!75] (6) at (-4,1) {$W_{X'}$};
\node[main] (7) at (-6,-1) {$\mu'$};
\node[main] (8) at (-6,1) {$\mu$};
\node[] (9) at (-8,-1) {$do(Y)$};
\node[] (10) at (-8,1) {$do(X)$};
\node[] (X) at (3.5,1) {$X$};
\node[] (Y) at (5.5,1) {$Y$};
\node[] (XX) at (3.5,-1) {$X'$};
\node[] (YY) at (5.5,-1) {$Y'$};
\node[] (L) at (6.5,-1.5) {$\mathcal{L}$};
\path[->,draw=blue] (2) edge [bend left=8] (1);
\path[->,draw=black] (2) edge [bend right=8] (1);
\draw[->,draw=red] (4) -- (1);
\draw[->,draw=black] (4) -- (2);
\path[->,draw=blue] (3) edge [bend right=45] (1);
\path[->,draw=black] (3) edge [bend right=40] (1);
\draw[->,draw=red] (5) -- (1);
\draw[->,draw=blue] (5) -- (2);
\path[->,draw=black] (6) edge [bend left=8] (4);
\path[->,draw=red] (6) edge [bend right=8] (4);
\path[->,draw=blue] (7) edge [bend left=8] (3);
\path[->,draw=black] (7) edge [bend right=8] (3);
\draw[->,draw=red] (8) -- (5);
\draw[->,draw=black] (8) -- (7);
\path[->,draw=black] (10) edge [bend left=25] (6);
\path[->,draw=red] (10) edge [bend left=20] (6);
\path[->,draw=red] (10) edge [bend left=8] (8);
\path[->,draw=black] (10) edge [bend right=8] (8);
\path[->,draw=blue] (9) edge (7);
\path[->,draw=blue] (9) edge [bend left=15] (5);
\path[->,draw=red] (X) to node[above,font=\small]{$\mu$} (Y);
\path[->,draw=red] (XX) to node[below,font=\small]{$\nu$} (YY);
\path[->,draw=red] (X) to node[left,font=\small]{$W_{X'}$} (XX);
\path[->,draw=red] (Y) to node[right,font=\small]{$W_{Y'}$} (YY);
\path[->,draw=red] (YY) edge [bend right=15] (L);
\path[->,draw=red] (YY) edge [bend left=15] (L);
\end{tikzpicture}
\end{center}
\caption{Neural network structure implied by the joint approach on an abstraction learning problem with $\mathcal{J} = \{ (X',Y'), (X',Z'), (Y',Z') \}$ (left). Circles represent known interventional distributions, solid rectangles learnable parameters, and the dashed box the loss function. Colors trace the diagrams defined by $(X',Y')$ (red), $(X',Z')$ (black), and $(Y',T')$ (blue); following, for instance, the red lines, it is possible reconstruct the upper and lower path as in the diagram of $(X',Y')$ (right).
The network structure highlights the possibility of learning at once, via backpropagation, all the weight matrices shared by multiple diagrams. Contrast with other approaches in Appendix \ref{app:NN}.}
\label{fig:nn}
\end{figure}
\paragraph{(vii) Weighting by interventional data.}
The loss term in Equation \ref{eqn:firstlossterm} implicitly weights each abstraction error $E_{\boldsymbol{\alpha}}(\mathbf{X'},\mathbf{Y'})$ uniformly. It may be desirable, however, to scale the error with the relevance of the different interventional distributions using a vector $\boldsymbol{\kappa} \in \mathbb{R}_{+}^{|\mathcal{J}|}$:
\begin{eqnarray}
&\min_{\mathbf{W}\in\mathcal{A}''}&\mathcal{L}_{1}(\mathbf{W},\boldsymbol{\kappa})=\min_{\mathbf{W}\in\mathcal{A}''}e(\boldsymbol{\alpha}(\tmp{\mathbf{W}},\boldsymbol{\kappa})) =\\
&\min_{\mathbf{W}\in\mathcal{A}''}&\sup\left\{ \kappa_{1}D_{JSD}(\alpha_{\mathbf{Y'}}\mu;\nu\alpha_{\mathbf{X'}}),\kappa_{2}D_{JSD}(\alpha_{\mathbf{Z'}}\mu';\nu'\alpha_{\mathbf{Y'}})...\right\}.
\end{eqnarray}
Assuming that more relevant interventions are collected more often, these weights $\boldsymbol{\kappa}$ may be set proportionally to the amount of interventions collected. Alternatively, importance schemes re-weighting samples or interventions \citep{xu2021understanding} or affecting the balance between JSD losses \citep{vandenhende2021multi} may be adopted.
\paragraph{(viii) Ensembling of models.}
Solving a relaxed combinatorial optimization problem by gradient descent does guarantee only the achievement of a local optimum \citep{papadimitriou1998combinatorial}. In particular, gradient descent is sensitive to the morphology of the loss landscape and parameter initialization. Ensembling has been shown to improve the performance and the uncertainty estimation of learning \citep{dietterich2000ensemble,lakshminarayanan2017simple}; we then rely on running an ensemble of models with different starting parameters $\mathbf{W}$ in order to learn better abstractions
\section{Experiments} \label{sec:Experiments}
In this section we report results from running our algorithm both on synthetic and real-world data. Data and code for all simulations are openly available at \url{https://github.com/mattdravucz/jointly-learning-causal-abstraction/}.
\subsection{Synthetic experiments}
In our synthetic simulations, we consider four scenarios featuring different aspects of abstraction and presenting different challenges.
Table \ref{tab:SyntheticExperiments} provides an overview of these scenarios; as evinced from it, our scenarios cover different forms of abstractions, while keeping the complexity of the problems limited for verification via enumeration. Appendix \ref{app:SyntheticData_models} provides details for all the models, the abstractions, the set $\mathcal{J}$, and the optimal solutions computed via enumeration.
\begin{table}
\caption{Summary of the scenarios. \emph{Scenario} and \emph{Type of abstraction} describe the scenario; \emph{Sub-pr.} refers to the number of sub-problems (and hence the number of diagrams) implied by each abstraction learning problem; \emph{Abstraction Matrices} lists the shapes of the abstraction matrices to be learned; \emph{\#Sol.} evaluates the number of possible solutions according to Equation \ref{eqn:all_n_surj_functions}; \emph{Optimum} reports the existence of a unique zero-error or non-zero error solution. \label{tab:SyntheticExperiments}}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c>{\centering}p{4.5cm}cccc}
\hline
\textbf{\emph{Scenario}} & \textbf{\emph{Type of abstraction}} & \textbf{\emph{Sub-pr.}} & \textbf{\emph{Abstraction Matrices}} & \textbf{\emph{\#Sol.}} & \textbf{\emph{Optimum}}\tabularnewline
\hline
Basic & Elimination of low-level var & 1 & $\{2\times2,2\times2\}$ & 4 & Unique zero-error\tabularnewline
\hline
Collapsing & Merging of low-level vars & 3 & $\{2\times4,2\times2,2\times2\}$ & 56 & Unique non-zero-error\tabularnewline
\hline
Extended & Reduction of resolution & 3 & $\{3\times4,2\times3,2\times2\}$ & 432 & Unique zero-error\tabularnewline
\hline
v-Structure & Elimination of low-level var\\
Merging of low-level vars\\
Reduction of resolution & 3 & $\{2\times2,2\times2,2\times4\}$ & 56 & Unique non-zero-error\tabularnewline
\hline
\end{tabular}}
\end{table}
For each one of these scenarios, we first compute the ground-truth solution via enumeration. Then we perform the following studies: (a) we run our algorithm, and compare its performance against an independent and a sequential baseline approach (see Section \ref{sec:Methodology}(i)); (b) we analyze how weighting can affect the quality of the solutions; (c) we perform an ablation study in which we remove the surjectivity penalty term; (d) we perform a sensitivity analysis in which we vary the value of the hyperparameters $T$ and $\lambda$ specified by our approach.
Performance is evaluated in terms of JSD, surjective penalty, $\ell_1$-distance from the optimal ground-truth solution, and wall-clock time. Results are averaged over 10 simulations. Algorithms for all the approaches are provided in Appendix \ref{app:SolutionApproaches} and details about the experimental settings and metrics in Appendix \ref{app:SyntheticData_experimental}.
Figure \ref{fig:syntexp_a_performance} exemplifies the learning process in the \emph{extended} scenario. The joint approach reliably learns a solution closer to the ground-truth optimum than the other approaches. As low levels of JSD and surjective penalty are necessary but not sufficient to reach the optimal ground-truth solution, the independent and sequential approach achieve a low loss, but their $\ell_1$-distance is significantly higher than the joint approach. We hypothesize that the better results of our algorithm are due to the reliance on the information shared between sub-problems that is discarded by the other approaches. Analogous plots for the other scenarios are available in Appendix \ref{app:SyntheticData_additional}.
Table \ref{tab:syntexp_a_performance} provides the performance of the three approaches aggregated over the four scenarios. Consistently with our hypothesis, the joint approach performs better or equally to the other algorithms in terms of $\ell_1$-distance from the ground truth by exploiting all the shared information; next comes the sequential approach which uses shared information only partially; last is the independent approach which completely discards it. For an analogous reason, ordering in terms of time efficiency is reversed: dropping shared information, the independent approach is fully parallelizable wrt the sub-problems; the sequential approach may deem some sub-problems redundant and avoid solving them; the joint approach considers all the sub-problems jointly at once.
Table \ref{tab:syntexp_b_performance} shows how the final result of the joint algorithm may be affected by weighting. In this instance, we considered only the \emph{v-structure} scenario, and we assumed that the JSD error related to cancer when intervening on smoking ($E_{\boldsymbol{\alpha}}(S',C')$) would be more important than the JSD error for fatigue when intervening on cancer ($E_{\boldsymbol{\alpha}}(C',F')$) or when intervening on smoking ($E_{\boldsymbol{\alpha}}(S',F')$). When placing $80\%$ of the weight on $E_{\boldsymbol{\alpha}}(S',C')$, the final JSD for this interventional diagram decreases, while other JSD do not significantly change and exhibit higher variance. Weighting can then be used to get better approximations on those parts of the problem the modeller is more concerned with.
Table \ref{tab:syntexp_c_performance} confirms the the critical role of the surjective penalty $\mathcal{L}_2$; in its absence the algorithm can learn a better solution in terms of JSD by ignoring values in the abstracted model, but it lands on a solution significantly further from the optimal solution. Additional discussion and sample learned matrices are provided in Appendix \ref{app:SyntheticData_additional}, together with results from the sensitivity analysis.
\begin{figure}
\centering
\includegraphics[scale=0.5]{img/a-perf-ext.png}
\caption{Performance during training on the \emph{extended} scenario. While all approaches minimize their loss, the joint approach achieves a significantly lower (Wilcoxon test, $p\textrm{-value}<0.05$) $\ell_1$-distance from the ground truth.}
\label{fig:syntexp_a_performance}
\end{figure}
\begin{table}
\caption{Performance at the end of training aggregated over the four scenarios. In all instances, the joint approach competes or overcomes the baseline approaches, although with longer runtime.}
\label{tab:syntexp_a_performance}
\centering
\resizebox{.7\columnwidth}{!}{%
\begin{tabular}{ccccc}
\hline
& \textbf{\emph{L1 Dist}} & \textbf{\emph{JSD Loss}} & \textbf{\emph{Surj Pen}} & \textbf{\emph{Time}}\tabularnewline
\hline
Independent & $3.25\pm2.06$ & $0.50\pm0.31$ & $0.40\pm0.37$ & ${3.74\pm0.05}$ \tabularnewline
\hline
Sequential & $1.45\pm1.55$ & $0.28\pm0.09$ & ${0.03\pm0.07}$ & $7.10\pm0.41$ \tabularnewline
\hline
Joint & ${0.85\pm0.64}$ & ${0.24\pm0.03}$ & $0.08\pm0.11$ & $8.12\pm0.20$ \tabularnewline
\hline
\end{tabular}}
\end{table}
\begin{table}
\centering
\begin{minipage}[t]{0.53\linewidth}\centering
\caption{JSD on the \emph{v-structure} scenario. Weighting causes a marginal redistribution of errors.}
\label{tab:syntexp_b_performance}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cccc}
\hline
& $E_{\boldsymbol{\alpha}}(S',C')$ & $E_{\boldsymbol{\alpha}}(C',F')$ & $E_{\boldsymbol{\alpha}}(S',F')$\tabularnewline
\hline
Unweighted & $0.28\pm0.07$ & $0.28\pm0.01$ & ${0.15\pm0.03}$ \tabularnewline
\hline
Weighted & ${0.25\pm0.05}$ & ${0.27\pm0.06}$ & ${0.15\pm0.04}$ \tabularnewline
\hline
\end{tabular}}
\end{minipage}\hfill%
\begin{minipage}[t]{0.43\linewidth}\centering
\caption{Performance on the \emph{v-structure} scenario. Surjectivity penalty is critical to achieve significantly better (Wilcoxon test, $p\textrm{-value}<0.05$) results.}
\label{tab:syntexp_c_performance}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ccc}
\hline
& \emph{\textbf{L1 Dist}} & \emph{\textbf{JSD Loss}} \tabularnewline
\hline
Joint & $\mathbf{3.00\pm1.34}$ & $0.72\pm0.08$ \tabularnewline
\hline
Ablated joint & $5.00\pm2.41$ & ${0.68\pm0.15}$ \tabularnewline
\hline
\end{tabular}}
\end{minipage}
\end{table}
\subsection{Real-world experiment}
We learn an abstraction between the implicit causal models for lithium-ion battery manufacturing developed by two research groups: the Laboratoire de Réactivité et Chimie des Solides (LRCS) group, and the Warwick Manufacturing Group (WMG).
Battery electrode manufacturing is a complex process involving several key stages (e.g. material selection, mixing and coating
To develop high-performing batteries, it is necessary to understand how each of the manufacturing parameters influence the subsequent product. The current approach relies heavily on experienced lab personnel with extensive knowledge in order to adjust the manufacturing parameters and achieve a desired battery performance. Considerable research effort is directed to develop models of the manufacturing stages \citep{RomanRamirez2022b,Niri2022} and optimised feedback control mechanisms to reduce the reliance on human expertise.
In this work we focus on the dry mass loading ($ML$) variable from the coating stage \citep{cunha2020artificial,liu2022interpretable}, which directly determines the energy density of the final battery. The dry ML is in part controlled by the comma-bar gap ($CG$) variable, which is set manually and guides the mass loading of the active material in its wet form; the wet coating subsequently passes through a drying stage resulting in the dry coating. Public datasets on this process are scarce and of limited dimensions, due to the cost and complexity of the measurements; however, the ability to predict the dry ML based on the upstream CG variable is vital in order to achieve a target battery energy density and increase manufacturing efficiency. To obviate this problem, we aim at learning an abstraction that may relate the models assumed by the two research groups and then integrate their data to significantly improve downstream inferences.
We use a dataset for battery coating from the LRCS group \citep{cunha2020artificial}, and recordings performed by the WMG group. We perform pre-processing in order to select the relevant variables.
As in many real-world scenarios, fully-specified SCMs are not available, so we define elementary SCMs $\mathcal{M}^{WMG}$ and $\mathcal{M}^{LRCS}$ with minimal assumptions (Figure \ref{fig:LRCR_SCM} and \ref{fig:SCM_WMG}).
We then elect the WMG model as the base model since it has a higher resolution in terms of number and domain of observed variables. Appendix \ref{app:RealData_models} provides details about the data, pre-processing, the SCMs and the abstraction.
We learn the abstraction $\boldsymbol{\alpha}{}$ from $\mathcal{M}^{WMG}$ to $\mathcal{M}^{LRCS}$ using our joint approach, similarly to what we have done with the synthetic experiments. We evaluate results in two ways. Qualitatively, we assess the shape of the learned matrices, to confirm they are binary and surjective, and to check whether identical values in the domain of the base and abstracted variables are mapped to each other. Quantitatively, we set three downstream regression problems, see Table \ref{tab:realexp_performance}, to assess whether transporting data via abstraction could help improve prediction. Task (a) represents a challenging scenario in which we use limited and expensive experimental data to perform predictions. Task (b) represents a scenario in which data transported via abstraction provides support for our predictions, while task (c) the case in which transported data augments the training set but does not provide support on the test set. See Appendix \ref{app:RealData_experimental} for details on experimental settings and metrics.
Figure \ref{fig:realexp_mat_cg} shows the learned matrix $W_{CG}$ corresponding to the abstraction map $\alpha_{CG}$ relating the $CG$ variables in the two models. The matrix approximates a binary matrix encoding a surjective function. A red border is used to denote identical value of the $CG$ matrix in the base and abstracted model; these values are successfully mapped to each other, while intermediate values align along the main diagonal.
Notice, however, that alternative patterns could also be learned, as discussed in Appendix \ref{app:RealData_additional}, together with shape of the learned matrix $W_{ML}$.
Table \ref{tab:realexp_performance} shows the mean-square error when learning only on the small LRCS dataset (task a) and when using data transported via abstraction from the WMG dataset; performance improves both when WMG data provides the missing support for prediction (task b) or when it just augments the dataset (task c).; when not providing the required missing support, we observe an improvement on selected cases; this is likely due to having learned a non-perfect abstraction and to the noise introduced during the transport of the data.
\begin{table}
\begin{minipage}[t]{0.33\linewidth}\centering
\resizebox{\columnwidth}{!}{%
\includegraphics{img/realexp_mat_cg.png}}
\captionof{figure}{Learned $W_{CG}$. The matrix is close to binary, surjective, and diagonal as expected.}
\label{fig:realexp_mat_cg}
\end{minipage}\hfill%
\begin{minipage}[b]{0.63\linewidth}\centering
\caption{Regression problems setup and MSE. Transporting data via abstraction reduces the error.}
\label{tab:realexp_performance}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cccc}
\hline
& \textbf{\emph{Training set}} & \textbf{\emph{Test Set}} & \textbf{\emph{MSE}} \tabularnewline
\hline
(a) & LRCS{[}$CG\neq k${]} & LRCS{[}$CG=k${]} & $1.86\pm1.75$\tabularnewline
\hline
(b) & LRCS{[}$CG\neq k${]} & LRCS{[}$CG=k${]} & $0.22\pm0.26$ \tabularnewline
& + WMG & & \tabularnewline
\hline
(c) & LRCS{[}$CG\neq k${]} & LRCS{[}$CG=k${]} & $1.22\pm0.95$ \tabularnewline
& + WMG{[}$CG\neq k${]} & + WMG{[}$CG=k${]} & \tabularnewline
\hline
\end{tabular}}
\end{minipage}
\end{table}
\section{Discussion and Conclusion} \label{sec:Discussion}
In this paper we have built an abstraction learning framework starting from the formalization of an abstraction between SCMs proposed in \citep{rischel2020category}. We have examined how we could use this definition to express the learning problem as an abstraction error minimization problem, and proposed an algorithm to solve it. Our approach has been based on relaxing the initial problem, parametrizing it and solving it via gradient descent. Results on synthetic datasets show that our algorithm perform better than other simpler approaches, and it reaches a solution closer to the optimum. Furthermore, application to EV manufacturing data provided a proof of concept of the usefulness of learning abstraction to relate models and transport data in low-data regimes.
The abstraction learning problem we have tackled is a particular instance of the very generic problem of learning a mapping not just between isolated sets or objects, but between sets and objects sharing a structure expressed by transformations or morphisms.
This problem has been given the form of a notoriously challenging discrete combinatorial optimization problem, for which other common heuristics exist in the optimization literature.
Application of our approach to the real-world data of a lithium-ion battery manufacturing stage provided a proof of concept of the potential of learning abstraction between models devised by different groups, although more research would be needed to make the proposed algorithm widely usable in real-world applications. The joint approach is sensitive to initialization, and using abstraction to transfer data may introduce noise; increasing robustness and extending it to the continuous domain are directions for future work.
As a first instance of an abstraction learning problem, our setup assumes perfect knowledge of the models. Dropping this requirement leads to problems in which we could learn both abstraction and distributions from data (incidentally, in our real-world scenario, we trivially learned mechanisms from data at setup time, but the learning of mechanisms and abstraction could happen jointly).
Similarly, we may further limit knowledge about the abstraction, dropping, for instance, the requirement of knowing $a$ or $R$ \citep{zennaro2022towards}.
Alternatively, it may also be possible to consider using domain knowledge on how specific variables and outputs are related to guide learning; the space of surjective function grows very rapidly, and restricting it by using available priors would simplify the learning problem.
Other relevant directions of work include theoretical evaluation of our relaxation or the definition of the set $\mathcal{J}$
(as discussed in Appendix \ref{app:J}, $\mathcal{J}$ may contain irrelevant or redundant interventional distributions, and an algorithm that selects relevant interventions may take advantage of ideas such as minimal intervention sets \citep{lee2018structural}).
|
\section{Introduction}
The transmission eigenvalue problem plays a role in the inverse scattering theory for inhomogeneous media. This eigenvalue problem is connected to the injectivity of the corresponding scattering operator \cite{CM88}, \cite{Kirsch86}. Transmission eigenvalues are related to interrogating frequencies for which there is an incident field that is not scattered by the medium. In the acoustic setting, the transmission eigenvalue problem is a system of two second-order elliptic equations of two unknowns equipped with the Cauchy data on the boundary. After four decades of extensive study, the spectral properties are known to depend on a type of contrasts of the media near the boundary. Natural and interesting questions on the interior transmission eigenvalue problem include: the {\it discreteness} of the spectrum (see e.g. \cite{CCG10, BCH11, Sylvester12, LV12, MinhHung1,CV21}), the {\it location} of transmission eigenvalues (see \cite{CGH10, CL12, Vodev15, Vodev18}, and also \cite{CMV20} for the application in time domain), the {\it Weyl law} of transmission eigenvalues and the {\it completeness} of the generalized eigenfunctions (see e.g. \cite{LV12, LV12-A, LV15, Robbiano16}). We refer the reader to \cite{CCH16} for a recent, and self-contained introduction to the transmission eigenvalue problem and its applications.
Let us describe its mathematical formulation.
Let $\Omega$ be a bounded, simply connected, open subset of $\mathbb{R}^d$ of class $C^3$ with $d\geq 2$. Let $A_1,A_2$ be two real, symmetric matrix-valued functions, and let $\Sigma_1,\Sigma_2$ be two bounded positive functions that are all defined in $\Omega$. Assume that $A_1$ and $A_2$ are uniformly elliptic, and $\Sigma_1$ and $\Sigma_2$ are bounded below by a positive constant in $\Omega$, i.e., for some constant $\Lambda \ge 1$, one has, for $\ell=1, 2$,
\begin{equation}\label{condi1}
\Lambda^{-1} |\xi|^2\leq \langle A_\ell (x) \xi, \xi \rangle \leq \Lambda |\xi|^2 \quad \mbox{ for all } \xi \in \mathbb{R}^d, \mbox{ for a.e. } x\in \Omega,
\end{equation}
and
\begin{equation}\label{condi2}
\Lambda^{-1} \leq \Sigma_\ell(x)\leq \Lambda \mbox{ for a.e. } x \in \Omega.
\end{equation}
Here and in what follows, $\langle \cdot, \cdot \rangle$ denotes the Euclidean scalar product in $\mathbb{C}^d$ and $|\cdot|$ is the corresponding norm.
A complex number $\lambda$ is called an eigenvalue of the transmission eigenvalue problem associated with the pairs $(A_1, \Sigma_1)$ and $(A_2, \Sigma_2)$ in $\Omega$ if there is a non-zero pair of functions $(u_1, u_2) \in [H^1(\Omega)]^2$ that satisfies the system
\begin{equation}\label{ITE}
\left\{\begin{array}{cl}
\operatorname{div}(A_1 \nabla u_1) - \lambda\Sigma_1 u_1= 0 &\text{ in}~\Omega, \\[6pt]
\operatorname{div}(A_2 \nabla u_2) - \lambda\Sigma_2 u_2= 0 &\text{ in}~\Omega, \\[6pt]
u_1 =u_2, \quad A_1 \nabla u_1\cdot \nu = A_2 \nabla u_2\cdot \nu & \text{ on }\Gamma.
\end{array}\right.
\end{equation}
Here and in what follows, $\Gamma$ denotes $\partial \Omega$, and $\nu$ denotes the outward, normal, unit vector on $\Gamma$. Such a pair $(u_1, u_2)$ is then called an eigenfunction pair.
Assume that $A_1, \, A_2, \, \Sigma_1, \, \Sigma_2$ are continuous in $\bar \Omega$, and the following conditions on the boundary $\Gamma$ hold, with $\nu = \nu(x)$ :
\begin{equation}\label{cond1}
\langle A_2(x) \nu, \nu\rangle \langle A_2(x) \xi, \xi \rangle - \langle A_2(x) \nu, \xi \rangle^2
\neq \langle A_1(x) \nu, \nu\rangle \langle A_1(x) \xi, \xi \rangle - \langle A_1(x) \nu, \xi \rangle^2,
\end{equation}
for all $x \in \Gamma$ and for all $\xi \in \mathbb{R}^d \setminus \{0 \}$ with $\langle \xi, \nu \rangle =0$, and
\begin{equation}\label{cond2}
\big\langle A_2(x) \nu, \nu \big\rangle \Sigma_2(x) \neq \big\langle A_1(x) \nu, \nu \big\rangle \Sigma_1(x), \; \forall \, x \in \Gamma.
\end{equation}
(Q. H.) Nguyen and the second author \cite{MinhHung2} established the Weyl law of eigenvalues and the completeness of the generalized eigenfunctions for transmission eigenvalue problem under conditions \eqref{cond1} and \eqref{cond2} via the Fourier analysis assuming that $A_1, \, A_2, \, \Sigma_1, \, \Sigma_2$ are continuous in $\bar \Omega$.
Condition \eqref{cond1} is equivalent to the celebrated complementing conditions due to Agmon, Douglis, and Nirenberg \cite{ADNII} (see also \cite{ADNI}). The explicit formula given here was derived in \cite{Ng-WP} in the context of the study of negative index materials. Conditions \eqref{cond1} and \eqref{cond2} were derived by (Q. H.) Nguyen and the second author in \cite{MinhHung1} in their study of the discreteness of the eigenvalues for transmission eigenvalue problem.
In the case
\begin{equation}\label{condi3}
A_1 = A_2 = A\mbox{ in $\Omega$},
\end{equation}
it was also shown by (Q. H.) Nguyen and the second author \cite{MinhHung1} (see also \cite{Sylvester12}) using the multiplier technique that the discreteness holds if
\begin{equation}\label{condi4}
\Sigma_1 \neq \Sigma_2 \mbox{ on } \Gamma.
\end{equation}
The goal of this paper is to study the Weyl law of the eigenvalues and the completeness of the generalized eigenfunctions under conditions \eqref{condi3} and \eqref{condi4}. It is worth noting that results in this direction have been obtained previously with more constraints on the coefficients than \eqref{condi3} and \eqref{condi4}. Robbiano \cite{Robbiano16} (see also \cite{Robbiano13}) gives the sharp order of the counting number when $A = I$ in $\Omega$, $\Sigma_1 = 1$, $\Sigma_2 \neq \Sigma_1$ near the boundary and $\Sigma_2$ is smooth. The analysis is based on both the microanalysis (see, e.g., \cite{GG94, Zworski12}) and the regularity theory for the transmssion eigenvalue problem. In the isotropic case, the Weyl law was established by Petkov and Vodev \cite{PV17} and Vodev \cite{Vodev18, Vodev18-2, Vodev19} for $C^\infty$ coefficients. Their analysis is heavily based on microanalysis and the smoothness condition is strongly required. In addition, their work involved a delicate analysis on the Dirichlet to Neumann maps using non-standard parametrix construction initiated by Vodev \cite{Vodev15}, which have their own interests. It is not clear how one can improve the $C^\infty$ condition and extend their results to the anisotropic setting using their approach. Concerning the completeness of the generalized eigenfunctions, we want to mention the work of Robbiano \cite{Robbiano13} where the case $A = I$ and $\Sigma_1 \neq \Sigma_2$ in $\bar \Omega$ was considered.
\medskip
We are ready to state the main results of this paper. From now on, we will assume in addition that
\begin{equation}\label{condi5}
\|(A_1,A_2)\|_{C^2(\bar \Omega)} + \|(\Sigma_1,\Sigma_2)\|_{C^1(\bar \Omega)} \leq \Lambda.
\end{equation}
We denote by $(\lambda_j)_j$ the set of transmission eigenvalues associated with the transmission eigenvalue problem \eqref{ITE}.
\medskip
Concerning the Weyl law, we have
\begin{theorem}
\label{thm1}
Assume \eqref{condi1}-\eqref{condi2} and \eqref{condi3}-\eqref{condi5}. Let ${\mathcal N}(t)$ denote the counting function, i.e.
\[
{\mathcal N}(t) = \#\{j\in \mathbb{N} : |\lambda_j|\leq t\}.
\]
Then
\[
\mathcal{N}(t) = {\bf c}t^{\frac{d}{2}} + o(t^{\frac{d}{2}}) \mbox{ as } t \to + \infty,
\]
where
\[
\mathbf{c} : = \frac{1}{(2\pi)^d} \sum_{\ell =1}^2 \int_\Omega\left | \Big \{ \xi \in \mathbb{R}^d; \; \langle A_\ell(x) \xi, \xi \rangle < \Sigma_\ell (x) \Big\} \right | dx.
\]
\end{theorem}
For a measurable subset $D$ of $\mathbb{R}^d$, we denote $|D|$ its (Lebesgue) measure.
\medskip
Concerning the completeness, we obtain
\begin{theorem}\label{thm2}
Assume \eqref{condi1}-\eqref{condi2} and \eqref{condi3}-\eqref{condi5}. The set of generalized eigenfunction pairs of \eqref{ITE} is complete in $L^2(\Omega) \times L^2(\Omega)$.
\end{theorem}
\begin{rem} \rm As a direct consequence of either \Cref{thm1} or \Cref{thm2}, the number of eigenvalues of the transmission eigenvalue problem is infinite. As far as we know, this fact is new under the assumption that $A$ is allowed to be anisotropic and the regularity of the coefficients are only required up to the order 2.
\end{rem}
\medskip
The analysis used in the proof of \Cref{thm1} and/or \Cref{thm2} also allows us to obtain the following result on the transmission eigenvalue free region of the complex plane ${\mathbb C}$.
\begin{proposition}\label{pro} Assume \eqref{condi1}-\eqref{condi2} and \eqref{condi3}-\eqref{condi5}. For $\gamma >0$, there exists $\lambda_0 > 0$ such that if $\lambda \in \mathbb{C}$ with $|\Im (\lambda)| \ge \gamma |\lambda|$ and $|\lambda| \ge \lambda_0$, then $\lambda$ is not a transmission eigenvalue.
\end{proposition}
Here and and in what follows, for $z \in \mathbb{C}$, let $\Im(z)$ denote the imaginary part of $z$.
\medskip
A more general result of \Cref{pro} is given in \Cref{pro-Free}.
\begin{rem} \rm \label{rem-cm}
Since $\gamma > 0$ can be chosen arbitrary small, combining the discreteness result in \cite{MinhHung1} mentioned above and Proposition \ref{pro}, one derives that all the transmission eigenvalues, but finitely many, lie in a wedge of arbitrary small angle.
\end{rem}
Some comments on \Cref{thm1} and \Cref{thm2} are in order. In the conclusion of \Cref{thm1}, the multiplicity of eigenvalues is taken into account and the multiplicity is associated with some operator $T_{\lambda^*}$, which is introduced in \Cref{sect-W} (see \eqref{pre-Tl} and \eqref{def-lambda^*}). Concerning $T_{\lambda^*}$, the following facts hold (see \Cref{rem-comp} and \Cref{rem-indep} for more information): if $\mu$ is a characteristic value of the operator $T_{\lambda^*}$ associated with an eigenfunction $(u,v)$ and $\lambda^*+\mu\neq 0$, then $ \lambda^*+ \mu$ is a transmission eigenvalue of \eqref{ITE} with an eigenfunction pair $(u_1,u_2)$ given by
\[
u_1 = (\lambda^*+\mu)u+v \quad \quad \mbox{ and }\quad \quad u_2 = v.
\]
Moreover, if $\lambda_j$ is a transmission eigenvalue problem, then $\lambda_j \neq \lambda^*$ and $\lambda_j - \lambda^*$ is a characteristic value of $T_{\lambda^*}$.
In \Cref{thm2}, the generalized eigenfunctions are also associated to such an operator $T_{\lambda^*}$. We recall that the generalized eigenfunctions are complete in $[L^2(\Omega)]^2$ if the subspace spanned by them is dense in $[L^2(\Omega)]^2$.
\medskip
\Cref{thm1} and \Cref{thm2} provide the Weyl laws and the completeness under the assumptions \eqref{condi3} and \eqref{condi4} assuming the regularity conditions in \eqref{condi5}. Our results hold for $A_1 = A_2 = A$ being anisotropic in contrast to the isotropic setting considered previously. Moreover, the regularity assumption \eqref{condi5} on the coefficients was out of reach previously.
\medskip
Our approach is in the spirit of \cite{MinhHung2}
and is hence different from the ones used to study these problems given in the previous works mentioned above. The key idea is to establish the {\it regularity theory} for the transmission eigenvalue problem under the stated assumptions (see \Cref{thm-WP}). Nevertheless, several new ingredients and observations are required for the regularity theory due to the fact that \eqref{condi3}, which is degenerate, is considered instead of \eqref{cond1}. One of the key steps to capture the phenomena is to derive appropriate estimates in a half plane setting. It is important to note that since $A_1 = A_2 = A$, the setting is non-standard, and the classical arguments pioneered in \cite{ADNI, ADNII} cannot be applied since the role of $\Sigma_1$ and $\Sigma_2$ are ignored there. To this end, our arguments for the Cauchy problems not only require the information of the first derivatives and their structure of the data but also involve the information of the second derivatives and their structure (see, e.g., \Cref{lem-vwp}). This is quite distinct from the complementing case where the arguments for the Cauchy problems only require the information of the first derivatives and no structure of the data is required \cite{MinhHung2} (see, e.g., \cite[Lemma 2 and Corollary 2]{MinhHung2}).
One might note that the arguments used to derive the discreteness in \cite{MinhHung1} requires less assumption on the regularity of the coefficients but only give the information for one direction of $\lambda$ ($\arg{\lambda} = \pi/2$) for large $\lambda$. This is not sufficient to apply the theory of Hilbert-Schmidt operators.
\medskip We have so far discuss the transmission eigenvalue problem in the acoustic setting. Known results for the transmission eigenvalue problem in the electromagnetic setting are much less. In this direction, we
mention the work of Cakoni and Nguyen \cite{Cakoni-Ng21} on the state of art on the discreteness of the eigenvalues, the work of Fornerod and Nguyen \cite{Fornerod-Ng1} on the completeness of generalized of eigenfunctions and the upper bound of the eigenvalues for the setting considered in \cite{Cakoni-Ng21}, and the work of Vodev \cite{VodevM21} on the free region of eigenvalues for a setting considered in \cite{Cakoni-Ng21}, and the references therein.
\medskip
The Cauchy problem also naturally appears in the context of negative-index materials after using reflections as initiated in \cite{Ng-Complementary} (see also \cite{Ng-Superlensing-Maxwell}). The well-posedness and the limiting absorption principle for the Helmholtz equation with sign-changing coefficients
were developed by the second author \cite{Ng-WP} using the Fourier and multiplier approach (see also \cite{NgSil}). The work \cite{Ng-WP} deals with the stability question of negative index materials, and is the starting point for the analysis of the transmission eigenvalue problems in \cite{MinhHung1, MinhHung2} (see also \cite{Cakoni-Ng21}). Other aspects and applications of negative-index materials as well as the stability and instability the Cauchy problem are discussed in \cite{Ng-Superlensing, Ng-Negative-Cloaking, Ng-CALR, Ng-CALR-O} and the references therein. A survey is given in \cite{Ng-Survey}.
\medskip
The paper is organized as follows. \Cref{sect-Notations} is devoted to define some notations used throughout the paper. In \Cref{sect-WP}, we establish the well-posedness and the regularity theory for the Cauchy systems associated with the transmission eigenvalue problems. The analysis is then developed in such a way that the theory of Hilbert-Schmidt operators can be used. This is given in \Cref{sect-W} where the Weyl laws are established. The completeness is considered in \Cref{sect-C}.
\section{Notations}\label{sect-Notations}
Here are some useful notations used throughput this paper.
We denote, for $\tau > 0$,
\begin{equation}\label{def-Omegatau}
\Omega_\tau=\Big\{x\in \Omega: \operatorname{dist}(x,\Gamma)<\tau \Big\}.
\end{equation}
For $d \ge 2$, set
$$
\mathbb{R}^d_+ = \Big\{x \in \mathbb{R}^d; x_d > 0 \Big\} \quad \mbox{ and } \quad \mathbb{R}^d_0 = \Big\{x \in \mathbb{R}^d; x_d = 0 \Big\}.
$$
We will identify $\mathbb{R}^d_0$ with $\mathbb{R}^{d-1}$ in several places. For $s>0$, we denote
$$B_s = \{x \in \mathbb{R}^d : |x|<s\}.$$ For $m \ge 1$, $p \ge 1$, and $\lambda \in \mathbb{C}^*$ and $u \in W^{m,p}(\Omega) $, we define
\begin{equation}\label{def-norm-lambda}
\|u\|_{W^{m,p}_{\lambda}(\Omega)} = \left( \sum_{j=0}^m \| |\lambda|^{\frac{m-j}{2}} \nabla^ju\|_{L^p(\Omega)}^p \right)^{1/p}.
\end{equation}
\section{Well-posedness and regularity theory for the transmission eigenvalue problems} \label{sect-WP}
In this section, we study the well-posedness and the regularity theory of the Cauchy problem
\begin{equation}\label{ITE-C}
\left\{\begin{array}{cl}
\operatorname{div}(A_1 \nabla u_1) - \lambda\Sigma_1 u_1= f_1 &\text{ in}~\Omega, \\[6pt]
\operatorname{div}(A_2 \nabla u_2) - \lambda\Sigma_2 u_2= f_2 &\text{ in}~\Omega, \\[6pt]
u_1 -u_2=0, \quad (A_1 \nabla u_1- A_2 \nabla u_2)\cdot \nu=0 & \text{ on }\Gamma,
\end{array}\right.
\end{equation}
under the assumptions \eqref{condi1}-\eqref{condi2}, and \eqref{condi4}-\eqref{condi5}, and
\begin{equation}\label{condi6}
A_1 = A_2 = A \mbox{ in $\Omega_\tau$},
\end{equation}
for some $\tau > 0$, instead of \eqref{condi3} for appropriate $\lambda \in \mathbb{C}$ and $(f_1,f_2)$ in $L^p$-scale.
\medskip
Here is the main result of this section.
\begin{theorem} \label{thm-WP}
Assume \eqref{condi1}-\eqref{condi2}, \eqref{condi4}-\eqref{condi5}, and \eqref{condi6}. Let $1<p<+\infty$ and $\gamma \in (0, 1)$. There exist constants $\lambda_0>0$ and $C>0$ depending on $\Omega, \, \Lambda$, $\tau$, $p$, and $\gamma$ such that for $\lambda \in \mathbb{C}$ with $|\lambda|>\lambda_0$ and $|\Im (\lambda)| \geq \gamma |\lambda|$ and for $(f_1,f_2)\in [L^p(\Omega)]^2$, there is a unique solution $(u_1,u_2)\in [L^p(\Omega)]^2$ with $u_1 - u_2 \in W^{2,p}(\Omega)$ of the Cauchy problem \eqref{ITE-C}. Moreover,
\begin{equation} \label{thm-WP-1}
|\lambda| \|(u_1,u_2)\|_{L^p(\Omega)} + \|u_1-u_2\|_{W^{2,p}_{\lambda}(\Omega)} \leq C \|(f_1,f_2)\|_{L^p(\Omega)}.
\end{equation}
Assume in addition that $f_1-f_2 \in W^{1,p}(\Omega)$. Then $(u_1,u_2) \in [W^{1,p}(\Omega)]^2$, $u_1-u_2 \in W^{3,p}(\Omega_{\tau/2})$, and
\begin{equation} \label{thm-WP-2}
|\lambda| \|(u_1,u_2)\|_{W^{1,p}_\lambda (\Omega)} + \|u_1-u_2\|_{W^{3,p}_{\lambda}(\Omega_{\tau/2})} \leq C \left ( |\lambda|^{1/2}\|(f_1,f_2)\|_{L^p(\Omega)}+\|f_1-f_2\|_{W^{1,p}_\lambda (\Omega)}\right ).
\end{equation}
\end{theorem}
\begin{remark} \rm The boundary conditions must be understood as
$$
u_1 - u_2 = 0 \mbox{ on } \Gamma \quad \mbox{ and } \quad A \nabla (u_1 - u_2) \cdot \nu = 0 \mbox{ on } \Gamma,
$$
which make sense since $u_1 - u_2 \in W^{2, p}(\Omega)$.
\end{remark}
\begin{remark} \rm In \eqref{thm-WP-2}, we only estimate $\|u_1-u_2\|_{W^{3,p}_{\lambda}(\Omega_{\tau/2})}$ not $\|u_1-u_2\|_{W^{3,p}_{\lambda}(\Omega)}$ since $f_1$ and $f_2$ are not supposed to be in $W^{1, p}(\Omega)$. Nevertheless, when $A_1 = A_2$ in $\Omega$, the estimate is also valid for $\|u_1-u_2\|_{W^{3,p}_{\lambda}(\Omega)}$.
\end{remark}
\begin{rem} \rm As a consequence of \eqref{thm-WP-1} and the theory of regularity of elliptic equations, one derives that $(u_1,u_2) \in [W^{2,p}_{loc}(\Omega)]^2$ and for $\Omega' \Subset \Omega$ \footnote{Recall that $\Omega' \Subset \Omega$ means $\overline{ \Omega'} \subset \Omega$.}, it holds
\[
\|(u_1,u_2)\|_{W^{2,p}_\lambda (\Omega')} \leq C \|(f_1,f_2)\|_{L^p(\Omega)},
\]
where $C$ depends also on $\Omega'$ (see, e.g., \cite[Lemma 17.1.5]{Hor3} and \cite[Theorem 9.11]{GilbargTrudinger}).
\end{rem}
As a consequence of \Cref{thm-WP}, we obtain the following result on the free-region of the eigenvalues.
\begin{proposition}\label{pro-Free} Assume \eqref{condi1}-\eqref{condi2}, \eqref{condi4}-\eqref{condi5}, and \eqref{condi6}. For $\gamma >0$, there exists $\lambda_0 > 0$ such that if $\lambda \in \mathbb{C}$ with $|\Im (\lambda)| \ge \gamma |\lambda|$ and $|\lambda| \ge \lambda_0$, then $\lambda$ is not a transmission eigenvalue.
\end{proposition}
The rest of this section, containing two subsections, is devoted to the proof of \Cref{thm-WP}. The first one is on the analysis in the half space. The proof of \Cref{thm-WP} is then given in the second subsection.
\subsection{Half space analysis}
Let $1< p < +\infty$. For $j=1, 2, \cdots$, and $\lambda \in \mathbb{C} \setminus \{0 \}$, we denote
\[
\|\psi\|_{W^{j-1/p,p}_\lambda(\mathbb{R}^d_0)} = |\lambda|^{1/2-1/(2p)} \|\psi\|_{W^{j-1,p}_\lambda(\mathbb{R}^d_0)} + | \nabla^{j-1} \psi |_{W^{1-1/p,p}(\mathbb{R}^d_0)},
\]
where $\|\psi\|_{W^{j-1,p}_\lambda(\mathbb{R}^d_0)}$ is defined as in \eqref{def-norm-lambda} with $\Omega = \mathbb{R}^d_0$, and
\[
| \psi |_{W^{1-1/p,p}(\mathbb{R}^d_0)}^p = \int_{\mathbb{R}^{d-1}} \int_{\mathbb{R}^{d-1}} \frac{|\psi(x')-\psi(y')|^p}{|x'-y'|^{d+p-2}} \, dx' \, dy'.
\]
By the trace theory, there exists a positive constant $C$ depending only on $p$ and $j$ such that
\[
\| u \|_{W^{j-1/p,p}_\lambda (\mathbb{R}^d_0)} \leq C \|u\|_{W^{j,p}_\lambda (\mathbb{R}^d_+)} \mbox{ for } u \in W^{j, p} (\mathbb{R}^d_+).
\]
In fact, this inequality holds for $\lambda \in \mathbb{C}$ with $|\lambda | = 1$; the general case follows by scaling.
\medskip
The starting point and the key ingredient of our analysis is \Cref{lem-vwp}. \Cref{lem-main} below is a special case of \Cref{lem-vwp} and is later used to derive \Cref{lem-vwp}.
\begin{lem} \label{lem-main}
Let $A \in \mathbb{R}^{d \times d}$ be a constant symmetric matrix and let $\Sigma_1,\Sigma_2$ be two positive constants such that
\[
\Lambda^{-1} |\xi|^2\leq \langle A \xi, \xi \rangle \leq \Lambda |\xi|^2 \mbox{ for all } \xi \in \mathbb{R}^d,
\]
\[
\Lambda^{-1} \leq \Sigma_1, \Sigma_2 \leq \Lambda, \quad \mbox{ and } \quad |\Sigma_1 - \Sigma_2| \ge \Lambda^{-1},
\]
for some $\Lambda \ge 1$. Let $\gamma \in (0, 1)$, $1<p<+\infty$, and let $\varphi \in W^{2-1/p,p}(\mathbb{R}^d_0)$. Given $\lambda \in \mathbb{C}$ with $|\lambda| \ge 1$ and $|\Im (\lambda)| \geq \gamma |\lambda|$, there exists a unique solution $(u_1,u_2) \in [L^p(\mathbb{R}^d_+)]^2$ with $u_1-u_2 \in W^{2,p}(\mathbb{R}^d_+)$ of the following Cauchy problem
\[
\left \{
\begin{array}{cl}
\operatorname{div} (A\nabla u_1)-\lambda \Sigma_1 u_1 = 0 & \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
\operatorname{div} (A\nabla u_2)-\lambda \Sigma_2 u_2 = 0 & \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
u_1-u_2 = \varphi , \quad A\nabla (u_1-u_2) \cdot e_d = 0 & \mbox{ on }\mathbb{R}_0^d.
\end{array}
\right .
\]
Moreover,
\begin{equation}
\label{lem-main-p1}
|\lambda|\|(u_1,u_2)\|_{L^p(\mathbb{R}^d_+)} + \|u_1-u_2 \|_{W^{2,p}_\lambda (\mathbb{R}^d_+)} \leq C \| \varphi \|_{W^{2-1/p,p}_\lambda (\mathbb{R}^d_0)}.
\end{equation}
Assume in addition that $\varphi \in W^{3-1/p,p}(\mathbb{R}^d_0)$. Then $(u_1,u_2) \in [W^{1,p}(\mathbb{R}^d_+)]^2$ with $u_1-u_2 \in W^{3,p}(\mathbb{R}^d_+)$, and
\begin{equation}
\label{lem-main-p2}
|\lambda|\|(u_1,u_2)\|_{W^{1,p}_\lambda(\mathbb{R}^d_+)} + \|u_1-u_2 \|_{W^{3,p}_\lambda (\mathbb{R}^d_+)} \leq C \|\varphi\|_{W^{3 - 1/p, p}_\lambda (\mathbb{R}^d_0)}.
\end{equation}
Here $C$ is a positive constant depending only on $\Lambda$, $\gamma$, $p$, and $d$.
\end{lem}
\begin{proof}
For a function $u : \mathbb{R}^d \to \mathbb{C}$ (resp. $\varphi :\mathbb{R}^{d-1} \to \mathbb{C}$) we denote by $\hat{u}$ the Fourier transform of $u$ with respect to the first $(d-1)$ variables (resp. by $\hat{\varphi}$ the Fourier transform of $\varphi$), i.e., for $(\xi',x_d) \in \mathbb{R}^{d-1}\times (0,\infty)$,
\[
\hat{u}(\xi',x_d) = \int_{\mathbb{R}^{d-1}}u(x',x_d)e^{-ix'\cdot \xi'} \, dx' \quad \mbox{ and } \quad \hat{\varphi}(\xi')= \int_{\mathbb{R}^{d-1}}\varphi (x') e^{-ix'\cdot \xi'} \, dx' .
\]
Since, for $\ell =1,2$,
\begin{equation*}
\operatorname{div}(A \nabla u_\ell) - \lambda \Sigma_\ell u_\ell = 0 \mbox{ in } \mathbb{R}^{d}_+,
\end{equation*}
it follows that
\begin{equation*}
a \hat u_\ell''(\xi',t) + 2 i b (\xi') \hat u_\ell'(\xi',t ) - (c (\xi') + \lambda \Sigma_\ell) \hat u_\ell(\xi',t) = 0 \mbox{ for } t > 0,
\end{equation*}
where
\begin{equation}\label{lem-main-abc}
a = \langle A e_d, e_d \rangle, \; \; b(\xi') = \sum_{j=1}^{d-1} A_{jd} \xi'_j , \; \; c (\xi') = \sum_{i,j=1}^{d-1}A_{ij}\xi'_i \xi'_j, \; \; \mbox{ and } \; \; a c (\xi') - b(\xi')^2 > 0,
\end{equation}
since $A$ is symmetric and positive. One then obtains, see, e.g., \cite[proof of Lemma 2]{MinhHung2} for the details,
\begin{equation}\label{lem-main-ul}
\quad \hat u_\ell(\xi', t) = \alpha_\ell (\xi') e^{\eta_\ell (\xi') t}
\end{equation}
where
\begin{equation}\label{lem-main-etal}
\eta_\ell (\xi') = \frac{1}{a} \big( - i b (\xi') - \sqrt{\Delta_\ell (\xi')} \big)
\end{equation}
and
\begin{equation}\label{lem-main-alphal}
\alpha_\ell (\xi') = \frac{\hat{\varphi}(\xi') \sqrt{\Delta_{\ell+1} (\xi')}}{\sqrt{\Delta_2(\xi')} - \sqrt{\Delta_1 (\xi')}} \quad \mbox{ with } \quad \Delta_\ell (\xi')= -b^2 (\xi') + a \big(c (\xi') + \lambda \Sigma_\ell \big).
\end{equation}
Here we use the convention $\Delta_{2 + \ell} = \Delta_\ell$, and $\sqrt{\Delta_\ell}$ denotes the square root of $\Delta_\ell$ with the positive real part.
Let $v_\ell \in W^{1, p} (\mathbb{R}^d_+)$ for $\ell = 1, \, 2$ be the unique solution of the system
\[
\left \{
\begin{array}{cl}
\operatorname{div} (A\nabla v_\ell)-\lambda \Sigma_\ell v_\ell = 0 & \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
v_\ell = \varphi & \mbox{ on } \mathbb{R}^d_0.
\end{array}
\right .
\]
We have\footnote{The results hold for $|\lambda| =1$, see, e.g. \cite[Theorem 14.1]{ADNI}, the general case follows by scaling.} , for $\ell = 1, \, 2$,
\begin{equation}
\label{lem-main-14-1}
\|v_\ell \|_{W^{j, p}_\lambda(\mathbb{R}^d_+)} \leq C \| \varphi \|_{W^{j - 1/p, p}_\lambda (\mathbb{R}^d_0)} \mbox{ for } j = 2, 3,
\end{equation}
and
\begin{equation}
\label{lem-main-14-2}
\hat v_\ell (\xi', t) = \hat{\varphi}(\xi') e^{\eta_\ell (\xi') t}.
\end{equation}
Extend $u_\ell(x', t)$ and $\partial_{tt}^2 v_\ell(x', t)$ by 0 for $t < 0$ for $\ell = 1, \, 2$ and {\it still} denote these extensions by $u_\ell(x', t)$ and $\partial_{tt}^2 v_\ell(x', t)$. Let ${\mathcal F}$ denote the Fourier transform in $\mathbb{R}^d$.
We then obtain from \eqref{lem-main-ul} and \eqref{lem-main-14-2} that, with $\xi=(\xi',\xi_d) \in \mathbb{R}^d$,
$$
{\mathcal F} u_\ell (\xi) = - \frac{\hat \varphi (\xi')}{\eta_\ell (\xi') - i \xi_d} \frac{\sqrt{\Delta_{\ell+1} (\xi')}}{\sqrt{\Delta_2 (\xi')} - \sqrt{\Delta_1(\xi')}} \quad \mbox{ and } \quad {\mathcal F} \partial_{tt}^2 v_\ell (\xi) = - \frac{\hat \varphi (\xi') \eta_\ell^2 (\xi') }{\eta_\ell (\xi') - i \xi_d} .
$$
It follows that
\[
\mathcal{F}u_\ell (\xi) = m_{\ell,\lambda} (\xi) \mathcal{F}\partial^2_{tt}v_\ell (\xi),
\]
where
\begin{equation}\label{def-ml}
m_{\ell,\lambda}(\xi) = \frac{\sqrt{\Delta_{\ell+1}(\xi')}}{\eta_\ell^2 (\xi') (\sqrt{\Delta_2(\xi')}-\sqrt{\Delta_1(\xi')})} .
\end{equation}
Note that
\[
\Delta_2 (\xi') -\Delta_1 (\xi') = a \lambda (\Sigma_2 - \Sigma_1) \neq 0
\]
and
\begin{multline*}
\frac{1}{\eta_\ell(\xi')} \mathop{=}^{\eqref{lem-main-etal}} \frac{a}{-i b(\xi') - \sqrt{\Delta_\ell(\xi')}} = \frac{a \big(-i b(\xi') + \sqrt{\Delta_\ell(\xi')} \big)}{-b(\xi')^2 - \Delta_\ell(\xi')}\\[6pt]
\mathop{=}^{\eqref{lem-main-alphal}} \frac{a \big(-i b(\xi') + \sqrt{\Delta_\ell(\xi')} \big)}{- a \big( c(\xi') + \lambda \Sigma_\ell \big)} = \frac{i b(\xi') - \sqrt{\Delta_\ell(\xi')}}{ c(\xi') + \lambda \Sigma_\ell}.
\end{multline*}
We derive from \eqref{def-ml} that
\begin{equation}\label{def-ml1}
m_{\ell,\lambda}(\xi) = \frac{\sqrt{\Delta_{\ell+1}(\xi')}(\sqrt{\Delta_1(\xi')} + \sqrt{\Delta_2(\xi')})(ib(\xi') - \sqrt{\Delta_\ell(\xi')})^2}{a\lambda (\Sigma_2-\Sigma_1)(c(\xi') + \lambda \Sigma_\ell)^2 }.
\end{equation}
We have, by \eqref{lem-main-abc} and \eqref{lem-main-alphal},\footnote{Given two functions, $p_1(\xi',\lambda)$ and $p_{2}(\xi',\lambda)$ the notation $p_1(\xi,\lambda) \sim p_2(\xi',\lambda)$ means that there exists a constant $C\geq 1$ independent of $\xi'$ and $\lambda$ such that $C^{-1}|p_1(\xi',\lambda)|\leq |p_2(\xi',\lambda)|\leq C |p_1(\xi',\lambda)|$. }
$$
|\Delta_{\ell}(\xi')| \sim (|\xi'|^2 + |\lambda|), \quad |b(\xi')| \le C |\xi'|, \quad \mbox{ and } \quad |c(\xi') + \lambda \Sigma_\ell| \sim |\xi'|^2 + |\lambda|.
$$
We then derive from \eqref{def-ml1} that
\begin{equation}
|\xi|^j |\nabla^j m_{\ell,\lambda}(\xi)| \le C_j |\lambda|^{-1} \mbox{ for } j \in \mathbb{N}.
\end{equation}
It follows from Mikhlin-H\"ormander's multiplier theorem, see, e.g., \cite[Theorem 7.9.5]{Hor1}, that
\begin{equation}\label{lem-main-P1}
|\lambda| \| u_\ell \|_{L^p(\mathbb{R}^d)} \le C \| \partial_{tt}^2 v_\ell \|_{L^p(\mathbb{R}^d)},
\end{equation}
which implies
\begin{equation}\label{lem-main-p1-1}
|\lambda| \| u_\ell \|_{L^p(\mathbb{R}^d)} \mathop{\le}^{\eqref{lem-main-14-1}} C \| \varphi \|_{W^{2-1/p, p}_\lambda(\mathbb{R}^d)}.
\end{equation}
On the other hand, one has
\begin{equation*}\left\{
\begin{array}{c}
\operatorname{div} \big(A \nabla (u_1 - u_2) \big) - \lambda \Sigma_1 (u_1 - u_2) = \lambda (\Sigma_1 - \Sigma_2) u_2 \mbox{ in } \mathbb{R}^d_+, \\[6pt]
u_1 - u_2 = 0 \mbox{ on } \mathbb{R}^d_0.
\end{array}\right.
\end{equation*}
This yields
\[
\| u_1 - u_2 \|_{W^{2, p}_\lambda(\mathbb{R}^d_+)} \le C \| \lambda (\Sigma_1 - \Sigma_2) u_2 \|_{L^p(\mathbb{R}^d_+)} \mathop{\le}^{\eqref{lem-main-p1-1}} C \| \varphi \|_{W^{2 - 1/p, p }_\lambda (\mathbb{R}^d_+)}.
\]
We next deal with \eqref{lem-main-p2}. By taking the derivative of the system with respect to $x_j$ for $1 \le j \le d -1$ and applying
\eqref{lem-main-p1}, we have, for $1 \le j \le d - 1$,
\begin{equation}
\label{lem-main-p2-1}
|\lambda| \| (\partial_{x_j} u_1, \partial_{x_j}u_2)\|_{L^p(\mathbb{R}^d_+)} + \|(\partial_{x_j}u_1- \partial_{x_j} u_2)\|_{W^{2,p}_{\lambda}(\mathbb{R}^d_+)} \leq C \| \partial_{x_j} \varphi\|_{W^{2- 1/p, p}_\lambda(\mathbb{R}^d_0)}.
\end{equation}
Extend $\partial_t u_\ell(x', t)$ and $\partial_{ttt}^3 v_\ell(x', t)$ by 0 for $t < 0$ for $\ell = 1, \, 2$ and {\it still} denote these extensions by $\partial_t u_\ell(x', t)$ and $\partial_{ttt}^3 v_\ell(x', t)$.
We then obtain from \eqref{lem-main-ul} and \eqref{lem-main-14-2} that, with $\xi=(\xi',\xi_d) \in \mathbb{R}^d$,
$$
{\mathcal F} \partial_t u_\ell (\xi) = - \frac{\hat \varphi (\xi') \eta_\ell(\xi')}{\eta_\ell (\xi') - i \xi_d} \frac{\sqrt{\Delta_{\ell+1} (\xi')}}{\sqrt{\Delta_2 (\xi')} - \sqrt{\Delta_1(\xi')}} \quad \mbox{ and } \quad {\mathcal F} \partial_{ttt}^3 v_\ell (\xi) = - \frac{\hat \varphi (\xi') \eta_\ell^3 (\xi') }{\eta_\ell (\xi') - i \xi_d} .
$$
This yields
\[
\mathcal{F} \partial_t u_\ell (\xi) = m_{\ell,\lambda} (\xi) \mathcal{F}\partial^3_{ttt}v_\ell (\xi).
\]
As in the proof of \eqref{lem-main-P1}, we obtain
\begin{equation*}
|\lambda| \| \partial_t u_\ell \|_{L^p(\mathbb{R}^d)} \le C \| \partial_{ttt}^3 v_\ell \|_{L^p(\mathbb{R}^d)},
\end{equation*}
which implies
\begin{equation}\label{lem-main-p2-2}
|\lambda| \| \partial_t u_\ell \|_{L^p(\mathbb{R}^d)} \mathop{\le}^{\eqref{lem-main-14-1}} C \| \varphi \|_{W^{3-1/p, p}_\lambda(\mathbb{R}^d)}.
\end{equation}
Combining \eqref{lem-main-p2-1} and \eqref{lem-main-p2-2}, we derive that
\begin{equation}
\label{lem-main-p2-3}
|\lambda| \| (\nabla u_1, \nabla u_2)\|_{L^p(\mathbb{R}^d_+)} + \|\nabla (u_1- u_2)\|_{W^{2,p}_{\lambda}(\mathbb{R}^d_+)} \leq C \| \varphi\|_{W^{3- 1/p, p}_\lambda(\mathbb{R}^d_0)}.
\end{equation}
Assertion \eqref{lem-main-p2} now follows from \eqref{lem-main-p2-3} and \eqref{lem-main-p1}. The proof is complete.
\end{proof}
We now state and prove a more general version of \Cref{lem-main}, which is the main ingredient of the proof of \Cref{thm-WP}.
\begin{lem} \label{lem-vwp}
Let $A \in \mathbb{R}^{d \times d}$ be a constant symmetric matrix and let $\Sigma_1,\Sigma_2$ be two positive constants such that
\[
\Lambda^{-1} |\xi|^2\leq \langle A \xi, \xi \rangle \leq \Lambda |\xi|^2 \mbox{ for all } \xi \in \mathbb{R}^d,
\]
and
\[
\Lambda^{-1} \leq \Sigma_1, \Sigma_2 \leq \Lambda, \quad \mbox{ and } \quad |\Sigma_1 - \Sigma_2| \ge \Lambda^{-1},
\]
for some $\Lambda \ge 1$. Let $\gamma \in (0, 1)$, $1<p<+\infty$, and let $f_1, f_2 \in L^p(\mathbb{R}^d_+)$, $G_1, G_2\in [L^p(\mathbb{R}^d_+)]^d$ with $G_1-G_2 \in [W^{1,p}(\mathbb{R}^d_+)]^d$, $\varphi \in W^{2-1/p,p}(\mathbb{R}^d_0)$, $\psi \in W^{1-1/p,p}(\mathbb{R}^d_0)$, and let $r_{1}^{(ij)}, r_{2}^{(ij)} \in L^p(\mathbb{R}^d_+)$ with $r_{1}^{(ij)} - r_{2}^{(ij)} \in W^{2,p}(\mathbb{R}^d_+)$ for $1 \le i, j \le d$. Given $\lambda \in \mathbb{C}$ with $|\lambda| \ge 1$ and $|\Im (\lambda)| \geq \gamma |\lambda|$, there exists a unique solution $(u_1,u_2) \in [L^p(\mathbb{R}^d_+)]^2$ with $u_1-u_2 \in W^{2,p}(\mathbb{R}^d_+)$ of the following Cauchy problem
\begin{equation} \label{lem-vwp-3}
\left \{
\begin{array}{cl}
\operatorname{div} (A\nabla u_1)-\lambda \Sigma_1 u_1 = f_1 + \operatorname{div} (G_1) + \sum_{i,j=1}^d\partial_{ij}^2 r_{1}^{(ij)}& \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
\operatorname{div} (A\nabla u_2)-\lambda \Sigma_2 u_2 = f_2 + \operatorname{div} (G_2) + \sum_{i,j=1}^d \partial_{ij}^2 r_{2}^{(ij)} & \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
u_1-u_2 =\varphi, \quad A\nabla (u_1-u_2) \cdot e_d = \psi & \mbox{ on }\mathbb{R}_0^d.
\end{array}
\right .
\end{equation}
Moreover,
\begin{align}
\label{lem-vwp-4}
C & \left(|\lambda|\|(u_1,u_2)\|_{L^p(\mathbb{R}^d_+)} + \|u_1-u_2 \|_{W^{2,p}_\lambda (\mathbb{R}^d_+)} \right) \\[6pt]
& \leq \|(f_1,f_2)\|_{L^p(\mathbb{R}^d_+)} + |\lambda|^{1/2}\|(G_1,G_2)\|_{L^p(\mathbb{R}^d_+)} + \sum_{i,j=1}^d |\lambda|\|(r_{1}^{(ij)}, r_{2}^{(ij)})\|_{L^p(\mathbb{R}^d_+)} \nonumber \\[6pt]
& + \|\varphi\|_{W^{2-1/p,p}_\lambda (\mathbb{R}^d_0)}+\|\psi\|_{W^{1-1/p,p}_{\lambda}(\mathbb{R}^d_0)} + \|G_1-G_2 \|_{W^{1,p}_\lambda (\mathbb{R}_+^d)}
+ \sum_{i,j=1}^d \|r_{1}^{(ij)} -r_{2}^{(ij)}\|_{W^{2,p}_{\lambda}(\mathbb{R}^d_+)} \nonumber.
\end{align}
Assume in addition that $f_1-f_2 \in W^{1,p}(\mathbb{R}^d_+)$, $G_1-G_2 \in W^{2,p}(\mathbb{R}^d_+)$, $\varphi \in W^{3-1/p,p}(\mathbb{R}^d_0)$, $\psi \in W^{2-1/p,p}(\mathbb{R}^d_0)$, and $r_{1}^{(ij)} = r_{2}^{(ij)} = 0$ for all $1\leq i,j \leq d$. Then $(u_1,u_2) \in W^{1,p}(\mathbb{R}^d_+)$ with $u_1-u_2 \in W^{3,p}(\mathbb{R}^d_+)$, and it holds
\begin{multline}
\label{lem-vwp-4bis}
C \left( |\lambda|\|(u_1,u_2)\|_{W^{1,p}_\lambda (\mathbb{R}^d_+)} + \|u_1-u_2 \|_{W^{3,p}_\lambda (\mathbb{R}^d_+)} \right) \\[6pt]
\leq |\lambda|^{1/2}\|(f_1,f_2)\|_{L^p(\mathbb{R}^d_+)} + |\lambda|\|(G_1,G_2)\|_{L^p(\mathbb{R}^d_+)} + \|\varphi\|_{W^{3-1/p,p}_\lambda (\mathbb{R}^d_0)} \\[6pt]
+\|\psi\|_{W^{2-1/p,p}_{\lambda}(\mathbb{R}^d_0)} + \|f_1-f_2\|_{W^{1,p}_\lambda (\mathbb{R}^d_+)} + \|G_1-G_2 \|_{W^{2,p}_\lambda (\mathbb{R}_+^d)} .
\end{multline}
Here $C$ denotes a positive constant depending only on $\Lambda$, $\gamma$, $d$, and $p$.
\end{lem}
\begin{remark} \rm Concerning \eqref{lem-vwp-4bis}, the assumption $r_{1}^{(ij)} = r_{2}^{(ij)} = 0$ for all $1\leq i,j \leq d$ is just to avoid the redundancy; the same estimate holds for the appropriate assumptions on $r_\ell^{(ij)}$ but this can be put into the conditions of $f_\ell$ and $G_\ell$ instead.
\end{remark}
\begin{proof} Since the problem is linear, \eqref{lem-vwp-4} and \eqref{lem-vwp-4bis} follow from the corresponding estimates in the following two cases:
\begin{enumerate}
\item[] $\bullet$ Case 1: $f_1=f_2 = 0$, $G_1 = G_2 = 0$, and $r_{1}^{(ij)} = r_{2}^{(ij)} = 0$ for all $1\leq i, j \leq d$.
\item[] $\bullet$ Case 2: $\varphi =0$ and $\psi=0$.
\end{enumerate}
We now proceed the proof for these cases.
\medskip
\noindent \underline{\it Case 1}: $f_1=f_2 = 0$, $G_1 = G_2 = 0$, and $r_{1}^{(ij)} = r_{2}^{(ij)}= 0$ for all $1\leq i, j \leq d$. We have
\[
\left \{
\begin{array}{cl}
\operatorname{div} (A\nabla u_1)-\lambda \Sigma_1 u_1 = 0& \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
\operatorname{div} (A\nabla u_2)-\lambda \Sigma_2 u_2 = 0& \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
u_1-u_2 =\varphi, \quad A\nabla (u_1-u_2) \cdot e_d = \psi & \mbox{ on }\mathbb{R}_0^d.
\end{array}
\right .
\]
Let $v \in W^{1,p}(\mathbb{R}^d_+)$ be the unique solution of
\[
\left \{
\begin{array}{rc}
\operatorname{div} (A\nabla v)-\lambda \Sigma_1 v = 0 & \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
A\nabla v \cdot e_d = \psi & \mbox{ on }\mathbb{R}^d_0.
\end{array}
\right .
\]
As a consequence of \cite[Theorem 2.3.2.7]{Grisvard} and a scaling argument, we have
\begin{equation}
\label{est-5}
\|v\|_{W^{2,p}_\lambda(\mathbb{R}^d_+)} \leq C\|\psi\|_{W^{1-1/p,p}_\lambda(\mathbb{R}^d_0)} \quad \mbox{ and } \quad \|v\|_{W^{3,p}_\lambda(\mathbb{R}^d_+)} \leq C\|\psi\|_{W^{2-1/p,p}_\lambda(\mathbb{R}^d_0)}.
\end{equation}
By the trace theory, it follows that
\begin{equation}
\label{est-5-bis}
\|v\|_{W^{2-1/p,p}_\lambda(\mathbb{R}^d_0)} \leq C\|\psi\|_{W^{1-1/p,p}_\lambda(\mathbb{R}^d_0)} \quad \mbox{ and } \quad \|v\|_{W^{3-1/p,p}_\lambda(\mathbb{R}^d_0)} \leq C\|\psi\|_{W^{2-1/p,p}_\lambda(\mathbb{R}^d_0)}.
\end{equation}
Considering the system of $(u_1-v,u_2)$ and using \eqref{est-5}, and \eqref{est-5-bis}, the conclusion of this case follows from \Cref{lem-main}.
\medskip
\noindent \underline{\it Case 2:} $\varphi =0$, $\psi=0$. In this case, we have
\[
\left \{
\begin{array}{cl}
\operatorname{div} (A\nabla u_1)-\lambda \Sigma_1 u_1 = f_1 + \operatorname{div} (G_1)+ \sum_{i,j=1}^d \partial_{ij}^2 r_{1}^{(ij)} & \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
\operatorname{div} (A\nabla u_2)-\lambda \Sigma_2 u_2 = f_2 + \operatorname{div} (G_2)+ \sum_{i,j=1}^d \partial_{ij}^2 r_{2}^{(ij)} & \mbox{ in }\mathbb{R}^d_+, \\[6 pt]
u_1-u_2 =0, \quad A\nabla (u_1-u_2) \cdot e_d = 0 & \mbox{ on }\mathbb{R}_0^d.
\end{array}
\right .
\]
For $\ell =1,2$, consider the following systems
\[
\left \{
\begin{array}{rll}
\operatorname{div} (A\nabla v_{\ell}^{(0)})-\lambda \Sigma_\ell v_{\ell}^{(0)} &= f_\ell &\mbox{ in }\mathbb{R}^d_+, \\[6 pt]
A \nabla v_{\ell}^{(0)} \cdot e_d &= 0 &\mbox{ on }\mathbb{R}^d_0,
\end{array}
\right .
\quad \quad \phantom{(1\leq j \leq d)}
\]
\[
\left \{
\begin{array}{rll}
\operatorname{div} (A\nabla v_{\ell}^{(j)})-\lambda \Sigma_\ell v_{\ell}^{(j)} &= (G_\ell)_j &\mbox{ in }\mathbb{R}^d_+, \\[6 pt]
A \nabla v_{\ell}^{(j)} \cdot e_d &= 0 &\mbox{ on }\mathbb{R}^d_0
\end{array}
\right .
\quad \quad (1\leq j \leq d),
\]
where $(G_\ell)_j$ denotes the $j$-th component of $G_\ell$, and
\[
\left \{
\begin{array}{rll}
\operatorname{div} (A\nabla v_{\ell}^{(ij)})-\lambda \Sigma_\ell v_{\ell}^{(ij)} &= r_{\ell}^{(ij)} &\mbox{ in }\mathbb{R}^d_+, \\[6 pt]
A \nabla v_{\ell}^{(ij)} \cdot e_d &= 0 &\mbox{ on }\mathbb{R}^d_0
\end{array}
\right .
\quad \quad (1\leq i,j \leq d).
\]
We have, see, e.g., \cite[Theorem 14.1]{ADNI}, for $1\leq i,j\leq d$,
\begin{equation}
\label{lem-vw-2}
\left\{\begin{array}{c}
\|v_{\ell}^{(0)} \|_{W^{2,p}_\lambda (\mathbb{R}^d_+)} \leq C \|f_\ell\|_{L^p(\mathbb{R}^d_+)}, \\[6pt]
\|v_{\ell}^{(j)} \|_{W^{2,p}_\lambda (\mathbb{R}^d_+)} \leq C \|G_\ell \|_{L^p(\mathbb{R}^d_+)}, \\[6pt]
\|v_{\ell}^{(ij)} \|_{W^{2,p}_\lambda (\mathbb{R}^d_+)} \leq C \| r_{\ell}^{(ij)} \|_{L^p(\mathbb{R}^d_+)}.
\end{array}\right.
\end{equation}
Since, we have
\[
\left \{
\begin{array}{rll}
\operatorname{div} (A\nabla( v_{1}^{(0)}-v_{2}^{(0)}))-\lambda \Sigma_1 (v_{1}^{(0)}-v_{2}^{(0)}) &= f_1-f_2 + \lambda (\Sigma_1 -\Sigma_2) v_2^{(0)} &\mbox{ in }\mathbb{R}^d_+, \\[6 pt]
A \nabla (v_{1}^{(0)}-v_{2}^{(0)}) \cdot e_d &= 0 &\mbox{ on }\mathbb{R}^d_0,
\end{array}
\right .
\]
and the equations for $v_1^{(j)} - v_2^{(j)}$ and $v_1^{(ij)} - v_2^{(ij)} $ are similar,
we also get, for $1\leq i,j\leq d$, by using \eqref{lem-vw-2},
\begin{equation}\label{lem-vw-cc1}
\left\{\begin{array}{c}
C\| v_1^{(0)} - v_2^{(0)} \|_{W^{2, p}_\lambda(\mathbb{R}^d_+)} \le \| (f_1, f_2 )\|_{L^p(\mathbb{R}^d_+)}, \\[6pt]
C\| v_1^{(j)} - v_2^{(j)} \|_{W^{3, p}_\lambda(\mathbb{R}^d_+)} \le \| G_1 - G_2 \|_{W^{1, p}_\lambda(\mathbb{R}^d_+)}+|\lambda|^{1/2}\|G_2\|_{L^p(\mathbb{R}^d_+)}, \\[6pt]
C\| v_1^{(ij)} - v_2^{(ij)} \|_{W^{4, p}_\lambda(\mathbb{R}^d_+)} \le \| r_1^{(ij)} - r_2^{(ij)} \|_{W^{2, p}_\lambda(\mathbb{R}^d_+)}+|\lambda|\|r_2^{(ij)}\|_{L^p(\mathbb{R}^d_+)},
\end{array}\right.
\end{equation}
and
\begin{equation}\label{lem-vw-cc1-bis}
\begin{aligned}
C\| v_1^{(0)} - v_2^{(0)} \|_{W^{3, p}_\lambda(\mathbb{R}^d_+)} &\le \|f_1 - f_2 \|_{W^{1, p}_\lambda(\mathbb{R}^d_+)}+|\lambda|^{1/2}\|f_2\|_{L^p(\mathbb{R}^d_+)}, \\
C\| v_1^{(j)} - v_2^{(j)} \|_{W^{4, p}_\lambda(\mathbb{R}^d_+)} &\le \| G_1 - G_2 \|_{W^{2, p}_\lambda(\mathbb{R}^d_+)}+|\lambda|\|G_2\|_{L^p(\mathbb{R}^d_+)}.
\end{aligned}
\end{equation}
For $\ell=1,2$, set
\[
w_\ell = v_{\ell}^{(0)} + \sum_{j=1}^d \partial_j v_{\ell}^{(j)} + \sum_{i,j=1}^d \partial^2_{ij} v_{\ell}^{(ij)}.
\]
We have
\[
\operatorname{div} (A\nabla w_\ell) - \lambda \Sigma_\ell w_\ell = f_\ell+ \operatorname{div} (G_\ell) + \sum_{i,j=1}^d \partial_{ij}^2 r_{\ell}^{(ij)} \mbox{ in }\mathbb{R}^d_+.
\]
Moreover,
\begin{multline}
\label{est-0}
C|\lambda|\|(w_1,w_2)\|_{L^p(\mathbb{R}^d_+)} \\[6pt]
\mathop{\leq}^{\eqref{lem-vw-2}} \|(f_1,f_2)\|_{L^p(\mathbb{R}^d_+)} + |\lambda|^{1/2}\|(G_1,G_2)\|_{L^p(\mathbb{R}^d_+)}+|\lambda|\sum_{i,j=1}^d\|(r_{1}^{(ij)}, r_{2}^{(ij)})\|_{L^p(\mathbb{R}^d_+)}.
\end{multline}
Using \eqref{lem-vw-cc1} and the trace theory, we derive that
\begin{multline}
\label{est-3}
\|w_1 - w_2 \|_{W^{2, p}_\lambda(\mathbb{R}^d_+)} + \|w_1-w_2\|_{W^{2-1/p,p}_{\lambda}(\mathbb{R}^d_0)} + \|A\nabla (w_1-w_2) \cdot e_d\|_{W^{1-1/p,p}_{\lambda}(\mathbb{R}^d_0)} \\[6pt]
\leq C \Big( \|f_2\|_{L^p(\mathbb{R}^d_+)}+|\lambda|^{1/2} \|(G_1,G_2)\|_{L^p(\mathbb{R}^d_+)} + \sum_{i,j=1}^d |\lambda| \|r_2^{ij}\|_{L^p(\mathbb{R}^d_+)} \\
+\|f_1 - f_2\|_{L^p(\mathbb{R}^d_+)}
+ \|G_1-G_2\|_{W^{1,p}_{\lambda}(\mathbb{R}^d_+)}+\sum_{i,j=1}^d \|r_{1}^{(ij)} - r_{2}^{(ij)}\|_{W^{2,p}_\lambda (\mathbb{R}^d_+)} \Big ).
\end{multline}
Considering the system of $(u_1-w_1,u_2-w_2)$, and using \eqref{est-0} and \eqref{est-3}, assertion \eqref{lem-vwp-4} now follows from case 1.
To deal with assertion \eqref{lem-vwp-4bis}, instead of \eqref{est-0} and \eqref{est-3}, we use, since $r_1^{(ij)} = r_2^{(ij)} = 0$,
\begin{equation}
\label{est-0-bis}
|\lambda|\|(w_1,w_2)\|_{W^{1, p}_\lambda(\mathbb{R}^d_+)}
\mathop{\leq}^{\eqref{lem-vw-2}} C \Big( |\lambda|^{1/2} \|(f_1,f_2)\|_{L^p(\mathbb{R}^d_+)} + |\lambda| \|(G_1,G_2)\|_{L^p(\mathbb{R}^d_+)} \Big),
\end{equation}
and
\begin{multline}
\label{est-3-bis}
\|w_1 - w_2 \|_{W^{3, p}_\lambda(\mathbb{R}^d_+)} + \|w_1-w_2\|_{W^{3-1/p,p}_{\lambda}(\mathbb{R}^d_0)} + \|A\nabla (w_1-w_2) \cdot e_d\|_{W^{2-1/p,p}_{\lambda}(\mathbb{R}^d_0)} \\[6pt]
\mathop{\leq}^{\eqref{lem-vw-cc1-bis}} C \Big( \|(f_1 - f_2)\|_{W^{1, p}_\lambda(\mathbb{R}^d_+)}
+ \|G_1-G_2\|_{W^{2,p}_{\lambda}(\mathbb{R}^d_+)} \Big).
\end{multline}
By considering the system of $(u_1-w_1,u_2-w_2)$, assertion \eqref{lem-vwp-4bis} now follows from case 1.
\medskip
The proof is complete.
\end{proof}
\subsection{Proof of \Cref{thm-WP}} The proof is divided into two steps:
\begin{itemize}
\item Step 1: Assuming the solution exists, we establish \eqref{thm-WP-1} and \eqref{thm-WP-2}.
\item Step 2: We establish the existence of the solutions.
\end{itemize}
We now proceed these two steps.
\medskip
\noindent {\underline{\it Step 1:}}
For $(f_1,f_2) \in [L^p(\Omega)]^2$, let $(u_1,u_2) \in [L^p(\Omega)]^2$ with $u_1-u_2 \in W^{2,p}(\Omega)$ be a solution of \eqref{ITE-C}. We prove that \eqref{thm-WP-1} and \eqref{thm-WP-2} hold.
Applying \Cref{lem-vwp} and the freezing coefficient technique, we deduce that there exists $\tau_* \in (0, \tau/2)$ depending only on $\Omega$, $\Lambda$, $\tau$, and $p$, such that
\begin{multline}
\label{thm-WP-pr1}
C \left ( |\lambda| \|(u_1,u_2) \|_{L^p(\Omega_{\tau_*})}+\|u_1-u_2\|_{W^{2,p}_\lambda (\Omega_{\tau_*})} \right ) \\[6 pt]
\leq \|(f_1,f_2)\|_{L^p(\Omega)} + |\lambda|^{1/2} \| (u_1,u_2) \|_{L^p(\Omega_\tau)}+ \|u_1-u_2\|_{W^{1,p}_\lambda (\Omega_\tau)}
\end{multline}
and
\begin{multline}
\label{thm-WP-pr1-bis}
C \left ( |\lambda| \|(u_1,u_2) \|_{W^{1, p}_\lambda(\Omega_{\tau_*})}+\|u_1-u_2\|_{W^{3,p}_\lambda (\Omega_{\tau_*})} \right )\\[6pt]
\leq |\lambda|^{1/2} \|(f_1,f_2)\|_{L^p(\Omega)} + \|f_1 - f_2 \|_{W^{1, p}_\lambda(\Omega)} \\[6pt]
+ |\lambda| \|(u_1,u_2) \|_{L^p (\Omega_\tau)}
+ |\lambda| \|(u_1,u_2) \|_{L^p (\Omega_\tau)}
+\|u_1-u_2\|_{W^{2,p}_\lambda (\Omega_\tau)},
\end{multline}
for every $\lambda \in \mathbb{C}$ with $|\Im (\lambda)| \geq c |\lambda|$ and $|\lambda| \ge 1 $. Here and in what follows, $C$ denotes a positive constant depending only on $\Omega$, $\Lambda$, $\tau$, and $p$.
Let us emphasize here that the terms $(r_{1,ij},r_{2,ij})$ in \Cref{lem-vwp}, play a crucial role in the proof of \eqref{thm-WP-pr1} since the solutions $(u_1, u_2)$ considered are only in $[L^p(\Omega)]^2$, but not in $[W^{1, p}(\Omega)]^2$. Indeed, let consider a small neighborhood of $x_0 \in \Gamma$. Using a change of variables, without loss of generality, one might assume that the boundary in this neighbourhood is {\it flat} already and $A_1 = A_2 = A$ there. In the freezing process, one has, in such a neighborhood,
\begin{multline*}
\operatorname{div} (A(x_0) \nabla u_\ell) - \lambda \Sigma_\ell (x_0) u_\ell = \operatorname{div} \big( (A(x_0) - A(x)) \nabla u_\ell \big) +
\operatorname{div} \big( A(x) \nabla u_\ell \big) - \lambda \Sigma_\ell (x_0) u_\ell \\[6pt]
=\sum_{i,j=1}^d \partial^2_{ij} \Big ( ( A_{ij}(x_0) - A_{ij}(x) ) u_\ell \Big ) - \sum_{i,j=1}^d \partial_i \Big ( u_\ell \partial_{j} (A_{ij}(x_0) - A_{ij}(x) ) \Big ) + f_\ell + \lambda \big( \Sigma_\ell (x) - \Sigma_\ell (x_0) \big) u_\ell.
\end{multline*}
Let $\chi \in C^\infty (\mathbb{R}^d)$ with the support in a sufficiently small neighborhood of $x_0$, then with $v_\ell = \chi u_\ell$ for
$\ell =1,2$, we have
\begin{multline}\label{thm-WP-cc}
\operatorname{div} (A(x_0) \nabla v_\ell) - \lambda \Sigma_\ell (x_0) v_\ell
\\[6pt]
= \chi \sum_{i,j=1}^d \partial^2_{ij} \Big ( ( A_{ij}(x_0) - A_{ij}(x) ) u_\ell \Big ) - \sum_{i,j=1}^d \chi \partial_i \Big ( u_\ell \partial_{j} (A_{ij}(x_0) - A_{ij}(x) ) \Big) \\[6pt]
+ \chi f_\ell + \lambda \big( \Sigma_\ell (x) - \Sigma_\ell (x_0) \big) v_\ell - u_\ell \operatorname{div} (A (x_0) \nabla \chi ) + 2 \operatorname{div} (u_\ell A (x_0) \nabla \chi).
\end{multline}
The terms $r_{\ell, ij}$ are then $( A_{ij}(x_0) - A_{ij}(x) ) \chi u_\ell = ( A_{ij}(x_0) - A_{ij}(x) ) v_\ell $.
Since $A_1 = A_2 = A$ in $\Omega_\tau$, $u_1 - u_2 = 0$ in $\Gamma$,
and $A \nabla (u_1 -u_2) \cdot \nu = 0$ on $\Gamma$, it follows that
$$
v_1 - v_2 = 0 \mbox{ on } \Gamma \quad \mbox{ and } \quad A(x_0) \nabla (v_1 - v_2) \cdot \nu = \chi (A(x_0)-A(x))\nabla (u_1-u_2) \cdot \nu \mbox{ on } \Gamma.
$$
We are thus in the situation to apply \Cref{lem-vwp} and the freezing coefficient technique to derive \eqref{thm-WP-pr1}.
Concerning \eqref{thm-WP-pr1-bis}, in \eqref{thm-WP-cc}, one writes $\partial^2_{ij} \Big ( ( A_{ij}(x_0) - A_{ij}(x) ) u_\ell(x) \Big ) $ under the form
$$
\partial_{i} \Big ( ( A_{ij}(x_0) - A_{ij}(x) ) \partial_j u_\ell \Big ) + \partial_i \Big( \partial_j ( A_{ij}(x_0) - A_{ij}(x) ) u_\ell \Big).
$$
We are thus in the situation to apply \Cref{lem-vwp} and the freezing coefficient technique to derive \eqref{thm-WP-pr1-bis}.
The details of the rest of the proof of \eqref{thm-WP-pr1} and \eqref{thm-WP-pr1-bis} are omitted.
On the other hand, since
\[
\operatorname{div} (A_\ell \nabla u_\ell) -\lambda \Sigma_\ell u_\ell = f_\ell \quad \mbox{ in }\Omega,
\]
we have, for $|\lambda| \ge 1$,
\begin{equation}
\label{thm-WP-pr2}
\|u_\ell\|_{W^{1, p}_\lambda(\Omega \setminus \Omega_{\tau_*/4})} \leq C \Big( |\lambda|^{-1/2}\|f_\ell\|_{L^p(\Omega)} + \|u_\ell \|_{L^p(\Omega_{\tau_*})} \Big),
\end{equation}
and
\begin{equation}
\label{thm-WP-pr3}
\|u_\ell\|_{W^{2, p}_\lambda(\Omega \setminus \Omega_{\tau_*/2})} \leq C \Big( \|f_\ell\|_{L^p(\Omega)} + \|u_\ell \|_{W^{1, p}_\lambda(\Omega_{\tau_*} \setminus \Omega_{\tau_*/4})} \Big).
\end{equation}
Combining \eqref{thm-WP-pr2} and \eqref{thm-WP-pr3} yields
\begin{equation}
\label{thm-WP-pr23}
\|u_\ell\|_{W^{2, p}_\lambda(\Omega \setminus \Omega_{\tau_*/2})} \leq C \left (\|f_\ell\|_{L^p(\Omega)} + \|u_\ell \|_{L^p(\Omega_{\tau_*})} \right ).
\end{equation}
From \eqref{thm-WP-pr1} and \eqref{thm-WP-pr23}, we obtain
\begin{equation*}
|\lambda| \|(u_1,u_2) \|_{L^p(\Omega)}+\|u_1-u_2\|_{W^{2,p}_\lambda (\Omega)} \le C \|(f_1,f_2)\|_{L^p(\Omega)}
\end{equation*}
for $|\lambda| \ge \lambda_0$ and for $\lambda_0$ large enough. This completes the proof of \eqref{thm-WP-1}.
From \eqref{thm-WP-pr1-bis}, \eqref{thm-WP-pr2}, after using \eqref{thm-WP-1}, we obtain
\begin{equation*}
|\lambda| \|(u_1,u_2) \|_{W^{1, p}_\lambda(\Omega)}+\|u_1-u_2\|_{W^{3,p}_\lambda (\Omega_{\tau_*})} \le C |\lambda|^{1/2} \|(f_1,f_2)\|_{L^p(\Omega)} + \|f_1 - f_2\|_{W^{1, p}_\lambda(\Omega)}
\end{equation*}
for $|\lambda| \ge \lambda_0$ and for $\lambda_0$ large enough. This completes the proof of \eqref{thm-WP-2}.
\medskip
\noindent {\underline{\it Step 2:}} Set
\begin{multline*}
X = \Big \{(u_1,u_2) \in [L^p(\Omega)]^2 : \operatorname{div} (A_1\nabla u_1),\operatorname{div} (A_2\nabla u_2) \in L^p(\Omega), \\
u_1-u_2 \in W^{2,p}(\Omega), u_1-u_2 = 0 \mbox{ on } \Gamma, \mbox{ and }(A_1\nabla u_1 - A_2 \nabla u_2)\cdot \nu = 0 \mbox{ on } \Gamma \Big \}.
\end{multline*}
The space $X$ is a Banach space endowed with the norm
\begin{equation}
\|(u_1,u_2)\|_X : = \|(u_1,u_2)\|_{L^p(\Omega)} + \|\operatorname{div} (A_1\nabla u_1),\operatorname{div} (A_2\nabla u_2) \|_{L^p(\Omega)} + \|u_1-u_2\|_{W^{2,p}(\Omega)}.
\end{equation}
Define
$$
B_\lambda : X \to [L^p(\Omega)]^2
$$
by
\[
B_\lambda(u_1,u_2) = (\operatorname{div} (A_1\nabla u_1)-\lambda \Sigma_1u_1, \operatorname{div} (A_2\nabla u_2)-\lambda \Sigma_2u_2) .
\]
Clearly, $B_\lambda$ is bilinear and continuous on $X$.
We claim that
\begin{equation}\label{thm-WP-claim}
\mbox{$B_\lambda$ has a closed and dense range}.
\end{equation}
Assuming this, we derive that
\begin{equation}\label{thm-WP-claim-1}
B_\lambda(X) = [L^p(\Omega)]^2,
\end{equation}
which yields the existence of the solutions.
\medskip
It remains to prove \eqref{thm-WP-claim}.
\medskip
We first prove that $B_\lambda$ has a closed range. Let $((u_{1,n},u_{2,n}))_n \subset X$ be such that $(f_{1,n},f_{2,n}): = B_\lambda(u_{1,n},u_{2,n}) \to (f_1,f_2)$ in $[L^p(\Omega)]^2$. It follows from \eqref{thm-WP-1} by Step 1 that $((u_{1,n},u_{2,n}))_n$ is a Cauchy sequence in $X$. Let $(u_1, u_2)$ denote its limit. One can then show that $(f_{1,n},f_{2,n}) \to (f_1, f_2) : = B_\lambda(u_1,u_2) $ since $B_\lambda$ is continuous. Thus $B_\lambda$ has a closed range.
\medskip
We next establish that the range of $B_\lambda$ is dense. To this end, it suffices to show that if $(f_1,f_2) \in [L^q(\Omega)]^2$ with $\frac{1}{p}+\frac{1}{q} = 1$ is such that
\begin{equation}
\label{thm-ex-111}
\int_{\Omega} \langle B_{\lambda}(u_1,u_2),(f_1,f_2)\rangle dx=0 \quad \quad \mbox{ for all }(u_1,u_2) \in X,
\end{equation}
then $(f_1, f_2) = (0, 0)$.
Since \eqref{thm-ex-111} holds for all $(u_1, u_2) \in [C_c^\infty (\Omega)]^2 \subset X$, it follows that, for $\ell =1,2$,
\begin{equation}
\label{thm-ex-12}
\operatorname{div} (A_\ell\nabla f_\ell) -\overline{\lambda}\Sigma_\ell f_\ell = 0 \mbox{ in }\Omega.
\end{equation}
Since $A_\ell \in C^1(\bar \Omega)$ and $f_\ell \in L^q(\Omega)$, using the standard regularity theory in $L^q$-scale, see also \cite[Lemma 17.1.5]{Hor3}, one has
$$
f_\ell \in W^{2, q}_{_{loc}} (\Omega).
$$
Set, in $\Omega$,
\begin{equation}\label{thm-ex-def-g}
g_1 = f_1 \quad \mbox{ and } \quad g_2 = - f_2.
\end{equation}
Then, by \eqref{thm-ex-12},
\begin{equation}
\label{thm-ex-13}
\operatorname{div} (A_\ell\nabla g_\ell) - \overline{\lambda}\Sigma_\ell g_\ell = 0 \mbox{ in } \Omega,
\end{equation}
and, by \eqref{thm-ex-111}, for $(u_1,u_2) \in X$,
\begin{equation}\label{thm-ex-pp0}
\int_{\Omega} \operatorname{div}(A_1 \nabla u_1) \bar g_1 - \lambda \Sigma_1 u_1 \bar g_1 - \int_{\Omega} \operatorname{div}(A_2 \nabla u_2) \bar g_2 - \lambda \Sigma_2 u_2 \bar g_2 = 0.
\end{equation}
From \eqref{thm-ex-pp0}, we have, taking $(u_1,u_2) \in X \cap [W^{2,p}(\Omega)]^2$,
\begin{multline}\label{thm-ex-pp1}
\int_{\Omega} \operatorname{div}\big(A_1 \nabla (u_1-u_2) \big) \bar g_1 + \operatorname{div}(A_1 \nabla u_2 ) (\bar g_1 - \bar g_2) + \operatorname{div} \big( (A_1- A_2) \nabla u_2 \big) \bar g_2
\\[6pt]- \lambda \Sigma_1 u_1 \bar g_1 +
\lambda \Sigma_2 u_2 \bar g_2 = 0.
\end{multline}
Using that $g_2 \in W^{2,q}_{loc}(\Omega)$ and $A_1 = A_2$ in $\Omega_\tau$, an integration by parts leads to
\begin{equation}\label{thm-ex-pp2}
\int_{\Omega} \operatorname{div} \big( (A_1- A_2) \nabla u_2 \big) \bar g_2 = \int_{\Omega}
\operatorname{div} \big( (A_1- A_2) \nabla \bar g_2 \big) u_2.
\end{equation}
Since $u_1 - u_2 \in W^{2, p}(\Omega)$, $u_1- u_2 = 0$ on $\Gamma$ and $A\nabla (u_1 - u_2) \cdot \nu = 0$ on $\Gamma$, there exists a sequence $(v_n)_n \subset C^2_c(\Omega)$ such that $v_n \to u_1 - u_2$ in $W^{2, p}(\Omega)$. An integration by parts yields
\begin{multline}\label{thm-ex-pp3}
\int_{\Omega} \operatorname{div}\big(A_1 \nabla (u_1-u_2) \big) \bar g_1
= \lim_{n \to + \infty} \int_{\Omega} \operatorname{div}\big(A_1 \nabla v_n \big) \bar g_1
\\[6pt]
= \lim_{n \to + \infty} \int_{\Omega} \operatorname{div}\big(A_1 \nabla \bar g_1 \big) v_n = \int_{\Omega} \operatorname{div}\big(A_1 \nabla \bar g_1 \big) (u_1 - u_2).
\end{multline}
Combining \eqref{thm-ex-pp1}, \eqref{thm-ex-pp2}, and \eqref{thm-ex-pp3} yields
\begin{multline}\label{thm-ex-pp4}
\int_{\Omega} \operatorname{div}(A_1 \nabla u_2 ) (\bar g_1 - \bar g_2)
= - \int_{\Omega}
\operatorname{div} \big( (A_1- A_2) \nabla \bar g_2 \big) u_2 - \int_{\Omega} \operatorname{div}\big(A_1 \nabla \bar g_1 \big) (u_1 - u_2) \\[6pt]
+ \int_{\Omega} \lambda \Sigma_1 u_1 \bar g_1 - \int_{\Omega}
\lambda \Sigma_2 u_2 \bar g_2 \mathop{=}^{\eqref{thm-ex-13}} \int_{\Omega} \operatorname{div}(A_1 \nabla (\bar g_1 - \bar g_2)) u_2.
\end{multline}
Since $u_2$ can be chosen arbitrary\footnote{Taking then $u_1:= u_2$ so that $(u_1,u_2) \in X$.} in $W^{2, p}(\Omega)$, and for every $\xi \in [C^1(\bar \Omega)]^d$ there exists $u_2 \in W^{2, p}(\Omega)$ with $u_2|_{\Gamma} =0$ such that $\operatorname{div}(A_1 \nabla u_2 ) = \operatorname{div} \xi $ with $\| u_2 \|_{W^{2, p}(\Omega)} \le C \|\xi \|_{L^p(\Omega)}$,
it follows that, see, e.g., \cite[Proposition 9.18]{Brezis-FA},
$$
g_1 - g_2 \in W^{1, q}_0 (\Omega).
$$
This in turn implies, by \eqref{thm-ex-pp4}, that $g_1 - g_2 \in W^{2, q}(\Omega)$ and $A \nabla ( g_1 - g_2) \cdot \nu = 0$ on $\Gamma$. It follows that $g_1 = g_2 = 0$ in $\Omega$ after applying Step 1 to $(g_1, g_2)$ and $\bar \lambda$ (instead of $\lambda$). Thus $f_1 = f_2 = 0$ by \eqref{thm-ex-def-g} and the proof of Step 2 is complete.
\qed
\section{The Weyl law for the transmission eigenvalues} \label{sect-W}
Throughout this section, we assume \eqref{condi3}.
\subsection{The operator $T_\lambda$ and its adjoint $T_{\lambda}^*$.}
We first reformulate the Cauchy problem \eqref{ITE-C} in a form for which we can apply the theory of Hilbert-Schmidt operators. Given $(f, g) \in [L^p(\Omega)]^2$ ($1<p<+\infty$), assume that $(u_1,u_2) \in [L^p(\Omega)]^2$ with $u_1-u_2 \in W^{2,p}(\Omega)$ is a solution of \eqref{ITE-C} with $\lambda \in \mathbb{C}^*$ where, in $\Omega$,
\[
f_1= \Sigma_1 f+\lambda^{-1}\Sigma_2 g\quad \quad \mbox{ and } \quad \quad f_2 = \lambda^{-1}\Sigma_2 g.
\]
Define, in $\Omega$,
\[
u = u_1-u_2 \quad \quad \quad \mbox{ and }\quad \quad \quad v = \lambda u_2.
\]
Then the pair $(u,v) \in W^{2,p}(\Omega) \times L^p(\Omega)$ is a solution of
\begin{equation}
\label{ITE-Cauchy-2}
\left \{
\begin{array}{rll}
\operatorname{div} (A \nabla u) -\lambda \Sigma_1 u - (\Sigma_1-\Sigma_2)v &= \Sigma_1 f & \text{ in }\Omega ,\\[6pt]
\operatorname{div} (A \nabla v) -\lambda \Sigma_2 v &= \Sigma_2 g & \text{ in }\Omega ,\\[6pt]
u = 0, \: \: A\nabla u\cdot \nu &=0 & \text{ on }\Gamma.
\end{array}
\right .
\end{equation}
As a direct consequence of \Cref{thm-WP} (see also \eqref{thm-WP-2}), we have
\begin{prop}
\label{prop-pre}
Assume \eqref{condi1}-\eqref{condi2}, and \eqref{condi3}-\eqref{condi5}. Let $c \in (0, 1)$ and $1<p<+\infty$. There exists $\lambda_0>0$ depending on $p$, $c$, $\Lambda$, and $\Omega$ such that the following holds: for $(f,g) \in [L^p(\Omega)]^2$
and for $\lambda \in \mathbb{C}$ with $|\Im (\lambda) | \geq c |\lambda| $ and $|\lambda|>\lambda_0$, there exists a unique solution $(u,v) \in W^{2,p}(\Omega) \times L^p(\Omega)$ of the Cauchy problem \eqref{ITE-Cauchy-2}; moreover, we have
\begin{equation}\label{prop-pre-p1}
\|v\|_{L^p (\Omega) } +\|u\|_{W^{2,p}_{\lambda}(\Omega)} \leq C|\lambda|^{-1/2} \left ( |\lambda|^{1/2}\|f\|_{L^p(\Omega)} + |\lambda|^{-1/2}\|g\|_{L^p(\Omega) } \right )
\end{equation}
and
\begin{equation}\label{prop-pre-p2}
\|v\|_{ W_{\lambda}^{1,p}(\Omega)} + \|u\|_{W^{3,p}_\lambda (\Omega)}\leq C \left ( \|f\|_{W^{1,p}_\lambda(\Omega)}+|\lambda|^{-1/2}\| g \|_{ L^p(\Omega)} \right ),
\end{equation}
for some positive constant $C$ independent of $\lambda$, $f$, and $g$.
\end{prop}
As a consequence, we have
\begin{cor}\label{cor-reg} Assume \eqref{condi1}-\eqref{condi2} and \eqref{condi3}-\eqref{condi5}. Let $c \in (0, 1)$, and $1 < p < + \infty$. There exists $\lambda_0>0$ depending on $p$, $c$, $\Lambda$, and $\Omega$ such that the following holds: for $(f, g) \in W^{1, p}(\Omega) \times L^p(\Omega)$, and for $\lambda \in \mathbb{C}$ with $|\Im (\lambda) | \geq c |\lambda| $ and $|\lambda|>\lambda_0$, there exists a unique solution $(u, v) \in W^{3,p}(\Omega) \times W^{1, p}(\Omega)$ of \eqref{ITE-Cauchy-2}; moreover, for
\begin{enumerate}
\item either $1 < p < d$ and $p \le q \le \frac{dp}{d-p}$,
\item either $d=p \le q < + \infty$,
\item or $p> d$ and $q = + \infty$,
\end{enumerate}
we have
\begin{equation}
\|v\|_{ L^q (\Omega)} +\|u\|_{W^{2,q}_\lambda(\Omega)}\leq C |\lambda|^{\frac{d}{2} \left ( \frac{1}{p}-\frac{1}{q} \right ) -\frac{1}{2}} \left ( \|f\|_{W^{1,p}_\lambda (\Omega)}+|\lambda|^{-1/2} \|g\|_{ L^p(\Omega)} \right ),
\end{equation}
for some positive constant $C$ independent of $\lambda$, $f$, and $g$.
\end{cor}
\begin{rem}\label{rem-cont} \rm
In case (3) of \Cref{cor-reg}, we derive that $(u,v) \in C^2(\overline{\Omega}) \times C(\overline{\Omega})$.
\end{rem}
\begin{proof}
Choose $\lambda_0$ such that the conclusion of \Cref{prop-pre} holds. By Gagliardo-Nirenberg's interpolation inequalities (see \cite{Gagliardo59,Nirenberg59}), we have
\begin{equation*}
\| v\|_{L^q(\Omega)} \le C_{p, q, \Omega} \| v\|_{L^p(\Omega)}^{1 - a} \| v\|_{W^{1, p}(\Omega)}^{a} \leq C_{p, q, \Omega} \| v\|_{L^p(\Omega)}^{1 - a} \| v\|_{W^{1, p}_\lambda(\Omega)}^{a},
\end{equation*}
where
$$
a = d \left(\frac{1}{p} - \frac{1}{q} \right).
$$
This implies
\begin{equation*}
\| v\|_{L^q(\Omega)} \le C_{p, q, \Omega} |\lambda|^{- \frac{1}{2} (1 - a)} \left ( \|f\|_{W^{1,p}_\lambda(\Omega)}+|\lambda|^{-1/2}\| g \|_{ L^p(\Omega)} \right ).
\end{equation*}
The other assertions can be proved similarly.
\end{proof}
\begin{definition} Assume \eqref{condi1}-\eqref{condi2} and \eqref{condi3}-\eqref{condi5}. Let $1 < p < + \infty$ and $\lambda \in \mathbb{C}$. System \eqref{ITE-Cauchy-2} is said to be {\rm well-posed} in $L^p(\Omega) \times L^p(\Omega)$ if the existence, the uniqueness, and \eqref{prop-pre-p1} and \eqref{prop-pre-p2} hold for $(f, g) \in L^p(\Omega) \times L^p(\Omega)$. For $p=2$ and $\lambda \in \mathbb{C}$ being such that \eqref{ITE-Cauchy-2} is well-posed in $L^2(\Omega) \times L^2(\Omega)$, we define
\begin{equation}
\label{pre-Tl}
\begin{array}{rccc}
T_\lambda : & L^2(\Omega) \times L^2(\Omega) &\to & L^2(\Omega) \times L^2(\Omega) \\[6pt]
& (f,g) & \mapsto & (u,v)
\end{array}
\end{equation}
where $(u,v)$ is the unique solution of \eqref{ITE-Cauchy-2}.
\end{definition}
\begin{rem}\rm Let $\lambda \in \mathbb{C}$ satisfy the conclusion of \Cref{prop-pre} with $p=2$. Then system \eqref{ITE-Cauchy-2} is well-posed in $L^2(\Omega) \times L^2(\Omega)$ and $T_\lambda$ is defined.
\end{rem}
\begin{rem} \label{rem-comp}\rm Let $\lambda^* \in \mathbb{C}$ be such that $T_{\lambda}^*$ is defined. If $\mu$ is a characteristic value of the operator $T_{\lambda^*}$ associated with an eigenfunction $(u,v)$ and if $\lambda^*+\mu \neq 0$ we have
\begin{equation}
\label{rem-comp-0}
\lambda^*+ \mu \mbox{ is a transmission eigenvalue of \eqref{ITE} }
\end{equation}
with an eigenfunction pair $(u_1,u_2)$ given by
\begin{equation}
\label{rem-comp-1}
u_1 = u + \frac{1}{\lambda^* + \mu} v \quad \quad \mbox{ and }\quad \quad u_2 = \frac{1}{\lambda^* + \mu} v.
\end{equation}
Moreover, the converse holds (see \Cref{rem-indep}).
\end{rem}
\begin{rem} \rm Let $\lambda^* \in \mathbb{C}$ be such that $T_{\lambda}^*$ is defined. By \eqref{prop-pre-p1} and \eqref{prop-pre-p2} the range of $T_{\lambda^*}^2$ is a subset of $H^1(\Omega)\times H^1(\Omega)$. It follows that the operator $T_{\lambda^*}^2$ is compact from $L^2(\Omega) \times L^2(\Omega)$ into itself. By the spectral theory of compact operators, see, e.g., \cite{Brezis-FA}, the spectrum of $T_{\lambda^*}^2$ consists in a discrete set of eigenvalues and the generalized eigenspace associated to each eigenvalue is of finite dimension. As a consequence, the set of eigenvalues of $T_{\lambda^*}$ is discrete. This in turn implies that the set of the transmission eigenvalues of \eqref{ITE} is discrete. This fact is previously established in \cite{MinhHung1} but the arguments presented here are different.
\end{rem}
\begin{rem}\label{rem-indep}\rm
Let $\lambda^* \in \mathbb{C}$ be such that $T_{\lambda}^*$ is defined.
If $\lambda_j$ is an eigenvalue of the transmission eigenvalue problem, then $\lambda_j \neq \lambda^*$ and $\lambda_j - \lambda^*$ is a characteristic value of $T_{\lambda^*}$. One can show that the multiplicity of the characteristic values of $\lambda_j - \lambda^*$ and $\lambda_j - \hat{\lambda}$ associated with $T_{\lambda^*}$ and $T_{\hat{\lambda}}$ are the same. Hence the multiplicity of the eigenvalues associated with $T_{\lambda^*}$ is independent of $\lambda^*$. With this observation we define the multiplicity of $\lambda_j$ as the one of the characteristic value $\lambda_j - \lambda^*$ of $T_{\lambda^*}$.
\end{rem}
The rest of this section is devoted to characterize the adjoint $T_\lambda^*$ of $T_\lambda$. This will be used in the proof of \Cref{prop-p}. To this end, for $(\widetilde{f},\widetilde{g}) \in [L^p(\Omega)]^2$ with $1 < p < + \infty$, we consider the system, for $(\widetilde{u}, \widetilde{v}) \in W^{1, p}(\Omega) \times L^p(\Omega)$ \footnote{We emphasize here that in the first equation of \eqref{ITE-Cauchy-3}, we have $\Sigma_2 \widetilde{u}$ not $\Sigma_1 \widetilde{u}$, compare with \eqref{ITE-Cauchy-2}.}
\begin{equation}\label{ITE-Cauchy-3}
\left \{\begin{array}{cl}
\operatorname{div} (A \nabla \widetilde{u}) -\lambda \Sigma_2 \widetilde{u} - (\Sigma_1-\Sigma_2)\widetilde{v} = \Sigma_2 \widetilde{f} & \text{ in }\Omega ,\\[6pt]
\operatorname{div} (A \nabla \widetilde{v}) -\lambda \Sigma_1 \widetilde{v} = \Sigma_1 \widetilde{g} & \text{ in }\Omega ,\\[6pt]
\widetilde{u} = 0, \quad A\nabla \widetilde{u}\cdot \nu =0 & \text{ on }\Gamma.
\end{array}\right .
\end{equation}
Assume \eqref{condi1}-\eqref{condi2}, and \eqref{condi3}-\eqref{condi5}. Let $c \in (0, 1)$ and $1<p<+\infty$. By \Cref{prop-pre}, there exists $\lambda_0>0$ depending on $p$, $c$, $\Lambda$, and $\Omega$ such that \eqref{ITE-Cauchy-3} is well-posed in $L^p(\Omega) \times L^p(\Omega)$ for $\lambda \in \mathbb{C}$ with $|\Im (\lambda) | \geq c |\lambda| $ and $|\lambda|>\lambda_0$, i.e., for $(f, g) \in L^p(\Omega) \times L^p(\Omega)$, there exists a unique solution $(\widetilde{u}, \widetilde{v}) \in W^{1,p}(\Omega) \times L^p(\Omega)$ of \eqref{ITE-Cauchy-3}; moreover,
\begin{equation*}
\|\widetilde{u}\|_{W^{2,p}_{\lambda}(\Omega)} + \|\widetilde{v}\|_{L^p (\Omega) }\leq C|\lambda|^{-1/2} \left ( |\lambda|^{1/2}\|\widetilde{f}\|_{L^p(\Omega)} + |\lambda|^{-1/2}\|\widetilde{g}\|_{L^p(\Omega) } \right )
\end{equation*}
and
\begin{equation*}
\|\widetilde{u}\|_{W^{3,p}_\lambda (\Omega)} + \|\widetilde{v}\|_{ W_{\lambda}^{1,p}(\Omega)} \leq C \left ( \|\widetilde{f}\|_{W^{1,p}_\lambda(\Omega)}+|\lambda|^{-1/2}\| \widetilde{g} \|_{ L^p(\Omega)} \right).
\end{equation*}
\begin{definition} Assume \eqref{condi1}-\eqref{condi2} and \eqref{condi3}-\eqref{condi5}. For $p=2$ and for $\lambda \in \mathbb{C}$ being such that \eqref{ITE-Cauchy-3} is well-posed in $L^2(\Omega) \times L^2(\Omega)$, we define
\begin{equation}
\begin{array}{rccc}
\widetilde{T}_\lambda : & L^2(\Omega) \times L^2(\Omega) &\to & L^2(\Omega) \times L^2(\Omega) \\[6pt]
& (\widetilde{f},\widetilde{g}) & \mapsto & (\widetilde{u}, \widetilde{v})
\end{array}
\end{equation}
where $(\widetilde{u}, \widetilde{v})$ is the unique solution of \eqref{ITE-Cauchy-3}.
\end{definition}
\begin{lem}\label{lem-T*}
Assume \eqref{condi1}-\eqref{condi2} and \eqref{condi3}-\eqref{condi5}. Let $p=2$ and let $\lambda$ be such that $T_\lambda$ and $\widetilde{T}_{\bar \lambda}$ are defined. Set, for $x \in \Omega$,
\begin{equation}
\label{pre-7}
P(x) =
\begin{pmatrix}
0 & \Sigma_1(x) \\[6pt]
\Sigma_2(x) & 0
\end{pmatrix}.
\end{equation}
We have
\begin{equation}
T_\lambda^* = P \widetilde{T}_{\overline{\lambda}} P^{-1}.
\end{equation}
\end{lem}
\begin{proof} Fix $(f, g) \in [L^2(\Omega)]^2$ and $(f^*,g^*) \in [L^2(\Omega)]^2$. Set $(u, v) = T_{\lambda}(f, g)$ and $(u^*,v^*) = \widetilde{T}_{\overline{\lambda}} P^{-1}(f^*,g^*)$. Then
\begin{equation}\label{lem-T*-p1}
\int_{\Omega} \langle (f,g), P \widetilde{T}_{\overline{\lambda}} P^{-1}(f^*,g^*) \rangle = \int_\Omega \Sigma_1 f \overline{v^*}+\Sigma_2 g \overline{u^*}.
\end{equation}
Since $(u, v) = T_{\lambda}(f, g)$, we have
\begin{equation}\label{lem-T*-p2}
\int_\Omega \Sigma_1 f \overline{v^*}+\Sigma_2 g \overline{u^*}
= \int_{\Omega} (\operatorname{div} (A\nabla u) -\lambda \Sigma_1 u - (\Sigma_1-\Sigma_2) v ) \overline{v^*} + \int_\Omega ( \operatorname{div} (A\nabla v ) -\lambda \Sigma_2 v ) \overline{u^*}.
\end{equation}
As in Step 2 of the proof of \Cref{thm-WP}, an integration by parts yields
\begin{multline}\label{lem-T*-p3}
\int_{\Omega} (\operatorname{div} (A\nabla u) -\lambda \Sigma_1 u - (\Sigma_1-\Sigma_2) v ) \overline{v^*} + \int_\Omega ( \operatorname{div} (A\nabla v ) -\lambda \Sigma_2 v ) \overline{u^*} \\[6pt]
= \int_{\Omega} u (\overline{\operatorname{div} (A\nabla v^*)-\overline{\lambda} \Sigma_1 v^*)} + \int_{\Omega} v (\overline{\operatorname{div} (A\nabla u^*)-\overline{\lambda} \Sigma_2 u^*-(\Sigma_1-\Sigma_2)v^*)}.
\end{multline}
Since $(u^*,v^*) = \widetilde{T}_{\overline{\lambda}} P^{-1}(f^*,g^*)$, we have
\begin{multline}\label{lem-T*-p4}
\int_{\Omega} u (\overline{\operatorname{div} (A\nabla v^*)-\overline{\lambda} \Sigma_1 v^*)} + \int_{\Omega} v (\overline{\operatorname{div} (A\nabla u^*)-\overline{\lambda} \Sigma_2 u^*-(\Sigma_1-\Sigma_2)v^*)} \\[6pt]
= \int_\Omega \langle T_\lambda (f,g), (f^*,g^*) \rangle .
\end{multline}
Combining \eqref{lem-T*-p1}-\eqref{lem-T*-p4} yields
\begin{equation}
\int_{\Omega} \langle (f,g), P \widetilde{T}_{\overline{\lambda}} P^{-1}(f^*,g^*) \rangle = \int_\Omega \langle T_\lambda (f,g), (f^*,g^*) \rangle,
\end{equation}
and the conclusion follows.
\end{proof}
\subsection{Hilbert-Schmidt operators}
In this section, we recall the definition and several properties of Hilbert-Schmidt operators. We begin with
\begin{definition}
\label{def-HS}
Let $H$ be a separable Hilbert space and let $(\phi_j)_{j=1}^\infty$ be an orthonormal basis of $H$.
\begin{enumerate}
\item Let $\mathcal T$ be a linear and bounded operator on $H$. We say that $\mathcal T$ is {\rm Hilbert-Schmidt} if its {\rm double norm} is finite, i.e.
\[
\vvvert \mathcal T \vvvert := \left ( \sum_{j=1}^\infty \|\mathcal T \phi_j\|_H^2 \right )^{1/2} < + \infty.
\]
\item Let $\mathcal T_1$ and $\mathcal T_2$ be two Hilbert-Schmidt operators on $H$. The {\rm trace} of the composition $\mathcal T_1 \mathcal T_2$ is defined by
\[
{\rm trace}(\mathcal T_1 \mathcal T_2) := \sum_{j=1}^\infty ( \mathcal T_1 \mathcal T_2 \phi_j,\phi_j )_H.
\]
\end{enumerate}
\end{definition}
\begin{rem} \rm One can check that \Cref{def-HS} does not depend on the choice of the basis $(\phi_j)_{j=1}^\infty$ and the trace of $\mathcal T_1 \mathcal T_2$ is well defined as an absolutely convergent series (see \cite[Theorems 12.9 and 12.12]{Agmon}).
\end{rem}
Let $m \in \mathbb{N}$ and ${\bf T} : [L^2(\Omega)]^m \to [L^2(\Omega)]^m$ be a Hilbert-Schmidt operator. There exists a unique kernel ${\rm \bold{K}} \in [L^2(\Omega \times \Omega)]^{m \times m}$, see e.g. \cite[Theorems 12.18 and 12.19]{Agmon}, such that
\begin{align}\label{Z2}
( \mathbf{T}u)(x)=\int_\Omega {\rm \bold{K}}(x,y) u(y) dy ~~~\mbox{ for a.e. } \, x\in \Omega, \mbox{ for all } u \in [L^2(\Omega)]^m.
\end{align}
Moreover,
\begin{equation}\label{norm-K-T}
\vvvert\mathbf{T} \vvvert^2=\mathop{\iint}_{\Omega\times\Omega}|{\rm \bold{K}}(x,y)|^2 \, dx \, dy.
\end{equation}
Note that \cite[Theorems 2.18 and 12.19]{Agmon} state for $m=1$, nevertheless, the same arguments hold for $m \in \mathbb{N}$ as noted in \cite{MinhHung2}.
\medskip
We have, see \cite{Agmon} (see also \cite[Lemma 4]{MinhHung2}):
\begin{lemma} \label{lem-T1T2} Let $m \in \mathbb{N}$ and let $\mathbf{T}_1,\mathbf{T}_2$ be two Hilbert-Schmidt operators in $[L^2(\Omega)]^m$ with the corresponding kernels ${\rm \bold{K}}_1$ and ${\rm \bold{K}}_2$. Then ${\bf T}: = \mathbf{T}_1\mathbf{T}_2$ is a Hilbert-Schmidt operator with the kernel $\mathbf{K}$ given by
\begin{equation}\label{lem-T1T2-1}
{\rm \bold{K}}(x, y) = \int_{\Omega} {\rm \bold{K}}_1 (x, z) {\rm \bold{K}}_2 (z, y) \, dz.
\end{equation}
Moreover,
\begin{equation}\label{traceT^2}
{\rm trace} (\mathbf{T}_1\mathbf{T}_2)=\int_{\Omega} {\rm trace}(\mathbf{K}(x,x))dx.
\end{equation}
\end{lemma}
We have, see, e.g., \cite[Lemma 3]{MinhHung2}.
\begin{lemma}\label{lem-HS1} Let $d \ge 2$, $m \in \mathbb{N}$, and ${\bf T} : [L^2(\Omega)]^m \to [L^2(\Omega)]^m$ be such that $\mathbf{T}(\phi)\in [C(\bar \Omega)]^m$ for $\varphi \in [L^2(\Omega)]^m$, and
\begin{equation}\label{lem-HS1-st0}
\|\mathbf{T}(\phi) \|_{L^\infty (\Omega)}\leq M \|\phi \|_{L^2 (\Omega)},
\end{equation}
for some $M \ge 0$.
Then $\mathbf{T}$ is a Hilbert-Schmidt operator,
\begin{align}\label{lem-HS1-st1}
\vvvert \mathbf{T} \vvvert \leq C_m |\Omega|^{1/2}M,
\end{align}
and the kernel ${\rm \bold{K}}$ of ${\bf T}$ satisfies
\begin{equation}
\label{lem-HS1-st1'}
\sup_{x\in \Omega}\left(\int_{\Omega}|{\rm \bold{K}}(x,y)|^2 dy\right)^{1/2}\leq C_m |\Omega|^{1/2}M.
\end{equation}
Assume in addition that
\begin{equation}\label{lem-HS1-st2}
\| \mathbf{T}(\phi) \|_{L^\infty (\Omega)}\leq \widetilde{M}||\phi||_{L^1 (\Omega)} \mbox{ for } \phi\in [L^2(\Omega)]^m,
\end{equation}
for some $\widetilde{M} \ge 0$,
then the kernel ${\rm \bold{K}}$ of ${\bf T}$ satisfies
\begin{align}\label{lem-HS1-st3}
|{\rm \bold{K}}(x,y)|\leq \widetilde{M} ~~ \mbox{ for a.e. } x, y \in \Omega.
\end{align}
Here $C_m$ denotes a positive constant depending only on $m$.
\end{lemma}
As a consequence of \Cref{lem-HS1}, we derive the following result.
\begin{cor}
\label{cor-K-cont}
Let $\mathbf{T}_1$ and $\mathbf{T}_2$ two Hilbert-Schmidt operators on $[L^2(\Omega)]^m$ be such that the ranges of $\mathbf{T}_1$ and $\mathbf{T}_2^*$ are in $[C(\bar{\Omega})]^m$ and \eqref{lem-HS1-st0} holds for $\mathbf{T}_1$ and $\mathbf{T}_2^*$. Assume that \eqref{lem-HS1-st2} holds for $\mathbf{T}= \mathbf{T}_1 \mathbf{T}_2$. Then the kernel $\mathbf{K}$ of $\mathbf{T}$ is continuous on $\bar \Omega \times \bar \Omega$ and \eqref{lem-HS1-st3} holds for every $(x,y) \in \Omega \times \Omega$.
\end{cor}
\begin{proof}
Let $\mathbf{K}_1$ (resp. $\mathbf{K}_2$) be the kernel of $\mathbf{T}_1$ (resp. $\mathbf{T}_2$) and let $\mathbf{K}_2^*$ be the kernel of $\mathbf{T}_2^*$. We claim that for $\varepsilon>0$, there exists $\delta>0$ such that for every $(x, x')\in \Omega \times \Omega$ with $|x-x'|<\delta$ we have
\begin{equation}
\label{cor-K-cont-1}
\left ( \int_{\Omega}|\mathbf{K}_1(x,z)-\mathbf{K}_1(x',z)|^2 dz \right )^{1/2} \le \varepsilon \quad \mbox{ and } \quad \left ( \int_{\Omega}|\mathbf{K}_2^*(x,z)-\mathbf{K}_2^*(x',z)|^2 dz \right )^{1/2} \le \varepsilon.
\end{equation}
Admitting \eqref{cor-K-cont-1}, we continue the proof.
We have, see, e.g., \cite[Theorem 12.20]{Agmon},
\begin{equation}
\label{cor-K-cont-11}
\mathbf{K}_2(z,y) = \overline{\mathbf{K}_2^*(y,z)}.
\end{equation}
Since
\[
\mathbf{K}(x,y) \mathop{=}^{\eqref{lem-T1T2-1},\eqref{cor-K-cont-11}} \int_\Omega \mathbf{K}_1(x,z) \overline{\mathbf{K}_{2}^*(z,y)} dz,
\]
it follows from \eqref{cor-K-cont-1} that ${\bf K}$ is continuous in $\bar \Omega \times \bar \Omega$. This in turn implies \eqref{lem-HS1-st3} by \Cref{lem-HS1} applied to $\mathbf{T}$.
It remains to prove \eqref{cor-K-cont-1}. We have
\begin{equation*}
\left ( \int_{\Omega}|\mathbf{K}_1(x,z)-\mathbf{K}_1(x',z)|^2 dz \right )^{1/2} \le C \sup_{\varphi \in [L^2(\Omega)]^m; \| \varphi\|_{L^2(\Omega)} \le 1} \left| \int_{\Omega}( \mathbf{K}_1(x,z)-\mathbf{K}_1(x',z) ) \varphi (z) dz \right|.
\end{equation*}
Given $\varepsilon > 0$, let $\varphi_\varepsilon \in [L^{2}(\Omega)]^m$ with $\|\varphi_\varepsilon\|_{L^2(\Omega)}\leq 1$ be such that
\begin{equation*}
\left ( \int_{\Omega}|\mathbf{K}_1(x,z)-\mathbf{K}_1(x',z)|^2 dz \right )^{1/2}\leq \left | \int_\Omega (\mathbf{K}_1 (x, z)-\mathbf{K}_1(x', z))\varphi_\varepsilon (z) dz \right | + \frac{\varepsilon}{2}.
\end{equation*}
This yields
\begin{equation}
\label{cor-K-cont-3}
\left ( \int_{\Omega}|\mathbf{K}_1(x,z)-\mathbf{K}_1(x',z)|^2 dz \right )^{1/2}\leq |(\mathbf{T}_1 \varphi_\varepsilon)(x)-(\mathbf{T}_1 \varphi_\varepsilon)(x')| + \frac{\varepsilon}{2}.
\end{equation}
The first inequality of \eqref{cor-K-cont-1} now follows from \eqref{cor-K-cont-3} and the fact that ${\bf T}_1 \varphi_\varepsilon \in [C(\bar \Omega)]^m$.
\medskip
Similarly, we obtain the second inequality of \eqref{cor-K-cont-1}.
\end{proof}
\subsection{The operators ${\bf T}_{\theta, t}$ and their properties}
Denote
\begin{equation}
\label{def-k}
k= \left [\frac{d}{2} \right ] + 1,
\end{equation}
the smallest integer greater than $d/2$. Fix
\begin{equation}
2=p_1 < p_2 < \cdots < p_{k}<+\infty
\end{equation}
such that
\begin{equation}
p_{j-1}<p_j < \frac{dp_{j-1}}{d-p_{j-1}} \quad \mbox{ and } \quad p_{k}>d.
\end{equation}
Denote
\begin{equation}
\label{def-lambda^*}
\lambda^* = t^*e^{i\frac{\pi}{2}},
\end{equation}
for some large $t^*>0$ such that, for $t \ge t^*$, \eqref{ITE-Cauchy-2} with $\lambda = t e^{i\frac{\pi}{2}}$ is well-posed in $L^p(\Omega) \times L^p(\Omega)$ and \eqref{ITE-Cauchy-3} with $\lambda = t e^{- i\frac{\pi}{2}}$ is well-posed in $L^p(\Omega) \times L^p(\Omega)$
with $p =p_1, \, \cdots, \, p_{k}$.
Let
\begin{equation}\label{omega-j}
\mbox{$\omega_j \in \mathbb{C}$ with $1 \le j \le k+1$ be the (distinct) $(k+1)$-th roots of $1$ (thus $\omega_j^{k+1} = 1$)}
\end{equation}
and let
\begin{equation}
\label{def-gam}
\Theta = \mathbb{R} \setminus \left \{ \frac{\pi}{k+1}\mathbb{Z} \right \}.
\end{equation}
\begin{definition}\label{def-t-theta}
For $\theta \in \Theta$, $1 \le j \le k+1$, and $t>0$, we define
\begin{equation}
\label{p4}
\lambda_{j, \theta , t} = \lambda^* + \omega_j t e^{i\theta},
\end{equation}
and \[t_\theta>t^*\] such that the following properties hold, for $t \ge t_\theta$,
\begin{multline}
\label{unif-t0-1}
\mbox{ \eqref{ITE-Cauchy-2} with $\lambda = \lambda_{j, \theta , t} $ is well-posed in $L^p(\Omega) \times L^p(\Omega)$} \\
\mbox{ and \eqref{ITE-Cauchy-3} with $\lambda = \bar \lambda_{j, \theta , t} $ is well-posed in $L^p(\Omega) \times L^p(\Omega)$ with $p =p_1, \, \cdots, \, p_{k}$},
\end{multline}
and
\begin{equation}
\label{unif-t0-2}
\frac{t}{2} \leq |\lambda_{j, \theta, t}| < 2t.
\end{equation}
\end{definition}
Such a $t_\theta>t^*$ exists by \Cref{prop-pre} after noting that, for $\theta \in \Theta$,
\[
\Im \left (\omega_j e^{i\theta} \right ) \neq 0,
\]
and, for $1 \le j \le k+1$,
\begin{equation*}
\lim_{t \to + \infty} \frac{\left |\Im \left (\lambda^*+ t\omega_j e^{i\theta} \right )\right |}{\left |\lambda^*+ t \omega_j e^{i\theta} \right |} = \left |\Im \left (\omega_j e^{i\theta} \right ) \right |>0.
\end{equation*}
Viewing \eqref{prop-pre-p1} and \eqref{prop-pre-p2}, it is convenient to modify $T_{\lambda_{j, \theta, t}}$ to capture the scaling with respect to $t \sim \lambda_{j,\theta,t}$ there, as in \cite{Robbiano16}. Denote
\begin{equation}\label{def-Mt}
M_t =
\begin{pmatrix}
t^{1/2} & 0 \\
0 & t^{-1/2}
\end{pmatrix}.
\end{equation}
Let $\theta\in \Theta$ and $t \ge t_\theta$. Define, for $1 \le j \le k+1$,
\begin{equation}
\label{p11}
T_{j, \theta, t} = M_t T_{\lambda_{j, \theta, t}} M_{t}^{-1} \quad \quad \mbox{ and } \quad \quad {\bf T}_{\theta, t} = T_{k+1, \theta, t}\circ T_{k, \theta, t} \circ \cdots \circ T_{1, \theta, t}.
\end{equation}
Here is the main result of this section.
\begin{prop}
\label{prop-p} Let $\theta\in \Theta$ and let $t_\theta$ be given in \Cref{def-t-theta}. Then, for $t \ge t_\theta$,
\begin{equation}
\label{prop-p-0}
\|{\bf T}_{\theta,t}\|_{L^2(\Omega)\to L^2(\Omega)} \leq Ct^{-k-1},
\end{equation}
the range of ${\bf T}_{\theta, t}$ is in $[C(\bar{\Omega})]^2$,
\begin{equation}
\label{prop-p-4}
\|{\bf T}_{\theta, t}\|_{L^2 (\Omega) \to L^\infty (\Omega)} \leq C t^{-k-1 + \frac{d}{4}},
\end{equation}
and
\begin{equation}
\label{prop-p-4*}
\|{\bf T}_{\theta, t}\|_{L^1 (\Omega) \to L^2 (\Omega)} \leq C t^{-k-1+\frac{d}{4}}
\end{equation}
for some positive constant $C$ independent of $t$. Similar facts hold for ${\bf T}_{\theta, t}^*$.
\end{prop}
As a direct consequence of \Cref{lem-HS1} and \Cref{prop-p}, we obtain
\begin{cor}\label{cor-p} Let $\theta\in \Theta$ and
let $t_\theta$ be given in \Cref{def-t-theta}. Then, for $t \ge t_\theta$, the operator ${\bf T}_{\theta, t}$ is Hilbert-Schmidt, and
\begin{equation}
\label{prop-p-5}
\vvvert {\bf T}_{\theta, t} \vvvert \leq Ct ^{-k-1+ \frac{d}{4}},
\end{equation}
for some positive constant $C$ independent of $t$.
\end{cor}
We now give
\begin{proof}[Proof of \Cref{prop-p}] We first deal with
\eqref{prop-p-4}.
By using \eqref{unif-t0-2}, we derive that
\[
\|T_{j,\theta,t}\|_{L^2(\Omega) \to L^2(\Omega)}\mathop{\leq}^{{\rm \Cref{prop-pre}}} Ct^{-1}
\]
and hence
\[
\|{\bf T}_{\theta,t}\|_{L^2(\Omega) \to L^2(\Omega)} \leq \prod_{j=1}^{k+1} \|T_{j,\theta,t}\|_{L^2(\Omega) \to L^2(\Omega)} \leq Ct^{-k-1}.
\]
This establishes \eqref{prop-p-0}.
Next we deal with \eqref{prop-p-4}. For $j=1,\cdots,k+1$ and $(f,g) \in [L^2(\Omega)]^2$, we write
\begin{equation}
\label{p12}
(u^{(j)},v^{(j)}) = T_{j, \theta, t}\circ T_{j-1, \theta,t} \circ \cdots \circ T_{1, \theta, t} (f,g).
\end{equation}
By \eqref{unif-t0-2}, we have
\begin{equation}
\label{prop-p-1}
t^{-1/2}\|u^{(1)}\|_{W^{1,2}_t(\Omega)} + \| v^{(1)} \|_{L^2(\Omega)} \mathop{\leq}^{{\rm \Cref{prop-pre}}} C t^{-1} \|(f, g)\|_{L^2(\Omega) \times L^2(\Omega)},
\end{equation}
and, for $2 \le j \le k$,
\begin{multline}
\label{prop-p-2}
t^{-1/2}\|u^{(j)}\|_{W^{1, p_j}_t(\Omega)}+\|v^{(j)}\|_{ L^{p_j}(\Omega)} \\[6pt]
\mathop{\leq}^{{\rm\Cref{cor-reg}}} C t^{-1+\frac{d}{2} \left ( \frac{1}{p_{j-1}} -\frac{1}{p_j} \right )}
\Big( t^{-1/2} \|u^{(j-1)}\|_{W^{1,p_{j-1}}_t(\Omega)} + \|v^{(j-1)}\|_{L^{p_{j-1}}(\Omega)} \Big),
\end{multline}
and
\begin{multline}
\label{prop-p-3}
t^{-1/2}\|u^{(k+1)}\|_{W^{1,\infty}_t(\Omega)} +\|v^{(k+1)} \|_{L^\infty(\Omega)} \\[6pt]
\mathop{\leq}^{{\rm\Cref{cor-reg}}} C t^{-1 + \frac{d}{2p_k}}
\Big( t^{-1/2} \| u^{(k)}\|_{W^{1,p_{k}}_t(\Omega)} + \| v^{(k)}\|_{L^{p_{k}}(\Omega)} \Big).
\end{multline}
We derive from \eqref{prop-p-1}, \eqref{prop-p-2} and \eqref{prop-p-3},
\begin{multline}
\label{prop-p-6}
\| u^{(k+1)}\|_{L^\infty(\Omega)}+\|v^{(k+1)}\|_{L^\infty(\Omega) } \\[6pt]
\le C t^{-1} t^{-1 + \frac{d}{2p_k}} \prod_{j=2}^k t^{-1+\frac{d}{2} \left ( \frac{1}{p_{j-1}}-\frac{1}{p_j} \right )} \|(f,g)\|_{L^2(\Omega) \times L^2(\Omega)} = C t^{-k-1 + \frac{d}{4}} \|(f,g)\|_{L^2(\Omega) \times L^2(\Omega)}.
\end{multline}
Thus \eqref{prop-p-4} is proved.
We next establish \eqref{prop-p-4*}. We have, by \Cref{lem-T*},
\[
T_{j, \theta, t}^* = M_t^{-1} P \widetilde{T}_{\overline{\lambda}_{j, \theta,t}} P^{-1} M_t,
\]
where $P$ is given by \eqref{pre-7}. This implies
\[
{\bf T}_{\theta, t}^* = M_{t}^{-1} P \widetilde{T}_{\overline{\lambda}_{1, \theta,t}} \circ \cdots \circ \widetilde{T}_{\overline{\lambda}_{k+1, \theta, t}} P^{-1}M_t.
\]
Similarly to \eqref{prop-p-6}, we have
\begin{equation}
\label{prop-p-7}
\|{\bf T}_{\theta, t}^*\|_{L^2 (\Omega) \to L^\infty(\Omega)} \leq C t^{-k-1 + \frac{d}{4}}.
\end{equation}
By a standard dual argument, we derive from \eqref{prop-p-7} that
\[
\|{\bf T}_{\theta, t}\|_{L^1(\Omega) \to L^2 (\Omega) } \leq C t^{-k-1 + \frac{d}{4}}.
\]
The properties for ${\bf T}_{\theta, t}$ are established.
\medskip
The properties for ${\bf T}_{\theta, t}^*$ can be derived similarly.
\end{proof}
\subsection{The approximation of the trace of a kernel}
Denote
\begin{equation}\label{p1}
\alpha = \frac{\pi}{4(k+1)} \quad \quad \mbox{ and } \quad \quad \beta = \frac{5\pi}{4(k+1)}.
\end{equation}
Then
\begin{equation}\label{p1*}
\alpha,\beta \in \Theta \quad \mbox{ and }\quad e^{i\alpha(k+1)}+e^{i\beta(k+1)} = 0.
\end{equation}
Recall that $\Theta$ is defined in \eqref{def-gam}.
\begin{lem}
\label{lem-Tt}
For $t \ge \max\{t_\alpha, t_\beta \}$, where $t_\alpha$ and $t_\beta$ are given in \Cref{def-t-theta}, we have
\begin{enumerate}
\item the operator ${\bf T}_{\alpha,t} {\bf T}_{\beta,t}$ is Hilbert-Schmidt, and
\begin{equation}
\label{lem-Tt-1}
\vvvert {\bf T}_{\alpha,t} {\bf T}_{\beta,t} \vvvert \leq C t^{-2k-2+\frac{d}{2}};
\end{equation}
\item the range of ${\bf T}_{\alpha,t} {\bf T}_{\beta,t}$ is in $[C(\bar{\Omega})]^2$, and
\begin{equation}
\label{lem-Tt-2}
\|{\bf T}_{\alpha,t} {\bf T}_{\beta,t}\|_{L^1(\Omega) \to L^\infty(\Omega)} \leq Ct^{-2k-2+\frac{d}{2}};
\end{equation}
\item the kernel ${\rm \bold{K}}_t$ of ${\bf T}_{\alpha,t} {\bf T}_{\beta,t}$ is continuous in $\Omega \times \Omega$, and
\begin{equation}
\label{lem-Tt-3}
|{\rm \bold{K}}_t(x,y)| \leq C t^{-2k-2+\frac{d}{2}} \quad \mbox{ for all }(x,y) \in \Omega \times \Omega;
\end{equation}
\end{enumerate}
for some positive constant $C$ independent of $t$.
\end{lem}
\begin{proof}
Assertion \eqref{lem-Tt-1} follows from \Cref{cor-p} and
\[
\vvvert {\bf T}_{\alpha,t} {\bf T}_{\beta,t} \vvvert \leq \vvvert {\bf T}_{\alpha , t} \vvvert \vvvert {\bf T}_{\beta, t} \vvvert.
\]
Applying \Cref{prop-p} and using the fact
\[
\|{\bf T}_{\alpha,t} {\bf T}_{\beta,t}\|_{L^1(\Omega) \to L^\infty(\Omega)} \leq \|{\rm \bold{T}}_{\alpha,t} \|_{L^2(\Omega) \to L^\infty(\Omega)}\|{\rm \bold{T}}_{\beta,t} \|_{L^1(\Omega) \to L^2(\Omega)},
\]
we obtain \eqref{lem-Tt-2}.
Since both the range of ${\bf T}_{\alpha,t}$ and ${\bf T}_{\beta,t}^*$ are contained in $[C(\bar{\Omega})]^2$, the continuity of ${\rm \bold{K}}_t$ and \eqref{lem-Tt-3} follow from \Cref{cor-K-cont} and \eqref{lem-Tt-2}.
\end{proof}
For $\ell =1, \, 2$, $\theta \in \Theta$, and $t>1$, consider, with $\lambda = t e^{i\theta}$,
\begin{equation}
\label{def-S}
\begin{array}{rccc}
S_{\ell,\lambda,x_0} :&L^2(\mathbb{R}^d)& \to &L^2(\mathbb{R}^d) \\[6 pt]
& f_\ell & \mapsto & v_\ell
\end{array}
\end{equation}
where $v_\ell \in H^1(\mathbb{R}^d)$ is the unique solution of
\begin{equation}\label{vl-Rd}
\operatorname{div} (A(x_0)\nabla v_\ell) -\lambda \Sigma_\ell (x_0) v_\ell = \Sigma_\ell (x_0) f_\ell \mbox{ in } \mathbb{R}^d.
\end{equation}
One then has
\begin{equation}
S_{\ell,\lambda,x_0} f(x) = \int_{\mathbb{R}^d}F_{\ell,\lambda}(x_0,x-y)f(y) dy,
\end{equation}
where
\begin{equation}
\label{def-F}
F_{\ell,\lambda}(x_0,z)= -\frac{1}{(2\pi)^d}\int_{\mathbb{R}^d} \frac{e^{i z \cdot \xi}}{\Sigma_{\ell}(x_0)^{-1}A(x_0) \xi \cdot \xi+\lambda}d\xi.
\end{equation}
Set, for $\ell =1,2$,
\[
\mathcal{S}_{\ell,t,x_0} = S_{\ell,\lambda_{k+1,\alpha,t},x_0}\circ \cdots \circ S_{\ell,\lambda_{1,\alpha,t},x_0} \circ S_{\ell,\lambda_{k+1,\beta,t},x_0}\circ \cdots \circ S_{\ell,\lambda_{1,\beta,t},x_0}.
\]
Define, for $\ell =1,2$,
\begin{equation}
\label{ker-33}
\mathcal{F}_{\ell, t}(x_0,z) = \frac{1}{(2\pi)^d} \int_{\mathbb{R}^d} \frac{e^{i z \xi} \, d \xi}{\prod_{j=1}^{k+1} \left ( \Sigma_\ell (x_0)^{-1} A(x_0) \xi \cdot \xi + \lambda_{j,\alpha, t} \right ) \left ( \Sigma_\ell (x_0)^{-1} A(x_0) \xi \cdot \xi + \lambda_{j,\beta, t} \right )}.
\end{equation}
Then
\begin{equation}
\label{def-mS}
\mathcal{S}_{\ell,t,x_0} f_\ell (x) = \int_{\mathbb{R}^d}\mathcal{F}_{\ell,t}(x_0,x-y)f_\ell (y) dy.
\end{equation}
Since $2k+2>d$, the integrand appearing in \eqref{ker-33} belongs to $L^1(\mathbb{R}^d) \cap L^2(\mathbb{R}^d)$, and thus
\begin{equation}
z \mapsto \mathcal{F}_{\ell, t}(x_0,z) \mbox{ is continuous and belongs to }L^2(\mathbb{R}^d).
\end{equation}
To introduce the freezing coefficient version of \eqref{ITE-Cauchy-2} in the whole space, we use the following result in which \eqref{lem-rd-S} is the system of $(u, v) : = (v_1 - v_2, \lambda v_1)$, where $v_\ell$ ($\ell = 1, 2$) is defined by \eqref{vl-Rd}.
\begin{lem}
\label{lem-rd}
Let $x_0 \in \Omega$, $c\in (0,1)$, $\lambda \in \mathbb{C}$ with $|\lambda|\ge 1$ and $|\Im (\lambda) |\geq c|\lambda|$. Let $p>1$ and let $(f,g) \in [L^p(\mathbb{R}^d)]^2$. Then there exists a unique solution $(u,v) \in [W^{1,p}(\mathbb{R}^d)]^2$ of
\begin{equation}
\label{lem-rd-S}
\left \{
\begin{array}{rll}
\operatorname{div} (A(x_0)\nabla u) -\lambda \Sigma_1(x_0)u-(\Sigma_1(x_0)-\Sigma_2(x_0))v&= \Sigma_1 (x_0) f & \mbox{ in } \mathbb{R}^d, \\ [6 pt]
\operatorname{div} ( A(x_0)\nabla v) -\lambda \Sigma_{2}(x_0) v &= \Sigma_2 (x_0) g & \mbox{ in }\mathbb{R}^d.
\end{array}
\right .
\end{equation}
Moreover,
\begin{equation}
\label{lem-rd-E}
\|u\|_{W^{2,p}_\lambda(\mathbb{R}^d)}+|\lambda|^{-1}\|v\|_{W^{2,p}_\lambda (\mathbb{R}^d)}
\leq C \left ( \|f\|_{L^p(\mathbb{R}^d)} +|\lambda|^{-1}\|g\|_{L^p(\mathbb{R}^d)} \right ),
\end{equation}
for some $C>0$ depending only on $\Lambda$,$c$ and $p$. As a consequence,
\begin{enumerate}
\item either $1 < p < d$ and $p \le q \le \frac{dp}{d-p}$,
\item either $d=p \le q < + \infty$,
\item or $p> d$ and $q = + \infty$,
\end{enumerate}
we have
\[
\|u\|_{W^{1,q}_\lambda (\mathbb{R}^d)} + |\lambda|^{-1}\|v\|_{W^{1,q}_\lambda (\mathbb{R}^d)} \leq C |\lambda|^{\frac{d}{2} \left ( \frac{1}{p}-\frac{1}{q} \right ) -\frac{1}{2}} \left ( \|f\|_{L^p(\mathbb{R}^d)} + |\lambda|^{-1}\|g\|_{L^p(\mathbb{R}^d)} \right ).
\]
\end{lem}
\begin{proof} We emphasize here that \eqref{lem-rd-S} is a system with constant coefficients imposed in $\mathbb{R}^d$. The proof is quite standard. The idea is first to obtain the existence, uniqueness, and the estimate for $v$ using the second equation of \eqref{lem-rd-S}, and then using these to derive the ones for $u$ using the first equation of \eqref{lem-rd-S}. The details are omitted.
\end{proof}
For $x_0 \in \Omega$, $j=1,\cdots,k+1$, $\theta \in \Theta$, and $t>1$, define, for $1 < p < + \infty$,
\[
\begin{array}{cccc}
R_{\lambda_{j, \theta, t},x_0} :& [L^p(\mathbb{R}^d)]^2 & \to & [L^p(\mathbb{R}^d)]^2 \\[6 pt]
& (f, g) & \mapsto & (u, v)
\end{array}
\]
where $(u, v) \in [W^{1,p}(\mathbb{R}^d)]^2$ is the unique solution \eqref{lem-rd-S} with $\lambda = \lambda_{j, \theta,t}$. Recall that $\lambda_{j, \theta,t}$ is defined in \eqref{p4}. We also introduce
\begin{equation}
\label{ker-8}
R_{j, \theta , t,x_0} = M_t R_{\lambda_{j, \theta,t}, x_0}M_t^{-1} \quad \quad \mbox{ and }\quad \quad \mathbf{R}_{\theta, t, x_0} = R_{k+1, \theta, t, x_0}\circ \cdots \circ R_{1, \theta, t, x_0}.
\end{equation}
As in the proof of \Cref{prop-p}, however, using \Cref{lem-rd} instead of
\Cref{prop-pre} and \Cref{cor-reg}, we obtain
\begin{lem}\label{lem-proS} Let $\theta \in \Theta$ and $t>1$. Then, the range of $\mathbf{R}_{\theta,t,x_0}$ and $\mathbf{R}_{\theta,t,x_0}^*$ are in $[C(\mathbb{R}^d)]^2$ for all $t>1$. Moreover,
\begin{equation}
\label{ker-37}
\| \mathbf{R}_{\theta,t,x_0}\|_{L^2 (\mathbb{R}^d) \to L^\infty (\mathbb{R}^d) }+ \|\mathbf{R}_{\theta,t,x_0}\|_{L^1 (\mathbb{R}^d) \to L^2 (\mathbb{R}^d)} \leq Ct^{-k-1 + \frac{d}{4}}
\end{equation}
and
\begin{equation}
\label{ker-371}
\| \mathbf{R}_{\theta,t,x_0}^*\|_{L^2 (\mathbb{R}^d) \to L^\infty (\mathbb{R}^d) }+ \|\mathbf{R}_{\theta,t,x_0}^*\|_{L^1 (\mathbb{R}^d) \to L^2 (\mathbb{R}^d)} \leq Ct^{-k-1 + \frac{d}{4}},
\end{equation}
for some positive constant $C$ independent of $t$.
\end{lem}
Define
\begin{equation}
\mathbf{R}_{t,x_0} = \mathbf{R}_{\alpha,t,x_0}\mathbf{R}_{\beta,t,x_0}.
\end{equation}
One can then write $\mathbf{R}_{t,x_0} $ under the form
\[
\mathbf{R}_{t,x_0}
=
\begin{pmatrix}
(\mathbf{R}_{t,x_0})_{11} & (\mathbf{R}_{t,x_0})_{12} \\
(\mathbf{R}_{t,x_0})_{21} & (\mathbf{R}_{t,x_0})_{22} \\
\end{pmatrix}.
\]
Note that, by the definition of $S_{\ell, \lambda, x_0}$,
\[
R_{\lambda_{j,\theta,t},x_0}
=
\begin{pmatrix}
S_{1,\lambda_{j,\theta,t},x_0} & \: \Sigma_1(x_0)^{-1}(\Sigma_1(x_0)-\Sigma_2(x_0))S_{1,\lambda_{j,\theta,t},x_0} S_{2,\lambda_{j,\theta,t},x_0} \\
0&\: S_{2,\lambda_{j,\theta,t},x_0}
\end{pmatrix}.
\]
It follows that $R_{\lambda_{j,\theta,t},x_0}$ is an upper triangular matrix operator, and so is $\mathbf{R}_{t,x_0}$. We deduce that
\[
(\mathbf{R}_{t,x_0})_{21} = 0
\]
and, for $\ell=1,2$,
\[
(\mathbf{R}_{t,x_0})_{\ell \ell} = \mathcal{S}_{\ell,t,x_0}.
\]
These simple observations are useful in computing the approximation of the trace of the kernel of $\mathbf{R}_{t, x_0}$.
As an immediate consequence of \eqref{def-mS}, $\mathbf{R}_{t,x_0}$ is an integral operator whose kernel verifies, for $\ell = 1, 2$,
\begin{equation}\label{kernel-Kll}
({\rm \bold{K}}_{t,x_0})_{\ell \ell}(x,y)=\mathcal{F}_{\ell}(x_0,x-y) \mbox{ for } x, y \in \mathbb{R}^d.
\end{equation}
Further properties of ${\rm \bold{K}}_{t,x_0}$ are given in the following lemma.
\begin{lem}
\label{lem-ker0} Let $t \ge 1$ and $x_0 \in \Omega$. Then ${\rm \bold{K}}_{t,x_0}$ is continuous on $\mathbb{R}^d \times \mathbb{R}^d$, and, for $(x,y) \in \mathbb{R}^d\times \mathbb{R}^d$, it holds, for $\ell =1, 2$,
\begin{equation}\label{lem-ker0-st1}
|({\rm \bold{K}}_{t,x_0})_{\ell \ell}(x,y)| \leq Ct^{-2k-2+\frac{d}{2}}.
\end{equation}
Moreover,
\begin{multline}\label{lem-ker0-st2}
\mbox{trace}({\rm \bold{K}}_{t,x_0} (x_0,x_0)) \\[6pt]
= \frac{t^{-2k-2+\frac{d}{2}} }{(2\pi)^d} \sum_{\ell = 1}^2 \int_{\mathbb{R}^d} \frac{d\xi}{(\Sigma_\ell (x_0)^{-1}A(x_0)\xi \cdot \xi)^{2k+2}-i} + o(t^{-2k-2+\frac{d}{2}}) \mbox{ as } t \to + \infty.
\end{multline}
\end{lem}
\begin{remark} \rm Assertion \eqref{lem-ker0-st2} holds uniformly with respect to $x_0 \in \Omega$.
\end{remark}
\begin{proof} From \eqref{kernel-Kll}, it follows that $({\rm \bold{K}}_{t,x_0})_{\ell \ell}(x,y)$ is continuous on $\mathbb{R}^d\times \mathbb{R}^d$. By the choice of $\alpha$, $\beta$, and $\omega_j$ in \eqref{p1}, \eqref{p1*}, and \eqref{omega-j}, one has
\begin{multline*}
\prod_{j=1}^{k+1} \left ( \Sigma_\ell (x_0)^{-1} A(x_0) \xi \cdot \xi + \lambda^* + \omega_j t e^{i\alpha} \right ) \left ( \Sigma_\ell (x_0)^{-1} A(x_0) \xi \cdot \xi + \lambda^* + \omega_j t e^{i\beta} \right ) \\[6pt]
= (\Sigma_\ell(x_0)^{-1} A(x_0) \xi \cdot \xi + \lambda^*)^{2(k+1)} - i t^{2(k+1)}.
\end{multline*}
It follows from \eqref{ker-33} that, for every $x_0 \in \Omega$ and every $z \in \mathbb{R}^d$,
\[
\mathcal{F}_{\ell,t}(x_0,z) = \frac{1}{(2\pi)^d} \int_{\mathbb{R}^d} \frac{e^{i z\cdot \xi} \, d \xi}{ (\Sigma_\ell(x_0)^{-1} A(x_0) \xi \cdot \xi + \lambda^*)^{2(k+1)} - i t^{2(k+1)}}.
\]
A change of variables yields
\begin{equation}\label{ker-41}
\mathcal{F}_{\ell,t}(x_0,z) = \frac{t^{-2k-2+\frac{d}{2} }}{(2\pi)^d} \int_{\mathbb{R}^d} \frac{e^{i t^{1/2}z\cdot \xi} \, d \xi}{ (\Sigma_\ell(x_0)^{-1} A(x_0) \xi \cdot \xi + t^{-1}\lambda^*)^{2(k+1)} - i}.
\end{equation}
Assertion \eqref{lem-ker0-st1} follows from \eqref{ker-41} since $|e^{i t^{1/2}z\cdot \xi}|=1$ and $\lambda^* t^{-1}$ is uniformly bounded with respect to $t\geq 1$.
By taking $z = 0$ in \eqref{ker-41}, we obtain \eqref{lem-ker0-st2} after using the dominated convergence theorem.
\medskip
The proof is complete.
\end{proof}
We now prove the main result of this section concerning the trace of ${\bf T}_{\alpha, t} {\bf T}_{\beta, t}$ where
$\alpha, \beta$ are given in \eqref{p1} and ${\bf T}_{\theta, t}$ is defined in \eqref{p11}.
\begin{prop}
\label{prop-ker}
We have
\[
{\rm trace} ({\bf T}_{\alpha, t} {\bf T}_{\beta, t} ) = {\bf c} t^{-2k-2+ \frac{d}{2}} + o (t^{-2k-2+ \frac{d}{2}}) \quad \mbox{ as } \quad t \to + \infty,
\]
where
\begin{equation}
\label{prop-ker-2}
{\bf c} = \frac{1}{(2\pi)^d} \sum_{\ell=1}^2 \int_\Omega \int_{\mathbb{R}^d} \frac{d\xi \, dx}{\left ( \Sigma_\ell^{-1}(x) A(x) \xi \cdot \xi \right )^{2k+2}-i} .
\end{equation}
\end{prop}
The proof of \Cref{prop-ker} uses the following result.
\begin{lem}
\label{lem-approx} Let $\delta_0 \in (0,1)$ and $\theta \in \Theta$. For every $\varepsilon > 0$, there exists $\delta_\varepsilon \in (0,\delta_0/2)$ depending on $\varepsilon$ such that the following holds: There exists $t_{\varepsilon}>0$ depending on $\varepsilon$ and $\delta_\varepsilon$ such that for every $t > t_{\varepsilon}$ and every $x_0 \in \Omega \setminus \overline{\Omega_{\delta_0}}$, we have
\begin{equation}
\label{lem-approx-p1}
\|{\bf T}_{\theta, t} -\mathbf{R}_{\theta,t, x_0} \mathds{1}_\Omega \|_{L^2(\Omega) \to L^\infty(B(x_0,\delta_\varepsilon))} \leq \epsilon t^{-k-1 + \frac{d}{4}}
\end{equation}
and
\begin{equation}\label{lem-approx-p2}
\|{\bf T}_{\theta, t} \mathds{1}_\Omega -\mathbf{R}_{\theta, t, x_0} \|_{L^2(\mathbb{R}^d) \to L^\infty(B(x_0,\delta_\varepsilon))} \leq \varepsilon t^{-k-1 + \frac{d}{4}},
\end{equation}
and similar facts for ${\bf T}_{\theta, t}^*$ and $\mathbf{R}_{\theta, t, x_0}^*$.
\end{lem}
Recall that ${\bf R}_{\theta, t, x_0}$ is defined in \eqref{ker-8}. We admit \Cref{lem-approx} and give the proof of
\Cref{prop-ker}. The proof of \Cref{lem-approx} is presented right after the one of \Cref{prop-ker}.
\begin{proof}[Proof of \Cref{prop-ker}]
For $\epsilon>0$, let $\delta_0>0$ be such that
\begin{equation}
\label{ker-1}
|\Omega_{2\delta_0}|<\epsilon,
\end{equation}
where $\Omega_\tau$ is given in \eqref{def-Omegatau}.
We claim that there exists $\tau_*>0$, depending on $\Omega$, and $\varepsilon$ but independent of $x_0$, and a positive constant $C$, independent of $\varepsilon$ and $x_0$, such that, for $t> \tau_*$,
\begin{equation}
\label{ker-42}
|\mbox{trace} ({\rm \bold{K}}_t (x_0, x_0))- \mbox{trace}({\bf K}_{t,x_0}(x_0,x_0))| \leq C \epsilon t^{-2k-2+\frac{d}{2}} \quad \mbox{ for } x_0 \in \Omega \setminus \Omega_{\delta_0}.
\end{equation}
Indeed, let $\chi \in C_c^\infty(\mathbb{R}^d)$ be such that $\chi = 1$ in $B_1$ and $\operatorname{supp} \chi \subset B_2$. Denote, for $\delta \in (0,\delta_0/10)$,
$$
\chi_{\delta, x_0} = \chi \big( ( \cdot - x_0) / \delta \big),
$$
and define
\begin{equation}
\label{ker-441}
\left\{\begin{array}{c}
{\rm \bold{P}}_{1,t,\delta} = \chi_{\delta, x_0} (\mathbf{R}_{\alpha,t,x_0} \mathds{1}_\Omega -{\bf T}_{\alpha,t}) {\bf T}_{\beta,t} \chi_{\delta, x_0}, \\[6pt]
{\rm \bold{P}}_{2,t,\delta} = \chi_{\delta, x_0} \mathbf{R}_{\alpha,t,x_0} (\mathbf{R}_{\beta,t,x_0} - \mathds{1}_\Omega {\bf T}_{\beta,t} )\chi_{\delta, x_0}.
\end{array} \right.
\end{equation}
Then
\begin{equation}
\label{ker-43}
\chi_{\delta, x_0} (\mathbf{R}_{\alpha,t,x_0} \mathbf{R}_{\beta,t,x_0} -{\bf T}_{\alpha,t} {\bf T}_{\beta,t} ) \chi_{\delta, x_0} = {\rm \bold{P}}_{1,t,\delta} + {\rm \bold{P}}_{2,t,\delta}.
\end{equation}
By applying \Cref{lem-approx} below with $\theta \in \{\alpha,\beta\}$, there exist $\delta_\varepsilon>0$ and $t_{\varepsilon}>0$ depending on $\varepsilon$ such that for every $t>t_{\varepsilon}$,
\begin{equation}
\label{ker-421}
\| \chi_{\delta_\varepsilon, x_0} ({\bf T}_{\alpha, t}-\mathbf{R}_{\alpha, t, x_0} \mathds{1}_\Omega ) \|_{L^2(\Omega) \to L^\infty(\Omega)} \leq \epsilon t^{-k-1 + \frac{d}{4}}
\end{equation}
and
\begin{equation}
\label{ker-422}
\| \chi_{\delta_\varepsilon, x_0} ( {\bf T}_{\beta, t}^* \mathds{1}_\Omega-\mathbf{R}_{\beta, t, x_0}^* ) \|_{L^2(\mathbb{R}^d) \to L^\infty(\Omega)} \leq \epsilon t^{-k-1 + \frac{d}{4}}.
\end{equation}
Since
\[
\Big( ( \mathds{1}_\Omega {\bf T}_{\beta,t} -\mathbf{R}_{\beta, t, x_0} ) \chi_{\delta_\varepsilon, x_0} \Big)^* = \chi_{\delta_\varepsilon, x_0} ( {\bf T}_{\beta, t}^* \mathds{1}_\Omega-\mathbf{R}_{\beta, t, x_0}^*),
\]
we derive from \eqref{ker-422}, using a dual argument, that
\begin{equation}
\label{ker-48}
\| (\mathds{1}_\Omega {\bf T}_{\beta,t}-\mathbf{R}_{\beta,t,x_0} ) \chi_{\delta_\varepsilon, x_0} \|_{L^1(\Omega) \to L^2(\mathbb{R}^d)} \le \epsilon t^{-k-1+\frac{d}{4}}.
\end{equation}
By \Cref{prop-p} and \Cref{lem-proS}, we have
\begin{equation}
\label{ker-481}
\| {\bf T}_{\beta,t} \chi_{\delta_\varepsilon, x_0} \|_{L^1(\Omega) \to L^2 (\Omega)} + \| \chi_{\delta_\varepsilon, x_0} \mathbf{R}_{\alpha,t,x_0} \|_{L^2 (\mathbb{R}^d) \to L^\infty (\Omega)} \le C t^{-k-1 + \frac{d}{4}}
\end{equation}
for some constant $C>0$ independent of $\epsilon$ and $t$.
Using the fact, for appropriate linear operators $L_1$ and $L_2$,
$$
\| L_1 L_2 \|_{L^1(\Omega) \to L^\infty(\Omega)} \le \| L_1 \|_{L^2 (\Omega) \to L^\infty (\Omega)} \| L_2 \|_{L^1(\Omega) \to L^2 (\Omega)},
$$
and
$$
\| L_1 L_2 \|_{L^1(\Omega) \to L^\infty(\Omega)} \le \| L_1 \|_{L^2 (\mathbb{R}^d) \to L^\infty (\Omega)} \| L_2 \|_{L^1(\Omega) \to L^2 (\mathbb{R}^d)},
$$
we derive from \eqref{ker-421}, \eqref{ker-48}, and \eqref{ker-481} that
\begin{equation*}
\| {\rm \bold{P}}_{1,t,\delta_\varepsilon} + {\rm \bold{P}}_{2,t,\delta_\varepsilon} \|_{L^1(\Omega)\to L^\infty(\Omega)} \le
\| {\rm \bold{P}}_{1,t,\delta_\varepsilon}\|_{L^1(\Omega)\to L^\infty(\Omega)} + \|{\rm \bold{P}}_{2,t,\delta_\varepsilon} \|_{L^1(\Omega)\to L^\infty(\Omega)} \leq C \epsilon t^{-2k-2+\frac{d}{2}}.
\end{equation*}
This yields, by \eqref{ker-43},
\begin{equation}
\label{ker-49}
\| \chi_{\delta_\varepsilon, x_0} (\mathbf{R}_{\alpha,t,x_0} \mathbf{R}_{\beta,t,x_0} -{\bf T}_{\alpha,t} {\bf T}_{\beta,t} ) \chi_{\delta_\varepsilon, x_0} \|_{L^1(\Omega)\to L^\infty(\Omega)} \leq C \epsilon t^{-2k-2+\frac{d}{2}}.
\end{equation}
Since, for $x \in \Omega$, $\ell =1,2$ and $f\in L^2(\Omega)$,
\begin{multline*}
\chi_{\delta_\varepsilon, x_0} \Big((\mathbf{R}_{\alpha,t,x_0} \mathbf{R}_{\beta,t,x_0})_{\ell \ell} -({\bf T}_{\alpha,t} {\bf T}_{\beta,t})_{\ell \ell} \Big) \chi_{\delta_\varepsilon, x_0} f (x) \\
= \chi_{\delta_\varepsilon, x_0} (x) \int_\Omega \chi_{\delta_\varepsilon, x_0} (y) \Big( ({\rm \bold{K}}_{t,x_0})_{\ell \ell}(x,y)- ({\rm \bold{K}}_t)_{\ell \ell} (x,y) \Big)f(y) dy,
\end{multline*}
it follows that $\chi_{\delta_\varepsilon, x_0} (x) \chi_{\delta_\varepsilon, x_0} (y) ( ({\rm \bold{K}}_{t,x_0})_{\ell \ell}(x,y)- ({\rm \bold{K}}_t(x,y))_{\ell \ell})$ is the kernel of the operator
$$
\chi_{\delta_\varepsilon, x_0} \Big((\mathbf{R}_{\alpha,t,x_0} \mathbf{R}_{\beta,t,x_0})_{\ell \ell} -({\bf T}_{\alpha,t} {\bf T}_{\beta,t})_{\ell \ell} \Big) \chi_{\delta_\varepsilon, x_0}.
$$
By \Cref{lem-Tt} and \Cref{lem-ker0}, this kernel is continuous on $\bar \Omega \times \bar \Omega$.
Using \eqref{ker-49} and applying \Cref{lem-HS1}, we derive that, since $\chi_{\delta_\varepsilon, x_0} (x_0) = 1$,
\begin{equation}
\label{ker-50}
|\mbox{trace}({\rm \bold{K}}_{t}(x_0,x_0))- \mbox{trace}({\rm \bold{K}}_{t, x_0}(x_0,x_0))| \leq C \epsilon t^{-2k-2+ \frac{d}{2}} \quad \mbox{ for all }t>t_{\varepsilon}.
\end{equation}
Since the LHS of \eqref{ker-50} does not depend on $\varepsilon>0$, the claim \eqref{ker-42} is proved.
By \Cref{lem-Tt} we have, for $t>0$ large enough,
\begin{equation}
\label{ker-501}
\int_{\Omega_{2\delta_0}} |\mbox{trace}({\rm \bold{K}}_t(x,x))| dx \leq C |\Omega_{2\delta_0} |t^{-2k-2+ \frac{d}{2}} \mathop{\leq}^{\eqref{ker-1}} C\epsilon t^{-2k-2+ \frac{d}{2}}
\end{equation}
and, similarly by \Cref{lem-ker0},
\begin{equation}\label{ker-502}
\int_{\Omega_{2\delta_0}} |\mbox{trace}({\rm \bold{K}}_{t, x}(x, x))| dx \leq C\epsilon t^{-2k-2+ \frac{d}{2}}.
\end{equation}
Combining \eqref{ker-42},\eqref{ker-501}, and \eqref{ker-502} yields
\begin{equation} \label{ker-521}
\int_\Omega |\mbox{trace}({\rm \bold{K}}_t(x,x))- \mbox{trace}({\rm \bold{K}}_{t, x}(x, x))| dx \leq C\varepsilon t^{-2k-2+\frac{d}{2}} \quad \mbox{ for all }t>t_\varepsilon.
\end{equation}
The conclusion follows from \Cref{lem-ker0} and \eqref{ker-521}.
\end{proof}
We now give
\begin{proof}[Proof of \Cref{lem-approx}] Let $\varepsilon >0 $ and $\theta \in \Theta$.
First, we prove \eqref{lem-approx-p1}. Fix $\chi \in C_c^\infty (\mathbb{R}^d)$ such that $\operatorname{supp} \chi \subset B_2$ and $\chi = 1$ in $B_1$. Set, for $0< \delta < \delta_0/ 100$,
\[
\chi_\delta = \chi \big( (\cdot -x_0)/\delta \big) \mbox{ in } \mathbb{R}^d.
\]
Define, for $(f,g) \in [L^2(\Omega)]^2$, and for $j=1,\cdots, k+1 = [d/2] + 2$,
\[
(u^j,v^j) = T_{\lambda_{j, \theta,t}}\circ \cdots \circ T_{\lambda_{1, \theta, t}} (u^0, v^0) \quad \mbox{ and } \quad (u^j_0,v^j_0) = S_{\lambda_{j, \theta, t},x_0}\circ \cdots \circ S_{\lambda_{1, \theta, t},x_0} (u^0_0,v^0_0),
\]
where
\begin{equation}
\label{case1}
(u^0,v^0)=(f, g) \quad \quad \mbox{ and } \quad \quad (u^0_0,v^0_0)= (\mathds{1}_\Omega f, \mathds{1}_\Omega g) \quad \mbox{ in } \Omega.
\end{equation}
Set, for $0 \le j \le k+1$,
$$
(u^{j, \delta}, v^{j, \delta}) = (\chi_\delta u^j, \chi_\delta v^j)
\quad \quad \mbox{ and } \quad \quad
(u^{j, \delta}_0, v^{j, \delta}_0) = (\chi_\delta u^j_0, \chi_\delta v^j_0).
$$
We have
\begin{equation}
\label{iter}
(u^{j, \delta}, v^{j, \delta}) = S_{\lambda_{j, \theta, t}, x_0} (u^{j-1, \delta}, v^{j-1, \delta}) + S_{\lambda_{j, \theta, t}, x_0} (f^{j, \delta}, g^{j, \delta}),
\end{equation}
where
\begin{multline}
\label{cc3}
\Sigma_1 (x_0) f^{j, \delta} = (\Sigma_1 (x) -\Sigma_1(x_0)) u^{j-1, \delta} -\lambda_{j, \theta,t} (\Sigma_1(x_0)-\Sigma_1(x)) u^{j, \delta}+
A(x)\nabla \chi_\delta \cdot \nabla u^j \\[6pt] - (\Sigma_1(x_0)-\Sigma_1(x)- \Sigma_2(x_0)+\Sigma_2(x)) v^{j, \delta} + \operatorname{div} \Big( (A(x_0)-A(x)) \nabla u^{j, \delta} + u^j A(x)\nabla \chi_\delta\Big)
\end{multline}
and
\begin{multline}
\label{cc4}
\Sigma_2(x_0) g^{j, \delta} = (\Sigma_2 (x) -\Sigma_2(x_0)) v^{j-1, \delta} - \lambda_{j, \theta,t} (\Sigma_2(x_0)-\Sigma_2(x)) v^{j, \delta} + A(x) \nabla \chi_\delta \cdot \nabla v^j \\[6pt]
+ \operatorname{div} \Big( (A(x_0)-A(x)) \nabla v^{j, \delta} + v^j A(x)\nabla \chi_\delta \Big).
\end{multline}
Similarly, we have
\[
(u^{j, \delta}_0, v^{j, \delta}_0) = S_{\lambda_{j, \theta, t}, x_0} (u^{j-1, \delta}_0, v^{j-1, \delta}_0) + S_{\lambda_{j, \theta, t},x_0} (f^{j, \delta}_0, g^{j, \delta}_0),
\]
where
\[
\Sigma_1(x_0) f^{j, \delta}_0 =
A(x_0) \nabla \chi_\delta \cdot \nabla u^j_0 + \operatorname{div} \Big(u^j_0 A (x_0)\nabla \chi_\delta\Big)
\]
and
\[
\Sigma_2(x_0) g^{j, \delta}_0 = A(x_0) \nabla \chi_\delta \cdot \nabla v^j_0 \\[6pt]
+ \operatorname{div} \Big( v^j_0 A(x_0) \nabla \chi_\delta \Big).
\]
For $r>0$, define
\[
\Phi (r) = \min \left \{1, \sup_{|x-y|<r} \left( |A(x)-A(y)|+ \sum_{\ell=1}^2 |\Sigma_\ell(x)-\Sigma_\ell(y)| \right) \right \}.
\]
We claim that
\begin{multline}
\label{cc1}
\|f^{j, \delta}\|_{L^{p_{j-1}}(\Omega \setminus \Omega_{\delta_0/2})} + t^{-1}\|g^{j, \delta}\|_{L^{p_{j-1}}(\Omega \setminus \Omega_{\delta_0/2})} \\[6pt] \le C_{\delta_0} \left(\Phi(\delta) + \frac{1}{\delta t^{1/2}} + \frac{1}{\delta^2 t} \right) \left ( \|u^{j-1}\|_{L^{p_{j-1}}(\Omega)} + t^{-1}\|v^{j-1}\|_{L^{p_{j-1}}(\Omega)} \right ),
\end{multline}
and
\begin{multline}
\label{cc2}
\|f^{j, \delta}_0\|_{L^{p_{j-1}}(\Omega \setminus \Omega_{\delta_0/2})} + t^{-1}\|g^{j, \delta}_0\|_{L^{p_{j-1}}(\Omega \setminus \Omega_{\delta_0/2})} \\[6pt] \le C_{\delta_0} \left( \frac{1}{\delta t^{1/2}} + \frac{1}{\delta^2 t} \right) \left ( \|u^{j-1}_0\|_{L^{p_{j-1}}(\Omega)} + t^{-1}\|v^{j-1}_0\|_{L^{p_{j-1}}(\Omega)} \right ).
\end{multline}
We first admit \eqref{cc1} and \eqref{cc2} and continue the proof.
Since, in $\Omega$,
$$
(u^{0, \delta}, v^{0, \delta}) = (u^{0, \delta}_0, v^{0, \delta}_0),
$$
using \eqref{iter}, \eqref{cc1} and \eqref{cc2} and \Cref{lem-rd}, for $j=1$ and then for $j=2, \dots, k+1$, we have
\begin{multline}\label{lem-approx-coucou}
\|u^{j, \delta} - u^{j, \delta}_0\|_{L^{p_{j}}(\Omega \setminus \Omega_{\delta_0/2})} + t^{-1}\|v^{j, \delta} - v^{j, \delta}_0\|_{L^{p_{j}}(\Omega \setminus \Omega_{\delta_0/2})} \\[6pt]
\le C_{\delta_0} \left(\Phi (\delta) + \frac{1}{\delta t^{1/2}} + \frac{1}{\delta^2 t} \right) t^{- \frac{d}{2 p_j}- j + \frac{d}{4} } \left( \| f\|_{L^2(\Omega)} + t^{-1 } \| g\|_{L^2(\Omega)} \right).
\end{multline}
Fix $\delta = \delta_\varepsilon>0$ such that $C_{\delta_0} \Phi(\delta_\varepsilon)<\varepsilon/2$. Take $t_\varepsilon>0$ sufficiently large such that $C_{\delta_0}(\delta_\varepsilon^{-1}t^{-1/2} + \delta^{-2}_\varepsilon t^{-1})<\varepsilon/2$ for every $t>t_\varepsilon$. Taking $j = k+1$ in \eqref{lem-approx-coucou} gives \eqref{lem-approx-p1}.
The proof of \eqref{lem-approx-p2} is similar to the one of \eqref{lem-approx-p1} by considering
$(u^0, v^0)$ and $(u^0_0, v^0_0)$ defined as follows
\[
(u^0,v^0)= \mathds{1}_\Omega (f, g) \quad \quad \mbox{ and } \quad \quad (u^0_0,v^0_0)= (f, g) \quad \mbox{ in } \mathbb{R}^d,
\]
instead of \eqref{case1}.
\medskip
Similar facts for ${\bf T}_{\theta, t}^*$ and $\mathbf{R}_{\theta, t, x_0}^*$ by analogous arguments.
\medskip
It remains to establish \eqref{cc1} and \eqref{cc2}.
From the definition of $(u^j, v^j)$ and the theory of elliptic equations (see e.g. \cite[Theorem 9.11]{GilbargTrudinger}), we have, for $\Omega_1 \Subset \Omega_2 \subset \Omega$,
\begin{equation}
\label{fact1}
\|u^{j}\|_{W^{2,p}_t(\Omega_1)} +t^{-1}\|v^{j}\|_{W^{2,p}_t (\Omega_1)} \leq C \left ( \|u^{j-1}\|_{L^p(\Omega_2)} + t^{-1}\|v^{j-1}\|_{L^p(\Omega_2)} \right )
\end{equation}
and, similarly,
\begin{equation}
\label{fact2}
\|u^{j}_0\|_{W^{2,p}_t(\Omega_1)} + t^{-1}\|v^{j}_0\|_{W^{2,p}_t (\Omega_1)} \leq C \left ( \|u^{j-1}_0\|_{L^p(\Omega_2)} + t^{-1}\|v^{j-1}_0\|_{L^p(\Omega_2)} \right ),
\end{equation}
for some positive constant $C$ independent of $f$, $g$, and $t$. It follows that
\begin{equation}
\label{fact3}
\|\nabla u^{j, \delta}\|_{L^p(\Omega_1)} +t^{-1}\|\nabla v^{j, \delta}\|_{L^p (\Omega_1)} \leq C \left( \frac{1}{\delta t} + \frac{1}{t^{1/2}} \right) \left ( \|u^{j-1}\|_{L^p(\Omega_2)} + t^{-1}\|v^{j-1}\|_{L^p(\Omega_2)} \right ),
\end{equation}
\begin{multline}
\label{fact4}
\|\nabla^2 u^{j, \delta}\|_{L^p(\Omega_1)} +t^{-1}\|\nabla^2 v^{j, \delta}\|_{L^p (\Omega_1)} \\[6pt]
\leq C \left( 1 + \frac{1}{\delta t^{1/2}} + \frac{1}{\delta^2 t} \right) \left ( \|u^{j-1}\|_{L^p(\Omega_2)} + t^{-1}\|v^{j-1}\|_{L^p(\Omega_2)} \right ),
\end{multline}
and
\begin{equation}
\label{fact5}
\|u^{j, \delta}_0\|_{W^{1, p}_t(\Omega_1)} +t^{-1}\|v^{j, \delta}_0\|_{W^{1,p}_t (\Omega_1)} \\[6pt]
\leq C \left( \frac{1}{\delta t} + \frac{1}{t^{1/2}} \right) \left ( \|u^{j-1}_0\|_{L^p(\Omega_2)} + t^{-1}\|v^{j-1}_0\|_{L^p(\Omega_2)} \right ).
\end{equation}
By \eqref{cc3} and \eqref{cc4}, we have
\begin{align*}
C \left ( \right. & \left. \|f^{j,\delta}\|_{L^{p_{j-1}}(\Omega_1)}+ t^{-1}\|g^{j,\delta}\|_{L^{p_{j-1}}(\Omega_1)} \right ) \le \Phi (\delta) \Big( \|u^{j-1}\|_{L^{p_{j-1}}}+t^{-1}\|v^{j-1}\|_{L^{p_{j-1}}(\Omega_1)} \Big ) \\
& + \Big ( \Phi(\delta) + \frac{1}{\delta^2 t}+\frac{1}{\delta t^{1/2}}\Big ) \Big ( \|u^j\|_{W^{2,p_{j-1}}_t(\Omega_1)}+t^{-1} \|v^j\|_{W^{2,p_{j-1}}_t(\Omega_1)} \Big ) \\
& +\|\nabla u^{j,\delta}\|_{L^{p_{j-1}}(\Omega_1)}+ t^{-1}\|\nabla v^{j,\delta}\|_{L^{p_{j-1}}(\Omega_1)} +\Phi(\delta) \Big ( \| \nabla^2 u^{j,\delta}\|_{L^{p_{j-1}}(\Omega_1)} + t^{-1}\|\nabla^2 v^{j,\delta}\|_{L^{p_{j-1}}(\Omega_1)} \Big ).
\end{align*}
Combining \eqref{fact1}-\eqref{fact5} yields \eqref{cc1}. Estimate \eqref{cc2} follows similarly.
\medskip
The proof is complete.
\end{proof}
\subsection{A connection of the counting function and the trace of ${\bf T}_{\alpha, t} {\bf T}_{\beta, t}$ for large $t$}
We start this section by recalling the definition of the modified resolvent of an operator (see, e.g., \cite[Definition 12.3]{Agmon}).
\begin{definition} \label{def-modRes}
Let $H$ be a Hilbert space and $\mathcal{T} : H \to H$ be a linear and bounded operator. The modified resolvent set $\rho_m(\mathcal{T})$ of $\mathcal{T}$ is the set of all non-zero complex numbers $s$ such that $I-s \mathcal{T}$ is bijective and $(I-s \mathcal{T})^{-1}$ is bounded on $H$. For $s \in \rho_m(\mathcal{T})$ the transformation $(\mathcal{T})_s = \mathcal{T}(I-s\mathcal{T})^{-1}$ is the modified resolvent of $\mathcal{T}$.
\end{definition}
Recall that, for $s \in \rho_m(\mathcal{T})$, we have
\begin{equation}\label{pro-mod}
(\mathcal{T})_s = \mathcal{T}(I-s\mathcal{T})^{-1} = (I-s\mathcal{T})^{-1}\mathcal{T}.
\end{equation}
Let $\mathcal{T}: L^2(\Omega) \times L^2(\Omega) \to L^2(\Omega) \times L^2(\Omega)$ be a linear and bounded operator. We have for $z \in \mathbb{C}$ (see e.g. \cite{Robbiano13})
\begin{equation}
\label{HS-3}
I-z^{k+1}\mathcal{T}^{k+1} = \prod_{j=1}^{k+1} (I-\omega_jz\mathcal{T})
\end{equation}
and
\begin{equation}
\label{HS-4}
I-z^{k+1}\mathcal{T}^{k+1} \mbox{ is invertible } \Longleftrightarrow I-\omega_jz\mathcal{T} \mbox{ is invertible for every }j.
\end{equation}
Recall that $\omega_j^{k+1}=1$. Using the decomposition \eqref{HS-3}, and the equivalence in \eqref{HS-4}, one can prove the following lemma.
\begin{lem}
\label{lem-mod-res}
Let $\widetilde{\theta} \in \mathbb{R}\setminus \{\pi \mathbb{Z}\}$. Set $\theta := \frac{\tilde{\theta}}{k+1} \in \Theta$.
There exists $t_\theta > 1$ such that, for every $t>t_\theta$, it holds
\begin{equation}
\label{lem-mod-res-2}
\gamma:=t^{k+1} e^{i\widetilde{\theta}} \in \rho_m(T_{\lambda^*}^{k+1})
\end{equation}
and
\begin{equation}
\label{lem-mod-res-3}
(T_{\lambda^*}^{k+1})_\gamma = M_t^{-1} \mathbf{T}_{\theta,t} M_t.
\end{equation}
\end{lem}
\begin{proof} We have, by the definition of $\gamma$,
\begin{equation}
\label{lem-mod-res-21}
I- \gamma T_{\lambda^*}^{k+1} \mathop{=}^{\eqref{HS-3}} \prod_{j=1}^{k+1} (I-\omega_j t e^{i\theta} T_{\lambda^*}).
\end{equation}
By \Cref{prop-pre}, there exists $t_\theta > 0$ such that $T_{\lambda_* + \omega t e^{i \theta}}$ is defined for $t \ge t_\theta$. Hence
\begin{equation}
\label{lem-mod-res-211}
\omega_j t e^{i\theta} \in \rho_m(T_{\lambda^*}) \quad \mbox{ and } \quad (T_{\lambda^*})_{\omega_j t e^{i\theta}} = T_{\lambda^* + \omega_j t e^{i\theta}} = T_{\lambda_{j,\theta,t}} \mbox{ for }t\geq t_\theta.
\end{equation}
(see, e.g. \cite[Lemma 3.1]{Fornerod-Ng1} for the arguments in a similar setting).
Combining \eqref{lem-mod-res-21} and \eqref{lem-mod-res-211} leads $\gamma \in \rho_m (T_{\lambda^*}^{k+1})$ for $t \ge t_\theta$.
It follows from \eqref{pro-mod} that, for $t \ge t_\theta$,
\begin{equation}
\label{HS-6}
T_{\lambda_{j,\theta,t}} = T_{\lambda^*} (I-\omega_j t e^{i\theta} T_{\lambda^*})^{-1}=(I-\omega_j t e^{i\theta} T_{\lambda^*})^{-1}T_{\lambda^*}
\end{equation}
and thus,
\begin{multline}
M_t^{-1} \mathbf{T}_{\theta,t}M_t \mathop{=}^{\eqref{p11}} \prod_{j=1}^{k+1} T_{\lambda_{j,\theta,t}} \mathop{=}^{\eqref{HS-6}} T_{\lambda^*}^{k+1}\prod_{j=1}^{k+1} (I-\omega_j e^{i\theta}t T_{\lambda^*})^{-1} \\
= T_{\lambda^*}^{k+1} (I-\gamma T^{k+1}_{\lambda^*})^{-1} \mathop{=}^{\rm \mbox{def.}} (T_{\lambda^*}^{k+1})_{\gamma}.
\end{multline}
The proof is complete.
\end{proof}
The following proposition establishes a connection between the trace of the operator ${\bf T}_{\alpha,t}{\bf T}_{\beta,t}$ and the counting function for large $t$.
The arguments of the proof are in the spirit of \cite{MinhHung2} (see also \cite{Robbiano13}).
\begin{prop}
\label{prop-spec} We have
\[
\mathcal{N}(t) = \frac{\Im ({\bf c})}{\frac{d}{8(k+1)} \int_0^\infty s^{\frac{d}{8(k+1)}-1}(1+s)^{-1}ds} t^{\frac{d}{2}} + o(t^{\frac{d}{2}})\quad \quad \mbox{ as }t \to + \infty,
\]
where $\mathbf{c}$ is given by \eqref{prop-ker-2}.
\end{prop}
\begin{proof}
For $t$ sufficiently large, by \Cref{lem-mod-res}, we have
$$
(T_{\lambda^*}^{k+1})_{t^{k+1} e^{ i (k+1) \alpha}} = M_t^{-1} \mathbf{T}_{\alpha, t} M_t.
$$
Note that
\begin{equation}\label{HS-Rm}
\Big((T_{\lambda^*}^{k+1})_{\gamma_1} \Big)_{\gamma_2} = (T_{\lambda^*}^{k+1})_{\gamma_1 + \gamma_2}
\end{equation}
provided that $\gamma_1$,$\gamma_1 + \gamma_2 \in \rho_m(T_{\lambda^*}^{k+1})$.
It follows from \Cref{lem-mod-res} that, for large $t$ and for $s \ge 0$,
$$
-2 (t+s)^{k+1} e^{i (k+1) \alpha} \in \rho_m \Big( M_t^{-1} \mathbf{T}_{\alpha, t} M_t \Big).
$$
Let $s_{1}, s_{2}, \dots$ be the characteristic values of $M_t^{-1} \mathbf{T}_{\theta,t} M_t$ repeated a number of times equal to their multiplicities. Applying \cite[Theorem 12.17]{Agmon}, we have
\begin{multline}
\label{HS-1500}
\mbox{trace}\Big(M_t^{-1}{\bf T}_{\alpha,t} M_t (M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2 (t+s)^{k+1} e^{i (k+1) \alpha}} \Big) = \sum_{j} \frac{1}{s_{j}(s_{j}+2e^{i\alpha (k+1)} (t+s)^{k+1})}+c_t.
\end{multline}
We claim that
\begin{equation}\label{prop-spec-ct}
c_t = 0.
\end{equation}
Assume this, we continue the proof. As a consequence of \eqref{HS-1500} with $s=0$, we have
\begin{equation}
\label{HS-1500*}
\mbox{trace}\Big(M_t^{-1}{\bf T}_{\alpha,t} M_t (M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2t^{k+1} e^{i (k+1) \alpha}} \Big) = \sum_{j} \frac{1}{s_{j}(s_{j}+2e^{i\alpha (k+1)} t^{k+1})}.
\end{equation}
Let $(\mu_j)_j$ be the set of characteristic values of $T_{\lambda^*}$ repeated according to their multiplicity. It is well-known that $\mu_j^{k+1}$ are the characteristic values of $T_{\lambda^*}^{k+1}$ and the multiplicity of $\mu_j^{k+1}$ is equal to the sum of the one of the characteristic values $\mu$ of $T_{\lambda^*}$ such that $\mu^{k+1} = \mu_j^{k+1}$. By \Cref{lem-mod-res}, for large $t$, $e^{i \alpha (k+1)}t^{k+1}$ is not a characteristic value of $T_{\lambda^*}^{k+1}$. We obtain, by \cite[Theorem 12.4]{Agmon}, that the set of the characteristic values of $(T_{\lambda^*}^{k+1})_{t^{k+1} e^{ i (k+1) \alpha}}$ is given by
$$
\Big\{ \mu_j^{k+1} - t^{k+1} e^{ i (k+1) \alpha}\: ; \:j \ge 1\Big\}.
$$
We now derive from \eqref{HS-1500*} that
\begin{multline*}
\mbox{trace} \Big(M_t^{-1}{\bf T}_{\alpha,t} M_t (M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2t^{k+1} e^{i (k+1) \alpha}} \Big) \\[6pt]
= \sum_{j} \frac{1}{(\mu_j^{k+1} - t^{k+1} e^{ i (k+1) \alpha})(\mu_j^{k+1} + t^{k+1} e^{ i (k+1) \alpha})},
\end{multline*}
which yields, since $\alpha = \frac{\pi}{4 (k+1)}$,
\begin{equation}
\label{HS-1500**}
\mbox{trace}\Big(M_t^{-1}{\bf T}_{\alpha,t} M_t (M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2t^{k+1} e^{i (k+1) \alpha}} \Big) = \sum_{j} \frac{1}{\mu_j^{2(k+1)} - i t^{2(k+1)}}.
\end{equation}
We have, by \Cref{prop-pre},
\[
\limsup_{|\mu_j| \to + \infty}\left | \frac{\Im (\mu_j)}{\mu_j} \right | =0.
\]
As a consequence and as in \cite[Proof of Corollary 3]{MinhHung2}, we derive that
\begin{equation}
\label{HS-151}
\sum_j \frac{1}{\mu_j^{2k+2} - i t^{2k+2}} - \sum_j \frac{1}{|\mu_j|^{2k+2} - i t^{2k+2}} = o(t^{ - 2k - 2+\frac{d}{2}}) \mbox{ as }t \to + \infty.
\end{equation}
Combining \eqref{HS-1500**} and \eqref{HS-151} yield
\begin{multline}
\label{HS-p1}
\mbox{trace}\Big(M_t^{-1}{\bf T}_{\alpha,t} M_t (M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2t^{k+1} e^{i (k+1) \alpha}} \Big) \\[6pt]
= \sum_{j} \frac{1}{|\mu_j|^{2(k+1)} - i t^{2(k+1)}} + o(t^{ - 2k - 2+\frac{d}{2}}) \mbox{ as }t \to + \infty.
\end{multline}
Applying \eqref{HS-Rm} with $\gamma_1 = t^{k+1} e^{i (k+1) \alpha}$ and $\gamma_2 = -2t^{k+1} e^{i (k+1) \alpha}$ and using \Cref{lem-mod-res}, we derive that
\begin{equation}
\label{HS-p111}
(M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2t^{k+1} e^{i (k+1) \alpha}} = M_t^{-1}{\bf T}_{\beta, t} M_t.
\end{equation}
Since
\[
\mbox{trace}\left (M_t^{-1} {\bf T}_{\alpha, t}{\bf T}_{\beta, t} M_t\right ) = \mbox{trace} \left ( {\bf T}_{\alpha, t}{\bf T}_{\beta, t}\right ),
\]
it follows from \eqref{HS-p1} and \eqref{HS-p111} that
\begin{equation}
\label{HS-p2}
\mbox{trace} \left ( {\bf T}_{\alpha, t}{\bf T}_{\beta, t}\right )
= \sum_{j} \frac{1}{|\mu_j|^{2(k+1)} - i t^{2(k+1)}} + o(t^{ - 2k - 2+\frac{d}{2}}) \mbox{ as }t \to + \infty.
\end{equation}
Applying \Cref{prop-ker}, we derive from \eqref{HS-p2} that, as $t \to + \infty$
\begin{equation}
\label{HS-271}
\sum_j \frac{1}{|\mu_j|^{2k+2}- i t^{2k+2}} = {\bf c}t^{-2k-2+\frac{d}{2}}+ o(t^{-2k-2+\frac{d}{2}}).
\end{equation}
Considering the imaginary part of \eqref{HS-271} we get, for $\tau = t^{4k+4}$,
\[
\sum_j \frac{1}{|\mu_j|^{4k+4}+\tau} = \Im({\bf c}) \tau^{\frac{d}{8k+8}-1}+ o(\tau^{\frac{d}{8k+8}-1})\mbox{ as }\tau \to + \infty.
\]
Since $\lambda_j = \mu_j+\lambda^*$, it follows that, as $\tau \to + \infty$,
\begin{equation}\label{HS-p4}
\sum_j \frac{1}{|\lambda_j|^{4k+4}+\tau} = \Im({\bf c}) \tau^{\frac{d}{8k+8}-1}+ o(\tau^{\frac{d}{8k+8}-1}).
\end{equation}
Using the fact
\begin{equation*}
\sum_j \frac{1}{|\lambda_j|^{4k+4}+\tau} = \int_0^{\infty} \frac{d\mathcal{N}(s^{\frac{1}{4(k+1)}})}{s+\tau},
\end{equation*}
we derive that
\begin{equation}\label{HS-p5}
\int_0^{\infty} \frac{d\mathcal{N}(s^{\frac{1}{4(k+1)}})}{s+\tau} = \Im({\bf c}) \tau^{\frac{d}{8k+8}-1}+ o(\tau^{\frac{d}{8k+8}-1}) \mbox{ as }\tau \to + \infty.
\end{equation}
Applying a Tauberian Theorem of Hardy and Littlewood (see, e.g., \cite[Theorem 14.5]{Agmon}), we obtain
\[
\mathcal{N}(t) = \frac{\Im ({\bf c})}{\frac{d}{8(k+1)} \int_0^\infty s^{\frac{d}{8(k+1)}-1}(1+s)^{-1}ds} t^{\frac{d}{2}} + o(t^{\frac{d}{2}})\quad \quad \mbox{ as }t \to + \infty,
\]
which is the conclusion.
It remains to prove \eqref{prop-spec-ct}.
Applying \eqref{HS-Rm} with $\gamma_1 = t^{k+1} e^{i (k+1) \alpha}$ and $\gamma_2 = -2(t+s)^{k+1} e^{i (k+1) \alpha}$ and using \Cref{lem-mod-res}, we derive that
$$
(M_t^{-1}{\bf T}_{\alpha,t} M_t )_{-2(t+s)^{k+1} e^{i (k+1) \alpha}} = M_r^{-1}{\bf T}_{\widetilde{\alpha}, r} M_r.
$$
where
$$
\widetilde{\alpha} = \alpha +\frac{\pi}{k+1} \quad \mbox{ and } \quad r = (2(t+s)^{k+1} - t^{k+1})^{\frac{1}{k+1}}.
$$
Thus by \cite[Theorem 12.14]{Agmon},
\begin{equation}
\label{chara-op-2}
\Big \{
s_j +2 (t+s)^{k+1}e^{i\alpha (k+1)}~;~ j\geq 1
\Big \} \mbox{ is the set of characteristic values of }M_r^{-1}{\bf T}_{\widetilde{\alpha}, r} M_r.
\end{equation}
Applying \Cref{cor-p} and using \eqref{def-Mt}, we have
\begin{equation}
\label{ct-p-0}
\vvvert M_{t}^{-1} {\bf T}_{\alpha, t} M_{t} \vvvert \leq C t^{-k + \frac{d}{4}} \quad \mbox{ and } \quad \vvvert M_{r}^{-1} {\bf T}_{\widetilde{\alpha}, r} M_{r} \vvvert \leq C r^{-k+\frac{d}{4}}
\end{equation}
for some constant $C>0$ which does not depend on $s$ (and $t$). By \cite[Theorem 12.12]{Agmon} we have
\begin{equation}
\label{ct-p-01}
|\mbox{trace}\Big(M_t^{-1}{\bf T}_{\alpha,t} M_t (M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2\alpha^{k+1} (t+s)^{k+1}} \Big)| \leq \vvvert M_{t}^{-1} {\bf T}_{\alpha, t} M_{t} \vvvert \vvvert M_{r}^{-1} {\bf T}_{\widetilde{\alpha} , r} M_{r} \vvvert.
\end{equation}
Since $-k+\frac{d}{4}<0$ it follows from \eqref{ct-p-0} and \eqref{ct-p-01} that
\begin{equation}\label{HS-p10}
\lim_{s \to + \infty} \mbox{trace}\Big(M_t^{-1}{\bf T}_{\alpha,t} M_t (M_t^{-1}{\bf T}_{\alpha,t}M_t )_{-2\alpha^{k+1} (t+s)^{k+1}} \Big) = 0.
\end{equation}
On the other hand,
by \cite[Theorem 12.14]{Agmon},
\begin{multline}\label{HS-p11}
\Big|\sum_{j} \frac{1}{s_{j}(s_{j}+2e^{i\alpha (k+1)} (t+s)^{k+1})} \Big|^2 \le \sum_{j} \frac{1}{|s_{j}|^2} \sum_{j} \frac{1}{|s_{j}+2e^{i\alpha (k+1)} (t+s)^{k+1}|^2} \\[6pt]
\mathop{\le}^{\eqref{chara-op-2}} \vvvert M_{t}^{-1} {\bf T}_{\alpha, t} M_{t} \vvvert^2 \vvvert M_{r}^{-1} {\bf T}_{\widetilde{\alpha} , r} M_{r} \vvvert^2 \mathop{\to}^{\eqref{ct-p-0}} 0 \mbox{ as } s \to + \infty.
\end{multline}
Combining \eqref{HS-p10} and \eqref{HS-p11} yields $c_{t} = 0$, which is \eqref{prop-spec-ct} .
\medskip
The proof is complete.
\end{proof}
\subsection{Proof of \Cref{thm1}} As in \cite[p.34]{MinhHung2}, we derive from \Cref{prop-ker} that
\[
\begin{aligned}
\Im ({\bf c}) &= \frac{1}{(2\pi)^d} \int_{\Omega} \sum_{\ell=1}^2 \int_{\mathbb{R}^d}\frac{1}{\left ( \Sigma_{\ell} (x_0)^{-1} A(x_0) \xi \cdot \xi \right )^{4k+4} +1 } d \xi dx \\
&= \frac{1}{(2\pi)^d} \sum_{\ell=1}^2\int_\Omega\left | \left \{ \xi : A(x) \xi \cdot \xi < \Sigma_{\ell} (x)\right \}\right | dx \frac{d}{8(k+1)}\int_0^\infty s^{\frac{d}{8(k+1)}-1}(1+s)^{-1} ds.
\end{aligned}
\]
The conclusion now follows from \Cref{prop-spec}. \qed
\section{Completeness of the generalized eigenfunctions of the transmission eigenvalue problem - Proof of \Cref{thm2}} \label{sect-C}
By \Cref{lem-mod-res}, for all $\widetilde{\theta} \in (0, 2 \pi) \setminus \{\pi \}$, there exists $t_{\widetilde{\theta}}>0$ such that, for $t>t_{\widetilde{\theta}}$,
\begin{equation}
\label{compl-20}
(T_{\lambda^*}^{k+1})_{t e^{i\widetilde{\theta}} } = M_{t_k}^{-1}{\bf T}_{\theta,t_k} M_{t_k},
\end{equation}
where
\[
\theta = \frac{\widetilde{\theta}}{k+1} \quad \mbox{ and }\quad t_k = t^{\frac{1}{k+1}}.
\]
By \Cref{prop-pre} and \Cref{cor-p},
\begin{equation}
\label{compl-3-1}
\vvvert M_{t_k}^{-1}{\bf T}_{\theta,t_k} M_{t_k} \vvvert \leq Ct_k^{-k+\frac{d}{4}}\quad \quad \mbox{ and } \quad \quad \| M_{t_k}^{-1}{\bf T}_{\theta,t_k} M_{t_k} \|_{L^2(\Omega) \to L^2(\Omega)}\leq Ct_k^{-k}.
\end{equation}
In particular, $T_{\lambda^*}^{k+1}$ is a Hilbert-Schmidt operator; moreover, for $t > t_\theta$,
\begin{equation}
\label{compl-3-2}
\vvvert (T_{\lambda^*}^{k+1})_{t e^{i\widetilde{\theta}}} \vvvert \leq C_{\widetilde{\theta}} t^{-1+\frac{1}{k+1}+\frac{d}{4(k+1)}}.
\end{equation}
Since $k = [d/2] + 1$, it follows that $-1+\frac{1}{k+1}+\frac{d}{4(k+1)} \leq 0$. This implies that
\begin{equation}
\label{compl-3-20}
\mbox{for all }\widetilde{\theta} \in (0, 2 \pi) \setminus \{\pi \} \mbox{ there exist }t_{\widetilde{\theta}}>0 \mbox{ and }C_{\widetilde{\theta}}>0 \mbox{ such that }\sup_{t>t_{\widetilde{\theta}}} \vvvert (T_{\lambda^*}^{k+1})_{t e^{i\widetilde{\theta}}} \vvvert \leq C_{\widetilde{\theta}}.
\end{equation}
Since $T_{\lambda^*}^{k+1}$ is a Hilbert-Schmidt operator, it follows from \cite[Theorem 16.4]{Agmon} that
\begin{itemize}
\item[i)] the space spanned by the generalized eigenfunctions of $T_{\lambda^*}^{k+1}$ is equal to $\overline{\mbox{range}(T_{\lambda^*}^{k+1})}$, the closure of the range of $T_{\lambda^*}^{k+1}$ with respect to the $L^2$-topology.
\end{itemize}
In fact, in order to be able to apply \cite[Theorem 16.4]{Agmon}, one requires the assumptions on the directions of the minimal growth of the modified resolvent of $T_{\lambda^*}^{k+1}$. We have only proved \eqref{compl-3-2} and \eqref{compl-3-20} instead of this requirement. Nevertheless, this is sufficient to derive 1) using almost the same arguments in \cite{Agmon} (see also \cite{Robbiano16}).
\medskip
The rest of the proof is as in \cite{MinhHung2, Fornerod-Ng1}. We have
\begin{itemize}
\item[ii)] $\mbox{range } T_{\lambda^*}^{k+1}$ is dense in $[L^2(\Omega)]^2$ since $\mbox{range }T_{\lambda^*}$ is dense in $[L^2(\Omega)]^2$ and $T_{\lambda^*}$ is continuous,
\item[iii)] the space spanned by the general eigenfunctions of $T_{\lambda^*}^{k+1}$ associated to the non-zero eigenvalues of $T_{\lambda^*}^{k+1}$ is equal to the space spanned by the general eigenfunctions of
$T_{\lambda^*}$ associated to the non-zero eigenvalues of $T_{\lambda^*}$ . This can be done as in the last part of the proof of \cite[Theorem 16.5]{Agmon}. Consequently, the space spanned by all generalized
eigenfunctions of $T_{\lambda^*}^{k+1}$ is equal to the space spanned by all generalized eigenfunctions of
$T_{\lambda^*}$.
\end{itemize}
The conclusion now follows from i), ii), and iii). \qed
\bigskip
\noindent {\bf Acknowledgement .} The authors thank Fioralba Cakoni for attracting their attention to the problem and stimulating discussions.
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
Access to transportation is a crucial factor in determining quality of life. In highly populated locations, public transit can provide mobility services on a large scale, connecting individuals to jobs, educational opportunities, nutritious food, and medical visits \cite{sanchez2004transit}. In low-density locations, however, public transit agencies find it financially difficult to maintain an acceptable level of service, resulting in restricted coverage or transit gaps \cite{borjesson2020rural}. In these underserved locations, rich inhabitants can still use their own vehicles to go around, but disadvantaged groups (e.g., low-income individuals) are more likely to rely on public transportation and have fewer options to meet their fundamental travel needs.
This transit equity gap hurts us all: when our neighbors are isolated and struggling, it undermines the local economy and weakens the social cohesion of the city. Today, these negative impacts are further compounded by the COVID-19 pandemic \cite{meredith2021relationship}. Closing this equity gap is more important than before.
However, transportation is on the verge of dramatic upheaval. Autonomous vehicle (AV) technology is being tested and deployed in numerous cities around the world. While level-5 fully autonomous driving could be decades away, ride-hailing services with level-4 AVs, offered by transportation network companies (TNCs) in urban and suburban areas, are expected to arrive much earlier \cite{Waymo2022SF}. Automation will bring radical changes to TNC business models: (a) compared to human drivers who selfishly reposition themselves to pursue individual earning opportunities, AVs can be centrally dispatched to achieve system-level objectives \cite{zhang2016control}; (b) compared to human drivers who share a significant portion of the trip fare, AVs can remove the costs associated with human drivers and reduce the trip fare substantially \cite{chen2016operations}. These favorable factors have profound implications on transport equity: with better control of their fleets and lower operational costs, TNCs can improve mobility services in underserved areas and enhance transportation access for disadvantaged populations while preserving financial sustainability.
Despite its potential in mitigating inequities, the actual equity impact of AV deployment in the TNC market remains uncertain. Unlike public transportation, AV technology is being led by multi-billion dollar companies in the private sector. The primary goal of these companies is to maximize the profits, which is naturally contradicted with the equity objectives of the social planner. As a result, the deployment of AV through TNC platforms may negatively affect spatial equity\footnote{ Spatial equity, also referred to as horizontal equity, requires that people in different regions of the city should
have access to the same quality of mobility services.} and social equity\footnote{Social equity, also referred to as vertical equity, requires that people that are disadvantaged in socioeconomic status (e.g., income, education, etc) should be favored by transport policies.} within the multimodal transportation network. On the spatial side, for-profit ride-hailing companies (e.g., Uber, Waymo, Baidu, etc.) tend to focus their services in regions with high demand \cite{hughes2016transportation}. These high-demand areas are often congested communities in the city core, which already have robust public transit services. The geographic concentration of TNC services will further enlarge the accessibility gap between wellserved and underserved areas, which exacerbates spatial inequity. On the social side, as for-hire AVs flock into these areas, they will compete with public transit, degrade transit service qualities, and disproportionately affect low-income individuals that are more transit-dependent \cite{hall2018uber,erhardt2022transportation}. This will enlarge the accessibility gap between passengers with distinct socioeconomic status (e.g., level of income), which exacerbates social inequity. These issues can be further reinforced as AV technology enables TNCs to have lower costs and stronger control over their fleet. Without regulatory interventions, market forces may drive us to a transport equity catastrophe: public transit ridership will decline, traffic congestion will increase, people who are already transportation-disadvantaged will be left further behind, and the benefits of AVs will be inaccessible to those who need them most.
This paper aims to assess the equity impacts of for-hire AVs in the ride-hailing network and investigate regulatory policies that ensure AV deployment can benefit underserved areas and disadvantaged groups. We consider a TNC platform that deploys a fleet of AVs to offer mobility-on-demand services to maximize its profit, a public transit agency that offers transit services to maximize transit ridership under a fixed budget, and a group of multiclass passengers with different income levels who can choose among ride-hailing, public transit, and bundled services that combines ride-hailing and public transit to reach their destinations. To maintain an equitable multimodal transportation network, it is crucial to investigate: (a) whether AV deployment will exaggerate spatial and social inequity in an unregulated environment; (b) whether it is necessary to impose regulatory intervention; and (c) how to design regulatory policies to mitigate the negative impacts of AVs and improve transport equity. To this end, we will develop a game-theoretic network equilibrium model to examine how travelers with heterogeneous socioeconomic attributes make mode choices in the multimodal transportation system, how the TNC platform and the transit agency interact with each other to reach their own objectives, and how these interactions are affected by the imposed regulations. The major contribution of this paper are summarized below:
\begin{itemize}
\item We develop a game-theoretic model to characterize the strategic interactions between the TNC platform and the public transit agency over the transport network. The model captures the essential economic aspects of the multimodal transportation network, including the spatial prices of the TNC platform, the service fare and frequencies of public transit, the spatial distribution of idle AVs, waiting times for ride-hailing and transit services, and the modal choice of multiclass passengers with distinct income groups. We use the model to evaluate the spatial and social equity in transport access among distinct population groups. {\em To our best knowledge, this is the first work that studies spatial and social equity for multimodal transportation networks considering the strategic interaction between TNC platforms and public transit agencies.}
\item The proposed game-theoretic model consists of a profit-maximizing subproblem for the TNC platform and a ridership maximization subproblem for the public transit agency, which are interdependent and highly non-concave. We use the best response algorithm to derive a candidate equilibrium solution, and evaluate the quality of these solutions by exploring the special structures of each subproblem. In particular, for TNC's profit-maximization subproblem, we use primal decomposition to compute the globally optimal solution to a relaxed reformulation of the problem, which establishes a tight upper bound to evaluate the performance of TNC's decisions. For the transit ridership maximization subproblem, we reformulate the problem and find conditions under which the reformulated problem is concave and its globally optimal solution can be efficiently obtained. Overall, this enable us to assess the quality of the derived candidate equilibrium solution by showing that it is at least as good as an $\epsilon$-Nash equilibrium \cite{daskalakis2006note, li2019connections}, where $\epsilon$ can be determined numerically through the above procedure.
\item We use Theil index to quantify both spatial and social inequity in transport accessibility and investigate the equity impacts of AV deployment in the absence of regulations. We find that although the proliferation of AVs improves overall accessibility, the benefits are not fairly distributed across different geographic locations and among distinct population groups, and the deployment of AVs will exacerbate both spatial and social inequity gaps at the same time. We point out that the increase in spatial inequity arises from the geographic concentration of ride-hailing services in high-demand areas due to the for-profit nature of the TNC platform, and the increased social inequity gap is due to the fact that the benefits of AV-enabled ride-hailing services are primarily enjoyed by individuals with higher income, whereas those with lower income are disproportionately transit-dependent.
\item We evaluate the equity impacts of two regulatory policies, including (a) a minimum service-level requirement on ride-hailing service, which improves the spatial equity in the transport network; (b) a subsidy on ride-hailing trips that serve as first/last-mile connection to public transit, which promotes the use of public transit and improves the social equity of the transport network. We show that the minimum service-level requirement entails a trade-off: as a higher minimum service level is imposed, the spatial inequity reduces, but the social inequity will be exacerbated. In contrast, subsidies on bundled trips can enhance accessibility, increase transit ridership, improve the profit of the ride-hailing platform, and more crucially, simultaneously bridge spatial and social inequity gaps. The reasons for the above results are identified and carefully discussed in Section \ref{service} and \ref{subsidies}, respectively.
\end{itemize}
The remainder of this article will proceed as follows. Related works in the existing literature will be surveyed in Section \ref{related works}. A game-theoretic model will be developed for the competition between TNC and transit in Section \ref{game-theoretic model}. A solution method will be proposed in Section \ref{solution method}, and numerical studies will be presented in Section \ref{market outcomes}-\ref{regulations}. Finally, concluding remarks are offered in Section \ref{conclusion}.
\section{Related Works} \label{related works}
There is a large body of literature that investigates the operational strategies of for-hire AV fleets, examines the interaction between TNC and public transit, quantifies their impacts on transport equity, and identifies measures to mitigate their negative externalities. Below we summarize the relevant literature from four aspects: (a) autonomous mobility-on-demand services; (b) interactions between TNC and public transit; (c) impacts of AVs and TNCs on transport equity; and (d) regulations on AVs and TNCs.
\subsection{Autonomous Mobility-on-Demand Services}
A recurrent topic of interest in the literature is the planning and operation of for-hire AV fleets to provide autonomous mobility-on-demand (AMoD) services. \cite{zhang2016control} presented a queuing-theoretic model to investigate the optimal rebalancing strategies of AMoD fleets. Their case study of New York showed that the current taxi demand in Manhattan can be met with about 8,000 AVs (roughly 70\% of the size of the current taxi fleet operating in Manhattan). \cite{wen2019value} investigated the value of demand information in AMoD systems and showed that aggregate demand information can lead to better service performance. \cite{zhang2016model} developed a model predictive control (MPC) approach to optimize vehicle scheduling and routing in AMoD systems. \cite{tsao2018stochastic} utilized short-term probabilistic forecasts for dispatching and rebalancing of AMoD systems and developed a stochastic MPC algorithm to efficiently solve the problem. \cite{alonso2017predictive} proposed a predictive method for the vehicle routing and assignment for an AMoD system with ridesharing. \cite{iglesias2019bcmp} formulated AMoD systems as a closed, multi-class Baskett-Chandy-Muntz-Palacios (BCMP) queuing network model to characterize the passenger arrival process, traffic, the state-of-charge of electric vehicles, and the availability of vehicles at the stations and derive the routing and charging policies. \cite{hyland2018dynamic} used agent-based simulations to evaluate the performance of AV-traveler assignment strategies. Their results showed that the spatial distribution of traveler requests significantly impacts the empty fleet miles generated by shared AVs. \cite{gueriau2018samod,guo2020deep} investigated the ride-sharing vehicle dispatching as well as the request-vehicle assignment problem for AMoD systems using reinforcement learning. \cite{chen2016operations} examined the operation of shared electric AV fleets under distinct vehicle range and charging infrastructure settings. \cite{turan2020dynamic} investigated the joint routing, battery charging, and pricing for electrified AMoD systems and used deep reinforcement learning to develop a near-optimal control policy. For a comprehensive literature review on the operation of AMoD systems, please refer to \cite{zardini2022analysis}.
{\em Note that all the aforementioned works primarily focus on the planning and operation of AMoD systems, while the interaction between AMoD and other modes is not explicitly considered.}
\subsection{Interactions Between AMoD and Public Transit}
An increasing body of literature revealed the interaction between AMoD/ridesourcing services and public transit using both data-driven and model-based methods. Based on real-world data, \cite{hall2018uber} estimated the effect of Uber on public transit ridership using a difference-in-differences design. Their results showed that Uber is a complement to the average transit agency, increasing ridership by five percent after two years. On the other hand, \cite{meredith2021relationship} analyzed detailed individual trip records at a spatially and temporally granular level, and found that only approximately 2\% of TNC trips complement public transit, while 45\% to 50\% of TNC trips substitute transit and this percentage drops during COVID-19 shutdowns. Based on realistic TNC data and public transit data, \cite{erhardt2022transportation} found that TNCs are responsible for a net ridership decline of about 10\%, offsetting net gains from other factors such as service increases and population growth. In modeling-wise, \cite{ke2021equilibrium} proposed a user equilibrium based mathematical model to explore the complement and substitution of ride-sourcing to public transit. \cite{zhu2020analysis} established a network equilibrium model to analyze the multimodal commute behavior with ridesplitting programs as both feeders and competitors to public transit. \cite{zhu2021competition} developed a bilevel game-theoretic approach to model the cooperative and competitive relationship between the TNC and the government. They highlighted that a carefully designed subsidy can benefits both the TNC and society, especially in areas with low public transit accessibility. \cite{salazar2019intermodal} and \cite{gurumurthy2020first} considered AMoD services for first/last-mile connection to public transport. \cite{salazar2018interaction} presented a network flow model to capture the interaction between AMoD and public and transit, and designed a pricing and tolling scheme that achieves the social optimum. \cite{pinto2020joint}, \cite{sieber2020improved} and \cite{kumar2022planning} studied the joint planning and operation of shared AV fleet and public transit to improve mobility services in low-density areas.
This paper significantly differs from the aforementioned works in two aspects: (a) the aforementioned studies primarily consider the multimodal transportation system at the aggregate level, and in contrast, our work considers a multimodal transportation network with multiclass travelers; (b) all aforementioned studies neglect the equity impacts of ride-hailing services in the multimodal transportation systems. {\em To our best knowledge, our paper is the first work that studies spatial and social equity for multimodal transportation networks considering the strategic interaction between TNC platforms and public transit agencies.}
\subsection{Impacts of AVs and TNCs on Transport Equity}
Emerging technologies, such as AVs and TNCs, have profoundly disrupted the transportation sector, but their impacts on transport equity remain unclear \cite{dianin2021implications}. On the one hand, AVs and TNCs have great potential to close existing inequity gaps \cite{shaheen2017travel}. \cite{brown2018ridehail} revealed that black riders were 73 percent more likely than white riders to have a taxi trip canceled and waited for 6-15 minutes longer than white riders, while TNC services nearly eliminate the racial-ethnic discrimination in service quality. \cite{palm2021equity} found significant academic evidence that new mobility technologies like ride-hailing have the potential to meaningfully address some disadvantaged travelers’ transportation problems or close existing gaps in transit services like the first/last mile. \cite{wen2018transit} simulated and evaluated integrated AV and public transportation systems and showed that encouraging ride-sharing, allowing in-advance requests, and combining fare with transit help enable service integration and encourage sustainable travel. \cite{chen2017connecting} compared two different relative spatial position (RSP) designs in an integrated e-hailing/fixed-route transit system and showed the great potential of e-hailing in filling transit gaps. \cite{cohn2019examining} and \cite{nahmias2021benefits} examined the equity impacts of AVs and showed that AMoD services could expand transportation access to car-free and underserved populations and bridge the mobility gaps between demographic groups. \cite{ahmed2020quantifying} revealed that shared AV mobility services can significantly increase job accessibility and low-income workers in low-density areas enjoy a higher benefit. On the other hand, there are growing concerns that TNCs and AVs may widen existing inequity gaps. \cite{ge2016racial} and \cite{yang2021equitable} investigated racial discrimination in TNCs and reported that customers with color experience more trip cancellations and longer waiting times. \cite{hughes2016transportation} studied the spatial distribution of TNC services and found that waiting time is lower in densely populated areas. In addition, the deployment of AV technology could also raise new issues such as induced demand \cite{cohn2019examining} and exacerbated geographic concentration \cite{jiao2021shared}. Given these concerns, it is widely agreed that regulatory policies should be imposed to ensure that AV technology promotes equity in future mobility systems \cite{milakis2017policy}.
{\em Distinct from most aforementioned works, we aim to reveal the fundamental mechanisms by which market incentives lead to systemic transport inequity under the interaction between TNC and public transit.} We will develop a mathematical framework that characterizes how market forces drive TNC services and public transit in the absence of regulatory intervention and quantifies both spatial and social inequity in the multimodal transportation system.
\subsection{Regulations on AVs and TNCs}
Regulations on TNCs and AVs are emerging topics that receive increasing attention from the transportation research community recently. In the existing literature, policy studies on human-driver TNCs are primarily focused on driver minimum wage \cite{parrott2018earnings,li2019regulating, shetty2022analysis}, congestion charges \cite{li2021impact,li2021spatial}, and cap on the platform commission rate \cite{vignon2021regulating,vignon2023regulating}. \cite{parrott2018earnings} examined the effects of the proposed minimum wage policy in New York City. \cite{li2019regulating} developed a queueing theoretic equilibrium model to assess the impacts of distinct regulations on TNCs including (1) A minimum wage for drivers, (2) a cap on the number of drivers or vehicles, and (3) a per-trip congestion tax. \cite{li2021impact} further compared the impacts of two types of congestion charge: (a) a charge per TNC trip, and (b) a charge per vehicle operating hour, on TNCs. \cite{li2021spatial} studied the spatial pricing for ride-sourcing platforms under congestion charge and evaluated the impacts of three forms of congestion charge: (a) a one-directional cordon charge on ride-sourcing vehicles that enter the congestion area; (b) a bi-directional cordon charge on ride-sourcing vehicles that enter or exit the congestion area; (c) a trip-based congestion charge on all ride-sourcing trips. \cite{vignon2021regulating} showed that a commission cap and a toll are effective in regulating the market and achieving a socially desirable outcome. \cite{vignon2023regulating} regulated the competitive market with ride-hailing services and street-hailing services and found that commission cap regulation can replicate the second-best in the market and it is effective even in a congested setting.
On the other hand, regulations on AMoD services are relatively under-examined, and we only find a handful of relevant works. For instance, \cite{freemark2020policies} surveyed municipal officials throughout the U.S. to investigate how they support various AV-related regulatory policies. \cite{simoni2019congestion} used agent-based simulations to investigate how toll strategies on AMoD systems impact traffic congestion and social welfare. \cite{salazar2019intermodal} considered an intermodal AMoD system that coordinates with public transit, and designed a tolling scheme to achieve the social optimum in a perfect market with selfish agents. \cite{dandl2021regulating} developed a tri-level framework to capture the interactions among regulatory agencies, AMoD service providers, and travelers.
{\em Although TNC regulations have garnered a growing amount of attention in recent years, neither the equity aspects of TNC regulation nor the interactions between distinct modes under TNC regulation have been investigated. Our paper differs from all aforementioned works as we investigate how to design policies that interact with the multimodal transportation system to mitigate spatial and social inequity. This has not been studied in the existing literature.}
\section{The Game-Theoretic Model} \label{game-theoretic model}
Consider a city divided into $M$ geographic zones. These zones are connected by a multimodal transport network which consists of a road network $\mathcal{G}_r(\mathcal{V},\mathcal{E}_r)$ and a transit network $\mathcal{G}_t(\mathcal{V},\mathcal{E}_t)$, where $\mathcal{V}$ denotes the set of zones, $\mathcal{E}_r$ denotes the set of roads, and $\mathcal{E}_t$ denotes the set of transit lines. We consider passenger trips at the zonal granularity and assign each trip with an origin zone $i\in\mathcal{V}$ and a destination zone $j\in\mathcal{V}$. The travel demand is subdivided into $K$ classes based on income levels in each zone. In the multimodal transport system, the TNC platform operates a fleet of AVs to provide mobility-on-demand services on the road network, and the public transit agency provides transit services through the transit network, and passengers in different classes make mode choices over the multimodal transport network. The decision-making of passengers, the TNC platform, and the public transit agency interact with each other and constitute the market equilibrium. This section formulates a mathematical model to capture the competition between the TNC platform and the public transit agency over a transportation network. The details of the model are presented below.
\subsection{Incentives of Passengers}
We consider a discrete choice model for passengers who choose among: (1) direct ride-hailing/AMoD services, indexed by $a$; (2) public transit, indexed by $p$; (3) bundled services combining AMoD and transit, indexed by $b$; and (4) outside option, indexed by $o$, to reach their destinations at the minimum cost. Using AMoD for the first-mile or/and last-mile connection, bundled services can be further categorized into three possible scenarios: (1) the bundle of first-mile AMoD service and public transit, indexed by $b_1$; (2) the bundle of last-mile AMoD service and public transit, indexed by $b_2$; and (3) the bundle of both first-mile and last-mile AMoD service and public transit, indexed by $b_3$. The costs of different mobility modes collectively determine passenger demand through the following relation:
\begin{equation} \label{general_demand_function}
\begin{cases}
\lambda_{ij,k}^a = \lambda_{ij,k}^0 F^a\left(c_{ij,k}^a,c_{ij,k}^p,c_{ij,k}^{b_1},c_{ij,k}^{b_2},c_{ij,k}^{b_3},c_{ij,k}^{o}\right) \\
\lambda_{ij,k}^p = \lambda_{ij,k}^0 F^p\left(c_{ij,k}^a,c_{ij,k}^p,c_{ij,k}^{b_1},c_{ij,k}^{b_2},c_{ij,k}^{b_3},c_{ij,k}^{o}\right) \\
\lambda_{ij,k}^{b_1} = \lambda_{ij,k}^0 F^{b_1}\left(c_{ij,k}^a,c_{ij,k}^p,c_{ij,k}^{b_1},c_{ij,k}^{b_2},c_{ij,k}^{b_3},c_{ij,k}^{o}\right) \\
\lambda_{ij,k}^{b_2} = \lambda_{ij,k}^0 F^{b_2}\left(c_{ij,k}^a,c_{ij,k}^p,c_{ij,k}^{b_1},c_{ij,k}^{b_2},c_{ij,k}^{b_3},c_{ij,k}^{o}\right) \\
\lambda_{ij,k}^{b_3} = \lambda_{ij,k}^0 F^{b_3}\left(c_{ij,k}^a,c_{ij,k}^p,c_{ij,k}^{b_1},c_{ij,k}^{b_2},c_{ij,k}^{b_3},c_{ij,k}^{o}\right) \\
\lambda_{ij,k}^{o} = \lambda_{ij,k}^0 F^{o}\left(c_{ij,k}^a,c_{ij,k}^p,c_{ij,k}^{b_1},c_{ij,k}^{b_2},c_{ij,k}^{b_3},c_{ij,k}^{o}\right)
\end{cases},
\end{equation}
where $\lambda_{ij,k}^{t},t\in \mathcal{T}=\{a,p,b_1,b_2,b_3,o\}$ represents the arrival rate of passengers of mode $t$ from origin zone $i$ to destination zone $j$ in income class $k$; $\lambda_{ij,k}^0$ is the arrival rate of potential passengers from zone $i$ to zone $j$ in income class $k$; $c_{ij,k}^{t},t\in \mathcal{T}$ denotes the average generalized travel costs for passengers from origin $i$ to destination $j$ in income class $k$ by choosing mode $t$; $F^{t}(\cdot),t\in \mathcal{T}$ is an aggregate function which captures the proportion of passengers choosing mode $t$. We assume that $F^t(\cdot)$ is a continuously differentiable function with $\frac{\partial F^t}{\partial c_{ij,k}^{t'}}<0$ if $t=t'$ and $\frac{\partial F^t}{\partial c_{ij,k}^{t'}}\geq 0$ otherwise. This indicates that the demand for a specific mode is negatively dependent on the generalized cost of this mode, but positively dependent on the generalized cost of its alternative modes. Under the well-established logit model, the passenger demand over distinct modes can be characterized as:
\begin{equation} \label{logit_demand_function}
\lambda_{ij,k}^t = \lambda_{ij,k}^0 \frac{\exp \left(-\epsilon c_{ij,k}^t\right)}{\sum_{{t'} \in \mathcal{T}} \exp \left(-\epsilon c_{ij,k}^{t'}\right)} \quad t \in \mathcal{T} ,
\end{equation}
where $\epsilon$ is the scaling parameter in the logit model. Note that (\ref{general_demand_function}) includes the logit model (\ref{logit_demand_function}) as a special case, and the logit model is a classical discrete choice model which has been extensively applied in traffic demand modeling \cite{ahmed2020quantifying,ke2021equilibrium}. For analytical tractability, we use the logit model (\ref{logit_demand_function}) as the passenger demand functions throughout rest of this paper.
{The generalized travel costs of trips by different mobility modes are defined as the weighted sum of waiting time, in-vehicle time, trip fare, and access/walking time (if any).} They satisfy:
\begin{equation} \label{generalized_travel_cost}
\begin{cases}
c_{ij,k}^a = \alpha_k w_i^a + \beta_k \frac{l_{ij}^a}{v_a} + \gamma_k (b + r_i^a l_{ij}^a) \\
c_{ij,k}^p = \alpha_k w_{ij}^p + \beta_k \frac{l_{ij}^p}{v_p} + \gamma_k r^p l_{ij}^p + \theta_k \left(\frac{d_i}{v_w} + \frac{d_j}{v_w}\right) \\
c_{ij,k}^{b_1} = \alpha_k \left(w_i^a + w_{ij}^p \right) + \beta_k \left(\frac{d_i}{v_a} + \frac{l_{ij}^p}{v_p}\right) + \gamma_k \left(b + r_i^a d_i + r^p l_{ij}^p\right) + \theta_k \frac{d_j}{v_w} \\
c_{ij,k}^{b_2} = \alpha_k \left( w_{ij}^p + w_j^a \right) + \beta_k \left( \frac{l_{ij}^p}{v_p} + \frac{d_j}{v_a}\right) + \gamma_k \left( r^p l_{ij}^p + b + r_j^a d_j\right) + \theta_k \frac{d_i}{v_w} \\
c_{ij,k}^{b_3} = \alpha_k \left(w_i^a + w_{ij}^p + w_j^a\right) + \beta_k \left(\frac{d_i}{v_a} + \frac{l_{ij}^p}{v_p} + \frac{d_j}{v_a}\right) + \gamma_k \left(b + r_i^a d_i + r^p l_{ij}^p + b + r_j^a d_j\right) \\
\end{cases} ,
\end{equation}
where $w_i^a$ is the average waiting time for AMoD services in zone $i$; $w_{ij}^p$ is the average waiting time for transit services from zone $i$ to zone $j$; $b$ is the average base fare of AMoD rides; $r_i^a$ is the average per-distance rate of AMoD rides originating from origin zone $i$; $r^p$ is the average per-distance transit fare; $l_{ij}^a$ and $l_{ij}^p$ denote the average trip distances from zone $i$ to zone $j$ using transit and AMoD, respectively; $d_i$ is the average first/last-mile distance (i.e., the average distance from the exact origin/destination to the nearest transit station) in zone $i$; $v_a$, $v_p$ and $v_w$ represent the average speed of TNC autonomous vehicles, the average operating speed of public transit, and the average walking speed, respectively; $\alpha_k$, $\beta_k$, $\gamma_k$ and $\theta_k$ are the average weights associated with waiting time, in-vehicle time and trip fare, and walking time for passengers in income class $k$, which correspond to passengers' valuation on waiting time, trip time, monetary cost and walking time, respectively.
\begin{remark}
The structure of the TNC platform's pricing scheme $b+r_i^a l_{ij}^a$ is consistent with the industry practice. For instance, both Uber and Lyft have a base fare, a fixed per-time fare and a per-distance fare. They first calculate the variable trip fare based on the trip time (time-based charge) and trip distance (distance-based charge), then the total trip fare is calculated as the sum of the fixed base fare and the variable trip fare multiplied by a surge multiplier that reflects the real-time imbalance between supply and demand in each zone $i$. Considering a uniform average speed of TNC vehicles on roads, the time-based charge is proportional to the distance-based charge. In this case, the pricing policy $b+r_i^a l_{ij}^a$ can be viewed as an approximation of the sum of the base fare, the time-based charge and the distance-based charge multiplied by the zone-specific surge multiplier.
\end{remark}
The average passenger waiting time for AMoD services and transit services depends on the operational decisions of the TNC platform and the public transit agency. Below we characterize these endogenous relations for AMoD services and public transit separately.
{\em Waiting time for AMoD services}: The passenger waiting time for AMoD services $w_i^a$ depends on the passenger demand and the vehicle supply in the TNC market. To characterize $w_i^a$, note that each trip is initiated by a passenger who requests a pickup from zone $i$ through the user app (for either direct AMoD service or first-mile/last-mile service). Upon receiving the request, the platform immediately matches the passenger to the closest idle vehicle in the same zone. The matched idle vehicle then travels to pick up the waiting passenger. In this case, the TNC platform does not prioritize either type of AMoD services in the vehicle-passenger matching process, and passengers are served on a first-come-first-served basis. Consequently, the direct AMoD rides and AMoD rides as first/last mile connections are indifferent in passenger waiting time. Typically, the ride confirmation time (from the ride request being submitted to a vehicle being assigned) is negligible compared to the pickup time (from a vehicle being assigned to passenger pickup) in the matching process. Therefore, we ignore the ride confirmation time and approximate the passenger waiting time as the passenger pickup time. The pickup time depends on the distance of the nearest idle vehicle to the passenger, and the passenger waiting time $w_i^a$ can be characterized as a monotone function of the average number of idle vehicles $N_i^I$ in zone $i$. With slight abuse of notation, we use $w_i^a\left(N_i^I\right)$ to denote the passenger waiting time function of AMoD services. In this paper, we assume the passenger waiting time follows the well-established "square root law", which has been widely used in street-hailing taxi market \cite{douglas1972price}, radio dispatching taxi market \cite{arnott1996taxi}, and online ride-hailing market \cite{li2021spatial}, among others. This leads to the following passenger waiting time function:
\begin{equation} \label{waiting_time_AMoD}
w_i^a\left(N_i^I\right) = \frac{A_i}{\sqrt{N_i^I}},
\end{equation}
where $A_i$ is the scaling parameter that captures possible factors in the matching between idle TNC vehicles and passengers, such as the size of the zone, the average traffic speed on roads, and the demand/supply distribution, etc. Equation (\ref{waiting_time_AMoD}) indicates that passengers' average waiting time is inversely proportional to the square root of the number of idle AVs. The intuition is that if both waiting passengers and idle vehicles are uniformly and independently distributed in each zone, the distance between a passenger and her/his closest idle vehicle is inversely proportional to the square root of the total number of idle vehicles, which further determines the average waiting time. A detailed justification of (\ref{waiting_time_AMoD}) can be found in \cite{li2019regulating}.
{\em Waiting time for transit services}: The passenger waiting time for transit services $w_{ij}^p$ from zone $i$ to zone $j$ depends on the service frequencies of transit lines traversed by the passenger. Note that each zone probably contains multiple transit stations, thus passengers may board the transit service at different stations in the origin zone, or exist the transit service at distinct stations in the destination zone. This leads to multiple transit routes that connect two zones, each of which may incur distinct waiting times for passengers. To characterize the {\em average} passenger waiting time of public transit from zone $i$ to zone $j$, we assume that passengers choose the nearest station to board the public transit, and get off at the station which is closest to the destination. In this case, if we further assume passenger origins and destinations are uniformly distributed in each zone, then passenger demand for public transit from zone $i$ to zone $j$ is uniformly distributed over different transit routes between $i$ and $j$. Without loss of generality, let $\mathcal{R}_{ij}$ denote the set of all possible transit routes from zone $i$ to zone $j$. A transit route may traverse multiple transit lines and thus each transit route $r\in \mathcal{R}_{ij}$ is a set of transit lines. We assume that:
\begin{assumption} \label{assump_waiting_time_transit_2}
(i) vehicles arrive exponentially with a mean of the transit headway; and (ii) passengers can be served by the first arriving vehicle.
\end{assumption}
Assumption \ref{assump_waiting_time_transit_2} states that vehicle arrival at stations follows the exponential distribution, and upon the vehicle arrives at the station, passengers can get on the vehicle immediately without waiting. Under the assumption, the mean waiting time of a specific transit line $l$ is the inverse of its service frequency \cite{spiess1989optimal}, i.e., $\frac{1}{f_l}$, which has been widely used in transit-related studies \cite{liu2018remote,kumar2021designing,du2022sensitivity,li2023strategy}. Finally, the {\em average} waiting time for public transit $w_{ij}^p$ from zone $i$ to zone $j$ can be approximated by the average of the waiting time along different transit routes connecting zone $i$ and zone $j$:
\begin{equation} \label{waiting_time_PT}
w_{ij}^p = \frac{1}{|\mathcal{R}_{ij}|} \sum_{r\in \mathcal{R}_{ij}} \sum_{l \in r} \frac{1}{f_l} ,
\end{equation}
where $\mathcal{R}_{ij}$ is the set of transit routes connecting zone $i$ and zone $j$ due to multiplicity of transit stations in the same zone; $|\mathcal{R}_{ij}|$ is the total number of transit routes from zone $i$ to zone $j$; and $f_l$ is the service frequency of transit line $l$.
\begin{remark}
We assume that origins/destinations of passengers, idle vehicles, and transit stations are uniformly distributed in each zone when deriving (\ref{waiting_time_AMoD}) and (\ref{waiting_time_PT}). However, for first-mile/last-mile AMoD services, the spatial distribution of waiting passengers and idle vehicles is affected by the locations of transit stations. We argue that in practice, each zone is usually traversed by multiple transit lines from distinct directions. The spatial distribution of transit stations along these transit lines can be reasonably approximated by the uniform distribution. In this case, passengers either waiting at transit stations or their origins and idle vehicles can be approximately regarded as uniformly distributed across the zone. However, we conjecture that even if they are not uniformly distributed, (\ref{waiting_time_AMoD}) should still hold with a different value of $A_i$ and (\ref{waiting_time_PT}) can be still approximated under differentiated weights of the waiting time of distinct transit routes. We leave it as future work to validate this conjecture.
\end{remark}
\subsection{Incentives of the TNC Platform}
Consider a TNC platform that hires a fleet of $N$ autonomous vehicles to provide mobility-on-demand services. Each TNC vehicle is in one of the following statuses: (a) cruising on the street and waiting for the passenger; (b) on the way to pickup a passenger; and (c) carrying a passenger. At the stationary state, the total number of vehicle hours $N$ should satisfy the following conservation law:
\begin{equation} \label{vehicle_hour_conservation}
\begin{split}
N = \sum_{i=1}^M N_i^I + & \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \left( \lambda_{ij,k}^a w_i^a + \lambda_{ij,k}^{b_1} w_i^a + \lambda_{ij,k}^{b_2} w_j^a + \lambda_{ij,k}^{b_3} \left(w_i^a + w_j^a\right)\right) + \\ & \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \left(\lambda_{ij,k}^a \frac{l_{ij}}{v_a} + \lambda_{ij,k}^{b_1} \frac{d_i}{v_a} + \lambda_{ij,k}^{b_2} \frac{d_j}{v_a} + \lambda_{ij,k}^{b_3} \left(\frac{d_i}{v_a} + \frac{d_j}{v_a}\right)\right) .
\end{split}
\end{equation}
Based on Little's law, the first term in the right-hand side of (\ref{vehicle_hour_conservation}) represents the operating hours of idle vehicles cruising on the street; the second term accounts for the operating hours of vehicles that are on the way to pick up passengers; and the third term sums the operating hours of vehicles that are occupied with a passenger. The total operating hours of picking up vehicles and occupied vehicles can be further decomposed into operating hours of vehicles that serve four different types of AMoD services (e.g., direct service, first-mile service, last-mile service, and first-mile and last-mile service), respectively.
The AMoD platform determines the base fare $b$, the per-distance rates $r_i^a$, the spatial distribution of idle AVs $N_i^I$, and the fleet size $N$ to maximize its profit subject to the equilibrium conditions. The profit maximization for the TNC platform can be formulated as follows:\footnote{We acknowledge that since our primary goal is to characterize competition between distinct modes, for simplicity, the repositioning of idle TNC vehicles is not explicitly considered in this model. We leave it as future work. }
\begin{equation} \label{Incentives_AMoD}
\begin{split}
\max_{b, \mathbf{r^{a}},\mathbf{N^I}, N} \quad & \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \lambda_{ij,k}^a \left(b+r_i^a l_{ij}^a\right) + \lambda_{ij,k}^{b_1}\left(b+r_i^a d_i\right) + \lambda_{ij,k}^{b_2} \left(b+r_j^a d_j\right) + \lambda_{ij,k}^{b_3} \left(2b + r_i^a d_i + r_j^a d_j\right) - N C_{av} \\
\mathrm{s.t.} \quad & (\ref{logit_demand_function}), (\ref{generalized_travel_cost}), (\ref{waiting_time_AMoD}), (\ref{vehicle_hour_conservation}) \\
\end{split}
\end{equation}
where $C_{av}$ is the hourly operating cost of an AV. The objective in (\ref{Incentives_AMoD}) defines the platform profit as the difference between the revenue and the total operating cost of AVs. The total revenue is the ride fares collected from passengers choosing four distinct types of AMoD services. The platform decisions are subject to the passenger demand model (\ref{logit_demand_function}) and (\ref{generalized_travel_cost}), the waiting time function (\ref{waiting_time_AMoD}), and vehicle conservation (\ref{vehicle_hour_conservation}).
\subsection{Incentives of the Public Transit Agency}
Consider a public transit agency that operates the transit system. Generally, a maximum per-distance fare and maximum acceptable waiting time are imposed by the regulatory agency for transit services connecting distinct areas of the city to guarantee an acceptable service cost and service quality of public transit \cite{SanFrancisco2022Muni}. Let $r_{max}^p$ and $w_{max}^p$ be the maximum allowable per-distance fare and waiting time for transit services, respectively, then we have the following upper bounds on $r^p$ and $w_{ij}^p$:
\begin{equation} \label{waiting_time_constraint}
\begin{cases}
r^p \leq r_{max}^p \\
w_{ij}^p \leq w_{max}^p \quad \forall i,j=1,\dots,M
\end{cases}.
\end{equation}
In the meanwhile, a profit constraint is imposed to guarantee the economic sustainability of the transit agency, which requires that the difference between the revenue and the operating cost be larger than a reservation level $\pi_0$:
\begin{equation} \label{profit_constraint}
\sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K r^p l_{ij}^p (\lambda_{ij,k}^p + \lambda_{ij,k}^{b_1} + \lambda_{ij,k}^{b_2} + \lambda_{ij,k}^{b_3}) - \sum_{l=1}^L f_l C_l \geq \pi_0 ,
\end{equation}
where $C_l$ is the per-vehicle hourly operating cost of transit line $l$.
The public transit agency determines the transit fare $r^p$, and the service frequencies $f_l$ to maximize the public transit ridership subject to the passenger demand model (\ref{logit_demand_function}) and (\ref{generalized_travel_cost}), the waiting time function (\ref{waiting_time_PT}), the transit fare and waiting time constraints (\ref{waiting_time_constraint}), and the profit constraint (\ref{profit_constraint}). The ridership maximization for public transit can be cast as:
\begin{equation} \label{Incentives_PT}
\begin{aligned}
\max_{r^p,\mathbf{f}} \quad & \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \lambda_{ij,k}^p + \lambda_{ij,k}^{b_1} + \lambda_{ij,k}^{b_2} + \lambda_{ij,k}^{b_3} \\
\mathrm{s.t.} \quad & (\ref{logit_demand_function}), (\ref{generalized_travel_cost}), (\ref{waiting_time_PT}), (\ref{waiting_time_constraint}), (\ref{profit_constraint})
\end{aligned}
\end{equation}
\begin{remark}
We postulate that the public transit agency aims to maximize the total ridership. This is consistent with the nature of public transport agencies (e.g. the government) aiming to provide reliable and sustainable mobility services to minimize negative externalities of vehicle travel (e.g. emissions and congestion). Although a large number of existing studies model the objective of transit agencies as (1) user benefit maximization (or equivalently, user cost minimization) \cite{hasselstrom1982public,ouyang2014continuum}, (2) total welfare maximization \cite{chien2002optimization,zhu2021competition}, (3) waiting time minimization \cite{chakroborty2003genetic,verbas2015integrated}, and (4) total cost minimization \cite{fan2008tabu,daganzo2010structure}, etc, we argue that ridership maximization \cite{yoo2010frequency,verbas2013optimal} is a good proxy for the maximization of social welfare or user benefit under a fixed budget. Maximizing total social welfare or passengers' surplus would complicate the formulation without providing extra insights.
\end{remark}
\subsection{Equilibrium of the Game}
The profit maximization for the TNC platform (\ref{Incentives_AMoD}) and the ridership maximization for the public transit (\ref{Incentives_PT}) constitute the game problem. Note that although the operational decision of the TNC platform $(b,\mathbf{r^a},\mathbf{N^I})$ and the public transit agency $(r^p,\mathbf{f})$ does not explicitly appear in the objective of the other, they interact with each other by affecting passengers' mode choices in the passenger demand model (\ref{logit_demand_function}) and (\ref{generalized_travel_cost}). Therefore, the platform profit of the TNC and the ridership of public transit not only depend on its own operational decisions, but also on the operational strategy of the counterpart. We assume that the TNC platform and the public transit agency interact simultaneously and they have perfect information on each other's strategy. The game is at equilibrium if the TNC platform cannot increase its profit and the public transit agency cannot increase the transit ridership by unilaterally changing their operational strategies. Without loss of generality, let $\xi^{a}=(b,\mathbf{r^a},\mathbf{N^I})$ and $\xi^{p}=(r^p,\mathbf{f})$ be the operational strategy of the TNC platform and the public transit agency, respectively, and denote $\pi^a(\xi^a,\xi^p)$ and $\pi^p(\xi^a,\xi^p)$ as the TNC platform profit and the transit ridership, respectively. We can formally define the equilibria $\xi^*=\left(\xi^{a^*},\xi^{p^*}\right)$ that satisfies:
\begin{equation} \label{equilibrium_condition}
\begin{cases}
\pi^a(\xi^{a^*},\xi^{p^*}) \geq \pi^a(\xi^a,\xi^{p^*}) \\
\pi^p(\xi^{a^*},\xi^{p^*}) \geq \pi^p(\xi^{a^*},\xi^p)
\end{cases}, \quad \forall \xi^a, \xi^p \geq 0 .
\end{equation}
which is consistent with the general definition of Nash equilibrium.
In practice, each player in the game may be indifferent to a small change in its objective function (e.g., the change only accounts for a negligible portion of the objective value). This motivates a more relaxed solution concept that encapsulates a broader range of equilibrium solutions, where players will not unilaterally change their strategies as far as the current solution is approximately optimal. More rigorously, we can define such relaxed solution as the $\epsilon$-Nash equilibrium \cite{daskalakis2006note, li2019connections}, denoted as $\xi^*=\left(\xi^{a^*},\xi^{p^*}\right)$, which satisfies:
\begin{equation} \label{equilibrium_condition}
\begin{cases}
\pi^a(\xi^{a^*},\xi^{p^*}) \geq \pi^a(\xi^a,\xi^{p^*})-\epsilon \\
\pi^p(\xi^{a^*},\xi^{p^*}) \geq \pi^p(\xi^{a^*},\xi^p)-\epsilon
\end{cases}, \quad \forall \xi^a, \xi^p \geq 0 .
\end{equation}
At an $\epsilon$-Nash equilibrium, each player can improve its objective value by at most $\epsilon$ via deviating from the equilibrium strategy $\xi^*=\left(\xi^{a^*},\xi^{p^*}\right)$, given that the other player follows the equilibrium strategy. Therefore, it indicates that the players are motivated to play the equilibrium strategy if they are indifferent to a change of $\epsilon$. Note that Nash equilibrium can be viewed as the special case of the $\epsilon$-Nash equilibrium, correpsonding to the case of $\epsilon=0$.
\section{The Solution Method} \label{solution method}
This section solves the game problem by computing the Nash equilibrium of the game using the best response method and conducting an ex-post evaluation of the performance of the obtained solution. Specifically, we will first derive a candidate Nash equilibrium solution by the best response algorithm, which iteratively solves the profit maximization problem of the TNC platform and the ridership maximization problem of the transit agency using standard interior point methods until convergence. We comment that this step is not difficult because both the best response algorithm and interior point methods are well-established approaches. However, since each of these subproblems is non-convex, the derived candidate solution is only locally optimal, and it is quite challenging to validate how good the derived solution is compared to the globally optimal solution for each decision maker. To address this concern, we treat each subproblem separately: (a) we first focus on the profit-maximization problem of the TNC platform, and relax the consistency of its decisions to derive a tight upper bound for characterizing the optimality gap of the derived solution; (b) we then reformulate the ridership maximization problem and derive the conditions under which the reformulated problem is concave and thus its global optimal solution can be efficiently computed. Overall, the proposed methods enable us to conduct an ex-post evaluation of the performance of the obtained solution by showing that the derived solution is at least as good as an $\epsilon$-Nash equilibrium, where the value of $\epsilon$ can be derived numerically. Below we investigate the profit-maximization problem for the TNC platform and the ridership-maximization problem for the transit agency to develop a numerical framework to compute $\epsilon$.
\subsection{Profit Maximization of the TNC Platform}\label{sectionTNC}
In the profit maximization problem (\ref{Incentives_AMoD}), the involved constraints are all equality constraints. Note that given $\xi^p$, (\ref{logit_demand_function}), (\ref{generalized_travel_cost}) and (\ref{waiting_time_AMoD}) jointly determine $\lambda_{ij,k}^t,t\in\left\{a,b_1,b_2,b_3\right\}$ as a function of $b$, $r_i^a$, $r_j^a$, $N_i^I$, and $N_j^I$. With a slight abuse of notation, we denote it as $\lambda_{ij,k}^t=\lambda_{ij,k}^t\left(b,r_i^a,r_j^a,N_i^I,N_j^I\right), t\in\{a,b_1,b_2,b_3\}$. By further substituting (\ref{vehicle_hour_conservation}) into the objective function, the original problem can be transformed into the following unconstrained optimization:
\begin{equation} \label{Incentives_AMoD_unconstrained}
\begin{split}
\max_{b, \mathbf{r^{a}},\mathbf{N^I}} \quad & \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \lambda_{ij,k}^a\left(b,r_i^a,r_j^a,N_i^I,N_j^I\right) \left[b+\left(r_i^a-\frac{C_{av}}{v_a}\right) l_{ij}^a - C_{av} \cdot w_i^a\left(N_i^I\right)\right] \\ + & \lambda_{ij,k}^{b_1}\left(b,r_i^a,r_j^a,N_i^I,N_j^I\right)\left[b+\left(r_i^a-\frac{C_{av}}{v_a}\right) d_i - C_{av} \cdot w_i^a\left(N_i^I\right)\right] \\ + & \lambda_{ij,k}^{b_2}\left(b,r_i^a,r_j^a,N_i^I,N_j^I\right) \left[b+\left(r_j^a-\frac{C_{av}}{v_a}\right) d_j - C_{av} \cdot w_j^a\left(N_j^I\right)\right] \\ + & \lambda_{ij,k}^{b_3}\left(b,r_i^a,r_j^a,N_i^I,N_j^I\right) \left[2b + \left(r_i^a-\frac{C_{av}}{v_a}\right) d_i + \left(r_j^a-\frac{C_{av}}{v_a}\right) d_j - C_{av} \cdot w_i^a\left(N_i^I\right) - C_{av} \cdot w_j^a\left(N_j^I\right)\right] \\ - & \frac{C_{av}}{M} \sum_{i=1}^M \sum_{j=1}^M N_i^I ,
\end{split}
\end{equation}
where we equivalently rewrite $C_{av}\sum_{i=1}^M N_i^I$ as $\frac{C_{av}}{M}\sum_{i=1}^M \sum_{j=1}^M N_i^I$.
Based on (\ref{Incentives_AMoD_unconstrained}), we introduce auxiliary variables $b_j$, $r_{ij}^a$, and $N_{ij}^a$, where $i,j=1,\dots,M$ to formulate the equivalent problem. The intuition behind the auxiliary variables is that the platform determines the base fare $b_j$, the ride fares $r_{ij}^a,i=1,\dots,M$, and the number of idle vehicles $N_{ij}^I,i=1,\dots,M$ for the specific destination zone $j$. Without loss of generality, let $\xi_j^a=\left(b_j,r_{1j}^a,N_{1j}^I,r_{2j}^a,N_{2j}^I,\dots,r_{Mj}^a,N_{Mj}^I\right)$ be the operational strategy exclusively for destination zone $j$, and define $\pi_j^a$ as the corresponding profit from destination zone $j$:
\begin{equation}
\begin{split}
& \pi_j^a\left(\xi_j^a\right) = \sum_{i=1}^M \sum_{k=1}^K \lambda_{ij,k}^a\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right) \left[b_j+\left(r_{ij}^a-\frac{C_{av}}{v_a}\right) l_{ij}^a - C_{av} \cdot w_i^a\left(N_{ij}^I\right)\right] \\ + & \lambda_{ij,k}^{b_1}\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right)\left[b_j+\left(r_{ij}^a-\frac{C_{av}}{v_a}\right) d_i - C_{av} \cdot w_i^a\left(N_{ij}^I\right)\right] \\ + &\lambda_{ij,k}^{b_2}\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right) \left[b_j+\left(r_{jj}^a-\frac{C_{av}}{v_a}\right) d_j - C_{av} \cdot w_j^a\left(N_{jj}^I\right)\right] \\ + &\lambda_{ij,k}^{b_3}\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right) \left[2b_j + \left(r_{ij}^a-\frac{C_{av}}{v_a}\right) d_i + \left(r_{jj}^a-\frac{C_{av}}{v_a}\right) d_j - C_{av} \cdot w_i^a\left(N_{ij}^I\right) - C_{av} \cdot w_j^a\left(N_{jj}^I\right)\right] \\ - & \frac{C_{av}}{M} \sum_{i=1}^M N_{ij}^I, \quad j=1,\dots,M.
\end{split}
\end{equation}
Note that the platform actually deploys a uniform operational strategy for distinct destination zones, therefore the destination-specific operational strategy $\xi_j^a$ should be consistent across different destination zones. In this case, the original problem (\ref{Incentives_AMoD_unconstrained}) is equivalent to the following optimization:
\begin{equation} \label{Incentives_AMoD_equivalent}
\begin{split}
\max_{\xi_1^a,\xi_2^a,\dots,\xi_M^a} \quad & \sum_{j=1}^M \pi_j^a(\xi_j^a) \\
\mathrm{s.t.} \quad & \xi_1^a = \xi_2^a = \dots =\xi_M^a
\end{split}
\end{equation}
Based on the equivalent formulation (\ref{Incentives_AMoD_equivalent}), we will slightly modify the decision consistency constraint $\xi_1^a=\xi_2^a=\dots=\xi_M^a$ to derive an upper bound. In particular, let $\overline{\mathcal{V}}=\left\{\mathcal{V}_1,\mathcal{V}_2,\dots,\mathcal{V}_m\right\}$ be a partition of $\mathcal{V}$, i.e., $\bigcup_{x=1,\dots,m} \mathcal{V}_x = \mathcal{V}$ and $\mathcal{V}_x \cap \mathcal{V}_y = \emptyset, \forall \mathcal{V}_x,\mathcal{V}_y\in \overline{\mathcal{V}}, x \neq y$, and define the relaxed problem as
\begin{equation} \label{Incentives_AMoD_relaxed}
\begin{split}
\max_{\xi_1^a,\xi_2^a,\dots,\xi_M^a} \quad & \sum_{j=1}^M \pi_j^a(\xi_j^a) \\
\mathrm{s.t.} \quad & \xi_j^a = \xi_{j'}^a \quad \forall j,j' \in \mathcal{V}_1 \\
& \xi_j^a = \xi_{j'}^a \quad \forall j,j' \in \mathcal{V}_2 \\
& \dots \\
& \xi_j^a = \xi_{j'}^a \quad \forall j,j' \in \mathcal{V}_m
\end{split} ,
\end{equation}
where we break the complete decision consistency and only require that the operational strategies are consistent across a subset of destination zones.
Note that (\ref{Incentives_AMoD_relaxed}) can be further decomposed over the partition $\overline{\mathcal{V}}$:
\begin{equation} \label{Incentives_AMoD_relaxed_subproblem}
\begin{split}
\max_{\xi_j^a,j \in \mathcal{V}_x} \quad & \sum_{j \in \mathcal{V}_x} \pi_j^a(\xi_j^a) \\
\mathrm{s.t.} \quad & \xi_j^a = \xi_{j'}^a \quad \forall j,j' \in \mathcal{V}_x
\end{split} ,
\quad x=1,\dots,m .
\end{equation}
The optimal solution to (\ref{Incentives_AMoD_relaxed_subproblem}), if obtained, provides an upper bound for the optimal value of the original problem, thus we have the following result:
\begin{proposition} \label{proposition_upper_lower_bound}
Suppose that $\xi^*=\left(\xi^{a^*},\xi^{p^*}\right)$ is a candidate Nash equilibrium. Let $\xi_j^{a^*},j\in\mathcal{V}_x$ be the optimal solution to (\ref{Incentives_AMoD_relaxed_subproblem}) given $\xi^{p^*}$, and let $\pi_{\mathcal{V}_x}^{a^*}$ be the corresponding optimal value. We have $\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*} \geq \pi^a\left(\xi^{a^*},\xi^{p^*}\right) \geq \pi^a \left(\frac{\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*} \xi_{j,j \in \mathcal{V}_x}^{a^*}}{\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*}}, \xi^{p^*} \right)$ .
\end{proposition}
\begin{proof}
(1) $\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*}$ is the optimal value of the relaxed problem (\ref{Incentives_AMoD_relaxed}) given $\xi^{p^*}$. Since (\ref{Incentives_AMoD_relaxed}) is a relaxed problem of (\ref{Incentives_AMoD_equivalent}), and $\xi^{a^*}$ is the optimal solution to (\ref{Incentives_AMoD_equivalent}) given $\xi^{p^*}$, then we have $\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*} \geq \sum_{j=1}^M \pi_j^a \left(\xi^{a^*}\right) = \pi^a\left(\xi^{a^*},\xi^{p^*}\right)$.
(2) Since $\xi_{j,j\in\mathcal{V}_x}^{a^*}, \forall x=1,\dots,m$, is a feasible solution to (\ref{Incentives_AMoD_equivalent}), a non-negative linear combination of $\xi_{j,j\in\mathcal{V}_x}^{a^*},x=1,\dots,m$ is also a feasible solution to (\ref{Incentives_AMoD_equivalent}). Therefore, $\sum_{j=1}^M \pi_j^a \left(\frac{\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*} \xi_{j,j \in \mathcal{V}_x}^{a^*}}{\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*}}\right) \leq \sum_{j=1}^M \pi_j^a \left(\xi^{a^*}\right) = \pi^a\left(\xi^{a^*},\xi^{p^*}\right) $.
\end{proof}
Proposition \ref{proposition_upper_lower_bound} establishes an upper bound and a lower bound for the platform profit at Nash equilibrium. The intuition behind the upper bound is that the relaxation of the consistency of platform decisions provides the platform a higher degree of freedom to manage the AMoD service and thus leads to higher profitability. In the meanwhile, we use a weighted sum of optimal solutions to (\ref{Incentives_AMoD_relaxed_subproblem}) to generate a feasible solution and a tight lower bound of the original problem (\ref{Incentives_AMoD_equivalent}). These weights are determined based on the intuition that the platform combines distinct operational strategies $\xi_{j,j\in\mathcal{V}_x}^{a^*}$ for distinct sets of destination zones $\mathcal{V}_x$ to generate a uniform strategy, and the operational strategy on the set of destination zones with a higher potential profit $\pi_{\mathcal{V}_x}^{a^*}$ is assigned with larger weight.
The upper bound derived by Proposition \ref{proposition_upper_lower_bound} is very useful for us because whenever a local solution to (\ref{Incentives_AMoD_equivalent}) is given, we can compare the optimal value with respect to the upper bound, which offers an upper bound on the distance between the optimal value of the local solution and the optimal value of the unknown globally optimal solution. This enables us to assert how good a solution is compared to the globally optimal one, and it will be an important intermediate step for assessing the overall quality of the Nash equilibrium. However, to use this result, we do need to numerically compute the globally optimal solution to the nonconvex program (\ref{Incentives_AMoD_relaxed_subproblem}). Fortunately, despite its non-convexity, (\ref{Incentives_AMoD_relaxed_subproblem}) has a decomposable structure, which can be addressed by primal decomposition. To this end, let $\xi_{ij}^a=\left(r_{ij}^a,N_{ij}^I\right)$ be the operational decision on the specific OD pair $ij$ and define $\pi_{ij}^a$ as the corresponding profit from OD pair $ij$:
\begin{equation}
\begin{split}
& \pi_{ij}^a\left(b_j,\xi_{jj}^a,\xi_{ij}^a\right) = \sum_{k=1}^K \lambda_{ij,k}^a\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right) \left[b_j+\left(r_{ij}^a-\frac{C_{av}}{v_a}\right) l_{ij}^a - C_{av} \cdot w_i^a\left(N_{ij}^I\right)\right] \\ + & \lambda_{ij,k}^{b_1}\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right)\left[b_j+\left(r_{ij}^a-\frac{C_{av}}{v_a}\right) d_i - C_{av} \cdot w_i^a\left(N_{ij}^I\right)\right] \\ + &\lambda_{ij,k}^{b_2}\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right) \left[b_j+\left(r_{jj}^a-\frac{C_{av}}{v_a}\right) d_j - C_{av} \cdot w_j^a\left(N_{jj}^I\right)\right] \\ + &\lambda_{ij,k}^{b_3}\left(b_j,r_{ij}^a,r_{jj}^a,N_{ij}^I,N_{jj}^I\right) \left[2b_j + \left(r_{ij}^a-\frac{C_{av}}{v_a}\right) d_i + \left(r_{jj}^a-\frac{C_{av}}{v_a}\right) d_j - C_{av} \cdot w_i^a\left(N_{ij}^I\right) - C_{av} \cdot w_j^a\left(N_{jj}^I\right)\right] \\ - & \frac{C_{av}}{M} N_{ij}^I, \quad i,j=1,\dots,M.
\end{split}
\end{equation}
Note that $\xi_j^a = \left(b_j,\xi_{1j}^a,\xi_{2j}^a,\dots,\xi_{Mj}^a\right)$ and $\pi_{j}^a\left(\xi_j^a\right)=\sum_{i=1}^M \pi_{ij}^a\left(b_j,\xi_{jj}^a,\xi_{ij}^a\right)$. Therefore, the optimization (\ref{Incentives_AMoD_relaxed_subproblem}) is equivalent to:
\begin{equation} \label{Incentives_AMoD_relaxed_equivalent}
\begin{split}
\max_{b_j,\xi_{ij}^a,j \in \mathcal{V}_x,i=1,\dots,M} \quad & \sum_{j \in \mathcal{V}_x} \sum_{i=1}^M \pi_{ij}^a (b_j,\xi_{jj}^a,\xi_{ij}^a) \\
\mathrm{s.t.} \quad & b_j = b_{j'} \quad \forall j,j' \in \mathcal{V}_x \\
& \xi_{ij}^a = \xi_{ij'}^a \quad \forall j,j' \in \mathcal{V}_x, i=1,\dots,M
\end{split} ,
\quad x = 1,\dots,m .
\end{equation}
For given $b_j,\xi_{jj}^a,j\in \mathcal{V}_x$, we define the subproblems:
\begin{equation} \label{Incentives_AMoD_primal_subproblem}
\begin{split}
\max_{\xi_{ij}^a, j \in \mathcal{V}_x} \quad & \sum_{j \in \mathcal{V}_x} \pi_{ij}^a(b_j,\xi_{jj}^a,\xi_{ij}^a) \\
\mathrm{s.t.} \quad & \xi_{ij}^a = \xi_{ij'}^a \quad \forall j,j' \in \mathcal{V}_x
\end{split} ,
\quad i = 1,\dots,M .
\end{equation}
with optimal values $\pi_{i,\mathcal{V}_x}^{a^*},i=1,\dots,M$. In this case, $\pi_{i,\mathcal{V}_x}^{a^*},i=1,\dots,M$ is a function of $b_j,\xi_{jj}^a,j\in \mathcal{V}_x$, and (\ref{Incentives_AMoD_relaxed_equivalent}) is equivalent to the master problem:
\begin{equation} \label{Incentives_AMoD_primal_master}
\begin{split}
\max_{b_j,\xi_{jj}^a,j\in\mathcal{V}_x} \quad & \sum_{i=1}^M \pi_{i,\mathcal{V}_x}^{a^*} \\
\mathrm{s.t.} \quad & b_j = b_{j'} \quad \forall j,j' \in \mathcal{V}_x
\end{split} .
\end{equation}
Note that (\ref{Incentives_AMoD_primal_subproblem}) only involves two decision variables $\xi_{ij}^a=\left(r_{ij}^a,N_{ij}^I\right)$ if we absorb the equality constraint into the objective. Similarly, (\ref{Incentives_AMoD_primal_master}) has $1+2\cdot |\mathcal{V}_x|$ decision variables if we absorb the equality constraint, where $|\mathcal{V}_x|$ is the number of zones in $\mathcal{V}_x$. Since both (\ref{Incentives_AMoD_primal_subproblem}) and (\ref{Incentives_AMoD_primal_master}) are small-scale problems, we can use grid search method to find the globally optimal solutions in parallel.
\begin{remark}
The partition $\bar{\mathcal{V}}$ is flexible and we can carefully choose the composition of $\bar{\mathcal{V}}$ to address the trade-off between computational complexity and the quality of the upper bound. In particular, if we partition the space $\mathcal{V}$ into a large number of smaller sets, then the decomposed problem (\ref{Incentives_AMoD_primal_master}) has a smaller dimension, which can be more efficiently solved. However, this induces more relaxation for the consistency of decisions, which leads to a worse upper bound. Our numerical simulation indicates that dividing all the zones $\mathcal{V}$ into pairwise groups (i.e., $|\mathcal{V}_x|=2$) will lead to fast computation and tight upper bound at the same time. More details will be reported in case studies.
\end{remark}
\subsection{Ridership Maximization of Transit Agency} \label{sectiontransit}
We next discuss how to solve the globally optimal solution to the ridership maximization problem for the public transit agency. Note that given $\xi^a$, (\ref{logit_demand_function}) and (\ref{generalized_travel_cost}) jointly determine $\lambda_{ij,k}^t,t\in\left\{p,b_1,b_2,b_3\right\}$ as a function of $r^p$ and $w_{ij}^p$. With a slight abuse of notation, we denote it as $\lambda_{ij,k}^t=\lambda_{ij,k}^t\left(r^p,w_{ij}^p\right),t\in\left\{p,b_1,b_2,b_3\right\}$. The ridership maximization problem (\ref{Incentives_PT}) is equivalent to
\begin{equation} \label{Incentives_PT_equivalent}
\begin{split}
\max_{r^p,\mathbf{f},\mathbf{w^p}} \quad & \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \lambda_{ij,k}^p \left(r^p,w_{ij}^p\right) + \lambda_{ij,k}^{b_1} \left(r^p,w_{ij}^p \right) + \lambda_{ij,k}^{b_2} \left(r^p,w_{ij}^p\right) + \lambda_{ij,k}^{b_3} \left(r^p,w_{ij}^p \right) \\
\mathrm{s.t.} \quad & w_{ij}^p = \Phi_{ij}^T \left[\frac{1}{f_1},\frac{1}{f_2},\dots,\frac{1}{f_L}\right]^T \quad \forall i,j=1,\dots,M \\
& r^p \leq r_{max}^p, \quad w_{ij}^p \leq w_{max}^p \quad \forall i,j=1,\dots,M \\
& \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K r^p l_{ij}^p \left[\lambda_{ij,k}^p \left(r^p,w_{ij}^p\right) + \lambda_{ij,k}^{b_1} \left(r^p,w_{ij}^p\right) + \lambda_{ij,k}^{b_2} \left(r^p,w_{ij}^p\right) + \lambda_{ij,k}^{b_3} \left(r^p,w_{ij}^p\right)\right] - \sum_{l=1}^L C_l f_l \geq \pi_0
\end{split}
\end{equation}
where $\mathbf{w^p}=(w_{11}^p,w_{12}^p,\dots,w_{MM}^p)^T$ is the vector of average waiting times for public transit of different OD pairs and $\Phi_{ij}$ is the vector that relates $w_{ij}^p$ to $\mathbf{f}$ based on (\ref{waiting_time_PT}). Specifically, $\Phi_{ij}$ is a $L$-dimensional vector in which each element accounts for the number of occurrences of a specific transit line in the set of transit routes $\mathcal{R}_{ij}$ connecting zone $i$ and zone $j$ divided by the total number of transit routes $|\mathcal{R}_{ij}|$. We note that the key challenge for (\ref{Incentives_PT_equivalent}) is that both the objective function and constraints are non-convex. To address this challenge, we first introduce a change of variable and reformulate (\ref{Incentives_PT_equivalent}) into an equivalent form as follows:
\begin{lemma} \label{lemma_reformulation}
Define $N_{ij}^p = \frac{1}{w_{ij}^p}$. The optimization (\ref{Incentives_PT_equivalent}) can be equivalently reformulated as:
\begin{equation} \label{Incentives_PT_reformulation}
\begin{aligned}
\max_{r^p,\mathbf{f},\mathbf{N^p}} \quad & \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \lambda_{ij,k}^p \left(r^p,\frac{1}{N_{ij}^p}\right) + \lambda_{ij,k}^{b_1} (r^p,\frac{1}{N_{ij}^p} ) + \lambda_{ij,k}^{b_2} \left(r^p,\frac{1}{N_{ij}^p}\right) + \lambda_{ij,k}^{b_3} \left(r^p,\frac{1}{N_{ij}^p}\right) \\
\mathrm{s.t.} \quad & N_{ij}^p \leq \dfrac{1}{\Phi_{ij}^T \left[\dfrac{1}{f_1},\ldots,\dfrac{1}{f_L}\right]^T} \quad \forall i,j=1,\dots,M \\
& r^p \leq r_{max}^p, \quad N_{ij}^p \geq \frac{1}{w_{max}^p} \quad \forall i,j=1,\dots,M \\
& \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K r^p l_{ij}^p \left[\lambda_{ij,k}^p \left(r^p,\frac{1}{N_{ij}^p}\right) + \lambda_{ij,k}^{b_1} \left(r^p,\frac{1}{N_{ij}^p}\right) + \lambda_{ij,k}^{b_2} \left(r^p,\frac{1}{N_{ij}^p}\right) + \lambda_{ij,k}^{b_3} \left(r^p,\frac{1}{N_{ij}^p}\right)\right] - \sum_{l=1}^L C_lf_l \geq \pi_0
\end{aligned}
\end{equation}
\end{lemma}
The proof of Lemma \ref{lemma_reformulation} is trivial since it only involves a change of variable. Intuitively, $N_{ij}^p$ is virtually similar to the number of idle vehicles $N_{ij}^I$ in TNC's AMoD services. Equivalent to adjusting the average waiting time $w_{ij}^p$ of distinct OD pairs, the public transit agency determines the number of vehicles $N_{ij}^p$ for distinct OD pairs to maximize the ridership. Note that the reformulated problem (\ref{Incentives_PT_reformulation}) is non-convex, which is difficult to address. However, we find that under certain mild conditions, if we fix $r_p$, then both the objective function and constraints of (\ref{Incentives_PT_equivalent}) can be concave with respect to $\mathbf{f}$ and $\mathbf{N^p}$. To make this more clear, we first introduce the following proposition as a preliminary result:
\begin{proposition} \label{proposition_concavity}
Given $\xi^a$ and $r^p$, (\ref{Incentives_PT_reformulation}) is concave with respect to $\mathbf{f}$ and $\mathbf{N^p}$ if
\begin{equation} \label{condition_concavity}
N_{ij}^p \geq \frac{\epsilon}{2} \frac{\sum_{k=1}^K \alpha_k^2 \lambda_{ij,k}^0 P_{ij,k} \left(1-P_{ij,k}\right) \left(1-2P_{ij,k}\right)}{\sum_{k=1}^K \alpha_k \lambda_{ij,k}^0 P_{ij,k} \left(1-P_{ij,k}\right)}, \quad \forall i,j=1,\dots,M,
\end{equation}
where $P_{ij,k}$ is the percentage of passengers choosing public transit ($p$) and bundled services ($b_1$, $b_2$ and $b_3$) from origin $i$ to destination $j$ in income class $k$.
\end{proposition}
The proof of Proposition \ref{proposition_concavity} is deferred to Appendix A. Proposition \ref{proposition_concavity} provides an important condition (\ref{condition_concavity}) under which the concavity of (\ref{Incentives_PT_reformulation}) with respect to $\mathbf{f}$ and $\mathbf{N^p}$ can be established. However, both sides of (\ref{condition_concavity}) involve decision/endogenous variables, which cannot be used directly without knowing the values of these variables. To address this concern, we will simplify (\ref{condition_concavity}) by reformulating it as a constraint on $N_{ij}^p$. Note that (\ref{condition_concavity}) indeed defines a constraint on $N_{ij}^p$, which is determined by the intersection of its left-hand side and the right-hand side. The left-hand side of (\ref{condition_concavity}) is a linear function of $N_{ij}^p$, whereas the right-hand side of (\ref{condition_concavity}) is a function of $r^p$ and $N_{ij}^p$. For notation convenience, we denote the right-hand side of (\ref{condition_concavity}) as $RHS(r^p,N_{ij}^p)$, and examine the two boundaries of $RHS(r^p,N_{ij}^p)$ when $N_{ij}^p$ approaches $0$ or $\infty$.
In particular, for $r^p \geq 0$, based on the definition of generalized travel cost (\ref{generalized_travel_cost}), we can write $c_{ij,k}^t$ as functions of $r^p$ and $N_{ij}^p$:
\begin{equation} \label{cost_given_AMoD_1}
\begin{cases}
c_{ij,k}^a = \text{const}_{ij,k}^a \\
c_{ij,k}^p = \gamma_k l_{ij}^p r^p + \frac{\alpha_k}{N_{ij}^p} + \text{const}_{ij,k}^p \\
c_{ij,k}^{b_1} = \gamma_k l_{ij}^p r^p + \frac{\alpha_k}{N_{ij}^p} + \text{const}_{ij,k}^{b_1} \\
c_{ij,k}^{b_2} = \gamma_k l_{ij}^p r^p + \frac{\alpha_k}{N_{ij}^p} + \text{const}_{ij,k}^{b_2} \\
c_{ij,k}^{b_3} = \gamma_k l_{ij}^p r^p + \frac{\alpha_k}{N_{ij}^p} + \text{const}_{ij,k}^{b_3} \\
c_{ij,k}^o = \text{const}_{ij,k}^{o} \\
\end{cases},
\end{equation}
Based on (\ref{cost_given_AMoD_1}) and (\ref{logit_demand_function}), it can be easily verified that $\lim_{N_{ij}^p \to 0} P_{ij,k} = 0$. In this case, by Bernoulli's rule, we have
\begin{equation} \label{RHS_0}
\lim_{N_{ij}^p \to 0} RHS(r^p,N_{ij}^p) = \lim_{P_{ij,k}\to 0} \frac{\epsilon}{2} \dfrac{\sum_{k=1}^K \alpha_k^2 \lambda_{ij,k}^0 (1-6P_{ij,k}+6P_{ij,k}^2)}{\sum_{k=1}^K \alpha_k \lambda_{ij,k}^0 (1-2P_{ij,k})}=\frac{\epsilon}{2} \frac{\sum_{k=1}^K \alpha_k^2 \lambda_{ij,k}^0}{\sum_{k=1}^K \alpha_k \lambda_{ij,k}^0}.
\end{equation}
Similarly, based on (\ref{cost_given_AMoD_1}) and (\ref{logit_demand_function}), when ${N_{ij}^p}$ approaches infinity, we have
\begin{equation} \label{Np_infty}
\lim_{N_{ij}^p \to \infty} P_{ij,k} =\dfrac{\sum_{t\in\{p,b_1,b_2,b_3\}}\exp(-\epsilon \cdot (\gamma_k l_{ij}^p r^p + \text{const}_{ij,k}^t))}{\sum_{t\in\{p,b_1,b_2,b_3\}}\exp(-\epsilon \cdot (\gamma_k l_{ij}^p r^p + \text{const}_{ij,k}^t)) + \sum_{t\in\{a,o\}} \exp(-\epsilon \cdot \text{const}_{ij,k}^t)} .
\end{equation}
The right-hand side of (\ref{Np_infty}) is a constant within the range $(0,1)$ and we denote it as $\hat{P}_{ij,k}\in(0,1)$. Given $\hat{P}_{ij,k}$, we have
\begin{equation}
\label{RHS_infty}
\lim_{N_{ij}^p \to \infty} RHS(r^p,N_{ij}^p) = \frac{\epsilon}{2} \frac{\sum_{k=1}^K \alpha_k^2 \lambda_{ij,k}^0 \hat{P}_{ij,k} \left(1-\hat{P}_{ij,k}\right) \left(1-2\hat{P}_{ij,k}\right)}{\sum_{k=1}^K \alpha_k \lambda_{ij,k}^0 \hat{P}_{ij,k} \left(1-\hat{P}_{ij,k}\right)} .
\end{equation}
Combining (\ref{RHS_0}) and (\ref{RHS_infty}), we conclude that the right-hand side of (\ref{condition_concavity}) is bounded, whereas the left-hand side of (\ref{condition_concavity}) is a linearly increasing function. Therefore, by continuity of $RHS(r^p,N_{ij}^p)$, there must exist a $\hat{N}_{ij}^p$ such that as far as $N_{ij}^p \geq \hat{N}_{ij}^p$, then $N_{ij}^p \geq RHS(r^p,N_{ij}^p)$. For this purpose, $\hat{N}_{ij}^p$ can be defined as the maximum solution to the following fixed-point equation
\begin{equation} \label{fixed_point_equation}
N_{ij}^p = RHS(r^p,N_{ij}^p) = \frac{\epsilon}{2} \frac{\sum_{k=1}^K \alpha_k^2 \lambda_{ij,k}^0 P_{ij,k} \left(1-P_{ij,k}\right) \left(1-2P_{ij,k}\right)}{\sum_{k=1}^K \alpha_k \lambda_{ij,k}^0 P_{ij,k} \left(1-P_{ij,k}\right)} .
\end{equation}
Clearly $\hat{N}_{ij}^p$ depends on $r^p$, and without loss of generality, we denote it as $\hat{N}_{ij}^p (r^p)$. Furthermore, let $\bar{N}_{ij}^p=\max_{0\leq r_p\leq r_{max}^p} \hat{N}_{ij}^p (r^p)$, we are then ready to present the major theorem regarding the concavity of ridership maximization problem (\ref{Incentives_PT_reformulation}):
\begin{theorem} \label{themrem_concavity}
Assume that $\xi^a$ is fixed, the ridership maximization problem (\ref{Incentives_PT_reformulation}) is concave with respect to $\mathbf{f}$ and $\mathbf{N^p}$ for any given $r^p\in\left[ 0,r_{max}^p \right]$, as far as $\bar{N}_{ij}^p\leq \dfrac{1}{w_{max}^p}$ $\forall i,j=1,\dots,M.$
\end{theorem}
Theorem \ref{themrem_concavity} is a direct consequence of Proposition \ref{proposition_concavity} and the aforementioned analysis: since the upper bound on transit waiting time requires that $N_{ij}^p\geq 1/w_{max}^p$, we can conclude that as long as $\bar{N}_{ij}^p\leq 1/w_{max}^p$, then we have $N_{ij}^p\geq \bar{N}_{ij}^p$, which satisfies the concavity conditions (\ref{condition_concavity}). This provides a condition based on which we can determine the concavity of (\ref{Incentives_PT_reformulation}) with respect to $\mathbf{f}$ and $\mathbf{N^p}$ without knowing the value of decision variables or endogenous variables. Our case study suggests that the premise of Theorem~\ref{themrem_concavity} holds under a large range of realistic model parameters.
It is important to note that in Theorem \ref{themrem_concavity}, the concavity only holds when $r^p$ is excluded from the decisions. In fact, the ridership maximization problem (\ref{Incentives_PT_reformulation}) is not jointly concave with respect to $r^p$ and other decision variables. However, Theorem \ref{themrem_concavity} is still very useful to compute the globally optimal solution to (\ref{Incentives_PT_reformulation}) because it enables us to consider a hierarchical algorithm, where we first fix $r^p$ and solve the remaining decision variables which are concave, then we repeat this process for distinct values of $r^p$ to search for the globally optimal solution. In summary, we consider a hierarchical algorithm in the following two steps: first we fix $r^p$ and compute the optimal value $\pi^{p^*}(\xi^a,r^p)$ to (\ref{Incentives_PT_reformulation}) by optimizing other decision variables aside from $r^p$; then we vary $r^p$ and enumerate all possible $r^p$ to find the optimal ${r}^{p^*}$ that maximizes $\pi^{p^*}(\xi^a,r^p)$. Since $r^p$ is only one-dimensional, and the problem is concave with respect to the rest of the decision variables, this algorithm can efficiently find the globally optimal solution to (\ref{Incentives_PT_reformulation}). Let us denote the optimal solution corresponding to ${r}^{p^*}$ as $\xi^{p^*}$, then we have the following corollary:
\begin{corollary} \label{proposition_lower_bound_PT}
Assume that $\xi^a$ is fixed, and suppose $\bar{N}_{ij}^p\leq \dfrac{1}{w_{max}^p}$ $\forall i,j=1,\dots,M$, then $\xi^{p^*}$ is the globally optimal solution to the ridership maximization problem (\ref{Incentives_PT_reformulation}).
\end{corollary}
To summarize, we run the best response algorithm to iteratively compute the profit maximization problem (\ref{Incentives_AMoD}) and the ridership maximization problem (\ref{Incentives_PT}), using the standard interior-point method \cite{mehrotra1992implementation}. After the algorithm converges, we obtain a candidate Nash equilibrium, which might be the local solution to each player's optimization problem. To validate the quality of these candidate solutions, we use the result of Proposition \ref{proposition_upper_lower_bound}, Proposition \ref{proposition_concavity}, and Theorem \ref{themrem_concavity}, so that the profit maximization problem (\ref{Incentives_AMoD}) can be solved approximately with a theoretical upper bound, whereas the ridership maximization problem (\ref{Incentives_PT}) can be solved exactly under certain mild conditions. Details of the algorithm are summarized in Algorithm \ref{algorithm1}.
\begin{algorithm}
\caption{Best response algorithm and the ex-post evaluation on Nash equilibrium} \label{algorithm1}
\begin{algorithmic}[1]
\REQUIRE Initial guess of operational strategy $\xi_{(0)}^a=\left(b,\mathbf{r^a},\mathbf{N^I}\right)$ and $\xi_{(0)}^p=\left(r^p,\mathbf{f}\right)$, the convergence tolerance $\sigma$, the maximum number of iterations $\hat{n}$, a partition $\overline{\mathcal{V}}=\left\{\mathcal{V}_1,\mathcal{V}_2,\dots,\mathcal{V}_m\right\}$ of $\mathcal{V}$.
\STATE Setup stopping criterion: $||\xi_{(n)}^a-\xi_{(n-1)}^a||_2 \leq \sigma$ and $||\xi_{(n)}^p-\xi_{(n-1)}^p||_2 \leq \sigma$
\FOR {$n=1,\dots,\hat{n}$}
\STATE Fix $\xi^p=\xi_{(n-1)}^p$, solve (\ref{Incentives_AMoD}) using interior-point algorithm, and obtain the optimal solution $\xi_{(n)}^a$.
\STATE Fix $\xi^a=\xi_{(n)}^a$, solve (\ref{Incentives_PT}) using interior-point algorithm, and obtain the optimal solution $\xi_{(n)}^p$.
\IF {stopping criterion satisfied}
\STATE Obtain the equilibrium strategy $\xi^{a^*}=\xi_{(n)}^a$ and $\xi^{p^*}=\xi_{(n)}^p$.
\STATE \algorithmicbreak
\ENDIF
\ENDFOR
\FOR {$x=1,\dots,m$}
\STATE Solve (\ref{Incentives_AMoD_relaxed_subproblem}) given $\xi^p=\xi^{p^*}$ using primal decomposition algorithm in Section \ref{sectionTNC}.
\STATE Obtain the optimal solution $\xi_{j,j\in\mathcal{V}_x}^{a^*}$ and the corresponding optimal value $\pi_{\mathcal{V}_x}^{a^*}$.
\ENDFOR
\STATE Calculate the upper bound $\overline{\pi}^a=\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*}$ and the lower bound $\underline{\pi}^a=\pi^a \left(\frac{\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*} \xi_{j,j \in \mathcal{V}_x}^{a^*}}{\sum_{x=1}^m \pi_{\mathcal{V}_x}^{a^*}}, \xi^{p^*} \right)$
\STATE Solve (\ref{Incentives_PT_reformulation}) given $\xi^a=\xi^{a^*}$ using the proposed hierarchical algorithm in Section \ref{sectiontransit} and obtain the optimal solution ${\xi}^{p^*}$ and corresponding optimal value $\pi^{p}\left(\xi^{a^*},{\xi}^{p^*}\right)$.
\ENSURE the equilibrium strategy $\xi^*=(\xi^{a^*},\xi^{p^*})$, the platform profit $\pi^a(\xi^{a^*},\xi^{p^*})$ and its upper bound $\overline{\pi}^a$ and lower bound $\underline{\pi}^a$, the transit ridership $\pi^p(\xi^{a^*},\xi^{p^*})$, the global optimality of $\xi^{p^*}$.
\end{algorithmic}
\end{algorithm}
The aforementioned numerical framework is very useful for us to evaluate the quality of the derived Nash equilibrium. This is because it indicates that the candidate Nash equilibrium is at least as good as an $\epsilon$-Nash equilibrium, where the value of $\epsilon$ can be given by the difference between the profit of the ride-hailing platform at the Nash equilibrium $\pi^a(\xi^{a^*},\xi^{p^*})$ and its upper bound $\bar{\pi}^a$. In particular, we summarize the above discussion as the following proposition:
\begin{proposition} \label{proposition_epsilon_nash}
Suppose the best response process (line 2-9) in Algorithm \ref{algorithm1} converges to a candidate Nash equilibrium $(\xi^{a^*},\xi^{p^*})$, which satisfies $\bar{N}_{ij}^p\leq \dfrac{1}{w_{max}^p}$ for $\forall i,j=1,\dots,M$, then $(\xi^{a^*},\xi^{p^*})$ is an $\epsilon$-Nash equilibrium of the game problem defined by (\ref{Incentives_AMoD}) and (\ref{Incentives_PT}), and we have:
\begin{equation}
\label{value_of_epsilon}
\epsilon=\bar{\pi}^a-\pi^a(\xi^{a^*},\xi^{p^*}).
\end{equation}
\end{proposition}
The above proposition can be easily derived based on Proposition \ref{proposition_upper_lower_bound}, Theorem \ref{themrem_concavity}, and the definition of the $\epsilon$-Nash equilibrium, thus its proof is omitted.
\section{Market Outcomes in the Absence of Regulation} \label{market outcomes}
This section investigates the equity impacts of AVs in the unregulated environment. To this end, we first propose evaluation metrics to quantify both spatial and social equity in the multimodal transportation system. Second, we demonstrate how AVs impact market outcomes and transport equity through numerical studies for San Francisco.
\subsection{Evaluation of Equity}
To characterize spatial and social equity, we first define $u_{ij,k}^t=-c_{ij,k}^t, t\in\mathcal{T}$ as the average utility of passengers from origin zone $i$ to destination zone $j$ in income class $k$ by choosing mode $t$. We further define the accessibility measure $A_{ij,k}$ as the expected maximum utility, which evaluates the expected received utility of passengers in income class $k$ from origin $i$ to destination $j$ in the multimodal transportation system. Based on the multinomial logit (MNL) model (\ref{logit_demand_function}), the maximum expected utility $A_{ij,k}$ can be written as the logsum formula \cite{ho2006combined,van2021evaluating}:
\begin{equation} \label{logsum accessibility}
A_{ij,k} = \frac{1}{\epsilon} \log \sum_{t\in\mathcal{T}} \exp\left(\epsilon u_{ij,k}^t\right) .
\end{equation}
Given the accessibility measure $A_{ij,k}$, we will use Theil coefficient to quantify the transport equity in the multimodal transport system. To this end, we first denote $\lambda_{i,k}$, $\lambda_{k}$, $\overline{\lambda}$ as the passenger demand from zone $i$ in income class $k$, the passenger demand in income class $k$, and the total passenger demand, respectively. They satisfy:
\begin{equation}
\begin{cases}
\lambda_{i,k} = \sum_{j=1}^M \sum_{t\in\mathcal{T}} \lambda_{ij,k}^t \\
\lambda_{k} = \sum_{i=1}^M \sum_{j=1}^M \sum_{t\in\mathcal{T}} \lambda_{ij,k}^t \\
\overline{\lambda} = \sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \sum_{t\in\mathcal{T}} \lambda_{ij,k}^t
\end{cases} .
\end{equation}
Correspondingly, we denote $A_{i,k}$, $A_k$, $\bar{A}$ as the average accessibility of passengers originating from zone $i$ in income class $k$, the average accessibility of passengers in income class $k$, and the average accessibility of all passengers, respectively. They are calculated as:
\begin{equation}
\begin{cases}
A_{i,k} = \frac{\sum_{j=1}^M \sum_{t\in\mathcal{T}} \lambda_{ij,k}^t A_{ij,k}}{\lambda_{i,k}} \\
A_k = \frac{\sum_{i=1}^M \sum_{j=1}^M \sum_{t\in\mathcal{T}} \lambda_{ij,k}^t A_{ij,k}}{\lambda_{k}} \\
\overline{A} = \frac{\sum_{i=1}^M \sum_{j=1}^M \sum_{k=1}^K \sum_{t\in\mathcal{T}} \lambda_{ij,k}^t A_{ij,k}}{\bar{\lambda}}
\end{cases} .
\end{equation}
Finally, we define the Theil (T) coefficient of the accessibility distribution in the multimodal transportation system as:
\begin{equation} \label{Theil_coefficient}
T = \text{WITHIN} + \text{BETWEEN} = \underbrace{\sum_{k=1}^K \sum_{i=1}^M \left(\frac{\lambda_{i,k}}{\overline{\lambda}} \frac{A_{i,k}}{A_k}\right) \ln \left(\frac{A_{i,k}}{A_k}\right)}_\text{spatial equity} + \underbrace{\sum_{k=1}^K \left(\frac{\lambda_k}{\bar{\lambda}} \frac{A_k}{\overline{A}}\right) \ln \left(\frac{A_k}{\bar{A}}\right)}_\text{social equity} .
\end{equation}
As shown in (\ref{Theil_coefficient}), the Theil coefficient can be decomposed into two components. The "WITHIN" component calculates the inequality in the distribution of passenger accessibility across distinct geographic zones {\em within} different groups. Consequently, it characterizes the spatial inequity due to the differentiated mobility service across distinct zones. The "BETWEEN" component evaluates the difference in the distribution of average passenger accessibility {\em between} different income classes, which captures the social inequity arising from the unequal socioeconomic status. Note that a larger Theil coefficient indicates a more inequitable distribution of benefits across different zones or socioeconomic groups. It can be easily verified that if all zones have the same level of accessibility, then the ``WITHIN" component of (\ref{Theil_coefficient}) is zero. Similarly, if all classes of passengers have the same level of accessibility, then the ``BETWEEN" component of (\ref{Theil_coefficient}) is zero.
\subsection{Case Studies} \label{case study}
Consider a case study for San Francisco, where a TNC platform operates autonomous vehicles for mobility-on-demand services and the public transit agency manages a transit network to provide public transit services. We conduct a numerical study using realistic synthetic data for San Francisco. The data consists of the origin and destination of TNC trips at the zip-code granularity, which is synthesized based on the historic total TNC pickups and dropoffs in each zone \cite{SanFrancisco2022TNC} together with a choice model calibrated with survey data. The zip code zones of San Francisco are shown in Figure \ref{fig:SF_multimodal}. Based on the TNC data, we remove zip code zones 94123, 94127, 94129, 94130, 94134 from our analysis since they have negligible trip volumes. We also merge zip code zones 94104, 94105, and 94111 into a single zone, and merge zip code zones 94108 and 94133 into a single zone since each of these individual zones is very small. We synthesize the public transit network based on the Muni system in San Francisco, which is a network of light-rail metro trains, rapid buses, regular buses, cable cars, etc \cite{SanFrancisco2022Muni}. The synthetic transit network consists of the five light rail lines ('F', 'J', 'KT', 'M', and 'N') and the three rapid bus lines ('5R', '9R', '38R') since they form the major skeleton of the Muni network and undertake the majority of trip volumes. The route information of these transit lines can be found in \cite{SanFrancisco2022Muniroutes} and the synthetic transit network is shown in Figure~\ref{fig:SF_multimodal}. To guarantee that passengers are at least accessible to public transit in each zone, we further aggregate zip code zones 94123 and 94115 since they are geographically adjacent\footnote{In practice, each zone can be covered by public transit although the density of stations and service frequency are spatially unequal.}.
We further define the underserved area $\mathcal{U}$ as consisting of zip code zones 94112, 94114, 94116, 94117, 94118, 94121, 94122, 94124, 94131, 94132 based on the density of transit stations in different zones. Note that the underserved area is generally the remote area in the city.
\begin{figure}[t]
\centering
\includegraphics[width=0.58\linewidth]{figures/SF_multimodal_network.pdf}
\caption{Zip code zones of San Francisco County and the synthetic public transit network.}
\label{fig:SF_multimodal}
\end{figure}
We consider heterogeneous passengers with different income levels. In particular, passengers are categorized into three distinct classes, namely, low-income class (class 1), medium-income class (class 2), and high-income class (class 3). Generally, low-income people have a lower valuation of time but a higher valuation of money compared to high-income people, and passengers have a higher valuation of waiting/walking time than that of in-vehicle time. Therefore, the following conditions hold:
\begin{equation}
\begin{cases}
\alpha_k,\theta_k > \beta_k \quad \forall k=1,2,3 \\
\alpha_1 < \alpha_2 < \alpha_3 \\
\beta_1 < \beta_2 < \beta_3 \\
\gamma_1 > \gamma_2 > \gamma_3
\end{cases} .
\end{equation}
In summary, the model parameters involved in the multimodal transportation system are
\begin{equation}
\Theta = \left\{\lambda_{ij,k}^0, \alpha_k, \beta_k, \gamma_k, \epsilon, c_{ij,k}^o, A_i, l_{ij}^a, l_{ij}^p, d_i, v_a, v_p, v_w, r_{max}^p, w_{max}^p, C_{av}, C_{l}, \pi_0 \right\} .
\end{equation}
Due to the lack of real data on AMoD, we calibrate the values of these model parameters based on human-driver-based TNC data and Muni data in San Francisco and map it to the future scenario. In particular, $\lambda_{ij}^0$ is set to satisfy $0.15\lambda_{ij}^0=\lambda_{ij}$, where $\lambda_{ij}$ is the current trip distribution of ride-sourcing trips in San Francisco from zone $i$ to zone $j$, such that 15\% of potential passengers choose to take AMoD services. We further split the potential passenger demand $\lambda_{ij}^0$ among three income classes to obtain $\lambda_{ij,k}^0$. To characterize the spatial difference in socioeconomic characteristics, we assume population distributions of low-income, middle-income, and high-income individuals as 0.4, 0.5, and 0.1 in remote areas and 0.2, 0.5, and 0.3 in core areas\footnote{For simplicity, we define the remote area as the underserved area $\mathcal{U}$. The core area consists of the zip code zones except those in $\mathcal{U}$.}. The potential passenger demand $\lambda_{ij,k}^0$ is then calculated proportionally. The trip distances $l_{ij}^a$ and $l_{ij}^p$ are obtained from Google map estimates. The average speed of TNC vehicles $v_a=17.937$ mph and the average operating speed of public transit $v_p=14.349$ mph are estimated based on the current TNC data and Muni data. For the first-mile/last-mile distance $d_i$, we assume that it follows the "square root law", which indicates that $d_i$ is inversely proportional to the number of transit stations $N_i^T$ in each zone:
\begin{equation}
d_i = \frac{B_i}{\sqrt{N_i^T}} ,
\end{equation}
where $B_i$ is the scaling parameter that captures the area of the zone and the distribution of transit stations. The average walking speed is $v_w$ is set as 3.48 mph. The travel cost of the outside option $c_{ij,k}^o$ is assumed to be proportional to $l_{ij}^a$. To account for both socioeconomic and spatial heterogeneity in $c_{ij,k}^o$, we assume that $c_{ij,1}^o=1.2c_{ij,2}^0$ and $c_{ij,2}^o=1.25c_{ij,3}^o$ such that the travel cost of the outside option of low-income passengers is 20\% higher than that of medium-income passengers and 50\% higher than that of high-income passengers. We further assume that the per-distance cost of the outside transport mode in the remote area is 50\% higher than that in the urban core. Based on data from the Muni system \cite{SanFrancisco2022Munifares,SanFrancisco2022Munifrequencies}, the per-vehicle operating costs of different transit lines $C_{l}$ are calibrated and we set the maximum acceptable per-distance fare $r_{max}^p$ and waiting time $w_{max}^p$ for transit services as \$3/\text{mile} and 20 minutes, respectively.
The rest of the model parameters are set as:
\begin{gather*}
\alpha_1 = 0.5; \quad \alpha_2 = 1.0; \quad \alpha_3 = 2.0; \quad \beta_1 = 0.15; \quad \beta_2=0.30; \quad \beta_3 = 0.65; \\ \gamma_1 = 3.0; \quad \gamma_2 = 1.5; \quad \gamma_3 = 0.75; \quad \theta_1 = 0.5; \quad \theta_2 = 1.0; \quad \theta_3 = 2.0; \\ \epsilon = 0.10; \quad A_i = 7.894 \times \text{Area of zone $i$}; \quad B_i = 1.609 \times \text{Area of zone $i$}; \quad \pi_0 = \$1\mathrm{e}4 \text{/hour} .
\end{gather*}
We adjust the values of these parameters such that when the hourly operating cost of AVs is roughly equivalent to the average driver wage of TNCs, the corresponding market outcomes at Nash equilibrium are close to real-world data of San Francisco (e.g., modal share, trip volumes, average trip fare and frequencies of transit lines). In particular, when $C_{av}=\$30\text{/hour}$, we execute Algorithm \ref{algorithm1} to compute the Nash equilibrium and conduct the ex-post evaluation on the obtained equilibrium. The best response algorithm converges to the same equilibrium $\xi^*=(\xi^{a^*},\xi^{p^*})$ in a few iterations under distinct initial guesses. At the equilibrium, the TNC profit $\pi^a(\xi^{a^*},\xi^{p^*})$ is \$95,613/hour. Under a pairwise partition $\bar{\mathcal{V}}$\footnote{$\mathcal{V}$ is partitioned into subsets, each of which contains two zones, e.g., $\mathcal{V}_1=\{1,2\}$, $\mathcal{V}_2=\{3,4\}$, $\dots$, and $\mathcal{V}_9=\{17,18\}$.}, the established upper bound $\overline{\pi}^a$ is \$97,526/hour. Based on Proposition \ref{proposition_epsilon_nash}, this indicates that $\epsilon$ only accounts for 2.0\% compared with the globally optimal solution to (\ref{Incentives_AMoD}), which is a very tight approximate Nash equilibrium. In the meanwhile, the corresponding lower bound $\underline{\pi}^a$ is \$95,146/hour, which is very close to the equilibrium profit $\pi^a(\xi^{a^*},\xi^{p^*})$. This validates the nature of TNC's decision-making: geographic zones with higher potential profits are given higher priority. {Besides, our numerical experiments showed that the concavity conditions $\bar{N}_{ij}^p\leq \frac{1}{w_{max}^p},\forall i,j=1,\dots,M$, hold and have fairly safe gaps under a large range of model parameters, which verifies the global optimality of $\xi^{p^*}$.} For the market outcomes under $C_{av}=\$30/\text{hour}$, the TNC base fare is \$9.6/trip and the average per-distance rate of AMoD services is \$3.61/mile. The calculated trip fare $9.6+3.61\times2.6$ is close to the fare estimates \cite{SanFrancisco2022rate} for a 2.6-mile trip \cite{castiglione2016tncs}. The per-distance fare of public transit is \$1.14/mile and the service frequencies of the transit lines range from 9.26 per hour to 20.10 per hour, which is consistent with the current fare and frequency setting of Muni system \cite{SanFrancisco2022Muniroutes,SanFrancisco2022Munifares}. Moreover, the modal share of public transit is 22\%, which also coincides with the reported data in the survey \cite{SanFrancisco2022modalshare}.
To investigate the impacts of AVs on the multimodal transportation system, we gradually reduce the cost of AVs, i.e., $C_{av}$, to simulate the scenario in which technological progress makes AVs less expensive over time, resulting in the ride-hailing industry's gradual acceptance of AVs. To this end, we fix all other model parameters, solve the game problem under distinct $C_{av}$ based on Algorithm \ref{algorithm1}, examine how the evolution of AV technology affects the market outcomes in the absence of regulations, and evaluate transport equity using Theil coefficient based on (\ref{Theil_coefficient}). The results are presented in Figure \ref{fig:TNC_no_reg}-\ref{fig:modal_share_no_reg}. Figure \ref{fig:TNC_no_reg} shows TNC's operational strategy and corresponding outcomes under distinct $C_{av}$. As the cost of AV reduces, the ride fare of TNC trips decreases (Figure \ref{fig:base_fare_no_reg}), the number of vehicles on the ride-hailing platform increases (Figure \ref{fig:total_number_vehicles_no_reg}), and the profit of the platform improves (Figure \ref{fig:TNC_profit_no_reg}). These results are intuitive since the reduction in AV cost lowers the marginal cost of AMoD services and thus both the TNC platform and passengers are direct beneficiaries of AV deployment. However, a less intuitive insight is that the spatial distribution of AMoD services is unbalanced:
\begin{itemize}
\item Idle AVs are highly concentrated in the urban core (solid lines in Figure \ref{fig:number_idle_vehicle_no_reg}), indicating that the platform prefers to deploy idle vehicles in high-demand areas. This is intuitive because TNC has higher utilization of AVs and higher potential profits in high-demand areas, which promotes it to deploy more idle vehicles to attract passengers from urban centers. However, the geographical concentration of ride-hailing vehicles will not only exacerbate traffic congestion in the urban core, but also leads to spatial inequality in the waiting time/service quality across distinct zones.
\item As AV cost reduces, the number of idle vehicles increases in both high-demand and low-demand areas of the city, but the gap of service quality between high-demand and low-demand areas is further enlarged (i.e., compare solid and dashed lines in Figure \ref{fig:number_idle_vehicle_no_reg}). This indicates that without regulatory intervention, lowering the cost of AV will promote the proliferation of AVs, but at the same time exacerbate the existing spatial inequality gaps in the multimodal transportation network.
\end{itemize}
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_no_regulations/average_trip_fare.pdf}
\caption{{Average TNC trip fare under distinct $C_{av}$.}}
\label{fig:base_fare_no_reg}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_no_regulations/total_number_AVs.pdf}
\caption{Total number of AVs under distinct $C_{av}$.}
\label{fig:total_number_vehicles_no_reg}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_no_regulations/TNC_profit.pdf}
\caption{Hourly TNC profit under distinct $C_{av}$.}
\label{fig:TNC_profit_no_reg}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_no_regulations/number_idle_AVs.pdf}
\caption{Number of idle vehicles in different zones under distinct $C_{av}$.}
\label{fig:number_idle_vehicle_no_reg}
\end{subfigure}
\caption{TNC's operational decisions and outcomes under distinct $C_{av}$.}
\label{fig:TNC_no_reg}
\end{figure}
To understand how the deployment of AVs affects social equity, we shall compare the impacts of AVs on distinct passenger income classes. Note that low-income populations are more transit-dependent, thus to characterize social equity, we will first examine the operational decisions of the transit agency.
Figure \ref{fig:PT_no_reg} presents the public transit agency's operational decisions and the corresponding outcomes under distinct $C_{av}$. Based on the simulation results, the proliferation of AVs has a moderate impact on the transit operational decisions: as $C_{av}$ reduces, the fare of public transit monotonically increases (Figure \ref{fig:transit_fare_no_reg}), the service frequencies of most transit lines decreases (Figure \ref{fig:transit_frequency_no_reg}), the revenue and operating cost of the transit network slightly decrease (Figure \ref{fig:transit_revenue_no_reg}), and the total ridership of public transit under distinct $C_{av}$ exhibits two regimes: when $C_{av}>\$36/\text{hour}$, the transit ridership increases as $C_{av}$ reduces; when $C_{av}\leq \$36/\text{hour}$, the transit ridership decreases as $C_{av}$ drops (Figure \ref{fig:transit_ridership_no_reg}). The above results reveal some interesting competitive and complementary effects of AMoD services on public transit. Note that the TNC provides on-demand door-to-door mobility services (direct AMoD services, i.e., mode $a$), which is a substitution for public transit. On the contrary, TNC trips serving as first/last-mile connections to public transit (bundled services, i.e., mode $b_1$, $b_2$, and $b_3$) could be a complement/feeder to public transit. The TNC and public transit agency carefully determine their operational strategies to balance the competitiveness and complementarity between AMoD services and public transit. When $C_{av}$ is relatively high, the competitiveness/substitution of AMoD services to public transit is weak, and the complementarity dominates. To strengthen the complementary effect and fully utilize the benefits of reduced AV cost, the public transit agency improves the service frequencies of some transit lines to attract more passengers to choose bundled services (Figure \ref{fig:transit_frequency_no_reg}). In this case, although the number of passengers choosing direct public transit drops, the passenger arrival rates of bundled services increase, which leads to increased total ridership of public transit (Figure \ref{fig:transit_ridership_no_reg}). The total operating cost of the transit network increases (Figure \ref{fig:transit_revenue_no_reg}), and therefore the agency raises the transit fare to maintain a balance of revenue and expenditure (Figure \ref{fig:transit_fare_no_reg}). On the other hand, when $C_{av}$ is relatively low, the competitiveness/substitution of direct AMoD services to public transit increases as $C_{av}$ reduces and gradually becomes the dominant effect. The competition between the TNC and public transit deteriorates transit services and procures passengers to deviate from public transit (Figure \ref{fig:transit_frequency_no_reg}). The total transit ridership significantly decreases as $C_{av}$ reduces (Figure \ref{fig:transit_ridership_no_reg}), and both the revenue and operating cost of the transit network decrease (Figure \ref{fig:transit_revenue_no_reg}). However, we acknowledge that the above impacts are insignificant because the AV cost only influences the transit agency indirectly through competition.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_no_regulation/transit_fare.pdf}
\caption{Per-distance transit fare under distinct $C_{av}$.}
\label{fig:transit_fare_no_reg}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_no_regulation/transit_frequency.pdf}
\caption{Frequencies of different transit lines under distinct $C_{av}$.}
\label{fig:transit_frequency_no_reg}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_no_regulation/PT_revenue_cost.pdf}
\caption{Transit revenue and operating cost under distinct $C_{av}$.}
\label{fig:transit_revenue_no_reg}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_no_regulation/PT_ridership.pdf}
\caption{Total public transit ridership under distinct $C_{av}$.}
\label{fig:transit_ridership_no_reg}
\end{subfigure}
\caption{Public transit's operational decisions and outcomes under distinct $C_{av}$.}
\label{fig:PT_no_reg}
\end{figure}
To further investigate the impact of AVs on social equity, we compare the mode choices of different income classes of the multimodal transportation system in Figure \ref{fig:modal_share_no_reg}. As AV cost reduces, the modal shares of direct AMoD services (mode $a$) and bundled services (mode $b_1$, $b_2$, and $b_3$) increase, and the proportion of passengers choosing public transit (mode $p$) decreases for distinct income classes (Figure \ref{fig:modal_share_low_no_reg}-\ref{fig:modal_share_high_no_reg}). Overall, passengers shift from the outside option (e.g., walking, bicycling, and driving) and public transit to direct AMoD services and bundled services as AV cost reduces (Figure \ref{fig:modal_share_overall_no_reg}). These results are intuitively straightforward: the decrease in AV cost lowers the marginal cost of AMoD services, which promotes cheaper and improved AMoD services and instead disrupts transit services (in the second regime). However, the impacts of AVs on the modal choices of distinct classes of travelers are different, indicating that the distribution of benefits of AV is unequal:
\begin{itemize}
\item Low-income individuals are highly transit-dependent. More than half of low-income individuals take public transit and over a third of them choose the outside option (e.g., walking or bicycling) (Figure \ref{fig:modal_share_low_no_reg}). As AV cost reduces, the increase in the modal share of AMoD services and bundled services is insignificant (Figure \ref{fig:modal_share_low_no_reg}). To understand this, note that low-income people have a relatively high valuation of monetary cost and a low valuation of time. As AV cost reduces, although ride fares and the waiting time of AMoD services decrease, the TNC deploys a small number of idle AVs and set relatively high ride fares (compared to transit) in remote areas with a high proportion of low-income individuals, who have to stick to public transit although its travel cost increases.
\item The evolution of AV technology significantly transforms medium-income individuals' mode choices. As AV cost reduces, the number of medium-income individuals taking transit decreases while the passenger arrival rate of direct AMoD services gradually increases and finally surpasses that of public transit (Figure \ref{fig:modal_share_medium_no_reg}). In the meanwhile, an increasing number of medium-income individuals choose bundled services (Figure \ref{fig:modal_share_medium_no_reg}), although they prefer to use AMoD as the first-mile or last-mile connection rather than for both the first-mile and last-mile connection (Figure \ref{fig:modal_share_medium_no_reg}).
\item High-income individuals prefer convenient and fast mobility options. As AV cost reduces, the majority of high-income individuals shift from the outside option (e.g., driving) to direct AMoD services and only a small proportion of individuals choose public transit and bundled services (Figure \ref{fig:modal_share_high_no_reg}). High-income people have a high valuation of time and therefore they are unwilling to experience a long trip time and prefer driving or direct AMoD services.
\end{itemize}
To summarize, the above analysis indicates that the proliferation of AVs will increase both spatial and social inequity gaps in the multimodal transportation network. This is consistent with the simulation results presented in Figure \ref{fig:Theil_no_reg}, where both the 'WITHIN' component (i.e, spatial inequity) and the 'BETWEEN' component (i.e., social inequity) increase and the Theil coefficient significantly increases as $C_{av}$ reduces. We comment that the reason for the expansion of both the spatial inequity gap and social inequity gap is twofold. On the one hand, the spatial inequality in accessibility across different zones is mainly caused by the geographic concentration of AMoD services. As mentioned above, due to the for-profit nature of the TNC platform, idle AVs are highly concentrated in the urban core (Figure \ref{fig:number_idle_vehicle_no_reg}), and passengers in the urban core enjoy more convenient services compared to those in remote areas. The geographical imbalance in AMoD services across distinct zones causes spatial inequity in accessibility. Besides, such geographic imbalance in AMoD services exacerbates as AV cost reduces (Figure \ref{fig:number_idle_vehicle_no_reg}), which enlarges the spatial inequality in accessibility and reinforces the spatial inequity gap (Figure \ref{fig:Theil_no_reg}). On the other hand, the increase in social inequity is not only due to the moderate disruptiveness of AMoD services to public transit, but also because the benefits of reduced AV costs are primarily enjoyed by medium-income and high-income individuals, and the increase in the accessibility of low-income individuals is less significant compared to that of medium-income and high-income individuals (Figure \ref{fig:accessibility_groups}). Therefore, the social inequity gap exacerbates as AV cost reduces (Figure \ref{fig:Theil_no_reg}).
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/passengers_no_regulation/modal_share_low.pdf}
\caption{Modal share in low-income class under distinct $C_{av}$.}
\label{fig:modal_share_low_no_reg}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/passengers_no_regulation/modal_share_medium.pdf}
\caption{Modal share in medium-income class under distinct $C_{av}$.}
\label{fig:modal_share_medium_no_reg}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/passengers_no_regulation/modal_share_high.pdf}
\caption{Modal share in high-income class under distinct $C_{av}$.}
\label{fig:modal_share_high_no_reg}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/passengers_no_regulation/modal_share_overall.pdf}
\caption{Overall modal share under distinct $C_{av}$.}
\label{fig:modal_share_overall_no_reg}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/equity_no_regulation/accessibility_groups.pdf}
\caption{Average accessibility of different income groups under distinct $C_{av}$.}
\label{fig:accessibility_groups}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/equity_no_regulation/Theil.pdf}
\caption{Theil coefficient, WITHIN component, and BETWEEN component under distinct $C_{av}$}
\label{fig:Theil_no_reg}
\end{subfigure}
\caption{Modal share and transport equity in the multimodal transportation system under distinct $C_{av}$.}
\label{fig:modal_share_no_reg}
\end{figure}
\section{Impacts of Regulatory Policies} \label{regulations}
Our numerical study so far suggests that spatial inequality and social inequality gap enlarges as AV cost reduces. This section proposes two regulatory policies, i.e., a minimum service-level requirement and a subsidy on bundled services, to mitigate the negative equity impacts of AVs in the multimodal transportation network.
\subsection{A Minimum Service-Level Requirement} \label{service}
To improve spatial equity, we consider a service-level requirement that mandates TNCs to guarantee an acceptable waiting time for AMoD services in all areas of the city. Such policy has already been implemented in practice, for instance, New York City requires that 90\% of wheelchair-accessible vehicle requests be fulfilled within 15 minutes \cite{NYCWAV}. Under the proposed model, this policy can be easily modeled as:
\begin{equation} \label{minimum_serivce}
w_i^a\left( N_i^I \right) \leq w_{max}^a, \forall i=1,2,\dots,M ,
\end{equation}
where $w_i^a(\cdot)$ is the average waiting time for AMoD trips starting in zone $i$, and $w_{max}^i$ is the maximum allowable waiting time set by the regulatory agency to ensure reasonable and equitable service quality for travelers across distinct zones. In this case, the TNC determines its operational strategy to maximize its profit subject to the market equilibrium conditions (\ref{logit_demand_function}), (\ref{generalized_travel_cost}), (\ref{waiting_time_AMoD}) and (\ref{vehicle_hour_conservation}) and the minimum service-level requirement (\ref{minimum_serivce}). The public transit agency strategically adjusts its operational decisions to maximize transit ridership subject to the equilibrium constraints (\ref{logit_demand_function}), (\ref{generalized_travel_cost}), (\ref{waiting_time_PT}), (\ref{waiting_time_constraint}) and (\ref{profit_constraint}). Note that the minimum service-level requirement does not change the structure of the profit maximization problem for the TNC and the ridership maximization problem for public transit. We can use Algorithm \ref{algorithm1} to solve the game problem and conduct an ex-post evaluation on the computed Nash equilibrium.
To investigate the impacts of the minimum service-level requirement, we fix $C_{av}$ as 20\$/hour and remain other model parameters unchanged, solve the game problem under distinct $w_{max}^a$, and present how it impacts the market outcomes and transport equity in Figure \ref{fig:TNC_service}-\ref{fig:equity_service}.
\begin{figure}[t!]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_service/average_trip_fare.pdf}
\caption{{Average trip fare of the ride-hailing services under distinct values of $w_{max}^a$.}}
\label{fig:base_fare_service}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_service/number_idle_AVs.pdf}
\caption{Number of idle vehicles in different zones of the transport network under distinct values of $w_{max}^a$.}
\label{fig:number_idle_vehicle_service}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_service/TNC_profit.pdf}
\caption{Hourly profit of the TNC platform under distinct values of $w_{max}^a$.}
\label{fig:TNC_profit_service}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_service/transit_fare.pdf}
\caption{ Trip fare (per-distance) of transit services under different values of $w_{max}^a$.}
\label{fig:transit_fare_service}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_service/transit_frequency.pdf}
\caption{Frequencies of different transit lines under distinct values of $w_{max}^a$.}
\label{fig:transit_frequency_service}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_service/PT_ridership.pdf}
\caption{Total public transit ridership per minute under distinct values of $w_{max}^a$.}
\label{fig:transit_ridership_service}
\end{subfigure}
\caption{Operational decisions of the TNC platform and transit agency under distinct $w_{max}^a$.}
\label{fig:TNC_service}
\end{figure}
Figure \ref{fig:TNC_service} shows the operational strategies of TNC platform and the public transit agency under distinct $w_{max}^a$. Under a stricter minimum service-level requirement, the number of idle vehicles in distinct zones remains unchanged initially and then increases, especially in the low-demand areas (Figure \ref{fig:number_idle_vehicle_service}). In this case, the TNC earns a reduced profit under a stricter minimum service-level requirement (Figure \ref{fig:TNC_profit_service}) and increases the average ride fare to improve profitability (Figure \ref{fig:base_fare_service})). Clearly, this lower bound enforces the TNC to improve the service quality in low-demand areas where the TNC is reluctant to place its idle AVs in the absence of regulations. This improves spatial equity while hurting the profitability of TNC's AMoD business.
Figure \ref{fig:transit_fare_service}-\ref{fig:transit_ridership_service} presents the transit agency's operational strategies and the corresponding outcomes under distinct $w_{max}^a$. As $w_{max}^a$ reduces, the agency slightly lowers the transit fare (Figure \ref{fig:transit_fare_service}) and increases the service frequencies of its transit lines (Figure \ref{fig:transit_frequency_service}). This leads to reduced waiting time for transit services and increased total transit ridership (Figure \ref{fig:transit_ridership_service}). Similarly, this can be explained as a consequence of the complementarity between AMoD services and public transit. Imposing a minimum service-level requirement improves the service quality of AMoD services, especially in underserved/remote areas (Figure \ref{fig:number_idle_vehicle_service}). It promotes the agency to improve the transit service and reinforce the AMoD-transit complementarity to attract more passengers to choose bundled services and stimulate a higher transit ridership.
Figure \ref{fig:equity_service} demonstrates the impacts of the minimum service-level requirement on accessibility and transport equity in the multimodal transportation system. As $w_{max}^a$ reduces, the inequality in accessibility across distinct zones reduces (Figure \ref{fig:Theil_service}), while the inequality gap between the accessibility of different income groups keeps widening (Figure \ref{fig:accessibility_groups_service}-\ref{fig:Theil_service}). Consequently, as $w_{max}^a$ reduces, the spatial inequity ('WITHIN' component) decreases, whereas the social inequity ('BETWEEN' component) keeps increasing, and the overall inequity gap (Theil coefficient) first reduces and then enlarges (Figure \ref{fig:Theil_service}). This indicates that imposing a minimum service-level requirement is double-edged: on the one hand, it reduces the geographic concentration of idle vehicles and improves spatial equity; on the other hand, enforcing more idle AVs in less lucrative areas primarily benefits high-income individuals in underserved areas, who rely on AMoD much more than low-income individuals in the same zone. In this case, the accessibility improvement of low-income individuals is insignificant due to the increased TNC ride fares and high-income individuals are the primary beneficiaries of the minimum service-level requirement, which further reinforces the existing social inequity gap.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/equity_service/accessibility_groups.pdf}
\caption{Average accessibility of different income groups under distinct valuse of $w_{max}^a$.}
\label{fig:accessibility_groups_service}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/equity_service/Theil.pdf}
\caption{Theil coefficient, WITHIN component, and BETWEEN component under distinct $w_{max}^a$}
\label{fig:Theil_service}
\end{subfigure}
\caption{Accessibility and transport equity under distinct $w_{max}^a$.}
\label{fig:equity_service}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_subsidy/average_trip_fare.pdf}
\caption{{Average trip fare of the ride-hailing services under distinct values of subsidy $s$.}}
\label{fig:base_fare_subsidy}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_subsidy/number_idle_AVs.pdf}
\caption{Number of idle vehicles in different zones of the transport network under distinct values of subsidy $s$.}
\label{fig:number_idle_vehicle_subsidy}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/TNC_subsidy/TNC_profit.pdf}
\caption{Hourly profit of the TNC platform under distinct values of subsidy $s$.}
\label{fig:TNC_profit_subsidy}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_subsidy/transit_fare.pdf}
\caption{Trip fare (per-distance) of transit services under different values of subsidy $s$.}
\label{fig:transit_fare_subsidy}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_subsidy/transit_frequency.pdf}
\caption{Frequencies of different transit lines under distinct values of subsidy $s$.}
\label{fig:transit_frequency_subsidy}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/PT_subsidy/PT_ridership.pdf}
\caption{Total public transit ridership under distinct values of subsidy $s$.}
\label{fig:transit_ridership_subsidy}
\end{subfigure}
\caption{Operational decisions of the TNC platform and the public transit agency under distinct $s$.}
\label{fig:TNC_subsidy}
\end{figure}
\subsection{Subsidies on Bundled Services} \label{subsidies}
To strengthen the synergy between TNC and transit, and to improve spatial and social equity at the same time, we consider a subsidy on TNC trips that connect passengers to public transit. We focus on the underserved/remote areas $\mathcal{U}$ where there is a large population of low-income transit-dependent households who are poorly served by public transit and AMoD services. A subsidy $s$ is provided for TNC trips that start from or end near a transit stop in zone $i\in \mathcal{U}$. Under the subsidy, the generalized travel costs of distinct mobility options can be captured by:
\begin{equation} \label{generalized_travel_cost_subsidy}
\begin{cases}
c_{ij,k}^a = \alpha_k w_i^a + \beta_k \frac{l_{ij}^a}{v_a} + \gamma_k (b + r_i^a l_{ij}^a) \\
c_{ij,k}^p = \alpha_k w_{ij}^p + \beta_k \frac{l_{ij}^p}{v_p} + \gamma_k r^p l_{ij}^p + \theta_k \left(\frac{d_i}{v_w}+\frac{d_j}{v_w}\right) \\
c_{ij,k}^{b_1} = \alpha_k \left(w_i^a + w_{ij}^p \right) + \beta_k \left(\frac{d_i}{v_a} + \frac{l_{ij}^p}{v_p}\right) + \gamma_k \left(b + r_i^a d_i - s \mathbbm{1}_i^{\mathcal{U}} + r^p l_{ij}^p\right) + \theta_k \frac{d_j}{v_w} \\
c_{ij,k}^{b_2} = \alpha_k \left( w_{ij}^p + w_j^a \right) + \beta_k \left( \frac{l_{ij}^p}{v_p} + \frac{d_j}{v_a}\right) + \gamma_k \left( r^p l_{ij}^p + b + r_j^a d_j - s \mathbbm{1}_j^{\mathcal{U}}\right) + \theta_k \frac{d_i}{v_w} \\
c_{ij,k}^{b_3} = \alpha_k \left(w_i^a + w_{ij}^p + w_j^a\right) + \beta_k \left(\frac{d_i}{v_a} + \frac{l_{ij}^p}{v_p} + \frac{d_j}{v_a}\right) + \gamma_k \left(b + r_i^a d_i - s \mathbbm{1}_i^{\mathcal{U}} + r^p l_{ij}^p + b + r_j^a d_j - s \mathbbm{1}_j^{\mathcal{U}} \right) \\
\end{cases}
\end{equation}
where $\mathbbm{1}_i^{\mathcal{U}}$ is an indicator function that equals 1 if $i\in \mathcal{U}$ and 0 otherwise. In this case, the TNC deploys its operational strategy to maximize its profit subject to the modified generalized travel costs (\ref{generalized_travel_cost_subsidy}) and other unchanged equilibrium conditions (\ref{logit_demand_function}), (\ref{waiting_time_AMoD}) and (\ref{vehicle_hour_conservation}). Similarly, the public transit agency determines the operational strategy to maximize the transit ridership subject to the modified generalized travel costs (\ref{generalized_travel_cost_subsidy}) and other unchanged equilibrium constraints (\ref{logit_demand_function}), (\ref{waiting_time_PT}), (\ref{waiting_time_constraint}) and (\ref{profit_constraint}). Note that the subsidy $s$ is a constant which does not alter the structure of the profit maximization problem for the TNC and the ridership maximization problem for public transit. Therefore, we can still use Algorithm \ref{algorithm1} to compute the Nash equilibrium and perform an ex-post evaluation on the obtained equilibrium.
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/equity_subsidy/accessibility_groups.pdf}
\caption{Average accessibility of different income groups under distinct values of $s$.}
\label{fig:accessibility_groups_subsidy}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/equity_subsidy/Theil.pdf}
\caption{Theil coefficient, WITHIN component, and BETWEEN component under distinct values of $s$}
\label{fig:Theil_subsidy}
\end{subfigure}
\caption{Accessibility and transport equity under distinct $s$.}
\label{fig:equity_subsidy}
\end{figure}
To investigate the impacts of subsidies for bundled services, we focus on the nominal case with $C_{av}=20\$/\text{hour}$ and remain other model parameters unchanged as in Section \ref{case study}, solve the game problem under distinct $s$, and demonstrate how it affects the market outcomes and transport equity in Figure \ref{fig:TNC_subsidy}-\ref{fig:equity_subsidy}.
Figure \ref{fig:base_fare_subsidy}-\ref{fig:TNC_profit_subsidy} shows TNC's operational strategy and the corresponding market outcomes under distinct $s$. As the subsidy increases, the TNC raises the trip fare (Figure \ref{fig:base_fare_subsidy}), increases the number of idle vehicles in distinct zones (Figure \ref{fig:number_idle_vehicle_subsidy}), and earns an increased profit under a higher subsidy level (Figure \ref{fig:TNC_profit_subsidy}). Interestingly, the majority of increased idle AVs are deployed in remote zones (dashed lines in Figure \ref{fig:number_idle_vehicle_subsidy}). This is because subsidizing TNC trips that connect passengers to public transit directly reduces the monetary costs of bundled trips in underserved/remote areas and attracts more passengers to choose bundled services, which enables the TNC to reduce the base fare and raise its per-distance fares in distinct zones to earn increased trip fares. To accommodate the increased demand for first/last-mile AMoD services, the TNC primarily deploys more idle vehicles in underserved areas, thus improving spatial equity.
Figure \ref{fig:transit_fare_subsidy}-\ref{fig:transit_ridership_subsidy} demonstrates the public transit agency's operational decisions and the corresponding outcomes under distinct $s$. As the subsidy level improves, the transit agency decreases its fare (Figure \ref{fig:transit_fare_subsidy}) and increases the frequencies of distinct transit lines (Figure \ref{fig:transit_frequency_subsidy}), which reduces passengers' waiting time for transit in distinct zones and leads to an increased total public transit ridership (Figure \ref{fig:transit_ridership_subsidy}). We argue that this is again attributed to the complementary effect of AMoD services on public transit: subsidizing TNC trips as first/last-mile connections strengthens the complementarity of AMoD services to public transit, thus the public transit agency reduces the transit fare and improves the service frequencies to attract more passengers to choose bundled services, which increases the total transit ridership.
Figure \ref{fig:equity_subsidy} shows the impacts of subsidies on accessibility and transport equity in the multimodal transportation system. As the subsidy increases, the accessibility of all different income groups increases at the same time (Figure \ref{fig:equity_subsidy}). More importantly, the accessibility increase in underserved/remote areas is more significant than that in core areas and the accessibility increase of low-income and medium-income individuals are much higher than that of high-income individuals, and the inequality in accessibility across distinct geographic zones and between distinct income groups reduce with the subsidy. Consequently, both spatial inequity ('WITHIN' component) and social inequity ('BETWEEN' component) reduce and the overall transport inequity (Theil coefficient) alleviates (Figure \ref{fig:Theil_subsidy}). The equity improvement is natural since subsidies are devoted to AMoD trips for first/last-mile connections in underserved areas with a large proportion of low-income transit-dependent individuals. It incentivizes the TNC to deploy more idle AVs in remote areas and reduces the geographic concentration of AMoD services (Figure \ref{fig:number_idle_vehicle_subsidy}). In the meanwhile, it promotes the transit agency to provide improved transit services (Figure \ref{fig:transit_fare_subsidy}-\ref{fig:transit_frequency_subsidy}). In this case, the subsidies reduce the travel costs of both public transit and bundled services, which improves accessibility and especially benefits low-income transit-dependent individuals. Consequently, subsidies on bundled services can bridge the existing spatial and social inequity gaps at the same time.
\section{Conclusion} \label{conclusion}
This paper assesses the equity impacts of AVs and investigates regulatory policies that guarantee the benefits of AV deployment reach underserved areas and transportation-disadvantaged groups. A network game-theoretic model is formulated to characterize the intimate interactions among TNC's location-differentiated price, the transit fare, services frequencies of distinct transit lines, waiting times for AMoD services, waiting times for transit services, TNC fleet size, and the modal share of different income groups in the multimodal transportation system, by which the incentives of multiclass passengers, the TNC, and the public transit agency are captured. An algorithm is developed to compute the Nash equilibrium and conduct an ex-post evaluation of the performance of the obtained Nash equilibrium. Theil coefficient is utilized to quantify both the spatial and social inequality in individuals' accessibility in the multimodal transportation system.
Based on the developed framework, we reveal the spatial and social inequity gaps without regulatory interventions and investigate regulatory policies that improve transport equity. Through numerical study, we showed that although individuals' accessibility improves, spatial and social inequity gaps enlarge as AV technology evolves. The for-profit nature of TNC leads to the geographic concentration of AMoD services and exacerbates spatial inequity. In the meanwhile, the competition between AMoD and public transit disrupts transit services and the benefits of reduced costs for AMoD services are primarily distributed to higher-income individuals who are less transit-dependent, which expands the social inequity gap. The impacts of two regulatory policies are evaluated: (a) a minimum service-level requirement on AMoD services; and (b) subsidies on AMoD trips for first/last-mile connections. We showed that a minimum service-level requirement incurs a trade-off: as a higher minimum service level is required, the spatial inequity decreases while the social inequity keeps increasing. Therefore, the regulatory agency should evaluate the trade-off between spatial equity and social equity and carefully control the minimum service-level requirement. On the other hand, subsidizing bundled services will benefit all stakeholders: higher subsidy yields increased TNC profit, transit ridership, and accessibility, and more importantly, improved spatial and social equity at the same time.
This paper delivers a comprehensive socioeconomic analysis of AV-enabled mobility future, with a special focus on transport equity. One future extension is extending the model to capture a mixture of AVs and human drivers in a TNC market and consider equity from both passenger and driver sides. Another extension would be incorporating other socio-economic factors (other than income levels) into the study of social equity.
\section*{Acknowledgments} {This research was supported by Hong Kong Research Grants Council under project 26200420, 16202922, and National Science Foundation of China under project 72201225.}
\bibliographystyle{unsrt}
|
\section{Introduction}
Federate Learning \cite{zhang2022fine, zhu2021data, li2020federated, chen2022anomalous, huang2022learn} has become a hot research topic in recent years due to applications in many fields where participants don't want to share private training data but want to have a co-trained model with high quality. However, the usually non-iid distribution of training data of participating clients harms the model convergence and model accuracy to a large extent. Supervised Contrastive Learning which tries to minimize the difference between feature space of points belonging to the same class and pushes away points from different classes was found to outperform cross-entropy \cite{simclr,simsiamese}. MOON \cite{moon} is one of the state-of-the-art regularization terms that is found to be very effective in solving heterogeneity problem in federated learning by utilizing the similarity between model representations to correct the local training of individual parties. Inspired by these two ideas, we propose Shared Supervised-Contrastive Federated Learning(FedSSC) to tackle the heterogeneity problem. In FedSSC, devices can share with each other the learned class-wise feature embeddings and add the supervised-contrastive learning loss as a regularization term to foster the feature space learning. The loss tries to minimize the cosine similarity distance between the current sample's feature embedding and the averaged feature embedding from another device if they are in the same class, and on the contrary maximizes the distance if they are in different classes. This new regularization term when added on top of the MOON regularization term is found to outperform the other state-of-the-art regularization terms by getting higher accuracy and converging in fewer rounds.
\section{Problem to Solve}
\subsection{Problem Statement}
We are particularly interested in the non-IID setting of federated learning. Specifically, we assume that there are $N$ devices, and each of them has the training data $D_i$ where $i \in \{1, 2, ..., N\}$. Each $D_i$ consists of a different distribution of class labels. Our goal is to learn a machine learning model without local devices directly sharing training data to solve
\begin{equation} \label{eq:loss term}
argmin_{w} L(w) = \sum_i^N \frac{|D_i|}{|D|} L_i(w)
\end{equation}
where $L_i(w)$ is the empirical loss of the local device.
\subsection{Background}
Federated learning is widely used to perform decentralized training of a global model on multiple devices while preserving the data privacy of each device. One of the most basic yet popular models is FedAvg \cite{fedavg} algorithm, in which a central server aggregates the local model weights on each device to build a global model without directly accessing the local training data. However, it suffers from heterogeneous local data on each training device which increases the difficulty of reaching the same level of accuracy as the centralized training. To tackle the heterogeneity, MOON proposes to add a regularization term to
prevent local feature representation of the image from being too far from the global feature representation of the same image \cite{moon}. MOON outperformed other regularization terms like FedProx \cite{fedprox}.
\subsection{Research Goal}
The goal of our research is to improve the federated learning algorithm in non-IID scenarios without sharing raw data across devices. Specifically, we would like to introduce an extra regularization for class-wise feature space through supervised contrastive loss on top of the MOON regularization term.
\subsection{Evaluation Metrics}
To compare our method with other existing approaches, we evaluate the model performance by the top-1 accuracy of global model on an isolated test set. Moreover, we use the number of communication rounds to achieve the same level of accuracy as our metric for convergence speed.
\section{Approach}
Inspired by the idea of Supervised Contrastive Learning, in addition to the MOON loss FedSSC utilizes class-wise average feature maps shared by other devices to correct local training and to tackle the heterogeneity problem. The objective function for the local device is composed of three parts: 1) typical supervised learning loss term calculated with cross-entropy ($l_{class}$), 2) MOON loss ($l_{moon}$), and 3) global class-wise contrastive loss ($l_{glob}$). With $\tau$ being the temperature, $l_{moon}$ takes the projected feature representations $z$, $z_{glob}$, and $z_{prev}$ by passing the same image into the current model, the current round's global model, and the previous epoch's model\cite{moon}.
\begin{equation} \label{eq:moon_loss}
l_{moon}= -log \frac{exp(sim(z,z_{glob})/\tau)}{exp(sim(z,z_{glob})/\tau)+exp(sim(z,z_{prev})/\tau)}\\
\end{equation}
Similarly $l_{glob}$ takes the temperature $\tau$, the current projected feature representations $z^i$ in class $i$ and the shared global class-wise projected feature representations $zs_{glob}$ with a total of $|K|$ classes. For the shared global representation in the same class as the image, we treat them as a positive pair, whereas any other shared global representations in a different class as negative pairs.
\begin{equation} \label{eq:global_loss}
l_{glob}= -log \frac{exp(sim(z^i,zs^i_{glob})/\tau)}{\sum_{k \in K} exp(sim(z^i,zs^k_{glob})/\tau)}\\
\end{equation}
To construct the $zs_{glob}$, we first have each local device report to global server the class-wise projected feature representations at the last round of epoch using Equation \ref{eq:global_rep} assuming that each device has $N$ training data with $N^k$ of them in class $k$. In each communication round, for each class the global server randomly selects a device who has at least 10 images of that class locally as a source of that class's class-wise feature representation.
\begin{equation} \label{eq:global_rep}
zs^k_{glob} = \frac{\sum_{j}^N z^i_j * 1_{i=k}}{N^k}
\end{equation}
As shown in Equation \ref{eq:loss}, we can tune the two parameters $\mu_{moon}$ and $\mu_{glob}$ to weight the MOON loss and global class-wise contrastive loss differently. The local objective is to minimize the $l$.
\begin{equation} \label{eq:loss}
l = l_{class} + \mu_{moon}*l_{moon} + \mu_{glob} * l_{glob}\\
\end{equation}
\begin{algorithm}
\caption{FedSSC Framework}\label{alg:cap}
\begin{algorithmic}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\Require number of communication rounds $T$,
number of devices $P$, number of local
epochs $E$, temperature $\tau$ , learning rate $\eta$,
hyper-parameter $\mu$, total number of data $N$, number of data in device$_i$ is $N_i$
\Ensure Final global model $w^T$
\\ \textit{\textbf{Global Server}} :
\\ \text{Initialize global model $w^0$}
\\ \text{and classwise feature representation $zs^0$}
\For{$t=0,1,...,T-1$}
\For{$i=1,...,P$}
\State send the $w^t$ and the $zs^t$ to device $i$
\State $w_i^{t+1} , zs_i^{t+1} \gets$ \textbf{LocalTraining}($i,w^t,zs^{t}$)
\EndFor
\State $w^{t+1}\gets \sum_i^P \frac{|N_i|}{|N|} w_i^t$
\State $zs^{t+1}\gets \sum_i^P \frac{zs_i^{t+1}}{P} $
\EndFor
\State return $w^T$
\\
\\ \textit{\textbf{LocalTraining}} :
\State $w_0^t = w^t$
\For{$i=0,...,E-1$}
\For{each batch b = ($x,y$) of $N_i$}
\State $l_{class} \gets CrossEntropyLoss(F_{w_i^t}(x),y)$
\State $z \gets Proj(Enc(w_i^t;x))$
\State $z_{glob} \gets Proj(Enc(w^t;x))$
\State $z_{prev} \gets Proj(Enc(w_{i}^{t-1};x))$
\State $l \gets l_{class} + \mu_{moon}*l_{moon}(z,z_{glob},z_{prev}) + \mu_{glob} * l_{glob}(z,zs^t)$
\State $w_{i+1}^t \gets w_i^t - \eta \Delta l $
\EndFor
\EndFor
\For{each class c $\in$ C}
\State $zs^c = \frac{1}{N^c}\sum_{x}^N Proj(Enc(w_E^t;x^j)) * 1_{j=c}$
\EndFor
\State return $w_E^{t+1}$, $zs$\\
\end{algorithmic}
\end{algorithm}
\section{Intellectual Points}
Our contributions are two-fold. First, most prior works focused on regularizing local devices' weights or only regularizing feature representations of local images\cite{moon},\cite{fedprox}. However, our approach directly takes advantage of feature representations from other devices without sharing the raw data by regularizing each sample's local representation with its corresponding global class-wise feature representation. The global class-wise feature map is from a randomly selected device for each round
Furthermore, our experiments show that the global representation contrastive loss and the MOON loss are complementary to each other. We have considered losses other than the MOON loss by modifying its negative pair, but the model could easily collapse or the performance would be worse than the FedAvg. Moreover, even without the MOON loss, simply adding our regularization term on top of the supervised learning loss can achieve the same level of accuracy with MOON. However, using one of them doesn't outperform the other. Combining them together is the key to our success.
\section{Work Performed}
\subsection{Dataset}
We used CIFAR-10 as our experiment dataset because it is relatively small and widely used in previous papers. To simulate the non-IID scenario, we follow MOON's setup by using the Dirichlet distribution to generate dataset $D_i$ for each device. Specifically, we sample ${p_k} \sim Dir_N(\beta)$ for each class and allocate $p_{kj}$ samples of class $k$ to device $j$. Our default is $\beta=0.5$, which simulates a severe non-IID situation. The larger the $\beta$ gets to, the more IID each device will be. Using Dirichlet, we can have each local device have the same total number of samples as each other, but different class distribution from each other.
\begin{figure*}[ht]
\centering
\begin{subfigure}{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/Beta=0.2.png}
\caption{$\beta=0.2$}
\label{fig:sub2A}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/Beta=0.5.png}
\caption{$\beta=0.5$}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/Beta=1.png}
\caption{$\beta=1$}
\label{fig:sub1A}
\end{subfigure}
\begin{subfigure}{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/Beta=5.png}
\caption{$\beta=5$}
\end{subfigure}
\caption{Performance comparison under different $\beta$ of non-IID scenarios.
}
\label{fig:beta_exp}
\end{figure*}
\subsection{Implementation Details}
\textbf{Main Differences Compared with MOON} Our approach is implemented as an extension to MOON with two main differences. In particular, we modify the loss function to include supervised contrastive loss for shared representations. Furthermore, we change the device-to-server communication to sharing class-wise average representations of the local model, in addition to the local model's weights. To avoid bias from limited data points, we only share the representation if the device has abundant samples for the corresponding class (i.e., more than 10 samples). Moreover, when distributing representations from the server to a device, we randomly sample $k$ representations for each class and take the average of them to represent the specific class. If the server has fewer than $k$ representations for a class, we will average everything we have to represent that class.
\textbf{Model Architecture} Considering the time limit of this study, we used a simple CNN with two convolution layers, two max-pooling layers, and two fully connected layers as the encoder. After each convolution layer and fully connected layers, we have ReLU for the activation.
\subsection{Experiment Setups}
To evaluate the performance of our method, we compared it with FedAvg and MOON, where MOON is expected to perform better than FedAvg in non-IID setting. In default, we set $\mu_{MOON}=5$ as it is the best parameter reported by the original paper \cite{moon}. Besides, we set the batch size to $64$, and use the SGD optimizer with a learning rate of $0.01$, a weight decay of $0.00001$, and a momentum of $0.9$.
Furthermore, to mimic traditional two-stage contrastive learning, we decrease the weight of the global representation contrastive loss throughout the communication rounds. Specifically, we use the following formula to control its weight, where $\mu_{glob, s}$ is the initial weight, the $\mu_{glob, e}$ is the end weight, and the $T_{0}$ is the number of warmup rounds. In our experiment, we set $\mu_{glob, s} = 1$, $\mu_{glob, e} = 0.0001$, $T=100$, and $T_0=5$.
$$
\mu_{glob, i} = \mu_{glob, s} - \frac{1}{T-T_{0}} (\mu_{glob, s} - \mu_{glob, e})
$$
To simplify our experiments, we assume that no device will reject the server's request, and we will not encounter communication failures. In other words, all devices will participate in each communication round. In our default setting, we set the number of devices to 10. We utilized $1$ GPU and $10$ CPUs for each experimental setting. The training took less than 1 day to finish for all the experiments.
\section{Results}
\begin{table}[h]
\label{results-summary}
\caption{Overall performance and efficiency for different methods.}
\begin{small}
\begin{sc}
\begin{tabular}{lcc}
\hline
Method & Top-1 Accuracy & Num of Comms (0.68 Acc)\\
\hline
FedAvg & 0.658 & more than 100\\
MOON & 0.686 & 61 \\
FedSSC & \textbf{0.693} & \textbf{41}\\
\hline
\end{tabular}
\end{sc}
\end{small}
\end{table}
\subsection{Overall Performance}
In the default setting, our method \textsc{FedSSC} outperforms \textsc{FedAvg} by $3.4\%$ and \textsc{MOON} by $0.7\%$. The improvement compared to \textsc{FedAvg} is significant, while the increase from \textsc{MOON} is smaller. However, \textsc{FedSSC} reaches the $68\%$ level of accuracy at just $41$ communication rounds, while \textsc{MOON} needs $61$ rounds, and \textsc{FedAvg} cannot reach it within $100$ rounds. In summary, our approach performs better and is more efficient than previous methods.
\begin{figure*}[ht]
\centering
\begin{subfigure}{0.3\linewidth}
\centering
\includegraphics[width=0.9\linewidth]{figs/Different_Loss_Functions.png}
\caption{Losses w/o model contrastive learning}
\label{fig:loss_exp_sub1}
\end{subfigure}
\begin{subfigure}{0.3\linewidth}
\centering
\includegraphics[width=0.9\linewidth]{figs/FedSCC_vsFedSCC_vsFedProc.png}
\caption{Losses w/ local contrastive learning}
\label{fig:loss_exp_sub2}
\end{subfigure}
\begin{subfigure}{0.3\linewidth}
\centering
\includegraphics[width=0.965\linewidth]{figs/Experiments_varied_global_representations.png}
\caption{Different $k$ of presentations shared}
\label{fig:loss_exp_sub3}
\end{subfigure}
\caption{Performance comparison for alternative loss components, where the blue line is our proposed approach. }
\label{fig:loss_exp}
\end{figure*}
\subsection{Different Non-IID Scenarios}
To better evaluate \textsc{FedSSC}, we compared its performance with other approaches under various non-IID scenarios, including $\beta \in \{0.2, 0.5, 1, 5\}$. From Figure \ref{fig:beta_exp}, we can see that as we increase $\beta$, the differences between the methods decrease since the heterogeneity is less severe. For both $\beta=0.2$ and $0.5$, \textsc{FedSSC} slightly outperforms \textsc{MOON} and is significantly better than \textsc{FedAvg}. Also, it converges faster than the others.
\subsection{Alternative Loss}
Beyond our proposed loss, we did extensive experiments with other variants that could potentially help us learn a better representation. In particular, we tried to use two-stage contrastive learning, where we first train the encoder for certain rounds while freezing the classifier in the first 90 rounds and then train the classifier while freezing the encoder in the last 10 rounds. Furthermore, we experimented with another alternative by adding negative pairs and positive pairs from the same local batch. Figure \ref{fig:loss_exp_sub1} shows that our proposed approach outperforms the alternatives a lot, especially in the earlier rounds. This is understandable because learning encoder with only supervised contrastive learning would take longer rounds to converge than using supervised classification loss. However, it could potentially outperform our current approach if we train it for more rounds.
Furthermore, we attempted to remove the $l_{MOON}$ completely by setting $\mu_{MOON}=0$ and it achieves the same level of accuracy. From Figure \ref{fig:loss_exp_sub2}, we can see that $l_{MOON}$ takes an important role. If we remove it, the performance will drop even if we share representations from all devices (i.e., \textsc{FedProc} \cite{contrafl-1}). Moreover, we experimented with different numbers of shared representations $k$ with the default $\mu_{MOON}=0$. Figure \ref{fig:loss_exp_sub3} shows that sharing representations does boost the performance, although increasing $k$ doesn't make a big difference. Therefore, the experimental results demonstrate that $l_{MOON}$ and $l_{glob}$ are complementary to each other.
\section{Related Work}
Recently many methods have been proposed to improve model accuracy and data usage at a heterogeneous distribution environment, as it has long been a key problem for Federated Learning in many fields such as finance, medicine, and social media where participants don't want to share private data.
\textbf{Federated Learning}
Based on the work of FedAvg \cite{fedavg}, many methods have been proposed to alleviate the heterogeneous distribution problem. FedProx \cite{fedprox} adds an extra regularization term to push together local model weights and global model weights. SCAFFOLD \cite{scaffold} method uses variance reduction to correct the heterogeneity during local training. SphereFed \cite{sphere} makes use of a freezed classification head to increase the similarity between global and local feature space. Generally, most of the previous methods focus on bring together local and global models.
\textbf{Contrastive learning.} Methods such as SimCLR \cite{simclr} and MOCO \cite{moco} have become promising self-supervised approaches in Computer Vision in recent years. BYOL \cite{byol} and Simsiam\cite{simsiamese} have extended the idea of contrastive learning to have zero negative samples, while SupCon \cite{supcon} proposed a supervised contrastive learning approach.
some researchers have combined contrastive learning with federated learning to mitigate the heterogeneous distribution problem\cite{moon}, \cite{contrafl-2}, \cite{contrafl-1}. Our approach is improved upon the idea of MOON \cite{moon}, where positive pair is the local representation and global representation of the same image, and the negative pair is set to be the current local representation and the local representation at the previous communication round.
\textbf{Representation sharing}. Recently, representation sharing becomes another direction to solve the problem of heterogeneous data distribution. There are quite some research works dedicated in this direction, by sharing clients' image-level or class-level representation with others. FedProc \cite{contrafl-1} proposed sharing both the local features and local model weights. The local features are averaged in a class-wise manner, which has achieved better performance than MOON or FedProx on CIFAR-10 or CIFAR-100. Another recent work, FedPCL \cite{contrafl-2} applies individual-level feature sharing on MINIST dataset.
\section{Conclusion}
In summary, our work proposes to utilize contrastive learning and representation sharing to mitigate the non-IID problem. The experiments show that our method is orthogonal to other federated learning methods, and can outperform state-of-the-art models in typical settings. Both the accuracy and convergence speed can be apparently raised. Admittedly, more experiments are needed to test the availability and performance of our method with different settings of neural network structures and datasets.
\section{Contribution Statement}
All four authors contributed equally to this work. All members participated fully in reviewing literatures, coming up with model ideas, coding different models, tuning hyperparameters and writing up the final report.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Let $[n]=\{1,2,\ldots,n\}$ be the standard $n$-element set, $2^{[n]}$ its powerset. For a family $\hf\subset 2^{[n]}$ and a subset $Y\subset [n]$ let $\hf_{\mid Y}=\{F\cap Y\colon F\in\hf\}$ denote the {\it trace} of $\hf$ on $Y$. Hajnal \cite{bondy} introduced the {\it arrow relation} $(n,m)\rightarrow (a,b)$ to denote that for all $\hf\subset 2^{[n]}$ with $|\hf|\geq m$ there exists an $a$-element set $Y\subset [n]$ such that $|\hf_{\mid Y}|\geq b$. For $\hf\subset 2^{[n]}$, let $\hf\rightarrow (a,b)$ denote that there exists an $a$-element set $Y\subset [n]$ such that $|\hf_{\mid Y}|\geq b$.
One of the most important results in extremal set theory, the Sauer-Shelah-Vapnik-Chervonenkis Theorem (\cite{S},\cite{SP},\cite{VC}) is equivalent to the arrow relation
\begin{align}\label{ineq-1.1}
\left(n,1+\sum_{i<k}\binom{n}{i}\right)\rightarrow (k,2^k) \mbox{ for all } n\geq k\geq 0.
\end{align}
Lov\'{a}sz \cite{L} conjectured and the first author \cite{F83} proved
\begin{align}\label{ineq-1.2}
\left(n,\left\lfloor \frac{n^2}{4}\right\rfloor+n+2\right)\rightarrow (3,7).
\end{align}
A family $\hf$ is called a {\it down-set} (or {\it complex}) if $F\in \hf$ always implies $2^F\subset \hf$. Both the above results are direct consequences of the following
\begin{lem}\label{lem-1.1}
If $\hf\not \rightarrow (a,b)$ for some family $\hf\subset 2^{[n]}$ then there is a down-set with the same property.
\end{lem}
\begin{example}
Let $\ell$ be a positive integer and $[n]=X_0\cup \ldots\cup X_{\ell-1}$ a partition with $|X_i|=\left\lfloor\frac{n+i}{\ell}\right\rfloor$, $0\leq i< \ell$. Define
\[
\hf(n,\ell)=\{F\subset [n]\colon|F\cap X_i|\leq 1, 0\leq i<\ell\}.
\]
Clearly, $|\hf(n,\ell)|=\prod\limits_{0\leq i<\ell}\left(1+\left\lfloor\frac{n+i}{\ell}\right\rfloor\right)$ and for $Y\in \binom{[n]}{\ell+1}$, $|\hf(n,\ell)_{\mid Y}|\leq 3\cdot 2^{\ell-1}$ is easy to verify.
In particular, $|\hf(n,2)|=\lfloor\frac{n^2}{4}\rfloor+n+1$ shows that $(n,\lfloor\frac{n^2}{4}\rfloor+n+1)\not\rightarrow (3,7)$, i.e., the corresponding arrow relation does not hold.
For general $\ell$, the example shows that
\begin{align}\label{ineq-1.3}
\left(n,\prod_{0\leq i<\ell} \left\lfloor\frac{n+\ell+i}{\ell}\right\rfloor\right)\not\rightarrow (\ell+1,3\cdot 2^{\ell-1}+1).
\end{align}
As \eqref{ineq-1.1} and \eqref{ineq-1.2} show \eqref{ineq-1.3} is best possible for $\ell=1$ and 2.
\end{example}
It is very limited evidence but let us make a conjecture for the general case.
\begin{conj}
\begin{align}\label{ineq-1.4}
\left(n,1+\prod_{0\leq i<\ell} \left\lfloor\frac{n+\ell+i}{\ell}\right\rfloor\right)\rightarrow (\ell+1,3\cdot 2^{\ell-1}+1)\mbox{ for all } n>\ell>0.
\end{align}
\end{conj}
As we will see in the next section, \eqref{ineq-1.4} is closely related to some classical results. Our main result settles the $\ell=3$ case for $n\geq 25$.
\begin{thm}\label{thm-main}
\eqref{ineq-1.4} holds for $\ell=3$ and $n\geq 25$.
\end{thm}
In view of Lemma \ref{lem-1.1} to check the veracity of \eqref{ineq-1.4} we can restrict ourselves to down-sets. Moreover, we may assume that $\hf\subset 2^{[n]}$ contains no members of size exceeding $\ell$. We shall use these facts without further mention.
We need the following notations:
$$\hf(i)=\{F\setminus\{i\}\colon i\in F\in \hf\}, \ \hf(\bar{i})= \{F\in\hf: i\notin F\}.$$
Note that $|\hf|=|\hf(i)|+|\hf(\bar{i})|$. For $i,j\in [n]$, we also use
\[
\hf(i,j)=\{F\setminus\{i,j\}\colon \{i,j\}\subset F\in \hf\}, \ \hf(\bar{i},\bar{j})=\{F\in \hf\colon F\cap \{i,j\}=\emptyset\}.
\]
For $\hf\subset 2^{[n]}$, let $\hf^{(\ell)}$ denote the subfamily $\{F\in \hf\colon |F|=\ell\}$.
\section{Cancellative families}
Let us recall that an $\ell$-graph $\hh\subset \binom{[n]}{\ell}$ is called {\it cancellative} if $\hh$ contains no three edges $H_1,H_2,H_3$ such that $|H_1\cap H_2|=\ell-1$ and $H_1\bigtriangleup H_2\subset H_3$ where $\bigtriangleup$ denotes the symmetric difference.
\begin{claim}\label{claim-2.1}
If $\hf\subset 2^{[n]}$ is a down-set and $\hf^{(\ell)}$ is not cancellative, then $\hf\rightarrow (\ell+1,3\cdot 2^{\ell-1}+1)$.
\end{claim}
\begin{proof}
Choose $F_1,F_2,F_3\in \hf^{(\ell)}$ such that $|F_1\cap F_2|=\ell-1$ and $F_1\bigtriangleup F_2 \subset F_3$. Set $Y=F_1\cup F_2$. Then $|Y|=\ell+1$ and both $2^{F_1}$ and $2^{F_2}$ are contained in $\hf_{\mid Y}$. Note that $|2^{F_1}\cup 2^{F_2}|=2\cdot 2^{\ell}-2^{\ell-1}=3\cdot 2^{\ell-1}$. Since the 2-element set $F_1\bigtriangleup F_2$ is in $2^{Y}\setminus (2^{F_1}\cup 2^{F_2})$ and $F_1\bigtriangleup F_2\subset F_3$, $F_1\bigtriangleup F_2\in \hf_{\mid Y}$ as well. Thus $|\hf_{\mid Y}|\geq 3\cdot 2^{\ell-1}+1$.
\end{proof}
The following statement was proved for $\ell=2$ by Mantel \cite{M}, for $\ell=3$ by Bollob\'{a}s \cite{bollobas} and for $\ell=4$ by Sidorenko \cite{Si}.
\begin{thm}
Let $2\leq \ell \leq 4$ and $\hh\subset \binom{[n]}{\ell}$. If $\hh$ is cancellative then
\begin{align}\label{ineq-1.5}
|\hf| \leq \prod_{0\leq i<\ell} \left\lfloor \frac{n+i}{\ell} \right\rfloor.
\end{align}
\end{thm}
Let us suppose that $\ell=3$ and $\hf\subset 2^{[n]}$ is a down-set with $\hf\not\rightarrow (4,13)$. Then $\hf^{(k)}=\emptyset$ for $k\geq 4$ and by \eqref{ineq-1.5} $|\hf^{(3)}|\leq \lfloor\frac{n+2}{3}\rfloor\lfloor\frac{n+1}{3}\rfloor\lfloor\frac{n}{3}\rfloor$. Consequently,
\[
|\hf| \leq \left\lfloor\frac{n+2}{3}\right\rfloor\left\lfloor\frac{n+1}{3}\right\rfloor\left\lfloor\frac{n}{3}\right\rfloor
+\binom{n}{2}+\binom{n}{1}+\binom{n}{0}.
\]
That is,
\begin{align}
\left(n,\left\lfloor\frac{n+2}{3}\right\rfloor\left\lfloor\frac{n+1}{3}\right\rfloor\left\lfloor\frac{n}{3}\right\rfloor
+\binom{n}{2}+n+2\right)\rightarrow (4,13).
\end{align}
This shows that \eqref{ineq-1.4} is ``asymptotically" true for $\ell=3$.
Similarly, the $\ell=4$ case of \eqref{ineq-1.5} and Lemma \ref{lem-1.1} imply
\begin{align}
\left(n,\prod_{0\leq i<4}\left\lfloor\frac{n+i}{4}\right\rfloor+\binom{n}{3}+\binom{n}{2}+n+2\right)\rightarrow (5,25).
\end{align}
Unfortunately, \eqref{ineq-1.5} is no longer true for $\ell\geq 5$. In particular for $\ell=5$ and 6 Frankl and F\"{u}redi \cite{FF} showed that the maximum possible size $m(n,\ell)$ of a cancellative family $\hf\subset \binom{[n]}{\ell}$ satisfies
\begin{align*}
&m(n,5) \leq \frac{6}{11^4}n^5 \mbox{ with equality iff } 11|n \mbox{ and }\\[5pt]
&m(n,6) \leq \frac{11}{12^5}n^6 \mbox{ with equality iff } 12|n,
\end{align*}
which is much larger than $(n/\ell)^{\ell}$.
Let us define $m^*(n,\ell)$ as the maximum size of $\hf\subset \binom{[n]}{\ell}$ where $\hf$ contains no three distinct edges satisfying $F_1\bigtriangleup F_2\subset F_3$. Unlike with cancellative families, we do not require $|F_1\cap F_2|=\ell-1$. Thus $m^*(n,\ell)\leq m(n,\ell)$.
Katona conjectured $m^*(n,\ell)=\prod\limits_{0\leq i<\ell} \left\lfloor\frac{n+i}{\ell}\right\rfloor$. However, Shearer \cite{Sh} disproved this conjecture for $\ell>10$.
\section{Proof of Theorem \ref{thm-main}}
We need the following inequality.
\begin{lem}
Let $a_1,a_2,\ldots,a_m\geq 0$. Then
\begin{align}\label{ineq-key0}
\prod_{1\leq i<j\leq m} a_ia_j \leq \frac{m-1}{2m}\left(\sum_{1\leq i\leq m}a_i\right)^2.
\end{align}
\end{lem}
\begin{proof}
Note that
\begin{align}\label{ineq-3.3}
\prod_{1\leq i<j\leq m} a_ia_j =\frac{1}{2}\left(\left(\sum_{1\leq i\leq m}a_i\right)^2-\sum_{1\leq i\leq m} a_i^2\right).
\end{align}
Since $x^2$ is convex, by Jensen's inequality
\[
\frac{1}{m}\sum_{1\leq i\leq m} a_i^2\geq \left(\frac{1}{m}\sum_{1\leq i\leq m} a_i\right)^2=\frac{1}{m^2}\left(\sum_{1\leq i\leq m} a_i\right)^2.
\]
It follows that
\[
\sum\limits_{1\leq i\leq m} a_i^2 \geq \frac{1}{m}\left(\sum\limits_{1\leq i\leq m} a_i\right)^2.
\]
By \eqref{ineq-3.3} we conclude that \eqref{ineq-key0} holds.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm-main}]
Let $\hf\subset 2^{[n]}$ be a down-set satisfying $\hf\not \rightarrow (4,13)$ and $|\hf|$ is maximal. Clearly, $|\hf|\geq |\hf(n,3)|= \lfloor\frac{n+3}{3}\rfloor\lfloor\frac{n+4}{3}\rfloor\lfloor\frac{n+5}{3}\rfloor$.
We showed that $\hf^{(3)}$ is cancellative however we are not going to use the bound \eqref{ineq-1.5}.
\begin{claim}\label{claim-3.1}
Let $\hf'$ be a family obtained from $\hf$ by removing all edges $F\in \hf$ with $y\in F$ and adding the edges $\{y\}\cup G$ for $G\in \hf(x,\bar{y})$. Then $\hf'\not\rightarrow (4,13)$.
\end{claim}
\begin{proof}
Indeed, otherwise let $C$ be a 4-set satisfying $|\hf'_{\mid C} |\geq 13$. Then clearly $y\in C$. If $x\in C$, then by $\hf'(x,y)=\emptyset$
\[
|\hf'_{\mid C} | \leq 2^{|C\setminus\{x\}|}+2^{|C\setminus\{y\}|}-2^{|C\setminus\{x,y\}|}=2^3+2^3-2^2=12,
\]
a contradiction. Thus $x\notin C$. Setting $C'=(C\setminus \{y\})\cup \{x\}$, $|\hf_{\mid C'}|=|\hf'_{\mid C}|\geq 13$, a contradiction again.
\end{proof}
There are two simple conditions to guarantee for a 4-set $C$ (with respect to a family $\hf$) that $|\hf_{\mid C} |\leq 12$.
\begin{itemize}
\item[(i)] $\exists \{x,y\}\in \binom{C}{2}$ such that no $F\in \hf$ contains $\{x,y\}$.
\item[(ii)] $\exists \{x,y\}\in \binom{C}{2}$ such that $\hf(x)=\hf(y)$.
\end{itemize}
Note that if $\{x,y\}\subset F\in \hf$ then $F\setminus \{x\}\in \hf(x)$ but $F\setminus \{x\}\notin \hf(y)$. Thus (ii) implies (i).
In view of these conditions if $\{x,y\}\not\subset F$ for all $F\in \hf$ then we can symmetrize $\hf$ by removing all $F\in \hf$ with $y\in F$ and adding all $\{y\}\cup G$ with $G\in \hf(x)$. Thereby $\hf(x)=\hf(y)$ for the new family. By Claim \ref{claim-3.1} the new family preserves the property $\hf\not\rightarrow (4,13)$. If $|\hf(x)|\geq |\hf(y)|$ then the new family has at least as many members as the old one. Thus we may assume that for all distinct $x,y\in [n]$ either $\exists F\in \hf$ with $\{x,y\}\subset F$ or $\hf(x)=\hf(y)$.
It is easy to see that $\hf(x)=\hf(y)$ is an equivalence relation. Thus we get a partition $[n]=Z_1\cup Z_2\cup \ldots\cup Z_r$ and an auxiliary family $\hh\subset 2^{[r]}$ such that each $Z_i$ is an equivalence class, $F\in \hf$ iff $|F\cap Z_i|\leq 1$ for all $i$ and $\{i\colon F\cap Z_i\neq \emptyset\}\in \hh$. Let us choose $\hf$ such that $r$ is minimal over all families $\hf$ with $\hf\not\rightarrow (4,13)$ and $|\hf|$ maximal.
Note that $\hf(x)=\hf(y)$ forces that $x$ and $y$ are in the same $Z_i$. Hence if $1\leq i<i'\leq r$, $x\in Z_i$, $y\in Z_{i'}$ then $\{x,y\}\subset F$ for some $F\in \hf$. Consequently, $\binom{[r]}{2}\subset \hh$.
\begin{claim}\label{claim-2.4}
If $H,H'\in \hh^{(3)}$ then $|H\cap H'|\leq 1$.
\end{claim}
\begin{proof}
Suppose the contrary. WLOG $H=(1,2,3)$, $H'=(1,2,4)$. Since $(3,4)\in \hh$, $|\hf{\mid_C}|\geq 13$ for the corresponding $C=\{z_1,z_2,z_3,z_4\}$ (where $z_i\in Z_i$), a contradiction.
\end{proof}
Let $b_i=|Z_i|$, $i=1,2,\ldots,r$. If $r=3$, then the theorem follows from the fact that $(b_1+1)(b_2+1)(b_3+1)$ is maximized when $b_1=\lfloor\frac{n+2}{3}\rfloor$, $b_2=\lfloor\frac{n+1}{3}\rfloor$ and $b_3=\lfloor\frac{n}{3}\rfloor$. Thus in the rest of the proof we assume $r\geq 4$.
\begin{claim}
For $x\in Z_i$, $y\in Z_j$ with $i\neq j$,
\begin{align}\label{ineq-key3}
|\hf(x,\bar{y})|< |\hf(y)|.
\end{align}
\end{claim}
\begin{proof}
If $|\hf(x,\bar{y})|\geq |\hf(y)|$ then for every $y\in Z_j$ we remove all edges $F\in \hf$ with $y\in F$ and add the edges $\{y\}\cup G$ for $G\in \hf(x,\bar{y})$. By Claim \ref{claim-3.1} the new family $\hf'$ satisfies $\hf'\not\rightarrow (4,13)$ and $|\hf'|\geq |\hf|$. However, $\hf'$ has $r-1$ classes, contradicting the minimality of $r$.
\end{proof}
\begin{claim}\label{claim-3.4}
There exists $z\in [n]$ such that
\begin{align}\label{ineq-key6}
|\hf^{(3)}(z)| > \frac{n^2}{9}-\frac{n}{2}+1.
\end{align}
\end{claim}
\begin{proof}
Let $z\in [n]$ be a vertex with $|\hf^{(3)}(z)|$ maximal. Note that $\hf=\hf^{(3)}\cup \hf^{(2)}\cup \hf^{(1)}\cup\hf^{(0)}$. Since $|\hf^{(0)}|+|\hf^{(1)}|=n+1$
and $|\hf^{(2)}|\leq \binom{n}{2}$,
\[
|\hf^{(3)}| =|\hf|-|\hf^{(0)}|-|\hf^{(1)}|-|\hf^{(2)}|\geq \left\lfloor\frac{n+3}{3}\right\rfloor\left\lfloor\frac{n+4}{3}\right\rfloor\left\lfloor\frac{n+5}{3}\right\rfloor -n-1-\binom{n}{2}.
\]
It follows that
\begin{align*}
|\hf^{(3)}(z)| \geq \frac{ 3|\hf^{(3)}| }{n}\geq \frac{3}{n} \left\lfloor\frac{n+3}{3}\right\rfloor\left\lfloor\frac{n+4}{3}\right\rfloor\left\lfloor\frac{n+5}{3}\right\rfloor
-\frac{3n}{2}-\frac{3}{2}-\frac{3}{n}.
\end{align*}
For $n=3t$,
\[
|\hf^{(3)}(z)| \geq\frac{3}{n}\frac{(n+3)^3}{27}
-\frac{3n}{2}-\frac{3}{2}-\frac{3}{n} = \frac{n^2}{9} -\frac{n}{2}+\frac{3}{2}.
\]
For $n=3t+1$ and $n\geq 8$,
\begin{align*}
|\hf^{(3)}(z)| \geq\frac{3}{n}\frac{(n+2)^2(n+5)}{27}
-\frac{3n}{2}-\frac{3}{2}-\frac{3}{n} &= \frac{n^2}{9} -\frac{n}{2}+\frac{7}{6}-\frac{7}{9n}>\frac{n^2}{9}-\frac{n}{2}+1.
\end{align*}
For $n=3t+2$ and $n\geq 8$,
\begin{align*}
|\hf^{(3)}(z)| \geq\frac{3}{n}\frac{(n+1)(n+4)^2}{27}
-\frac{3n}{2}-\frac{3}{2}-\frac{3}{n} &= \frac{n^2}{9} -\frac{n}{2}+\frac{7}{6}-\frac{11}{9n}>\frac{n^2}{9}-\frac{n}{2}+1.
\end{align*}
\end{proof}
Let $x\in Z_i$ and assume $z\in Z_j$. If $i=j$ then clearly $\hf^{(3)}(x)=\hf^{(3)}(z)$. If $i\neq j$ then $|\hf(z,x)|\leq n-b_i-b_j+1$. By \eqref{ineq-key3},
\[
|\hf(z)|-|\hf(x)|\leq |\hf(z,x)|+|\hf(z,\bar{x})|-|\hf(x)| \leq |\hf(x,z)|-1\leq n-b_i-b_j.
\]
Since $|\hf^{(2)}(x)|=n-b_i$ and $|\hf^{(2)}(z)|=n-b_j$,
\begin{align*}
|\hf^{(3)}(z)|-|\hf^{(3)}(x)|&=(|\hf(z)|-|\hf^{(2)}(z)|-1)-(|\hf(x)|-|\hf^{(2)}(x)|-1)\nonumber\\[5pt]
&\leq n-b_i-b_j+(b_j-b_i)\\[5pt]
&=n-2b_i.
\end{align*}
By \eqref{ineq-key6} and $n\geq 25$, it follows that for all $x\in [n]$
\begin{align}\label{ineq-key7}
|\hf^{(3)}(x)| \geq |\hf^{(3)}(z)|-(n-2b_i)> \frac{n^2}{9}-\frac{3n}{2}+1+2b_i\geq \frac{n^2}{9}-\frac{3n}{2}+3\geq \frac{n^2}{18}.
\end{align}
If $r=4$, then by Claim \ref{claim-2.4} we may assume that $\hh^{(3)}=\{(1,2,3)\}$. Then $\hf^{(3)}(x)=\emptyset$ for all $x\in Z_4$, contradicting \eqref{ineq-key7}.
Let us fix $x_i\in Z_i$, $i=1,2,\ldots,r$.
\begin{claim}\label{claim-3.5}
$r\neq 5$.
\end{claim}
\begin{proof}
By Claim \ref{claim-2.4} and symmetry, we may assume that $\hh^{(3)}\subset\{(1,2,3),(1,4,5)\}$, $b_2\geq b_3$ and $b_4\geq b_5$.
Then
\[
|\hf|=b_1(b_2b_3+b_4b_5) +\sum_{1\leq i<j\leq 5}b_ib_j+n+1.
\]
Let $\hf'$ be the family obtained from $\hf$ by merging $Z_2$ and $Z_5$, $Z_3$ and $Z_4$. Then
\[
|\hf'|=b_1(b_2+b_5)(b_3+b_4)+b_1(b_2+b_5+b_3+b_4)+(b_2+b_5)(b_3+b_4)+n+1.
\]
Using $b_1\geq 1$, we obtain that
\[
|\hf'|-|\hf| =b_1(b_2b_4+b_3b_5)-b_2b_5-b_3b_4\geq (b_2-b_3)(b_4-b_5)\geq 0.
\]
Clearly $\hf'\not\rightarrow (4,13)$ and $\hf'$ is 3-partite. This contradicts the minimality of $r$.
\end{proof}
\begin{claim}\label{claim-3.6}
$r\neq 6$.
\end{claim}
\begin{proof}
If there are two disjoint edges in $\hh^{(3)}$, then by Claim \ref{claim-2.4} $|\hh^{(3)}|=2$. Without loss of generality, assume that $\hh^{(3)}=\{(1,2,3),(4,5,6)\}$ and $b_4+b_5+b_6\leq \frac{n}{2}$. Then by \eqref{ineq-key7}
\begin{align}\label{ineq-key8}
|\hf^{(3)}(x_4)|+|\hf^{(3)}(x_5)|+|\hf^{(3)}(x_6)|=b_5b_6+b_4b_6+ b_4b_5 >\frac{n^2}{6}.
\end{align}
By \eqref{ineq-key0}, we infer that
\[
b_4b_5+b_4b_6+b_5b_6 \leq \frac{(b_4+b_5+b_6)^2}{3} \leq \frac{n^2}{12},
\]
contradicting \eqref{ineq-key8}.
Thus $|H\cap H'|=1$ for all $H,H'\in \hh^{(3)}$. Up to isomorphism there is only one triple-system with four triples on six vertices. By symmetry we may assume that
\[
\hh^{(3)}\subset\{(1,3,5),(1,4,6),(2,3,6),(2,4,5)\}.
\]
Then
\begin{align*}
&b_1b_5+b_2b_6\geq |\hf^{(3)}(x_3)|> \frac{n^2}{18}, &b_1b_6+b_2b_5\geq |\hf^{(3)}(x_4)| > \frac{n^2}{18},\\[5pt]
&b_1b_3+b_2b_4\geq |\hf^{(3)}(x_5)|> \frac{n^2}{18}, &b_1b_4+b_2b_3 \geq |\hf^{(3)}(x_6)|> \frac{n^2}{18}.
\end{align*}
Adding these inequalities, we get
\begin{align}\label{ineq-3.1}
(b_1+b_2)(b_3+b_4+b_5+b_6) >\frac{2n^2}{9}.
\end{align}
Moreover,
\begin{align*}
&b_3b_5+b_4b_6\geq |\hf^{(3)}(x_1)|> \frac{n^2}{18}, &b_3b_6+b_4b_5\geq |\hf^{(3)}(x_2)| > \frac{n^2}{18}.
\end{align*}
It implies that
\begin{align}\label{ineq-3.2}
(b_3+b_4)(b_5+b_6) >\frac{n^2}{9}.
\end{align}
Note that $b_1+b_2+b_3+b_4+b_5+b_6=n$. If $b_1+b_2\geq \frac{n}{3}$, then $b_3+b_4+b_5+b_6\leq \frac{2n}{3}$. It follows that $(b_3+b_4)(b_5+b_6) \leq \frac{n^2}{9}$, contradicting \eqref{ineq-3.2}. If $b_1+b_2< \frac{n}{3}$, then
\[
(b_1+b_2)(b_3+b_4+b_5+b_6)<\frac{2n}{9},
\]
contradicting \eqref{ineq-3.1}.
\end{proof}
For $H\in 2^{[r]}$, let $b_H=\prod_{i\in H} b_i$.
\begin{claim}
$r=7$.
\end{claim}
\begin{proof}
Suppose that $r\neq 7$. Then by Claims \ref{claim-3.5} and \ref{claim-3.6}, $r\geq 8$. For each $i=1,2,\ldots,r$,
\[
|\hf^{(3)}(x_i)| =\sum_{P\in \hh^{(3)}(i)} b_P>\frac{n^2}{18}.
\]
By Claim \ref{claim-2.4}, $\hf^{(3)}(x_i)\cap \hf^{(3)}(x_j)=\emptyset$ for all $1\leq i<j\leq r$. By \eqref{ineq-key0}, we obtain that
\begin{align}\label{ineq-key10}
\frac{rn^2}{18}< \sum_{1\leq i\leq r}|\hf^{(3)}(x_i)| \leq \sum_{1\leq i<j\leq r} b_ib_j \leq \frac{r-1}{2r}\left(b_1+b_2+ \ldots+ b_r\right)^2=\frac{r-1}{2r}n^2.
\end{align}
It follows that
\[
\frac{r}{9}< 1- \frac{1}{r},
\]
which leads to a contradiction for $r\geq 8$.
\end{proof}
Now we assume that $r=7$.
\begin{claim}\label{claim-2.5}
For $n\geq 17$, $\max\limits_{1\leq i\leq 7} b_i\leq n/2$.
\end{claim}
\begin{proof}
Assume that $b_1\geq b_2\geq \ldots \geq b_7$. By \eqref{ineq-key7}, for all $x\in [n]$
\[
|\hf^{(3)}(x)| > \frac{n^2}{9}-\frac{3n}{2}+3.
\]
It is easy to check that for $n\geq 17$ the RHS is greater than $\frac{n^2}{32}$. Assume that $b_1\geq \frac{n}{2}$ and we distinguish two cases.
{\bf \noindent Case 1. } $(1,2,3)\in \hh$.
Then
\[
|\hf^{(3)}(x_3)|\geq b_1b_2,\ |\hf^{(2)}(x_3)|=n-b_3,\ |\hf(x_1,x_3)|= b_2+1,
\]
and
\[
|\hf^{(3)}(x_1)| \leq b_2b_3+\left(\frac{n-b_1-b_2-b_3}{2}\right)^2,\ |\hf^{(2)}(x_1)|=n-b_1.
\]
By \eqref{ineq-key3}, $|\hf(x_3,\overline{x_1})|< |\hf(x_1)|$. It follows that
\[
b_1b_2+n-b_3-(b_2+1)\leq b_2b_3+n-b_1+\left(\frac{n-b_1-b_2-b_3}{2}\right)^2.
\]
Equivalently,
\begin{align}\label{ineq-newkey1}
b_2(b_1-b_3)+b_1-b_2-b_3-1\leq \left(\frac{n-b_1-b_2-b_3}{2}\right)^2.
\end{align}
Note that $b_1\geq \frac{n}{2}$ implies $b_1\geq b_2+b_3$. If $b_1=b_2+b_3$ then $b_1+b_2+b_3=n$ and \eqref{ineq-newkey1} cannot hold. Thus $b_1>b_2+b_3$. Then \eqref{ineq-newkey1} implies
\begin{align}\label{ineq-newkey2}
b_2(b_1-b_3)\leq \frac{n-b_1-b_2-b_3}{4} (n-b_1-b_2-b_3).
\end{align}
If $b_2\geq \frac{n-b_1-b_2-b_3}{4}$, then \eqref{ineq-newkey2} implies
\[
b_1-b_3\leq n-b_1-b_2-b_3.
\]
It follows that $2b_1+b_2\leq n$, contradicting $b_1\geq \frac{n}{2}$. Thus $b_2< \frac{n-b_1-b_2-b_3}{4}$. That means $5b_2+b_3<n-b_1$. Then $6\frac{b_2+b_3}{2} \leq n-b_1\leq \frac{n}{2}$. It implies $\frac{b_2+b_3}{2}\leq \frac{n}{12}$. Therefore
\[
|\hf^{(3)}(x_1)|\leq \left(\frac{n}{12}\right)^2+\left(\frac{n}{12}\right)^2+\left(\frac{n}{12}\right)^2<\frac{n^2}{32},
\]
a contradiction.
{\bf \noindent Case 2. } $(1,2,3)\notin \hh$.
Then
\[
|\hf^{(3)}(x_1)|\leq b_2b_4+b_3b_5+b_6b_7.
\]
The maximum should be for $b_6=b_7=0$, $b_3=b_4$. Set $b_2=\alpha n$, $b_3=b_4=\beta n$, $b_5=\gamma n$, then
\begin{align}\label{ineq-newkey8}
\frac{|\hf^{(3)}(x_1)|}{n^2}\leq \alpha \beta +\beta \gamma =\frac{1}{2}[2\beta(\alpha+\gamma)] \leq \frac{1}{2}\left(\frac{2\beta+\alpha+\gamma}{2}\right)^2=\frac{1}{2}\left(\frac{n-b_1}{2n}\right)^2.
\end{align}
The RHS is at most $\frac{1}{2}\left(\frac{1}{4}\right)^2=\frac{1}{32}$ for $b_1\geq \frac{n}{2}$. Thus we get $|\hf^{(3)}(x_1)|\leq \frac{n^2}{32}$, a contradiction.
\end{proof}
By \eqref{ineq-key10}, we have
\begin{align}\label{ineq-key11}
\sum_{1\leq i\leq 7} |\hf^{(3)}(x_i)| \leq \frac{3n^2}{7}.
\end{align}
By \eqref{ineq-key6},
\[
|\hf^{(3)}(z)| > \frac{n^2}{9}-\frac{n}{2}+1.
\]
Assume $z\in Z_j$. Then Claim \ref{claim-2.5} implies $n-b_j\geq \frac{n}{2}$. By \eqref{ineq-key7}, we obtain that
\begin{align*}
\sum_{1\leq i\leq 7} |\hf^{(3)}(x_i)| &= |\hf^{(3)}(z)|+\sum_{i\neq j} \left(\frac{n^2}{9}-\frac{3n}{2}+1+2b_i\right)\\[5pt]
&> \frac{n^2}{9}-\frac{n}{2}+1+ 6\left(\frac{n^2}{9}-\frac{3n}{2}+1\right)+2(n-b_j)\\[5pt]
&\geq \frac{7n^2}{9}-\frac{19n}{2}+7+n\\[5pt]
&\geq \frac{7n^2}{9}-\frac{17n}{2}+7.
\end{align*}
It is easy to check that the RHS is greater than $\frac{3n^2}{7}$ for $n\geq 24$, contradicting \eqref{ineq-key11}.
\end{proof}
\section{Other results for $(n,m)\rightarrow (4,b)$}
Let us introduce the general notation
\[
m(n,a,b)=\min\left\{m\colon (n,m)\rightarrow (a,b)\right\}.
\]
In this section we consider $m(n,4,b)$ for $b\leq 16$. With this notation the Sauer-Shelah-Vapnik-Chervonenkis Theorem is equivalent to $m(n,4,16)=1+\sum\limits_{0\leq i\leq 3} \binom{n}{i}$.
For $5<b<16$, it is easy to see that all extremal families satisfying $\hf\not\rightarrow (4,b)$ span $[n]$, i.e., $\cup_{F\in \hf} F=[n]$. Hence $\binom{[n]}{\leq 1}\subset \hf$. This motivates us to introduce the following auxiliary definitions. Set $\tilde{\hf}=\hf^{(2)}\cup \hf^{(3)}$ and say that $\tilde{\hf}$ is {\it complete} if $\partial \hf^{(3)}\subset \hf^{(2)}$, that is, if $P\subset T\in \hf^{(3)}$ and $|P|=2$ then $P\in \hf^{(2)}$.
Let us introduce the notation $\tilde{\hf}\hookrightarrow (4,c)$ if there exists a 4-set $C$ with $|\hf^{(2)}\cap \binom{C}{2}|+|\hf^{(3)}\cap \binom{C}{3}|\geq c$. If $\hf$ is a down-set with $|F|\leq 3$ for all $F\in \hf$ and $\cup \hf=[n]$ then $\tilde{\hf}\hookrightarrow (4,c)$ is equivalent to $\hf\rightarrow (4,c+5)$.
Finally, for $1\leq c<11$ we introduce the notation
\[
\tilde{m}(n,4,c)=\min\left\{\tilde{m}\colon |\tilde{\hf}|\geq \tilde{m} \mbox{ implies } \tilde{\hf} \hookrightarrow (4,c) \mbox{ for a complete family }\tilde{\hf}\subset 2^{[n]} \right\}.
\]
Clearly, $\tilde{m}(n,4,c)=m(n,4,c+5)-n-1$.
Let $\hht(r,n)$ be a complete $r$-partite graph on $n$ vertices with each part of size $\lfloor \frac{n}{r} \rfloor$ or $\lceil \frac{n}{r} \rceil$ and let $t(r,n)$ be the number of edges in $\hht(r,n)$. We have the following results.
\begin{table}[H]
\centering
\caption{$\tilde{m}(n,4,c)$ for $1\leq c\leq 8$, $n\geq 5$}
\begin{tabular}{c}
\toprule
$\tilde{m}(n,4,1)=1$\\[3pt]
\midrule
$\tilde{m}(n,4,2)=2$\\[3pt]
\midrule
$\tilde{m}(n,4,3)=\lfloor \frac{2}{3}n\rfloor+1$\\[3pt]
\midrule
$\left(\frac{n}{2}\right)^{3/2}+o(n^{3/2})\leq \tilde{m}(n,4,4)\leq \frac{1}{2}n^{3/2}+O(n)$\\[3pt]
\midrule
$\tilde{m}(n,4,5)=\lfloor\frac{n^2}{4}\rfloor+1$\\[3pt]
\midrule
$\tilde{m}(n,4,6)=t(3,n)+1$\\[3pt]
\midrule
$\tilde{m}(n,4,7)=\binom{n}{2}+1$ for $n\neq 6$, $\tilde{m}(6,4,7) = 17$\\[3pt]
\midrule
$\tilde{m}(n,4,8)=\left\lfloor\frac{n+2}{3}\right\rfloor\left\lfloor\frac{n+1}{3}\right\rfloor\left\lfloor\frac{n}{3}\right\rfloor+1$ for $n\geq 25$\\[3pt]
\bottomrule
\end{tabular}
\end{table}
The cases $c=9, 10$ will be discussed later. To prove the above statements for each particular choice of $c$, we assume that $\tilde{\hf}=\hf^{(2)}\cup \hf^{(3)}$ is a complete family with $\tilde{\hf}\not\hookrightarrow (4,c)$.
Since $\hf^{(3)}\neq \emptyset$ forces $\tilde{\hf}\hookrightarrow (4,4)$, for the case $1\leq c\leq 4$ we may assume $\hf^{(3)}=\emptyset$. Then $\tilde{m}(n,4,c)=c$ is trivial for $c=1$ and $c=2$.
Let $c=3$. Consider $\hf^{(2)}$, a graph in which no four vertices span more than 2 edges. Thus $\hf^{(2)}$ has maximum degree at most two and without a path or cycle of length three. Hence each connected component of $\hf^{(2)}$ is a single edge or a path of length two. Consequently, $|\hf^{(2)}|\leq \frac{2}{3}n$, proving $\tilde{m}(n,4,3)=\lfloor \frac{2}{3}n\rfloor+1$.
For $c=4$, $\hf^{(2)}$ is a graph that contains no subgraph on 4 vertices with 4 or more edges. Let $C_3^+$ be a triangle plus a pendant edge. It follows that $\hf^{(2)}$ is $C_3^+$-free and $C_4$-free. Consequently if $\hf^{(2)}$ contains a triangle, then it is a connected component. It follows that each connected component of $\hf^{(2)}$ with at least 4 vertices is $\{C_3,C_4\}$-free. For a given family $\mathscr{F}$ of graphs, let $ex(n, \mathscr{F})$ denote the maximum number of edges in an $n$-vertex graph which does not contain any member in $\mathscr{F}$ as its subgraph. The {\it Zarankiewicz number} $z(n, C_4)$ is the maximum number of edges in an $n$-vertex bipartite graph without containing a $C_4$. It is well known that $z(n, C_4)=\left(\frac{n}{2}\right)^{3/2}+o(n^{3/2})$ (see \cite{DHS}, \cite{FS}). Since bipartite graphs are $C_3$-free, we see that $ex(n, \{C_3,C_4\})\geq z(n, C_4)$.
Erd\H{o}s-R\'{e}nyi-S\'{o}s \cite{ERS} and Brown \cite{brown} showed that $ex(q^2+q+1,C_4)\geq \frac{1}{2}q(q+1)^2$ for all prime powers $q$. F\"{u}redi \cite{Furedi1,Furedi2} proved that $ex(q^2+q+1,C_4)=\frac{1}{2}q(q+1)^2$ for all prime powers $q\geq 14$. As
it is shown in \cite{KST} this implies $ex(n,C_4)=\frac{1}{2}n^{3/2}+O(n)$ all $n$. Thus, $ex(n, \{C_3,C_4\})\leq ex(n,C_4)=\frac{1}{2}n^{3/2}+O(n)$. These results imply that $\left(\frac{n}{2}\right)^{3/2}+o(n^{3/2})\leq \tilde{m}(n,4,4)\leq \frac{1}{2}n^{3/2}+O(n)$.
\begin{prop}
\[
\tilde{m}(n,4,5) = \left\lfloor\frac{n^2}{4}\right\rfloor+1.
\]
\end{prop}
\begin{proof}
Note that $\hht(2,n)\not\hookrightarrow (4,5)$. This shows that $\tilde{m}(n,4,5)\geq \lfloor\frac{n^2}{4}\rfloor+1$.
Let $\tilde{\hf}=\hf^{(2)}\cup \hf^{(3)}$ be a complete family satisfying $\tilde{\hf}\not\hookrightarrow (4,5)$.
We prove $|\tilde{\hf}|\leq \lfloor\frac{n^2}{4}\rfloor$ by induction on $n$. Clearly it holds for $n=4$. Now we assume that it holds for $4,5,\ldots,n-1$ and prove it for $n$. If there exists $F_0\in \tilde{\hf}$ with $|F_0|=3$, then for any $y\in [n]\setminus F_0$, $\{x,y\}\notin \tilde{\hf}$ for all $x\in F_0$. It follows that $|\tilde{\hf}|\leq \lfloor\frac{(n-3)^2}{4}\rfloor+4\leq \lfloor\frac{n^2}{4}\rfloor$. Thus we may assume that $\hf^{(3)}=\emptyset$.
If there are two triangles with a common edge in $\hf^{(2)}$, let $C$ be the set of these 4 vertices. Then $|\tilde{\hf}_{\mid C}|\geq 5$, a contradiction. Thus $\hf^{(2)}$ contains no two triangles with a common edge. Assume $\hf^{(2)}$ contains a triangle, say $\{x_1,x_2,x_3\}$. Then each $y\in [n]\setminus \{x_1,x_2,x_3\}$ has at most one neighbor in $\{x_1,x_2,x_3\}$. Therefore,
\[
|\{P\in \hf^{(2)}\colon P\cap \{x_1,x_2,x_3\}\neq \emptyset\}| \leq 3+(n-3)=n.
\]
Using the induction hypothesis, it follows that $|\tilde{\hf}|\leq \lfloor\frac{(n-3)^2}{4}\rfloor+n\leq \lfloor\frac{n^2}{4}\rfloor$. Finally if $\hf^{(2)}$ is triangle-free, then by Mantel's theorem \cite{M}, $|\tilde{\hf}|= |\hf^{(2)}|\leq \lfloor\frac{n^2}{4}\rfloor$.
\end{proof}
\begin{prop}
\[
\tilde{m}(n,4,6) = t(3,n)+1.
\]
\end{prop}
\begin{proof}
Clearly $\hht(3,n)\not\hookrightarrow (4,6)$. We see that $\tilde{m}(n,4,6)\geq t(3,n)+1$.
Let $\tilde{\hf}=\hf^{(2)}\cup \hf^{(3)}$ be a complete family satisfying $\tilde{\hf}\not\hookrightarrow (4,6)$.
We prove $|\tilde{\hf}|\leq t(3,n)$ by induction on $n$. Clearly it holds for $n=4$. Now assume that it holds for $4,5,\ldots,n-1$ and we prove it for $n$. If there exists $F_0=\{x_1,x_2,x_3\}\in \tilde{\hf}$, then by $\tilde{\hf}\not\hookrightarrow (4,6)$ for every $y\in [n]\setminus F$ at most one of $\{x_1,y\},\{x_2,y\},\{x_3,y\}$ is in $\tilde{\hf}$. It follows that
\[
|\{P\in \hf^{(2)}\colon P\cap F_0\neq \emptyset\}|\leq 3+n-3=n.
\]
Note that $\tilde{\hf}\not\hookrightarrow (4,6)$ implies $|F\cap F'|\leq 1$ for all distinct $F,F'\in \hf^{(3)}$. We infer that $\hf^{(3)}(x_i)$ is a matching. Let $\hg(x_i)=\hf^{(3)}(x_i)\cap \binom{[n]\setminus F_0}{2}$, $i=1,2,3$.
We claim that for $1\leq i<j\leq 3$, $\hg(x_i)$ and $\hg(x_j)$ are disjoint. For otherwise since $\tilde{\hf}$ is complete, we shall find $y\in [n]\setminus F_0$ such that two of $\{x_1,y\},\{x_2,y\},\{x_3,y\}$ are in $\tilde{\hf}$, a contradiction. Hence $\hg(x_1)\cup \hg(x_2)\cup \hg(x_3)$ is a matching. Therefore,
\[
|\{F\in \hf^{(3)}\colon F\cap F_0\neq \emptyset\}|\leq 1+\left\lfloor\frac{n-3}{2}\right\rfloor=\left\lfloor\frac{n-1}{2}\right\rfloor.
\]
By the induction hypothesis,
\begin{align}\label{ineq-4.2}
|\tilde{\hf}|\leq t(3,n-3)+n+\left\lfloor\frac{n-1}{2}\right\rfloor.
\end{align}
\begin{claim}
\begin{align}\label{ineq-4.1}
t(3,n)-t(3,n-3)=2n-3.
\end{align}
\end{claim}
\begin{proof}
Note that
\[
t(3,n) =\binom{n}{2}-\binom{\lfloor\frac{n}{3}\rfloor}{2}-\binom{\lfloor\frac{n+1}{3}\rfloor}{2}-\binom{\lfloor\frac{n+2}{3}\rfloor}{2}.
\]
Then
\begin{align*}
t(3,n)- t(3,n-3)&=\binom{n}{2}-\binom{n-3}{2}-\left(\left\lfloor\frac{n}{3}\right\rfloor-1+\left\lfloor\frac{n+1}{3}\right\rfloor-1
+\left\lfloor\frac{n+2}{3}\right\rfloor-1\right)\\[5pt]
&=(3n-6)-(n-3)=2n-3.
\end{align*}
\end{proof}
Since $n\geq 5$ implies $n+\lfloor\frac{n-1}{2}\rfloor\leq 2n-3$, by \eqref{ineq-4.2} and \eqref{ineq-4.1} we obtain $|\tilde{\hf}|\leq t(3,n)$.
Thus we may assume that $\hf^{(3)}=\emptyset$.
Since $\tilde{\hf}\not\hookrightarrow (4,6)$ implies $\hf^{(2)}$ is $K_4$-free, by Tur\'{a}n's Theorem \cite{turan} $|\tilde{\hf}|= |\hf^{(2)}| \leq t(3, n)$.
\end{proof}
\begin{prop}
$\tilde{m}(n,4,7) = \binom{n}{2}+1$ for $n\neq 6$ and $\tilde{m}(6,4,7) = 17$.
\end{prop}
\begin{proof}
Note that $\binom{[n]}{2}\not\hookrightarrow (4,7)$. It follows that $\tilde{m}(n,4,7)\geq \binom{n}{2}+1$.
For $n=6$, define $\hf^{(3)} =\{ \{1,3,5\}, \{1,4,6\}, \{2,3,6), \{2,4,5\}\}$, $\hf^{(2)}=\partial \hf^{(3)}$ and $\tilde{\hf}=\hf^{(2)}\cup \hf^{(3)}$. Then $\hf^{(2)}$ is a complete 3-partite graph on parts $\{1,2\}$, $\{3,4\}$ and $\{5,6\}$. Now every 4-set $C\subset [6]$ contains at least one full part and at most one edge in $\hf^{(3)}$. It follows that $|\tilde{\hf}_{\mid C}|\leq 1+(6-1)=6$. Thus $\tilde{m}(6,4,7) \geq 4+12+1=17$.
Suppose that $\tilde{\hf}=\hf^{(2)}\cup \hf^{(3)}$ is a complete family of the maximal size satisfying $\tilde{\hf}\not\hookrightarrow (4,7)$.
\begin{claim}\label{claim-4.4}
For any $P=\{z_1,z_2\}\notin \hf^{(2)}$, $|\hf^{(3)}(z_1)|\leq \lfloor\frac{n-2}{2}\rfloor$ and $|\hf^{(3)}(z_2)|\leq \lfloor\frac{n-2}{2}\rfloor$ .
\end{claim}
\begin{proof}
Note that $\tilde{\hf}\not\hookrightarrow (4,7)$ implies $|F\cap F'|\leq 1$ for all distinct $F,F'\in \hf^{(3)}$. It follows that $\hf^{(3)}(x)$ is a matching for all $x\in [n]$.
Let $T_1,T_2,\ldots,T_r$ be the triples in $\tilde{\hf}$ that contain $z_1$. Since $P\notin \hf^{(2)}$, none of them contain $z_2$ and $T_1\setminus\{z_1\},\ldots,T_r\setminus\{z_1\}$ are pairwise disjoint. Hence $|\hf^{(3)}(z_1)|=r\leq \lfloor\frac{n-2}{2}\rfloor$. Similarly, $|\hf^{(3)}(z_2)|\leq \lfloor\frac{n-2}{2}\rfloor$.
\end{proof}
Let us construct a bipartite graph $\hb$ between $\hf^{(3)}$ and $\binom{[n]}{2}\setminus \hf^{(2)}$ by connecting $T\in \hf^{(3)}$ and $P\in \binom{[n]}{2}\setminus \hf^{(2)}$ iff $T\cap P\neq \emptyset$. Note that in this case $|T\cap P|=1$ by completeness of $\tilde{\hf}$.
For $x\notin T\in \hf^{(3)}$, $\tilde{\hf}\not\hookrightarrow (4,7)$ implies that at least one of the edges $\{x,y\}$, $y\in T$ is missing from $\hf^{(2)}$. Thus the degree of $T$ in $\hb$ is at least $n-3$. Should the maximum degree of $P\in \binom{[n]}{2}\setminus \hf^{(2)}$ in $\hb$ be at most $n-3$, $|\hf^{(3)}|\leq \binom{n}{2}-|\hf^{(2)}|$ and thereby $|\tilde{\hf}|\leq \binom{n}{2}$ would follow.
Assume next that $P=\{z_1,z_2\}\in \binom{[n]}{2}\setminus \hf^{(2)}$ and it has degree at least $n-2$. By Claim \ref{claim-4.4} $|\hf^{(3)}(z_i)|\leq \lfloor\frac{n-2}{2}\rfloor$, $i=1,2$. If $n$ is odd we infer $|\hf^{(3)}(z_1)|+|\hf^{(3)}(z_2)|\leq n-3$, a contradiction. The only remaining possibility is that $n$ is even and $\hf^{(3)}(z_i)$ is a perfect matching for $i=1,2$. We need only one of them.
Let $\hf^{(3)}(z_1)=\{E_i\colon 1\leq i\leq \frac{n-2}{2}\}$. We claim that at least two of the possible four edges between $E_i$ and $E_j$ are missing from $\hf^{(2)}$. Indeed otherwise we fix $x\in E_j$ that is joined (in $\hf^{(2)}$) to both vertices of $E_i$. However this forces that $\{z_1,x\}\cup E_i$ span a $K_4$ in $\hf^{(2)}$ whence $\tilde{\hf} \hookrightarrow (4,7)$.
Consequently, together with $P$ there are at least $2\binom{\frac{n-2}{2}}{2}+1$ missing edges from $\hf^{(2)}$. As to $T\in \hf^{(3)}$, $T\cap P=\emptyset$ would force that $T\cup\{z_i\}$ spans a $K_4$ in $\hf^{(2)}$ and $\tilde{\hf}\hookrightarrow (4,7)$. Thus $|\hf^{(3)}|=|\hf^{(3)}(z_1)|+|\hf^{(3)}(z_2)|\leq n-2$. For $n\geq 8$, $2\binom{\frac{n-2}{2}}{2}+1>n-2$ implies $|\hf^{(2)}|+|\hf^{(3)}|<\binom{n}{2}$ and we are done.
For $n=6$ we infer $|\tilde{\hf}|=|\hf^{(2)}|+|\hf^{(3)}|\leq \binom{6}{2}-3+4=16$.
\end{proof}
What remains are $m(n,4,14)$ and $m(n,4,15)$. These are closely related to the famous unsolved problems of Tur\'{a}n on $3$-graphs: $K_4^{(3)}$ and $K_4^{(3)-}$, where $K_4^{(3)}$ denotes the complete 3-graph on 4 vertices and $K_4^{(3)-}$ denotes $K_4^{(3)}$ minus an edge. For a $k$-graph $F$, let $ex_k(n, F)$ denote the maximum number of edges in an $n$-vertex $k$-graph which does not contain $F$ as a subgraph. It is well known that $\lim_{n\rightarrow \infty}\binom{n}{k}^{-1}ex_k(n,F)$ exists. It is called the {\it Tur\'{a}n density} of $F$ and denoted by $\pi(F)$. Tur\'{a}n \cite{turan} proposed a construction showing that $\pi(K_4^{(3)})\geq \frac{5}{9}$. Chung and Lu \cite{CL} proved $\pi(K_4^{(3)})\leq\frac{3+\sqrt{17}}{12} \approx 0.593592\cdots$. By applying the flag algebra method invented by Razborov, Razborov \cite{Ra} showed $\pi(K_4^{(3)})\leq 0.561666$. For $K_4^{(3)-}$, Frankl and F\"{u}redi \cite{FF0} proved that $\frac{2}{7}\leq \pi(K_4^{(3)-})\leq \frac{1}{3}$. In \cite{FV}, by using the flag algebra method Falgas-Ravry and Vaughan showed $\pi(K_4^{(3)-})\leq 0.286889$.
Let us derive the formula for $m(n,4,14)$ and $m(n,4,15)$ from a more general statement. In analogy with $3$-graphs let $K_r^{(k)}$ and $K_r^{(k)-}$ denote the complete $k$-graph and complete $k$-graph minus an edge on $r$ vertices, respectively.
\begin{prop}
Let $\hf\subset 2^{[n]}$ be a down-set. Then (i) and (ii) hold.
\begin{itemize}
\item[(i)] $\hf\not\rightarrow (k+1,2^{k+1}-1)$ iff $\hf^{(k)}$ is $K_{k+1}^{(k)}$-free.
\item[(ii)] $\hf\not\rightarrow (k+1,2^{k+1}-2)$ iff $\hf^{(k)}$ is $K_{k+1}^{(k)-}$-free.
\end{itemize}
\end{prop}
\begin{proof}
Since the proofs are almost identical let us show (ii) only. If $Y\in \binom{[n]}{k+1}$ spans $K_{k+1}^{(k)-}$ in $\hf$, then being a down-set forces $\binom{Y}{\ell}\subset \hf$ for all $0\leq \ell<k$. Hence $|\hf_{\mid Y}|\geq 2^{k+1}-2$.
On the other hand if $\hf$ is a $K_{k+1}^{(k)-}$-free down-set then $|F|\leq k$ for all $F\in \hf$ and $|\hf\cap \binom{Y}{k}|\leq \binom{k+1}{k}-2$ for all $Y\in \binom{[n]}{k+1}$. Thus $\hf\not\rightarrow (k+1,2^{k+1}-2)$.
\end{proof}
\begin{cor}
\begin{itemize}
\item[(i)] $m(n,k+1,2^{k+1}-1)=1+\sum\limits_{0\leq \ell<k}\binom{n}{\ell}+ex_k(n,K_{k+1}^{(k)})$.
\item[(ii)] $m(n,k+1,2^{k+1}-2)=1+\sum\limits_{0\leq \ell<k}\binom{n}{\ell}+ex_k(n,K_{k+1}^{(k)-})$.
\end{itemize}
\end{cor}
Let us close this paper by stating an old but attractive conjecture. Recall that $\hf$ is {\it antichain} if $F\subset F'$ never holds for distinct members $F,F'\in \hf$.
\begin{conj}[\cite{F89}]
Let $k$ be a non-negative integer, $n\geq 2k$. Suppose that $\hf\subset 2^{[n]}$ is an antichain with $\hf\not\rightarrow (k+1,2^{k+1})$. Then $|\hf|\leq \binom{n}{k}$.
\end{conj}
Let us note that the statement was proved in \cite{F89} for $k\leq 2$ and by Anstee and Sali \cite{AS} for $k=3$.
|
\section{Introduction}
Authentication, or the act of verifying the identity of the source of information, is a crucial aspect of security; especially in scenarios where the information leads to an observable action (e.g., calling in a missile strike or executing a stock market trade).
For information-theoretic authentication a decoder must be able to decode a message from the legitimate encoder while rejecting messages from a computationally-unbounded adversary.
More specifically, when an adversary is actively tampering with a message, the decoder only needs to output that the message is fake and does not need to output a message estimate.
Of course, a decoder declaration that the message is fake when it has not been tampered with is still considered an error.
For the scenario here, we will consider information-theoretic authentication in the context of a classical communication system where the encoder, decoder, and adversary are connected by a noisy channel.
In such a context, information-theoretic authentication is generally achieved by exploiting a feature of the communication model that is unique between encoder and decoder and which the adversary cannot imitate.
In the existing literature two features are used: either exploiting the channel in such a way that the adversary cannot mimic a valid transmission, or by use of a secret key shared by encoder and decoder.
This work is classified in the former category, as we will not allow the encoder and decoder to share a secret key.
For readers interested in secret key equipped information-theoretic authentication see Perazzone et al.~\cite{perazzone2020secret,graves2020secret} for an in-depth discussion on prior works and the best results to date.
In cases where no secret key is available, information-theoretic authentication can be obtained by exploiting (if possible) the uniqueness of the channel from the encoder to decoder.
This exploitation generally takes the form of choosing an encoder whose output when passed through the channel produces a set of observations that cannot be reliably reproduced by the adversary.
Obviously then, the information that the adversary may act on and how they are allowed to act is crucial, as it determines how well they can mimic the legitimate encoder.
Previous work~\cite{jiang2014keyless,jiang2015optimality,graves2016keyless,gungor2016basic,tu2018keyless,kosut2018authentication,beemer2019authentication,Sangwan19} on this topic is mainly differentiated by these particular formulation decisions.
A few of these decisions to be made are (without vs. with): the allowance of joint transmission by adversary and encoder~\cite{jiang2014keyless,jiang2015optimality,tu2018keyless,gungor2016basic} vs. \cite{graves2016keyless,kosut2018authentication,beemer2019authentication,Sangwan19}, side information about the encoder's message at the adversary \cite{jiang2014keyless,jiang2015optimality,gungor2016basic,tu2018keyless,kosut2018authentication} vs. \cite{graves2016keyless,beemer2019authentication,Sangwan19}, and a noisy copy of the encoder's output at the adversary \cite{jiang2014keyless,jiang2015optimality,graves2016keyless,tu2018keyless,kosut2018authentication,Sangwan19} vs. \cite{gungor2016basic,beemer2019authentication}.
It is not surprising then that most of this work is similar in formulation, methodology, and results while still being diverse in terminology. {Our work here makes all three ``with'' allowances.}
For simplicity, we will broadly characterize~\cite{jiang2014keyless,jiang2015optimality,graves2016keyless,gungor2016basic,tu2018keyless,kosut2018authentication,beemer2019authentication,Sangwan19}.
To permit more formal discussion, let $p(y|x,v)$ be the conditional distribution of the decoder's input ($y$) given the encoder's output $(x)$ and the adversary's output $(v)$, and let $\mcf{Q}$ be the (model dependent) set of joint probability distributions for the encoder and adversary's output. Further assume that there is some symbol $\emptyset$ for a not-transmitting state.
As an example of how the formulation affects $\mcf{Q}$, if the encoder and adversary are not allowed simultaneous transmission then $\mcf{Q}$ will contain only distributions such that $q(x,v)>0$ only if $x= \emptyset$ or $v= \emptyset$.
With this in mind, the previous literature divides the set of channels into sets based upon the property\footnote{Typically this property is denoted by an ``-able''-suffixed term, such as simulatable~\cite{jiang2014keyless,jiang2015optimality,gungor2016basic,tu2018keyless} (pace Maurer~\cite{maurer1994strong}), or overwritable~\cite{kosut2018authentication,Sangwan19}, $U/I$-overwritable~\cite{beemer2019authentication}. We abstained from naming the channel condition in~\cite{graves2016keyless}.} that
$$\min_{q \in \mcf{Q}} |p(y|x',\emptyset) - \sum_{x,v} p(y|x,v) q_{x'}(v,x)|_1 > 0 $$
for at least one encoder output $x$.
To understand how this property equips channels for information-theoretic authentication, view the distribution $w_{x'}(y) = \sum_{x,v} p(y|x,v) q(v,x)$ as the adversary's attempt to make the distribution at the input to the decoder close to the distribution that occurs when $x'$ is sent by the encoder and the adversary does not transmit, i.e. $p(y|x',\emptyset)$.
If $w_{x'}(y) \neq p(y|x',\emptyset)$ then the set of typical $k$-sequences of $y$ for this given $x'$ when the adversary is non-malicious (here denoted $\mcf{T}(x)$) has an exponentially decaying probability given the adversary is trying to imitate it, that is
$$w(\mcf{T}(x)) \leq 2^{-O(k \sqrt{ |w_{x'}(y) - p(y|x',\emptyset) |_1 } ) }. $$
Hence a typical set detector can generally suffice to detect the manipulation.
While not necessarily clear from the above discussion, the distribution of the encoder's output is important in determining $\mcf{Q}.$
That the capacity-achieving distribution at the output of the encoder also allows for authentication cannot be taken for granted.
Hence much of the previous work, specifically \cite{jiang2014keyless,jiang2015optimality,gungor2016basic,tu2018keyless,kosut2018authentication,Sangwan19,beemer2019structured}, opts for a two-code concatenated approach.
For this approach, one of the codes is a long (in terms of symbols) capacity-achieving code, while the other is a short low-rate code equipped with information-theoretic authentication.
Generally, the message is transmitted with the capacity achieving code, while a randomly generated number and a hash of the message and this randomly generated number are transmitted with the low-rate code that provides information-theoretic authentication.
At this point it is important to note that the previous works that use this two stage approach do not give the adversary a noisy copy of the encoder output, and hence the adversary cannot possibly determine the randomly generated number used to construct the hash value.
Without knowledge of this number, and hence without the ability to modify the low-rate code, the adversary can at best hope that they will choose a message that when combined with the randomly generated number will result in the same hash value.
In this work, we will allow the adversary a noisy copy of the transmission; the addition of this extra channel is motivated by a common wireless communication scenario with an overwhelmingly strong adversary.
This overwhelmingly strong adversary will non-causally observe noisy versions of the encoder's output while also knowing the message that the encoder is transmitting.
On the other end, the decoder will observe the superimposed transmissions of the adversary and encoder.
Both observations will be corrupted with independent \emph{additive white Gaussian noise} (AWGN), as is the tradition for first-order approximation to practical continuous-value channels dating back to Shannon~\cite{shannon1948mathematical}.
The allowance of non-causal observations at the adversary, in particular, is crucial for modeling since in practice it would be impossible to know the delay from the adversary to the encoder and from the adversary to the decoder.
Without knowing these delays, it would likewise be impossible to know how much of the encoder's output the adversary has observed and can therefore use in constructing their attack.
Allowing a non-causal observation by the adversary thus corresponds to a worst-case scenario where the adversary has enough time to observe all of the encoder's output and then choose their own outputs accordingly.
Our desire to model realistic channels under extremely adverse conditions costs us both aspects of the traditional analysis.
Indeed, recalling that a typical set detector is used to detect the manipulation, it is not surprising that the traditional analysis makes use of the fact that there are only a polynomial (in block length) number of different types.
By considering continuous channels, as opposed to~\cite{jiang2014keyless,jiang2015optimality,graves2016keyless,gungor2016basic,tu2018keyless,kosut2018authentication,beemer2019authentication,Sangwan19}, we can no longer take this approach.
Furthermore, by allowing non-causal observations at the adversary, we are generally eliminating the option to use a two-code approach to obtain capacity.
To be sure, consider the case where the adversary has less noisy observations: here the adversary would be able to decode the random number used, and hence could determine the set of messages that would result in the same hash as the transmitted message.
Thus our choice of model, motivated by practical implementation, also requires a completely new approach to solve the problem.
Despite these adversarial advantages, our scheme will achieve information-theoretic authentication with the following notable features:
\begin{itemize}
\item a construction based upon modifying almost any existing deterministic channel code;
\item does not require a shared secret key or common randomness;
\item will detect an adversary's manipulation as long as the adversary's observations of the encoder output's are not completely noiseless;
\item achieves rates arbitrarily close to the non-adversarial channel capacity.
\end{itemize}
Thus, despite the austere channel model, our scheme still allows for a robust detection.
We have specifically chosen the modification of arbitrarily given channel codes to provide a path forward for implementation.
Our work allows researchers to concentrate on modifying existing codes already having good encoders and decoders, such as low-density parity-check codes, turbo codes, polar codes, or repetition codes.
We achieve the outcomes above by building on the insights of Graves et al.~\cite{graves2016keyless} and Beemer at al.~\cite{beemer2019authentication}, where authentication was enabled by introducing artificial noise at the output of the encoder.
In fact, both works show something even more surprising: there exist channels for which deterministic codes do not allow for information-theoretic authentication, but information-theoretic authentication can be enabled by adding artificial noise to the output of the encoder.
For a simple example, consider a channel where the encoder can output $0$ or $1$, the adversary can output $-1$, $0$, and $1$, and where the decoder receives the sum of the two.
Given any deterministic encoder $\mbf{x} : \mcf{M} \rightarrow \{0,1\}^n$, if the adversary has knowledge of the transmitted message they may in turn choose their transmitted sequence as $\mbf{z}(M) = \mbf{x}(a)-\mbf{x}(M)$ so that the decoder receives $$\mbf{y}(M) = \mbf{x}(M) + \mbf{z}(M) = \mbf{x}(a),$$
which is indistinguishable from the case where the encoder sends $a\in \mcf{M}$ and the adversary does not interfere.
But now, instead consider a stochastic encoder constructed by simply taking a deterministic encoder and passing the output through a binary symmetric channel with positive crossover probability $p<\nicefrac{1}{2}$.
Now, the probability of detection can be characterized as a function of the number of coordinates $i$ for which $z_i \neq 0.$
Indeed, assume that $z_i=1$, regardless of message the probability that the encoder outputs $1$ for the $i$-th coordinate is at least $p$, hence the probability that $y_i = 2$ is at least $p$, and $y_i = 2$ can only happen if the adversary is not sending $0$.
Thus it is easy to see that the probability of false authentication is at most $(1-p)^{|\{i | z_i \neq 0\}|}$.
This probability can be made arbitrarily small by starting with a well-chosen channel code.
Hence a simple stochastic code gives us the ability to authenticate.
Of course our situation will be more complicated here because the adversary will have their own observation, but the premise remains the same.
Without complete knowledge of the encoder's output, the adversary's actions will result in decoder inputs that are not expected.
To take advantage of this insight, our code modification strategy consists of first adding carefully constructed message-dependent noise and then decimating the message set.
The message dependent noise is determined by a novel coding scheme that guarantees the adversary must always remove some of the noise added to the channel in order to forge a message.
As long as the adversary's observations themselves are noisy, the adversary will not be able to completely eliminate the message-dependent noise the encoder has added to the channel.
Thus, by detecting the presence of this noise the decoder can detect the adversary's presence.
This additional noise will guarantee that the adversary cannot modify a message to a specific message of their own choosing.
From there, decimating the message set (a concept borrowed from Ahlswede and Dueck's local strong converse~\cite{bcgc}) extends this guarantee to ensure a small maximum probability of false authentication.
To begin the formal treatment of this problem, the notation, model, and operational measures will be presented in Section~\ref{sec:nm}.
The results will be presented in Section~\ref{sec:results}, with many of the proofs being removed to the appendices for readability. Section \ref{sec:futures} includes a discussion of topics for further investigation, as well as a comparison of our scheme to secret key-based authentication schemes.
Conclusions are presented in the Section~\ref{sec:con}.
\section{Model and notation}\label{sec:nm}
\subsection{Notation}\label{sec:notation}
Uppercase letters will denote random variables, lowercase constants, and script sets.
In particular $\mcf{R}$ denotes the set of real numbers.
Bold font always denotes $n$-fold Cartesian products, with $n$ to be later defined as the block length of the code, and given $\mbf{x}$, $x_i $ is the $i$-th coordinate.
In other words $\mbf{x} = \bigtimes_{i=1}^n x_i.$
While Cartesian products of random variables and constants may have unique coordinates, a set which is a Cartesian products of sets will not (i.e., $\mbcf{X} = \bigtimes_{i=1}^n \mcf{X}$).
Throughout the paper, $\mbf{G}_{\mbf{\rho}} = \bigtimes_{i=1}^n G_{\mbf{\rho},i}$
will be used to denote Cartesian product of $n$ independent Gaussian random variables with mean $0$ where the $i$-th coordinate has variance $\rho_i$.
When all the variances are equal, (i.e., $\mbf{\rho} = \bigtimes_{i=1}^n \rho$) just the single variance will be listed (i.e., $\mbf{G}_{\rho}$).
Sometimes $\mbf{G}_{\mathrm{some~qualitative~value}}$ will be used in place of $\mbf{G}_{\mbf{\rho}_{\mathrm{some~qualitative~value}}}$ so that it is easier to specify the source of this randomness in the math.
Finally all values of $G$, unless otherwise explicitly stated, should be assumed independent.
All logarithms are natural, and the following functions will be used:
\begin{align}
\mathbb{E}[ X] &= \int_{\mcf{R}} x f_{X}(x) \mathrm{d} x
\notag \\
\mathbb{D}(X||Y) &= \int_{\mcf{R}} f_{X}(x) \log \frac{f_{X}(x)}{f_{Y}(x)} \mathrm{d} x
\notag \\
\mathbb{D}_2(a||b) &= a \log \frac{a}{b} + (1-a) \log \frac{1-a}{1-b}
\notag \\
\mathbb{H}_2(a) &= -a \log a - (1-a) \log (1-a)
\notag \\
\mathbb{I}_2\left( a || b \right) & =
b \mathbb{D}_2(a||b) + (1-b) \mathbb{D}_2\left( b \frac{1-a}{1-b} \middle| \middle| b \right)
\notag \\
&= \mathbb{H}_2(b) - b\mathbb{H}_2(a) - (1-b) \mathbb{H}_2\left( b \frac{1-a}{1-b} \right)
\notag\\
\mathbb{1}_{\mcf{A}}(b) &= \begin{cases} 1 &\text{if } b \in \mcf{A} \\ 0 &\text{else} \end{cases}
\notag \\
\Phi(x) &= \int_{-\infty}^x \frac{1}{\sqrt{2 \pi}} e^{- \frac{t^2}{2}} \mathrm{d} t
\notag
\end{align}
where $ f_X$ is used to denote the probability density function of $X$.
Furthermore, we will use $\left( \begin{matrix} \mcf{A} \\ b \end{matrix} \right)$ to denote the set of all $b$-element subsets of $\mcf{A}.$
For instance $$\left( \begin{matrix} \{1,2,3\} \\ 2 \end{matrix} \right) = \left\{ \{1,2\}, \{1,3\}, \{2,3\} \right\}.$$
\subsection{Model}\label{sec:model}
\tikzstyle{circ} = [draw, fill=white, circle, node distance=1cm]
\tikzstyle{block} = [draw, fill=white, rectangle,
minimum height=30pt, minimum width=30pt, text centered]
\tikzstyle{bigblock} = [draw, fill=white, rectangle,
minimum height=100pt, minimum width=20pt, text centered]
\begin{figure*}[t!]
\begin{center}
\begin{tikzpicture}[thick, every node/.style={transform shape}]
\node[block] (enc) at (0,0) {$\begin{array}{c} \text{Encoder} \\ \mbf{X}(M) \end{array}$};
\node[circ] (add1) at (2,-.75) {$+$};
\node[circ] (add2) at (6,0) {$+$};
\node[block] (dec) at (4,-1.7){$\begin{array}{c} \text{Adversary} \\ \mbf{Z}(\mbf{V},M) \end{array}$};
\node[block] (dec2) at (9,0){$\begin{array}{c} \text{Decoder} \\ \hat m( \mbf{Y} ) \end{array}$};
\draw[->,thick,dashed] (3.5,-.75) node[above]{$\mbf{G}_{\mathrm{Adv}}$} -- (add1.east) ;
\draw[->,thick] (6,.75) node[right]{$\mbf{G}_{\mathrm{Dec}}$} -- (add2.north) ;
\draw[->,thick] (dec.east) -- (6,-1.7) node[below] {$\mbf{Z}$} -- (add2.south);
\draw[->,thick] (enc.east) -- (add2.west);
\draw[->,thick] (add2.east) -- (6.75,0) node[above]{$\mbf{Y}$} -- (dec2.west);
\draw[->,thick] (dec2.east) -- (11,0) node[above]{$\hat M$};
\draw[->,thick] (-2,0) node[above] {$M$} -- (enc.west);
\draw[->,thick,dashed] (2,0) node[above] {$\mbf{X}$} -- (add1.north);
\draw[->,thick,dashed] (add1.south) -- (2,-1.45) node[left] {$\mbf{V}$} -- (2.96,-1.45)
\draw[->,thick,dashed] (-1.5,0) -- (-1.5,-1.95) -- (2.96,-1.95);
\end{tikzpicture}
\end{center}
\caption{Channel with encoder $\mbf{X}:\mcf{M} \rightarrow \mbcf{R}$ and decoder $\hat m: \mbcf{R} \rightarrow \mcf{M} \cup \{!\}$, where $\mbf{G}_{\mathrm{Dec}} \sim \text{Gaussian}(0,\rho_{\mathrm{Dec}})$ and $\mbf{G}_{\mathrm{Adv}} \sim \text{Gaussian}(0,\rho_{\mathrm{Adv}})$. The dashed lines represent non-causal links.} \label{fig:chan}
\end{figure*}
The model for communications (pictured in Figure~\ref{fig:chan}) studied here consists of three entities: an encoder, decoder, and an adversary.
\com{\eric{Used ``in this model'' instead of deleting ``here'' to avoid the same word starting back to back sentences}} In this model, the encoder is tasked with sending a message $M$ to the decoder, where the message is assumed to be uniform\footnote{The distribution of this message will not play a role in the results.} over $\mcf{M}$.
To do this the encoder will map the message to an $n$-symbol sequence $\mbf{X}(M)$ and send it across the communications channel. It is important to note that the code is allowed to be a random function of the message.
When the encoder sends its codeword, the adversary will receive a non-causal noisy copy of the $n$-symbol sequence
$$\mbf{V} = \mbf{X}(M) + \mbf{G}_{\mathrm{Adv}},$$
where $\rho_{\mathrm{Adv}} \in (0,\infty)$ represents variance of the adversary's noise.
Using this received information, the adversary will craft their own $n$-symbol sequence to inject into the channel.
In general this function will be modeled by $\mbf{Z} : \mbcf{R} \bigtimes \mcf{M} \rightarrow \mbcf{R}$.
Towards discussion purposes, it can be generally assumed that the adversary chooses this function optimally: that is, to minimize the performance metrics of the system.
On the other hand, the decoder will receive a noisy copy of the combination of $n$-symbol sequences sent by encoder and adversary,
$$\mbf{Y} = \mbf{X}(M) + \mbf{Z}(\mbf{V},M) + \mbf{G}_{\mathrm{Dec}},$$
where $\rho_{\mathrm{Dec}}\in (0,\infty)$ represents the noise variance at the decoder.
From there, the decoder will attempt to estimate the message the encoder sent as $\hat m(\mbf{Y})$ or will output $\mbf{!}$ to indicate that the adversary has altered the message.
\subsection{Operational Parameters} \label{sec:opparam}
The objective of this work is to construct a good code for authenticated communications.
\begin{define}\textbf{(Code)}
\com{\eric{Stochastic code in the sense that the output is stochastic, not in the sense that a code is randomly chosen before hand. Although, I feel that the difference between those two is more philosophical than mathematical. Like, if a coin is flipped but you do not know the results, is it still random? Regardless, the terminology can be changed to ``random code,'' but this creates problems with code construction.}} A \emph{code} is a set of paired functions $\mbf{X} : \mcf{M} \rightarrow \mbcf{R}$, $\hat m : \mbcf{R} \rightarrow \mcf{M} \cup \{\mbf{!}\}$ representing the \emph{encoder} and \emph{decoder} respectively. The symbol $\mbf{!}$ specifically represents the case that the decoder labels the observation as not authentic.
\end{define}
\begin{remark}
A code not designed for authenticated communications can be considered as a special case where $\hat m(\mbf{y}) \neq \mbf{!}$ for all $\mbf{y} \in \mbcf{R}.$
\end{remark}
\begin{remark}
Codes are assumed to have block length (number of symbols output) $n$, unless otherwise stated.
\end{remark}
Codes will be measured by the rate at which they can send information, the power required to do so, the reliability with which information is decoded when there is no adversarial interference, and the likelihood the adversary can manipulate the decoder into accepting a false message.
Formal definitions for the first three follow.
\begin{define}\label{def:rate}\textbf{(Rate)} The \emph{rate} of a code $\mcf{H} = (\mbf{X}, \hat m)$ is
$$r_{\mcf{H}} = \frac{1}{n} \log |\mcf{M}|.$$
\end{define}
\begin{define}\label{def:power}\textbf{(Power Constraint)} The \emph{power constraint} of a code $\mcf{H} = (\mbf{X}, \hat m)$ is
$$\omega_{\mcf{H}} = \max_{m \in \mcf{M}} \sum_{i=1}^n \frac{1}{n} \mathbb{E}\left[ X_i^2(m) \right] .$$
\end{define}
\begin{define}\label{def:prerr}\textbf{(Error Probability)}
For code $\mcf{H} = (\mbf{X}$, $\hat m$) the \emph{arithmetic average error probability} at noise variance $\rho_{\mathrm{Dec}} \in (0,\infty)$ is
$$
\varepsilon_{\mcf{H}}(\rho_{\mathrm{Dec}}) = \sum_{m \in \mcf{M}} \frac{1}{|\mcf{M}|} \Pr \left( \hat m ( \mbf{X}(m) + \mbf{G}_{\mathrm{Dec}} ) \neq m \right) .
$$
\end{define}
\noindent Note that the error probability is indeed a measure of reliability when not under adversarial influence, since if $\mbf{Z}(\mbf{V},M) = \mbf{0}$ then $\mbf{Y} = \mbf{X}(M) + \mbf{G}_{\mathrm{Dec}}$.
Two measures of the adversary's ability to interfere will be considered.
The weaker of these two measures considers the adversary's ability to have the decoder accept a specific message.
\begin{define}\label{def:tfa}\textbf{(Maximum Probability of Targeted False Authentication)}
The \emph{maximum probability of targeted false authentication} for code $\mcf{H}= (\mbf{X}, \hat m)$ with decoder noise variance $\rho_{\mathrm{Dec}} \in ( 0 , \infty)$ and adversary noise variance $\rho_{\mathrm{Adv}} \in (0,\infty)$ is
\begin{align}
&\alpha^*_{\mcf{H}}(\rho_{\mathrm{Dec}}, \rho_{\mathrm{Adv}}) =
\notag \\
&\sup_{\substack{\mbf{Z}:\mbcf{R} \bigtimes \mcf{M} \rightarrow \mbcf{R} \\ a \in \mcf{M} \\ b \in \mcf{M} \setminus \{a\}} } \Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) = b \right)
\notag ,
\end{align}
where
$$\mbf{V} = \mbf{X}(a) + \mbf{G}_{\mathrm{Adv}}.$$
\end{define}
A small probability of targeted false authentication does not guarantee the decoder will not output a false message, instead it guarantees that the adversary cannot \textit{choose} which message it is.
This weaker metric will only play a brief role in this study, with the main goal being to obtain codes which measure favorably under the following, stronger metric.
\begin{define}\label{def:pa}\textbf{(Maximum Probability of False Authentication)}
The \emph{maximum probability of false authentication} for code $\mcf{H}= (\mbf{X},\hat m)$ with decoder noise variance $\rho_{\mathrm{Dec}}$ and adversary noise variance $ \rho_{\mathrm{Adv}}$ is
\begin{align}
&\alpha_{\mcf{H}}(\rho_{\mathrm{Dec}}, \rho_{\mathrm{Adv}}) =
\notag \\
&\sup_{\substack{\mbf{Z}:\mbcf{R} \bigtimes \mcf{M} \rightarrow \mbcf{R} \\ a \in \mcf{M}} } \Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) \notin \{ a,\mbf{!}\} \right) ,
\notag
\end{align}
where
$$\mbf{V} = \mbf{X}(a) + \mbf{G}_{\mathrm{Adv}}.$$
\end{define}
Unlike the targeted false authentication probability, a vanishing probability of false authentication does asymptotically guarantee the decoder will not output a false message in the presence of an adversary.
\begin{remark}
To better understand the relationship between the two metrics observe that
\begin{align}
&\Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) \notin \{ a,\mbf{!}\} \right)
\notag \\ &\quad
=
\sum_{b \in \mcf{M}\setminus\{a,\mbf{!}\}} \Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) = b \right),
\notag
\end{align}
from which it is clear that having a small maximum probability of targeted false authentication does not guarantee a small maximum probability of false authentication, but a small maximum probability of false authentication does guarantee a small maximum probability of targeted false authentication.
\end{remark}
\begin{remarkstar}\label{remark:iwonderifthiswill work}
Readers familiar with information-theoretic authentication literature\footnote{Primarily, the information-theoretic authentication literature whose genesis is Simmons~\cite{Auth}; most of these works are secret key-based {which marks a striking difference from our work here}.} may be wondering why we have not defined the impersonation attack.
For those unfamiliar, an impersonation attack is one where the adversary does not wait for the encoder to produce an output, but directly sends a value to the decoder.
If we were to formally define this as an operational measure it would be
\begin{align}
&\sup_{\mbf{Z}:\mbcf{R} \bigtimes \mcf{M} \rightarrow \mbcf{R} } \Pr \left( \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{G}_{\mathrm{Adv}},\emptyset) \notin \{ \emptyset, \mbf{!}\} \right) .
\notag
\end{align}
We do not feel the need to define this metric separately, since it can already be accounted for in the definition of the encoder and decoder.
That is, in $\mcf{M}$ we may assume that there is a special symbol (call it $\emptyset$) that corresponds to the case where the encoder has no message to transmit.
If, for instance, $\mbf{X}(\emptyset) = \mbf{0}$ then clearly
\begin{align}
&\alpha_{\mcf{H}}(\rho_{\mathrm{Dec}}, \rho_{\mathrm{Adv}})
\notag \\ &\quad
\geq
\sup_{\mbf{Z}:\mbcf{R} \bigtimes \mcf{M} \rightarrow \mbcf{R} } \Pr \left( \hat m( \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{G}_{\mathrm{Adv}},\emptyset) ) \notin \{ \emptyset ,\mbf{!}\} \right) ,
\notag
\end{align}
and hence our formulation (for appropriately defined codes) already encompasses impersonation attacks.
This observation will have an important consequence in the context of our results.
In preview of this, in order to ensure authentication, our results will require the encoder still output low levels of additive white Gaussian noise when it has no message to transmit.
Clearly the assumption of such a possibility will be model-dependent and would not be valid for a situation like a wired channel where the channel may be physically severed.
On the other hand, severing a link would be very difficult in a wireless environment and hence an eternally active encoder in our model is justifiable.
\end{remarkstar}
One of the primary goals of our work will be to characterize the authenticated capacity.
Intuitively, the authenticated capacity is the maximum rate possible under a given power constraint and the requirement that the probability of error and maximum probability of false authentication converge to zero.
In order to present the exact definition, the notation
\com{\eric{Do not understand comment. Defining the message set size as a function of the block length allows for a sequence of codes with near constant rate.}}
$\mbf{X}_{(n)}: \mcf{M}_{(n)} \rightarrow \mbcf{R}_{(n)}, \hat m_{(n)} : \mbcf{R}_{(n)} \rightarrow \mcf{M}_{(n)} \cup \{\mbf{!}\}$ will be used to denote codes with block length $n$.
\begin{define}\label{def:cap}\textbf{(Authenticated capacity)}
The \emph{authenticated channel capacity} is
\begin{align}
&c(\rho, \rho_{\mathrm{Dec}},\rho_{\mathrm{Adv}}) =
\notag \\
& \sup \left\{ r \in \mcf{R} \middle|
\begin{array}{rl} \exists \mcf{H}_{(n)} = \{ \mbf{X}_{(n)}, \hat m_{(n)} \}_{n=1}^\infty & \\
\text{ such that }& \\
\displaystyle \limsup_{n \rightarrow \infty} \omega_{\mcf{H}_{(n)}} &\leq \rho \\
\displaystyle\liminf_{n \rightarrow \infty} r_{\mcf{H}_{(n)}} &\geq r \\
\displaystyle\limsup_{n \rightarrow \infty} \varepsilon_{\mcf{H}_{(n)}}(\rho_{\mathrm{Dec}}) &= 0 \\
\displaystyle\limsup_{n \rightarrow \infty} \alpha_{\mcf{H}_{(n)}}(\rho_{\mathrm{Dec}},\rho_{\mathrm{Adv}}) &= 0
\end{array} \right\}.
\notag
\end{align}
\end{define}
\begin{remark}
If the authentication requirement were removed, the capacity would be
$$\frac{1}{2} \log \left( 1 + \frac{\rho}{\rho_{\mathrm{Dec}}} \right) $$
following from Shannon~\cite{shannon1948mathematical}.
Indeed, removing the authentication measure leaves the operational definitions for point-to-point communications over an AWGN channel without an adversary, as is to be expected.
\end{remark}
On a final note, as mentioned in the introduction, Section~\ref{sec:results}'s code construction results will be presented in terms of a given channel code.
These results will, however, require the initial channel code be deterministic, i.e., the encoder output is not random given the message.
\section{Results}\label{sec:results}
In this section we will build a number of consecutive results which lead to the conclusion that information-theoretic authentication is possible in AWGN channels without the need for a secret key.
Not only this, but we will also show that existing codes can be equipped with information-theoretic authentication at a small cost to the rate, power, and error probability of the code.
This is achieved by modifying the codes with two complementary modifications.
The first of these modifications will (literally) add to the encoder's output a type of code which enables detection of targeted authentication attacks.
The second modification of eliminating messages will then extend this to cover all attacks.
With these results in hand, we show that the costs asymptotically vanish while the ability to detect manipulation remains; furthermore, we show that this is true regardless of the difference between the adversary and decoder's noise variance, instead only requiring that the noise variance at the adversary be non-zero.
In pursuit of a modular scheme, we begin by constructing a new type of code, termed an \textit{overlay code}.
Conceptually, these codes are used to control the amount of a persistent\footnote{By ``persistent'' we mean that it is difficult to remove.} resource added to each transmission symbol for each message.
The overlay code guarantees that a portion of this persistent resource must be removed by the adversary before they can falsify a message.
If the adversary is unable to remove the persistent resource, then its presence can be used by the decoder to detect the intrusion.
For the given channel model, the persistent resource will take the form of Gaussian noise, and the adversary will have to attempt noise cancellation in order to remove the persistent resource's presence.
Before introducing overlay codes in Definition~\ref{def:overlay}, it will be helpful to introduce the intuition behind their conception.
These codes are structured to enable basic statistical testing practices to detect the overabundance of the persistent resource.
This is done by first limiting to a discrete set the possible levels of persistent resource added per symbol.
All symbols that have a given amount of persistent resource (e.g., all symbols which have had half of the maximum amount of resource added) can be thought of as the ``test sets'' since these sets will eventually form the sets over which we perform hypothesis testing {in order to determine the presence of an adversary}.
The most important property of the overlay code is that for any given message and any alternative message, one of the test sets for the given message will correspond to symbols whose persistent resource level is always less than or equal to (with a certain amount guaranteed to be strictly less than) the persistent resource level of the alternative message.
Consider this set-up in the context of authentication, where the alternative message represents the actual transmitted message and the given message the one produced by the decoder.
In this case, one of the test set for the given (decoded) message will correspond to a set of symbols for which the encoder added more of the persistent resource for the alternative (transmitted) message.
If the adversary cannot remove this resource efficiently enough, then its presence can be used to detect the message is false.
We now define the overlay code.
\begin{define} \label{def:overlay}
Given finite set $\mcf{K} \subset[0,1)$, and $\mcf{\tilde K} = \mcf{K} \cup 1,$ positive real number $r$, and $\gamma \in \left( \frac{1}{2}, 1 \right)$, a function $\mbf{f}: \mcf{M} \rightarrow \mbcf{\tilde K}$ is a \emph{$(r,\mcf{K} ,\gamma)$-overlay code} when
\begin{itemize}
\item $$\frac{1}{n}\log |\mcf{M}|\geq r;$$
\item $$ \sum_{i=1}^n \idc{ f_i(m)}{\{ k \} } = \ell := \left \lfloor \frac{n}{|\mcf{\tilde K}|} \right \rfloor
$$
for all $m \in \mcf{M}$ and $k \in \mcf{K}$;
\item and for each distinct $m, m' \in \mcf{M}$ there exists a $k \in \mcf{K}$ such that
$$ \sum_{i=1}^n \idc{ f_i(m)}{\{k\}} \idc{ f_i(m')}{\{ k\}} \leq \gamma \ell $$
and for all $j \in \mcf{K}$ such that $j < k$
$$ \sum_{i=1}^n \idc{ f_i(m)}{ \{k\} } \idc{ f_i(m')}{\{j\}} = 0 .$$
\end{itemize}
\emph{Uniform overlay codes} are overlay codes with $\mcf{K} = \left\{0, |\mcf{\tilde K}|^{-1}, \dots, 1 - |\mcf{\tilde K}|^{-1} \right\}$.
\end{define}
\begin{remark}
For the remainder of the paper, let $\mcf{\tilde K} := \mcf{K} \cup 1$ and $\ell:= \left \lfloor \frac{n}{|\mcf{\tilde K}|} \right \rfloor.$
\end{remark}
\begin{remark}
If $\mbf{f}$ is an $(r,\mcf{K},\gamma)$-overlay code, then for each $\mcf{\tilde M}\subset \mcf{M}$ the function $\mbf{\tilde f} : \mcf{\tilde M} \rightarrow \mbcf{R}$ defined by $\mbf{\tilde f}(m) = \mbf{ f}(m)$ is a $(\frac{1}{n} \log |\mcf{\tilde M}|,\mcf{K},\gamma)$-overlay code.
\end{remark}
\com{\eric{$\mcf{M}$ and $n$ are defined in the model section as the set of messages $M$ is distributed over, and the number of symbols transmitted respectively. Also, notation defined earlier is that $\mbf{x} = \times_{i=1}^n x_i $, which should imply $\mbf{f} = \times_{i=1}^n f_i$. I would be ok with changing it to $f(m)_i$, which would be taking the $i$-th coordinate of the output of $\mbf{f}(m).$} }
\begin{remark}
It would certainly be possible to define overlay codes to allow a non-uniform number of symbols per persistent resource level (less than the maximum). That we did not do so is merely for the sake of simplicity.
\end{remark}
Note, fewer resource levels $|\mcf{\tilde K}|$ implies more symbols share each level, hence fewer levels implies that there are more symbols to test per set.
Obviously though, fewer resource levels also means fewer unique output sequences for the overlay code, hence the overlay code will support fewer messages.
To quickly see this, observe that if $\mcf{\tilde K}$ consisted of two elements, then there would be at most $2^n$ different possible code combinations.
The existence of overlay codes should not be taken for granted a priori.
For instance, consider a traditional random coding argument where for each message the encoder outputs are chosen at random from a predefined distribution.
For any two messages $a$ and $b$, let $F_i(a)$ and $F_i(b)$ denote the randomly chosen value of $i$-th coordinate resource level for messages $a$ and $b$.
Observe that $ \Pr \left( F_i(a) < F_i(b) \right) = \frac{1 - \sum_{k \in \mcf{\tilde K}} \Pr\left( F_i(a) = k \right)^2 }{2}$.
Thus for any choice of distribution other than a deterministic one, $\Pr \left( F_i(a) < F_i(b) \right)>0$, and hence when rate $r> - n^{-1}\log \Pr \left( F_i(a) < F_i(b) \right)$ this construction will (with near certainty) produce a code such that for every message $a$, there exists a message $b$ whose resource levels are always greater than or equal to $a$'s.
Increasing the size of $\mcf{K}$ would exacerbate this problem.
Nevertheless, overlay codes do exist given certain conditions outlined in Theorem~\ref{thm:codevid19} and Corollary~\ref{cor:wedontwantthemtogetofftheshipbecausethatmaydoubleournumbers}.
\begin{theorem}\label{thm:codevid19}
For any positive real number $r$, finite $\mcf{K}\subset [0,1)$, and $\gamma \in \left( \frac{1}{2}, 1 \right)$ such that
$$r \leq \frac{1}{n} \sum_{k \in \mcf{K}} n_k \left| \mathbb{I}_2\left(\gamma \middle|\middle| \frac{\ell}{n_k} \right) - \frac{4}{3n_k} - \frac{2}{n_k}\log n_k \sqrt{\ell} \right|^+,$$
where $n_k = n - \ell |\{ j \in \mcf{K}|j< k\}|$, there exists a $(r,\mcf{K},\gamma)$-overlay code.
\end{theorem}
\begin{cor}\label{cor:wedontwantthemtogetofftheshipbecausethatmaydoubleournumbers}
For all $\gamma \in \left( \frac{1}{2}, 1 \right)$ and finite $\mcf{K}\subset [0,1)$, if positive number
$$r < \gamma \log (|\mcf{\tilde K}|) - \gamma - \mathbb{H}_2(\gamma) , $$
then for large enough $n$ there exists a $(r,\mcf{K},\gamma)$-overlay code.
\end{cor}
\begin{proofsketch}
The full proofs of Theorem~\ref{thm:codevid19} and Corollary~\ref{cor:wedontwantthemtogetofftheshipbecausethatmaydoubleournumbers} can be found in Appendix~\ref{app:codevid19}. Also to be found in Appendix~\ref{app:codevid19} is a detailed example of the overlay code construction.
We prove the theorem using an iterated random coding procedure.
First we represent $\mcf{M}$ as {in bijection with} a product of smaller sets, that is
$\mcf{M} = \bigtimes_{i\in \{1,\dots,|\mcf{K}|\}} \mcf{M}_{i} .$
Next, independently for each $m_1 \in \mcf{M}_1$ we randomly select an $\ell$-coordinate subset out of the total $n$ coordinates.
These $\ell$ coordinates are those for which the overlay code outputs the smallest resource concentration (i.e., the minimum value in $\mcf{K}$).
This process is repeated for all $(m_1, m_2) \in \mcf{M}_1 \times \mcf{M}_2$, with the difference being that the set of $n-\ell$ coordinates not selected for $m_1$ is used {for selection of coordinates of the second-smallest resource concentration}.
This process of removing the selected coordinates and then randomly selecting a new set of coordinates is repeated until there are fewer than $\ell$ coordinates remaining; at this point the remaining coordinates are assigned an overlay output of $1$.
From this process, the resulting form of Theorem~\ref{thm:codevid19} should be clear.
The summand over each $k \in \mcf{K}$ is simply the maximum rate at which our analysis can guarantee that two messages {match on} at most $\gamma \ell$-chosen coordinates.
To see why this method works, consider the following.
For any two messages $m,m' \in \mcf{M}$ there exist representations ($m_1,\ldots, m_{|\mcf{K}|})$ and $(m_1',\ldots, m_{|\mcf{K}|}')$ respectively.
Clearly, there exists a smallest value $j\in \{1,\dots, |\mcf{K}|\}$ such that $m_j \neq m_j'.$
Now, for $m$ and $m'$ the overlay code coordinates corresponding to the $1$st through $(j-1)$th resource levels will be equal since $(m_1,\ldots, m_{j-1}) = (m_{1}',\ldots, m_{j-1}')$.
For the $j$th level though, the two messages will have different coordinates.
Furthermore, whenever the output overlay concentration for message $m$ is equal to the $j$th level, the concentration for message $m'$ must be greater than or equal to the $j$th level since all coordinates for resource levels less than that level are shared.
Using the appropriate random coding techniques, we can then guarantee a certain percentage of coordinates that do not share a level for $m_{j}$ and $m_{j}'.$
\end{proofsketch}
\begin{remark}
Of extreme importance here is that for a fixed rate $r$ and fixed $\gamma$, there is a fixed $|\mcf{K}|$ that guarantees the existence of a overlay code for large enough $n$.
Thus, the value of $|\mcf{K}|$ should be intuitively viewed as a constant when dealing with asymptotic results.
\end{remark}
\begin{remark}
We will not be concerned with choosing the optimal values for inclusion in $\mcf{K}$ in this paper.
This is primarily because the optimal values will depend on the adversary's noise variance, and we wish to have our code construction be independent of this knowledge.
We will return to this discussion in Section~\ref{sec:futures}.
\end{remark}
Given the existence of overlay codes, we now go about applying them to arbitrary codes to enable authentication.
Importantly, a secret key is not necessary in this application, since authentication is enabled by the persistence of the resource added.
For our communication model, the persistent resource is additive Gaussian noise.
Our code modification will make use of the overlay code to determine the variance of the Gaussian noise added to the encoder's output.
For primarily clerical reasons, another message-dependent signal, $\mbf{t}(M)$, will also be added to the output of the encoder.
We strongly suspect it is not necessary for most practical codes, although it is necessary for a result that is agnostic of the original code.
\begin{codingmod} \label{code:addnoise}
~\\
Suppose
\begin{itemize}
\item a deterministic code $\mbf{x}: \mcf{M} \rightarrow \mbcf{R}$, $\hat m : \mbcf{R} \rightarrow \mcf{M}$,
\item injection noise power $\rho_{\Delta} \in (0,\infty)$,
\item tolerance $\delta\in (0,1)$, and
\item an $\left( \frac{1}{n} \log |\mcf{M}|, \mcf{K}, \gamma \right)$-overlay code $\mbf{f}:\mcf{M} \rightarrow \mbcf{\tilde K}$, for some finite $\mcf{K} \subset [0,1)$ and $\gamma \in \left(\frac{1}{2},1\right)$,
\end{itemize}
are given.
Independently for each $m \in \mcf{M}$ and $i\in\{1,\dots,n\}$ randomly choose $t_i(m)\in \mcf{R}$ according to a Gaussian distribution with mean $0$ and variance $(1-f_i^2(m)) \rho_{\Delta}$.
Define the modified encoder $\mbf{ X}': \mcf{M} \rightarrow \mbcf{R}$ by
$$\mbf{X}'(M) = \mbf{x}(M) + \mbf{t}(M)+ \mbf{f}(M) \cdot \mbf{G}_{\Delta},$$
where $\cdot$ is the coordinate-wise product and $\mbf{G}_{\Delta} = \mbf{G}_{\rho_{\Delta}}$.
Define the modified decoder $\hat m': \mbcf{R}\rightarrow \mcf{M} \cup \{ \mbf{!}\}$ by
\begin{align}
&\hat m' (\mbf{y})
\notag \\ &=
\begin{cases}
\hat m(\mbf{y}) & \text{if } \forall k \in \mcf{K} \\
& \displaystyle \sum_{i \in \mcf{I}_k} \frac{[y_i -t_i(\hat m(\mbf{y}))- x_i(\hat m (\mbf{y}))]^2}{k^2 \rho_{\Delta} + \rho_{\mathrm{Dec}}} \leq \ell ( 1+ \delta) \\
\mbf{!} &\text{else}
\end{cases}, \notag
\end{align}
where $\mcf{I}_k$ is the set of coordinates $i$ such that $f_i(\hat m (\mbf{y}) ) = k.$
The resulting modified code is defined by $\mbf{X}',\hat m'.$
\end{codingmod}
\begin{remarkstar}\label{remark:alwayson}
Recall Remark~\ref{remark:iwonderifthiswill work} from Section~\ref{sec:model}.
In this remark we noted how our single metric could handle both impersonation attacks and substitution (or inference) attacks by assuming that the code had a message that corresponded to a ``not transmitting'' state.
The application of this code modification must also apply to this ``not transmitting'' state.
In other words, Code Modification \ref{code:addnoise} requires that the encoder still send a low level noise when there is no message to transmit.
\end{remarkstar}
\begin{remark}
Note the modified decoder is the original decoder with the extra requirement that
$$ \sum_{i \in \mcf{I}_k} \frac{[y_i -t_i(\hat m(\mbf{y}))- x_i(\hat m (\mbf{y}))]^2}{k^2 \rho_{\Delta} + \rho_{\mathrm{Dec}}} \leq \ell ( 1+ \delta) $$
for all $k \in \mcf{K}$.
In this sense, the modified decoder can be viewed as first using the original decoder to decode the message, and then checking for manipulation by ensuring that the extra requirement is met.
For reference purposes, we shall adopt this two-stage decoder view, and refer to the checking of the extra requirement as the \emph{detector}.
\end{remark}
To see why Code Modification~\ref{code:addnoise} provides a small probability of \textit{targeted} false authentication, consider the steps an adversary would have to perform in order to fool the decoder into authenticating a particular message.
First, the adversary, given their a priori knowledge of the message and codebook, would subtract out the output of the unmodified encoder for the transmitted message as well as the $\mbf{t}$ term.
Next they would add in the unmodified encoder's output and the $\mbf{t}$ term for the alternative message they wished the decoder to accept.
Finally, the adversary would try to ensure that the correct amount of noise is applied to the correct symbols {so as to avoid detection}.
But, while the adversary knows the variance of the encoder-added noise per symbol, they will not know the exact value of this added noise since their measurement is itself noisy.
As the injected noise power becomes smaller, the variance of the adversary's estimate will become increasingly large relative to the encoder-added noise's own variance.
Eventually, the adversary's estimate will be so poor that if the adversary tries to cancel out the added noise the resulting variance would not be significantly less than that of the encoder-added noise alone.
Thus the scheme protects against any message being forged into a different particular message, since this different message will be guaranteed to have a set of coordinates that have less noise variance per symbol than the adversary can manage.
Later we will extend this scheme using Code Modification~\ref{code:decimate}/Code Modification Corollary~\ref{code:decimate_alt} to protect against all types of attacks.
While this does provide a form of information-theoretic authentication, adding noise to the output of the encoder will degrade the signal-to-noise ratio.
In turn, this decrease in the signal-to-noise ratio will reduce the maximum achievable rate or, alternatively, increase the probability of decoding error.
Our analysis favors the increase in the probability of error.
Additionally, the increase in noise will increase the power needed by the encoder.
But, as Theorem~\ref{thm:1stcode}/Corollary~\ref{cor:1stcode} formally shows, these costs can vanish while still allowing detection of \emph{targeted} authentication attacks.
\begin{cor}[Theorem~\ref{thm:1stcode}]\label{cor:1stcode}
Given injection noise power $\rho_{\Delta} = o(1) $ and tolerance $\delta = o(\rho_{\Delta})$, then for all
\begin{itemize}
\item deterministic codes $\mcf{H} = (\mbf{x}: \mcf{M} \rightarrow \mbcf{R}$, $\hat m : \mbcf{R} \rightarrow \mcf{M})$,
\item $\left( r_{\mcf{H}} , \mcf{K}, \gamma \right)$-uniform overlay code $\mbf{f}: \mcf{M} \rightarrow \mbcf{\tilde K}$ for any $\gamma \in (1/2, 1)$ and viable $|\mcf{K}|$,
\item and large enough $n$ ,
\end{itemize}
Code Modification~\ref{code:addnoise} yields with high probability a code $\mcf{J} = (\mbf{X}' : \mcf{M} \rightarrow \mbcf{R}$, $\hat m': \mbcf{R} \rightarrow \mcf{M} \cup \{\mbf{!} \} )$ such that
\begin{align}
r_{\mcf{J}} &= r_{\mcf{H}}
\notag \\
\omega_{\mcf{J}} &\leq \omega_{\mcf{H}} + O( \sqrt{\rho_{\Delta}} )
\notag \\
\varepsilon_{\mcf{J}}(\rho_{\mathrm{Dec}} ) &\leq \varepsilon_{\mcf{H}}(\rho_{\mathrm{Dec}}+ \rho_{\Delta}) + e^{-O(n \delta^2 )}
\notag \\
\alpha^*_{\mcf{J}}(\rho_{\mathrm{Dec}}, \rho_{\mathrm{Adv}})
&\leq
e^{ -O(n \rho_{\Delta}^2)}.
\notag
\end{align}
\end{cor}
\begin{remark}
Corollary~\ref{cor:1stcode} is a corollary of Theorem~\ref{thm:1stcode} located in Appendix~\ref{app:gen}.
Theorem~\ref{thm:1stcode}, unlike the above corollary, does not fix the injection noise power, tolerance, or the values in $\mcf{K}.$
Furthermore, Theorem~\ref{thm:1stcode} specifies the error terms instead of using order terms.
\end{remark}
\begin{proofsketch}
The proof of Theorem~\ref{thm:1stcode} is found in Appendix~\ref{app:1stcode}, and Corollary~\ref{cor:1stcode} trivially follows.
Proving the rate is immediate, since it is unchanged from the original code.
For the average power, we have to deal with the deterministic value of $\mbf{t}(M)$ added to the code, in particular analyzing the probability that a spurious value of $\mbf{t}(m)$ is chosen with a large amount of correlation with the related $\mbf{x}(m).$
For the probability of error, we have to consider both the probability of error of the original decoder with the added noise and $\mbf{t}$ as well as the probability of error introduced with the detector.
To upper bound the probability of error of the original decoder, we use the fact that the randomly chosen value of $\mbf{t}$ plus the message-dependent additive white Gaussian noise terms is effectively a message-independent additive white Gaussian noise term with variance $\rho_{\Delta}.$
Hence, the error averaged over all possible choices of $\mbf{t}(M)$ is $\varepsilon_{\mcf{H}}(\rho_{\mathrm{Dec}} + \rho_{\Delta})$.
Using Hoeffding's inequality, it follows that the random choice of $\mbf{t}$ must yield a probability of error close to the average.
On the other hand, the probability of error of the detector is straightforward to calculate since, under no manipulation, the detector is checking to see if a sum of independent random variables has the correct mean.
Finally for the probability of targeted false authentication, we note that if the adversary does try to attack, then the distribution of the received sequence at the decoder will consist of independent Gaussian random variables where the variance of the $i$th coordinate is
$$ \tau_i(m) = \tau^\star (f_i(m)) := \frac{f_i^2(m) \rho_{\Delta} \rho_{\mathrm{Adv}}}{f_i^2(m) \rho_{\Delta} + \rho_{\mathrm{Adv}}} + \rho_{\mathrm{Dec}} ,$$
and the mean is of the adversary's choosing.
For visualization purposes, note that when $\rho_{\Delta}$ becomes small this variance term converges to $f_i^2(m) \rho_{\Delta} + \rho_{\mathrm{Dec}}.$
By properties of the overlay code though, for each message and alternative message, there exists one set of overlay output coordinates whose output for the decoded message is less than or equal to an alternative message.
The probability of detecting this increase in noise variance (under the assumption that the decoded message is not the one transmitted by the encoder is calculated and used to determine the probability of detecting the adversary's manipulation.
\end{proofsketch}
While Code Modification~\ref{code:addnoise} does not allow the adversary to impersonate any specific message, it does not guarantee that the adversary cannot impersonate any message at all.
This difference is made plain by referring to the operational definitions and observing again that
\begin{align}
&\Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) \notin \{ a,\mbf{!}\} \right) \notag \\
&\quad = \sum_{b \in \mcf{M} \setminus \{a , \mbf{!}\}} \Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) = b \right) . \label{eq:ai0}
\end{align}
While Code Modification~\ref{code:addnoise} produces codes such that each summand $ \Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) = b \right)$ is small, it does not guarantee the production of a code for which the sum itself is small.
Some reflection, though, shows that the case where the summand is small but this sum is not can only occur if there is (in some sense) a densely packed set of decoding regions.
Under this notion, it makes sense to randomly decimate the message set, similar to how (and why) Ahlswede and Dueck~\cite{bcgc} chose to demonstrate the local strong converse.
While this does reduce the rate of the code, only a negligible amount of loss (in terms of rate) is needed to guarantee the decoding regions are much less dense.
We will resume with a slightly more formal description of why this works after we introduce the coding modification.
For now, we must mention that the amount of decimation the message set needs is dependent on operational measures of the underlying code.
Therefore, to improve readability we have opted to produce a simplified version of the code modification here, and leave the more precise result for Appendix~\ref{app:gen}.
\begin{codingmodcor}[Code Modification~\ref{code:decimate}]\label{code:decimate_alt}
Suppose
\begin{itemize}
\item deterministic code $\mcf{H} = (\mbf{x}: \mcf{M} \rightarrow \mbcf{R}$, $\hat m : \mbcf{R} \rightarrow \mcf{M})$,
\item injection noise power $\rho_{\Delta}= o(1)$,
\item tolerance $\delta = o(\rho_{\Delta})$,
\item $\left( \frac{1}{n} \log |\mcf{M}|, \mcf{K}, \gamma \right)$-overlay code $\mbf{f}:\mcf{M} \rightarrow \mbcf{\tilde K}$, for finite $\mcf{K} \subset [0,1)$ and $\gamma \in \left(\frac{1}{2},1\right)$,
\end{itemize}
are given.
First apply Code Modification~\ref{code:addnoise} to code $\mcf{H}$, to obtain code $ \mbf{X}': \mcf{M} \rightarrow \mbcf{R}$, $\hat m' : \mbcf{R} \rightarrow \mcf{M} \cup \{\mbf{!}\}$.
Next, select $\mcf{M}^\ddagger$ uniformly at random from $\left( \begin{matrix} \mcf{M} \\ \left \lfloor \exp (nr^\ddagger) \right \rfloor \end{matrix} \right)$, where
$$r^\ddagger = r_{\mcf{H}} - O\left(\rho_{\Delta}^2+\frac{\log n}{n} \right) .$$
Define the modified encoder $\mbf{X}^\ddagger : \mcf{M}^\ddagger \rightarrow \mbcf{R}$ by
$$\mbf{X}^\ddagger (M) = \mbf{X}'(M) .$$
Define the modified decoder $\hat m^\ddagger : \mbcf{R} \rightarrow \mcf{M}^\ddagger \cup \{\mbf{!}\}$ by
$$\hat m^\ddagger (\mbf{Y}) = \begin{cases} \hat m'(\mbf{Y}) & \text{ if } \hat m'(\mbf{Y}) \in \mcf{M}^\ddagger \\
\mbf{!} & \text{ else}
\end{cases}.$$
The resulting modified code is given by $\mbf{X}^\ddagger, \hat m^\ddagger.$
\end{codingmodcor}
\begin{remark}
Code Modification Corollary~\ref{code:decimate_alt} is a corollary of Code Modification~\ref{code:decimate} located in Appendix~\ref{app:gen}.
There, the decimation terms are made explicit.
\end{remark}
\begin{remarkstar}
Decimating the message set reduces the rate of the code.
\end{remarkstar}
We now return to a more formal description of why this works, which follows from two important facts.
First, decimating the message set will not impact the maximum probability of targeted false authentication for any two non-decimated messages.
Second, by decimating the message set to $\mcf{M}^\ddagger$, the probability of false authentication for a given encoded message $a \in \mcf{M}$ and fixed adversary function $\mbf{Z}$ can be written as
\begin{equation}\label{eq:adint}
\sum_{b \in \mcf{M} \setminus \{a , \mbf{!}\}} \idc{b}{\mcf{ M}^{\ddagger}}\Pr \left( \hat m(\mbf{X}(a) + \mbf{G}_{\mathrm{Dec}} + \mbf{Z}(\mbf{V},a) ) = b \right).
\end{equation}
Equation~\eqref{eq:adint}, when considered jointly with the decimated message set $\mcf{M}^\ddagger$ being randomly chosen, takes a form whose concentration is analytically tractable.
More specifically Equation~\eqref{eq:adint} should with high probability be close to the mean, which is at most $|\mcf{M}^{\ddagger}|/|\mcf{M}|$ since this is the probability a message is not decimated.
The above intuition is overly-simplistic because all possible attacks must be simultaneously considered.
Nevertheless, the technique is sufficient to prove the next theorem/corollary.
\begin{cor}[Theorem~\ref{thm:2ndcode}]\label{cor:2ndcode}
Setting injection noise power $\rho_{\Delta} = o(1) $ and tolerance $\delta = o(\rho_{\Delta})$, then for all
\begin{itemize}
\item deterministic codes $\mcf{H} = (\mbf{x}: \mcf{M} \rightarrow \mbcf{R}$, $\hat m : \mbcf{R} \rightarrow \mcf{M})$ with rate $r_{\mcf{H}} = \Omega( n^{-1} \log n) $,
\item $\left( r_{\mcf{H}}, \mcf{K}, \gamma \right)$-uniform overlay code $\mbf{f}: \mcf{M} \rightarrow \mbcf{\tilde K}$ for any $\gamma \in (1/2, 1)$,
\item and large enough $n$
\end{itemize}
Code Modification~\ref{code:decimate} with high probability yields a code $\mcf{J} = (\mbf{X}^\ddagger: \mcf{M}^\ddagger \rightarrow \mbcf{R}$, $\hat m^\ddagger : \mbcf{R} \rightarrow \mcf{M}^\ddagger \cup \{\mbf{!}\})$ such that
\begin{align}
r_{\mcf{J}} &\geq r_{\mcf{H}} - O\left(\rho_{\Delta}^2+\frac{\log n}{n} \right)
\notag \\
\omega_{\mcf{J}} &\leq \omega_{\mcf{H}} + O( \sqrt{\rho_{\Delta}} )
\notag \\
\varepsilon_{\mcf{J}}(\rho_{\mathrm{Dec}} ) &\leq \varepsilon_{\mcf{H}}(\rho_{\mathrm{Dec}}+ \rho_{\Delta}) + e^{-O(n \delta^2)}
\notag \\
\alpha_{\mcf{J}}(\rho_{\mathrm{Dec}},\rho_{\mathrm{Adv}} ) &\leq e^{-O(n \rho_{\Delta}^2)}.
\notag
\notag
\end{align}
\end{cor}
\begin{remark}
Corollary~\ref{cor:2ndcode} is a corollary of Theorem~\ref{thm:2ndcode} located in Appendix~\ref{app:gen}.
Theorem~\ref{thm:2ndcode}, unlike the above corollary, does not fix the injection noise power, tolerance, or the values in $\mcf{K}.$
Furthermore, Theorem~\ref{thm:2ndcode} specifies the error terms instead of using order terms.
\end{remark}
\begin{proofsketch}
The proof of Theorem~\ref{thm:2ndcode} is found in Appendix~\ref{app:2ndcode}; note that it relies on elements of the proof of Theorem~\ref{thm:1stcode} since Code Modification~\ref{code:decimate} relies on Code Modification~\ref{code:addnoise}.
The rate and power for the new code are straightforward, while the probability of error calculation essentially follows from Hoeffding's inequality.
The major difficulty in the proof is proving the bound on the probability of false authentication.
As the first step in proving this bound, we recall a result from the proof of Theorem~\ref{thm:1stcode}; specifically, that the decoder's observation when conditioned on a particular message, adversary observation, and adversary attack is equal to a sequence of independent random variables with the mean of the adversary's choosing but the variance fixed, i.e.,
$$\mbf{Y}|\{ M,\mbf{Y} ,\mbf{Z} = m, \mbf{v}, \mbf{z} \} = \mbf{G}_{\mbf{\tau}(m)} + \mbf{u}(m,\mbf{v},\mbf{z}),$$
where $\mbf{u}(m,\mbf{v},\mbf{z})$ is an arbitrary function (whose specification is unimportant for this proof) and
$$ \tau_i(m) = \frac{f_i^2(m) \rho_{\Delta} \rho_{\mathrm{Adv}}}{f_i^2(m) \rho_{\Delta} + \rho_{\mathrm{Adv}}} + \rho_{\mathrm{Dec}} $$
for each symbol $i \in \{1,\dots,n\}.$
Clearly, we can effectively ignore the values of $\mbf{V}$ and $\mbf{Z}$ by jointly considering $ \mbf{G}_{\mbf{\tau}(m)} + \mbf{\mu}$ for all $m \in \mcf{M}$ and $\mbf{\mu} \in \mbcf{R}$.
Now for any given $\mbf{\mu}\in \mbcf{R}$ and $m \in \mcf{M}^\ddagger$, we start by noting the probability of false authentication can be written
\begin{equation}\label{eq:ps:thm:dec}
\sum_{b \in \mcf{M} \setminus \{m , \mbf{!}\}} \idc{b}{\mcf{ M}^{\ddagger}}\Pr \left( \hat m'( \mbf{G}_{\mbf{\tau}(m)} + \mbf{\mu}) = b \right),
\end{equation}
where $\hat m'$ is the modified decoder resulting from the application of Code Modification~\ref{code:addnoise} in Code Modification~\ref{code:decimate}.
Using a modified version of the Hoeffding lemma we then bound the concentration of equation~\eqref{eq:ps:thm:dec}.
The problem that remains is to extend above concentration to simultaneously work for all $\mbf{\mu} \in \mbcf{R}$.
Here we take a divide-and-conquer approach by separately considering the sets of $\mbf{\mu} \in \mbcf{U}^\dagger$ and $\mbf{\mu} \notin \mbcf{U}^\dagger$, where $\mcf{U}^\dagger$ is a bounded interval on the real number line.
These bounds are set sufficiently large so that $\mbf{\mu} \notin \mbcf{U}^\dagger$ guarantees that for the coordinate such that $\mu_i \notin \mcf{U}^\dagger$, the probability of passing the detector for each message is less than $e^{-n r_{\mcf{H}} - O(n \rho_{\Delta}^2)},$ and hence the probability of passing any message detector is less than $e^{-O(n\rho_{\Delta}^2)}.$
For $\mbf{\mu} \in \mbcf{U}^\dagger$, we show that there exists a finite set $\mbcf{U}^\ddagger \subset \mbcf{R}$ such that bounding all $\mbf{\mu} \in \mbcf{U}^\ddagger$ will suffice to bound all $\mbf{\mu} \in \mbcf{U}^\dagger$.
From there, we use the union bound to simultaneously guarantee the concentration of all $\mbf{\mu} \in \mbcf{U}^\ddagger$ (hence all $\mbf{\mu} \in \mbcf{U}^\dagger$) and all $m \in \mcf{M}.$
\end{proofsketch}
At this point, it is important to reflect on the form of Theorem~\ref{thm:2ndcode}/Corollary~\ref{cor:2ndcode}.
Specifically, consider Corollary~\ref{cor:2ndcode} where $\delta$ is chosen such that $\lim n\delta^2 = \infty$.
For example $\rho_{\Delta} = \sqrt[-4]{n} \log n$ and $\delta = \sqrt[-4]{n}.$
In this case, the code modifications have necessitated a loss in rate, an increase in power, and require the code to be operational at a larger noise level than the original code.
However, each of these changes disappear as $n$ increases, and hence the rate converges back to the original rate, the new power converges to the original power, and the level of noise the code must be robust against converges to the original noise level.
Suppose then we start with a capacity-achieving sequence of codes with average power $\omega - O(\sqrt{\rho_{\Delta}}),$ and which are robust to a noise variance of $\rho_{\Delta} + \rho_{\mathrm{Dec}}$.
Applying Theorem~\ref{thm:2ndcode}/Corollary~\ref{cor:2ndcode} should give us a sequence of codes with rate
$$\lim_{n \rightarrow \infty} \frac{1}{2} \log \left( 1 + \frac{\omega - O(\sqrt{\rho_{\Delta}})}{\rho_{\Delta} + \rho_{\mathrm{Dec}}} \right) - o(1) = \frac{1}{2} \log \left( 1 + \frac{\omega}{\rho_{\mathrm{Dec}}} \right), $$
which is capacity.
At the same time, plugging the values into the maximum probability of false authentication yields
$$\lim_{n \rightarrow \infty} \alpha_{\mcf{J}}(\rho_{\mathrm{Dec}},\rho_{\mathrm{Adv}} ) \leq \lim_{n \rightarrow \infty} e^{-O(n \rho_{\Delta}^2)} = 0,$$
and thus we have the ability to authenticate.
This essentially proves the following theorem.
\begin{theorem}\label{thm:cap}
$$c(\rho,\rho_{\mathrm{Dec}}, \rho_{\mathrm{Adv}}) = \begin{cases} \frac{1}{2} \log \left( 1 + \frac{\rho}{\rho_{\mathrm{Dec}}} \right) & \text{if } \rho_{\mathrm{Adv}}>0 \\ 0 & \text{else.} \end{cases}$$
\end{theorem}
\begin{proofsketch}
The proof Theorem~\ref{thm:cap} is found in Appendix~\ref{app:cap} and is essentially a more formal version of the discussion preceding the theorem.
Additionally, we show that if $\rho_{\mathrm{Adv}} = 0$ then the capacity is zero.
This is somewhat obvious since the adversary knows the encoder's output perfectly in this case.
\end{proofsketch}
Notice that the capacity experiences a sharp jump at $\rho_{\mathrm{Adv}} = 0,$ but is otherwise independent of the value.
From a practical perspective, this is ideal.
A perfect continuous channel is a physical impossibility, thus allowing us to assume that $\rho_{\mathrm{Adv}}$ is greater than zero. Hence, our result implies that in practical wireless scenarios, information-theoretic authentication is possible without use of a secret key.
It is also important to observe that the code modifications themselves do not rely on knowledge of the adversary's channel.
Interestingly, our results indicate that obtaining information-theoretic authentication from a channel differs significantly from obtaining information-theoretic secrecy from a channel.
Indeed, all practically relevant schemes for the wiretap channel, dating back to Wyner's seminal work~\cite{wyner75wtc}, require both knowledge of the adversary's channel as well certain guarantees on this channel which make implementation a difficult proposition.
In the relevant analog to our model\footnote{Specifically, from Figure~\ref{fig:chan} remove the message side information given to the adversary and remove the adversary's output.}, information-theoretic secrecy cannot be guaranteed when the noise to the adversary is less than the noise to the decoder.
This is not an impediment to information-theoretic authentication though, as our results demonstrate; {importantly, knowledge of the message is distinct from knowledge of the transmitted sequence.}
In the next section we will discuss the path forward in more detail.
Among other things, we will discuss unexplored alternatives for implementation, barriers to practical implementation, difficulties in other channels, and different implementation scenarios.
\section{Discussion \& Future Directions}\label{sec:futures}
While we derive a scheme that leads to information-theoretic authentication, there remains much to be done.
It is worth discussing these remaining questions with some candor, so that those so motivated have a clear understanding of areas for improvement.
We also provide here further discussion on the distinction (beyond the obvious) between secret-key based authentication and what we accomplish here.
\subsection{Overlay code improvements}
When first formulating overlay codes, the goal was to ensure the unique relationship of the output symbols for different messages.
In the construction, there were a number of different parameters that could have been varied. In particular: the number of coordinates for a given output concentration, the overlap amount per coordinate, and the output levels themselves (i.e., $\mcf{K}$).
To simplify our analysis, we chose to fix the first two considerations, while leaving $\mcf{K}$ variable.
Surprisingly, the actual values for $\mcf{K},$ while they do impact the efficiency of the authentication scheme, are actually rather immaterial to achieving authentication.
Further analysis showed the optimal values of $\mcf{K}$ depend on the value of $\rho_{\mathrm{Adv}}.$
As a result, we chose not to optimize over $\mcf{K}$ since important to our claims is that the value of $\rho_{\mathrm{Adv}}$ need not be known when constructing the code.
During the review process, the question of the optimal value of $\mcf{K}$ was raised.
To that end, when $\rho_{\Delta} = o(1)$ the optimal choice of $\mcf{K}= \{0,k_{(1)},\dots, k_{(|\mcf{K}|-1)} \}$ converges to
$$ k_{(a)} = \left[\frac{(1-c\gamma)}{c(1-\gamma)}\right]^{a -1} \frac{\rho_{\mathrm{Dec}}}{\rho_{\Delta}} ,$$
where
$$c = \frac{\sqrt[|\mcf{K}|]{\rho_{\mathrm{Dec}}} }{\gamma \sqrt[|\mcf{K}|]{\rho_{\mathrm{Dec}}} + (1-\gamma) \sqrt[|\mcf{K}|]{ \rho_{\Delta}+\rho_{\mathrm{Dec}} }},$$
and yields a maximum probability of false authentication (subject to our analysis) of essentially
$$\exp\left(-\frac{1}{8} (1-\gamma) [1 - (1+\delta) c ]^2 \right).$$
Still, use of this asymptotically optimal value did not simplify our analysis and hence was not instituted.
This does, however, raise the question of what is being lost (in terms of authentication ability) by choosing a sub-optimal values for $\mcf{K}$. Specifically, it would be interesting to quantify that loss in such a way as to allow for choosing $\mcf{K}$ to minimize the maximum of the maximum of the probability of false authentication.
Additionally, it remains an open question whether allowing variable $\gamma$ and variable coordinates per output symbol could further improve the final results.
\subsection{Practical implementation of code modifications}\label{sec:fut:pi}
To enable authentication, message-dependent noise must be added and then certain distance properties between the codewords must be ensured.
These two tasks appear here as Code Modifications~\ref{code:addnoise} and~\ref{code:decimate}.
Our original intent was practicality in these code modifications; we were moderately successful with regards to Code Modification~\ref{code:addnoise}, but not so with~\ref{code:decimate}.
That Code Modification~\ref{code:addnoise} could be reasonably implemented guided our decision to include here the non-asymptotic versions of Theorem~\ref{thm:1stcode} and~\ref{thm:2ndcode}.
Still, it is worthwhile to discuss alternatives to our code modifications that could allow for an analytical bounds on the operational parameters, as well as a practical implementation.
For Code Modification~\ref{code:addnoise}, the only real concern in terms of practicality is the construction of the $\mbf{t}$ function.
Indeed, since the initial decoder is used in the first stage of the updated decoder, the output of the decoder can be used to determine what the appropriate value of $\mbf{t}$ should be for the estimated message.
It is worth mentioning that we suspect that setting $\mbf{t}$ equal to zero will suffice in most cases.
Our suspicion derives from the fact that $\mbf{t}$ is only needed to ensure that the code appears to have uniform noise across all coordinates.
In practical decoders though, less noise per symbol is usually to the decoder's benefit.
Setting $\mbf{t}$ to zero would yield $\omega_{\mcf{J}} \leq \omega_{\mcf{H}} + \rho_{\Delta}$, with the rate and probability of targeted false authentication remaining as in Theorem~\ref{thm:1stcode}.
On the other hand, the average arithmetic error could be estimated empirically.
Hence, this should result in a practical implementation of Code Modification~\ref{code:addnoise} for which Theorem~\ref{thm:1stcode} is relevant.
Code Modification~\ref{code:decimate}, on the other hand, cannot be directly implemented as currently stated.
Choosing such a large subset uniformly at random from the set of all such subsets is clearly impossible in practice.
There may of course be feasible alternatives. For instance, the subset selection could be accomplished using a universal hash function, and the Hoeffding concentration analysis replaced with one deriving from the leftover hash lemma.
Alternatively, it may be possible to show that some codes do not actually require a rate reduction.
Indeed, our analysis for Code Modification~\ref{code:decimate} relies heavily on the maximum probability of targeted false authentication established by Code Modification~\ref{code:addnoise}.
But the adversary can only obtain this maximum by choosing a very specific output, and cannot obtain it for multiple alternative messages at one time.
As a result, it seems likely that a more sophisticated analysis, using the amount of perturbation from the optimal output, could yield a maximum distance between codewords required for there to be a successful attack.
Ensuring that the code's minimum distance was greater than this maximum would be sufficient to skip Code Modification~\ref{code:decimate} entirely.
\subsection{Comparison with secret key-based authentication}
The major advantage our authentication scheme has over one that is secret key-dependent is that the secret key becomes a finite resource when the channel to the adversary is better than the channel to the decoder.
Hence, in some channel models, our scheme could operate in perpetuity while one which is key-based would have a finite life span.
That this is particularly true in any case where the adversary has a better channel is shown in Graves et al.~\cite{graves2020secret} whose converse proves the key has a finite duration of use.
On the other hand, secret key-based authentication still allows for two advantages over the non-secret key-based authentication of this paper.
First, it is still operational when there is no noise over the channel to the adversary\footnote{A physical impossibility.} and when the adversary knows, and can therefore cancel, the decoder's noise\footnote{Also a physical impossibility.}.
Second, and more important, secret key-based authentication experiences a better trade-off between rate loss and how quickly the probability of false authentication converges to zero.
For secret key-based authentication, we know that there must exist a trade-off between the channel capacity and the exponent for the probability of false authentication due to the converse results from Graves and Wong~\cite{graves2019inducing} and Graves et al.~\cite{graves2020secret}.
For some measures of false authentication, this trade-off is linear, and in that sense the message rate and probability of false authentication must share the channel capacity.
Our results do not allow for this type of trade off.
That is, while our results require a reduction in rate from the channel capacity in order to achieve authentication, the exponent for the probability of false authentication is at most $\exp(-o(n))$ whereas secret key-based authentication allows $\exp(-O(n)).$
If we assume that the encoder knows the channel to the adversary\footnote{This comparison to secret key-based authentication is not entirely fair, since knowledge of the channel to the adversary is not needed in that case.} then it is possible to also achieve $\exp(-O(n))$ with our results.
Indeed, this is because in this case we do not need $\rho_{\Delta}$ to vanish, but instead just be sufficiently small.
Regardless, even under this unfair comparison, and further assuming the more generous result on the power constraint raised in Section~\ref{sec:fut:pi} and that the second code modification was unnecessary, to obtain a maximum probability of false authentication of $\exp(-O(n\rho_{\Delta}^2))$ requires that the difference between the maximum rate and capacity be at least
\begin{align}
&\frac{1}{2} \log \left( 1 + \frac{\rho}{\rho_{\mathrm{Dec}}} \right) - \frac{1}{2} \log \left( 1 + \frac{\rho- \rho_{\Delta}}{\rho_{\mathrm{Dec}} + \rho_{\Delta}} \right)
\notag \\ & \quad
=
\sum_{i = 1}^\infty \frac{1}{i} (c \rho_{\Delta})^i
\notag
\end{align}
where
$$c = \frac{\rho+\rho_{\mathrm{Dec}}}{\rho + \rho_{\mathrm{Dec}} + \rho_{\Delta} \left( 1 + \frac{\rho}{\rho_{\mathrm{Dec}}} \right)}.$$
Hence, a loss of rate does not lead to a linear increase in the exponent of maximum probability of false authentication using our scheme.
\subsection{Higher order wireless channel approximations}
While Gaussian channels are great approximations for free-space fixed point single antenna communications\footnote{This point is discussed by Massey~\cite{massey1992deep} regarding deep-space communications.}, there exist other scenarios of wireless communications with their own corresponding best channel approximations.
Some of these alternative channels consider \emph{multi-input multi-output} (MIMO) antenna arrays, fading channels, and multi-path channels.
Outright, we do not see any reason that the overlay code concept cannot be modified and applied to these channels to create codes that provide information-theoretic authentication.
However, any such modification will be highly dependent on the assumptions placed on the encoder and decoder with regards to knowledge of their own channels.
\section{Conclusion}\label{sec:con}
In this work we have shown that physical layer authentication is possible for a channel that models wireless communication.
Not only is physical layer authentication possible, but our scheme can be used to detect any adversary as long as the block length is sufficiently large and the adversary does not have access to a completely noiseless copy of the transmission.
Our scheme achieves this by adding artificial noise into the system using the novel concept of overlay codes.
This approach allows for authentication by forcing the adversary to remove the added noise when they hope to insert a fake message of their own.
Although random coding elements were used in the proofs, many of the difficulties in practical implementation do not exist in our modular scheme.
That is, only the encoder needs to be constructed, since part of the concept of the modular scheme is that the message can still be decoded using the original decoder {(see Section \ref{sec:fut:pi})}.
Furthermore, our modular scheme establishes that every deterministic channel code has a variant which can provide physical layer authentication.
We expect this to lower the implementation barrier since we therefore do not require a completely new channel code be added to the system design.
{Open problems include those outlined in Section \ref{sec:futures}, as well as investigating further scenarios where adding artificial noise can provide authentication.}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
A key assumption of Shannon's mathematical theory of communication \cite{shannon48} was that communication was limited to random samples from a finite set of symbols, such as for example the alphabet, or zeros and ones. As Shannon, stated: \textit{Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem.} However, modern communication have obviated this limitation. It is common in modern communication systems to interact with chat-bots, virtual assistants, and different sorts of conversational agents.
More dramatically, modern virtual reality systems enable us to capture, stream and render full three dimensional representations of human actors, across the auditory and the visual domains. Alternatively, such performances can also be created with digital production tools, on a labour-intensive practice by character-animation specialists. We can therefore create, store and replay animations of cartoon or realistic characters. Moreover, we can render interactive performances using Virtual or Augmented Reality headsets, thus recreating a communication scenario that, despite being technologically mediated, can be quite similar to spontaneous natural communication.
For such a communication scenario to match our expectations, artificial cognitive agents must infer (or appear to infer) the meaning of a certain message and, from a combination (or an abstract representation) of previously recorded performances, generate a response that is appropriate to the message received, and that is (or appears to be) aligned with the communication intentions of the cognitive agent. Several techniques can be adapted for this purpose, either from natural language processing, or from artificial intelligence literature focused on video game characters. These techniques, though, are often in stark contrast with the way such mechanisms occur in humans. We lack a communication theory comparable to the precision of Shannon's work to characterise precisely inferential communication, as well as to quantify the benefits and drawbacks of different approaches to create artificial agents capable of generating interactive behaviour consistent with these principles.
To address this gap, in this article we introduce a Mathematical Theory of Inferential Communication (MaTIC). The article is structured as follows.
After this introduction, we review how inferential communication is described in cognitive pragmatics and review broader cognitive principles. We review notions such as predication, natural categories, conceptual metaphors, and competence acquisition.
Next, we briefly review how cognitive and computational neurosciences characterise the embodiment of such principles. We review the concept of massive modularity as a cognitive organisation principle and discuss how such functional modules inter-relate in inhibitory and excitatory networks.
We then summarise the previous picture in a formal construct called the general cognitive module (GCM). We argue it summarises how cognitive sciences characterise the building blocks of embodied cognition. We also try to
illustrate the generality of a GCM we show how well-known computational models for autonomous agents can be interpreted as particular cases of such a construct.
We then turn to our main case of interest: how cognitive agents formed of networks of GCMs can learn to interpret the actions of other agents and generate appropriate responses. For this purpose, we introduce three cognitive principles for inferential communication that we label $I$, $S$ and $T$. We interpret these principles as mechanisms that shape the topology of connections among different GCM.
These guide \emph{competence acquisition} in an agent, and our proposal is that such principles are as valid for humans as for artificial agents.
A particularity of this theory is that it tries to reconcile the assumption that subjective inferences, as they occur in inferential communication, emerge from embodied mechanisms, with a formalist approach. We try to illustrate how these two aspects make a quite compelling picture in a domain where subjective inferences and a formalist approach converge: mathematical foundations as depicted from the perspective of embodied cognition. In this context, we show that the resulting theory under specific assumptions produces a set theory that, contrary to the axiomatic assumptions of ZFC set theory, the set theory assumed by default, is consistent with the picture of embodied cognition as a foundation for mathematics.
Finally, we discuss whether such a model allows adapting the rich tool set of information theory to characterise information exchanges between two inferential agents, how it compares with different proposals that try to define cognitive systems with general computational principles, and future research directions, both theoretical and practical.
\section{Inferential communication}
Contrary to the code model \cite[]{shannon48}, in which a communicator encodes a message to be decoded by an audience, the \emph{inferential} model of communication considers that \emph{communicative actions}, both verbal and non-verbal, should be understood as the expression and recognition of intentions: the communicator provides evidence of his intention to convey a certain meaning, and the audience infers the meaning on the basis of the evidence provided. This evidence can be something said together with some gestures and voice inflexions, but it can also be non-verbal, such as pointing, smiling, approaching, or any other action that the audience is susceptible of associating a meaning to it. The association of the evidence with a meaning is strongly influenced by the social and material context.
If classical information theory was largely inspired by the image of two agents coding and decoding signs through a telegraph wire, the inferential model of communication can be exemplified by the communication of two fully embodied agents sharing a three dimensional environment portraying in real time bidirectional behaviour and speech with the purpose of exchanging information and intentions related with this information. This model has never been formalised or tested in quantitative and rigorous terms such as information theory has. On the other side, it can give simple explanations to quite complicated phenomena. For example, in this view the reason why we are constantly playing with different layers of meaning when we communicate is simply for communicative economy. To give a better picture of this theory, we summarise some of the notions on which it is based:
\begin{enumerate}
\item \textbf{Language games}\cite{wittgenstein1953pit} characterised everyday communication as a \emph{language game}. To do so, he outlined a large number of examples on how the meaning of words and actions are determined by how we use them, in what context and with what intention. Therefore, we can define the class of \emph{Language Games} as all situations in which several people (or things like virtual actors) can perform \emph{communicative actions}.
\item \textbf{Natural Categories}
The notion of Natural Categories \cite[]{rosch1996frs} has been widely adopted in cognitive sciences, ranging from cognitive psychology to applied linguistics. Essentially, natural categories are the basic element that allows us to categorise the world and organise mentally cognitive knowledge. Formally speaking, they were generally identified with fuzzy sets \cite[]{zadeh1965fuzzy}.
This implies there are better and worse examples of each category, which it is often named as the level of prototypicality. Obvious examples of this can be found in colours: in face of a spectrum of different red samples we will attribute a ``typical red" example or a ``not so good red" example to different samples.
\item \textbf{Conceptual Metaphors}
In cognitive linguistics, the notion of conceptual metaphor as used by \cite{lakoff1980mwl} builds upon the idea of natural categories. In this view, conceptual metaphors are mappings between domains that give a richer meaning to how we communicate and think. Usual metaphors such as \emph{Life is a Journey}, or categories of different cultural systems, such as the category of \emph{Women, Fire, and Dangerous things} \cite{lakoff2008women}, to cite only the most famous examples, rely on the notion of Natural Category.
This interpretative framework has had numerous developments. For example, in arguing that basic mathematical intuitions rely on conceptual metaphors that are derived from experience, and that it is this which gives a lot of what we call mathematical understanding, independently of the demonstration process \cite[]{lakoff2000mc}.
\item \textbf{The Context of an action}
\emph{Communicative actions} related with interpersonal communication have a context which can include factors such as:
part of the physical situation, but also the cultural background assumed to be shared, the gender, social relation of the speakers and in general anything that contributes implicitly to determine what is meant by a certain communicative action.\footnote{The role of contexts in natural communication can be traced back to 1923 in the work of Malinowsky, but see chapter 2, section 1 in \cite{kramsch1993cac}, or the entry \emph{context et situation} in \cite{houde1998vsc}} A typical example of this is irony. If someone says \emph{I am very happy}, his assertion can be interpreted as meaning to express his happiness, or exactly the inverse, his extreme sadness. This will depend not only on how this is said, but also according to what we know about the person from what he did in the past. Some elements of the present situation can also have an influence in what is interpreted from what is said. Therefore, in natural communication agents must control very carefully the context in which \emph{communicative actions} are performed. Otherwise it will be difficult to ensure the meaning intended is appropriately conveyed.
\item \textbf{Intentions}
A different aspect of interpreting events when performed by actors is that we associate \emph{intentions} to them.
If we see someone in a certain situation or conversation, especially if we know him well or he is in a familiar socio-cultural context, we can spontaneously say what he will probably do or say, or at least a small set of different options, according to the situation, taking into account previous experienced contexts. This is a general cognitive ability of humans called \emph{intention attribution} or \emph{mind-reading}, which some impaired people seam to lack \cite[]{baron2000understanding}.
For example, consider the expression: \emph{This paper is white.}
Depending on the context in which it is said, the same text can convey a meaning similar to \emph{I’m surprised because in his exam he did not write a word}, or \emph{Fantastic, finally I found a blank paper and I can draw what I was wanting to} or \emph{so, you finally did not fill in the papers of our divorce, did you?}. The process by which we go from \emph{what is said} to \emph{what is meant} is determined not only by the meaning of the words and in what situation the communicative action is performed, but also by implicit assumptions in everyday interaction. All these are part of the context.
With the previous notions, we can summarise inferential communication in non-mathematical terms with the \emph{Principle of Cooperation} \cite[]{grice200lac} and the two principles of \emph{Relevance Theory} \cite[]{sperber2004rt}.
\item \textbf{The Cooperative Principle}
\label{coop-principle}
Grice's \emph{Cooperative Principle} was the first principle of communication proposed, and it would involve any everyday linguistic interaction --even non-verbal. The clearest explanation might be introduced by it's proponent \cite[]{grice200lac}):
\begin{quote}
Our talk exchanges do not normally consist of a succession of disconnected remarks, and would not be
rational if they did. They are characteristically, to some degree at least, cooperative efforts;
and each participant recognises in them, to some extent, a common purpose or set of
purposes, or at least a mutually accepted direction. This purpose or direction may be fixed
from the start (e.g., by an initial proposal of a question for discussion), or it may evolve
during the exchange [...]. We might then formulate a rough general principle which participants will be expected (ceteris paribus) to observe, namely: Make your conversational contribution such as is required, at the stage at which it occurs, by the accepted purpose or direction of the talk exchange in which you are
engaged. One might label this the \emph{Cooperative Principle}.
\end{quote}
A classical example adapted from \cite{grice200lac} will frame better the problem. Consider this scenario:
\begin{quote}
A is standing by an obviously immobilised car. \\
B approaches him. \\
A says: I am out of petrol. \\
B says: There is a garage around the corner. \\
\end{quote}
In the previous example, the last reply of B would be a consequence of the previous exchanges.
To formalise such a mechanism we will need a definition of implication ($\Rightarrow$) general enough to include these \emph{implicatures}.
\item \textbf{Relevance Theory}
Sperber and Wilson's Relevance Theory \cite[]{sperber2004rt} develops two basic principles to explain the inferential model of communication. The first principle of relevance theory or cognitive principle, states that \emph{human cognition tends to be geared towards the maximisation of the cognitive effect}, and this cognitive effect is quantified with a relevance measure. This implies that, given a certain context, a communicative action will be performed in order to maximise the relevance that the receiver will associate to it.
The second principle of relevance theory or communication principle adds an extra layer. It states that \emph{every ostensive stimulus conveys a presumption of its own optimal relevance}. This is best shown with an example. If two people, Alice and Bob, are chatting in Alice's living room, Bob can put his empty glass in Alice's line of sight to suggest that he might want more water, but this is only exploiting Alice's cognitive tendency to maximise relevance. Bob still has to show you that he put the glass there \emph{on purpose}, that it was his intention, and therefore that it is worth processing this stimuli in Alice's cognitive system, which is something not trivial in a world of stimuli competing for scarce cognitive attention.
The previous implies that a communicative event is not one in which a maximum amount of information is given, but rather one in which a maximally \emph{relevant} information is given.
It is somewhat a principle operating at a second level: it implies any event triggered will aim at giving the minimal information but with the maximal relevance. At reception, it also implies that determining the goal(s) conveyed by a certain action should assume optimal relevance. But to be able to quantify such relevance, agents need not only to determine causal relations, but also have the skills to determine the underlying goals of a certain behaviour.
\item \textbf{Competence Acquisition}
The ability to associate intentions with human behaviour is developed by sustained social interaction from the very first stages of baby learning \cite[]{halliday1985saw,tomasello2003clu}. Competence acquisition in language and communication seems to be based on imitation combined with the development of mental schemes. Language competence would come progressively by associating the performance of different chunks of language with real-world contexts and consequences. Progressively the learner would use these chunks of language and their combinations. Therefore, the connection between sounds, grammar,and meaning is not an universal process, but a social and situated one \cite{kendrick2020sequence}. According to Kendriks, there appear to be strong universals of interactive language usage,
namely the sequencing of social actions across neighbouring turns, and this allows to identify the universality of the sequence organisation observable in informal human conversational interaction. There is some experimental evidence this describes accurately how kids acquire grammatical knowledge \cite[]{bannard2009modeling}.
This also seems to correspond with the old notion of \emph{internalisation} as a means to acquire communicative competence \cite[]{vygotsky1964tal}. For example, one learns how to ride a bicycle by seeing it and trying it, independently of an understanding of how it works: you internalise what to do to get a certain effect. Similarly, one learns to use a word by determining in what context and with what purpose someone else used it, and then shifts to use it in order to obtain a certain effect.
The term \emph{internalisation} therefore refers to the process by which, given a set of examples on a particular domain, we construct mental categories that abstract particular examples and give us the skills to operate, in a variety of circumstances, on that particular domain. In this picture, kids and adults acquire novel mental categories by internalising from few examples given \cite[]{Vygotsky1986}.
\end{enumerate}
The previous notions give a (quite brief) account of the general assumptions of inferential communication. To summarise this section, we are assuming that in \emph{Language games} the meaning of a \emph{Communicative action} is determined from how it is used. We call the meaning an \emph{interpretation}. The \emph{interpretation} of an action requires associating intentions or goals to it. To determine these an agent needs to categorise the action with a large set of \emph{Natural Categories} interrelated with \emph{Conceptual Metaphors}. The agent also needs to monitor the \emph{Context} in which each \emph{Communicative action} will occur.
\emph{Causal inferences} are core mechanisms of the agents to do so. These inferences take the form of \emph{Conversational implicatures} and they are based on \emph{Relevance principles} shared among the agents.
All these skills are acquired through interactive mechanisms of imitation and cognitive development in a process called \emph{Internalisation}.
\section{A general cognitive module}
\label{embodied-model}
\label{cog-module}
\label{embodied-reasoning}
\subsection{The model}
Sperber convincingly argues that the specifics of how we draw cognitive inferences is best explained through massive modularity , i.e., the existence of a massive amount of interrelated functional modules \cite[]{sperber1996ecn}. An example he draws is the use of masks in primitive rituals, and how the functional module specialised in faces is here used in an altered way. In this sense, the evolution of culture is to be understood as how brain plasticity shapes such functional modules to better fit their use among social agents. Building on Sperber's vision, we assume that in inferential communication an agent is composed of a massive amount of interrelated functional modules.
To formalise inferential communication among humans and artificial agents, we need to characterise such functional modules, and do so in a way that works both for humans and for artificial agents. To do so we introduce a formal model that we call a \emph{general cognitive module} (GCM). Despite it is non-linear and difficult to treat, it has the advantage of being quite general.
Our working hypothesis is that such a structure is general enough to be used to implement the different cognitive mechanisms previously outlined, in the same way that combinations of NAND logical gates can be used to compile any digital algorithm.
Despite we do not provide a proof of such an assumption, we do try to show that under certain assumptions such a model fits with well understood cognitive systems.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{block-diagram
\end{center}
\caption{ The different elements of a general cognitive module. On the higher half, in bold, the fast pathway. On the lower half, in plain text, the slow pathway. }\label{fig:1}
\end{figure}
A GCM is composed of two pathways, the fast one and the slow one. The fast pathway is defined as a transfer function that is an invariant, non-linear filter with a response function $\boldsymbol{h}$. This pathway is defined by the equation $\boldsymbol{o(t) = h(p(t),n(t))}$, and it is composed of:
\begin{enumerate}
\item $\boldsymbol{o(t)}$: A stochastic output arriving through from electrical synapses. In the case of a neuron this is the axonal output firing action potentials. In the case of a piece of brain tissue this can be a statistical distribution of action potentials.
\item $\boldsymbol{p(t)}$: An array of stochastic inhibitory inputs . For example, in the case of a neuron or a piece of brain tissue these will be inputs from inter-neurons arriving through from electrical synapses. This signal would define whether the output is \emph{possible} or not.
\item $\boldsymbol{n(t)}$: An array of stochastic excitatory inputs . For example, in the case of a neuron or a piece of brain tissue these will be inputs from pyramidal neurons arriving through from electrical synapses. This signal would define whether a possible output would \emph{necessarily} occur.
\end{enumerate}
The slow pathway updates the previous response function.
It uses the following signals:
\begin{enumerate}
\item $r(t)$: An array of reward inputs. For example, in the case of a neuron or a piece of brain tissue these will be inputs from chemical synapses.
\item $l(t)$: A learning input which, for simplicity, we assume can activate or disable learning. In a neuron or brain tissue this would be chemical inputs from the neuroglia, enabling or disabling metabolic plasticity.
\item $m(l,r)$ A metabolic response, which determines the update of the transfer function: $h'=m(l(t),r(t))$. When considering $h'$ we may also want to consider the previous transfer function $h$ as a constant in the metabolic response $m(l,r)$.
\end{enumerate}
Overall we can think of general cognitive modules (GCMs) as a scale independent, simplified representation of a computational mechanism for cognitive processing. As such, a correspondence can be established with a biologically realistic neuron, but also with a fragment of brain tissue, to a unicellular animal, or even to more sophisticated natural cognitive systems. It can also be matched to an artificial neuron, or to different machine learning architectures.
\subsection{GCMs in natural cognition}
To illustrate how a GCM fits with the idea of a human brain analysed as a massive amount of GCM modules, we show how it fits with cognitive and computational neuroscience. Neuroscience textbooks give us some hint on how a brain tissue might instantiate a functional module in practice. The nervous tissue constantly balances inhibitory and excitatory signals \cite[]{buzsaki}. Typical examples of excitatory pathways are pyramidal cells in the brain cortex, which transmit signals at long distances. Typical inhibitory signals in the brain cortex are inter-neurons, which are shorter, inhibitory pathways, and more abundant than excitatory ones. In a GCM module, inhibitory pathways correspond to $\boldsymbol{p(t)}$, and excitatory pathways correspond to $\boldsymbol{n(t)}$.
We do not have a clear understanding on how information is encoded in the central nervous system. Despite a common assumption in modelling approaches is to study neural firing rates, it has been shown that temporal patterns also play a major role in coding perceptual information (see, for example, \cite{Gutig2006}). We therefore need to consider different encoding options are possible. However, what we do know that in brain tissue outputs are determined by, first, not having inhibitory inputs and, second, having enough excitatory input to stimulate an output. In a GCM, outputs are defined by $\boldsymbol{o(t)}$, and the relation between input and output determines a, comparatively, fast, non-linear transfer function, denoted as $\boldsymbol{h(p,n)}$.
Opposite to the fast pathways, chemical synapses are complex chemical reactions, with a slower processing time, and determine the metabolic consequences of excitatory and inhibitory input.
The dopamine system, along with several other chemical signals, can inhibit or activate the functional plasticity of the tissue (in a GCM, this corresponds to $l(t)$). Other signals will make the response of the cell change to respond better or worse to fast stimuli (in a GCM, this corresponds to $r(t)$).Chemical synapses can also trigger proteine transduction from gene expression, which ultimately change the shape of cell and the non-linear transfer function between inputs and outputs. In a GCM, this corresponds to $m(l,r)$.
The previous picture is quite general, and it is consistent with brain areas with very different functions, different connectivity, and different stimuli-response electrical patterns (for example, between the cerebellum, the prefrontal cortex and the basal ganglia).
Another role of chemical communication, beyond synapses, is to shape network connectivity, mostly through chemical gradients spread across brain tissue.
Arguably, the mechanisms that are behind the establishment of network connectivity are what determine the brain's extreme parallelism, high level of recursivity patterns, as well as specific recursive circuits such as, for example, thalamo cortical loops for multimodal integration, or cortico-basal and ganglia-thalamo-cortical loops related with the dopamine reward system.
Moreover, the capacity of the brain to change function and connectivity is not only related with neural activity. It also involves the support system. For example, astrocyte signaling regulates plasticity \cite{ota2013role,haydon2015astrocytes}. Other phenomena, such as brain cell migration in early neurodevelopment \cite[]{tierney2009brain}, also has an important role in brain plasticity. The mechanisms supporting network connectivity are not defined in the GCM. We will later introduce some principles to guide connectivity between GCM modules. However, before doing this, we turn towards artificial systems, to show how GCM modules can formalise, also, computational agents devoid of any biological realism.
\subsection{GCMs in artificial systems}
\label{specific-cases}
To illustrate the generality of this formalism we now discuss some specific, well known, cases, from the perspective of such a model.
\subsubsection{A Classic Communication System}
If we target at a classic communication system \cite[]{shannon48}, we can design a receptor as a single instance of the General Cognitive Module previously outlined. This can be done quite simply by doing the following additional assumptions:
\begin{enumerate}
\item The slow pathway is completely static ($h'=h$), Reward inputs are null ($r(t)=0$) and Learning inputs are also null ($l(t)=0$).
\item The excitatory input array ($\boldsymbol{n(t)}$) becomes, simply the baseline signal, i.e., the signal transmitted after the radio-frequency filter and the demodulation step.
\item The inhibitory inputs ($\boldsymbol{p(t)}$) are a digital switch, encoding the connectivity in the graph of possible symbol transitions
\item The transfer function ($\boldsymbol{h(t)}$) is a simple array of linear and invariant matched filters responding to the excitatory input, turned on or off by the inhibitory signal
\item The fast pathway has a stochastic output ($\boldsymbol{o(t)}$) defined by the modulation technique (defined in the matched filters) the coding (defined by the possibility signal in the inhibitory inputs), and the received input (defined in the necessity, excitatory input).
\end{enumerate}
\subsubsection{A Video game Character}
A typical way to create an interactive character in video game industry is to combine a behaviour tree, which makes the decisions, together with a hierarchical state machine which takes care of blending the relevant animations.
A behaviour tree is defined as a tree-like graph, where different control nodes will select which branches are active. Some of these control nodes will evaluate binary conditions, and some will simply select leaf nodes sequentially, or randomly.
Leaf nodes, in turn, will execute a behaviour, and return a binary outcome.
Executing a behaviour means a virtual character being rotated or displaced in a 3D environment, often also triggering a kinematic animation, often through a hierarchical state machine. The transition graph of the state machine will define the possible transitions between animations. Additional parameters will define when these transitions can be triggered.
Skeletal animation techniques \cite[]{magnenat1988joint} will be used to render the body mesh. \footnote{ Different techniques for facial animation are possible, but for the purpose of this analysis we will assume there are no facial expressions involved.}
Such an architecture can also be considered a particular case of our GCM. Below we outline a possible implementation:
\begin{enumerate}
\item The input from the environment ($\boldsymbol{n(t)}$) will simply be a set of binary conditions, perceived from the virtual environment inhabited by the character.
\item The inhibitory input ($\boldsymbol{p(t)}$) will be defined by the animator. It will define the graph of possible transitions, i.e., when an animation can transition to another animation.
\item The output ($\boldsymbol{o(t)}$) will have three parts:
A) a skeleton, i.e., a graph specifying the connectivity between bones, and the size of each.
B) the 3D position of the root node
C) a set of 3D rotations, corresponding to each node in the bone topology, generally expressed as quaternions.
\item the transfer function ($\boldsymbol{h(t)}$) can be thought as a set of impulse responses of linear invariant filters, one for each kinematic animation stored. The selection of each impulse response will depend both on the possibilities enabled by the transition graph.
\end{enumerate}
Of course, it would also be possible to consider the decision and animation parts as separate GCM modules, where the output of the first is connected to the input of the second. However, for this purpose we need some kind of logic that helps us connect the different modules, which we will address in next section. Before this, we illustrate with an example how the slow path of a GCM can be used.
\begin{figure}
\includegraphics[width=8cm]{RL-diagram}
\caption{Our diagram for a GCM integrated in the general reinforcement learning framework }
\label{fig:2}
\end{figure}
\subsubsection{Character animation based on reinforcement learning}
Some of the most impressive demonstrations for computational cognitive systems are
based on Deep Reinforcement Learning. Recent systems have demonstrated super-human features in classical games such as go or chess. However, to further illustrate specific cases of the GCM introduced we choose to analyse an example closer to our main interest: the use of reinforcement learning for physics-based interactive character animation \cite[]{peng2018deepmimic}\footnote{A general introduction and demonstration videos can also be found in https://xbpeng.github.io/projects/DeepMimic/index.html}. Such a system uses the same structure than a general reinforcement learning problem \cite[]{sutton2018reinforcement}. In figure \ref{fig:2} we show how our GCM fits in a RL framework. Compared to the previous section, the main changes are:
\begin{enumerate}
\item The slow pathway is not static ($h' \neq h$). Reward inputs ($r(t)=0$) are defined from the connectivity of the agent with the environment. Learning inputs ($l(t)=0$) will be an inverted step function: one for the training phase, null for the animation synthesis step.
\item In the fast pathway, the transitions between animations are not determined by an external state machine. Therefore the inhibitory input ($\boldsymbol{p(t)}$) is not defined. The transfer function ($\boldsymbol{h(t)}$) is not anymore a set of linear invariant filters, one per each animation. Rather, it is a non-linear function with internal recursivity: triggering an animation means increasing the importance of an attractor in a dynamic system. The stable state of the attractor is the synthesis of an animation. The excitatory inputs ($\boldsymbol{n(t)}$) are not only the binary variables that previously triggered a given animation, but also dynamic inputs such as the last pose of the character, as well as arbitrary parameters, typically target positions and orientations of skeletal nodes such as the root, a feet or a hand.
\end{enumerate}
The previous examples show how we can consider a GCM as a generalisation of diverse specific, well known, cognitive systems. However, as already stated, we do not know anything about the connectivity and dynamic exchanges between different GCMs and agents (made of GCMs). We address this in next section.
\section{A Formalism for Inferential communication}
In this section we introduce formal definitions to relate the events and the pragmatic inferences characteristic of inferential communication. Theorem proofs are in the appendixes.
\subsection{Events and Predicates}
First, we formalise events and predicates associated to them.
\begin{definition} A specific \textbf{event} is denoted with letters such as
`$a, b, c, ...$'. A generic event is denoted with the letters `$x, y, z$'.
We may use sub-indices if further variables are needed, such as $x_1$. We assume an event $e$ has a beginning in time $t_e$, and a duration $d_e$.
\end{definition}
Notice that these events occur in a specific time period, and as such they are not the abstract, instantaneous events, found typically in probability theory.
\begin{definition} A \textbf{property} is denoted with letters such as `$A, B, C, ..., P, Q, ...$'. We may also sub-indices if further variables are needed, such as $P_1$.
\end{definition}
We also introduce an epistemic definition of logical quantifiers relative to the notion of \emph{a cognitive system}.
\begin{definition}
A \textbf{cognitive system} is defined as the combination of one or more GCM instances connected between them through at least some of their inputs and outputs.
\end{definition}
\begin{definition}
The symbol `$\exists^{st}$'
denotes the existence of an event external to a cognitive system. We read it as \textbf{standard existence}. For an event `$e$' we write `$\exists^{st}e$'. For a property `$P(e)$' we write `$\exists^{st}P(e)$'. An event `$i$' whose existence is represented internally by the cognitive system, but not necessarily associated with an external event, it is denoted as `$\exists i$'. Conversely, a property associated to such an event is denoted `$\exists P(i)$'.
\end{definition}
To denote an event with standard existence we call it a \textbf{standard event}. It is important to notice that this does \emph{not} correspond to the classical idea of events as atomic elements, and properties as equivalent to predicates defined as sets of atomic elements. Here \emph{standard} events are associated to the input of a cognitive system, and properties are associated to its outputs. As such, if the outputs of a cognitive system is connected to the input of a second cognitive system, the property will become an event \emph{from the perspective of the second cognitive system}. The transition from standard existence to general existence will be discussed in detail in later sections.
Consistently with the previous definition, we also introduce the standard universal quantifier:
\begin{definition}
The symbol `$\forall^{st}$'
denotes \textbf{all standard events}. The predicate `$\forall$' denotes all events.
\end{definition}
The epistemic nature of the standard existential and universal quantifiers is particularly visible when the cognitive system under consideration is an entire agent. In this case, `$\exists^{st}$' denotes existence of an event \emph{in the world}, as perceived by an agent. The event is the external stimulus, and the property is the perception. Different from this, `$\exists$' denotes existence in more abstract terms, as is done in usual mathematical logic. For example, it could refer to a hypothetical event, independently of whether it is associated to an external stimuli or not.
A class of properties that will become very important are those that can also be considered sets:
\begin{theorem}
The outputs of all non circular cognitive systems are Sets.
\label{predication-theorem}
\end{theorem}
\begin{definition}
\textbf{Predicates} are properties that are also sets (i.e., that are the outcome of non-circular cognitive systems).
\end{definition}
Since predicates are also sets, they are close to the sense of predicate given in predicate logic and first order logic. We can therefore introduce the usual operations of set theory for predicates:
\begin{definition}
The operator `$\in$' denotes a relation between an event and a predicate. It reads as \textbf{belongs to}, or \textbf{in}. For example, $ a \in P $ denotes `$a$ satisfies $P$' (i.e., $a$ is associated with predicate $P$). We also denote this relation as $P(e)$.
We may also consider fuzzy belonging, denoted as $\in_{\alpha}$, with $\alpha \in [0..1]$. We also use $\notin$ to denote the negation of $\in$.
\end{definition}
\begin{definition} To combine predicates we use the the union ($\cup $) and intersection ($\cap$) operators. To denote two predicates are equivalent we use `$\equiv$'. The opposite is denoted as `$\not\equiv$'.
\end{definition}
We now have a complete formalism to combine and compare \emph{predicates}.
\subsection{Causal Implicatures}
To formalise \emph{causal implicatures} we introduce event sequences:
\begin{definition}
A \textbf{chain of events} is denoted with the `$.$' operator. For example `$a.b.c$' denotes a chain of three events. The only constraint introduced by this operator is that the occurrence of the events satisfies a partial order, i.e., that $a$ starts occurring before or at the same time than $b$, etc. Therefore, if $t_a$ denotes when $a$ starts occurring, `$a.b.c$' implies $t_a \leq t_b \leq t_c $.
\end{definition}
There are no constraints regarding when events end (i.e., since events have a duration, $a$ may finish occurring after $b$). From the previous, we can define:
\begin{definition}
The \textbf{context} $x$ of an event $y$ is denoted as $y^*$. It is defined as a chain: $y^* \equiv x_1 . x_2 . ... . x_n$. It does satisfy the condition that $t_{y^*} \leq t_y$.
To denote a context $x$ as perceived by an agent we write $X$, i.e., the predicates associated with the events in $x$.
\end{definition}
\begin{definition}
The \textbf{implied cause} $x$ of an event $y$ is denoted as $y^{**}$.
It does satisfy the condition that $t_{y^{**}} \leq t_y$.
\end{definition}
\begin{definition}
The relation between an event $e$ and its implied cause $e^{**}$ is called a \textbf{causal implicature}, and we denote it as `$e^{**} \Rightarrow e$'.
\end{definition}
\begin{theorem}
For a given event $e$, it is always possible to build a cognitive system that can determine the most likely implied cause using an information-based metric such as surprise minimization.
\end{theorem}
\begin{definition}
An \textbf{inferential agent} is an agent that builds causal implicatures based on information-based metrics such as surprise minimisation.
\end{definition}
In this picture, an inferential agent trying to interpret a communicative exchange will build pseudo-complemented semi-lattices to establish causal implicatures and select the construct that minimises uncertainty.
\begin{theorem}
An information exchange between two inferential agents with goals is not stationary.
\end{theorem}
The previous result shows in what sense this theory generalises classical information theory: contrary to traditional information theory where events are part of a stochastic stationary process, here the probability of occurrence of an event will vary across time.
\subsection{Cognitive processes for competence acquisition}
\label{internalization-principles}
An important aspect of inferential communication is that it presupposes \emph{competence acquisition}. How do we internalise mental categories and build conceptual metaphors to process everyday events? We have introduced an epistemic difference between standard and non-standard events, but not determined how they relate, or how do we build abstract representations (like `furniture') from specific examples (like `this table', or `that chair').
How do we go from \emph{standard} events, i.e., from experiencing concrete events associated to one or several predicates ($\exists^{st} P(a)$), to a predicate \emph{in general} ($\exists P(a)$), which generalises across concrete and hypothetical events? We break down this learning process in three cognitive processes:
\begin{principle}
We denote the process of \textbf{Idealisation (\emph{I})} as `$\overset{I}{\Longrightarrow}$', and define it as:
\begin{equation}
\Big[ \forall^{ st fin} y' \exists x \ \ \Big/ \ \ \forall y \in A(y') \Big] \overset{I}{\Longrightarrow} \Big[ \exists x \forall^{st} y \in A(y) \Big]
\end{equation}
\end{principle}
In the previous $ \forall^{ st fin}$ reads \emph{for all standard finite}. The intuition is that if a cognitive system has a property that applies to all standard finite examples to which it has been exposed, then it is possible to implement a process, which we call \textbf{Idealisation}, that makes this predicate apply to \emph{any} standard examples (thus, the ones seen, but also possible examples not yet seen).
Of course, this occurs spontaneously in humans and other animals. However, the implementation of such a generalisation mechanism in artificial systems is a central problem of machine learning: finding a function that fits best the training set but also generalises to stimuli beyond the training set is the main challenge when training artificial neural networks in classification problems.
In the previous expression $x$ stands for the context in which such an idealisation process occurs. The main idea is that this idealisation will only work in some contexts, not all of them.
This highlights the importance of contextual factors when associating predicates to events. Indeed, contextual factors are of paramount importance in inferential communication. One of their main roles is to help determine subsets of the domain of a predicate.
We can express this idea as follows:
\begin{principle}
We denote the process of \textbf{Selection (\emph{S})} as `$\overset{S}{\Longrightarrow}$' and define it as:
\begin{equation}
\forall^{st} X \exists^{st} Y \forall^{st} z \Big[ z \in Y \overset{S}{\Longrightarrow} z \in X \cap Y(z) \Big]
\end{equation}
\end{principle}
The intuition behind the \textbf{Selection} process is that, given a predicate $Y$ associated to standard events, we can implement a cognitive process, which we call \textbf{Selection} that, given any predicate $X$, is able to determine which subset of events $z$ satisfy predicate $Y$ and also $X$.
If we think of properties $X$ and $Y$ as implemented in separate GCMs (which we can call $GCM_X$ and $GCM_Y$), this principle can be implemented at least in two ways. First, with the output of $GCM_X$ feeding the inhibitory inputs $p$ of $GCM_Y$, and therefore affecting the domain on which $GCM_Y$ responds to stimuli consistent with predicate $Y$. Second, introducing a third module that processes the outputs of both $GCM_X$ and $GCM_Y$ and responds only to the combination of both. However, implementing a separate module for \emph{any} combination of predicates that have some domain overlap can produce a combinatorial explosion. Therefore, to implement this principle it seems more efficient to opt for contextual influence of GCM modules that have domain overlap, despite the dynamics of contextual influence will be more challenging to understand.
The third cognitive principle extends the domain of predicates from \emph{standard} events to events in general:
\begin{principle}
We denote the process of \textbf{Transference (\emph{T}) } as `$\overset{T}{\Longrightarrow}$' and define it as follows:
\begin{equation}
\Big[ \forall^{st} y A(y) \Big] \overset{T}{\Longrightarrow} \Big[ \forall y A(y) \Big]
\end{equation}
\end{principle}
The intuition behind the \textbf{Transference} principle is that if in a cognitive system there is a predicate that applies to all standard members, then we can implement a cognitive process, which we call \textbf{Transference}, by which we can make the predicate apply to all members. This mechanism allows to transfer the properties of standard members to all members. For example, to hypothetical sequences of events.
This undeniably happens in natural reasoning, where we learn to make inferences on hypothetical situations and events, and we even build conceptual metaphors to enrich the predicates associated with factual or hypothetical events. However, current artificial cognitive systems do not learn to generalise a predicate beyond the domain of standard examples on which it is trained. In artificial systems it is easy to simulate standard events using stored media to be presented as stimuli to a machine learning algorithm.
However, it is difficult to see how we may implement in an artificial agent the notion of \emph{conceptual metaphor} to interpret the events involved in everyday communication. To do so, we would first need to clarify what are, in an artificial system, events that exist but do not have a standard existence. We can however assume such agents can be built, and explore their properties:
\begin{definition}
A \textbf{competence acquirer} is an agent where the 3 cognitive processes
\textbf{I} ($\overset{I}{\Longrightarrow}$), \textbf{S} ($\overset{S}{\Longrightarrow}$) and \textbf{T} ($\overset{T}{\Longrightarrow}$) are implemented.
\end{definition}
\subsection{The consistency of inferential agents}
\label{maths}
If we implement artificial agents that have inferential skills and are also capable of competence acquisition, will they develop consistent representations? We believe this is not the case in general, in the sense that \emph{any} inferential agent will develop consistent representations. However, here we show that a particular subset of inferential agents can develop consistent representations.
\begin{definition}
A binary agent is one where all the processes are binary rules. This applies to the transference functions in the GCM modules that form it and also, if implemented, to the three cognitive processes for competence acquisition previously defined.
\end{definition}
\begin{theorem}
There exists a binary inferential agent that is also a competence acquirer which is a consistent embodiment of mathematical foundations.
\end{theorem}
More qualitatively, the way this theory is built also seems closer to an embodied account of mathematical foundations \cite{lakoff2000mc}: numbers are defined as Frege numerals, which is not only the first introduced historically, it is also more intuitive: for example, the number 3 can be defined as a property shared by all sets which have 3 elements. This seems particularly appropriate here: if we are to build mathematical foundations consistently with how we build our internal representations, it seems reasonable that the definitions of notions for which we have an intuition should match such intuitions (intuitions which are, presumably, built from embodied representations). Similarly, in this set theory the universal set exists, the complement to a given set always exists, and it is possible to create any set simply through the union of two existing sets. In addition, with the 3 IST principles (see appendix) we can formalise calculus closer to how it was introduced historically, with infinitesimals, circumventing the need for definitions based on limits, and stating closer to the, arguably, intuitive notion of change at arbitrarily small increments. It can also be used to formalise probability theory in a much more simple way than the usual way, based on complicated results from measure theory \cite{nelson1987rep}.
\section{Discussion}
\label{discuss}
\subsection{Inferential communication}
The publication of \emph{A mathematical theory of communication} \cite[]{shannon48} spread the view of communication as the transmission of a message encoded by an emitter and decoded by a receptor.
Hardly ever a theory developed for engineering purposes has had such influence in fundamental research fields such as physics, linguistics, biology and cognitive sciences. However, people studying natural communication started to raise some objections to such view of communication and they eventually proposed the inferential communication model \cite[]{sperber1986relevance} to give some account of aspects of everyday conversations not described by classical information theory. The benefit of such a theory
is that it allowed to explain why communicators convey much more information with their utterances than what is contained in their literal expressions. We have introduced a formal model that addresses this challenge and hints at some possible ways to engineer practical solutions to such interaction scenarios.
\subsection{Model summary}
In inferential communication an agent is an entity which processes information. This information is received as a stream of events. \emph{Events} are mainly actions of other communicative agents, for which we assume such events are intentional actions, i.e., events generated by an agent to which other agents associate intentions.
However, part of the events can also be natural events that may be perceived as relevant for communication purposes. In both cases, when they receive information in the form of events, inferential agents associate an interpretative meaning to these events. This cognitive process, that we call \emph{interpretation}, consists in associating the events with communicative intentions, within a context that includes a given situation, and a cultural background assumed as shared among the different agents. It can, sometimes, provoke agents to perform additional actions, either as an immediate reaction to them, or as a delayed answer.
We have argued this picture can be formalised in MaTIC: within the space of \emph{Language games}, communication is a collaborative game where each agent has a set of goals, some of which are shared. Grice's cooperative principle can be formalised as a goal shared among the participants in the game. The intentions of individual agents can be formalised as a set of goals. The events defining the game are communicative actions, and are intended to be perceived, and associated with predicates.
In MaTIC the idea of a standard event is associated with an epistemic definition of event, but also with a rigorous logical treatment. The particular predicates associated with an event are selected according to the general principles of relevance theory, with information theoretic measures. These associations are learnt through a combination of adaptive GCM modules, together with connectivity changes guided by the I, S and T principles. The existence of a property is associated with the output of a cognitive module. Predicates are properties organised in non-circular connectivity, something that allows building richer representations.
In this picture, cognitive agents infer causal implicatures for the occurrence of an event $e$ by, first, finding the conditions in which $e$ \emph{can} occur and, second, finding a motivation for $e$ to occur. In other terms: the context determines the \emph{possibility} of $e$. Correspondingly, the event or intention that determines the \emph{necessity} that $e$ occurs, constitutes the cause that we associate with it. The way a cognitive agent will do so will be based on heuristic reasoning, under general principles of uncertainty minimization.
The main difference between cognitive agents involved in inferential communication with the other existing systems introduced is that in those examples of GCM instances
the inhibitory input $p(t)$ was either null, or either it simply restricted the response of a system to a subset of possible outcomes. In addition, the inhibitory input function was determined \emph{beforehand}, i.e., before any excitatory input occurred. This does not seem appropriate to address the dynamic interplay that occurs in inferential communication. In the conversation example previously introduced, the occurrence of an event such as `\emph{B says: There is a garage around the corner}' was, arguably, made possible by a larger context, such as B knowing this fact, as well as the previous events. In addition, assuming none of the participants had hidden intentions, an event such as \emph{A says: I am out of petrol} would be generally interpreted as the cause of B's statement.
More generally, as new exchanges occur in a conversation, the context of the conversation will evolve, and therefore the set of possible actions will too. \emph{Contexts in natural communication are dynamic}, therefore we need to consider GCM instances with response functions which combine the $p$ and $n$ inputs in a more intricate way. Overall, MaTIC does not define completely how causal implicatures occur, but it does define a set of \emph{requirements} that must be satisfied by the agents exchanging intentional actions. Under these requirements, and contrary to classical information theory, the probability of occurrence of an event changes across time.
\subsection{A neuroscience perspective}
Is MaTIC based on reasonable assumptions, from which we can build further details? Contemporary cognitive neuroscience suggest there may be empirical support for a computational theory of inferential communication. The mirror system is assumed to enable imitation, and empathy, and to be based on sensor-motor neural mechanisms \cite{kohler2002hearing}. Such bodily disposition would explain the skill of learning other's intentions \cite{barron2020neuronal}. In fact, embodied and enactivist approaches to cognition are connecting with reasonable plausibility the spheres of bodies, societies, and cultural manifestations, with (inferential) communication as a multi-modal skill \cite{gallagher2018active}.
The modern study of intentionality
encompasses both mental states towards things as well as the ways by which human agents understand others intentional dispositions \cite[]{brentano2014psychology}. This atonement process is located in our neural system, working from very basic sensor-motor level to symbolic processing \cite[]{rizzolatti2007mirror}, \cite[]{bretherton1991intentional}. For such reasons, there are mechanisms by which such intentional actions can be observed and therefore the related actions can be inferred to. In this sense, the study of intentionality allow us to predict and understand the next actions of human agents, allowing us to design a suitable answer. Indeed, in the model we propose agents will also, sometimes, perform additional actions as a reaction to the events and the predicates associated to them. According to this theory, this should be done in line with the first principle of relevance theory, i.e., the fact that \emph{human cognition tends to be geared towards the maximisation of the cognitive effect}.
There are other accounts on how abstract representations can be built through the combination of bottom up and top down processes.
Notably, predictive coding and active inference \cite[]{constant2020representation,friston2007free}, suggest information-theoretic measures are relevant to account for how the brain learns to interact with its environment.
Despite MaTIC is compatible with predictive coding and the idea that uncertainty minimisation can go quite far in determining cognitive processes, there are two aspects that make MaTIC strikingly different: first, in how it reconciles the idea of embodied reasoning with a formalism for mathematical foundations, and second in how it seems compatible with cognitive accounts of everyday exchanges, as conceptualised in cognitive pragmatics.
\subsection{Cognitive Pragmatics}
The process of \emph{interpretation} has also been interpreted in a Bayesian framework
\cite{vallverdu2015bayesians}.
We believe the model introduced here may complement Bayesian models of thinking processes by taking into account the other embodied variables present into communicative processes. One of the key assumptions of Bayesian inference is that a belief and its opposite (in our notation, $A$ and $\neg A$) exclude each other: they form a complementary set. However, this is not a general assumption, for example, in deep learning inference systems, where Bayesian models are just a subset of the solutions that, in practice, work.
In this context, local restrictions may explain why the agents opt for selecting some informational sets and processing models instead than others: cultural and training/academic variables are fundamental thresholds of the way of dealing with information
(see the notion of blended cognition \cite{vallverdu2019blended}).
Future work should analyse whether this model does different predictions than Bayesian modelling, and which gives better support to the analysis of the mental mechanisms that are used to process intentional communication.
In this context, we also point to Rational Speech Act Theory \cite[]{yuan2018understanding}, which provides a fully formalised framework for pragmatic inference under a Bayesian framework. For example, \cite{goodman2013knowledge} shows that pragmatic agents do inferences that seem more adjusted to reality than their literal counterparts. Beyond its insistence on maximising utility, Rational Speech Act Theory seems consistent with a Bayesian picture, and is reasonably aligned with general principles related with reducing uncertainty in information exchanges, close to the idea of minimising surprisal, as discussed in the previous section. However, beyond the question of whether Bayesian modelling covers the whole picture, there are two aspects of MaTIC that do not seem addressed in the Rational Speech Act Theory: first, how such inferences can be supported by an embodied mechanism (for which we have proposed the GCM, and shown how biological and artificial systems can be seen as particular instances of such a modular embodiment) and, second, how the internal representations supporting pragmatic inference can be developed and maintained through time.
\subsection{Embodied reasoning and representation consistency}
A particularly original aspect of MaTIC is its link with formal set theory. In this context it is striking that actual infinity is not assumed. Instead, we have assumed the existence of external events, and predication, and argued we can build most of the usual construction for mathematical engineering (sets, numbers, vectors, functions, metrics, derivatives) from it.
If our model of competence acquisition is correct, the mechanisms that we posited for acquiring communicative competence should also apply to acquiring mathematical competence. However, this is at odds with how the foundations of mathematics are generally assumed by mathematicians. For example, in \cite{cohen1966set} we find an exposition of the default set theory, also called Zermelo-Frankel with Choice (ZFC). Such a theory assumes there only exist sets, and that there exist an actual infinite amount of them.
This is absolutely at odds with the notion of an agent developing internal representations through interactive sensor-motor loops in a finite world. In this context, it is particularly striking that our theory does not assume the actual existence of an infinite number of sets to build solid mathematical foundations. Rather, it is enough to exist the existence of internal and external events, and GCM modules processing them.
Future work should explore if MaTIC can also can be used to build agents that develop consistent internal representations from data, beyond the domain of mathematics. For this purpose, we believe again virtual reality provides an ideal experimental environment, one where artificial agents can easily build internal representations from their virtual environment. This would strongly simplify the perception challenges generally involved in robotics, and allow exploring directly whether agents in an environment can build internal representations of it, and whether this derives in exchanging information under pragmatic principles.
\subsection{A falsable theory}
Any good scientific model must do accurate predictions that can be validated experimentally. The main implication of MaTIC might be the idea that by building artificial agents that satisfy the three cognitive principles introduced, then using data and machine learning these agents should be able to achieve inferential communication as competently as native speakers, and generate responses appropriate to their behaviour and also to their own communication intentions. We believe the best way to show this is through the iterative development of VR scenarios involving social interaction between autonomous agents and humans, based on these premises.
Another aspect of a scientific model is that it should be falsable. There are several ways to show this theory is wrong. For example, by:
\begin{enumerate}
\item showing that the three cognitive principles introduced are not relevant for the purpose at hand.
\item showing that massive modularity does not occur, after all, in animals nor humans, or that it is not needed to build artificial systems that are capable of inferential communication.
\item showing that predication can be built with circular systems, that there is no difference between circular and non-circular systems
\item showing that contextual processing does not occur through inhibitory inputs of GCM modules.
\end{enumerate}
Further computational systems could also be analysed within MaTIC. Here we are not short of options: Sperber has often stated that relevance theory is actually a general model of cognition. Since MaTIC is based on formalising some aspects of this theory, and on how to embody it, there is ground for challenging this theory with virtually any modelling approach used in systems neuroscience or in artificial intelligence focused on machine learning.
A separate stream of work analysing interpersonal interaction using methods derived from the modelling of dynamic physical systems that should be explored\cite[]{
zhai2016design, alderisio2017interaction, lombardi2019deep}.
Future work should also focus on validating MaTIC in behavioural studies of people interacting with agents embodied in interactive virtual characters.
Another way to explore whether MaTIC does valid predictions is to to review insights from behavioural economics and cognitive psychology on decision making. The discounting effect and similar paradoxes have been well characterised in behavioural economics, but are difficult to explain with classical probability theory. We may study whether an agent learning on these premisses will also adopt heuristic decisions that reflect these cognitive biases.
Another direction to explore is to clarify whether MaTIC gives a better account for inferences and predicates as they occur \emph{in the world}, i.e., consistently with how cognitive psychology characterises natural categorisation and how cognitive pragmatics argues we reason and make decisions. From a developmental perspective: do the interpretation of the I, S and T principles introduced capture part or all of the picture in Vyggotsky's concept of \emph{Internalisation}? These topics are being addressed by ongoing neuroscientific work \cite[]{teufel2020forms}, and it would be interesting to explore whether our theory is compatible or helps explain how such developmental processes work. Empirically, we could also look at the extent to which artificial agents developing their skills under such principles incur in phenomena such as "hyper-regularisation", i.e., the trend kids to learn general rules for syntax, and then have to re-learn exceptions to the rule that they had initially learnt.
\subsection{Ethics}
Last but not least, a different question to consider is what ethical implications should be addressed if we were to implement such systems in practical, functional, consumer products. The arrival of immersive virtual and augmented reality to the consumer market, is likely to have a considerable social impact particularly if combined with autonomous agents capable of everyday pragmatical inferences as humans do everyday spontaneously and often unconsciously. Depending on how it is implemented it can provide opportunities or challenges to the realisation of human rights in society. Let's introduce a simple example: the always winning Janken robot from Ishikawa Oku Lab (Japan), which has absolute winning rates playing at rock-paper-scissors against humans \cite[]{katsuki2015high}. The robot is able to capture and interpret bodily movements of the human player and then react more quickly always beating the human, which has not been able to understand the unfairness of the situation, because the robot moves itself following a similar temporal pattern than that of humans. This mechanism applied to other contexts would make possible to open a new way of cheating humans in benefit of machine owners.
Is it reasonable to allow these super-human affordances to be used without the users' knowledge? \cite[]{vallverdu2016emotional}.
What considerations should be taken into account when our governments approach these topics?
\bibliographystyle{apacite}
\setlength{\bibleftmargin}{.125in}
\setlength{\bibindent}{-\bibleftmargin}
\section{Introduction}
A key assumption of Shannon's mathematical theory of communication \cite{shannon48} was that communication was limited to random samples from a finite set of symbols, such as for example the alphabet, or zeros and ones. As Shannon, stated: \textit{Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem.} However, modern communication have obviated this limitation. It is common in modern communication systems to interact with chat-bots, virtual assistants, and different sorts of conversational agents.
More dramatically, modern virtual reality systems enable us to capture, stream and render full three dimensional representations of human actors, across the auditory and the visual domains. Alternatively, such performances can also be created with digital production tools, on a labour-intensive practice by character-animation specialists. We can therefore create, store and replay animations of cartoon or realistic characters. Moreover, we can render interactive performances using Virtual or Augmented Reality headsets, thus recreating a communication scenario that, despite being technologically mediated, can be quite similar to spontaneous natural communication.
For such a communication scenario to match our expectations, artificial cognitive agents must infer (or appear to infer) the meaning of a certain message and, from a combination (or an abstract representation) of previously recorded performances, generate a response that is appropriate to the message received, and that is (or appears to be) aligned with the communication intentions of the cognitive agent. Several techniques can be adapted for this purpose, either from natural language processing, or from artificial intelligence literature focused on video game characters. These techniques, though, are often in stark contrast with the way such mechanisms occur in humans. We lack a communication theory comparable to the precision of Shannon's work to characterise precisely inferential communication, as well as to quantify the benefits and drawbacks of different approaches to create artificial agents capable of generating interactive behaviour consistent with these principles.
To address this gap, in this article we introduce a Mathematical Theory of Inferential Communication (MaTIC). The article is structured as follows.
After this introduction, we review how inferential communication is described in cognitive pragmatics and review broader cognitive principles. We review notions such as predication, natural categories, conceptual metaphors, and competence acquisition.
Next, we briefly review how cognitive and computational neurosciences characterise the embodiment of such principles. We review the concept of massive modularity as a cognitive organisation principle and discuss how such functional modules inter-relate in inhibitory and excitatory networks.
We then summarise the previous picture in a formal construct called the general cognitive module (GCM). We argue it summarises how cognitive sciences characterise the building blocks of embodied cognition. We also try to
illustrate the generality of a GCM we show how well-known computational models for autonomous agents can be interpreted as particular cases of such a construct.
We then turn to our main case of interest: how cognitive agents formed of networks of GCMs can learn to interpret the actions of other agents and generate appropriate responses. For this purpose, we introduce three cognitive principles for inferential communication that we label $I$, $S$ and $T$. We interpret these principles as mechanisms that shape the topology of connections among different GCM.
These guide \emph{competence acquisition} in an agent, and our proposal is that such principles are as valid for humans as for artificial agents.
A particularity of this theory is that it tries to reconcile the assumption that subjective inferences, as they occur in inferential communication, emerge from embodied mechanisms, with a formalist approach. We try to illustrate how these two aspects make a quite compelling picture in a domain where subjective inferences and a formalist approach converge: mathematical foundations as depicted from the perspective of embodied cognition. In this context, we show that the resulting theory under specific assumptions produces a set theory that, contrary to the axiomatic assumptions of ZFC set theory, the set theory assumed by default, is consistent with the picture of embodied cognition as a foundation for mathematics.
Finally, we discuss whether such a model allows adapting the rich tool set of information theory to characterise information exchanges between two inferential agents, how it compares with different proposals that try to define cognitive systems with general computational principles, and future research directions, both theoretical and practical.
\section{Inferential communication}
Contrary to the code model \cite[]{shannon48}, in which a communicator encodes a message to be decoded by an audience, the \emph{inferential} model of communication considers that \emph{communicative actions}, both verbal and non-verbal, should be understood as the expression and recognition of intentions: the communicator provides evidence of his intention to convey a certain meaning, and the audience infers the meaning on the basis of the evidence provided. This evidence can be something said together with some gestures and voice inflexions, but it can also be non-verbal, such as pointing, smiling, approaching, or any other action that the audience is susceptible of associating a meaning to it. The association of the evidence with a meaning is strongly influenced by the social and material context.
If classical information theory was largely inspired by the image of two agents coding and decoding signs through a telegraph wire, the inferential model of communication can be exemplified by the communication of two fully embodied agents sharing a three dimensional environment portraying in real time bidirectional behaviour and speech with the purpose of exchanging information and intentions related with this information. This model has never been formalised or tested in quantitative and rigorous terms such as information theory has. On the other side, it can give simple explanations to quite complicated phenomena. For example, in this view the reason why we are constantly playing with different layers of meaning when we communicate is simply for communicative economy. To give a better picture of this theory, we summarise some of the notions on which it is based:
\begin{enumerate}
\item \textbf{Language games}\cite{wittgenstein1953pit} characterised everyday communication as a \emph{language game}. To do so, he outlined a large number of examples on how the meaning of words and actions are determined by how we use them, in what context and with what intention. Therefore, we can define the class of \emph{Language Games} as all situations in which several people (or things like virtual actors) can perform \emph{communicative actions}.
\item \textbf{Natural Categories}
The notion of Natural Categories \cite[]{rosch1996frs} has been widely adopted in cognitive sciences, ranging from cognitive psychology to applied linguistics. Essentially, natural categories are the basic element that allows us to categorise the world and organise mentally cognitive knowledge. Formally speaking, they were generally identified with fuzzy sets \cite[]{zadeh1965fuzzy}.
This implies there are better and worse examples of each category, which it is often named as the level of prototypicality. Obvious examples of this can be found in colours: in face of a spectrum of different red samples we will attribute a ``typical red" example or a ``not so good red" example to different samples.
\item \textbf{Conceptual Metaphors}
In cognitive linguistics, the notion of conceptual metaphor as used by \cite{lakoff1980mwl} builds upon the idea of natural categories. In this view, conceptual metaphors are mappings between domains that give a richer meaning to how we communicate and think. Usual metaphors such as \emph{Life is a Journey}, or categories of different cultural systems, such as the category of \emph{Women, Fire, and Dangerous things} \cite{lakoff2008women}, to cite only the most famous examples, rely on the notion of Natural Category.
This interpretative framework has had numerous developments. For example, in arguing that basic mathematical intuitions rely on conceptual metaphors that are derived from experience, and that it is this which gives a lot of what we call mathematical understanding, independently of the demonstration process \cite[]{lakoff2000mc}.
\item \textbf{The Context of an action}
\emph{Communicative actions} related with interpersonal communication have a context which can include factors such as:
part of the physical situation, but also the cultural background assumed to be shared, the gender, social relation of the speakers and in general anything that contributes implicitly to determine what is meant by a certain communicative action.\footnote{The role of contexts in natural communication can be traced back to 1923 in the work of Malinowsky, but see chapter 2, section 1 in \cite{kramsch1993cac}, or the entry \emph{context et situation} in \cite{houde1998vsc}} A typical example of this is irony. If someone says \emph{I am very happy}, his assertion can be interpreted as meaning to express his happiness, or exactly the inverse, his extreme sadness. This will depend not only on how this is said, but also according to what we know about the person from what he did in the past. Some elements of the present situation can also have an influence in what is interpreted from what is said. Therefore, in natural communication agents must control very carefully the context in which \emph{communicative actions} are performed. Otherwise it will be difficult to ensure the meaning intended is appropriately conveyed.
\item \textbf{Intentions}
A different aspect of interpreting events when performed by actors is that we associate \emph{intentions} to them.
If we see someone in a certain situation or conversation, especially if we know him well or he is in a familiar socio-cultural context, we can spontaneously say what he will probably do or say, or at least a small set of different options, according to the situation, taking into account previous experienced contexts. This is a general cognitive ability of humans called \emph{intention attribution} or \emph{mind-reading}, which some impaired people seam to lack \cite[]{baron2000understanding}.
For example, consider the expression: \emph{This paper is white.}
Depending on the context in which it is said, the same text can convey a meaning similar to \emph{I’m surprised because in his exam he did not write a word}, or \emph{Fantastic, finally I found a blank paper and I can draw what I was wanting to} or \emph{so, you finally did not fill in the papers of our divorce, did you?}. The process by which we go from \emph{what is said} to \emph{what is meant} is determined not only by the meaning of the words and in what situation the communicative action is performed, but also by implicit assumptions in everyday interaction. All these are part of the context.
With the previous notions, we can summarise inferential communication in non-mathematical terms with the \emph{Principle of Cooperation} \cite[]{grice200lac} and the two principles of \emph{Relevance Theory} \cite[]{sperber2004rt}.
\item \textbf{The Cooperative Principle}
\label{coop-principle}
Grice's \emph{Cooperative Principle} was the first principle of communication proposed, and it would involve any everyday linguistic interaction --even non-verbal. The clearest explanation might be introduced by it's proponent \cite[]{grice200lac}):
\begin{quote}
Our talk exchanges do not normally consist of a succession of disconnected remarks, and would not be
rational if they did. They are characteristically, to some degree at least, cooperative efforts;
and each participant recognises in them, to some extent, a common purpose or set of
purposes, or at least a mutually accepted direction. This purpose or direction may be fixed
from the start (e.g., by an initial proposal of a question for discussion), or it may evolve
during the exchange [...]. We might then formulate a rough general principle which participants will be expected (ceteris paribus) to observe, namely: Make your conversational contribution such as is required, at the stage at which it occurs, by the accepted purpose or direction of the talk exchange in which you are
engaged. One might label this the \emph{Cooperative Principle}.
\end{quote}
A classical example adapted from \cite{grice200lac} will frame better the problem. Consider this scenario:
\begin{quote}
A is standing by an obviously immobilised car. \\
B approaches him. \\
A says: I am out of petrol. \\
B says: There is a garage around the corner. \\
\end{quote}
In the previous example, the last reply of B would be a consequence of the previous exchanges.
To formalise such a mechanism we will need a definition of implication ($\Rightarrow$) general enough to include these \emph{implicatures}.
\item \textbf{Relevance Theory}
Sperber and Wilson's Relevance Theory \cite[]{sperber2004rt} develops two basic principles to explain the inferential model of communication. The first principle of relevance theory or cognitive principle, states that \emph{human cognition tends to be geared towards the maximisation of the cognitive effect}, and this cognitive effect is quantified with a relevance measure. This implies that, given a certain context, a communicative action will be performed in order to maximise the relevance that the receiver will associate to it.
The second principle of relevance theory or communication principle adds an extra layer. It states that \emph{every ostensive stimulus conveys a presumption of its own optimal relevance}. This is best shown with an example. If two people, Alice and Bob, are chatting in Alice's living room, Bob can put his empty glass in Alice's line of sight to suggest that he might want more water, but this is only exploiting Alice's cognitive tendency to maximise relevance. Bob still has to show you that he put the glass there \emph{on purpose}, that it was his intention, and therefore that it is worth processing this stimuli in Alice's cognitive system, which is something not trivial in a world of stimuli competing for scarce cognitive attention.
The previous implies that a communicative event is not one in which a maximum amount of information is given, but rather one in which a maximally \emph{relevant} information is given.
It is somewhat a principle operating at a second level: it implies any event triggered will aim at giving the minimal information but with the maximal relevance. At reception, it also implies that determining the goal(s) conveyed by a certain action should assume optimal relevance. But to be able to quantify such relevance, agents need not only to determine causal relations, but also have the skills to determine the underlying goals of a certain behaviour.
\item \textbf{Competence Acquisition}
The ability to associate intentions with human behaviour is developed by sustained social interaction from the very first stages of baby learning \cite[]{halliday1985saw,tomasello2003clu}. Competence acquisition in language and communication seems to be based on imitation combined with the development of mental schemes. Language competence would come progressively by associating the performance of different chunks of language with real-world contexts and consequences. Progressively the learner would use these chunks of language and their combinations. Therefore, the connection between sounds, grammar,and meaning is not an universal process, but a social and situated one \cite{kendrick2020sequence}. According to Kendriks, there appear to be strong universals of interactive language usage,
namely the sequencing of social actions across neighbouring turns, and this allows to identify the universality of the sequence organisation observable in informal human conversational interaction. There is some experimental evidence this describes accurately how kids acquire grammatical knowledge \cite[]{bannard2009modeling}.
This also seems to correspond with the old notion of \emph{internalisation} as a means to acquire communicative competence \cite[]{vygotsky1964tal}. For example, one learns how to ride a bicycle by seeing it and trying it, independently of an understanding of how it works: you internalise what to do to get a certain effect. Similarly, one learns to use a word by determining in what context and with what purpose someone else used it, and then shifts to use it in order to obtain a certain effect.
The term \emph{internalisation} therefore refers to the process by which, given a set of examples on a particular domain, we construct mental categories that abstract particular examples and give us the skills to operate, in a variety of circumstances, on that particular domain. In this picture, kids and adults acquire novel mental categories by internalising from few examples given \cite[]{Vygotsky1986}.
\end{enumerate}
The previous notions give a (quite brief) account of the general assumptions of inferential communication. To summarise this section, we are assuming that in \emph{Language games} the meaning of a \emph{Communicative action} is determined from how it is used. We call the meaning an \emph{interpretation}. The \emph{interpretation} of an action requires associating intentions or goals to it. To determine these an agent needs to categorise the action with a large set of \emph{Natural Categories} interrelated with \emph{Conceptual Metaphors}. The agent also needs to monitor the \emph{Context} in which each \emph{Communicative action} will occur.
\emph{Causal inferences} are core mechanisms of the agents to do so. These inferences take the form of \emph{Conversational implicatures} and they are based on \emph{Relevance principles} shared among the agents.
All these skills are acquired through interactive mechanisms of imitation and cognitive development in a process called \emph{Internalisation}.
\section{A general cognitive module}
\label{embodied-model}
\label{cog-module}
\label{embodied-reasoning}
\subsection{The model}
Sperber convincingly argues that the specifics of how we draw cognitive inferences is best explained through massive modularity , i.e., the existence of a massive amount of interrelated functional modules \cite[]{sperber1996ecn}. An example he draws is the use of masks in primitive rituals, and how the functional module specialised in faces is here used in an altered way. In this sense, the evolution of culture is to be understood as how brain plasticity shapes such functional modules to better fit their use among social agents. Building on Sperber's vision, we assume that in inferential communication an agent is composed of a massive amount of interrelated functional modules.
To formalise inferential communication among humans and artificial agents, we need to characterise such functional modules, and do so in a way that works both for humans and for artificial agents. To do so we introduce a formal model that we call a \emph{general cognitive module} (GCM). Despite it is non-linear and difficult to treat, it has the advantage of being quite general.
Our working hypothesis is that such a structure is general enough to be used to implement the different cognitive mechanisms previously outlined, in the same way that combinations of NAND logical gates can be used to compile any digital algorithm.
Despite we do not provide a proof of such an assumption, we do try to show that under certain assumptions such a model fits with well understood cognitive systems.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{block-diagram
\end{center}
\caption{ The different elements of a general cognitive module. On the higher half, in bold, the fast pathway. On the lower half, in plain text, the slow pathway. }\label{fig:1}
\end{figure}
A GCM is composed of two pathways, the fast one and the slow one. The fast pathway is defined as a transfer function that is an invariant, non-linear filter with a response function $\boldsymbol{h}$. This pathway is defined by the equation $\boldsymbol{o(t) = h(p(t),n(t))}$, and it is composed of:
\begin{enumerate}
\item $\boldsymbol{o(t)}$: A stochastic output arriving through from electrical synapses. In the case of a neuron this is the axonal output firing action potentials. In the case of a piece of brain tissue this can be a statistical distribution of action potentials.
\item $\boldsymbol{p(t)}$: An array of stochastic inhibitory inputs . For example, in the case of a neuron or a piece of brain tissue these will be inputs from inter-neurons arriving through from electrical synapses. This signal would define whether the output is \emph{possible} or not.
\item $\boldsymbol{n(t)}$: An array of stochastic excitatory inputs . For example, in the case of a neuron or a piece of brain tissue these will be inputs from pyramidal neurons arriving through from electrical synapses. This signal would define whether a possible output would \emph{necessarily} occur.
\end{enumerate}
The slow pathway updates the previous response function.
It uses the following signals:
\begin{enumerate}
\item $r(t)$: An array of reward inputs. For example, in the case of a neuron or a piece of brain tissue these will be inputs from chemical synapses.
\item $l(t)$: A learning input which, for simplicity, we assume can activate or disable learning. In a neuron or brain tissue this would be chemical inputs from the neuroglia, enabling or disabling metabolic plasticity.
\item $m(l,r)$ A metabolic response, which determines the update of the transfer function: $h'=m(l(t),r(t))$. When considering $h'$ we may also want to consider the previous transfer function $h$ as a constant in the metabolic response $m(l,r)$.
\end{enumerate}
Overall we can think of general cognitive modules (GCMs) as a scale independent, simplified representation of a computational mechanism for cognitive processing. As such, a correspondence can be established with a biologically realistic neuron, but also with a fragment of brain tissue, to a unicellular animal, or even to more sophisticated natural cognitive systems. It can also be matched to an artificial neuron, or to different machine learning architectures.
\subsection{GCMs in natural cognition}
To illustrate how a GCM fits with the idea of a human brain analysed as a massive amount of GCM modules, we show how it fits with cognitive and computational neuroscience. Neuroscience textbooks give us some hint on how a brain tissue might instantiate a functional module in practice. The nervous tissue constantly balances inhibitory and excitatory signals \cite[]{buzsaki}. Typical examples of excitatory pathways are pyramidal cells in the brain cortex, which transmit signals at long distances. Typical inhibitory signals in the brain cortex are inter-neurons, which are shorter, inhibitory pathways, and more abundant than excitatory ones. In a GCM module, inhibitory pathways correspond to $\boldsymbol{p(t)}$, and excitatory pathways correspond to $\boldsymbol{n(t)}$.
We do not have a clear understanding on how information is encoded in the central nervous system. Despite a common assumption in modelling approaches is to study neural firing rates, it has been shown that temporal patterns also play a major role in coding perceptual information (see, for example, \cite{Gutig2006}). We therefore need to consider different encoding options are possible. However, what we do know that in brain tissue outputs are determined by, first, not having inhibitory inputs and, second, having enough excitatory input to stimulate an output. In a GCM, outputs are defined by $\boldsymbol{o(t)}$, and the relation between input and output determines a, comparatively, fast, non-linear transfer function, denoted as $\boldsymbol{h(p,n)}$.
Opposite to the fast pathways, chemical synapses are complex chemical reactions, with a slower processing time, and determine the metabolic consequences of excitatory and inhibitory input.
The dopamine system, along with several other chemical signals, can inhibit or activate the functional plasticity of the tissue (in a GCM, this corresponds to $l(t)$). Other signals will make the response of the cell change to respond better or worse to fast stimuli (in a GCM, this corresponds to $r(t)$).Chemical synapses can also trigger proteine transduction from gene expression, which ultimately change the shape of cell and the non-linear transfer function between inputs and outputs. In a GCM, this corresponds to $m(l,r)$.
The previous picture is quite general, and it is consistent with brain areas with very different functions, different connectivity, and different stimuli-response electrical patterns (for example, between the cerebellum, the prefrontal cortex and the basal ganglia).
Another role of chemical communication, beyond synapses, is to shape network connectivity, mostly through chemical gradients spread across brain tissue.
Arguably, the mechanisms that are behind the establishment of network connectivity are what determine the brain's extreme parallelism, high level of recursivity patterns, as well as specific recursive circuits such as, for example, thalamo cortical loops for multimodal integration, or cortico-basal and ganglia-thalamo-cortical loops related with the dopamine reward system.
Moreover, the capacity of the brain to change function and connectivity is not only related with neural activity. It also involves the support system. For example, astrocyte signaling regulates plasticity \cite{ota2013role,haydon2015astrocytes}. Other phenomena, such as brain cell migration in early neurodevelopment \cite[]{tierney2009brain}, also has an important role in brain plasticity. The mechanisms supporting network connectivity are not defined in the GCM. We will later introduce some principles to guide connectivity between GCM modules. However, before doing this, we turn towards artificial systems, to show how GCM modules can formalise, also, computational agents devoid of any biological realism.
\subsection{GCMs in artificial systems}
\label{specific-cases}
To illustrate the generality of this formalism we now discuss some specific, well known, cases, from the perspective of such a model.
\subsubsection{A Classic Communication System}
If we target at a classic communication system \cite[]{shannon48}, we can design a receptor as a single instance of the General Cognitive Module previously outlined. This can be done quite simply by doing the following additional assumptions:
\begin{enumerate}
\item The slow pathway is completely static ($h'=h$), Reward inputs are null ($r(t)=0$) and Learning inputs are also null ($l(t)=0$).
\item The excitatory input array ($\boldsymbol{n(t)}$) becomes, simply the baseline signal, i.e., the signal transmitted after the radio-frequency filter and the demodulation step.
\item The inhibitory inputs ($\boldsymbol{p(t)}$) are a digital switch, encoding the connectivity in the graph of possible symbol transitions
\item The transfer function ($\boldsymbol{h(t)}$) is a simple array of linear and invariant matched filters responding to the excitatory input, turned on or off by the inhibitory signal
\item The fast pathway has a stochastic output ($\boldsymbol{o(t)}$) defined by the modulation technique (defined in the matched filters) the coding (defined by the possibility signal in the inhibitory inputs), and the received input (defined in the necessity, excitatory input).
\end{enumerate}
\subsubsection{A Video game Character}
A typical way to create an interactive character in video game industry is to combine a behaviour tree, which makes the decisions, together with a hierarchical state machine which takes care of blending the relevant animations.
A behaviour tree is defined as a tree-like graph, where different control nodes will select which branches are active. Some of these control nodes will evaluate binary conditions, and some will simply select leaf nodes sequentially, or randomly.
Leaf nodes, in turn, will execute a behaviour, and return a binary outcome.
Executing a behaviour means a virtual character being rotated or displaced in a 3D environment, often also triggering a kinematic animation, often through a hierarchical state machine. The transition graph of the state machine will define the possible transitions between animations. Additional parameters will define when these transitions can be triggered.
Skeletal animation techniques \cite[]{magnenat1988joint} will be used to render the body mesh. \footnote{ Different techniques for facial animation are possible, but for the purpose of this analysis we will assume there are no facial expressions involved.}
Such an architecture can also be considered a particular case of our GCM. Below we outline a possible implementation:
\begin{enumerate}
\item The input from the environment ($\boldsymbol{n(t)}$) will simply be a set of binary conditions, perceived from the virtual environment inhabited by the character.
\item The inhibitory input ($\boldsymbol{p(t)}$) will be defined by the animator. It will define the graph of possible transitions, i.e., when an animation can transition to another animation.
\item The output ($\boldsymbol{o(t)}$) will have three parts:
A) a skeleton, i.e., a graph specifying the connectivity between bones, and the size of each.
B) the 3D position of the root node
C) a set of 3D rotations, corresponding to each node in the bone topology, generally expressed as quaternions.
\item the transfer function ($\boldsymbol{h(t)}$) can be thought as a set of impulse responses of linear invariant filters, one for each kinematic animation stored. The selection of each impulse response will depend both on the possibilities enabled by the transition graph.
\end{enumerate}
Of course, it would also be possible to consider the decision and animation parts as separate GCM modules, where the output of the first is connected to the input of the second. However, for this purpose we need some kind of logic that helps us connect the different modules, which we will address in next section. Before this, we illustrate with an example how the slow path of a GCM can be used.
\begin{figure}
\includegraphics[width=8cm]{RL-diagram}
\caption{Our diagram for a GCM integrated in the general reinforcement learning framework }
\label{fig:2}
\end{figure}
\subsubsection{Character animation based on reinforcement learning}
Some of the most impressive demonstrations for computational cognitive systems are
based on Deep Reinforcement Learning. Recent systems have demonstrated super-human features in classical games such as go or chess. However, to further illustrate specific cases of the GCM introduced we choose to analyse an example closer to our main interest: the use of reinforcement learning for physics-based interactive character animation \cite[]{peng2018deepmimic}\footnote{A general introduction and demonstration videos can also be found in https://xbpeng.github.io/projects/DeepMimic/index.html}. Such a system uses the same structure than a general reinforcement learning problem \cite[]{sutton2018reinforcement}. In figure \ref{fig:2} we show how our GCM fits in a RL framework. Compared to the previous section, the main changes are:
\begin{enumerate}
\item The slow pathway is not static ($h' \neq h$). Reward inputs ($r(t)=0$) are defined from the connectivity of the agent with the environment. Learning inputs ($l(t)=0$) will be an inverted step function: one for the training phase, null for the animation synthesis step.
\item In the fast pathway, the transitions between animations are not determined by an external state machine. Therefore the inhibitory input ($\boldsymbol{p(t)}$) is not defined. The transfer function ($\boldsymbol{h(t)}$) is not anymore a set of linear invariant filters, one per each animation. Rather, it is a non-linear function with internal recursivity: triggering an animation means increasing the importance of an attractor in a dynamic system. The stable state of the attractor is the synthesis of an animation. The excitatory inputs ($\boldsymbol{n(t)}$) are not only the binary variables that previously triggered a given animation, but also dynamic inputs such as the last pose of the character, as well as arbitrary parameters, typically target positions and orientations of skeletal nodes such as the root, a feet or a hand.
\end{enumerate}
The previous examples show how we can consider a GCM as a generalisation of diverse specific, well known, cognitive systems. However, as already stated, we do not know anything about the connectivity and dynamic exchanges between different GCMs and agents (made of GCMs). We address this in next section.
\section{A Formalism for Inferential communication}
In this section we introduce formal definitions to relate the events and the pragmatic inferences characteristic of inferential communication. Theorem proofs are in the appendixes.
\subsection{Events and Predicates}
First, we formalise events and predicates associated to them.
\begin{definition} A specific \textbf{event} is denoted with letters such as
`$a, b, c, ...$'. A generic event is denoted with the letters `$x, y, z$'.
We may use sub-indices if further variables are needed, such as $x_1$. We assume an event $e$ has a beginning in time $t_e$, and a duration $d_e$.
\end{definition}
Notice that these events occur in a specific time period, and as such they are not the abstract, instantaneous events, found typically in probability theory.
\begin{definition} A \textbf{property} is denoted with letters such as `$A, B, C, ..., P, Q, ...$'. We may also sub-indices if further variables are needed, such as $P_1$.
\end{definition}
We also introduce an epistemic definition of logical quantifiers relative to the notion of \emph{a cognitive system}.
\begin{definition}
A \textbf{cognitive system} is defined as the combination of one or more GCM instances connected between them through at least some of their inputs and outputs.
\end{definition}
\begin{definition}
The symbol `$\exists^{st}$'
denotes the existence of an event external to a cognitive system. We read it as \textbf{standard existence}. For an event `$e$' we write `$\exists^{st}e$'. For a property `$P(e)$' we write `$\exists^{st}P(e)$'. An event `$i$' whose existence is represented internally by the cognitive system, but not necessarily associated with an external event, it is denoted as `$\exists i$'. Conversely, a property associated to such an event is denoted `$\exists P(i)$'.
\end{definition}
To denote an event with standard existence we call it a \textbf{standard event}. It is important to notice that this does \emph{not} correspond to the classical idea of events as atomic elements, and properties as equivalent to predicates defined as sets of atomic elements. Here \emph{standard} events are associated to the input of a cognitive system, and properties are associated to its outputs. As such, if the outputs of a cognitive system is connected to the input of a second cognitive system, the property will become an event \emph{from the perspective of the second cognitive system}. The transition from standard existence to general existence will be discussed in detail in later sections.
Consistently with the previous definition, we also introduce the standard universal quantifier:
\begin{definition}
The symbol `$\forall^{st}$'
denotes \textbf{all standard events}. The predicate `$\forall$' denotes all events.
\end{definition}
The epistemic nature of the standard existential and universal quantifiers is particularly visible when the cognitive system under consideration is an entire agent. In this case, `$\exists^{st}$' denotes existence of an event \emph{in the world}, as perceived by an agent. The event is the external stimulus, and the property is the perception. Different from this, `$\exists$' denotes existence in more abstract terms, as is done in usual mathematical logic. For example, it could refer to a hypothetical event, independently of whether it is associated to an external stimuli or not.
A class of properties that will become very important are those that can also be considered sets:
\begin{theorem}
The outputs of all non circular cognitive systems are Sets.
\label{predication-theorem}
\end{theorem}
\begin{definition}
\textbf{Predicates} are properties that are also sets (i.e., that are the outcome of non-circular cognitive systems).
\end{definition}
Since predicates are also sets, they are close to the sense of predicate given in predicate logic and first order logic. We can therefore introduce the usual operations of set theory for predicates:
\begin{definition}
The operator `$\in$' denotes a relation between an event and a predicate. It reads as \textbf{belongs to}, or \textbf{in}. For example, $ a \in P $ denotes `$a$ satisfies $P$' (i.e., $a$ is associated with predicate $P$). We also denote this relation as $P(e)$.
We may also consider fuzzy belonging, denoted as $\in_{\alpha}$, with $\alpha \in [0..1]$. We also use $\notin$ to denote the negation of $\in$.
\end{definition}
\begin{definition} To combine predicates we use the the union ($\cup $) and intersection ($\cap$) operators. To denote two predicates are equivalent we use `$\equiv$'. The opposite is denoted as `$\not\equiv$'.
\end{definition}
We now have a complete formalism to combine and compare \emph{predicates}.
\subsection{Causal Implicatures}
To formalise \emph{causal implicatures} we introduce event sequences:
\begin{definition}
A \textbf{chain of events} is denoted with the `$.$' operator. For example `$a.b.c$' denotes a chain of three events. The only constraint introduced by this operator is that the occurrence of the events satisfies a partial order, i.e., that $a$ starts occurring before or at the same time than $b$, etc. Therefore, if $t_a$ denotes when $a$ starts occurring, `$a.b.c$' implies $t_a \leq t_b \leq t_c $.
\end{definition}
There are no constraints regarding when events end (i.e., since events have a duration, $a$ may finish occurring after $b$). From the previous, we can define:
\begin{definition}
The \textbf{context} $x$ of an event $y$ is denoted as $y^*$. It is defined as a chain: $y^* \equiv x_1 . x_2 . ... . x_n$. It does satisfy the condition that $t_{y^*} \leq t_y$.
To denote a context $x$ as perceived by an agent we write $X$, i.e., the predicates associated with the events in $x$.
\end{definition}
\begin{definition}
The \textbf{implied cause} $x$ of an event $y$ is denoted as $y^{**}$.
It does satisfy the condition that $t_{y^{**}} \leq t_y$.
\end{definition}
\begin{definition}
The relation between an event $e$ and its implied cause $e^{**}$ is called a \textbf{causal implicature}, and we denote it as `$e^{**} \Rightarrow e$'.
\end{definition}
\begin{theorem}
For a given event $e$, it is always possible to build a cognitive system that can determine the most likely implied cause using an information-based metric such as surprise minimization.
\end{theorem}
\begin{definition}
An \textbf{inferential agent} is an agent that builds causal implicatures based on information-based metrics such as surprise minimisation.
\end{definition}
In this picture, an inferential agent trying to interpret a communicative exchange will build pseudo-complemented semi-lattices to establish causal implicatures and select the construct that minimises uncertainty.
\begin{theorem}
An information exchange between two inferential agents with goals is not stationary.
\end{theorem}
The previous result shows in what sense this theory generalises classical information theory: contrary to traditional information theory where events are part of a stochastic stationary process, here the probability of occurrence of an event will vary across time.
\subsection{Cognitive processes for competence acquisition}
\label{internalization-principles}
An important aspect of inferential communication is that it presupposes \emph{competence acquisition}. How do we internalise mental categories and build conceptual metaphors to process everyday events? We have introduced an epistemic difference between standard and non-standard events, but not determined how they relate, or how do we build abstract representations (like `furniture') from specific examples (like `this table', or `that chair').
How do we go from \emph{standard} events, i.e., from experiencing concrete events associated to one or several predicates ($\exists^{st} P(a)$), to a predicate \emph{in general} ($\exists P(a)$), which generalises across concrete and hypothetical events? We break down this learning process in three cognitive processes:
\begin{principle}
We denote the process of \textbf{Idealisation (\emph{I})} as `$\overset{I}{\Longrightarrow}$', and define it as:
\begin{equation}
\Big[ \forall^{ st fin} y' \exists x \ \ \Big/ \ \ \forall y \in A(y') \Big] \overset{I}{\Longrightarrow} \Big[ \exists x \forall^{st} y \in A(y) \Big]
\end{equation}
\end{principle}
In the previous $ \forall^{ st fin}$ reads \emph{for all standard finite}. The intuition is that if a cognitive system has a property that applies to all standard finite examples to which it has been exposed, then it is possible to implement a process, which we call \textbf{Idealisation}, that makes this predicate apply to \emph{any} standard examples (thus, the ones seen, but also possible examples not yet seen).
Of course, this occurs spontaneously in humans and other animals. However, the implementation of such a generalisation mechanism in artificial systems is a central problem of machine learning: finding a function that fits best the training set but also generalises to stimuli beyond the training set is the main challenge when training artificial neural networks in classification problems.
In the previous expression $x$ stands for the context in which such an idealisation process occurs. The main idea is that this idealisation will only work in some contexts, not all of them.
This highlights the importance of contextual factors when associating predicates to events. Indeed, contextual factors are of paramount importance in inferential communication. One of their main roles is to help determine subsets of the domain of a predicate.
We can express this idea as follows:
\begin{principle}
We denote the process of \textbf{Selection (\emph{S})} as `$\overset{S}{\Longrightarrow}$' and define it as:
\begin{equation}
\forall^{st} X \exists^{st} Y \forall^{st} z \Big[ z \in Y \overset{S}{\Longrightarrow} z \in X \cap Y(z) \Big]
\end{equation}
\end{principle}
The intuition behind the \textbf{Selection} process is that, given a predicate $Y$ associated to standard events, we can implement a cognitive process, which we call \textbf{Selection} that, given any predicate $X$, is able to determine which subset of events $z$ satisfy predicate $Y$ and also $X$.
If we think of properties $X$ and $Y$ as implemented in separate GCMs (which we can call $GCM_X$ and $GCM_Y$), this principle can be implemented at least in two ways. First, with the output of $GCM_X$ feeding the inhibitory inputs $p$ of $GCM_Y$, and therefore affecting the domain on which $GCM_Y$ responds to stimuli consistent with predicate $Y$. Second, introducing a third module that processes the outputs of both $GCM_X$ and $GCM_Y$ and responds only to the combination of both. However, implementing a separate module for \emph{any} combination of predicates that have some domain overlap can produce a combinatorial explosion. Therefore, to implement this principle it seems more efficient to opt for contextual influence of GCM modules that have domain overlap, despite the dynamics of contextual influence will be more challenging to understand.
The third cognitive principle extends the domain of predicates from \emph{standard} events to events in general:
\begin{principle}
We denote the process of \textbf{Transference (\emph{T}) } as `$\overset{T}{\Longrightarrow}$' and define it as follows:
\begin{equation}
\Big[ \forall^{st} y A(y) \Big] \overset{T}{\Longrightarrow} \Big[ \forall y A(y) \Big]
\end{equation}
\end{principle}
The intuition behind the \textbf{Transference} principle is that if in a cognitive system there is a predicate that applies to all standard members, then we can implement a cognitive process, which we call \textbf{Transference}, by which we can make the predicate apply to all members. This mechanism allows to transfer the properties of standard members to all members. For example, to hypothetical sequences of events.
This undeniably happens in natural reasoning, where we learn to make inferences on hypothetical situations and events, and we even build conceptual metaphors to enrich the predicates associated with factual or hypothetical events. However, current artificial cognitive systems do not learn to generalise a predicate beyond the domain of standard examples on which it is trained. In artificial systems it is easy to simulate standard events using stored media to be presented as stimuli to a machine learning algorithm.
However, it is difficult to see how we may implement in an artificial agent the notion of \emph{conceptual metaphor} to interpret the events involved in everyday communication. To do so, we would first need to clarify what are, in an artificial system, events that exist but do not have a standard existence. We can however assume such agents can be built, and explore their properties:
\begin{definition}
A \textbf{competence acquirer} is an agent where the 3 cognitive processes
\textbf{I} ($\overset{I}{\Longrightarrow}$), \textbf{S} ($\overset{S}{\Longrightarrow}$) and \textbf{T} ($\overset{T}{\Longrightarrow}$) are implemented.
\end{definition}
\subsection{The consistency of inferential agents}
\label{maths}
If we implement artificial agents that have inferential skills and are also capable of competence acquisition, will they develop consistent representations? We believe this is not the case in general, in the sense that \emph{any} inferential agent will develop consistent representations. However, here we show that a particular subset of inferential agents can develop consistent representations.
\begin{definition}
A binary agent is one where all the processes are binary rules. This applies to the transference functions in the GCM modules that form it and also, if implemented, to the three cognitive processes for competence acquisition previously defined.
\end{definition}
\begin{theorem}
There exists a binary inferential agent that is also a competence acquirer which is a consistent embodiment of mathematical foundations.
\end{theorem}
More qualitatively, the way this theory is built also seems closer to an embodied account of mathematical foundations \cite{lakoff2000mc}: numbers are defined as Frege numerals, which is not only the first introduced historically, it is also more intuitive: for example, the number 3 can be defined as a property shared by all sets which have 3 elements. This seems particularly appropriate here: if we are to build mathematical foundations consistently with how we build our internal representations, it seems reasonable that the definitions of notions for which we have an intuition should match such intuitions (intuitions which are, presumably, built from embodied representations). Similarly, in this set theory the universal set exists, the complement to a given set always exists, and it is possible to create any set simply through the union of two existing sets. In addition, with the 3 IST principles (see appendix) we can formalise calculus closer to how it was introduced historically, with infinitesimals, circumventing the need for definitions based on limits, and stating closer to the, arguably, intuitive notion of change at arbitrarily small increments. It can also be used to formalise probability theory in a much more simple way than the usual way, based on complicated results from measure theory \cite{nelson1987rep}.
\section{Discussion}
\label{discuss}
\subsection{Inferential communication}
The publication of \emph{A mathematical theory of communication} \cite[]{shannon48} spread the view of communication as the transmission of a message encoded by an emitter and decoded by a receptor.
Hardly ever a theory developed for engineering purposes has had such influence in fundamental research fields such as physics, linguistics, biology and cognitive sciences. However, people studying natural communication started to raise some objections to such view of communication and they eventually proposed the inferential communication model \cite[]{sperber1986relevance} to give some account of aspects of everyday conversations not described by classical information theory. The benefit of such a theory
is that it allowed to explain why communicators convey much more information with their utterances than what is contained in their literal expressions. We have introduced a formal model that addresses this challenge and hints at some possible ways to engineer practical solutions to such interaction scenarios.
\subsection{Model summary}
In inferential communication an agent is an entity which processes information. This information is received as a stream of events. \emph{Events} are mainly actions of other communicative agents, for which we assume such events are intentional actions, i.e., events generated by an agent to which other agents associate intentions.
However, part of the events can also be natural events that may be perceived as relevant for communication purposes. In both cases, when they receive information in the form of events, inferential agents associate an interpretative meaning to these events. This cognitive process, that we call \emph{interpretation}, consists in associating the events with communicative intentions, within a context that includes a given situation, and a cultural background assumed as shared among the different agents. It can, sometimes, provoke agents to perform additional actions, either as an immediate reaction to them, or as a delayed answer.
We have argued this picture can be formalised in MaTIC: within the space of \emph{Language games}, communication is a collaborative game where each agent has a set of goals, some of which are shared. Grice's cooperative principle can be formalised as a goal shared among the participants in the game. The intentions of individual agents can be formalised as a set of goals. The events defining the game are communicative actions, and are intended to be perceived, and associated with predicates.
In MaTIC the idea of a standard event is associated with an epistemic definition of event, but also with a rigorous logical treatment. The particular predicates associated with an event are selected according to the general principles of relevance theory, with information theoretic measures. These associations are learnt through a combination of adaptive GCM modules, together with connectivity changes guided by the I, S and T principles. The existence of a property is associated with the output of a cognitive module. Predicates are properties organised in non-circular connectivity, something that allows building richer representations.
In this picture, cognitive agents infer causal implicatures for the occurrence of an event $e$ by, first, finding the conditions in which $e$ \emph{can} occur and, second, finding a motivation for $e$ to occur. In other terms: the context determines the \emph{possibility} of $e$. Correspondingly, the event or intention that determines the \emph{necessity} that $e$ occurs, constitutes the cause that we associate with it. The way a cognitive agent will do so will be based on heuristic reasoning, under general principles of uncertainty minimization.
The main difference between cognitive agents involved in inferential communication with the other existing systems introduced is that in those examples of GCM instances
the inhibitory input $p(t)$ was either null, or either it simply restricted the response of a system to a subset of possible outcomes. In addition, the inhibitory input function was determined \emph{beforehand}, i.e., before any excitatory input occurred. This does not seem appropriate to address the dynamic interplay that occurs in inferential communication. In the conversation example previously introduced, the occurrence of an event such as `\emph{B says: There is a garage around the corner}' was, arguably, made possible by a larger context, such as B knowing this fact, as well as the previous events. In addition, assuming none of the participants had hidden intentions, an event such as \emph{A says: I am out of petrol} would be generally interpreted as the cause of B's statement.
More generally, as new exchanges occur in a conversation, the context of the conversation will evolve, and therefore the set of possible actions will too. \emph{Contexts in natural communication are dynamic}, therefore we need to consider GCM instances with response functions which combine the $p$ and $n$ inputs in a more intricate way. Overall, MaTIC does not define completely how causal implicatures occur, but it does define a set of \emph{requirements} that must be satisfied by the agents exchanging intentional actions. Under these requirements, and contrary to classical information theory, the probability of occurrence of an event changes across time.
\subsection{A neuroscience perspective}
Is MaTIC based on reasonable assumptions, from which we can build further details? Contemporary cognitive neuroscience suggest there may be empirical support for a computational theory of inferential communication. The mirror system is assumed to enable imitation, and empathy, and to be based on sensor-motor neural mechanisms \cite{kohler2002hearing}. Such bodily disposition would explain the skill of learning other's intentions \cite{barron2020neuronal}. In fact, embodied and enactivist approaches to cognition are connecting with reasonable plausibility the spheres of bodies, societies, and cultural manifestations, with (inferential) communication as a multi-modal skill \cite{gallagher2018active}.
The modern study of intentionality
encompasses both mental states towards things as well as the ways by which human agents understand others intentional dispositions \cite[]{brentano2014psychology}. This atonement process is located in our neural system, working from very basic sensor-motor level to symbolic processing \cite[]{rizzolatti2007mirror}, \cite[]{bretherton1991intentional}. For such reasons, there are mechanisms by which such intentional actions can be observed and therefore the related actions can be inferred to. In this sense, the study of intentionality allow us to predict and understand the next actions of human agents, allowing us to design a suitable answer. Indeed, in the model we propose agents will also, sometimes, perform additional actions as a reaction to the events and the predicates associated to them. According to this theory, this should be done in line with the first principle of relevance theory, i.e., the fact that \emph{human cognition tends to be geared towards the maximisation of the cognitive effect}.
There are other accounts on how abstract representations can be built through the combination of bottom up and top down processes.
Notably, predictive coding and active inference \cite[]{constant2020representation,friston2007free}, suggest information-theoretic measures are relevant to account for how the brain learns to interact with its environment.
Despite MaTIC is compatible with predictive coding and the idea that uncertainty minimisation can go quite far in determining cognitive processes, there are two aspects that make MaTIC strikingly different: first, in how it reconciles the idea of embodied reasoning with a formalism for mathematical foundations, and second in how it seems compatible with cognitive accounts of everyday exchanges, as conceptualised in cognitive pragmatics.
\subsection{Cognitive Pragmatics}
The process of \emph{interpretation} has also been interpreted in a Bayesian framework
\cite{vallverdu2015bayesians}.
We believe the model introduced here may complement Bayesian models of thinking processes by taking into account the other embodied variables present into communicative processes. One of the key assumptions of Bayesian inference is that a belief and its opposite (in our notation, $A$ and $\neg A$) exclude each other: they form a complementary set. However, this is not a general assumption, for example, in deep learning inference systems, where Bayesian models are just a subset of the solutions that, in practice, work.
In this context, local restrictions may explain why the agents opt for selecting some informational sets and processing models instead than others: cultural and training/academic variables are fundamental thresholds of the way of dealing with information
(see the notion of blended cognition \cite{vallverdu2019blended}).
Future work should analyse whether this model does different predictions than Bayesian modelling, and which gives better support to the analysis of the mental mechanisms that are used to process intentional communication.
In this context, we also point to Rational Speech Act Theory \cite[]{yuan2018understanding}, which provides a fully formalised framework for pragmatic inference under a Bayesian framework. For example, \cite{goodman2013knowledge} shows that pragmatic agents do inferences that seem more adjusted to reality than their literal counterparts. Beyond its insistence on maximising utility, Rational Speech Act Theory seems consistent with a Bayesian picture, and is reasonably aligned with general principles related with reducing uncertainty in information exchanges, close to the idea of minimising surprisal, as discussed in the previous section. However, beyond the question of whether Bayesian modelling covers the whole picture, there are two aspects of MaTIC that do not seem addressed in the Rational Speech Act Theory: first, how such inferences can be supported by an embodied mechanism (for which we have proposed the GCM, and shown how biological and artificial systems can be seen as particular instances of such a modular embodiment) and, second, how the internal representations supporting pragmatic inference can be developed and maintained through time.
\subsection{Embodied reasoning and representation consistency}
A particularly original aspect of MaTIC is its link with formal set theory. In this context it is striking that actual infinity is not assumed. Instead, we have assumed the existence of external events, and predication, and argued we can build most of the usual construction for mathematical engineering (sets, numbers, vectors, functions, metrics, derivatives) from it.
If our model of competence acquisition is correct, the mechanisms that we posited for acquiring communicative competence should also apply to acquiring mathematical competence. However, this is at odds with how the foundations of mathematics are generally assumed by mathematicians. For example, in \cite{cohen1966set} we find an exposition of the default set theory, also called Zermelo-Frankel with Choice (ZFC). Such a theory assumes there only exist sets, and that there exist an actual infinite amount of them.
This is absolutely at odds with the notion of an agent developing internal representations through interactive sensor-motor loops in a finite world. In this context, it is particularly striking that our theory does not assume the actual existence of an infinite number of sets to build solid mathematical foundations. Rather, it is enough to exist the existence of internal and external events, and GCM modules processing them.
Future work should explore if MaTIC can also can be used to build agents that develop consistent internal representations from data, beyond the domain of mathematics. For this purpose, we believe again virtual reality provides an ideal experimental environment, one where artificial agents can easily build internal representations from their virtual environment. This would strongly simplify the perception challenges generally involved in robotics, and allow exploring directly whether agents in an environment can build internal representations of it, and whether this derives in exchanging information under pragmatic principles.
\subsection{A falsable theory}
Any good scientific model must do accurate predictions that can be validated experimentally. The main implication of MaTIC might be the idea that by building artificial agents that satisfy the three cognitive principles introduced, then using data and machine learning these agents should be able to achieve inferential communication as competently as native speakers, and generate responses appropriate to their behaviour and also to their own communication intentions. We believe the best way to show this is through the iterative development of VR scenarios involving social interaction between autonomous agents and humans, based on these premises.
Another aspect of a scientific model is that it should be falsable. There are several ways to show this theory is wrong. For example, by:
\begin{enumerate}
\item showing that the three cognitive principles introduced are not relevant for the purpose at hand.
\item showing that massive modularity does not occur, after all, in animals nor humans, or that it is not needed to build artificial systems that are capable of inferential communication.
\item showing that predication can be built with circular systems, that there is no difference between circular and non-circular systems
\item showing that contextual processing does not occur through inhibitory inputs of GCM modules.
\end{enumerate}
Further computational systems could also be analysed within MaTIC. Here we are not short of options: Sperber has often stated that relevance theory is actually a general model of cognition. Since MaTIC is based on formalising some aspects of this theory, and on how to embody it, there is ground for challenging this theory with virtually any modelling approach used in systems neuroscience or in artificial intelligence focused on machine learning.
A separate stream of work analysing interpersonal interaction using methods derived from the modelling of dynamic physical systems that should be explored\cite[]{
zhai2016design, alderisio2017interaction, lombardi2019deep}.
Future work should also focus on validating MaTIC in behavioural studies of people interacting with agents embodied in interactive virtual characters.
Another way to explore whether MaTIC does valid predictions is to to review insights from behavioural economics and cognitive psychology on decision making. The discounting effect and similar paradoxes have been well characterised in behavioural economics, but are difficult to explain with classical probability theory. We may study whether an agent learning on these premisses will also adopt heuristic decisions that reflect these cognitive biases.
Another direction to explore is to clarify whether MaTIC gives a better account for inferences and predicates as they occur \emph{in the world}, i.e., consistently with how cognitive psychology characterises natural categorisation and how cognitive pragmatics argues we reason and make decisions. From a developmental perspective: do the interpretation of the I, S and T principles introduced capture part or all of the picture in Vyggotsky's concept of \emph{Internalisation}? These topics are being addressed by ongoing neuroscientific work \cite[]{teufel2020forms}, and it would be interesting to explore whether our theory is compatible or helps explain how such developmental processes work. Empirically, we could also look at the extent to which artificial agents developing their skills under such principles incur in phenomena such as "hyper-regularisation", i.e., the trend kids to learn general rules for syntax, and then have to re-learn exceptions to the rule that they had initially learnt.
\subsection{Ethics}
Last but not least, a different question to consider is what ethical implications should be addressed if we were to implement such systems in practical, functional, consumer products. The arrival of immersive virtual and augmented reality to the consumer market, is likely to have a considerable social impact particularly if combined with autonomous agents capable of everyday pragmatical inferences as humans do everyday spontaneously and often unconsciously. Depending on how it is implemented it can provide opportunities or challenges to the realisation of human rights in society. Let's introduce a simple example: the always winning Janken robot from Ishikawa Oku Lab (Japan), which has absolute winning rates playing at rock-paper-scissors against humans \cite[]{katsuki2015high}. The robot is able to capture and interpret bodily movements of the human player and then react more quickly always beating the human, which has not been able to understand the unfairness of the situation, because the robot moves itself following a similar temporal pattern than that of humans. This mechanism applied to other contexts would make possible to open a new way of cheating humans in benefit of machine owners.
Is it reasonable to allow these super-human affordances to be used without the users' knowledge? \cite[]{vallverdu2016emotional}.
What considerations should be taken into account when our governments approach these topics?
\bibliographystyle{apacite}
\setlength{\bibleftmargin}{.125in}
\setlength{\bibindent}{-\bibleftmargin}
|
\chapter{Introduction}
\chapter{Introduction to Molecular Dynamics}
\renewcommand{\chaptername}{Chapter}
The computational methodology carrying the name of Molecular Dynamics (MD) is entirely built on the laws of classical Statistical Mechanics, that allow to describe with a surprisingly good approximation a large class of microscopic system, without resorting to the tools of Quantum Mechanics. One of the founding concepts of Statistical Mechanics is the notion of \emph{statistical ensemble}, which can be defined as a set of infinite copies of a system with different \emph{microscopic} states but sharing the same \emph{macroscopic} state. The macroscopic state is defined by fixing a few thermodynamic quantities, which univocally identify the corresponding statistical ensemble (e.g. number of particles $N$, volume $V$ and internal energy $E$ in the microcanonical ensemble), while the microscopic state includes the particles' positions $\{\mathbf{q}_i\}$ and momenta $\{\mathbf{p}_i\}$ ($i=1,...,N$), which compose the $6N$-dimensional phase space $\Omega$. Importantly, the algorithms employed in a MD simulation depend on the statistical ensemble the system belongs to. Among all the possible statistical ensembles, the most common ones in MD simulations are the microcanonical, the canonical and the isothermal-isobaric ensembles. In the following, these three ensembles are reviewed in their fundamental concepts, together with the most popular MD algorithms that have been developed over the years to reproduce them in computer simulations.
\section{Microcanonical ensemble (NVE)}\label{hamilton}
When a system is isolated, its dynamical evolution is completely determined by the Hamilton's equations of motion,
\begin{subequations}\label[pluralequation]{hamiltons_eqs}
\begin{align}
\label{hamilton_eqs1}
\text{d}\mathbf{q}_i &= \frac{\mathbf{p}_i}{m_i}\text{d}t\,, \\
\label{hamilton_eqs2}
\text{d}\mathbf{p}_i &= -\frac{\partial U}{\partial\mathbf{q}_i}\text{d}t\,,
\end{align}
\end{subequations}
where $U=U\big(\{\mathbf{q}_i\}\big)$ is the potential energy of the system, $m_i$ is the mass of the particle $i$ and $t$ is the time. It is possible to prove that \cref{hamiltons_eqs} have a conserved quantity $\mathcal{H}$, which is called Hamiltonian or energy and whose value only depends on the initial conditions $\{\mathbf{q}_i(t_0),\mathbf{p}_i(t_0)\}$:
\begin{equation}
\mathcal{H}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) \equiv \sum_i \frac{|\mathbf{p}_i|^2}{2 m_i} + U\big(\{\mathbf{q}_i\}\big)
= \mathcal{H}\big(\{\mathbf{q}_i(t_0),\mathbf{p}_i(t_0)\}\big) \equiv E\,.
\end{equation}
The term $K\big(\{\mathbf{p}_i\}\big) = \sum_i |\mathbf{p}_i|^2/(2 m_i)$ is called kinetic energy of the system.
As an additional property, \cref{hamiltons_eqs} are time-reversible, meaning that if the trajectory $\{\mathbf{q}_i(t),\mathbf{p}_i(t)\}$ is a solution of Hamilton's equations, then also the \emph{time-reversed trajectory}
\begin{equation}
\{\mathbf{q^*}_i(t),\mathbf{p^*}_i(t)\} \equiv \{\mathbf{q}_i(-t),\mathbf{p}_i(-t)\}
\end{equation}
is a solution.
The bridge between the deterministic Hamiltonian mechanics and the statistical approach embedded in the microcanonical ensemble can be constructed with three fundamental steps, i.e. the ergodic hypothesis, the Liouville's theorem and the principle of equal a priori probabilities.
\begin{ergodic_hyp}
Given a region d$\Omega$ of the phase space, over large times the amount of time d$t$ that the system spends in d$\Omega$ is proportional to the volume of this region:
\begin{equation}
\label{erg_hyp}
\lim_{\tau\rightarrow\infty}\frac{\text{d}t}{\tau} = \rho\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)\, \text{d}\Omega\,.
\end{equation}
\end{ergodic_hyp}
\noindent The quantity $\rho$ that appears in \cref{erg_hyp} is called \emph{phase space density}, and if properly normalized it can be interpreted as the probability (density) of finding the system in a given point of phase space. This allows to replace time averages of physical observables with ensemble averages over the phase space: if $a\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)$ is the instantaneous microscopical estimator of a macroscopic quantity $A$, i.e.
\begin{equation}
A = \lim_{\tau\rightarrow\infty}\frac{1}{\tau}\int_t^{t+\tau}\text{d}t'\, a\big(\{\mathbf{q}_i(t'),\mathbf{p}_i(t')\}\big)\, ,
\end{equation}
by applying the ergodic hypothesis we can replace the integral over time with an integral over the phase space, weighted by the density $\rho$:
\begin{align}
A &= \lim_{\tau\rightarrow\infty}\int_t^{t+\tau}\frac{\text{d}t'}{\tau}\, a\big(\{\mathbf{q}_i(t'),\mathbf{p}_i(t')\}\big) \\
&= \int \text{d}\Omega\, \rho\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) \, a\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) \,.
\label{ens_average}
\end{align}
\begin{liouville_th}
The phase space density $\rho$ is conserved:
\begin{equation}\label{liouv_th}
\frac{\text{d}\rho}{\text{d}t}=0\,.
\end{equation}
\end{liouville_th}
\noindent The so-called Liouville's equation includes a total time derivative, which describes how $\rho$ changes in time as a function of how positions and momenta change in time. Considering also an explicit time dependence, i.e. $\rho = \rho\big(\{\mathbf{q}_i,\mathbf{p}_i\},t\big)$, \cref{liouv_th} can be rewritten by applying the chain rule for derivatives, namely
\begin{equation}\label{chain_rule}
\frac{\text{d}\rho}{\text{d}t} = \frac{\partial\rho}{\partial t} + \sum_{i,\alpha} \Big(\frac{\partial\rho}{\partial q_{i}^{\alpha}}\,\dot{q}_{i}^{\alpha} + \frac{\partial\rho}{\partial p_{i}^{\alpha}}\,\dot{p}_{i}^{\alpha} \Big)\,.
\end{equation}
Defining the vector of all phase space coordinates as $\mathbf{x} = \left(q_1^x,q_1^y,q_1^z,...\,,p_N^x,p_N^y,p_N^z\right)$, this condition turns out to be equivalent to
\begin{equation}\label{incompressibility}
\nabla_{\mathbf{x}}\cdot\dot{\mathbf{x}} = 0\,,
\end{equation}
where $\nabla_{\mathbf{x}} = \left(\partial/\partial q_1^x,...\,,\partial/\partial p_N^z\right)$. For this reason Liouville's theorem is also said to describe the property of \emph{phase space incompressibility}, since \cref{incompressibility} implies the absence of sources or sinks for a fluid with velocity flow field $\dot{\mathbf{x}}$. Providing a geometrical interpretation, this condition implies that the copies of the system initialized within the phase space element $\text{d}\mathbf{x}_0$ will evolve in a phase space element $\text{d}\mathbf{x}_t$ with the same volume, at any future time $t$ (see \cref{fig:liouville}).
\begin{figure}[h!]
\centering
\includegraphics[width = 0.75\textwidth]{core/first_chapter/Liouville.jpg}
\caption{Graphical representation of the phase space \emph{volume conservation} prescribed by Liouville's theorem for a one-dimensional system. Source: \cite{tuckerman}.}
\label{fig:liouville}
\end{figure}
\noindent Using \cref{chain_rule} together with Hamilton's equations, \cref{liouv_th} can be rewritten as:
\begin{equation}
0 = \frac{\partial\rho}{\partial t} + \sum_{i,\alpha} \Big(\frac{\partial\rho}{\partial q_{i}^{\alpha}}\,\frac{\partial\mathcal{H}}{\partial p_{i}^{\alpha}} - \frac{\partial\rho}{\partial p_{i}^{\alpha}}\,\frac{\partial\mathcal{H}}{\partial q_{i}^{\alpha}} \Big)
= \frac{\partial\rho}{\partial t} + \Big\{\rho,\,\mathcal{H}\Big\}\,.
\label{liouv}
\end{equation}
Here $\alpha$ identifies the cartesian component ($\alpha=x,y,z$), the dot notation is used for the total time derivative and $\{\cdot\,,\cdot\}$ define the Poisson brackets. Therefore in a system at equilibrium, for which
\begin{equation}
\frac{\partial\rho}{\partial t} = 0\,,
\end{equation}
\cref{liouv} brings to $\{\rho,\,\mathcal{H}\} = 0$, which implies that the density $\rho$ can only depend on the phase space coordinates $\{\mathbf{q}_i,\mathbf{p}_i\}$ through the Hamiltonian $\mathcal{H}$:
\begin{equation}
\rho = \rho \Big(\mathcal{H}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)\Big)\,.
\end{equation}
\noindent Among all the possible functional forms satisfying \cref{rho_H}, the one that defines the density $\rho$ in the microcanonical ensemble is fixed by the following principle.
\begin{equiprobab_pr}
The probability associated to each microstate, i.e. to each point of phase space, is the same for all the microstates compatible with the energy $E$ of the system, fixed by the initial conditions.
\end{equiprobab_pr}
\noindent In other words, the probability distribution of the microcanonical ensemble can be written as a $\delta$-function:
\begin{equation}
\mathcal{P}_{NVE}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) = \frac{1}{\Gamma}\, \delta\Big(\mathcal{H}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) - E\Big)\,,
\label{rho_H}
\end{equation}
where $\Gamma = \Gamma(N,V,E)$ is the normalization factor of the distribution, fixed by the constraint
\begin{equation}
\int_{\Omega} \Big(\prod_{i=1}^N d\mathbf{q}_id\mathbf{p}_i\Big)\, \mathcal{P}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) = 1\,.
\end{equation}
\subsection{Velocity Verlet and leapfrog algorithms}\label{velocity_verlet}
Since Liouville's theorem is derived by means of Hamilton's equations, the dynamical evolution of the system can be obtained by solving \cref{liouv_th}, that can be alternatively written as
\begin{equation}\label{liouvilleq}
\frac{\partial\rho}{\partial t} = -i\mathcal{\hat{L}}\,\rho\,,
\end{equation}
where $\mathcal{\hat{L}}$, called \emph{Liouville operator}, is typically split in the \emph{drift operator} $\mathcal{\hat{L}}_q$ and the \emph{kick operator} $\mathcal{\hat{L}}_p$:
\begin{equation}
i\mathcal{\hat{L}}\equiv i\mathcal{\hat{L}}_q + i\mathcal{\hat{L}}_p \equiv\sum_{i,\alpha} \left(\frac{\partial \mathcal{H}}{\partial p_{i}^{\alpha}}\,\frac{\partial}{\partial q_{i}^{\alpha}}-\frac{\partial \mathcal{H}}{\partial q_{i}^{\alpha}}\,\frac{\partial }{\partial p_{i}^{\alpha}}\right)\,.
\end{equation}
Isolating the action of $\mathcal{\hat{L}}_q$ and $\mathcal{\hat{L}}_p$ separately, it is easy to show that the two following relations hold exactly:
\begin{subequations}
\begin{align}
\label{Lq}
e^{-i \Delta t\mathcal{\hat{L}}_q}\rho(\mathbf{q}_i,\mathbf{p_i},t) &= \rho\left(\mathbf{q}_i -\frac{\mathbf{p}_i}{m_i}\Delta t,\mathbf{p_i},t\right)\,, \\
\label{Lp}
e^{-i \Delta t\mathcal{\hat{L}}_p}\rho(\mathbf{q}_i,\mathbf{p_i},t) &= \rho\left(\mathbf{q}_i,\mathbf{p_i}-\mathbf{F}_i\Delta t,t\right)\,,
\end{align}
\end{subequations}
where $\mathbf{F}_i=-\frac{\partial U}{\partial \mathbf{q}_i}$ is the force acting on atom $i$ and $\Delta t$ is the integration time step.
However, the formal solution of \cref{liouvilleq},
\begin{equation}
\label{exact_sol}
\rho(\mathbf{q}_i,\mathbf{p_i},t+\Delta t) = e^{-i \Delta t\mathcal{\hat{L}}}\rho(\mathbf{q}_i,\mathbf{p_i},t)\,,
\end{equation}
includes the exponential of a sum of two non-commuting operators and cannot be rewritten separating the action of the two operators. A popular approximation of \cref{exact_sol}, which is correct up to second-order in $\Delta t$, employs the so-called \emph{Trotter splitting}:
\begin{equation}\label{vv_trotter}
\exp\left(-i \Delta t\mathcal{\hat{L}}\right) = \exp\left(-i \frac{\Delta t}{2}\mathcal{\hat{L}}_p\right) \exp\left(-i \Delta t\mathcal{\hat{L}}_q\right) \exp\left(-i \frac{\Delta t}{2}\mathcal{\hat{L}}_p\right) + \mathcal{O}(\Delta t^3)
\end{equation}
In words, \cref{vv_trotter} gives a simple recipe to evolve the system according to Hamilton's equations, with an accuracy depending on the timestep $\Delta t$. A single iteration of the algorithm for atom $i$ is reported below.\\
\begin{algorithm}[H]\label{velocity_verlet_alg}
\SetAlgoLined
$\mathbf{p}_i(t+\Delta t/2) = \mathbf{p}_i(t) + \mathbf{F}_i(t) \Delta t/2$\;
$\mathbf{q}_i(t+\Delta t) = \mathbf{q}_i + \frac{\mathbf{p}_i}{m_i}(t+\Delta t/2) \Delta t$\;
recompute forces: $\mathbf{F}_i(t+\Delta t) \longleftarrow\mathbf{F}_i(t)$\;
$\mathbf{p}_i(t+\Delta t) = \mathbf{p}_i(t+\Delta t/2) + \mathbf{F}_i(t+\Delta t) \Delta t/2$\;
\caption{Velocity Verlet}
\end{algorithm}
\vspace{0.4cm}
\noindent This integration scheme is named \emph{velocity Verlet}, as opposed to the algorithm where drift and kick operators are exchanged in the Trotter splitting (called \emph{position Verlet}), and it's easy to show that it satisfies both the time-reversibility of Hamilton's equations and the property of density conservation in phase space, but clearly violates the conservation of energy.
Since the last momentum update and the first one in the next iteration employ the same forces, these two steps can be combined in order to improve the efficiency of the algorithm. This approach defines the \emph{leapfrog integrator}, where positions and momenta are propagated with a time lag of half timestep:\\
\begin{algorithm}[H]
\SetAlgoLined
$\mathbf{p}_i(t+\Delta t/2) = \mathbf{p}_i(t-\Delta t/2) + \mathbf{F}_i(t) \Delta t$\;
$\mathbf{q}_i(t+\Delta t) = \mathbf{q}_i + \frac{\mathbf{p}_i}{m_i}(t+\Delta t/2) \Delta t$\;
recompute forces: $\mathbf{F}_i(t+\Delta t) \longleftarrow\mathbf{F}_i(t)$\;
\caption{Leapfrog integrator}
\end{algorithm}
\section{Canonical ensemble (NVT)}
A physical system belongs to the canonical ensemble if it is closed but can exchange energy with the external environment in the form of heat, i.e. if its number of particles $N$, its volume $V$ and its temperature $T$ are fixed. In this context, the system is typically said to be coupled with an external bath at temperature $T$, which is the intensive thermodynamic quantity controlling the energy fluctuations. Indeed, by considering both the system and the external bath as a single system belonging to the microcanonical ensemble, i.e. with fixed total energy, it is possible to prove that the probability distribution as a function of the phase space coordinates takes the form
\begin{equation}
\label{boltzmann}
\mathcal{P}_{NVT}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) = \frac{1}{\mathcal{Z}} \exp\left[ -\frac{1}{k_B T}\mathcal{H}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)\right]\,,
\end{equation}
where $k_B$ is the Boltzmann constant and $\mathcal{Z} = \mathcal{Z}(N,V,T)$ is the normalization factor, called \emph{canonical partition function}. This functional form defines what is called the \emph{Boltzmann distribution} (or canonical distribution), in term of which the generic ensemble average of the physical observable $a$ can be written as:
\begin{equation}
\label{ens_average_canonical}
\langle a \rangle = \frac{1}{\mathcal{Z}}\int_{\Omega} \Big(\prod_{i=1}^N d\mathbf{q}_id\mathbf{p}_i\Big)\, \exp\left[ -\frac{1}{k_B T}\mathcal{H}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)\right]\,a\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)\,,
\end{equation}
Some relevant relations following from \cref{boltzmann} and \cref{ens_average_canonical} are the expression for the average energy of the system,
\begin{equation}
\langle \mathcal{H} \rangle = -\frac{\partial \log \mathcal{Z}}{\partial\beta}\,,
\end{equation}
where $\beta=1/(k_B T)$, and its relation with the temperature and the standard deviation $\sigma_{\mathcal{H}}$ of the energy distribution:
\begin{equation}
\frac{\partial \langle \mathcal{H} \rangle}{\partial T} = \frac{1}{k_B T^2}\sigma_{\mathcal{H}}^2\,.
\end{equation}
As a function of the kinetic energy $K$ only, the canonical distribution becomes a gamma distribution:
\begin{equation}\label{kineng_distr}
\mathcal{P}_{NVT}(K) \propto K^{\frac{N_f}{2}-1}\,e^{-\frac{K}{k_B T}}\,.
\end{equation}
Here $N_f$ is the number of degrees of freedom in the system. As a consequence, the average kinetic energy $\langle K\rangle$ and its standard deviation $\sigma_K$ are:
\begin{subequations}
\begin{align}
\langle K \rangle &= \frac{1}{2}N_f k_B T \,,\label{avg_kin}\\
\sigma_K &= \langle K \rangle \sqrt{\frac{2}{N_f}}\,.
\end{align}
\end{subequations}
It is worth observing that \cref{avg_kin} is in perfect agreement with the classical equipartition theorem. Instead of studying the behaviour of the instantaneous kinetic energy $K$, an equivalent description can be given in terms of the instantaneous temperature $T_s$, defined as:
\begin{equation}\label{instantaneous_temp}
\frac{1}{2}N_d k_B T_s = K\,.
\end{equation}
\section{Thermostats}
Simulating a system in the canonical ensemble means generating a sequence of samplings from the kinetic energy distribution defined in \cref{kineng_distr}, and the algorithms to accomplish this task are called \emph{thermostats} in the MD language. Thermostats can be either \emph{global}, when all the single atom kinetic energies are rescaled with the same factor - the same used to rescale the total kinetic energy - or \emph{local}, when the change in the total kinetic energy is a result of a different rescaling for each atom \cite{global_vs_local}. A second distinction is between \emph{deterministic} and \emph{stochastic} thermostats, where the second category employs random number generators.
\subsection{Properties of sampling algorithms}\label{properties_sampling}
A good thermostat should have at least three properties, which are stated below for a generic sampling algorithm and a generic probability distribution.
\begin{enumerate}[label=\textbf{\arabic*}, align = left, labelwidth = 2 em, labelsep = 0 em]
\item \textbf{Stationarity of the target distribution (or balance).} A sampling algorithm S is said to satisfy the balance condition with respect to the probability distribution $\mathcal{P}(x)$ if the application of S on a set of $N$ samplings $\{x_i\}_{i=1}^N$ independently drawn from $\mathcal{P}(x)$ produces a new set of samplings $\{x'_i\}_{i=1}^N$ that are still independently drawn from $\mathcal{P}(x)$, in the limit of large $N$. Equivalently, $\mathcal{P}(x)$ is said to be stationary with respect to S.
\item \textbf{Ergodicity.} A sampling algorithm S is said to be ergodic with respect to its stationary distribution $\mathcal{P}(x)$ if, starting from any point $x_0$ in the domain of $\mathcal{P}$, it allows to reach in a finite number of steps any point $\overbar{x}$ such that $\mathcal{P}(\overbar{x}) \neq 0$.
It is possible to prove that a necessay condition for the ergodicity of a sampling algorithm is the existence of a single stationary distribution.
\item \textbf{Fast decorrelation of the samplings.} An ideal sampling algorithm should generate independent samplings of the target distribution. The degree of correlation after a time lag $\tau$ can be measured via the autocorrelation function
\begin{equation}
C(\tau)=\frac{\langle x_t x_{t+\tau}\rangle - \langle x\rangle^2}{\sigma_x^2}\,,
\end{equation}
which has the properties $C(\tau) \in [-1,1]$, $C(0) = 1$ and for which $C(\tau) = 0$ denotes independent samplings after a time lag $\tau$.
\end{enumerate}
\subsection{Berendsen thermostat}
The Berendsen thermostat \cite{berendsen} is a global and first-order deterministic thermostat, i.e. it can be formulated in terms of a first-order differential equation that evolves the total kinetic energy $K$ towards the target value $\overbar{K} = \frac{1}{2}N_f k_B T$:
\begin{equation}
\label{berendsen_eq}
\text{d}K = \frac{\overbar{K}-K}{\tau}\,\text{d}t
\end{equation}
The parameter $\tau$ plays the role of the relaxation time of the kinetic energy and it is also related to the autocorrelation time of the samplings; with the proper timestep it can be chosen arbitrarily small without any drawback.
The implementation of the Berendsen thermostat simply consists in the following momentum rescaling, applied to each atom $i$:
\begin{equation}
\label{berendsen_alg}
\mathbf{p}_i(t+\Delta t) = \mathbf{p}_i(t)\,\sqrt{\frac{e^{-\Delta t/\tau}\,K + (1-e^{-\Delta t/\tau})\,\overbar{K}}{K}}\,.
\end{equation}
In fact, it is possible to show that \cref{berendsen_alg} brings to \cref{berendsen_eq} in the limit $\Delta t/\tau\rightarrow 0$.
Although the Berendsen thermostat is efficient in the equilibration of the system it is typically not employed in the production phase, since it reproduces the correct average of the kinetic energy but not the higher order moments of its canonical distribution, i.e. it targets a probability distribution which is not the one defined in \cref{kineng_distr}.
\subsection{Andersen thermostat}
The temperature coupling method developed by Andersen \emph{et al.} \cite{andersen} consists in a local and stochastic thermostat that does not admit a continuous formulation, i.e. it cannot be described in terms of a differential equation for the kinetic energy or the momenta. In its easiest formulation, the idea of the algorithm is to propagate the system at constant energy, i.e. in the microcanonical ensemble, and to redefine the momentum of each atom once every $n_s$ steps by extracting its components $p_i^{\alpha}$ from their reference distribution in the canonical ensemble,
\begin{equation}\label{gaussian_momentum}
\mathcal{P}_{NVT}(p_i^{\alpha}) = \frac{1}{\sqrt{2\pi m_i k_B T}}\,\exp\left(-\frac{(p_i^{\alpha})^2}{2 m_i k_B T}\right)\,,
\end{equation}
which can be obtained from \cref{boltzmann} by marginalizing the Boltzmann distribution over all the atom positions $\{\mathbf{q}_i\}$ and the remaining momenta $\{\mathbf{p}_j\}_{j\neq i}$. The parameter $n_s$ is called \emph{stride} of the thermostat, and the resampling can also be applied using a different stride for each atom or choosing randomly a different atom at each step. Thus, the algorithm can be easily implemented using a Gaussian random number generator. Using the formulation with a common stride for all the $N$ atoms, the scheme of the algorithm is the following: \\
\begin{algorithm}[H]
\SetAlgoLined
velocity Verlet (or leapfrog integrator) for $n_s$ steps\;
\For{$i=1$ \KwTo $N$}{
\For{$\alpha = x,y,z$} {
extract $\mathcal{R}_{i}^{\alpha}\sim\mathcal{N}(0,1)$\;
update momentum component: $p_{i}^{\alpha} \gets \sqrt{m_i\, k_B T}\,\mathcal{R}_{i}^{\alpha}$\;
}
}
\caption{Andersen thermostat}
\label{andersen_alg}
\end{algorithm}
\vspace{0.4cm}
It is worth underlining that in the original formulation of the Andersen algorithm \cite{andersen} particles' momenta are not randomized with a fixed stride; instead, for each particle the time intervals $\Delta t_j$ between successive collisions with the bath are extracted from the distribution
\begin{equation}
\mathcal{P}(\Delta t_j) = \nu e^{-\nu \Delta t_j}\,,
\end{equation}
where the input parameter $\nu$ is the the mean rate of the collisions.
\subsection{Langevin thermostat}\label{langevin_thermo}
The Langevin thermostat \cite{langevin_bussi} evolves each momentum $\mathbf{p_i}$ according to the stochastic equation (see Appendix~\ref{appendix_stocdiffeq})
\begin{equation}
\label{langevin_eq}
\text{d}\mathbf{p}_i = -\gamma \mathbf{p}_i\text{d}t + \sqrt{2\gamma m_i\,k_B T}\text{d}\mathbf{W}_i\,,
\end{equation}
where the scalar parameter $\gamma$ is called \emph{friction}. Since \cref{langevin_eq} includes a noise term and is referred to the momentum of the single atom $i$, this thermostat is stochastic and local. It is possible to interpret the Langevin thermostat as a continuous version of the Andersen one shown in \cref{andersen_alg}, where the exact resampling from the target distribution is replaced by the "smoother" update
\begin{equation}
\label{mom_update_andersen}
\mathbf{p}_i (t+\Delta t) = e^{-\Delta t/\tau}\mathbf{p}_i(t) + \sqrt{\left(1-e^{-2\Delta t/\tau} \right)m_i\, k_B T}\,\mathbfcal{R}_i\,.
\end{equation}
In fact, by defining $\gamma\equiv 1/\tau$ it is possible to recover \cref{langevin_eq} from \cref{mom_update_andersen} in the limit $\Delta t/\tau\rightarrow 0$. By coupling the Langevin thermostat with Hamilton's equations one gets the \emph{underdamped Langevin equations}
\begin{subequations}
\begin{align}
\text{d}\mathbf{q}_i &= \frac{\mathbf{p}_i}{m_i}\text{d}t\,, \\
\text{d}\mathbf{p}_i &= \mathbf{F}_i\text{d}t - \gamma \mathbf{p}_i\text{d}t + \sqrt{2\gamma m_i\,k_B T}\text{d}\mathbf{W}_i\,,
\end{align}
\end{subequations}
which can be integrated in a time-reversible way with the following scheme:\\
\begin{algorithm}[H]
\SetAlgoLined
\For{$i=1$ \KwTo $N$}{
extract $\mathbf{\mathbfcal{R}}_{i}\sim\mathcal{N}_3\left((0,0,0),(1,1,1)\right)$\;
update $\mathbf{p}_i$ with time step $\frac{\Delta t}{2}$: \indent$\hspace{1cm}\mathbf{p}_i \gets e^{-\Delta t/(2\tau)}\mathbf{p}_i + \sqrt{\left(1-e^{-\Delta t/\tau} \right)m_i\, k_B T}\,\mathbf{\mathbfcal{R}}_i$\;
}
Velocity Verlet (or leapfrog integrator)\;
\For{$i=1$ \KwTo $N$}{
extract $\mathbf{\mathbfcal{R}}_{i}\sim\mathcal{N}_3\left((0,0,0),(1,1,1)\right)$\;
update $\mathbf{p}_i$ with time step $\frac{\Delta t}{2}$: \indent$\hspace{1cm}\mathbf{p}_i \gets e^{-\Delta t/(2\tau)}\mathbf{p}_i + \sqrt{\left(1-e^{-\Delta t/\tau} \right)m_i\, k_B T}\,\mathbf{\mathbfcal{R}}_i$\;
}
\caption{Langevin thermostat}
\label{langevin_alg}
\end{algorithm}
\vspace{0.4cm}
\noindent In the Langevin thermostat the value of the friction $\gamma$ affects in a non-trivial way the efficiency of the algorithm, especially when the system has to be equilibrated. Since a large friction accounts for a strong coupling with the external bath, one could think that increasing $\gamma$ always leads to accelerate the thermalization of the system. Conversely, when $\gamma$ is chosen too large the equilibration time actually increases \cite{global_vs_local}; as a consequence, for each system it exists an optimal value of $\gamma$ that minimizes the relaxation time, but there is no general recipe to set it a priori. The physical interpretation of this behaviour is that when the friction is too large the "collisions" of the particles with the bath become so frequent that they suppress the collisions of the particles among themselves, reducing the momentum exchanges and slowing down the dynamics of the system. Hence, since different choices of $\gamma$ can alter significantly the Hamiltonian dynamics, this algorithm cannot be used to compute dynamical properties, unless an extremely small friction is used \cite{global_vs_local}.
The Langevin thermostat can be shown to satisfy the \emph{detailed balance} condition (see Appendix~\ref{appendix_fp}) with respect to the canonical distribution in \cref{boltzmann} and it allows to calculate the \emph{effective energy drift} (see Appendix~\ref{appendix_effenergy}), which quantifies the detailed balance violations \cite{langevin_bussi}.
\subsection{Nosé-Hoover thermostat}
The Nosé-Hoover method \cite{nose-original,hoover} is a global thermostat based on a second-order deterministic equation, which can be derived from an extended Hamiltonian that includes a new degree of freedom related to the bath coupling. Split into two first-order equations, the Nosé-Hoover dynamics reads
\begin{subequations}
\begin{align}
\text{d}K &= -2 \gamma K \text{d}t\,, \\
\text{d} \gamma &= \frac{2}{M}\left(K -\overbar{K} \right)\text{d}t\,,
\end{align}
\end{subequations}
where $\gamma$ is a time-dependent friction and $M$ - usually called \emph{thermostat mass} - is a non-trivial parameter which affects the behaviour in the equilibration phase as well as the autocorrelation time of the samplings. In particular, a non-optimal choice of $M$ can bring to large and slow oscillations when the system is initialized far from equilibrium; for this reason, first-order thermostats are in general more efficient when a system needs to be thermalized.
Using the formalism of Fokker-Planck equations (see Appendix~\ref{appendix_fp}) it is possible to show that the canonical distribution defined in \cref{kineng_distr} satisfies the condition of stationarity described in \cref{properties_sampling}. However the Nosé-Hoover algorithm is not ergodic, especially for small and stiff systems and harmonic oscillators. Ergodicity can be achieved by coupling the friction variable to a second external bath, typically implemented as a new Nosé-Hoover thermostat with its own friction. This coupling procedure can be repeated several times, in a scheme named \emph{Nosé-Hoover chains} \cite{nosehoover_chains}. If the additional coupling is performed with of a Langevin thermostat, the temperature coupling method is called \emph{Nosé-Hoover-Langevin thermostat} \cite{nose-hoover-langevin}.
\subsection{Stochastic velocity rescaling}\label{SVR}
Stochastic velocity rescaling (SVR), also known as Bussi-Donadio-Parrinello thermostat \cite{SVR}, is a global and first-order algorithm based on the following stochastic equation:
\begin{equation}
\text{d}K=-(K-\overbar{K})\frac{\text{d}t}{\tau} + 2\sqrt{\frac{K\overbar{K}}{\tau N_f}}\text{d}W\,,
\end{equation}
where $N_f$ is the number of degrees of freedom and $\overbar{K}$ is the target kinetic energy, related to the external temperature via the usual equipartition theorem $\overbar{K}=N_f k_B T/2$. It is worth noting that the deterministic part of the equation is exactly the Berendsen thermostat; hence the time constant $\tau$ can be interpreted also in this case as the relaxation time of the system and the autocorrelation time of the samplings. The additional noise term brings to a Fokker-Planck equation that can be shown to satisfy the \emph{detailed balance} condition (see Appendix \cref{appendix_fp}) with respect to the canonical distribution in \cref{kineng_distr}, meaning that SVR generates the correct canonical ensemble. The algorithm can be formulated in a time-reversible way that allows the calculation of the effective energy drift (see Appendix \ref{appendix_effenergy}).
\subsection{Monte Carlo thermostat}\label{MC_thermostat}
Canonical sampling can be also achieved through the popular technique of \emph{Markov chain Monte Carlo} (MCMC), that will be here discussed in a general framework and considering a discrete sample space. The aim of MCMC is to achieve the ergodic sampling of a probability distribution $\mathcal{P}(x)$ by repeatedly applying a stochastic rule embedded in the transition matrix $\Pi_{x,x'} \equiv \Pi(x\rightarrow x')$, which represents the conditional probability of sampling $x'$ starting from the previous sample $x$. The fact that this conditional probability only depends on the previous step is called \emph{Markov property}. $\Pi(x\rightarrow x')$ is a stochastic matrix, i.e. it satisfies the two properties
\begin{subequations}
\begin{align}
\label{normaliz_trmatrix}
\sum_{x'}\Pi(x\rightarrow x') &= 1\,,\\
\Pi(x\rightarrow x') &\geq 0\,,
\end{align}
\end{subequations}
and it can be used to relate the marginal probabilities of consecutive steps, via the so-called \emph{Master equation}:
\begin{equation}\label{master_eq_mcmc}
\mathcal{P}_{i+1}(x) = \mathcal{P}_{i}(x) + \sum_{x'\neq x} \mathcal{P}_{i}(x') \Pi(x'\rightarrow x) - \sum_{x'\neq x} \mathcal{P}_{i}(x) \Pi(x\rightarrow x')\,.
\end{equation}
In \cref{master_eq_mcmc} the index $i$ represents the discrete time step in the sampling chain, and the last two terms in the RHS of the equation are called \emph{gain} and \emph{loss} terms respectively. The stationarity or balance condition reads $\mathcal{P}_{i+1}(x) = \mathcal{P}_i(x)$, hence it brings to the equation
\begin{equation}\label{balance_MC}
\sum_{x'} \mathcal{P}(x') \Pi(x'\rightarrow x) = \sum_{x'} \mathcal{P}(x) \Pi(x\rightarrow x')\,,
\end{equation}
where the case $x'= x$ is now included in the sums since it gives the same contribution to the RHS and LHS of the equation. By applying the property in \cref{normaliz_trmatrix} this condition becomes
\begin{equation}\label{lefteig}
\sum_{x'} \mathcal{P}(x') \Pi(x'\rightarrow x) = \mathcal{P}(x) \,.
\end{equation}
In words, \cref{lefteig} tells that $\mathcal{P}$ is a stationary distribution with respect to the sampling algorithm embedded in $\Pi$ if $\mathcal{P}$ is a a left eigenvector of $\Pi$ with eigenvalue equal to 1. The existence of such a distribution is a consequence of the Perron-Frobenius theorem, and the ergodicity of the sampling algorithm can only be achieved if the left eigenvector of $\Pi$ is unique. A simpler condition that implies the stationarity of $\mathcal{P}$ in \cref{balance_MC} is the so-called \emph{detailed balance} or \emph{equilibrium} condition, that in the framework of MCMC reads:
\begin{equation}\label{detailed_balance_MC}
\mathcal{P}(x') \Pi(x'\rightarrow x) = \mathcal{P}(x) \Pi(x\rightarrow x')\,.
\end{equation}
This stricter condition is usually more employed than the balance one since it is generally easier to construct algorithms satisfying \cref{detailed_balance_MC} than \cref{balance_MC}; however, detailed balance is not a necessary condition for MCMC \cite{db_notnecessary}. The first step to create a rule satisfying the detailed balance condition is to split the transition matrix $\Pi$ into a \emph{proposal matrix} $M$ and an \emph{acceptance matrix} $\alpha$:
\begin{equation}
\Pi(x\rightarrow x') = M(x\rightarrow x') \,\alpha(x\rightarrow x')\,.
\end{equation}
By decomposing $\Pi$ in this way, \cref{detailed_balance_MC} can be rewritten as
\begin{equation}\label{db_ratioacc}
\frac{\alpha(x\rightarrow x')}{\alpha(x'\rightarrow x)} = \frac{\mathcal{P}(x')\,M(x'\rightarrow x)}{\mathcal{P}(x)\,M(x\rightarrow x')}\,.
\end{equation}
Among several possibilities to satify \cref{db_ratioacc}, the one giving the highest acceptance is the Metropolis-Hastings rule:
\begin{equation}
\alpha(x\rightarrow x') \equiv \min \left(1,\,\frac{\mathcal{P}(x')\,M(x'\rightarrow x)}{\mathcal{P}(x)\,M(x\rightarrow x')}\right)
\end{equation}
Although not strictly necessary, the proposal rule embedded in $M$ is typically constructed in order to make $M$ a symmetric matrix, so that the acceptance can be simply calculated as
\begin{equation}
\alpha(x\rightarrow x') \equiv \min \left(1,\,\frac{\mathcal{P}(x')}{\mathcal{P}(x)}\right)\,.
\end{equation}
Since the acceptance $\alpha$ is always calculated with ratios of probabilities, the Metropolis-Hastings rule allows to sample a generic distribution $\mathcal{P}$ without knowing its normalization, which can be hard to compute in a high-dimensional space. In general, the trial move embedded in the proposal matrix $M$ should be designed in order to obtain the highest possible value for the average value of $\alpha$ over the simulation. A scheme of the method is reported in \cref{MC_algorithm}.\\
\begin{algorithm}[H]
\SetAlgoLined
Propose the move $x\rightarrow x'$\;
Compute the acceptance: $\alpha=\min \left(1,\,\frac{\mathcal{P}(x')}{\mathcal{P}(x)}\right) $ \;
Extract a uniform random number in $[0,1]$: $\mathcal{R}\sim U(0,1)$\;
\eIf{$\alpha> \mathcal{R}$}{
Accept the move: $x\gets x'$\;
}{
Refuse the move: $x\gets x$\;
}
\caption{Markov chain MC with Metropolis-Hastings}
\label{MC_algorithm}
\end{algorithm}
\vspace{0.4cm}
When a system is simulated in the canonical ensemble this scheme can be used to perform single-particle trial moves on their positions, where at each step the particle involved can be selected both with a deterministic \emph{sweep} strategy and with random selections \cite{sweep}. However, when constraints are present the construction of an efficient trial move becomes a complex task, and more efficient Monte Carlo strategies can be employed. As an additional limitation, standard MCMC applied to the canonical distribution carries no information about momentum variables, since the proposal move only involves atomic positions. \\
\subsection{Hybrid Monte Carlo}\label{hybrid_MC}
Hybrid (or smart) Monte Carlo \cite{hybrid_MC} is a technique to sample the canonical ensemble that employs a molecular dynamics algorithm in the microcanonical ensemble - such as velocity Verlet - to propose a new move, and the Metropolis-Hastings rule to accept it or refuse it.
As a starting point to illustrate the method, it is possible to show that the canonical distribution, and more in general every distribution $\mathcal{P}$ depending on the phase space coordinates via the Hamiltonian $\mathcal{H}$,
\begin{equation}\label{p_of_H}
\mathcal{P} = \mathcal{P}\Big(\mathcal{H}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)\Big)\,,
\end{equation}
is stationary with respect to the Hamiltonian dynamics. Indeed the Fokker-Planck equation that one can write from \cref{hamilton_eqs1,hamilton_eqs2}, which does not include any diffusion term due to the deterministic nature of Hamilton's equations (see Appendix \ref{appendix_fp}), is
\begin{equation}
\frac{\partial \mathcal{P}}{\partial t} = - \sum_{i,\alpha} \frac{p_i^{\alpha}}{m_i} \frac{\partial\,\mathcal{P}}{\partial q_i^{\alpha}} - \sum_{i,\alpha}F_i^{\alpha}\frac{\partial\,\mathcal{P}}{\partial p_i^{\alpha}}\,
\end{equation}
and it is straightforward to show that $\frac{\partial \mathcal{P}}{\partial t} = 0$ if \cref{p_of_H} holds. Hence, the canonical distribution is stationary with respect to Hamilton's equations. Nevertheless this is not a sufficient condition to sample the canonical distribution, because the ergodicity condition is clearly not satisfied. The idea of Hybrid MC is to achieve ergodicity by employing a microcanonical integrator that exactly satisfies the time-reversibility property of Hamilton's equations (see \cref{hamilton}) but violates the energy conservation, so that energy variations can be accepted or refused according to the Metropolis-Hastings rule. In order to apply this rule, a detailed balance condition such as the one in \cref{detailed_balance_MC} should be satisfied. Calling a point in phase space $\mathbf{x} \equiv \big(\{\mathbf{q}_i,\mathbf{p}_i\}\big)$ and embedding the (deterministic) Hamiltonian dynamics in the matrix $\Pi$, it is clear that the standard detailed balance condition does not hold with respect to the canonical distribution, since in general $\Pi(\mathbf{x'}\rightarrow\mathbf{x}) = 0$ if $\Pi(\mathbf{x}\rightarrow\mathbf{x'}) \neq 0$.
On the other hand, defining $\mathbf{x^*} \equiv \left(\{\mathbf{q}_i,-\mathbf{p}_i\}\right)$ and recalling the time-reversibility property of Hamilton's equations, the transition matrix $\Pi$ satisfies:
\begin{equation}
\Pi(\mathbf{x}\rightarrow\mathbf{x'}) = \Pi(\mathbf{x'^*}\rightarrow\mathbf{x^*})\,.
\end{equation}
Observing that the canonical distribution $\mathcal{P}_{NVT}$ in \cref{boltzmann} does not depend on the sign of momenta, it is possible to introduce a condition called \emph{generalized detailed balance},
\begin{equation}\label{generalized_db}
\mathcal{P}_{NVT}(\mathbf{x})\,\Pi(\mathbf{x}\rightarrow\mathbf{x'}) = \mathcal{P}_{NVT}(\mathbf{x'^*})\,\Pi(\mathbf{x'^*}\rightarrow\mathbf{x^*})\,,
\end{equation}
which justifies the use of the Metropolis-Hastings rule. In a schematic way, a single iteration of the hybrid MC method is reported in \cref{hybrid_MC_alg}.
It is worth Note that the parameter $\Delta t$ employed in the proposal move acquires a whole new meaning in this context: in the propagation of a system in the microcanonical ensemble $\Delta t$ is the physical time step in the dynamics and it is related to the error made by the algorithm; in hybrid MC instead $\Delta t$ is a parameter which affects the average acceptance and the efficiency of the sampling, but not its correctness. In other words, hybrid MC (as well as standard MC) does not carry any information about the time dependence of the fluctuations, because the discrete "MC time" has no relation with the physical time in the dynamics of the system.
\begin{algorithm}[h!]
\SetAlgoLined
Compute initial energy: $E=\mathcal{H}(\mathbf{x})$\;
Propose new state with velocity Verlet (or leapfrog integrator): $\mathbf{x}\rightarrow\mathbf{x'}$\;
Compute energy of proposed state: $E'=\mathcal{H}(\mathbf{x'})$\;
Compute acceptance: $\alpha=\min \left[1,\,\exp{\left(-\frac{1}{k_B T}(E'-E)\right)}\right] $ \;
Extract uniform random number: $\mathcal{R}\sim U(0,1)$\;
\eIf{$\alpha> \mathcal{R}$}{
Accept the move: $\mathbf{x}\gets\mathbf{x'}$
}{
Refuse the move: $\mathbf{x}\gets\mathbf{x}$
}
\caption{Hybrid Monte Carlo}
\label{hybrid_MC_alg}
\end{algorithm}
\section{Isothermal-isobaric ensemble (NPT)}\label{npt_theory}
A system in the isothermal-isobaric ensemble is defined by a fixing the number of particles $N$, the external pressure $P_0$ and the external temperature $T$. In the isotropic formulation of the ensemble the volume $V$ is allowed to fluctuate according to the probability distribution
\begin{equation}\label{npt_isotropic}
\mathcal{P}_{NP_0T}\big(\{\mathbf{q}_i,\mathbf{p}_i\},V\big) = \frac{1}{\Delta}\exp\left[-\frac{1}{k_B T}\Big(K\big(\{\mathbf{p}_i\}\big) + U\big(\{\mathbf{q}_i\}\big) + P_0 V\Big)\right]\,,
\end{equation}
where $\Delta=\Delta(N,P_0,T)$ is the isothermal-isobaric partition function. In the limit case of an ideal gas it is easy to show that the marginal distribution of the volume is reduced to a Gamma distribution, $\mathcal{P}_{NP_0T}^{\text{id}}(V) \propto V^N \exp \left(-\frac{1}{k_B T}P_0 V \right)$.
As an instantaneous temperature was defined in the $NVT$ ensemble, it is also possible to define an instantaneous \emph{internal pressure} in the $NP_0 T$ ensemble, which is calculated via the Clausius virial theorem as
\begin{equation}\label{internal_pressure}
P_{\text{int}} = \frac{2 K}{3 V} - \frac{\partial U}{\partial V}\,.
\end{equation}
At a given temperature $T$ and number of particles $N$, the variation of the average volume with respect to the pressure, which is an intrinsic property of the system, is quantified by the \emph{isothermal compressibility}:
\begin{equation}
\beta_T = -\frac{1}{\langle V \rangle}\frac{\partial \langle V\rangle}{\partial P_0}\,,
\end{equation}
This thermodynamic quantity is tightly related to volume fluctuations, since it can be computed as
\begin{equation}\label{beta_T_formula}
\beta_T = \frac{k_B T}{\langle V\rangle}\sigma_V^2\,,
\end{equation}
where $\sigma_V^2 = \langle \left(V-\langle V\rangle \right)^2\rangle$ is the variance of the volume distribution at pressure $P_0$. The same information carried by $\beta_T$ is sometimes expressed in terms its reciprocal, called bulk modulus:
\begin{equation}
k_T = \frac{1}{\beta_T}\,.
\end{equation}
While \cref{npt_isotropic} assumes that the volume fluctuates in a isotropic way, namely that the box where the system is confined changes its size but not its shape, there are several situations where anisotropic fluctuations can be relevant. For instance, in solid-state physics a fully flexible description of the system allows to predict crystal structures \cite{crystal_structure} and to study conformational transitions between them \cite{parr-rahman}. Moreover, semi-isotropic volume fluctuations can play a central role to study liquid-liquid interfaces \cite{constant_surface_tension} or to simulate membranes. In the fully flexible isothermal-isobaric ensemble the system is typically contained within a general parallelepiped, which represents the most general box shape and appears appropriate to describe, for example, solids whose unit cells are generally triclinic \cite{tuckerman}. Such a box can be described in terms of three vectors $\mathbf{a},\,\mathbf{b},\,\text{and}\,\mathbf{c}$ that lie along the three edges starting from a certain vertex (see \cref{fig:box}). Their nine components can be collected in the 3$\times$3 \emph{box matrix} (or \emph{cell matrix}) $\mathbf{h}$, which contains the three box vectors along its columns according to the convention adopted here:
\begin{equation}
\mathbf{h}=
\begin{pmatrix}
a_x & b_x & c_x\\
a_y & b_y & c_y\\
a_z & b_z & c_z
\end{pmatrix}
\end{equation}
Using Greek and Latin letters to label the cartesian component and the number of the cell vector espectively, the elements of the box matrix will also be written as $h_{\alpha i}$ ($\alpha=x,y,z,\,i=1,2,3$).
\noindent The volume of the box is given by the triple product of the three cell vectors $\mathbf{a},\,\mathbf{b},\,\text{and}\,\mathbf{c}$, assuming that they form in that order a right-handed triad:
\begin{equation}\label{volume_box}
V = \mathbf{a}\cdot\mathbf{b}\times\mathbf{c} = \det \mathbf{h}\,.
\end{equation}
From \cref{fig:box} it is clear that the shape of the box can actually be described with six scalar numbers only, i.e. the moduli $|\mathbf{a}|,\,|\mathbf{b}|,\,|\mathbf{c}|$ of the three cell vectors and the angles $\alpha,\,\beta,\,\gamma$ between them. The three additional degrees of freedom in the box matrix $\mathbf{h}$ account for overall rotations of the cell, which leave both the moduli and the angles untouched. Since rotations of the entire system are not of interest in MD simulations, the three redundant degrees of freedom can be eliminated by employing different methods, which are discussed in \cref{eliminate_rotations}.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.25\textwidth]{core/first_chapter/box.PNG}
\caption{Construction of the box using the three cell vectors $\mathbf{a},\,\mathbf{b},\,\text{and}\,\mathbf{c}$.}
\label{fig:box}
\end{figure}
In the anisotropic formulation of the isothermal-isobaric ensemble, the reference pressure $P_0$, also called external \emph{hydrostatic pressure}, is generalized to a symmetric 3$\times$3 tensor $\mathbf{S}$, where the off-diagonal components account for possible shear stresses applied to the system while the diagonal ones are related to the isotropic counterpart via the relation
\begin{equation}\label{hydro_pressure}
P_0 = \frac{\text{Tr}\left(\mathbf{S}\right)}{3}\,.
\end{equation}
$\mathbf{S}$ is commonly referred to as \emph{external pressure tensor} or \emph{stress tensor}, and the anisotropic iosthermal-isobaric ensemble is also denoted as $N\mathbf{S}T$ ensemble. Given \cref{hydro_pressure}, $\mathbf{S}$ can be split into the sum of a hydrostatic contribution and a trace-less \emph{deviatoric stress tensor} $\mathbf{S}_{\text{dev}}$:
\begin{equation}
\mathbf{S} = P_0\mathbf{I} + \mathbf{S}_{\text{dev}} \,,
\end{equation}
where $\mathbf{I}$ is the 3$\times$3 identity matrix.
Similarly, the scalar internal pressure $P_{\text{int}}$ is generalized to a pressure tensor $\mathbf{P}_{\text{int}}$ that is calculated with a tensorial formulation of the virial theorem,
\begin{equation}\label{internal_pressure_tensor}
\mathbf{P}_{\text{int}} = \frac{2}{V}\left(\mathbf{K}-\Xi\right)\,,
\end{equation}
where $\mathbf{K}$ and $\Xi$ are called \emph{kinetic energy tensor} and \emph{virial tensor} respectively:
\begin{subequations}
\begin{align}
K_{\alpha\beta} &= \sum_{i=1}^N \frac{p_i^{\alpha}p_i^{\beta}}{2m_i}\,, \\
\Xi_{\alpha\beta} &= -\frac{1}{2}\sum_{i=1}^N F_i^{\alpha}\,q_i^{\beta}\,.
\end{align}
\end{subequations}
\noindent The relations between these tensors and the corresponding scalar quantities are straightforward:
\begin{subequations}
\begin{align}
K &= \frac{\text{Tr}\left(\mathbf{K}\right)}{3}\,, \\
P_{\text{int}} &= \frac{\text{Tr}\left(\mathbf{P}_{\text{int}}\right)}{3}\,.
\end{align}
\end{subequations}
Given these definitions, in presence of an external hydrostatic pressure - i.e. $\mathbf{S} = P_0 \mathbf{I}$ - the $N\mathbf{S}T$ ensemble is described by the following distribution:
\begin{equation}\label{NPT_anisotropic}
\mathcal{P}_{N\mathbf{S}T}\big(\{\mathbf{q}_i,\mathbf{p}_i\},\mathbf{h}\big) = \frac{\left(\det \mathbf{h}\right)^{-2}}{\Delta} \exp\left[-\frac{1}{k_B T}\Big(K + U + P_0 \det \mathbf{h}\Big)\right]\,.
\end{equation}
\noindent To derive this expression and explain the origin of the factor $\left(\det\mathbf{h}\right)^{-2}$ it is possible to start from the expression of the isothermal-isobaric partition function $\Delta(N,P_0,T)$ in terms of the isotropic $NP_0 T$ distribution:
\begin{equation}
\Delta(N,P_0,T) = \frac{1}{V_0}\int \text{d}V \exp \left(-\frac{P_0 V}{k_B T}\right)\,\mathcal{Z}(N,V,T)\,,
\end{equation}
where $V_0$ is a reference volume that is necessary to make the partition function dimensionless.
Introducing the matrix $\mathbf{h}_u$ such that $\mathbf{h}=V^{1/3}\mathbf{h}_u$, $\mathcal{Z}(N,V,T)$ can be rewritten in terms of a constrained canonical ensemble - with partition function $\mathcal{Z}(N,V,T,\mathbf{h}_u)$ - where not only the volume of the system is fixed but also its shape:
\begin{align}
\mathcal{Z}(N,V,T) &= \frac{1}{V_0}\int \text{d}\mathbf{h}_u\, {Z}(N,V,T,\mathbf{h}_u)\,\delta(\det\mathbf{h}_u -1) \\
&= \frac{1}{V_0}\int \text{d}\mathbf{h}_u \int\left(\prod_i\text{d}\mathbf{q}_i\text{d}\mathbf{p}_i\right)\,\exp\left(-\frac{1}{k_B T}\mathcal{H}\big(\{\mathbf{q}_i,\mathbf{p}_i\}\big) \right) \delta(\det\mathbf{h}_u -1) \nonumber
\label{from_iso_to_flex}
\end{align}
Since $\text{d}\mathbf{h}_u = \text{d}\mathbf{h}/V^3$ and $\delta(\det\mathbf{h}_u -1) = V\delta(V-\det\mathbf{h})$, substituting in \cref{from_iso_to_flex} and integrating the $\delta$-function in $\text{d}V$ it is possible to express the isothermal-isobaric partition function as
\begin{align}
\Delta(N,P_0,T) = \frac{1}{V_0}\int& \text{d}\mathbf{h}\int\left(\prod_i\text{d}\mathbf{q}_i\text{d}\mathbf{p}_i\right)\, \left(\det\mathbf{h}\right)^{-2} \nonumber \\
&\times\exp\left[-\frac{1}{k_B T}\Big(K\big(\{\mathbf{p}_i\}\big) + U\big(\{\mathbf{q}_i\}\big) + P_0 \det{\mathbf{h}}\Big)\right]\,,
\end{align}
which is consistent with the expression of the $N\mathbf{S}T$ distribution given in \cref{NPT_anisotropic}. In an arbitrary number $d$ of spatial dimensions, the relation between the original box and the unitary one is $\mathbf{h}=V^{1/d}\mathbf{h}_u$ and the factor $\left(\det\mathbf{h}\right)^{-2}$ is generalized to $\left(\det\mathbf{h}\right)^{1-d}$ \cite{tuckerman}.
In presence of a generic external stress $\mathbf{S}$ - namely if $\mathbf{S}_{\text{dev}}\neq \mathbf{0}$ - it is necessary to introduce an additional \emph{strain energy}
\begin{equation}\label{strain_energy}
E_{\text{S}} = \frac{1}{2}\text{Tr}\left(\bm{\Sigma}\mathbf{G}\right)\,,
\end{equation}
where $\mathbf{G} = \mathbf{h}^T\mathbf{h}$ is called \emph{metric tensor} and the matrix $\bm{\Sigma}$ is defined with respect to a reference system with cell $\mathbf{h}_0$ and volume $V_0=\det\mathbf{h}_0$, which is typically identified with the initial system in MD simulations:
\begin{equation}\label{sigma_matrix}
\bm{\Sigma} = V_0\, \mathbf{h}_0^{-1}\,\mathbf{S}_{\text{dev}}\,\big(\mathbf{h}_0^T\big)^{-1} = V_0\, \mathbf{h}_0^{-1}\left(\mathbf{S}-P_0\mathbf{I} \right)\big(\mathbf{h}_0^T\big)^{-1}\,.
\end{equation}
This energy contribution can be derived in the framework of elasticity theory and was fist proposed by Parrinello and Rahman \cite{parr-rahman}. Hence the $N\mathbf{S}T$ distribution in presence of a generic external stress becomes:
\begin{equation}\label{NPT_anisotropic_shear}
\mathcal{P}_{N\mathbf{S}T} \propto V^{-2}\exp\left[-\frac{1}{k_B T}\Big(K + U + P_0 V + \frac{1}{2}\text{Tr}\left(\bm{\Sigma}\mathbf{G}\right)\Big)\right]\,.
\end{equation}
An intermediate case between the fully flexible and isotropic ensembles is given by the \emph{constant normal pressure and surface-tension} ensemble $NP_0^\perp\gamma_0 T$, where volume fluctuations result from only two decoupled degrees of freedom, i.e. the area $A$ of the simulation box in the $xy$-plane and its height $L$. Supposing that the system is contained in an orthorhombic box, i.e. that the box matrix $\mathbf{h}$ is diagonal, $A = h_{x1} h_{y2}$ and $L = h_{z3}$. In this case the external stress can be written as the diagonal tensor
\begin{equation}
\mathbf{S}=
\begin{pmatrix}
P_{0,xx} & 0 & 0\\
0 & P_{0,yy} & 0\\
0 & 0 & P_{0,zz}
\end{pmatrix}\,,
\end{equation}
and the fixed control parameters of the ensemble, i.e. the normal pressure $P_0^\perp$ to the surface $A$ and the surface tension $\gamma$ (multiplied by the number of surfaces), are related to $\mathbf{S}$ by the following relations \cite{constant_surface_tension}:
\begin{subequations}
\begin{align}
P_0^\perp &= P_{0,zz}\,, \\
P_0^{\parallel} &= \frac{P_{0,xx} + P_{0,yy}}{2}\,,\\
\gamma_0 &= L\left(P_0^\perp - P_0^{\parallel}\right)\,. \label{gamma_def}
\end{align}
\end{subequations}
The additional degree of freedom brings a new energy contribution with respect to the isotropic case, resulting in the $NP_0^\perp\gamma_0 T$ probability distribution
\begin{equation}\label{npt_constant_surface_tension}
\mathcal{P}_{NP_0^\perp\gamma_0 T}\big(\{\mathbf{q}_i,\mathbf{p}_i\},A,L\big) \propto\exp\left[-\frac{1}{k_B T}\Big(K + U + P_0^\perp A L -\gamma_0 A\Big)\right]\,.
\end{equation}
\section{Barostats}
In the MD language, the algorithms which allow to simulate volume fluctuations at constant pressure are called barostats. Common features of all these algorithms are that they must be coupled with a thermostat in order to reproduce the correct isothermal-isobaric ensemble, and they all rescale the particles' positions together with the volume; instead, rescaling of momenta is not in general a necessary feature. As in the case of thermostats, also barostats can be distinguished in deterministic and stochastic ones.
\subsection{Monte Carlo barostat}\label{MC_barostat}
The first strategy to simulate a system at constant pressure and with isotropic volume fluctuations is to employ a MCMC approach (see \cref{MC_thermostat}), namely to propose a move on the volume that is accepted or refused according to the Metropolis-Hastings rule. This kind of algorithm is called MC barostat \cite{MC_barostat1,MC_barostat2}. The proposal move is typically implemented as
\begin{equation}
V' = V + (2\,\mathcal{R}-1)\delta V\,,
\end{equation}
where $\mathcal{R}\sim U(0,1)$ and $\delta V$ is a parameter that quantifies the maximum volume variation achievable in a single MC step. Before computing the acceptance, all the particles' positions are rescaled as
\begin{equation}\label{pos_rescaling_MC}
\mathbf{q}_i' = \left(\frac{V'}{V}\right)^{1/3}\mathbf{q}_i\,,
\end{equation}
with the assumption that the system is confined in a cubic box. This step is necessary to avoid large modifications of the distance - and thus of the interaction - between each atom and the copies of the others (see Appendix \ref{PBCs}), which would drastically reduce the acceptance. Rescaling the physical positions as in \cref{pos_rescaling_MC} is equivalent to say that the volume is propagated at constant \emph{rescaled positions}
\begin{equation}\label{rescaled_pos_iso}
\mathbf{s}_i \equiv \frac{\mathbf{q}_i}{V^{1/3}}\,.
\end{equation}
In terms of these variables, the $NP_0T$ distribution in \cref{npt_isotropic} becomes
\begin{equation}
\mathcal{P}_{NP_0T}\big(\{\mathbf{s}_i,\mathbf{p}_i\},V\big) = \frac{V^N}{\Delta}\exp\left[-\frac{1}{k_B T}\Big(K\big(\{\mathbf{p}_i\}\big) + U\big(\{\mathbf{s}_i\},V\big) + P_0 V\Big)\right]\,,
\end{equation}
therefore - when smaller than 1 - the acceptance $\alpha = \mathcal{P'}_{NP_0T}/ \mathcal{P_{NP_0T}}$ satisfies:
\begin{equation}
-k_B T \log\alpha = \Delta U + P_0\Delta V -N k_B T \Delta(\log V)\,,
\end{equation}
where $\Delta U = U\big(\{\mathbf{s}_i\},V'\big)-U\big(\{\mathbf{s}_i\},V\big)$. Taking the limit $\Delta V = V'-V\rightarrow0$ one finds that $\alpha$ can be written as
\begin{equation}\label{acceptance_MC_barostat}
-k_B T \log\alpha = \Delta V \left(P_0-P_{\text{int}}^{\langle K \rangle}\right)\,,
\end{equation}
where $P_{\text{int}}^{\langle K \rangle}$ is the internal pressure defined in \cref{internal_pressure} but computed with the average kinetic energy instead of the instantaneous one:
\begin{equation}\label{internal_pressure_avgkin}
P_{\text{int}}^{\langle K \rangle} = \frac{N k_B T}{V}-\frac{\partial U}{\partial V}\,.
\end{equation}
\noindent An alternative formulation of the MC barostat employs the additional rescaling
\begin{equation}
\mathbf{p}_i' = \left(\frac{V}{V'}\right)^{1/3}\mathbf{p}_i\,,
\end{equation}
namely the volume is propagated at constant \emph{rescaled momenta}
\begin{equation}\label{rescaled_mom_iso}
\bm{\pi}_i = V^{1/3} \mathbf{p}_i\,.
\end{equation}
Since the rescaling factor in \cref{rescaled_mom_iso} is the inverse of the one in \cref{rescaled_pos_iso}, the $NP_0 T$ distribution as a function of $\{\mathbf{s}_i,\bm{\pi}_i\}$ gains no additional prefactors, i.e. the Jacobians of the two tranformations cancel each other. It is possible to show that in this case the acceptance can be computed as in \cref{acceptance_MC_barostat} but with the actual expression of the internal pressure, i.e. substituting the average kinetic energy with the instantaneous one.
Since it does not require the calculation of the virial, the MC barostat is one of the easiest algorithms to control volume fluctuations at constant pressure; however there are situations where it appears less efficient than virial-based barostats \cite{MC_barostat_limitations}. Moreover, it does not allow to interpret the volume dynamics in a physical way, similarly to what was discussed in \cref{MC_thermostat}.
\nopagebreak
\subsection{Berendsen barostat}\label{berendsen_barostat}
The Berendsen barostat \cite{berendsen} is a deterministic pressure coupling algorithm based on the following first-order differential equation:
\begin{equation}\label{berendsen_barostat_iso}
\dot{V} = -\frac{\beta_T V}{\tau_p} \left(P_0 - P_{\text{int}}\right)\,,
\end{equation}
where $\tau_p$ and $\beta_T$, namely the relaxation time of the volume and the isothermal compressibility, are the input parameters of the barostat.
Therefore, in a single step of the algorithm (with time step $\Delta t$) the box edges and the positions' components are rescaled by a factor
\begin{equation}
\mu = \left[1 - \frac{\beta_T \Delta t}{\tau_p}\left(P_0 -P_{\text{int}} \right)\right]^{1/3} \simeq 1 - \frac{\beta_T \Delta t}{3\tau_p}\left(P_0 -P_{\text{int}} \right)\,.
\end{equation}
Although the Berendsen barostat allows to sample the correct average volume, it does not reproduce the correct volume fluctuations; in other words, the stationary distribution that it samples is not the isothermal-isobaric one. Anyway, since it does not produce instabilities when the system is initialized far from equilibrium, the Berendsen algorithm is largely employed in the first part of constant pressure simulations, i.e. in the equilibration phase.
Since $\tau_p$ only appears in $\beta_T/\tau_p$, this ratio is the only effective input parameter of the barostat, and an error in the isothermal compressibility - which may not be accurately known - only influences the accuracy of the time constant $\tau_p$ but has no consequence for the dynamics.
As a consequence, it is sufficient to use as input a rough estimate of $\beta_T$; then, after computing the correct value of the isothermal compressibility in the production run - by means of a barostat that generates the correct isothermal-isobaric ensemble - the actual value of the volume relaxation time can be calculated \emph{a posteriori} with a simple rescaling, as shown \cref{LJ_crystal}. When simulating solvated molecules, the common practice is to use as input the isothermal compressibility of the solvent.
In the flexible formulation of the algorithm - but still assuming an isotropic external stress - \cref{berendsen_barostat_iso} is generalized to
\begin{equation}\label{berendsen_aniso}
\dot{\mathbf{h}} = -\frac{\beta_T}{3\tau_p} \left(P_0 \mathbf{I} - \mathbf{P}_{\text{int}}\right)\,\mathbf{h}\,,
\end{equation}
where $\mathbf{I}$ is the 3$\times$3 identity matrix. Equivalently, at each step the box matrix and the positions are rescaled by the matrix
\begin{equation}
\bm{\mu} = \mathbf{I} - \frac{\beta_T \Delta t}{3\tau_p}\left(P_0\mathbf{I} -\mathbf{P}_{\text{int}} \right)\,,
\end{equation}
namely by applying the matrix products $\mathbf{h}' = \bm{\mu} \mathbf{h}$ and $\mathbf{q}_i' = \bm{\mu}\mathbf{q}_i$. In the anisotropic implementation, the isothermal compressibility can be in principle generalized to a 3$\times$3 tensor $\bm{\beta}_T$, although this only affects the time constants with which the various components of $\mathbf{P}_{\text{int}}$ relax to the reference value $P_0$ \cite{berendsen}. For instance, in the GROMACS \cite{gromacs} implementation of the Berendsen barostat, the anisotropic box rescaling is performed by means of a matrix $\Tilde{\mu}$ with elements
\begin{equation}
\Tilde{\mu}_{\alpha\beta} = \delta_{\alpha\beta} - \frac{\Delta t}{3\tau_p}\beta_{T,\alpha\beta}\left(P_0 \delta_{\alpha\beta} - P_{\text{int},\alpha\beta} \right)\,.
\end{equation}
\subsection{Andersen barostat}\label{andersen_barostat}
The Andersen barostat \cite{andersen} is a deterministic algorithm based on a second-order equation for the dynamics of the volume. Employing both rescaled coordinates as in \cref{rescaled_pos_iso} and rescaled momenta as in \cref{rescaled_mom_iso}, the
idea of the barostat is to treat the volume as a dynamical variable by adding to the system a new degree of freedom, which appears in the Hamiltonian as a new position-like variable $V$ and its conjugated momentum $p_V$:
\begin{equation}\label{hamiltonian_andersen_barostat}
\mathcal{H}\big(\{\mathbf{s}_i,\bm{\pi}_i\},V,p_V\big) \equiv \sum_i \frac{V^{-2/3}|\bm{\pi}_i|^2}{2 m_i} + U\big(\{V^{1/3}\mathbf{s}_i\}\big) + P_0 V +\frac{p_V^2}{2W}\,.
\end{equation}
The first two terms of the augmented Hamiltonian in \cref{hamiltonian_andersen_barostat} are the kinetic and potential energy as functions of the rescaled variables, while the third and fourth terms are respectively the potential and kinetic energy terms associated to the volume. The parameter $W$ is called \emph{barostat mass} and quantifies the inertia associated to $V$. The equations of motion derived from $\mathcal{H}$ are the following:
\begin{subequations}\label[pluralequation]{andersen_barostat_eqs}
\begin{align}
\dot{\mathbf{s}}_i &= V^{-2/3}\frac{\bm{\pi}_i}{m_i}\,,\label{1st_eq_andersen} \\
\dot{\bm{\pi}}_i &= V^{1/3}\mathbf{F}_i\,,\label{2nd_eq_andersen} \\
\dot{V} &= \frac{p_V}{W}\,\label{volume_eq_andersen} \\
\dot{p}_V &= P_{\text{int}} - P_0\,, \label{p_V_eq_andersen}
\end{align}
\end{subequations}
where $\mathbf{F}_i = -\frac{\partial U}{\partial \mathbf{q}_i}$ and $P_{\text{int}}$ is calculated as in \cref{internal_pressure}. Equivalently, applying the inverse transformations from the rescaled variables to the physical positions and momenta, \cref{1st_eq_andersen,2nd_eq_andersen} can also be written as:
\begin{subequations}\label{}
\begin{align}
\dot{\mathbf{q}}_i &= \frac{\mathbf{p}_i}{m_i} + \frac{1}{3}\frac{\dot{V}}{V}\mathbf{q}_i\,, \label{q_eq_andersen}\\
\dot{\mathbf{p}}_i &= \mathbf{F}_i - \frac{1}{3}\frac{\dot{V}}{V}\mathbf{p}_i \label{p_eq_andersen}\,.
\end{align}
\end{subequations}
Integrating these equations is equivalent to sample the distribution of an extended \emph{isoenthalpic-isobaric} ensemble, namely
\begin{equation}
\mathcal{P}_{NVH'}\Big(\{\mathbf{q}_i,\bm{p}_i\},V,p_V\Big) \propto \delta\Big(K + U + P_0 V + \frac{p_V^2}{2W} - E\Big)\,,
\end{equation}
where $H = K + U + P_0 V$ is called \emph{enthalpy} and the barostat kinetic energy $p_V^2/(2W)$ is an additional term with respect to the standard ensemble, leading to the conserved quantity
\begin{equation}\label{conserved_enthalpy}
H' = H + \frac{p_V^2}{2W}\,.
\end{equation}
However, according to the equipartition theorem the average contribution of this additional kinetic term is $\langle p_V^2/(2W)\rangle = k_B T/2$; as a consequence its effect becomes irrelevant when the number of degrees of freedom is large.
Moreover, similarly to what discussed in \cref{hybrid_MC}, also the distribution
\begin{equation}\label{distrib_andersen}
\mathcal{P}_{N P_0 T}\big(\{\mathbf{q}_i,\bm{p}_i\},V,p_V\big) \propto \exp\left[-\frac{1}{k_BT}\left(K + U + P_0 V + \frac{p_V^2}{2W} \right)\right]\,
\end{equation}
is stationary with respect to \cref{andersen_barostat_eqs}. $\mathcal{P}_{NP_0 T}$ is exactly the target $NP_0 T$ distribution because the additional term depending on $p_V$ factorizes and can be integrated out, as $p_V$ is not a variable of the isothermal-isobaric ensemble. In order to sample $\mathcal{P}_{NP_0 T}$ in a ergodic way it is possible to exploit that $\mathcal{P}_{N P_0 T}$ is also stationary with respect to any thermostat generating the correct canonical ensemble, since also the Boltzmann distribution can be recovered from \cref{distrib_andersen} - in this case by integrating out both $V$ and $p_V$, which are not dynamical variables of the $NVT$ ensemble. Thus the ergodic sampling of the $NP_0 T$ distribution can be achieved by coupling the Andersen's equations with a thermostat, acting on the particles' degrees of freedom. Furthermore, the volume itself can be coupled with an external bath by means of a Langevin thermostat (see \cref{langevin_thermo}), in order to damp the volume oscillations and accelerate the equilibration of the system. With this additional coupling, \cref{volume_eq_andersen,p_V_eq_andersen} become
\begin{subequations}
\begin{align}
\dot{V} &= \frac{p_V}{W}\,, \\
\dot{p}_V &= (P_{\text{int}} - P_0) -\gamma p_V + \sqrt{2\gamma W k_B T}\,\mathcal{R}(t)\,,
\end{align}
\end{subequations}
where $\mathcal{R}(t)$ is a white noise (see Appendix \cref{appendix_stocdiffeq}). This method is typically referred to as \emph{Langevin piston} \cite{langevin_piston}.
As all second-order algorithms, when the Andersen barostat is employed far from equilibrium the volume shows damped oscillations, which can decay slow if the barostat mass $W$ is not chosen properly. Nevertheless, setting the optimal value of $W$ is a system-dependent problem and there is not a general recipe to accomplish it. For this reason, the typical pipeline for a simulation in the isothermal-isobaric ensemble employs the Berendsen barostat for the equilibration phase, and a second-order barostat such as the Andersen one for the production run.
\subsection{Parrinello-Rahman barostat}\label{pr_barostat}
The Parrinello-Rahman method \cite{parr-rahman0,parr-rahman} extends the Andersen barostat to allow for changes in volume and shape of the system box, i.e. it generalizes the Andersen equations \cref{andersen_barostat_eqs} to the fully flexible $N\mathbf{S}T$ ensemble. This generalization includes a new definition of rescaled coordinates $\mathbf{s}_i$, namely
\begin{equation}
\mathbf{s}_i = \mathbf{h}^{-1}\mathbf{q}_i\,.
\end{equation}
Note that the matrix $\mathbf{h}$ is reasonably assumed to be invertible, since the three box vectors defining its columns are linearly independent. The augmented Hamiltonian is constructed by introducing nine new degrees of freedom, corresponding to the nine components of the box matrix $\mathbf{h}$:
\begin{equation}
\mathcal{H} = K + U + \frac{W}{2}\text{Tr}\left(\dot{\mathbf{h}}^T\dot{\mathbf{h}}\right) + P_0 \det \mathbf{h}\,.
\end{equation}
Considering as in \cite{parr-rahman0,parr-rahman} a pair-wise potential of the form
\begin{equation}
U\left(\{ \mathbf{q}_i \}\right) = \frac{1}{2} \sum_{i,j} \Phi (q_{ij})\,,
\end{equation}
where $q_{ij} = |\mathbf{q}_i-\mathbf{q}_j|$, it is possible to derive the following equations of motion:
\begin{subequations}
\begin{align}
\label{first_eq_PR}
\ddot{\mathbf{s}}_i &= -\sum_{j\neq i}\frac{1}{m_i q_{ij}}\frac{\text{d}\Phi}{\text{d}q_{ij}} \left(\bm{s}_i-\bm{s}_j \right) - \mathbf{G}^{-1}\dot{\mathbf{G}}\,\dot{\mathbf{s}}_i\,, \\
\ddot{\mathbf{h}} &= V W^{-1}\,\big(\mathbf{P} - P_0\mathbf{I}\big)\big( \mathbf{h}^{T}\big)^{-1}\,. \label{second_eq_PR}
\end{align}
\end{subequations}
where $\mathbf{G} \equiv \mathbf{h}^T\mathbf{h}$ is called \emph{metric tensor}.
\indent In case of a general anisotropic external stress, i.e. when $\mathbf{S}\neq P_0 \mathbf{I}$, the Hamiltonian is further augmented with the strain energy term defined in \cref{strain_energy}:
\begin{equation}
\mathcal{H}_S = \mathcal{H} + \frac{1}{2}\text{Tr}\left(\bm{\Sigma}\mathbf{G}\right)\,.
\end{equation}
As a consequence, \cref{first_eq_PR} remains unchanged while the dynamics of $\mathbf{h}$ is now described by
\begin{equation}
\ddot{\mathbf{h}} = V W^{-1} \big(\mathbf{P}_{\text{int}} - P_0\mathbf{I}\big)\big( \mathbf{h}^{T}\big)^{-1} - W^{-1}\mathbf{h}\bm{\Sigma}\,.
\end{equation}
\indent Even if in the original formulation of the Parrinello-Rahman barostat $W^{-1}$ is a scalar, in some implementations - such as in GROMACS \cite{gromacs} - it is treated as a 3$\times$3 symmetric tensor related to a tensorial expression of the isothermal compressibility. As already commented in \cref{berendsen_barostat}, this only affects the relaxation times of the various components of $\mathbf{P}_{\text{int}}$ towards the reference values of the external stress $\mathbf{S}$.
\subsection{Martyna-Tobias-Klein barostat}\label{mtk_barostat}
The pressure coupling method developed by Martyna, Tobias and Klein (MTK) \cite{mtk} develops a set of equations similar but not equivalent to the Andersen's ones, namely the Hoover's equations \cite{hoover,hoover2}
\begin{subequations}
\begin{align}
\dot{\mathbf{q}}_i &= \frac{\mathbf{p}_i}{m_i} + \frac{p_\varepsilon}{W}\mathbf{q}_i\,, \label{pos_eq_mtk}\\
\dot{\mathbf{p}}_i &= \mathbf{F}_i - \frac{p_\varepsilon}{W}\mathbf{p}_i\,, \label{mom_eq_mtk}\\
\dot{V} &= \frac{3Vp_\varepsilon}{W}\,, \label{vol_eq_mtk} \\
\dot{p}_\varepsilon &= 3 V (P_{\text{int}}-P_0) \label{peps_eq_mtk}\,,
\end{align}
\end{subequations}
where the variables $\varepsilon$ and $p_{\varepsilon}$ are defined as:
\begin{subequations}
\begin{align}
\varepsilon &= \frac{1}{3}\log\frac{V}{V_0}\,, \\
p_{\varepsilon} &= W \dot{\varepsilon} = \frac{W\dot{V}}{3V}\,.
\end{align}
\end{subequations}
It is easy to show that the first three equations above are equivalent to Andersen's \cref{q_eq_andersen,p_eq_andersen,volume_eq_andersen}, while the one for $p_\varepsilon$ is different from \cref{p_V_eq_andersen}. As a consequence, when coupled to a thermostat Hoover's equations actually generate a slightly different ensemble than the isothermal-isobaric one:
\begin{equation}
\mathcal{P}_{\text{Hoover}} \propto \frac{1}{V}\mathcal{P}_{NP_0 T}
\end{equation}
As an additional problem, in the extended phase space Hoover's equations do not satisfy anymore the incompressibility condition defined in \cref{incompressibility}, namely
\begin{equation}
\sum_{i,\alpha}\left(\frac{\partial \dot{q}_i^{\alpha}}{\partial q_i^{\alpha}} + \frac{\partial \dot{p}_i^{\alpha}}{\partial p_i^{\alpha}} \right) + \frac{\partial\dot{V}}{\partial V}\neq 0\,.
\end{equation}
In order to fix this problem without changing the conserved enthalpy defined in \cref{conserved_enthalpy} it is possible to modify \cref{mom_eq_mtk,peps_eq_mtk} with two corrections whose energy contributions cancel each other:
\begin{subequations}
\begin{align}
\dot{\mathbf{p}}_i &= \tilde{\mathbf{F}}_i - \left(1+\frac{3}{N_f} \right)\frac{p_\varepsilon}{W}\mathbf{p}_i\,, \label{mom_eq_corr_mtk}\\
\dot{p}_\varepsilon &= 3 V (P_{\text{int}}-P_0) + \frac{3}{N_f}\sum_i \frac{\mathbf{p}_i^2}{m_i}\,. \label{peps_eq_corr_mtk}
\end{align}
\end{subequations}
In this case, $\tilde{\mathbf{F}}_i$ is the total force acting on atom $i$, including the contribution of constraints. The full set of MTK equations is obtained by coupling these modified equations with two Nosé-Hoover chains, one for the particles and one for the volume, to keep into account that positions and momenta thermalize at a considerably faster time than the volume \cite{tuckerman}. With this additional coupling, it is possible to show that the MTK equations just defined generate the correct (isotropic) isothermal-isobaric distribution defined in \cref{npt_isotropic}.
The MTK method can be generalized to anisotropic cell fluctuations, promoting $V$ to the nine box variables in $\mathbf{h}$ and the conjugated momentum $p_\varepsilon$ to the 3$\times$3 matrix of box momenta $\mathbf{p}_g$, such that $\mathbf{p}_g / W_g = \dot{\mathbf{h}}\mathbf{h}^{-1}$. Then \cref{pos_eq_mtk,mom_eq_corr_mtk,vol_eq_mtk,peps_eq_corr_mtk} become respectively:
\begin{subequations}\label[pluralequation]{mtk_aniso}
\begin{align}
\dot{\mathbf{q}}_i &= \frac{\mathbf{p}_i}{m_i} + \frac{\mathbf{p}_g}{W_g}\mathbf{q}_i\,, \\
\dot{\mathbf{p}}_i &= \tilde{\mathbf{F}}_i - \frac{\mathbf{p}_g}{W_g}\mathbf{p}_i - \frac{1}{N_f}\frac{\text{Tr}(\mathbf{p}_g)}{W_g}\mathbf{p}_i \,, \\
\dot{\mathbf{h}} &= \frac{\mathbf{p}_g}{W_g} \mathbf{h}\,, \\
\dot{\mathbf{p}}_g &= \left(\det\mathbf{h}\right) \left(\mathbf{P}_{\text{int}}-P_0 \mathbf{I}\right) + \frac{1}{N_f}\sum_i \frac{\mathbf{p}_i^2}{m_i}\mathbf{I} \,.
\end{align}
\end{subequations}
Also in this case, the full set of MTK equations is obtained by coupling particles and cell components with two separate Nosé-Hoover chains, resulting in the generation of the correct anisotropic $N\mathbf{S}T$ ensemble defined in \cref{NPT_anisotropic}.
Since the time-reversible integration scheme for the MTK equations has been derived by Tuckerman \emph{et al.}, the algorithm is also referred to as Martyna-Tuckerman-Tobias-Klein (MTTK) barostat.
\subsection{Shinoda barostat}
The equations developed by Shinoda \emph{et al.} \cite{shinoda} combine the hydrostatic MTK \cref{mtk_aniso} with the strain energy calculated as in \cref{strain_energy} within the Parrinello-Rahman barostat. Apart from the details in the time-reversible integration scheme, the only modification to the MTK method involves the equation for the matrix of box momenta $\mathbf{p}_g$:
\begin{equation}
\dot{\mathbf{p}}_g = \left(\det\mathbf{h}\right) \left(\mathbf{P}_{\text{int}}-P_0 \mathbf{I}\right) - \mathbf{h}\bm{\Sigma}\mathbf{h}^T + \frac{1}{N_f}\sum_i \frac{\mathbf{p}_i^2}{m_i}\mathbf{I}\,,
\end{equation}
with $\bm{\Sigma}$ defined as in \cref{sigma_matrix}.
\subsection{Stochastic cell rescaling}\label{section:crescale_iso}
Stochastic cell rescaling (SCR) \cite{crescale_iso} is a first-order stochastic barostat that generates the correct isothermal-isobaric ensemble when coupled to the Hamilton's equations and to a thermostat. SCR employs a Berendsen-like deterministic part and a suitable noise term, which is responsible for the correct volume fluctuations. In its isotropic version, the stochastic equation driving the dynamics of the volume is
\begin{equation}\label{crescale_iso_V}
\text{d}V = -\frac{\beta_T V}{\tau_p}\left(P_0 - P_{\text{int}} - \frac{k_B T}{V}\right)\text{d}t + \sqrt{\frac{2k_B T \beta_T V}{\tau_p}}\text{d}W\,,
\end{equation}
where $P_{\text{int}}$ is computed with the instantaneous kinetic energy as in \cref{internal_pressure} if momenta are rescaled, or with the average kinetic energy as in \cref{internal_pressure_avgkin} if they are not. The stationarity of the $NP_0 T$ distribution in \cref{npt_isotropic} can be proved by considering the associated Fokker-Planck equation (see Appendix \ref{appendix_fp}). As in the Berendsen barostat, the time constant $\tau_p$ defines the equilibration time of the volume and its autocorrelation time in equilibrium conditions. By defining the variables $\varepsilon = \log\left(V/V_0 \right)$, where $V_0$ is a reference volume, and $\lambda = \sqrt{V}$, it is possible to derive two equivalent ways of writing \cref{crescale_iso_V} by means of the It\^{o} chain rule (see Appendix \ref{appendix_stocdiffeq}):
\begin{subequations}
\begin{align}
\text{d}\varepsilon &= -\frac{\beta_T}{\tau_p}\left(P_0 - P_{\text{int}} \right)\text{d}t + \sqrt{\frac{2 k_B T \beta_T}{V \tau_p}}\text{d}W\,, \label{eps_eq_SCR} \\
\text{d}\lambda &= -\frac{\beta_T\lambda}{2\tau_p}\left(P_0 - P_{\text{int}} - \frac{k_B T}{2 V}\right)\text{d}t + \sqrt{\frac{k_B T \beta_T}{2\tau_p}}\text{d}W\,. \label{lambda_eq_SCR}
\end{align}
\end{subequations}
In particular, \cref{lambda_eq_SCR} allows to write a time-reversible integrator for which the effective energy drift can be computed (see Appendix~\ref{appendix_effenergy}). Moreover, it is possible to derive \cref{eps_eq_SCR} as the high-friction limit of a Langevin piston algorithm (see \cref{andersen_barostat}) with a volume-dependent friction.
SCR has also been formulated in a semi-isotropic version, namely to generate the constant surface-tension ensemble $N P_0^\perp \gamma_0 T $ described by \cref{npt_constant_surface_tension}. In this case, by defining the variables $\varepsilon_{xy} = \log(A/A_0)$ and $\varepsilon_z = \log(L/L_0)$ the dynamics can be written in terms of two decoupled stochastic equations:
\begin{subequations}\label[pluralequation]{crescale-semi-isotropic-eq}
\begin{align}
\text{d}\varepsilon_{xy} &= -\frac{2\beta_T}{3\tau_p}\left(P_0^\perp - \frac{\gamma_0}{L} - \frac{P_{\text{int},xx} + P_{\text{int},yy}}{2} \right)\text{d}t + \sqrt{\frac{4 k_B T \beta_T}{3 V \tau_p}}\text{d}W_{xy}\,, \label{eps_xy}\\
\text{d}\varepsilon_z &= -\frac{\beta_T}{3\tau_p}\left(P_0^\perp - P_{\text{int},zz}\right)\text{d}t + \sqrt{\frac{2 k_B T \beta_T}{3 V \tau_p}}\text{d}W_z\,, \label{eps_z}
\end{align}
\end{subequations}
where $W_{xy}$ and $W_z$ are two distinct and independent Wiener processes. As one might expect, by summing the two equations above with $\gamma_0 = 0$ it is possible to recover the isotropic \cref{eps_eq_SCR}, with $P_0 = P_0^\perp$.
This pressure coupling method preserves all the good properties of the Berendsen barostat, namely it is efficient in the equilibration phase (as it is not affected by instabilities or oscillations if the system is far from equilibrium), it allows to easily tune the relaxation time of the volume and it is easier to implement than second-order barostats. Moreover, since SCR generates the correct isothermal-isobaric ensemble it can be used also in production runs in place of second-order algorithms, replacing the typical pipeline for constant pressure simulations with a more efficient one, where a single algorithm is employed (see \cref{fig:pipeline}).
\chapter{Fully flexible formulation of SCR}\label{chapter:formulation}
The aim of this chapter is to derive and discuss the equations for the anisotropic version of SCR, which allows to generate the correct $N\mathbf{S}T$ ensemble both in presence of a hydrostatic external pressure and in case of a generic external stress.
\section{Derivation of the equations}\label{derivation_aniso}
In order to formulate the anisotropic version of the stochastic cell rescaling method, starting from the case of a diagonal external stress $\mathbf{S} = P_0\mathbf{I}$, we look for a multidimensional It\^{o} equation (see Appendix \ref{appendix_stocdiffeq}) for the box matrix $\mathbf{h}$,
\begin{equation}\label{SCR_flex_start}
\text{d}\mathbf{h} = \mathbf{A}(\mathbf{h})\, \text{d}t + \mathbf{B}(\mathbf{h})\,\text{d}\mathbf{W} \,,
\end{equation}
such that the two following requirements are satisfied:
\begin{enumerate}[label=(\roman*)]
\item the anisotropic isothermal-isobaric distribution
\begin{equation}
\mathcal{P}_{N\mathbf{S}T}\big(\{\mathbf{q}_i,\mathbf{p}_i\},\mathbf{h}\big) \propto \left(\det\mathbf{h}\right)^{-2}\exp\left[-\frac{1}{k_B T}\Big(K + U + P_0 \det\mathbf{h}\Big)\right]
\end{equation}
should be stationary with respect to the multidimensional FP equation (see Appendix \ref{appendix_fp}) corresponding to \cref{SCR_flex_start};\label{first_requirement}
\item the deterministic part of these equations, $\text{d}\mathbf{h}^{\text{det}} = \mathbf{A}(\mathbf{h})\, \text{d}t$, should contain a Berendsen-like term as written in \cref{berendsen_aniso}.\label{second_requirement}
\end{enumerate}
Written explicitely, this multidimensional It\^{o} equation reads:
\begin{equation}\label[pluralequation]{SCR_eq_A_B}
\text{d}h_{\alpha i} = A_{\alpha i}(\mathbf{h})\, \text{d}t + \sum_{\beta j} B_{\alpha i \beta j}(\mathbf{h})\,\text{d}W_{\beta j} \,,
\end{equation}
where we remind that $\alpha = x,y,z$ and $i=1,2,3$.
Note that $\mathbf{A}$ and $\mathbf{B}$ have already been defined with no explicit time dependence in order to satisfy \ref{first_requirement}.
Instead of imposing the stationarity condition for $\mathcal{P}_{N\mathbf{S}T}$ it is easier to require the stricter condition of detailed balance, as defined in \cref{db_multidim}. As a starting point, the multidimensional FP equation corresponding to \cref{SCR_flex_start} reads:
\begin{equation}
\frac{\partial}{\partial t} \mathcal{P}(\mathbf{h},t) = -\sum_{\alpha i}\frac{\partial}{\partial h_{\alpha i}}\Big(A_{\alpha i}(\mathbf{h})\mathcal{P}(\mathbf{h},t)\Big) + \sum_{\alpha i \beta j} \frac{\partial^2}{\partial h_{\alpha i}\partial h_{\beta j}} \Big(D_{\alpha i \beta j} (\mathbf{h}) \mathcal{P}(\mathbf{h},t)\Big)\,,
\end{equation}
where $\mathbf{D} = \frac{1}{2}\mathbf{B}\mathbf{B}^T$, namely
\begin{equation}\label{D_B_relation}
D_{\alpha i \beta j} = \frac{1}{2}\sum_{\gamma k} B_{\alpha i \gamma k} B_{\beta j \gamma k}\,.
\end{equation}
Imposing the detailed balance condition for the generic distribution $\mathcal{P} = \mathcal{P}(\mathbf{h})$ means requiring that each component of the probability density tensor $\mathbf{J}$ vanishes, namely
\begin{equation}
J_{\alpha i} = A_{\alpha i}\,\mathcal{P} - \sum_{\beta j} \frac{\partial}{\partial h_{\beta j}} \Big(D_{\alpha i \beta j}\,\mathcal{P}\Big) = 0\,.
\end{equation}
By solving these equations with respect to the components of $\mathbf{A}$ one gets:
\begin{equation}
A_{\alpha i} = \sum_{\beta j} D_{\alpha i \beta j}\frac{\partial}{\partial h_{\beta j}}\Big(\log\big(D_{\alpha i \beta j}\,\mathcal{P}\big)\Big)\,.
\end{equation}
As a consequence, the condition \ref{first_requirement} is satisfied writing the target \cref{SCR_eq_A_B} as:
\begin{equation}\label[pluralequation]{target_eqs}
\text{d}h_{\alpha i} = \sum_{\beta j} D_{\alpha i \beta j} \frac{\partial}{\partial h_{\beta j}}\Big(\log\big(D_{\alpha i \beta j}\,\mathcal{P}_{N\mathbf{S}T}\big)\Big)\,\text{d}t + \sum_{\beta j} B_{\alpha i \beta j}\,\text{d}W_{\beta j}\,,
\end{equation}
\indent In order to satisfy the requirement \ref{second_requirement}, let's make the following \emph{ansatz} on the functional form of the diffusion tensor:
\begin{equation}
D_{\alpha i\beta j} = \frac{\beta_T k_B T}{3V\tau_p}\delta_{\alpha\beta}\sum_\gamma h_{\gamma i}h_{\gamma j}\,.
\end{equation}
Leaving the complete calculations to Appendix \ref{appendix_derivation}, with this choice of $\mathbf{D}$ the deterministic part of \cref{target_eqs} becomes:
\begin{equation}\label{dh_det_final}
\text{d}h_{\alpha i}^{\text{det}} = -\frac{\beta_T}{3\tau_p} \Bigg[\sum_\beta \Big( P_0\,\delta_{\alpha\beta} - P_{\text{int},\alpha\beta} \Big)h_{\beta i} - \frac{k_B T}{V}h_{\alpha i}\Bigg]\,\text{d}t\,.
\end{equation}
The expression of the internal pressure tensor appearing in $\text{d}h_{\alpha i}^{\text{det}}$ is different if the propagation of $\mathbf{h}$ is performed at both constant rescaled positions $\mathbf{s}_i$ and rescaled momenta $\bm{\pi}_i$, or instead by keeping fixed rescaled positions and physical momenta. Representing physical positions $\mathbf{q}_i$ and momenta $\mathbf{p}_i$ as column vectors, the rescaled counterparts are defined via the following relations:
\begin{subequations}\label[pluralequation]{rescaled_variables}
\begin{align}
\mathbf{q}_i &= \mathbf{h} \mathbf{s}_i \,, \label{rescaled_positions}\\
\mathbf{p}^T_i &= \bm{\pi}^T_i \mathbf{h}^{-1}\,.\label{rescaled_momenta}
\end{align}
\end{subequations}
As it is shown in Appendix \ref{appendix_derivation}, when both positions and momenta are rescaled the internal pressure tensor $\mathbf{P}_{\text{int}}$ is defined as in \cref{internal_pressure_tensor}, while in the formulation with only rescaled positions the kinetic energy tensor is replaced by an average contribution, namely
\begin{equation}\label{internal_pressure_avgkineng}
P^{\langle K\rangle}_{\text{int},\alpha\beta} = \frac{Nk_B T}{V}\delta_{\alpha\beta} + \frac{1}{V}\sum_{i=1}^N F_i^{\alpha}\,q_i^{\beta} \,.
\end{equation}
Regardless of how the internal pressure is computed, the first part of \cref{dh_det_final} is exactly the anisotropic formulation of the Berendsen barostat; thus the initial choice of the diffusion tensor $\mathbf{D}$ appears meaningful in order to satisfy \ref{second_requirement}.
The additional term containing $k_B T/V$ can be seen as a correction that becomes more and more negligible as the system size increases, since the average kinetic contribution included in $\mathbf{P}_{\text{int}}$ is the dominant term when the number of atoms $N$ is large. Without considering the stochastic part of \cref{target_eqs}, it is obvious that this correction is not sufficient to generate the correct isothermal-isobaric ensemble.
As a comment on the definition of the rescaled momenta $\bm{\pi}_i$, there are two reasons to multiply by the inverse box matrix $\mathbf{h}^{-1}$ \emph{on the right side}:
\begin{itemize}
\item this is the only way to have a consistency in the class of labels for physical and rescaled variables, which appear with latin and greek indices respectively if we write \cref{rescaled_variables} explicitely:
\begin{subequations}\label[pluralequation]{rescaled_variables_explicit}
\begin{align}
q_i^{\alpha} &= \sum_{k=1}^3 h_{\alpha k}\, s_i^k \,, \\
p_i^{\alpha} &= \sum_{k=1}^3 \pi_i^{k}\, h_{k\alpha}^{-1}\,;
\end{align}
\end{subequations}
\item defined in this way, it is possible to show that the variables $\bm{\pi}_i$ are the actual \emph{conjugated momenta} of the rescaled coordinates $\mathbf{q}_i$, or equivalently that the transformation from physical to rescaled variables is \emph{canonical}, i.e. it preserves the form of Hamilton's equations \cite{tuckerman}.
\end{itemize}
While the first motivation is necessary to be consistent with the definition of the box matrix $\mathbf{h}$, the second one is more of aesthetic nature, since the SCR method is not formulated within a Hamiltonian framework. As a comparison, in the MTK \cref{mtk_aniso} - which cannot be obtained from a Hamiltonian as well - positions and momenta are rescaled with a matrix multiplication \emph{on the same side}.
Let's now consider the stochastic part of the target equations, $\text{d}\mathbf{h}^{\text{stoc}} = \mathbf{B}(\mathbf{h})\,\text{d}\mathbf{W}$. The tensor $\mathbf{B}$ that satisfies $\mathbf{D} = \frac{1}{2}\mathbf{B}\mathbf{B}^T$ has components
\begin{equation}\label{B_tensor}
B_{\alpha i \beta j} = \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}}\,h_{\beta i}\,\delta_{\alpha j}\,;
\end{equation}
as a consequence, the stochastic term of the target \cref{target_eqs} becomes:
\begin{equation}
\text{d}h_{\alpha i}^{\text{stoc}} = \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_{\beta}h_{\beta i}\,\text{d}W_{\alpha\beta}\,.
\end{equation}
Putting together the deterministic and stochastic parts, we finally find the equations that generalize the SCR method to anisotropic cell fluctuations,
\begin{equation}\label[pluralequation]{crescale_aniso}
\text{d}h_{\alpha i} = -\frac{\beta_T}{3\tau_p} \Bigg[\sum_\beta \Big( P_0\delta_{\alpha\beta} - P_{\text{int},\alpha\beta} \Big)h_{\beta i} - \frac{k_B T}{V}h_{\alpha i}\Bigg]\,\text{d}t + \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_{\beta}h_{\beta i}\,\text{d}W_{\alpha\beta}\,,
\end{equation}
or in matrix notation:
\begin{equation}
\boxed{\text{d}\mathbf{h} = -\frac{\beta_T}{3\tau_p} \Big[ \big( P_0\mathbf{I} - \mathbf{P}_{\text{int}} \big) - \frac{k_B T}{V}\mathbf{I}\Big]\mathbf{h}\,\text{d}t + \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \text{d}\mathbf{W}\,\mathbf{h}\,.\,}
\end{equation}
It is worth observing that the diffusion tensor $\mathbf{D}$ satisfying \ref{second_requirement} is not unique; for instance, it has been found that also the choice
\begin{equation}
\widetilde{D}_{\alpha i \beta j} = \frac{\beta_T k_B T}{3V\tau_p} h_{\alpha j} h_{\beta i}
\end{equation}
brings to the same expression for the Berendsen-like deterministic part of the equations. However, this choice has been discarded as it seems not possible to find an analytic expression for $\widetilde{\mathbf{B}}$ satisfying \cref{D_B_relation}.
\subsection{Generic external stress}
In case of a generic external stress, i.e. $\mathbf{S}\neq P_0\mathbf{I}$, the target distribution in \cref{NPT_anisotropic_shear} includes an additional term that only enters in the equations in an additive way, and that is independent on the choice of rescaling momenta or not. Indeed, as it is shown in Appendix \ref{derivation_strain}, the deterministic part of the equations has to be expanded with the contribution
\begin{equation}\label{dh_strain_maintext}
\text{d}h_{\alpha i}^{\text{strain}} = -\frac{\beta_T}{3V\tau_p}\big(\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T \mathbf{h} \big)_{\alpha i}\,,
\end{equation}
where $\bm{\Sigma} = V_0\, \mathbf{h}_0^{-1}\big(\mathbf{S}-P_0\mathbf{I}\big)\left(\mathbf{h}_0^{-1}\right)^T$. Then the full equations for a generic external stress are the following:
\begin{equation}\label[pluralequation]{crescale_eqs_strain}
\boxed{\text{d}\mathbf{h} = -\frac{\beta_T}{3\tau_p} \Big[ \big( P_0\mathbf{I} - \mathbf{P}_{\text{int}} \big) - \frac{k_B T}{V}\mathbf{I} + \frac{1}{V}\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T \Big]\,\mathbf{h}\,\text{d}t + \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \text{d}\mathbf{W}\,\mathbf{h}\,.\,}
\end{equation}
\section{Properties}\label{properties}
As a first important observation, both the isotropic and semi-isotropic formulations of SCR - namely \cref{crescale_iso_V} and \cref{crescale-semi-isotropic-eq} - can be derived from the anisotropic SCR equations (see Appendix \ref{appendix_flex2iso}), by changing the propagated variables with the multidimensional It\^{o} chain rule. Hence these equations represent a self-consistent generalization of the previous formulations of the method.
A relevant feature of the anisotropic SCR \cref{crescale_eqs_strain} is that they are invariant under a redefinition of the box vectors leaving the Bravais lattice structure untouched. In MD simulations, the notion of Bravais lattice is used to periodically replicate the system in space, in order to minimize edge effects (see Appendix \ref{PBCs}). In general, if the three box vectors $\mathbf{a},\mathbf{b},\mathbf{c}$ (or $\mathbf{a}_1,\mathbf{a}_2,\mathbf{a}_3$) are thought as the primitive vectors of a Bravais lattice, the most general transformation leaving the Bravais lattice invariant is a subclass of the following mapping:
\begin{equation}\label{change_cell_vectors}
\mathbf{a}_i \longmapsto \mathbf{a}'_i = \sum_{j=1}^3 n_j^{(i)}\,\mathbf{a}_j\,
\end{equation}
where $n_j^{(i)}$ are integer numbers. The actual transformation is only a subclass of \cref{change_cell_vectors} because additional constraints should be imposed on the three integers, in order to obtain three vectors $\{\mathbf{a}'_i\}$ that are still linearly independent and that generate a cell with the same volume of the original one. Recalling that $\mathbf{a}_1,\mathbf{a}_2,\mathbf{a}_3$ define the columns of the box matrix $\mathbf{h}$, the transformation above can be also written as:
\begin{equation}
h_{\alpha i} \longmapsto h'_{\alpha i} = \sum_{j=1}^3 n_j^{(i)}\,h_{\alpha j}\,.
\end{equation}
By employing once again the multidimensional It\^{o} chain rule, it is possible to show that the anisotropic SCR equations assume the same form when written in terms of the transformed box matrix $\mathbf{h}'$ (see Appendix \ref{appendix_changecellvectors}).
\section{SCR as limit case of Parrinello-Rahman equations}\label{limit_PR}
As already shown in the isotropic case \cite{crescale_iso}, also the anisotropic SCR equations can be derived as the high-friction limit of a second-order barostat, namely the Parrinello-Rahman equations coupled to a Langevin thermostat with a variable-dependent friction tensor $\bm{\gamma} = \bm{\gamma}(\mathbf{h})$. Writing the second-order Parrinello-Rahman \cref{second_eq_PR} as two first-order equations and adding both friction and noise terms, the equations of interest are
\begin{subequations}\label[pluralequation]{PR_friction_eqs}
\begin{align}
\dot{\mathbf{h}} &= \mathbf{v}\,, \\
\dot{\mathbf{v}} &= \frac{V}{W}\big(\mathbf{P}_{\text{int}} - P_0\mathbf{I}\big)\big(\mathbf{h}^{-1}\big)^T - \frac{1}{W}\bm{\gamma}\,\mathbf{v} + \frac{1}{W}\bm{\sigma}\,\bm{\eta}\,, \label{2nd_eq_PR}
\end{align}
\end{subequations}
where $\bm{\eta}$ is a tensor of independent white noise processes and $\bm{\sigma}$ is a tensor satisfying the multidimensional \emph{fluctuation-dissipation theorem} \cite{high_friction_limit}:
\begin{equation}\label{fluct_diss_multidim}
\bm{\sigma}\bm{\sigma}^T = 2k_B T \bm{\gamma}\,.
\end{equation}
As shown in Appendix \ref{appendix_highfriction}, with a suitable choice of the tensor $\bm{\gamma}(\mathbf{h})$,
\begin{equation}
\gamma_{\alpha i\beta j} = \frac{3V\tau_p}{\beta_T}\delta_{\alpha\beta}\sum_{\eta}h_{i\eta}^{-1}\,h_{j\eta}^{-1}\,,
\end{equation}
the equations above bring to the anisotropic SCR \cref{crescale_aniso} in the high-friction limit described in \cite{high_friction_limit}, apart for a small correction in the deterministic part:
\begin{equation}
\text{d}\mathbf{h}^{PR} = \frac{\beta_T}{3\tau_p}\frac{2 k_B T}{V}\mathbf{I}\,\text{d}t\,.
\end{equation}
This additional term comes from a slight difference in the $N\mathbf{S}T$ distribution sampled by the Parrinello-Rahman method, namely the absence of the factor $(\det\mathbf{h})^{-2}$ in the target distribution defined in \cref{NPT_anisotropic}. Anyway this term is negligible if the system includes a large number of atoms $N$, as in this case the main contribution in the deterministic part of the SCR equations comes from the internal pressure tensor. Indeed the average contribution of the kinetic energy included in $\mathbf{P}_{\text{int}}$ scales linearly in $N$, as it is clear from \cref{internal_pressure_tensor}.
\Cref{PR_friction_eqs} hold in case of a isotropic external stress ($\mathbf{S} = P_0\mathbf{I}$). The inclusion of a generic stress, namely of the additional term $\frac{1}{W}\bm{\Sigma}\mathbf{h}$ in the (RHS) of \cref{2nd_eq_PR}, brings to the general anisotropic SCR \cref{crescale_eqs_strain} in the same high-friction limit considered before. Also this generalization is better discussed in Appendix \ref{appendix_highfriction}.
\section{Euler integrator}\label{euler_integrator}
The simplest way to integrate \cref{crescale_eqs_strain} is to use the Euler method, which is a simple finite time step approximation:
\begin{equation}\label[pluralequation]{Delta_h}
\Delta\mathbf{h} = -\frac{\beta_T}{3\tau_p} \Big[ \big( P_0\mathbf{I} - \mathbf{P}_{\text{int}} \big) - \frac{k_B T}{V}\mathbf{I} + \frac{1}{V}\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T \Big]\,\mathbf{h}\,\Delta t + \sqrt{\frac{2\beta_T k_B T\Delta t}{3V\tau_p}} \mathbfcal{R}\,\mathbf{h}
\end{equation}
Here $\mathbfcal{R}$ is a 3$\times$3 matrix of i.i.d. standard Gaussian numbers. The propagation $\mathbf{h}\mapsto\mathbf{h}+\Delta\mathbf{h}$ is equivalent to the rescaling $\mathbf{h}\mapsto\bm{\mu}\mathbf{h}$, which gives the name to the algorithm, where the rescaling matrix is
\begin{equation}\label{mu_matrix}
\bm{\mu} = \mathbf{I} -\frac{\beta_T}{3\tau_p} \Big[ \big( P_0\mathbf{I} - \mathbf{P}_{\text{int}} \big) - \frac{k_B T}{V}\mathbf{I} + \frac{1}{V}\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T \Big]\,\Delta t + \sqrt{\frac{2\beta_T k_B T\Delta t}{3V\tau_p}} \mathbfcal{R}\,.
\end{equation}
The same matrix is also employed to rescale positions and momenta, according to the formulation chosen:
\begin{subequations}\label[pluralequation]{pos_mom_rescaling}
\begin{align}
\mathbf{q}_i &\longmapsto \bm{\mu}\,\mathbf{q}_i\,, \\
\mathbf{p}_i^T &\longmapsto \mathbf{p}_i^T\bm{\mu}^{-1}\,.
\end{align}
\end{subequations}
The rescaling can be performed at each MD step or every $n_s$ steps in a multiple-time-step fashion \cite{multiple_timestep}, in order to speed up the simulation.
Even if the continuous \cref{crescale_aniso} satisfy detailed balance, this condition is violated when they are integrated with a finite time step algorithm. To quantify this violation and find out if the time step and the other parameters were chosen correctly, in principle it is possible to compute a quantity called \emph{effective energy drift}, which can be interpreted as the work performed by the integration algorithm on the system (see Appendix \ref{appendix_effenergy}). However, as already observed in the isotropic case \cite{crescale_iso}, the effective energy drift has a "bad scaling" with the time step when - as in \cref{Delta_h} - the noise prefactor is variable-dependent. In other words, in such a situation variations of the effective energy appear not much sensitive to variations of the time step, and this makes the effective energy an unsuitable quantity to evaluate the quality of the integration. For this reason, no effective energy is computed for the simple Euler integrator in \cref{Delta_h}. In order to explain this behaviour in presence of a variable-dependent noise prefactor, it is possible to observe from \cref{effective_energy} that the effective energy is computed as the ratio between the probability of generating the forward move and the probability of generating the backward one, and these transition probabilities are dominated by the stochastic term when the time step is small, or equivalently when $\tau_p$ is large. Indeed, in this limit it is clear from the $\Delta t/\tau_p$-dependencies in \cref{Delta_h} that the deterministic contribution goes to zero faster than the stochastic one. Neglecting the deterministic term, the integration appears perfectly time-reversible if we suppose that the noise prefactor is constant; therefore this condition is expected to maximize the probability of the backward move at fixed time step.
In the isotropic case, it is easy to find a change of variable - namely $\lambda = \sqrt{V}$ - that brings to a formulation with a constant noise prefactor, as shown in \cref{lambda_eq_SCR}. In the anisotropic case, instead, such a transformation appears not feasible, since the variable dependence in the stochastic term is notably complicated by a matrix product involving all the nine cell components. A possibility that has been taken into account but finally discarded is to perform a transformation such that one of the propagated variables is exactly $\lambda$, while the other eight variables are propagated according to equations with a $\lambda$-dependent noise prefactor, which is "symmetrized" with a geometric mean between consecutive steps in order to enhance the time-reversibility of the generated trajectory. More information about the attempt of constructing this integrator and its limitations are reported in Appendix \ref{appendix_epsvariables}.
\section{Time-reversible integrator}\label{sec:TR}
The Euler integrator of the previous section does not allow to use the effective energy drift (see Appendix \cref{appendix_effenergy}) to efficiently monitor the quality of the integration, since the volume dependence in the noise prefactor makes the box matrix update $\mathbf{h}\mapsto\mathbf{h}+\Delta\mathbf{h}$ non-reversible even when $\Delta t$ is small enough to allow neglecting the deterministic part of the move. In order to derive a time-reversible integration scheme in the limit of small $\Delta t$, let's first rewrite the anisotropic SCR \cref{crescale_eqs_strain} as
\begin{equation}
\text{d}\mathbf{h} = \left(\mathbf{A}+\frac{b^2}{2} \right)\,\mathbf{h}\,\text{d}t + b\,\text{d}\mathbf{W}\,\mathbf{h}\,,
\end{equation}
where $\mathbf{A}$ and $b$ are
\begin{subequations}
\begin{align}
\mathbf{A} &= -\frac{\beta_T}{3\tau_p}\Big(P_0\mathbf{I}-\mathbf{P}_{\text{int}} + \frac{1}{V}\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T\Big)\,, \\
b &= \sqrt{\frac{2\beta_T k_B T\Delta t}{3V\tau_p}}\,.
\end{align}
\end{subequations}
Note that $\mathbf{A}$ depends on the full box matrix $\mathbf{h}$, while $b$ only depends on its determinant, which is the volume of the system.
In terms of these quantities, the Euler integrator reads:
\begin{equation}
\mathbf{h}^{t+\Delta t} = \left(\mathbf{A}^t+\frac{\left(b^t\right)^2}{2} \right)\,\mathbf{h}^t\Delta t + b^t\Delta\mathbf{W}^t\,\mathbf{h}^t\,,
\end{equation}
where $\Delta\mathbf{W}^t = \sqrt{\Delta t}\, \bm{\mathcal{R}}^t$. Superscripts are referred to the MD time at which each quantity is computed. This expression can be seen as the first order approximation of
\begin{equation}\label{eq:expmatrix}
\mathbf{h}^{t+\Delta t} = \exp\big(\mathbf{A}^t\Delta t + b^t\Delta\mathbf{W}^t\big)\,\mathbf{h}^t\,,
\end{equation}
where the \emph{matrix exponential} of a generic matrix $\mathbf{M}$ is defined via the power series
\begin{equation}\label{eq:exp_matrix_def}
\exp\big(\mathbf{M}\big) = \sum_{k=0}^{\infty}\frac{1}{k!}M^k\,.
\end{equation}
Let's now decompose $\mathbf{A}$ and $\Delta \mathbf{W}$ as
\begin{subequations}
\begin{align}
\mathbf{A} &= a_1\mathbf{I} + \mathbf{A}_2\,,\\
\Delta\mathbf{W} &= \Delta W_1\mathbf{I} + \Delta \mathbf{W}_2\,,
\end{align}
\end{subequations}
where the scalars $a_1$ and $\Delta W_1$ are the averages of the diagonal elements of the matrices $\mathbf{A}$ and $\Delta \mathbf{W}$ respectively:
\begin{subequations}
\begin{align}
a_1 &= \frac{\text{Tr}\,\mathbf{A}}{3}\,, \\
\Delta W_1 &= \frac{\text{Tr}\,\Delta\mathbf{W}}{3}\,.
\end{align}
\end{subequations}
As a consequence, $\mathbf{A}_2$ and $\Delta \mathbf{W}_2$ are by construction traceless matrices. With this decomposition, \cref{eq:expmatrix} becomes:
\begin{subequations}
\begin{align}
\mathbf{h}^{t+\Delta t} &= \exp\big(\left(a_{1}^t\mathbf{I} + \mathbf{A}_2^t\right)\Delta t + b^t\,\left(\Delta W_1^t\mathbf{I} + \Delta \mathbf{W}_2^t\right)\big)\,\mathbf{h}^t \\
&= \exp\big(\left(a_1^t\Delta t + b^t\,\Delta W_1^t\right)\,\mathbf{I}+ \left(\mathbf{A}_2^t\,\Delta t + b^t\Delta \mathbf{W}_2^t\right)\big)\,\mathbf{h}^t \\
&= \exp\big(a_1^t\Delta t + b^t\,\Delta W_1^t\big)\,\exp\big(\mathbf{A}_2^t\,\Delta t + b^t\Delta \mathbf{W}_2^t\big)\,\mathbf{h}^t\,, \label{eq:split_volume_shape}
\end{align}
\end{subequations}
where in the last passage holds with no approximations, since the identity matrix $\mathbf{I}$ commutes with any other matrix and $\exp\big(c\mathbf{I}\big) = \exp(c)\, \mathbf{I}$ for any scalar $c$. Note that the matrix exponential in \cref{eq:split_volume_shape} has unitary determinant, as the matrix in the argument is traceless and the determinant of a matrix exponential is given by the well-known relation
\begin{equation}
\det\left[\exp\big(\mathbf{M}\big)\right] = e^{\text{Tr}\,\mathbf{M}}\,.
\end{equation}
As a consequence, \cref{eq:split_volume_shape} allows to separate in the overall rescaling two different contributions, the former related to the change of volume and given by the first (scalar) exponential $\exp\big(a_1^t\Delta t + b^t\,\Delta W_1^t\big)$, and the latter connected to the change of shape and given by the second (matrix) exponential $\bm{\mu}_s =\exp\big(\mathbf{A}_2^t\,\Delta t + b^t\Delta \mathbf{W}_2^t\big)$, which leaves the determinant of $\mathbf{h}_t$ untouched. In other words, the first operation propagates the isotropic degree of freedom, namely $V=\text{det}\mathbf{h}$, while the second rescaling evolves the remaining eight degrees of freedom, which are responsible for anisotropic box fluctuations and for global rotations. To obtain a time-reversible move, we can write the box matrix update with the following Trotterization:
\begin{equation}\label{eq:trotterization}
\mathbf{h}^{t+\Delta t} = \exp\left(a_1^t\frac{\Delta t}{2} + b^t\frac{\Delta W_1^t}{2}\right)\exp\left(\mathbf{A}_2^t\Delta t + b^{t+\frac{\Delta t}{2}}\Delta \mathbf{W}_2^t\right)\exp\left(a_1^t\frac{\Delta t}{2} + b^t\frac{\Delta W_1^t}{2}\right)\mathbf{h}^t
\end{equation}
The only difference with respect to \cref{eq:split_volume_shape} is that in the rescaling involving the matrix exponential, responsible for the change of shape of the box, the noise prefactor $b$ is computed after propagating the volume for half time step. Although not strictly necessary, this modification is expected to increase the reversibility of the integrator, since the noise term $\Delta \mathbf{W}_2^t$ will be scaled with the same prefactor in the forward and backward trajectories.
A scheme of the box matrix propagation according to \cref{eq:trotterization} is reported in \cref{alg:trotterized_integrator}.\\
\begin{algorithm}[H]
\SetAlgoLined
propagate volume for $\Delta t/2$\;
propagate box matrix shape for $\Delta t$\;
propagate volume for $\Delta t/2$\;
\caption{Time-reversible integrator for anisotropic SCR equations}
\label{alg:trotterized_integrator}
\end{algorithm}
\vspace{0.4cm}
The rescaling matrix $\bm{\mu}$ containing the information on both volume and shape update, used to rescale positions and momenta as shown in \cref{pos_mom_rescaling}, is computed after step 3 as
\begin{equation}
\bm{\mu} = \left(\frac{V^{t+\Delta t}}{V^t} \right)^{1/3}\bm{\mu}_s\,,
\end{equation}
where $\bm{\mu}_s$ is evaluated at step 2. The matrix exponential $\bm{\mu}_s$ is computed by means of a \emph{Padé approximation} \cite{pade_approximation} that reproduces \cref{eq:exp_matrix_def} up to the sixth-order. Calling $\Delta\bm{\varepsilon} = \mathbf{A}_2^t\,\Delta t + b^t\Delta \mathbf{W}_2^t$, so that $\bm{\mu}_s =\exp\big(\Delta\bm{\varepsilon}\big)$, this approximation reads:
\begin{equation}
\bm{\mu}_s \simeq \Big(\mathbf{I}-\frac{1}{2}\Delta\bm{\varepsilon}+\frac{1}{10}\Delta\bm{\varepsilon}^2 - \frac{1}{120}\Delta\bm{\varepsilon}^3\Big)^{-1}\Big(\mathbf{I}+\frac{1}{2}\Delta\bm{\varepsilon}+\frac{1}{10}\Delta\bm{\varepsilon}^2 + \frac{1}{120}\Delta\bm{\varepsilon}^3\Big)\,.
\end{equation}
The isotropic moves at step 1 and 3 can be performed by propagating the variable $\lambda = \sqrt{V}$ instead of $V$, as this choice brings to a volume-independent noise prefactor and, as a consequence, to a "well behaved" effective energy (see \cref{section:crescale_iso}).
In principle, the random numbers employed for the isotropic propagations at step 1 and 3 depend on the "diagonal" random numbers extracted at step 2. Actually, it is easy to show that, given three i.i.d. Gaussian numbers $\mathcal{R}_{xx},\mathcal{R}_{yy},\mathcal{R}_{zz}$,
\begin{equation}
\Big\langle \mathcal{R}_{\alpha\alpha}-\frac{\mathcal{R}_{xx}+\mathcal{R}_{yy}+\mathcal{R}_{zz}}{3}\Big\rangle = 0
\end{equation}
for any $\alpha=x,y,z$; as a consequence, the random numbers at step 1 and 3 can be extracted independently from the ones at step 2, with the only request that $\text{Tr}\,\Delta\mathbf{W}_2 = 0$.
Note that steps 1+3 are equivalent to propagate $V$ for a full time step, and the split is only employed to compute the noise prefactor $b$ in an "intermediate" time at step 2. In other words, the update of $\lambda$ given by steps 1+3 is equivalent to
\begin{equation}\label{eq:lambda_equivalent_1+3}
\lambda \longmapsto \lambda
-\frac{\beta_T\lambda}{2\tau_P}\left(P_0-P_{\text{int}}+
\frac{\text{Tr}(\mathbf{h}\mathbf{\Sigma}\mathbf{h}^T )}{3\lambda^2}-\frac{k_BT}{2\lambda^2}\right)\Delta t +\sqrt{\frac{k_BT\beta_T\Delta t}{2\tau_P}}\mathcal{R}\,,
\end{equation}
which is the Euler propagation of the isotropic \cref{lambda_eq_SCR} for a time $\Delta t$, except for the additional $\bm{\Sigma}$-dependent term, that is related to a possible deviatoric stress.
As a consequence, the effective energy drift resulting from steps 1+3 is the same that one gets in the isotropic formulation of the barostat \cite{crescale_iso}, where $\lambda$ is propagated for a full time step with no further splitting:
\begin{align}\label{eq:eff_eng_iso}
\Delta \widetilde{H}_{1+3} =& \,\Delta K_{1+3} + \Delta U_{1+3} + \Delta E_{s,1+3} + P_0\Delta\lambda^2 - k_B T\Delta\log\lambda \nonumber\\
&+ \Delta\lambda\left(\frac{f(\lambda^t)+f(\lambda^{t+\Delta t})}{2}\right) + \frac{\beta_T\Delta t}{16\tau_p}\Delta f^2,\,
\end{align}
where $f(\lambda) = -2\lambda\left(P_0-P_{\text{int}}-\frac{k_B T}{2\lambda^2} + \frac{\text{Tr}\left(\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T\right)}{3\lambda^2}\right)$. Here $E_s$ is the strain energy defined in \cref{strain_energy}, and the terms $\Delta K_{1+3},\, \Delta U_{1+3}\,\text{and}\,\Delta E_{s,1+3}$ accumulate the energy increments due to the volume rescaling half-steps. In this case, the differences with respect to the effective energy drift reported in \cite{crescale_iso} are the presence of a term accounting for a possible anisotropic external stress in $f(\lambda)$ and the additional strain energy contribution.
The contribution of the anisotropic step 2 is instead given by
\begin{align}\label{eq:eff_eng_aniso}
\Delta \widetilde{H}_{2} = \Delta K_{2} + \Delta U_{2} + \Delta E_{s,2} + \frac{k_B T}{2}\sum_{\alpha\beta}\left[\frac{\Delta t}{(b^t)^2}\Delta A_{2,\alpha\beta}^2 + \frac{2}{(b^{t})^2}\Delta\varepsilon_{\alpha\beta}\left( A_{2,\alpha\beta}^t + A_{2,\alpha\beta}^{t+\Delta t}\right) \right]\,,
\end{align}
where the increments $\Delta K_{2},\, \Delta U_{2}\,\text{and}\,\Delta E_{s,2}$ account for energy differences due to the change of shape.
For the full derivations of \cref{eq:eff_eng_iso} and \cref{eq:eff_eng_aniso}, see Appendix \ref{appendix:iso_effeng} and Appendix \ref{appendix:aniso_effeng}.
\section{Elimination of box rotations}\label{eliminate_rotations}
As mentioned in \cref{npt_theory}, three among the nine degrees of freedom in $\mathbf{h}$ only account for the global orientation of the box and their evolution describes overall rotations of the system, which are typically not of interest in MD simulations. At least two different strategies can be employed to eliminate box rotations.
The first possibility, as suggested by Martyna \emph{et al.} \cite{mtk}, is to rescale the box matrix with a symmetric tensor, so that no torque is applied to the cell causing it to rotate. Note that a 3$\times$3 symmetric matrix only has six independent elements, coherently with the remaining degrees of freedom after eliminating three of them. In the specific case of the anisotropic MTK barostat (see \cref{mtk_barostat}), this is accomplished by symmetrizing the internal pressure tensor, which is the only source of possible asymmetries in the rescaling of the box matrix:
\begin{equation}
P_{\text{int},\alpha\beta} \longmapsto \frac{P_{\text{int},\alpha\beta} + P_{\text{int},\beta\alpha}}{2}\,.
\end{equation}
In our case, symmetrizing only the internal pressure tensor would be not sufficient, since by chance also the stochastic term in \cref{crescale_aniso} could be responsible for global rotations. Hence, a possibility is to symmetrize both $\mathbf{P}_{\text{int}}$ and the tensor $\text{d}\mathbf{W}$ containing nine independent Wiener processes:
\begin{equation}
W_{\alpha\beta} \longmapsto \frac{W_{\alpha\beta} + W_{\beta\alpha}}{2}\,.
\end{equation}
Equivalently, instead of performing two symmetrization it is possible to directly symmetrize the rescaling matrix $\mathbf{\mu}$ defined in \cref{mu_matrix}. Note that with this method the three redundant degrees of freedom are integrated out by means of the symmetrization procedure, but the number of propagated variables is still nine. As a final observation, using this approach the box will be in general \emph{triclinic}, (see Appendix \ref{PBCs}) since no constraints are imposed on its shape.
The second possibility to eliminate rotations is to constrain $\mathbf{h}$ to be upper-triangular, so that the box vector $\mathbf{a}$ is always oriented along the $x$-axis, $\mathbf{b}$ lies in the $xy$-plane and only $\mathbf{c}$ is free to evolve in all its three components. It is important to underline that this constraint does \emph{not} imply any limitation on the shape of the box, since any triclinic box can be represented by an upper-triangular matrix $\mathbf{h}$ with a suitable choice of the chartesian frame. Hence, also this method allows to represent the box in the most general way. Since some of the most popular software for MD simulations employ such a representation for $\mathbf{h}$, this is also the method used for the anisotropic SCR algorithm, in all the implementations discussed in \cref{implementations}. This constraint can be imposed either by evolving only the six degrees of freedom corresponding to the upper-triangular part of $\mathbf{h}$ (as for instance in the implementation of the MTK barostat in LAMMPS \cite{lammps}), or by evolving all the box matrix components and imposing the constraint afterwards (as for the Parrinello-Rahman and the anisotropic Berendsen barostats in GROMACS \cite{gromacs}). In this second implementation, at each step the matrix $\mathbf{h}' = \bm{\mu}\mathbf{h}$ obtained by evolving all the nine components,
\begin{equation}
\mathbf{h'}=
\begin{pmatrix}
a'_x & b'_x & c'_x\\
a'_y & b'_y & c'_y\\
a'_z & b'_z & c'_z
\end{pmatrix}\,,
\end{equation}
has to be "rotated back" in order to eliminate the rotations acquired and restore the upper-triangular shape (see \cref{fig:box_rot_elimination}):
\begin{equation}\label{backrotate}
\mathbf{h''}= \mathbf{R}\mathbf{h'} =
\begin{pmatrix}
a'_x & b'_x & c'_x\\
0 & b'_y & c'_y\\
0 & 0 & c'_z
\end{pmatrix}\,.
\end{equation}
$\mathbf{R}$ is the rotation matrix the accounts for this operations, and it is applied on the left since it has to act on the columns of $\mathbf{h'}$, namely on the box vectors $\mathbf{a'},\mathbf{b'},\mathbf{c'}$. Note that \cref{backrotate} can be written as
\begin{equation}
\mathbf{h''}= \mathbf{R}\mathbf{h'} = (\mathbf{R}\bm{\mu})\mathbf{h}\,;
\end{equation}
as a consequence, rotating back $\mathbf{h'}$ is equivalent to rescale the initial box matrix $\mathbf{h}$ with the rotated rescaling matrix $\bm{\mu'} = \mathbf{R}\bm{\mu}$, using the same rotation matrix $\mathbf{R}$ that makes $\mathbf{h''}$ upper-triangular. Clearly, this rotation also makes $\bm{\mu'}$ upper-triangular. The problem of determining $\bm{\mu'}$ knowing $\bm{\mu}$ is part of the so-called $\emph{QR factorization}$. In all the implementations of the anisotropic SCR algorithm, this problem is accomplished by considering a simple method equivalent to the Gram-Schmidt procedure, explained in Appendix \ref{appendix_rotations}.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.9\textwidth]{core/third_chapter/box_rot_eliminations.PNG}
\caption{Graphical representation of the method applied to eliminate box rotations ($\mathbf{h}'\mapsto\mathbf{h}''$), after propagating all the nine degrees of freedom according to the anisotropic SCR \cref{crescale_aniso} ($\mathbf{h}\mapsto\mathbf{h}'$).}
\label{fig:box_rot_elimination}
\end{figure}
\section{Equations with tensorial isothermal compressibility}
The anisotropic SCR \cref{crescale_aniso} has been derived taking as a reference the Berendsen \cref{berendsen_aniso} for the deterministic part. In the derivation, the isothermal compressibility has been always considered as a scalar quantity, but the Berendsen algorithm also allows for a formulation where $\bm{\beta}_T$ is a 3$\times$3 tensor, which is multiplied by the pressure tensor in a element-wise manner (see \cref{berendsen_barostat}):
\begin{equation}\label{berendsen_betat_tensor}
\text{d}h_{\alpha i}^{\text{Ber}} = -\frac{1}{3\tau_p} \sum_\gamma \beta_{T,\alpha\gamma}\Big( P_0\,\delta_{\alpha\gamma} - P_{\text{int},\alpha\gamma} \Big)h_{\gamma i}\,\text{d}t\,.
\end{equation}
Moreover, the GROMACS software \cite{gromacs} employs exactly this formulation for the anisotropic Berendsen barostat, by considering in its input six independent components of $\bm{\beta}_T$, which is always constructed as a symmetric matrix. Therefore, even if such a formulation should only affect the relaxation times for the different components of $\mathbf{h}$ and $\mathbf{P}_{\text{int}}$, it is worth giving a theoretical formulation of the SCR method with a tensorial expression of $\bm{\beta}_T$, which is treated on the model of \cref{berendsen_betat_tensor}.
The derivation is completely equivalent to the one described in \cref{derivation_aniso}, but starting from a different \emph{ansatz} for the diffusion tensor,
\begin{equation}
D_{\alpha i\beta j}' = \frac{k_B T}{3V\tau_p}\delta_{\alpha\beta}\sum_\gamma \beta_{T,\alpha\gamma}\,h_{\gamma i}\,h_{\gamma j}\,,
\end{equation}
and as a consequence employing a different tensor $\mathbf{B}'$ for the stochastic term:
\begin{equation}
B'_{\alpha i\beta j} = \sqrt{\frac{2\beta_{T,\alpha\beta}k_B T}{3V\tau_p}}h_{\beta i}\,\delta_{\alpha j}\,.
\end{equation}
With calculations similar to those reported in Appendix \ref{appendix_derivation}, the equations that one obtains with these two tensors are the following:
\begin{align}
\text{d}h_{\alpha i} = -&\frac{1}{3\tau_p} \Bigg[\sum_\gamma \beta_{T,\alpha\gamma}\Big( P_0\delta_{\alpha\gamma} - P_{\text{int},\alpha\gamma} \Big)h_{\gamma i} - \beta_{T,\alpha\alpha}\frac{k_B T}{V}h_{\alpha i}\Bigg]\,\text{d}t \nonumber\\
&+ \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_{\gamma}\sqrt{\beta_{T,\alpha\gamma}}\,h_{\gamma i}\,\text{d}W_{\alpha\gamma}\,,
\end{align}
Including the correction for a generic external stress, the equations become:
\begin{align}
\text{d}h_{\alpha i} = -&\frac{1}{3\tau_p} \Bigg[\sum_\gamma \beta_{T,\alpha\gamma}\Big( P_0\delta_{\alpha\gamma} - P_{\text{int},\alpha\gamma} \Big)h_{\gamma i} - \beta_{T,\alpha\alpha}\frac{k_B T}{V}h_{\alpha i} + \frac{1}{V}\big(\bm{\beta}_T\odot\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T \mathbf{h} \big)_{\alpha i}\Bigg]\,\text{d}t \nonumber\\
&+ \sqrt{\frac{2 k_B T}{3V\tau_p}} \sum_{\gamma}\sqrt{\beta_{T,\alpha\gamma}}\,h_{\gamma i}\,\text{d}W_{\alpha\gamma}\,,
\end{align}
where the notation $\mathbf{A}\odot\mathbf{B}$ is used for the element-wise (or \emph{Hadamard}) product, namely
\begin{equation}
\big(\mathbf{A}\odot\mathbf{B}\big)_{ij} = A_{ij}\,B_{ij}\,.
\end{equation}
\chapter{Implementations and tests}\label{implementations}
The anisotropic SCR method developed in \cref{chapter:formulation} was tested with three different implementations, built by modifying the MD software SimpleMD, GROMACS 2021.2 \cite{gromacs} and LAMMPS (2 July 2021 release) \cite{lammps}. All the implementations employ the Euler integrator described in \cref{euler_integrator} and eliminate box rotations by constraining $\mathbf{h}$ to be a triclinic upper triangular matrix, first by propagating all the nine box components, and then by rotating back the cell vectors using the procedure described in Appendix \ref{appendix_rotations}.
The SimpleMD program is an educational code to perform MD simulations of a Lennard-Jones (LJ) system, namely using a simple pair potential of the form
\begin{equation}
V_{LJ}(q_{ij}) = 4\varepsilon\left[\left(\frac{\sigma}{q_{ij}}\right)^{12} - \left(\frac{\sigma}{q_{ij}}\right)^{6} \right]\,,
\end{equation}
where $q_{ij} = |\mathbf{q}_i - \mathbf{q}_j|$.
The modified code includes both the isotropic and the anisotropic implementations of the SCR algorithm. The available temperature coupling methods are the Langevin and the SVR thermostats, discussed in \cref{langevin_thermo,SVR} respectively. The scheme applied to propagate the SCR equations together with Hamilton's equations and the selected thermostat is the one referred to as \emph{Trotter-based integrator} in \cite{crescale_iso}:\\
\begin{algorithm}[H]\label{trotter_based_integrator}
\SetAlgoLined
apply thermostat for $\Delta t/2$\;
propagate momenta for $\Delta t/2$: $\mathbf{p}_i \gets \mathbf{p}_i + \mathbf{F}_i\Delta t/2$\;
\label{modulo}
\If{$\big(\text{n}_{\text{MD}}\text{ \% n}_s =0\big)$} {
apply SCR barostat to compute $\bm{\mu}$\; \label{line3}
rescale box matrix: $\mathbf{h} \gets \bm{\mu}\mathbf{h}$\;
rescale and propagate positions with (rescaled) momenta: \cref{rescale_propagate_positions}\; \label{line6}
(rescale momenta: $\mathbf{p}_i^T \gets \mathbf{p}_i^T\bm{\mu}^{-1}$)\; \label{line7}
}
recompute forces\;
propagate momenta for $\Delta t/2$: $\mathbf{p}_i \gets \mathbf{p}_i + \mathbf{F}_i\Delta t/2$\;
apply thermostat for $\Delta t/2$\;
\caption{Trotter-based integrator implemented in SimpleMD}
\end{algorithm}
\vspace{0.4cm}
\noindent The symbol \% at \cref*{modulo} stands for the modulo operation, meaning that the barostat is applied once every $n_s$ steps, and the brackets at \cref*{line6,line7} refer to the formulation where both positions and momenta are rescaled. Step \ref*{line6} is performed with a further Trotter splitting, where positions are first propagated with momenta for half time step, then rescaled and finally propagated again for half time step. Depending on the formulation chosen, this splitting results in:
\begin{subequations}\label[pluralequation]{rescale_propagate_positions}
\begin{align}
\mathbf{q}_i &\gets \bm{\mu}\,\mathbf{q}_i + \frac{\bm{\mu}\mathbf{p}_i}{2m_i} \Delta t + \frac{\mathbf{p}_i^T\bm{\mu}^{-1}}{2m_i} \Delta t \hspace{0.4cm}\text{(if momenta are rescaled)}\,, \\
\mathbf{q}_i &\gets \bm{\mu}\,\mathbf{q}_i + \left(\bm{\mu}+\mathbf{I}\right)\frac{\mathbf{p}_i}{2m_i} \Delta t \hspace{0.54cm}\text{(if momenta are \emph{not} rescaled)}\,.
\end{align}
\end{subequations}
The SimpleMD implementation allows to apply the anisotropic SCR barostat both with and without cell rotations, and also includes the time-reversible integration scheme outlined in \cref{sec:TR}.
A comparison between the main features of the three implementations of the anisotropic SCR algorithm is reported in \cref{tab:implementations}.
\begin{table}[h!]
\centering
\caption{Comparison between the SimpleMD\footnote{\href{https://github.com/bussilab/crescale.git}{https://github.com/bussilab/crescale.git}}, GROMACS\footnote{\href{https://github.com/bussilab/crescale-gromacs.git}{https://github.com/bussilab/crescale-gromacs.git}} and LAMMPS\footnote{\href{https://github.com/bussilab/crescale-lammps.git}{https://github.com/bussilab/crescale-lammps.git}} implementations of the anisotropic SCR algorithm (top lines), with additional software specifities (bottom lines).}
\begin{threeparttable}
\label{tab:implementations}
\begin{tabular}{l c c c}
\toprule[0.5pt]\toprule[0.5pt]
{} & \small\textbf{SimpleMD} & \small\textbf{GROMACS} & \small\textbf{LAMMPS} \\\midrule
Number of propagated variables & 9 & 9 & 9 \\
Multiple time step & yes & yes & no \\
Rotations or not & both & no & no \\
Rescaling of momenta or not & both & yes & yes \\
Time-reversible implementation & yes & no & no \\
Isotropic implementation & yes & yes & yes \\
Semi-isotropic implementation & no & yes & yes \\
Coupling of arbitrary components & no & no & yes \\\midrule
Electrostatics & no & yes & yes \\
Constraints & no & yes & yes \\
Potentials for solid-state materials & no & no & yes \\
Parrinello-Rahman barostat & no & yes & no \\
MTTK barostat & no & yes\tnotex{tnote:MTTK_gromacs} & yes
\\\bottomrule[0.5pt]\bottomrule[0.5pt]
\end{tabular}
\begin{tablenotes}
\item\label{tnote:MTTK_gromacs}Only isotropic version and in absence of constraints
\end{tablenotes}
\end{threeparttable}
\end{table}
The tests discussed in the following are performed on a variety of crystal systems, including a Lennard-Jones (LJ) solid, ice, gypsum (chemical formula: $\text{(CaSO)}_\text{4}\cdot\text{2(H}_\text{2}\text{O)}$) and gold (Au). Since in all the simulations the reference distributions for $\mathbf{h}$ are unknown, the results are validated against the reference barostats available in the MD software employed.
\section{Lennard-Jones crystal}\label{LJ_crystal}
The first tests were performed with the modified SimpleMD program on a LJ crystal with $N = 256$ particles arranged in a face-centered-cubic (fcc) lattice. For each choice of the external parameters ($\tau_p$ and $n_s$), the system was simulated for $10^6$ steps with time step $\Delta t = 0.005$, external hydrostatic pressure $P_0=1$ and temperature $T=0.1$, using a SVR thermostat with relaxation time $\tau_T=0.05$. All the parameters are reported here in reduced LJ units. The input isothermal compressibility was set to $\beta_T = 0.3$, as estimated for the same system in the liquid phase at $T=1.5$ \cite{crescale_iso}. The simulations were carried out by using a cut-off distance $r_{\text{cut}}=2.5$ for the interactions and accumulating statistics at each step.
First, the behaviour of the volume distribution was studied using different values of $\tau_p$ at fixed barostat stride $n_s = 1$. Results were validated by comparing the volume distributions generated by the anisotropic barostat, both in the Euler and the time-reversible (TR) implementations, with the ones generated by the isotropic barostat, employing the same input parameters (see \cref{fig:taups_LJ}). All the analysis were carried out by discarding the first $10^5$ steps of the simulations.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.1 cm]{core/third_chapter/volume_avgs_LJ.png}\\[1 ex]
\caption{\RaggedLeft Volume average vs. $\tau_p\hspace{0.5cm}$}
\label{fig:volume_taups_LJ}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.1cm]{core/third_chapter/volume_vars_LJ.png}\\[1 ex]
\caption{\RaggedLeft Volume variance vs. $\tau_p\hspace{0.7cm}$}
\label{fig:std_taups_LJ}
\end{subfigure}
\caption{Results from the simulations of a LJ crystal system in the $N\mathbf{S}T$ ensemble, as a function of $\tau_p$. Error bars were computed by means of block bootstrap analysis (see Appendix \ref{appendix_block_bootstrap}). Blue and orange lines refer to the anisotropic simulations using respectively the Euler and the time-reversible (TR) integrators in SimpleMD.}
\label{fig:taups_LJ}
\end{figure}
\noindent As \cref{fig:volume_taups_LJ} shows, the average volumes obtained with the isotropic and the anisotropic SCR barostats converge by increasing $\tau_p$, although this behaviour is more accentuated in the Euler case. The reason of this behaviour is that integration errors become more relevant when $\tau_p$ is small, since decreasing this parameter is equivalent to increase the time step at fixed relaxation time. In this regime, the integration errors on the average volume appear identical for the Euler and the TR integrators. As expected, the average volume appears less dependent on $\tau_p$ for the distributions generated by the isotropic barostat; indeed, integration errors in the anisotropic case generate larger deviations from the exact volume distribution, as they result from the propagation of nine variables instead of one. These deviations have a systematic nature and do not enter in the error bars, which only depend on the autocorrelation time of the volume. As shown in \cref{fig:acfs_volume_LJ} and discussed in detail in Appendix \ref{appendix_relaxation_autocorrelation}, this autocorrelation time is completely dictated by the barostat relaxation time, namely it can be effectively identified with $\tau_p$. Since the uncertainty of an average value is known to increase with the autocorrelation time of the series from which it is computed (see Appendix \ref{appendix_autocorrelation}), the behaviour of the error bars of $\langle V\rangle$ is meaningful. Looking at \cref{fig:std_taups_LJ}, the volume fluctuations generated by the anisotropic SCR method appear cosnistent with the ones obtained with the isotropic barostat, regardless of the relaxation time employed. The deviations of the volume variance due to integrations errors in the small-$\tau_p$ regime are more evident for the Euler integrator, suggesting an increased accuracy in the TR scheme.
Only for the anisotropic integrators, a further comparison was carried out between some relevant distributions extracted from the box matrix components, namely for the squared moduli of the cell vectors $|\mathbf{a}|^2,\,|\mathbf{b}|^2,\,|\mathbf{c}|^2$ and their three scalar products $\mathbf{a}\cdot\mathbf{b}\,, \mathbf{a}\cdot\mathbf{c}\,, \mathbf{b}\cdot\mathbf{c}$. Also in this case (see \cref{fig:box_comp_LJ,fig:box_vars_LJ} in the Appendices) the performances of the Euler and the TR integrators appear almost identical.
For the same simulations, the volume autocorrelation function (ACF) was computed for each value of $\tau_p$ (see \cref{fig:acfs_volume_LJ}). Note that, since the input isothermal compressibility is the one in the liquid phase, the real $\beta_T$ for the crystal system is larger than the input one. As a consequence, the values of $\tau_p$ are smaller than the actual volume relaxation times. Anyway, the correct values can be calculated \emph{a posteriori} with the simple rescaling
\begin{equation}
\tau_{p,\text{exp}} = \frac{\beta_{T,\text{exp}}}{\beta_{T,\text{input}}}\tau_{p,\text{input}}\,,
\end{equation}
after estimating the real isothermal compressibility according to \cref{beta_T_formula}, resulting in $\beta_{T,\text{exp}}\simeq 0.015$.
\noindent It is also possible to study how fast an estimate of the volume variance decorrelates as a function of $\tau_p$, since this quantity is essential to calculate physical observables such as the isothermal compressibility. Note that in principle it is \emph{not} possible to define the ACF of the variance, since this quantity is a global property of the trajectory and is not calculated as the average of consecutive "instantaneous variances". If we assume that the average volume $\overbar{V}$ is known exactly, however, the variance can be computed as the mean value over the time series $\{(V_1-\overbar{V})^2),...,(V_n-\overbar{V})^2 \}$, namely
\begin{equation}
\sigma_V^2 = \frac{1}{n}\sum_{j=1}^n\Big(V_j-\overbar{V}\Big)^2\,,
\end{equation}
where $j$ is the index for the MD step and $n$ is the total number of samplings. In the unbiased estimator of the variance $n$ should be replaced by $n-1$, but this correction is negligible for a large number of samplings. The ACFs of the volume variance are shown \cref{fig:acfs_var_LJ}; for each $\tau_p$, the value of $\overbar{V}$ is taken as the average over the entire volume trajectory.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/third_chapter/ACFs_volume_LJ.png}\\[1 ex]
\caption{\RaggedLeft Volume ACFs$\hspace{1.4cm}$}
\label{fig:acfs_volume_LJ}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/third_chapter/ACF_var_LJ.png}\\[1 ex]
\caption{\RaggedLeft ACFs of volume variance$\hspace{0.6cm}$}
\label{fig:acfs_var_LJ}
\end{subfigure}
\caption{Left panel: ACFs of the volume time series from the simulations of a LJ crystal with the anisotropic SCR barostat, using the Euler integrator in SimpleMD. The values of $\tau_p$ reported in the legend are the input ones, while the dashed lines represent the exponentially decaying functions $e^{-t/\tau_{p,\text{exp}}}$, where the correct values of the relaxation times are computed \emph{a posteriori}. Right panel: variance ACFs computed under the assumption of known volume averages. The dashed lines are in this case the exponential functions $e^{-2t/\tau_{p,\text{exp}}}$, which represent the anlytical ACFs of the variance in the limit case of a Gaussian-distributed volume following a Langevin dynamics (see Appendix \ref{appendix_relaxation_autocorrelation}).}
\label{fig:ACFs_LJ}
\end{figure}
\noindent Note that these ACFs decay faster than the volume ones (ideally, with halved characteristic time); as a consequence, a calculation of the variance of the volume converges faster than a calculation of its average when the SCR barostat is employed. Similar results were obtained in the isotropic case. The ACFs shown in \cref{fig:ACFs_LJ} are obtained with the anisotropic barostat in the Euler integration scheme; employing the TR integrator, the deviations of the variance ACFs in the large-$\tau_p$ regime are no present anymore (see \cref{fig:ACFs_LJ_TR} in the Appendices).
For each tested value of $\tau_p$, the ACFs of the squared moduli and the scalar products of the box vectors were also computed, resulting qualitatively in the same behaviour shown in \cref{fig:ACFs_LJ}. These ACFs go to zero slower than the volume ACFs, but their limiting analytical behaviour cannot be easily predicted. As an example, see for instance \cref{fig:ACFs_mod2_LJ} in the Appendices.
For the same simulations, the effective energy drift is calculated along the trajectories generated by the isotropic and the TR anisotropic implementations (see \cref{fig:slopes_LJ}).
\begin{figure}[h!]
\centering
\includegraphics[width = 0.55\textwidth]{core/third_chapter/slopes.png}
\caption{Absolute value of the effective energy drift per step from the simulations of the LJ crystal in SimpleMD. Each value is the slope of a line interpolating the effective energy drift on the entire trajectory. The green $\times$ markers are actually negative values, for which the absolute value is taken in order to represent all data on a logarithmic scale. The drift consistently increase in the small-$\tau_p$ regime, i.e. when integration errors have a larger impact. For $\tau_p \gtrapprox 1$, the energy drift appears irrelevant on the time scale of the trajectories, resulting in noisy estimates of the slope above and around zero.}
\label{fig:slopes_LJ}
\end{figure}
Both \cref{fig:taups_LJ,fig:ACFs_LJ} are obtaind from simulations where momenta are rescaled and, in the anisotropic case, rotations are eliminated. However, all the possible four combinations of these two options were tested, showing equivalent results to the ones discussed here.
Additional simulations were performed at fixed volume relaxation time ($\tau_p = 10$) and different barostat strides $n_s$, in order to validate the application of a multiple-time-step approach to the anisotropic SCR barostat. The results are reported in \cref{fig:volume_vs_stride}.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.2 cm]{core/third_chapter/avgs_vs_stride_LJ.png}\\[1 ex]
\caption{\RaggedLeft Volume average vs. $n_s\hspace{0.5cm}$}
\label{fig:avg_vs_stride}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.2 cm]{core/third_chapter/vars_vs_stride_LJ.png}\\[1 ex]
\caption{\RaggedLeft Volume variance vs. $n_s\hspace{0.5cm}$}
\label{fig:std_vs_stride}
\end{subfigure}
\caption{Average and variance of volume distributions as a function of the barostat stride $n_s$, using the anisotropic SCR barostat in SimpleMD. Error bars are computed with block bootstrap analysis. The simulations were performed with $\tau_p = 10$.}
\label{fig:volume_vs_stride}
\end{figure}
Interestingly, the average volume appears to be well reproduced even for large barostat strides (up to $n_s = 100)$, while systematic errors for the volume fluctuations become evident for $n_s \gtrapprox 20$. Also in this test, the results obtained with the Euler and the TR integration scheme appear equivalent.
Similar simulations were also performed with the modified version of GROMACS 2021.2, using a larger crystal with $N=1000$ Argon atoms, employing the GROMOS 54A7 force field and setting $T=5$ K and $P_0 = 1$ bar. The crystal structure was obtained first by equilibrating the system in the $NVT$ ensemble for 500 ps, and then with a simulated annealing protocol in the $NP_0 T$ ensemble from $T = 80$ K to $T=5$ K, resulting in an hexagonal-close-packed (hcp) structure with defects. Scanning different values of $\tau_p$, production runs of $10$ ns were carried out using both the SCR barostat and the other pressure coupling methods available in GROMACS (see \cref{tab:implementations}), except for the Berendsen one, which is known to generate wrong volume fluctuations. Temperature was controlled by a SVR thermostat with a relaxation time $\tau_T = 0.01$ ps. Both thermostat and barostat were applied every 10 steps, and statistics were accumulated every 20 steps (40 fs). The time step was set to $\Delta t = 2$ fs, the cut-off distance for the LJ interactions to 1 nm and the input isothermal compressibility to $\beta_T = 3.53\times 10^{-5}$ bar$^{-1}$, using a rough estimation on a preliminary run. Instead of using the standard GROMACS leap frog integrator, the simulations with the Parrinello-Rahman and the MTTK barostats were carried out with a velocity Verlet scheme, which is more accurate for these coupling methods and partially reduces (but does not eliminate) the pathological behaviours discussed in the following. All the analysis were carried out by discarding the first 2.5$\times$10$^5$ steps of the simulations.
Averages and variance of the volume distributions from the GROMACS simulations are shown in \cref{fig:avg_vs_taup_LJ_GMX,fig:std_vs_taup_LJ_GMX}.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.9 cm]{core/third_chapter/avgs_aniso_LJ_GMX.png}\\[1 ex]
\caption{\RaggedLeft Anisotropic barostats$\hspace{0.5cm}$}
\label{fig:avg_aniso_vs_taup_LJ_GMX}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.9 cm]{core/third_chapter/avgs_iso_LJ_GMX.png}\\[1 ex]
\caption{\centering Isotropic barostats}
\label{fig:avg_iso_vs_taup_LJ_GMX}
\end{subfigure}
\caption{Average of volume distributions as a function of the barostat relaxation time $\tau_p$, using the barostats available in GROMACS 2021.2 (SCR: Stochastic Cell Rescaling, PR: Parrinello-Rahman, MTTK: Martyna-Tuckerman-Tobias-Klein). In the PR and MTTK methods, $\tau_p$ is related to the barostat mass $W$ as $W\propto \tau_p^2$.}
\label{fig:avg_vs_taup_LJ_GMX}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 5.1 cm]{core/third_chapter/vars_aniso_LJ_GMX.png}\\[1 ex]
\caption{\RaggedLeft Anisotropic barostats$\hspace{0.5cm}$}
\label{fig:std_aniso_vs_taup_LJ_GMX}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 5.1 cm]{core/third_chapter/vars_iso_LJ_GMX.png}\\[1 ex]
\caption{\centering Isotropic barostats}
\label{fig:std_iso_vs_taup_LJ_GMX}
\end{subfigure}
\caption{Variance of volume distributions as a function of the barostat relaxation time $\tau_p$. The zoomed regions of the plots refer to the range of $\tau_p$ values where the PR and the MTTK methods do not produce pathological volume distributions.}
\label{fig:std_vs_taup_LJ_GMX}
\end{figure}
\noindent As a necessary observation, the time constant $\tau_p$ is not equivalently defined for the different barostats employed; as a consequence, comparing the distributions generated by different methods for a given value of $\tau_p$ is not completely meaningful. However, it is useful to study how the volume distributions generated by different barostats are affected by equivalent variations of $\tau_p$. As \cref{fig:avg_aniso_vs_taup_LJ_GMX} shows, the average volume reproduced by the anisotropic SCR barostat appears perfectly in agreement with the one obtained from the anisotropic PR method, as well as in the case of isotropic volume fluctuations (see \cref{fig:avg_iso_vs_taup_LJ_GMX}). Moreover, in both the cases the SCR method shows a reduced sensitivity to the input parameter $\tau_p$ than the PR barostat, for which sampling problems for large $\tau_p$ are more evident. Using the MTTK method, which works only in the isotropic case in GROMACS 2021.2, these sampling problems are already present from $\tau_p \gtrapprox 2$ ps. In general, above $\tau_p\approx 5$ the volume trajectories generated by the PR and the MTTK barosats do not equilibrate properly and show non-stationary behaviours (see Appendix \cref{appendix_pathological_distributions}). As a consequence, the results of these simulations are reliable and actually comparable with the SCR method only up to $\tau_p \approx 2-5$. Looking at \cref{fig:std_vs_taup_LJ_GMX}, the SCR method appears more robust than the PR barostat in reproducing the correct volume fluctuations against variations of $\tau_p$. Moreover, the PR method is more affected by accuracy problems for small relaxation times. It has not to be excluded that the errors observed for the MTTK and the PR barostats - especially in the anisotropic case - could depend on the technical details of the GROMACS implementation.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.95\textwidth]{core/third_chapter/ACFs_volume_LJ_GMX.png}
\caption{Comparison of the volume ACFs obtained with the anisotropic SCR barostat and the PR method in GROMACS. The dashed lines represent the decaying functions $\text{exp}(-t/\tau_{p,\text{exp}})$, where $\tau_{p,\text{exp}}$ is computed from the correct value of the isothermal compressibility.}
\label{fig:ACFs_volume_LJ_GMX}
\end{figure}
For the anisotropic SCR barostat, the ACFs of the volume show the same behaviour observed in the SimpleMD simulations, with autocorrelation times scaling linearly with the input values of $\tau_p$.
\Cref{fig:ACFs_volume_LJ_GMX} shows a comparison between the volume ACFs obtained with the anisotropic SCR and PR barostats, considering values of $\tau_p$ within the reliable range previously discussed. At fixed $\tau_p$, the anisotropic SCR barostat appears more efficient in terms of decorrelation speed. However, this does \emph{not} imply that the uncertainty associated to an estimate of the average volume is larger with the PR method. Indeed, the actual autocorrelation time related to the statistical uncertainty of $\langle V\rangle$ is defined as the integral of the ACF (see Appendix \ref{appendix_autocorrelation}); as a consequence, if the damped oscillations of the ACF are symmetric with respect to zero, the integral can be very small even if the envelope of this function does not go to zero as fast as in the SCR case. This is the reason why the standard errors on the average volume, estimated with a block bootstrap analysis, are smaller from the trajectories generated by the PR barostat for $\tau_p< 5$ (see \cref{fig:SE_avg_LJ_GMX}). On the other hand, the ACFs of the volume variance from the PR trajectories do not show the same symmetry (see \cref{fig:ACFs_var_LJ_GMX} in the Appendices), resulting in an integrated autocorrelation time that is larger than in the SCR trajectories. As a consequence, the standard error on the estimate of volume fluctuations is smaller in the SCR simulations (see \cref{fig:SE_std_LJ_GMX}).
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 3.9 cm]{core/third_chapter/SE_avg_LJ_GMX.png}\\[1 ex]
\caption{\RaggedLeft Errors of volume average$\hspace{0.3cm}$}
\label{fig:SE_avg_LJ_GMX}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 3.9 cm]{core/third_chapter/SE_var_LJ_GMX.png}\\[1 ex]
\caption{\RaggedLeft Errors of volume variance$\hspace{0.2cm}$ }
\label{fig:SE_std_LJ_GMX}
\end{subfigure}
\caption{Standard errors of volume average and variance computed with block bootstrap analysis (see Appendix \ref{appendix_block_bootstrap}). The dashed lines represent the expected behaviour as a function of $\tau_p$, computed with \cref{standard_error_autocorr}.}
\label{fig:SEs_LJ_GMX}
\end{figure}
\section{Crystal Ice I$_h$}\label{IceIh_analysis}
In order to test the performance of the anisotropic SCR barostat in presence of constraints, simulations were performed on a system of Ice I$_h$ (see \cref{fig:IceIh}) composed of $N=3072$ atoms, using the TIP4P/Ice model \cite{tip4pice}.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.34\textwidth]{core/third_chapter/iceih.png}
\caption{\centering Orthographic perspective of the simulated hexagonal crystal ice I$_h$.}
\label{fig:IceIh}
\end{figure}
In GROMACS, for each $\tau_p$ the crystal system was simulated for 4 $ns$ with a time step of $1$ fs, accumulating statistics every 10 steps (10 fs) and applying both barostat and thermostat at each step. The simulations were carried out with external hydrostatic pressure $P_0 = 1$ bar and external temperature $T = 270$ K, employing a SVR thermostat with relaxation time $\tau_T = 0.1$ ps. The input isothermal compressibility, estimated over a preliminary run of $1$ ns, was set to $\beta_T = 9.49\times 10^{-6}$ bar$^{-1}$. All the simulations were performed by employing as integrator of Hamilton's equations a modified version of the velocity Verlet algorithm, where the kinetic energy is determined as the average of the two half step kinetic energies. The use of this integrator appears to reduce the pathological behaviour of some distributions generated by the PR method, as discussed in \cref{LJ_crystal}. In all the simulations, electrostatics was treated with a Particle-Mesh Ewald (PME) approach. The same system was also simulated in LAMMPS with equivalent MD options, and performing a comparison with the anisotropic MTTK barostat. The behaviour of average and variance of the volume distributions, computed by discarding the first 4$\times 10^4$ steps, are shown in \cref{fig:vol_IceIh} for different $\tau_p$ values.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 4.3 cm]{core/third_chapter/avgs_Ice.png}\\[1 ex]
\label{fig:avg_IceIh}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 4.3 cm]{core/third_chapter/vars_Ice.png}\\[1 ex]
\label{fig:std_IceIh}
\end{subfigure}
\caption{Results from the simulations of the Ice$_{h}$ system. Average (left panel) and variance (right panel) of volume distributions are shown as functions of the barostat relaxation time $\tau_p$, comparing the anisotropic SCR barostat with the anisotropic pressure coupling methods available in GROMACS 2021.2 and LAMMPS (2 July 2021 release).}
\label{fig:vol_IceIh}
\end{figure}
\noindent Both in GROMACS and in LAMMPS, the SCR barostat generates volume distributions that are consistent with the MTTK ones. Moreover, the LAMMPS implementation of the anisotropic SCR method is the one showing the smallest sensibility to variations of the relaxation time. The anisotropic PR barostat in GROMACS, instead, generates volume fluctuations that are not consistent with the other pressure coupling methods, and it is the one showing the largest deviation for small $\tau_p$. Similar results are obtained by studying how the squared moduli of the box vectors and their scalar products are distributed (see \cref{fig:box_comp_ice,fig:box_vars_ice} in the Appendices).
The ACFs of the volume for the SCR simulations are consistent with the expected exponential behaviour as a function of $\tau_p$ (see \cref{fig:ACFs_avgs_Ice}). The largest deviation is shown for $\tau_p = 0.1$ ps, when the observed decaying is slower than the reference exponential one. This slowdown of the first-order relaxation occurs when $\tau_p$ is of the same order or smaller than the timescale in the rearrangement of atoms, which then becomes the bottleneck for volume dynamics.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.95\textwidth]{core/third_chapter/ACFs_volume_Ice.png}
\caption{ACFs of the volume for different values of $\tau_p$, compared among the pressure coupling methods employed in the simulations of the Ice I$_h$ crystal.}
\label{fig:ACFs_avgs_Ice}
\end{figure}
\noindent Also the ACFs of the volume variance match the expected decaying functions (see \cref{fig:ACFs_vars_Ice} in the Appendices). In this case, statistical errors associated to average and variance of volume distributions are larger for the SCR trajectories, at fixed $\tau_p$ (see \cref{fig:SE_Ice} in the Appendices). However, as already commented for the simulations of the Argon crystal, the results obtained from different barostats are not directly comparable for the same value of the relaxation time, since it does not exist a clear mapping between the corresponding definitions of $\tau_p$. Moreover, since the SCR method performs well up to $\tau_p=0.1$, this analysis suggests that small values of the relaxation time could be used to reduce the statistical error without introducing systematic ones.
Using the GROMACS implementation, additional simulations were performed by varying the barostat stride $n_s$ at fixed relaxation time (see \cref{fig:variable_stride_ice}). The value selected is $\tau_p = 1$ ps, for which the SCR barostat in GROMACS appears consistent with the other methods. With the only exception of $n_s=100$, average and variance of the volume distributions are consistent with the results obtained by applying the barostat at each step.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 3.8 cm]{core/third_chapter/avgs_vs_stride_Ice.png}\\[1 ex]
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 3.8 cm]{core/third_chapter/vars_vs_stride_Ice.png}\\[1 ex]
\end{subfigure}
\caption{Results from GROMACS simulations of the ice I$_h$ system, employing the anisotropic SCR barostat at fixed relaxation time $\tau_p=1$ ps. Average (left panel) and variance (right panel) of volume distributions are shown as functions of the barostat stride $n_s$. The error region, computed with block bootstrap analysis, is associated to the smallest stride tested ($n_s = 1$).}
\label{fig:variable_stride_ice}
\end{figure}
\section{Gypsum crystal}
Additional tests of the anisotropic SCR barostat in LAMMPS were performed by simulating a gypsum crystal (see \cref{fig:gypsum}) composed of $N=3456$ atoms, setting $P_0 = 1$ bar and $T=270$ K.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.55\textwidth]{core/third_chapter/gypsum.png}
\caption{Crystal structure of $\text{(CaSO)}_\text{4}\cdot\text{2(H}_\text{2}\text{O)}$, commonly known as gypsum. White, red, blue and yellow beads represent H, O, Ca and S atoms respectively. Hydrogen bonds are shown as dashed lines. Source: \cite{gypsum_img}.}
\label{fig:gypsum}
\end{figure}
Simulations of $4$ ns were carried out using the SCR and the MTTK barostats, both in the the anisotropic implementations, accumulating statistics every $10$ steps (4 ps). Barostat and thermostat relaxation times were both set to $\tau_p = \tau_T = 0.1$ ps. For the simulations performed with the SCR barostat, the input isothermal compressibility was fixed to $\beta_T = 2.3\times 10^{-6}$ bar$^{-1}$. The statistics of interest are reported in \cref{tab:gypsum_lammps_aniso} for a comparison with the MTTK pressure coupling method. For additional results related to the single components of the box matrix $\mathbf{h}$ and the internal pressure tensor $\mathbf{P}_{\text{int}}$, see \cref{tab:gypsum_lammps_aniso_box} in the Appendices.
\begin{table}[h!]
\centering
\caption{Results from the simulations of the gypsum crystal in LAMMPS. The pairs of raws refer respectively to the distributions of volume, internal pressure, density, potential energy and temperature, for which average and standard deviations are reported.}
\begin{threeparttable}
\label{tab:gypsum_lammps_aniso}
\begin{tabular}{l c c c c}
\toprule[0.5pt]\toprule[0.5pt]
{} & \small\textbf{Anisotropic SCR} & \small\textbf{Anisotropic MTTK} \\\midrule
$\langle V\rangle\,$ (nm$^3$) & $36.8682\pm0.0008$ & $36.8691\pm0.0010$ \\
$\sigma_V\,$ & $0.0624\pm0.0003$ & $0.0622\pm0.0002$ \\ \midrule
$\langle P_{\text{int}}\rangle\,$ (bar) & $-7\pm 5$ & $1.0\pm 1.4$ \\
$\sigma_{P_{\text{int}}}\,\,$ & $946\pm3$ & $940\pm 10$ \\ \midrule
$\langle\rho\rangle\,$ (kg/cm$^3$) & $2233.44\pm0.05$ & $2233.38\pm0.06$ \\
$\sigma_\rho\,\,$ & $3.78\pm0.02$ & $3.765\pm0.014$ \\ \midrule
$\langle U\rangle\,$ (kJ/mol) & $-789541.4\pm1.4$ & $-789542.3\pm1.4$ \\
$\sigma_U\,\,$ & $171.0\pm0.7$ & $170.5\pm0.7$ \\ \midrule
$\langle T\rangle\,$ (K) & $270.00\pm0.03$ & $270.00\pm0.03$ \\
$\sigma_T\,\,$ & $3.762\pm0.012$ & $3.759\pm0.012$
\\\bottomrule[0.5pt]\bottomrule[0.5pt]
\end{tabular}
\end{threeparttable}
\end{table}
\noindent The distributions generated by the two barostats appear consistent, with the usual \emph{caveat} that errors cannot be directly compared for the same choice of $\tau_p$. As already discussed in \cref{LJ_crystal} for the PR barostat, the symmetric oscillations in the MTTK volume ACF can result in a shorter autocorrelation time than the one obtained with the SCR method, due to a cancellation effect between positive and negative contributions when the volume ACF is integrated. When this happens, as in the case of the Ice I$_h$ simulations, the MTTK barostat achieves smaller statistical errors on the average of volume-related quantities. However, when a small enough $\tau_p$ is employed (as in this case, i.e. $\tau_p = 0.1$ ps) the MTTK volume ACF (see \cref{fig:ACFs_gypsum}) can suffer from a damping that makes the cancellation effect less efficient, resulting in statistical errors on volume-related averages that are comparable with the SCR method or even larger (see for instance $\langle V\rangle$ and $\langle \rho\rangle$ in the table above).
\newpage
Also for this system, the the ACFs of the volume and its variance show an exponential decay in agreement with the expected behaviour in the SCR simulation (see \cref{fig:ACFs_gypsum}).
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 4. cm]{core/third_chapter/ACFs_volume_gypsum.png}\\[1 ex]
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 4. cm]{core/third_chapter/ACFs_var_gypsum.png}\\[1 ex]
\end{subfigure}
\caption{ACFs of the volume (left panel) and its variance (right panel) from the simulations of the gypsum crystal system in LAMMPS. The dashed lines in the left and right panels are respectively the exponential functions $e^{-t/\tau_{p,\text{exp}}}$ and $e^{-2t/\tau_{p,\text{exp}}}$.}
\label{fig:ACFs_gypsum}
\end{figure}
\section{Au crystal}
In order to test the anisotropic SCR method in presence of a generic external stress, namely for $\mathbf{S}_{\text{dev}}\neq 0$ (see \cref{npt_theory}), two simulations of a gold (Au) crystal system with $N=4000$ atoms were carried out in LAMMPS, applying the SCR and the MTTK barostats at each step and employing the modified embedded-atom model (EAM) described in \cite{embedded_atom}. In both the runs the crystal, intialized to an FCC structure, was simulated with time step $\Delta t=1$ fs, keeping the external hydrostatic pressure constant to $P_0 = 1$ bar and increasing linearly the $xz$ shear stress (and its symmetric $zx$) of 0.05 bar at each step, starting from $S_{xz} = S_{zx} = 0$ bar. The purpose of the simulations is to identify the extreme shear that the crystal can bear before its breaking. The external temperature was set to $T=298.15$ K and controlled with a SVR thermostat with relaxation time $\tau_T = 0.1$ ps. The barostat relaxation time was fixed to $\tau_p = 1$ ps for the MTTK barostat and $\tau_p = 0.1$ ps for the SCR run, using as input bulk modulus $k_T = \beta_T^{-1} = 1.7\times 10^6$ bar. Statistics were saved every 100 steps ($0.1$ ps). In the two simulations, the breaking of the crystal structure occurs consistently at $t\approx 3.03-3.05$ ns, when the external shear stress is $S_{xz}\approx 15.1-15.3$ kbar (see \cref{fig:gold_slip_plot}). This result is in agreement with the value predicted by \cite{slip_gold_reference}, performing static calculations with the GULP software \cite{GULP}.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.75\textwidth]{core/third_chapter/gold_slip_plot.png}
\caption{Results from the simulations of the gold crystal systems in LAMMPS, employing the MTTK and SCR anisotropic barostats in presence of a linearly increasing shear stress $S_{xz}$. The plot shows the trajectory for the modulus of the cell vector $\mathbf{b}$, which appears as a good variable to describe the crystal breaking.}
\label{fig:gold_slip_plot}
\end{figure}
A visualization of the crystal breaking, which occurs through the slipping of the crystallographic planes (1\,1\,1), is reported in \cref{fig:gold_slip_vmd}.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.44\textwidth}
\RaggedLeft
\includegraphics[height = 3.5 cm]{core/appendices/gold_0.png}\\[1 ex]
\caption{\RaggedLeft $\,t=3.035$ ns $\hspace{0.5cm}$}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.44\textwidth}
\RaggedRight
\includegraphics[height = 3.5 cm]{core/appendices/gold_1.png}\\[1 ex]
\caption{\RaggedLeft $\,t=3.260$ ns$\hspace{1.5cm}$}
\end{subfigure}
\vfil
\vspace{0.3cm}
\begin{subfigure}[b]{0.44\textwidth}
\RaggedLeft
\includegraphics[height = 3.5 cm]{core/appendices/gold_2.png}\\[1 ex]
\caption{\RaggedLeft $\,t=3.515$ ns $\hspace{0.2cm}$}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.44\textwidth}
\RaggedRight
\includegraphics[height = 3.5 cm]{core/appendices/gold_3.png}\\[1 ex]
\caption{\RaggedLeft $\,t=3.517$ ns$\hspace{0.8cm}$}
\end{subfigure}
\caption{Frames from the simulations of the gold crystal in LAMMPS, using the anisotropic SCR barostat.}
\label{fig:gold_slip_vmd}
\end{figure}
\chapter{Conclusions}
In this work, the stochastic cell rescaling (SCR) barostat \cite{crescale_iso} is generalized to anisotropic volume fluctuations, allowing the system box to change its shape during MD simulations in the isothermal-isobaric ensemble. This pressure coupling method is formulated in terms of a nine-dimensional and first-order stochastic differential equation, whose deterministic part resembles the equation of the anisotropic Berendsen barostat \cite{berendsen}, except for a corrective term that becomes negligible in the thermodynamic limit. Unlike second-order methods, the SCR barostat can be used effectively during the equilibration phase. When coupled with a thermostat, the algorithm generates the correct anisotropic isothermal-isobaric ensemble, both when a isotropic external pressure is applied and in the case of a deviatoric external stress, i.e. with anisotropic external conditions. The anisotropic SCR method is shown to be equivalent to the high-friction limit of a second-order barostat, namely the Parrinello-Rahman barostat coupled with a Langevin thermostat.
Two equivalent formulations of the method are presented, where the components of the system box are propagated by keeping constant either the physical or the rescaled momenta. These two possibilities result respectively in the use of the average and instantaneous kinetic energy to compute the virial.
As in the isotropic case and in the Berendsen barostat, the algorithm has effectively a single input parameter, namely the ratio between the system isothermal compressibility $\beta_T$ and the barostat relaxation time $\tau_p$. If the accurate value of the isothermal compressibility is unknown \emph{a priori}, it is possible to use an estimate over a short preliminary run for its input, re-computing \emph{a posteriori} the correct values of both $\beta_T$ and $\tau_p$.
For the integration of the anisotropic SCR equations, two schemes are proposed and tested. The first method employs a simple Euler propagation of the box matrix components, generating trajectories that are not reversible. The second one is built on a conceptual separation between the propagation of the isotropic degree of freedom, i.e. the volume, and the evolution of the remaining eight degrees of freedom, which are responsible for the change of shape of the box. The two operations are applied in a way that allows generating time-reversible trajectories. As a consequence, a quantity called \emph{energy drift} \cite{langevin_bussi}, which behaves as a constant of motion is the limit of small time step, can be defined and computed for this integrator, so that its variations along the trajectory can be used to monitor the violation of detailed balance.
The algorithm has been tested with three different implementations, using the SimpleMD, GROMACS 2021.2 and LAMMPS (2 July 2021 release) MD softwares. Consistently among all the implementations, box rotations are eliminated after propagating all the nine box variables, with an orthogonal transformation that compensates for the three redundant degrees of freedom. Simulations were carried out on a variety of solid-state systems, both in presence and in absence of inter-molecular interactions and constraints. In all the tests, results appear consistent with the ones from the reference barostats employed in the comparison, especially as regards the distributions of the volume and the ones extracted from the single box matrix components. In terms of accuracy, the two integration schemes proposed (Euler and time-reversible) show an equivalent performance. The generated distributions appear to be stable in the entire range of $\tau_p$ values tested, covering 3 orders of magnitude in the LJ crystal simulations and 1-2 orders of magnitude in the Ice I$_h$ tests. In some cases, the method appears to be more robust than other second-order pressure coupling methods against systematic sampling errors in the small-$\tau_p$ regime. For this reason, the inaccuracy in the input value of $\tau_p$, due to a possible error in the input isothermal compressibility, is considered as a minor drawback of the method. The algorithm was also tested in a multiple-time-step fashion, showing stable results up to a stride $n_s = 10-20$.
In all the simulations, the autocorrelation functions of the volume and of its variance match the expected exponential decays as functions of the corrected relaxation time $\tau_p$, namely $\exp(-t/\tau_p)$ and $\exp(-2t/\tau_p)$. In other words, the input parameter $\tau_p$ can be used to estimate \emph{a priori} how fast the statistical error of any volume-dependent quantity approaches zero with the length of the trajectory. The current formulation of the algorithm does not allow to control the autocorrelation times of the single box components with the same accuracy, as the their limiting behaviours cannot be analytically predicted. A further formulation of the algorithm that could allow for this additional feature, employing a tensorial expression for the isothermal compressibility, is proposed in this work in view of future refinements of the method.
As other anisotropic pressure coupling methods, the proposed barostat could find applications in MD simulation of two classes of systems, namely in the presence of internal anisotropies, as in the case of crystal systems, or in the case of an external anisotropic stress, that can result in modifications of the crystal structure and eventually in conformational phase transitions \cite{crystal_phase_trans,crystal_phase_trans2}. In this work, an example of this realistic applications is shown by simulating a gold crystal system in presence of a variable external shear stress, identifying the limiting stress causing the breaking of the structure.
\begin{appendices}
\chapter{Stochastic differential equations}\label{appendix_stocdiffeq}
Stochastic differential equations (SDEs) allow to model and describe stochastic processes, but can also employed to construct non-deterministic sampling algorithms. Considering a generic one-dimensional variable $x$, the most common SDE that one finds in the literature is the first-order equation
\begin{equation}\label{Langevin_colored}
\dot{x}(t) = a(x,t) + b(x,t) \mathcal{\eta}(t)\,,
\end{equation}
where $\eta(t)$ is a stochastic process called \emph{Gaussian white noise}, namely it follows a zero-mean and unit-variance Gaussian distribution at any fixed time $t$ and it has an instantaneously decaying autocorrelation:
\begin{subequations}\label[pluralequation]{gaussian_white_noise}
\begin{align}
\eta(t) &\sim \mathcal{N}(0,1)\,, \\
\langle \eta(t) \eta(t') \rangle &= \delta (t-t')\,.
\end{align}
\end{subequations}
\Cref{Langevin_colored} is a generalized Langevin equation including a \emph{multiplicative noise}, since in the (RHS) the white noise $\eta(t)$ is multiplied by a function of the dynamical variable $x$. A more rigorous way of writing this equation is
\begin{equation}\label{ito_1d}
\text{d}x(t) = a(x,t)\text{d}t + b(x,t) \text{d}W\,,
\end{equation}
where $W(t)$ satisfies $\frac{\text{d}W}{\text{d}t} = \eta(t)$. The stochastic process $W(t)$ is called \emph{Wiener noise} and it is characterized by independent and Gaussian-distributed time increments,
\begin{equation}\label{wiener_incement}
\Delta W (t_0,\Delta t) = W(t_0 + \Delta t) - W(t_0) \sim \mathcal{N}(0,\Delta t)\,,
\end{equation}
where $\sigma^2 = \Delta t$ is the variance of the distribution and \cref{wiener_incement} holds for any initial time $t_0$ and any finite increment $\Delta t$.
\noindent The formal solution of \cref{ito_1d} is
\begin{equation}
x(t) = x(0) + \int_0^t \text{d}t' a\big(x(t'),t'\big) + \underbrace{\int_0^t \text{d}W(t') b\big(x(t'),t'\big)}_{= I}
\end{equation}
and it depends on how the \emph{stochastic integral} $I$ is computed. The result of this integral can be defined through an approach similar to Riemann integration, starting from a partition $\{t_0\equiv0,t_1,...,t_n,t_{n+1}\equiv t\}$ of the time interval $\left[0,t\right]$. Given this partition, $I$ can be approximated as the sum of $n$ (signed) rectangular areas:
\begin{equation}
I_n = \sum_{i=1}^n \big[W(t_{i+1}) -W(t_i)\big]b\big(x(\tau_i),\tau_i\big)\,,
\end{equation}
with the condition $\tau_i \in \left[t_i,t_{i+1}\right]$. Due to the stochastic nature of $W(t)$, the classical limit of the Riemann sums $I_n$ is replaced by the notion of \emph{mean-square convergence}:
\begin{equation}
I = \mathop{\text{ms-lim}}_{n\rightarrow\infty} I_n \hspace{0.3cm}\Longleftrightarrow \hspace{0.3cm} \lim_{n\rightarrow\infty} \langle \left(I-S_n\right)^2\rangle = 0\,.
\end{equation}
The main difference with respect to Riemann integration is that the result depends on where the points $\tau_i$ are located in the partition $\{t_i\}$, and different conventions give rise to different rules of \emph{stochastic calculus}. Among all the possible choices, the two limiting cases are $\tau_i = t_i$ and $\tau_i = (t_i + t_{i+1})/2$ are what define the \emph{It\^{o} convention} and the \emph{Stratonovich convention} respectively. In this work, all the SDEs are interpreted \emph{à la} It\^{o}.
A relevant result in \emph{It\^{o} calculus} \cite{gardiner} is the so-called \emph{It\^{o}'s lemma} (or chain rule), which allows to write the differential of a function of a stochastic process.
\begin{ito_lemma}
Let $y = y(x)$ be a function of a stochastic process $x$ described by the It\^{o} equation \cref{ito_1d}. Then:
\begin{equation}\label{ito_chain}
\text{d}y(x) = \left[\frac{\text{d}y}{\text{d}x}a(x,t) + \frac{\text{d}^2 y}{\text{d}x^2}\frac{b^2(x,t)}{2}\right]\text{d}t + \frac{\text{d}y}{\text{d}x} b(x,t) \text{d}W\,.
\end{equation}
\end{ito_lemma}
\noindent Hence, if the transformation $y = y(x)$ is invertible it is possible to obtain a new It\^{o} SDE for $y$ by substituting $x = x(y)$ in the expression above.
It\^{o} \cref{ito_1d} can be generalized to the following multidimensional SDE for the variable $\mathbf{x}^T = (x_1,...,x_n)$:
\begin{equation}\label{ito_multidim}
\text{d}\mathbf{x}(t) = \mathbf{a}(\mathbf{x},t)\text{d}t + \mathbf{b}(\mathbf{x},t) \text{d}\mathbf{W}\,,
\end{equation}
where now $\mathbf{a}$ and $\text{d}\mathbf{W}$ are vectors while $\mathbf{b}$ is a matrix. Equivalently, written for the single component $i$:
\begin{equation}
\text{d}x_i(t) = a_i(x,t)\text{d}t + \sum_{j=1}^n b_{ij}(x,t) \text{d}W_j\,.
\end{equation}
The multidimensional Wiener process $\text{d}\mathbf{W}$ is simply composed by independent scalar Wiener processes. In the derivation of the anisotropic SCR \cref{crescale_aniso}, a further generalization is considered by promoting $\mathbf{a}$,$\mathbf{b}$ and $\text{d}\mathbf{W}$ to tensors. The corresponding generalization of It\^{o}'s lemma is the following.
\begin{ito_lemma_multidim}
Let $y = y(\mathbf{x})$ be a function of a multidimensional stochastic process $\mathbf{x}$ described by the It\^{o} \cref{ito_multidim}. Then
\begin{equation}\label{ito_chain_multidim}
\text{d}y(x) = \left\{\big(\nabla_{\mathbf{x}}y\big)^T \mathbf{a} + \frac{1}{2}\text{Tr}\left[\mathbf{b}^T \mathbf{H}_{\mathbf{x}}(y)\mathbf{b}\right]\right\}\text{d}t + (\nabla_{\mathbf{x}}y)^T \mathbf{b}\, \text{d}\mathbf{W}\,,
\end{equation}
where $\nabla_{\mathbf{x}}y$ and $\mathbf{H}_{\mathbf{x}}(y)$ are the gradient and the Hessian matrix of y respectively:
\begin{subequations}
\begin{align}
\big(\nabla_{\mathbf{x}}y\big)_i &= \frac{\partial y}{\partial x_i}\,, \\
\big(\mathbf{H}_{\mathbf{x}}(y)\big)_{ij} &= \frac{\partial^2 y}{\partial x_i \partial x_j}\,.
\end{align}
\end{subequations}
\end{ito_lemma_multidim}
\chapter{Fokker-Planck equations}\label{appendix_fp}
Given a stochastic variable $x$ that evolves in time according to It\^{o} \cref{ito_1d}, it is possible to give an equivalent description of its stochastic dynamics in terms of a time-dependent probability density $\mathcal{P} = \mathcal{P}(x,t)$, such that
\begin{equation}
\mathcal{P}(x,t)\,\text{d}x\text{d}t
\end{equation}
is the probability of finding the system in $(x,x+dx)$ within the time interval $(t,t+\text{d}t)$.
\noindent The partial differential equation that defines the time evolution of $\mathcal{P}$ is called Fokker-Planck (FP) equation:
\begin{equation}\label{fp_1d}
\frac{\partial}{\partial t }\mathcal{P}(x,t) = -\frac{\partial}{\partial x}\Big(a(x,t)\mathcal{P}(x,t)\Big) + \frac{\partial^2}{\partial x^2}\Big(D(x,t)\mathcal{P}(x,t)\Big)\,.
\end{equation}
Here the \emph{drift coefficient} $a(x,t)$ is the same function appearing in \cref{ito_1d}, while the \emph{diffusion coefficient} $D(x,t)$ is given by:
\begin{equation}
D(x,t) = \frac{b^2(x,t)}{2}\,.
\end{equation}
In this context, the stationarity (or \emph{balance}) condition reads $\frac{\partial}{\partial t }\mathcal{P}(x,t) = 0$. A stronger condition - namely \emph{detailed balance} - can be defined by writing \cref{fp_1d} as the continuity equation
\begin{equation}
\frac{\partial}{\partial t }\mathcal{P}(x,t) = -\frac{\partial}{\partial x} J(x,t)
\end{equation}
and requiring the \emph{probability current} $J$ to be zero at any time $t$:
\begin{equation}
J(x,t) = a(x,t)\mathcal{P}(x,t) - \frac{\partial}{\partial x}\Big(D(x,t)\mathcal{P}(x,t)\Big) = 0\,.
\end{equation}
The relation between the It\^{o} and FP equations can be shown by considering a generic function $y=y(x)$ and employing It\^{o}'s lemma. Fixing the time $t$ and taking the average over all the possible configurations of $x$, \cref{ito_chain} becomes:
\begin{equation}\label{proof_fp}
\big\langle \text{d}y \big\rangle = \Big\langle a(x,t)\frac{\text{d}y}{\text{d}x}\Big\rangle\,\text{d}t + \Big\langle\frac{b^2(x,t)}{2}\frac{\text{d}^2 y}{\text{d}x^2}\Big\rangle\,\text{d}t + \Big\langle \frac{\text{d}y}{\text{d}x} b(x,t) \text{d}W\Big\rangle\,.
\end{equation}
In the third term of the (RHS), $\text{d}W = W(t+\text{d}t) - W(t)$ is independent on the remaining part - that only depends on the time $t$ but not on consecutive times - and as a consequence:
\begin{equation}
\Big\langle \frac{\text{d}y}{\text{d}x} b(x,t) \text{d}W\Big\rangle = \Big\langle \frac{\text{d}y}{\text{d}x} b(x,t)\Big\rangle \big\langle \text{d}W\big\rangle = 0\,.
\end{equation}
Therefore \cref{proof_fp} can be rewritten as:
\begin{equation}\label{proof_fp2}
\underbrace{\frac{\text{d}}{\text{d}t} \big\langle y \big\rangle}_{(A)} = \underbrace{\Big\langle a(x,t)\frac{\text{d}y}{\text{d}x}\Big\rangle}_{(B)} + \underbrace{\Big\langle\frac{b^2(x,t)}{2}\frac{\text{d}^2 y}{\text{d}x^2}\Big\rangle}_{(C)}
\end{equation}
The three terms appearing in this equation can be written separately as
\begin{subequations}
\begin{align}
(A) &= \frac{\text{d}}{\text{d}t} \int \text{d}x\, \mathcal{P}(x,t) y(x) = \int \text{d}x\, \frac{\partial\mathcal{P}(x,t)}{\partial t} y(x) \mathcal{P}(x,t)\,, \\
(B) &= \int \text{d}x\,\mathcal{P}(x,t)a(x,t)\frac{\text{d}y}{\text{d}x}
= -\int \text{d}x\,\frac{\partial}{\partial x}\left(\mathcal{P}(x,t)a(x,t)y(x)\right)\,, \label{(B)}\\
(C) &= \int \text{d}x\,\mathcal{P}(x,t)\frac{b^2(x,t)}{2}\frac{\text{d}^2 y}{\text{d}x^2}
= \int \text{d}x\,\frac{\partial^2}{\partial x^2}\left(\mathcal{P}(x,t)\frac{b^2(x,t)}{2}y(x)\right)\,,\label{(C)}
\end{align}
\end{subequations}
where the results in \cref{(B),(C)} come from integrating by parts and assuming that $P(x,t)$ vanishes at the boundaries, which is a necessary condition in order to have a normalizable probability distribution. The initial statement immediately follows by substituting $(A)$, $(B)$ and $(C)$ in \cref{proof_fp2}, since this condition must hold for any $y=y(x)$.
A similar reasoning can be followed starting from the multidimensional It\^{o} \cref{ito_multidim}; in this case, the Fokker-Planck equation that one obtains for the probability density $\mathcal{P} = \mathcal{P}(\mathbf{x},t)$ is
\begin{equation}\label{fp_multidim}
\frac{\partial}{\partial t} \mathcal{P}(\mathbf{x},t) = -\sum_{i=1}^n \frac{\partial}{\partial x_i}\Big(a_i(\mathbf{x},t)\mathcal{P}(\mathbf{x},t)\Big) + \sum_{i,j=1}^n \frac{\partial^2}{\partial x_i \partial x_j} \Big( D_{ij}(\mathbf{x},t)\mathcal{P}(\mathbf{x},t)\Big)\,,
\end{equation}
where here the \emph{diffusion matrix} $\mathbf{D}$ is related to $\mathbf{b}$ via the relation
\begin{equation}
\mathbf{D} = \frac{1}{2}\mathbf{b}\mathbf{b}^T\,,
\end{equation}
or equivalently, written for the single component $ij$:
\begin{equation}
D_{ij} = \frac{1}{2}\sum_k b_{ik}b_{jk}.
\end{equation}
By introducing a probability current vector $\mathbf{J}$, also \cref{fp_multidim} can be written as a continuity equation,
\begin{equation}
\frac{\partial}{\partial t} \mathcal{P}(\mathbf{x},t) = \sum_{i=1}^n \frac{\partial}{\partial x_i} J_i (\mathbf{x},t)\,,
\end{equation}
and in this case the detailed balance condition reads, for any component $i$:
\begin{equation}\label{db_multidim}
J_i (\mathbf{x},t) = a_i(\mathbf{x},t)\mathcal{P}(\mathbf{x},t) - \sum_{j=1}^n \frac{\partial}{\partial x_j } \Big( D_{ij}(\mathbf{x},t)\mathcal{P}(\mathbf{x},t)\Big) = 0\,.
\end{equation}
All these relations are generalized to a tensorial formulation in the derivation of the anisotropic SCR \cref{crescale_aniso} (see \cref{derivation_aniso}).
\chapter{Complete derivations}
This appendix includes calculations and mathematical technicalities that are not reported in the main text.
\section{Full derivation of the anisotropic SCR equations}\label{appendix_derivation}
Starting from the case of a isotropic external stress, we report here the calculations to derive the deterministic part of the anisotropic SCR equations, namely the expression in \cref{dh_det_final}.
Let's consider for the moment the formulation where both positions and momenta are rescaled according to \cref{rescaled_variables_explicit}. The Jacobians of the two changes of variable are respectively
\begin{subequations}
\begin{align}
J(\mathbf{q}_i\mapsto\mathbf{s}_i)_{\alpha k} = \frac{\partial q_i^\alpha}{\partial s_i^k} = h_{\alpha k} \,, \label{J_qs} \\
J(\mathbf{p}_i\mapsto\bm{\pi}_i)_{\alpha k} = \frac{\partial p_i^{\alpha}}{\partial \pi_i^k} = h^{-1}_{k\alpha}\,, \label{J_ppi}
\end{align}
\end{subequations}
and the $N\mathbf{S}T$ distribution as a function of $\{\mathbf{s}_i,\bm{\pi}_i\}$ gains no additional prefactor, as
\begin{subequations}
\begin{align}
\text{d}\mathbf{q}_i\,\text{d}\mathbf{p}_i &= \det \big(\mathbf{J}(\mathbf{q}_i\mapsto\mathbf{s}_i)\big) \text{d}\mathbf{s}_i\, \det\big(\mathbf{J}(\mathbf{p}_i\mapsto\bm{\pi}_i)\big) \text{d}\bm{\pi}_i \\
&= \big(\det \mathbf{h}\big)\big(\det \mathbf{h}^{-1}\big) \text{d}\mathbf{s}_i\, \text{d}\bm{\pi}_i \\
&= \left(\det \mathbf{h h}^{-1}\right) \text{d}\mathbf{s}_i\, \text{d}\bm{\pi}_i \\ &= \text{d}\mathbf{s}_i\, \text{d}\bm{\pi}_i\,.
\end{align}
\end{subequations}
Hence, using the expression of the $N\mathbf{S}T$ distribution as reported in \cref{NPT_anisotropic}, the deterministic part of \cref{target_eqs} can be decomposed as it follows:
\begin{equation}\label{dh_det_terms}
\text{d}h_{\alpha i}^{\text{det}} = \sum_{\beta j} D_{\alpha i\beta j}\Bigg[\underbrace{\frac{\partial \log D_{\alpha i\beta j}}{\partial h_{\beta j}}}_{(a)} + \underbrace{\frac{\partial \log (V^{-2})}{\partial h_{\beta j}}}_{(b)} -\frac{1}{k_B T}\Big(\underbrace{\frac{\partial K}{\partial h_{\beta j}}}_{(c)} +\underbrace{\frac{\partial U}{\partial h_{\beta j}}}_{(d)} +\underbrace{P_0\frac{\partial V}{\partial h_{\beta j}}}_{(e)} \Big)\Bigg]\,\text{d}t
\end{equation}
Let's evaluate these terms one by one.
\begin{subequations}
\begin{align}
(a) &= \sum_{\beta j} D_{\alpha i\beta j}\frac{\partial \log D_{\alpha i\beta j}}{\partial h_{\beta j}} = \sum_{\beta j} \frac{\partial D_{\alpha i\beta j}}{\partial h_{\beta j}}\\
&= \frac{\beta_T k_B T}{3 \tau_p}\sum_{\beta j \eta}
\delta_{\alpha \beta} \frac{\partial }{\partial h_{\beta j}} \left( \frac{1}{V}h_{\eta i}h_{\eta j}\right) \\
&= \frac{\beta_T k_B T}{3 \tau_p}\sum_{j \eta}
\left(-\frac{1}{V^2}\frac{\partial V}{\partial h_{\beta j}}h_{\eta i}h_{\eta j} +\frac{1}{V}\delta_{\alpha \eta}\,\delta_{ji}h_{\eta j} +\frac{1}{V} h_{\eta i}\,\delta_{\alpha \eta}\delta_{jj}\right) \\
&= \frac{\beta_T k_B T}{3 \tau_p}
\left( -\frac{1}{V^2}\sum_{j \eta}V\,h_{j\alpha}^{-1}\,h_{\eta i}h_{\eta j} +\frac{1}{V}h_{\alpha j} +\frac{3}{V} h_{\alpha i}\right) \\
&= \frac{\beta_T k_B T}{3V \tau_p} \left(-h_{\alpha j} + h_{\alpha j} + 3h_{\alpha i}\right)
= \frac{3\beta_T k_B T}{3V \tau_p}h_{\alpha i}\,;
\end{align}
\end{subequations}
\begin{subequations}
\begin{align}
(b) &= \sum_{\beta j} D_{\alpha i\beta j}\frac{(-2)}{V}\frac{\partial V}{\partial h_{\beta j}} = -\frac{2\beta_T k_B T}{3V^2 \tau_p}\sum_{\beta j \eta}\delta_{\alpha\beta}\,h_{\eta i}\,h_{\eta j}\,\text{cof}(\mathbf{h})_{\beta j} \\
&= -\frac{2\beta_T k_B T}{3V^2 \tau_p}\sum_{\eta}h_{\eta i}\,\left(\mathbf{h}\,\text{cof}(\mathbf{h})^T\right)_{\eta\alpha} = -\frac{2\beta_T k_B T}{3V^2 \tau_p}\sum_{\eta}h_{\eta i}\left(\det\mathbf{h}\right)\delta_{\eta\alpha} \\
&= -\frac{2\beta_T k_B T}{3V \tau_p}h_{\alpha i}\,. \label{term_b_derivation}
\end{align}
\end{subequations}
Both to evaluate $(a)$ and $(b)$, Jacobi's formula for the derivative of a matrix determinant has been used:
\begin{equation}\label{jacobi}
\frac{\partial V}{\partial h_{\beta j}} = \frac{\partial \det\mathbf{h}}{\partial h_{\beta j}} = \text{cof}(\mathbf{h})_{\beta j} = (\det\mathbf{h})\,h^{-1}_{j\beta}\,,
\end{equation}
where $\text{cof}(\mathbf{h})$ is the cofactor matrix of $\mathbf{h}$. In order to evaluate $(c)$, let's focus on the derivative of the kinetic energy, which is a function of the rescaled momenta $\bm{\pi}_i$ in the formulation that we are considering:
\begin{subequations}
\begin{align}
\frac{\partial K}{\partial h_{\beta j}} &= \frac{\partial }{\partial h_{\beta j}} \sum_{k=1}^N \frac{1}{2m_k}\sum_{\alpha}\big(p_k^\alpha(\bm{\pi}_k)\big)^2 \\
&= \frac{\partial }{\partial h_{\beta j}} \sum_{k=1}^N \frac{1}{2m_k}\sum_{\alpha l m} h^{-1}_{l\alpha}\,h^{-1}_{m\alpha}\,\pi_k^{l}\,\pi_l^m \\
&= \sum_{k=1}^N \frac{1}{2m_k}\sum_{\alpha l m}\Big[ \frac{\partial h^{-1}_{l\alpha}}{\partial h_{\beta j}}\,h^{-1}_{m\alpha}\,\pi_k^{l}\,\pi_l^m + h^{-1}_{l\alpha}\,\frac{\partial{h^{-1}_{m\alpha}}}{\partial h_{\beta j}}\,\pi_k^{l}\,\pi_l^m \Big] \label{identical_terms}\\
&= \sum_{k=1}^N \frac{1}{m_k}\sum_{\alpha l m} \frac{\partial h^{-1}_{l\alpha}}{\partial h_{\beta j}}\,h^{-1}_{m\alpha}\,\pi_k^{l}\,\pi_l^m \,.\label{kineng_derivative}
\end{align}
\end{subequations}
The last passage is possible since the two terms in \cref{identical_terms} are identical, as it is possible to observe by exchanging the summed indices $i$ and $j$ in the second one. The derivative of the inverse box matrix can be evaluated by using the following property, that holds in general for square matrices:
\begin{equation}
\frac{\partial h^{-1}_{l\alpha}}{\partial h_{\beta j}} = -\sum_{\gamma k} h^{-1}_{l\gamma}\frac{\partial h_{\gamma k}}{\partial h_{\beta j}}h^{-1}_{k\alpha} = -\sum_{\gamma k} h^{-1}_{l\gamma}\,\delta_{\gamma\beta}\,\delta_{k j}\,h^{-1}_{k\alpha} = -h^{-1}_{l\beta}\,h^{-1}_{j\alpha}\,.
\end{equation}
Then, by substituting this last expression in \cref{kineng_derivative}, applying the Kronecker deltas and recomposing the physical momenta one gets:
\begin{equation}
\frac{\partial K}{\partial h_{\beta j}} = -\sum_{k = 1}^N \frac{1}{m_k} p^{\beta}_k \sum_\gamma h^{-1}_{j\gamma}\,p_k^{\gamma}\,.
\end{equation}
We can now evaluate the term $(c)$ appearing in \cref{dh_det_terms}:
\begin{subequations}
\begin{align}
(c) &= \sum_{\beta j}\frac{ D_{\alpha i\beta j}}{k_B T}\frac{\partial K}{\partial h_{\beta j}} = \frac{\beta_T}{3V \tau_p} \sum_{\beta j \eta k \gamma}\delta_{\alpha\beta}\,h_{\eta i}\,h_{\eta j}\,\frac{1}{m_k} p^{\beta}_k\, h^{-1}_{j\gamma}\,p_k^{\gamma} \\
&= \frac{\beta_T}{3V \tau_p} \sum_{\eta k \gamma} h_{\eta i}\,\delta_{\eta \gamma}\,\frac{1}{m_k} p^{\alpha}_k\,p_k^{\gamma} = \frac{\beta_T}{3V \tau_p} \sum_{\gamma}\Big(\sum_k \frac{1}{m_k} p^{\alpha}_k\,p_k^{\gamma}\Big)h_{\gamma i} \\
&= \frac{\beta_T}{3\tau_p} \sum_{\gamma}P_{\text{int},\alpha \gamma}^{(p)}\,h_{\gamma i}\,
\end{align}
\end{subequations}
where $\mathbf{P}_{\text{int}}^{(p)}$ is the kinetic part of the internal pressure tensor. Let's now focus on term $(d)$, starting from the calculation of the derivative of the potential energy. For the sake of simplicity we restrict the derivation to the case of a two-body potential:
\begin{equation}
U = \frac{1}{2}\sideset{}{'}\sum_{l,m =1}^{N} u(|\mathbf{q}_{lm}|) = \frac{1}{2}\sideset{}{'}\sum_{l,m =1}^N u(q_{lm})\,,
\end{equation}
where $\mathbf{q}_{lm} = \mathbf{q}_l - \mathbf{q}_m = \mathbf{h}\,\mathbf{s}_{lm}$ and the symbol ' in the sum is a notation for the constraint $l\neq m$. Then we have
\begin{subequations}
\begin{align}
\frac{\partial U}{\partial h_{\beta j}} &= \frac{1}{2}\sideset{}{'}\sum_{l,m =1}^N \sum_\gamma \frac{\partial q_{lm}^\gamma}{\partial h_{\beta j}}\frac{\partial u({q}_{lm})}{\partial q_{lm}^\gamma}
= \frac{1}{2}\sideset{}{'}\sum_{l,m =1}^N \sum_\gamma \left(\frac{\partial }{\partial h_{\beta j}}\sum_k h_{\gamma k} s_{lm}^k\right) \frac{\partial u({q}_{lm})}{\partial q_{lm}^\gamma} \\
&= \frac{1}{2}\sideset{}{'}\sum_{l,m =1}^N s_{lm}^j \frac{\partial u({q}_{lm})}{\partial q_{lm}^\beta} =
\frac{1}{2}\sideset{}{'}\sum_{l,m =1}^N \left(\sum_\gamma h_{j \gamma}^{-1}\,q_{lm}^\gamma \right)\left(-F_{lm}^\beta\right) \\
&= - V\sum_\gamma h_{j \gamma}^{-1} \Bigg(\frac{1}{2V}\sideset{}{'}\sum_{l,m =1}^N q_{lm}^\gamma \,F_{lm}^\beta\Bigg) = -V\sum_\gamma h_{j \gamma}^{-1}\,P_{\text{int},\beta\gamma}^{(q)} \label{virial_variant} \,,
\end{align}
\end{subequations}
where $\mathbf{F}_{lm} = \mathbf{F}_{l} - \mathbf{F}_{m} = - \frac{\partial u({q}_{lm})}{\partial \mathbf{q}_{lm}}$ and $\mathbf{P}_{\text{int}}^{(q)}$ is the virial part of the internal pressure tensor, since the expression in \cref{virial_variant} can be shown to be equivalent to the second term in \cref{internal_pressure_avgkineng}. We can finally evaluate the term $(d)$ of \cref{dh_det_terms}:
\begin{subequations}
\begin{align}
(d) &= \sum_{\beta j}\frac{D_{\alpha i \beta j}}{k_B T}\frac{\partial U}{\partial h_{\beta j}} = -\frac{\beta_T}{3\tau_p} \sum_{\beta j \eta \gamma}\delta_{\alpha\beta}\,h_{\eta i}\,h_{\eta j}\,h_{j\gamma}^{-1}\,P_{\text{int},\beta\gamma}^{(q)} \\
&= -\frac{\beta_T}{3\tau_p} \sum_{\eta \gamma}\,\delta_{\eta\gamma}\,h_{\eta i}\,P_{\text{int},\alpha\gamma}^{(q)} = -\frac{\beta_T}{3\tau_p} \sum_{ \gamma}\,P_{\text{int},\alpha\gamma}^{(q)}\,h_{\gamma i}\,.
\end{align}
\end{subequations}
We are left with the last term of \cref{dh_det_terms}:
\begin{subequations}
\begin{align}
(e) &= \sum_{\beta j}\frac{D_{\alpha i \beta j}}{k_B T}P_0\frac{\partial V}{\partial h_{\beta j}} = \frac{\beta_T P_0}{3V\tau_p}\sum_{\beta j\eta}\delta_{\alpha\beta}\,h_{\eta i}\,h_{\eta j}\,\text{cof}(h)_{\beta j} \\
&= \frac{\beta_T P_0}{3V\tau_p}\sum_{\eta}h_{\eta i}\left(\sum_j h_{\eta j}\,V\,h_{j\alpha}^{-1}\right) = \frac{\beta_T P_0}{3\tau_p}\sum_{\eta}h_{\eta i}\,\delta_{\eta\alpha} \\
&= \frac{\beta_T P_0}{3\tau_p}h_{\alpha i}\,,
\end{align}
\end{subequations}
where we have used once again the property in \cref{jacobi}. Putting all these terms together, namely
\begin{equation}
\text{d}h_{\alpha i}^{\text{det}} = \big[(a) + (b) - (c) - (d) - (e)\big]\,\text{d}t\,,
\end{equation}
the result that one gets is the one in \cref{dh_det_final}.
Let's now consider the formulation where the components of $\mathbf{h}$ are propagated at constant rescaled positions $\mathbf{s}_i$ and physical momenta $\mathbf{p}_i$. The $N\mathbf{S}T$ distribution as a function of this variables acquires in this case an additional prefactor, since the Jacobian in \cref{J_ppi} is no more present and it does not cancel the one in \cref{J_qs}. Hence the distribution to be considered is:
\begin{subequations}
\begin{align}
\mathcal{P}_{N\mathbf{S}T}\big(\{\mathbf{s}_i,\mathbf{p}_i \},V \big) &= \left(\prod_{i=1}^N \det \mathbf{J}(\mathbf{q}_i\mapsto\mathbf{s}_i)\right)\, \mathcal{P}_{N\mathbf{S}T}\big(\{\mathbf{q}_i(\mathbf{s}_i),\mathbf{p}_i \},V \big) \\
&= V^N\,\mathcal{P}_{N\mathbf{S}T}\big(\{\mathbf{q}_i(\mathbf{s}_i),\mathbf{p}_i \},V \big)
\end{align}
\end{subequations}
Then the derivative of the kinetic energy - i.e. the term $(c)$ in the previous derivation - is now zero, while an additional term $(f)$ appears in \cref{dh_det_terms} as a consequence of the new factor $V^N$:
\begin{subequations}
\begin{align}
(f) &= \sum_{\beta j} D_{\alpha i \beta j} \frac{\partial\log\left(V^N\right)}{\partial h_{\beta j}} = \sum_{\beta j} \frac{\beta_T k_B T}{3V\tau_p}\delta_{\alpha\beta}\sum_\eta h_{\eta i}\,h_{\eta j} \frac{N}{V} \frac{\partial V}{\partial h_{\beta j}} \\
&= \frac{N \beta_T k_B T}{3V^2\tau_p} \sum_{j\eta}h_{\eta i}\,h_{\eta j}\,\text{cof}(\mathbf{h})_{\alpha j} = \frac{N \beta_T k_B T}{3V^2\tau_p} \sum_{\eta}h_{\eta i}\sum_j h_{\eta j}\,V\,h_{j\alpha}^{-1} \\
&= \frac{N \beta_T k_B T}{3V\tau_p} \sum_{\eta}h_{\eta i}\,\delta_{\eta\alpha} = \frac{N \beta_T k_B T}{3V\tau_p} h_{\alpha i}\,.
\end{align}
\end{subequations}
As a result, \cref{dh_det_final} remains the same except for the expression of the internal pressure $\mathbf{P}_{\text{int}}$, where the kinetic energy tensor $\mathbf{K}$ is replaced by the diagonal tensor $\frac{N k_B T}{V}\mathbf{I}$, as stated in \cref{derivation_aniso}.
\subsection*{Derivation with a generic external stress}\label{derivation_strain}
If we consider the $N\mathbf{S}T$ dsitribution defined in \cref{NPT_anisotropic_shear}, the additional strain energy brings the following contribution to the deterministic part of the equations:
\begin{equation}\label{dh_strain}
\text{d}h_{\alpha i}^{\text{strain}} = -\frac{1}{k_B T}\sum_{\beta j} D_{\alpha i \beta j} \frac{\partial}{\partial h_{\beta j}}\left( \frac{1}{2}\text{Tr}(\bm{\Sigma}\,\mathbf{G})\right)\,\text{d}t\,,
\end{equation}
where we remind that $\mathbf{G}$ and $\bm{\Sigma}$ are defined as:
\begin{subequations}
\begin{align}
\mathbf{G} &= \mathbf{h}^T\mathbf{h} \,,\\
\bm{\Sigma} &= V_0\,\mathbf{h}_0^{-1}\big(\mathbf{S}-P_0\mathbf{I}\big)\left(\mathbf{h}_0^{-1}\right)^T\,.
\end{align}
\end{subequations}
Let's focus on the derivative:
\begin{subequations}
\begin{align}
\frac{\partial}{\partial h_{\beta j}}\Bigg( &\frac{1}{2}\text{Tr}(\bm{\Sigma} \,\mathbf{G})\Bigg) = \frac{1}{2}\frac{\partial}{\partial h_{\beta j}} \sum_{l m \gamma} \Sigma_{l m}\,h_{\gamma l}\,h_{\gamma m} \\
&= \frac{1}{2}\sum_{l m \gamma}\left( \Sigma_{l m}\,\delta_{\gamma \beta}\,\delta_{l j}\,h_{\gamma m} + \Sigma_{l m}\,h_{\gamma l}\,\delta_{\gamma\beta}\,\delta_{m j}\right) \\
&= \frac{1}{2}\left(\sum_{m}\Sigma_{j m}\,h_{\beta m} + \sum_{l}\Sigma_{l j}\,h_{\beta l}\right) \\
&= \sum_{m}\Sigma_{j m}\,h_{\beta m}\,.
\end{align}
\end{subequations}
To perform the last passage we renamed the summed index $l$ as $m$ in the second sum, and we used the fact that $\bm{\Sigma}$ is a symmetric tensor ($\Sigma_{j m} = \Sigma_{m j}$) as $\mathbf{S}$ is symmetric as well. By substituting in \cref{dh_strain} with the explicit expression of the diffusion tensor we get:
\begin{subequations}
\begin{align}
\text{d}h_{\alpha i}^{\text{strain}} &= -\frac{\beta_T}{3V\tau_p}\sum_{\beta j \eta m} \delta_{\alpha\beta}\,h_{\eta i}\,h_{\eta j}\,\Sigma_{j m}\,h_{\beta m}\,\text{d}t\\
&= -\frac{\beta_T}{3V\tau_p}\sum_{j \eta m}\,h_{\eta i}\,h_{\eta j}\,\Sigma_{j m}\,h_{\alpha m}\,\text{d}t\\
&= -\frac{\beta_T}{3V\tau_p} \big(\mathbf{h}^T\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T \big)_{i\alpha}\,\text{d}t = -\frac{\beta_T}{3V\tau_p} \big(\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T\mathbf{h} \big)_{\alpha i}\,\text{d}t\,,
\end{align}
\end{subequations}
where in the last line we used the property $\big(\mathbf{A}\,\mathbf{B}\big)^T = \mathbf{B}^T\mathbf{A}^T$ and once again the symmetry of $\bm{\Sigma}$. This proves the expression of $\text{d}h_{\alpha i}^{\text{strain}}$ given in \cref{dh_strain_maintext}.
\section{Self-consistency with isotropic SCR equations}\label{appendix_flex2iso}
Let's show the first feature of the anisotropic SCR equations stated in \cref{properties}, namely that they are consistent with the isotropic \cref{crescale_iso_V}. By applying the multidimensional It\^{o}'s lemma in \cref{ito_chain_multidim} to the anisotropic \cref{crescale_aniso} with respect to the variable $y = V(\mathbf{h}) = \det\mathbf{h}$, we get:
\begin{equation}\label{ito_lemma_volume}
\text{d}V = \Big[\underbrace{\left(\nabla_{\mathbf{h}}V\right)^T\mathbf{A}}_{(a)} + \underbrace{\frac{1}{2}\text{Tr}\big(\mathbf{B}^T\,\mathbf{H}_{\mathbf{h}}(V)\,\mathbf{B}\big)}_{(b)}\Big]\,\text{d}t + \underbrace{\left(\nabla_{\mathbf{h}}V\right)^T \mathbf{B}\, \text{d}\mathbf{W}}_{(c)}\,.
\end{equation}
Using Jacobi's formula in \cref{jacobi} the gradient and the Hessian matrix turn out to be:
\begin{subequations}
\begin{align}
\left(\nabla_{\mathbf{h}}V\right)_{\alpha i} &= \frac{\partial V}{\partial h_{\alpha i}} = V\,h_{i \alpha}^{-1}\,,\\
\big(\mathbf{H}_{\mathbf{h}}(V)\big)_{\alpha i\beta j} &= \frac{\partial V}{\partial h_{\alpha i}\partial h_{\beta j}} = V\left(h_{i \alpha}^{-1}\,h_{j \beta}^{-1} - h_{j \alpha}^{-1}\,h_{i \beta}^{-1}\right)\,.
\end{align}
\end{subequations}
Let's evaluate the three terms of \cref{ito_lemma_volume} separately, starting from $(a)$:
\begin{subequations}
\begin{align}
(a) &= -\frac{\beta_T}{3\tau_p}\sum_{\alpha i} V\,h_{i \alpha}^{-1}\Big[\sum_\beta \Big( P_0\delta_{\alpha\beta} - P_{\text{int},\alpha\beta} \Big)h_{\beta i} - \frac{k_B T}{V}h_{\alpha i}\Big] \\
&= -\frac{\beta_T V}{3\tau_p}\sum_{\alpha} \Big[\sum_\beta \Big( P_0\delta_{\alpha\beta} - P_{\text{int},\alpha\beta} \Big)\,\delta_{\alpha\beta} - \frac{k_B T}{V}\,\delta_{\alpha\alpha}\Big] \\
&= -\frac{\beta_T V}{\tau_p} \left(P_0 - P_{\text{int}} - \frac{k_B T}{V}\right)\,.
\end{align}
\end{subequations}
In the last passage we used that $P_{\text{int}} = \text{Tr}\left(\mathbf{P}_{\text{int}}\right)$. The second term of \cref{ito_lemma_volume} is actually zero, in fact:
\begin{subequations}
\begin{align}
(b) &= \frac{1}{2} \frac{2\beta_T k_B T}{3V\tau_p}\sum_{\alpha i\beta j\gamma k}B_{\alpha i\gamma k}\,\big(\mathbf{H}_{\mathbf{h}}(V)\big)_{\alpha i\beta j}\,B_{\beta j \gamma k} \\
&= \frac{\beta_T k_B T}{3V\tau_p}\sum_{\alpha i\beta j\gamma k}h_{\gamma i}\,\delta_{\alpha k}\left(h_{i\alpha}^{-1}\,h_{j\beta}^{-1} - h_{j\alpha}^{-1}\,h_{i\beta}^{-1} \right)h_{\gamma j}\,\delta_{\beta k} \\
&= \frac{\beta_T k_B T}{3V\tau_p}\sum_{\alpha i\gamma j} \left(h_{\gamma i}\,h_{i\alpha}^{-1}\,h_{j\alpha}^{-1}\,h_{\gamma j} - h_{\gamma i}\,h_{i\alpha}^{-1}\,h_{j\alpha}^{-1}\,h_{\gamma j}\right) = 0\,.
\end{align}
\end{subequations}
Finally we can evaluate the term $(c)$ in \cref{ito_lemma_volume}:
\begin{subequations}
\begin{align}
(c) &= \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}}\sum_{\alpha i \beta j} V\,h_{i\alpha}^{-1}\,h_{\beta i}\,\delta_{\alpha j}\,\text{d}W_{\beta j} \\
&= \sqrt{\frac{2\beta_T k_B T V}{3\tau_p}}\sum_{\alpha \beta j} \delta_{\alpha\beta}\,\delta_{\alpha j}\,\text{d}W_{\beta j} \\
&= \sqrt{\frac{2\beta_T k_B T V}{\tau_p}}\frac{\text{Tr}(\text{d}\mathbf{W})}{\sqrt{3}}\,.
\end{align}
\end{subequations}
Since $\text{\text{d}}\mathbf{W}$ is a Wiener noise whose variance is three times that of the single components, $\text{\text{d}}\mathbf{W}/\sqrt{3} = \eta(t)\text{d}t$ where $\eta(t)$ is a Gaussian white noise. Hence we can do the replacement
\begin{equation}
\text{d}W = \frac{\text{Tr}(\text{d}\mathbf{W})}{\sqrt{3}}\,,
\end{equation}
where $\text{d}W$ is the standard Wiener noise as defined in Appendix \ref{appendix_stocdiffeq}.
Putting together the terms $(a)$ and $(b)$ we obtain the isotropic SCR \cref{crescale_iso_V}.
\section{Self-consistency with semi-isotropic SCR equations}\label{appendix_flex2semi-iso}
Also this derivation employs the multidimensional It\^{o} chain rule in \cref{ito_chain_multidim}. The starting equations are actually a modified version of the anisotropic \cref{crescale_aniso}, namely the equations that one obtains with the derivation in Appendix \ref{appendix_derivation} but using the of the $NP_0^\perp \gamma_0 T$ distribution as a function of the box matrix components,
\begin{equation}\label{semi-isotropic-target}
\mathcal{P}_{NP_0^\perp \gamma_0 T}\big(\{\mathbf{q}_i,\mathbf{p}_i\},\mathbf{h}\big) \propto \left(\det \mathbf{h}\right)^{-2} \exp\left[-\frac{1}{k_B T}\Big(K + U + P_0^\perp \det \mathbf{h} - \gamma_0\,A\Big)\right]\,.
\end{equation}
Although the $NP_0^\perp \gamma_0 T$ was introduced considering an orthorhombic box in \cref{npt_theory}, it is possible to consider here a slightly more general case with two additional nonzero off-diagonal elements:
\begin{equation}
\mathbf{h}=
\begin{pmatrix}
h_{x1} & h_{x2} & 0\\
h_{y1} & h_{y2} & 0\\
0 & 0 & h_{z3}
\end{pmatrix}\,.
\end{equation}
Obviously $A = h_{x1}\,h_{y2} - h_{x2}\,h_{y1}$ and $L = h_{z3}$. By repeating the derivation of \cref{derivation_aniso} with \cref{semi-isotropic-target} as target distribution, the same equations are obtained but with two differences, namely the hydrostatic pressure $P_0$ is substituted by the normal pressure $P_0^\perp$ and an additional term with the reference surface tension $\gamma_0$ appears:
\begin{equation}
\text{d}h_{\alpha i}^{\gamma_0} = - \sum_{\beta j}\frac{D_{\alpha i \beta j}}{k_B T}\,\left(-\gamma_0\frac{\partial A}{\partial h_{\beta j}}\right)\text{d}t\,. \\
\end{equation}
Observing that the derivative of $A$ can be written as
\begin{equation}\label{dh_gamma0}
\frac{\partial A}{\partial h_{\beta j}} = \delta_{\beta x}\,\delta_{j1}\,h_{y2} + h_{x1}\,\delta_{\beta y}\,\delta_{j2} - \delta_{\beta x}\,\delta_{j2}\,h_{y1} - h_{x2}\,\delta_{\beta y}\,\delta_{j1}\,,
\end{equation}
we obtain, by substituting this expression in \cref{dh_gamma0} and applying the Kronecker deltas:
\begin{subequations}
\begin{align}
\text{d}h_{\alpha i}^{\gamma_0} &= \frac{\beta_T \gamma_0}{3V\tau_p} \Big[\left(\delta_{\alpha x}\,h_{xi} + \delta_{\alpha y}\,h_{yi}\right)\left(h_{x1}\,h_{y2} - h_{x2}\,h_{y1}\right) \Big] \\
&= \frac{\beta_T \gamma_0 A}{3V\tau_p} \left(\delta_{\alpha x}\,h_{xi} + \delta_{\alpha y}\,h_{yi}\right) =\frac{\beta_T \gamma_0}{3L\tau_p} h_{\alpha i}\left(1-\delta_{\alpha z}\right)\,.
\end{align}
\end{subequations}
Then the modified anisotropic equations for the $NP_0^\perp \gamma_0 T$ ensemble read:
\begin{align}
\text{d}h_{\alpha i} = -&\frac{\beta_T}{3\tau_p} \Bigg[\sum_\beta \Big( P_0^\perp\delta_{\alpha\beta} - P_{\text{int},\alpha\beta} \Big)h_{\beta i} - \frac{k_B T}{V}h_{\alpha i} - \frac{\gamma_0}{L}h_{\alpha i}\left(1-\delta_{\alpha z}\right)\Bigg]\,\text{d}t \nonumber \\
&+ \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_{\beta}h_{\beta i}\,\text{d}W_{\alpha\beta}\,.
\end{align}
Starting from these equations, let's derive the semi-isotropic ones for the variables $\varepsilon_{xy} = \log A/A_0$ and $\varepsilon_z = \log L/L_0$, namely \cref{crescale-semi-isotropic-eq}. The equation for $L$ is simply the one for $h_{3z}$:
\begin{equation}
\text{d}L = \underbrace{-\frac{\beta_T L}{3\tau_p} \left( P_0^\perp - P_{\text{int},zz} - \frac{k_B T}{V}\right)}_{a(L)}\,\text{d}t + \underbrace{\sqrt{\frac{2\beta_T k_B T}{3V\tau_p}}\,L}_{b(L)}\,\text{d}W_{zz}\,.
\end{equation}
Then, by applying It\^{o}'s lemma for $\varepsilon_z$, we find the following SDE:
\begin{equation}
\text{d}\varepsilon_z = \left(\frac{\partial \varepsilon_z}{\partial L}\,a + \frac{b^2}{2}\frac{\partial^2 \varepsilon_z}{\partial L^2} \right)\text{d}t + b\frac{\partial \varepsilon_z}{\partial L}\,\text{d}W \\
\end{equation}
Using that $\frac{\partial \varepsilon_z}{\partial L} = \frac{1}{L}$ and $\frac{\partial^2 \varepsilon_z}{\partial L^2} = -\frac{1}{L^2}$ we finally obtain
\begin{equation}
\text{d}\varepsilon_z = -\frac{\beta_T}{3\tau_p}\left(P_0^\perp - P_{\text{int},zz}\right)\text{d}t + \sqrt{\frac{2 k_B T \beta_T}{3 V \tau_p}}\text{d}W_z\,,
\end{equation}
which is exactly \cref{eps_xy}. In order to obtain the second equation, we can first isolate the four equations that evolve the box matrix components entering in $A$:
\begin{equation}
\text{d}\mathbf{h}^{(xy)} = \left(\mathbf{A}^{(xy)} + \frac{\gamma_0\beta_T}{3L\tau_p}\mathbf{I}_{2\times2}\right)\,\mathbf{h}^{(xy)}\,\text{d}t + \mathbf{B}^{(xy)}\,\text{d}\mathbf{W}^{(xy)}\,.
\end{equation}
Here the upperscript $(xy)$ identifies the 2$\times$2 upper-left submatrices, and $A$ can be written as $A = \det \mathbf{h}^{(xy)}$. The equation for $A$ can be obtained by applying the multidimensional It\^{o} chain rule to \cref{ito_chain_multidim}, performing the same calculations shown in Appendix \ref{appendix_flex2iso} to derive the equation for $V$, but in two dimensions instead of three, and with the additional diagonal term containing $\gamma_0$. The equation that one obtains is:
\begin{equation}
\text{d}A = -\frac{2A\beta_T}{3\tau_p}\left(P_0^\perp - \frac{P_{\text{int},xx}+P_{\text{int},yy}}{2} - \frac{k_B T}{V} - \frac{\gamma_0}{L}\right)\,\text{d}t + A\sqrt{\frac{4 k_B T \beta_T}{3 V \tau_p}}\,\text{d}W\,,
\end{equation}
where here the Wiener noise comes from $\text{d}W = (\text{d}W_{xx} + \text{d}W_{yy})/\sqrt{2}$. By applying It\^{o}'s lemma for the variable $\varepsilon_{xy}$ as already done for $\varepsilon_z$, we finally get
\begin{equation}
\text{d}\varepsilon_{xy} = -\frac{2\beta_T}{3\tau_p}\left(P_0^\perp - \frac{\gamma_0}{L} - \frac{P_{\text{int},xx} + P_{\text{int},yy}}{2} \right)\text{d}t + \sqrt{\frac{4 k_B T \beta_T}{3 V \tau_p}}\text{d}W_{xy}\,,
\end{equation}
namely \cref{eps_xy}.
\section{Change of box vectors}\label{appendix_changecellvectors}
We show here that the anisotropic SCR equations are invariant under the transformation
\begin{equation}\label{change_box}
h_{\alpha i} \longmapsto h'_{\alpha i} = \sum_{j=1}^3 n_j^{(i)}\,h_{\alpha j}\,.
\end{equation}
As shown in \cref{derivation_aniso}, the equations can be written as
\begin{equation}
\text{d}\mathbf{h} = \mathbf{A}(\mathbf{h})\, \text{d}t + \mathbf{B}(\mathbf{h})\,\text{d}\mathbf{W}\,,
\end{equation}
where $\mathbf{B}$ is defined in \cref{B_tensor} and $\mathbf{A}$ is given by:
\begin{equation}
\mathbf{A} = -\frac{\beta_T}{3\tau_p} \Big[ \big( P_0\mathbf{I} - \mathbf{P}_{\text{int}} \big) - \frac{k_B T}{V}\mathbf{I} \Big]\,\mathbf{h}\,.
\end{equation}
Once again the demonstration is based on the multidimensional It\^{o}'s chain rule, applied to the equations above with respect to each transformed variable $h'_{\alpha i}$:
\begin{equation}\label{dh_prime}
\text{d}h'_{\alpha i} = \Big[\underbrace{\left(\nabla_{\mathbf{h}}h'_{\alpha i}\right)^T\mathbf{A}}_{(a)} + \underbrace{\frac{1}{2}\text{Tr}\Big(\mathbf{B}^T\,\mathbf{H}_{\mathbf{h}}\left(h'_{\alpha i}\right)\,\mathbf{B}\Big)}_{(b)}\Big]\,\text{d}t + \underbrace{\left(\nabla_{\mathbf{h}}h'_{\alpha i}\right)^T \mathbf{B}\, \text{d}\mathbf{W}}_{(c)}\,,
\end{equation}
Since the transformation \cref{change_box} is linear, the first derivatives are
\begin{equation}
\left(\nabla_{\mathbf{h}}h'_{\alpha i}\right)_{\beta j} = \frac{\partial h'_{\alpha i}}{\partial h_{\beta j}} = n_j^{(i)}\delta_{\alpha\beta} \,,
\end{equation}
while the second derivatives are zero:
\begin{equation}
\big(\mathbf{H}_{\mathbf{h}}(h'_{\alpha i})\big)_{\beta j\gamma k} = \frac{\partial h'_{\alpha i}}{\partial h_{\beta j}\partial h_{\gamma k}} = 0\,.
\end{equation}
As a consequence, $(b) = 0$. Let's evaluate the other two terms:
\begin{align}
(a) &= -\frac{\beta_T}{3\tau_p} \sum_{\beta j} n_j^{(i)}\,\delta_{\alpha\beta}\, \Bigg[\sum_\gamma \Big( P_0\delta_{\beta\gamma} - P_{\text{int},\beta\gamma} \Big)h_{\gamma j} - \frac{k_B T}{V}h_{\beta j}\Bigg]\nonumber \\
&= -\frac{\beta_T}{3\tau_p} \left[\sum_{\gamma } \Big( P_0\delta_{\alpha\gamma} - P_{\text{int},\alpha\gamma} \Big)\left(\sum_{j}n_j^{(i)}\,h_{\gamma j}\right) - \frac{k_B T}{V}\sum_{j}n_j^{(i)}h_{\alpha j}\right] \nonumber \\
&= -\frac{\beta_T}{3\tau_p} \Bigg[\sum_\gamma \Big( P_0\delta_{\alpha\gamma} - P_{\text{int},\alpha\gamma} \Big)h'_{\gamma i} - \frac{k_B T}{V}h'_{\alpha i}\Bigg]
\end{align}
\begin{align}
(c) &= \sum_{\beta j\gamma k} n_j^{(i)}\,\delta_{\alpha\beta} \,B_{\beta j\gamma k}\,\text{d}W_{\gamma k} = \sum_{j}n_j^{(i)}\,\sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_\gamma h_{\gamma j}\,\text{d}W_{\alpha\gamma} \nonumber\\
&= \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_\gamma\left(\sum_{j}n_j^{(i)}\, h_{\gamma j}\right)\text{d}W_{\alpha\gamma} =\sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_\gamma h'_{\gamma j}\text{d}W_{\alpha\gamma}
\end{align}
Then the transformation $\mathbf{h}\mapsto\mathbf{h}'$ does not change the form of the anisotropic SCR equations.
\section{Method for QR factorization}\label{appendix_rotations}
Given a rescaling matrix $\bm{\mu}$ with all non-zero elements, obtained by propagating nine degrees of freedom according to \cref{mu_matrix}, let's show the procedure to rotate the columns of $\mu$ in order to obtain an upper-triangular matrix $\bm{\mu'}$:
\begin{equation}
\bm{\mu'}= \mathbf{R}\bm{\mu} =
\begin{pmatrix}
\mu^{\prime}_{xx} & \mu^{\prime}_{xy} & \mu^{\prime}_{xz}\\
0 & \mu^{\prime}_{yy} & \mu^{\prime}_{yz}\\
0 & 0 & \mu^{\prime}_{zz}
\end{pmatrix}\,.
\end{equation}
The rotation of the first column is simply achieved by imposing
\begin{subequations}\label{rot_column1}
\begin{align}
\mu^{\prime\,2}_{xx} &= \mu_{xx}^2 + \mu_{yx}^2 + \mu_{zx}^2\,, \\
\mu^{\prime\,2}_{yx} &= \mu^{\prime\,2}_{zx} = 0\,.
\end{align}
\end{subequations}
To rotate the second column of $\bm{\mu}$ we have to take into account the invariance both of the norm and of the scalar product with the first column:
\begin{subequations}\label{rot_column2}
\begin{align}
\mu^{\prime\,2}_{xy} + \mu^{\prime\,2}_{yy} &= \mu_{xy}^2 + \mu_{yy}^2 + \mu_{zy}^2\,, \\
\mu^{\prime}_{xy}\,\mu^{\prime}_{xx} &= \mu_{xy}\,\mu_{xx} + \mu_{yy}\,\mu_{yx} + \mu_{zy}\,\mu_{zx}\,, \\
\mu^{\prime}_{zy} &= 0\,.
\end{align}
\end{subequations}
Finally, the conditions to rotate the third column include the invariance of the scalar product both with the first and the second column:
\begin{subequations}\label{rot_column3}
\begin{align}
\mu^{\prime\,2}_{xz} + \mu^{\prime\,2}_{yz} + \mu^{\prime\,2}_{zz} &= \mu_{xz}^2 + \mu_{yz}^2 + \mu_{zz}^2\,, \\
\mu^{\prime}_{xz}\,\mu^{\prime}_{xx} &= \mu_{xz}\,\mu_{xx} + \mu_{yz}\,\mu_{yx} + \mu_{zz}\,\mu_{zx}\,, \\
\mu^{\prime}_{xz}\,\mu^{\prime}_{xy} + \mu^{\prime}_{yz}\,\mu^{\prime}_{yy} &= \mu_{xz}\,\mu_{xy} + \mu_{yz}\,\mu_{yy} + \mu_{zz}\,\mu_{zy}\,.
\end{align}
\end{subequations}
By solving together \cref{rot_column1,rot_column2,rot_column3} one gets the six non-zero elements of $\bm{\mu'}$:
\begin{subequations}
\begin{align}
\mu^{\prime}_{xx} &= \sqrt{\mu_{xx}^2 + \mu_{yx}^2 + \mu_{zx}^2} \,, \\
\mu^{\prime}_{xy} &= \frac{\mu_{xy}\,\mu_{xx} + \mu_{yy}\,\mu_{yx} + \mu_{zy}\,\mu_{zx}}{\mu^{\prime}_{xx}}\,, \\
\mu^{\prime}_{yy} &= \sqrt{\mu_{xy}^2 + \mu_{yy}^2 + \mu_{zy}^2 - \mu^{\prime\,2}_{xy}}\,, \\
\mu^{\prime}_{xz} &= \frac{\mu_{xz}\,\mu_{xx} + \mu_{yz}\,\mu_{yx} + \mu_{zz}\,\mu_{zx}}{\mu^{\prime}_{xx}}\,, \\
\mu^{\prime}_{yz} &= \frac{\mu_{xz}\,\mu_{xy} + \mu_{yz}\,\mu_{yy} + \mu_{zz}\,\mu_{zy} - \mu^{\prime}_{xz}\,\mu^{\prime}_{xy}}{\mu^{\prime}_{yy}}\,, \\
\mu^{\prime}_{zz} &= \sqrt{\mu_{xz}^2 + \mu_{yz}^2 + \mu_{zz}^2 -\mu^{\prime\,2}_{xz} - \mu^{\prime\,2}_{yz}}\,.
\end{align}
\end{subequations}
As a final observation, in the GROMACS \cite{gromacs} implementation of the anisotropic Berendsen barostat, the same operation is performed using a first order approximation of the equations above, namely
\begin{equation}
\bm{\mu'}=
\begin{pmatrix}
\mu_{xx} & \mu_{xy} + \mu_{yx} & \mu_{xz} + \mu_{zx}\\
0 & \mu_{yy} & \mu_{yz} +\mu_{zy}\\
0 & 0 & \mu_{zz}
\end{pmatrix}\,.
\end{equation}
\section{High-friction limit of Parrinello-Rahman equations}\label{appendix_highfriction}
We report here the main calculations to derive the anisotropic SCR \cref{crescale_aniso} from the extended Parrinello-Rahman \cref{PR_friction_eqs}.
The limit taken into account is known as \emph{Smoluchowski-Kramers limit} in the field of SDEs, and it is performed sending the friction to infinity and the mass to zero, such that their product stays finite. A detailed discussion of this limit in case of a variable-dependent friction is reported in \cite{high_friction_limit}. To apply the limit, we first rewrite the extended Parrinello-Rahman equations as
\begin{subequations}\label[pluralequation]{PR_plusfriction}
\begin{align}
\text{d}\mathbf{h} &= \mathbf{v}\,\text{d}t\,, \\
\text{d}\mathbf{v} &= \left(\frac{\mathbf{F}(\mathbf{h})}{W} - \frac{\bm{\gamma}(\mathbf{h})}{W}\mathbf{v} \right)\,\text{d}t + \frac{\bm{\sigma}(\mathbf{h})}{W}\,\,\text{d}\mathbf{W}\,,
\end{align}
\end{subequations}
where $\mathbf{F}(\mathbf{h}) = V\big(\mathbf{P}_{\text{int}} - P_0\mathbf{I}\big)\big(\mathbf{h}^{-1}\big)^T$. In case of a variable-dependent friction \cite{high_friction_limit}, the result of the limit is
\begin{equation}\label{hf_limit}
\text{d}\mathbf{h} = \Big(\underbrace{\bm{\gamma}^{-1}(\mathbf{h})\,\mathbf{F}(\mathbf{h})}_{(a)} + \underbrace{\mathbf{T}(\mathbf{h})}_{(b)}\Big)\,\text{d}t + \underbrace{\bm{\gamma}^{-1}(\mathbf{h})\,\bm{\sigma}(\mathbf{h})\,\text{d}\mathbf{W}}_{(c)}\,,
\end{equation}
with the \emph{noise-induced drift tensor} $\mathbf{T}(\mathbf{h})$ determined via the following relations:
\begin{subequations}
\begin{align}
&T_{\alpha i} = \sum_{\beta j\gamma k}\left(\frac{\partial}{\partial h_{\beta j}}\gamma_{\alpha i\gamma k}^{-1} \right)\,J_{\gamma k \beta j}\,, \\
&\mathbf{J}\bm{\gamma}^T + \bm{\gamma}\mathbf{J} = \bm{\sigma}\bm{\sigma}^T\,. \label{lyapunov}
\end{align}
\end{subequations}
\Cref{lyapunov} is called \emph{Lyapunov equation} for $\mathbf{J}$. Let's start by considering the term $(a)$:
\begin{subequations}
\begin{align}
(a)_{\alpha i} &= \sum_{\beta j}\gamma^{-1}_{\alpha i\beta j}\,F_{\beta j} \\
&= \sum_{\beta j}\gamma^{-1}_{\alpha i\beta j}\,V\sum_{\gamma}\left(P_{\text{int},\beta\gamma} - P_0\,\delta_{\beta\gamma}\right)h_{j\gamma}^{-1}
\end{align}
\end{subequations}
We can now arbitrary set the following functional form for $\bm{\gamma}^{-1}$:
\begin{equation}\label{gamma_inverse}
\gamma_{\alpha i\beta j}^{-1} = \frac{\beta_T}{3V\tau_p}\delta_{\alpha\beta}\sum_{\eta}h_{\eta i}\,h_{\eta j}\,.
\end{equation}
In spirit, this \emph{ansatz} is similar to the one for the diffusion tensor in the derivation of \cref{derivation_aniso}, namely it breaks the generality of the equations in order to reproduce a Berendsen-like deterministic term, but without affecting the sampled distribution.
Substituting $\bm{\gamma}^{-1}$ in the previous expression we get:
\begin{subequations}
\begin{align}
(a)_{\alpha i} &= \frac{\beta_T V}{3V\tau_p}\sum_{\beta j \gamma \eta}\delta_{\alpha\beta}h_{\eta i}\,h_{\eta j}\,h_{j\gamma}^{-1}\left(P_{\text{int},\beta\gamma} - P_0\,\delta_{\beta\gamma}\right) \\
&= -\frac{\beta_T}{3\tau_p}\sum_{\gamma} \left(P_0\,\delta_{\alpha\gamma} - P_{\text{int},\alpha\gamma}\right)h_{\gamma i} \,.
\end{align}
\end{subequations}
Hence we have recovered the first term in the deterministic part of the anisotorpic SCR equations. Let's now focus on the term (b). To compute the drift tensor $\mathbf{T}$ we first need to solve the Lyapunov equation \cref{lyapunov}, where both $\bm{\gamma}$ and $\bm{\sigma}$ appear. The tensor $\bm{\gamma}$ is obtained by inverting \cref{gamma_inverse}:
\begin{equation}\label{gamma}
\gamma_{\alpha i\beta j} = \frac{3V\tau_p}{\beta_T}\delta_{\alpha\beta}\sum_{\eta}h_{i\eta}^{-1}\,h_{j\eta}^{-1}\,.
\end{equation}
The tensor $\bm{\sigma}$ is obtained instead by imposing the fluctuation-dissipation theorem in \cref{fluct_diss_multidim}, resulting in:
\begin{equation}
\sigma_{\alpha i\beta j} = \sqrt{\frac{6 V\tau_p k_B T}{\beta_T}}\,h_{i\beta}^{-1}\delta_{\alpha j}\,.
\end{equation}
With these expressions for $\bm{\gamma}$ and $\bm{\sigma}$, the solution of the Lyapunov equation turns out to be
\begin{equation}
J_{\alpha i \beta j} = k_B T \delta_{\alpha\beta}\,\delta_{ij}\,.
\end{equation}
We can now evaluate the term $(b)$:
\begin{subequations}\label{almost_finished}
\begin{align}
(b)_{\alpha i} &= T_{\alpha i} = \sum_{\beta j\gamma k}\left(\frac{\partial}{\partial h_{\beta j}}\gamma_{\alpha i\gamma k}^{-1} \right)\,J_{\gamma k \beta j} = k_B T\sum_{\gamma k}\left(\frac{\partial}{\partial h_{\gamma k}}\gamma_{\alpha i\gamma k}^{-1} \right)\,.
\end{align}
\end{subequations}
Performing the calculations with Jacobi's formula in \cref{jacobi} one finds for the derivative
\begin{equation}
\frac{\partial}{\partial h_{\gamma k}}\gamma_{\alpha i\gamma k}^{-1} = \frac{1}{V}\left( -h_{k\gamma}^{-1}\,\delta_{\alpha\gamma}\sum_\eta h_{\eta i}\,h_{\eta k} + \delta_{\alpha\gamma}\,\delta_{k i}\,h_{\gamma k} + \delta_{\alpha\gamma}\,\delta_{kk}\,h_{\gamma i}\right) \,,
\end{equation}
and substituting in \cref{almost_finished}:
\begin{equation}
(b)_{\alpha i} = \frac{\beta_T}{3\tau_p}\frac{3 k_B T}{V}\,h_{\alpha i}\,.
\end{equation}
Finally, the last term is:
\begin{subequations}
\begin{align}
(c)_{\alpha i} &= \sum_{\beta j\gamma k}\gamma_{\alpha i\beta j}^{-1}\,\sigma_{\beta j\gamma k}\,\text{d}W_{\gamma k} \\
&= \sum_{\beta j\gamma k}\frac{\beta_T}{3V\tau_p}\delta_{\alpha\beta}\sum_\eta h_{\eta i}\,h_{\eta j}\sqrt{\frac{6V\tau_p k_B T}{\beta_T}}\,h_{\gamma j}^{-1}\,\delta_{\beta k}\,\text{d}W_{\gamma k} \\
&= \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}}\sum_\gamma h_{\gamma i}\,\text{d}W_{\alpha \gamma}\,.
\end{align}
\end{subequations}
Evaluating $\left[(a) + (b)\right]\,\text{d}t + (c)$ we obtain the equations
\begin{equation}
\text{d}h_{\alpha i} = -\frac{\beta_T}{3\tau_p} \Bigg[\sum_\beta \Big( P_0\delta_{\alpha\beta} - P_{\text{int},\alpha\beta} \Big)h_{\beta i} - \frac{3 k_B T}{V}h_{\alpha i}\Bigg]\,\text{d}t + \sqrt{\frac{2\beta_T k_B T}{3V\tau_p}} \sum_{\beta}h_{\beta i}\,\text{d}W_{\alpha\beta}\,,
\end{equation}
which perfectly match the anisotropic SCR \cref{crescale_aniso} except for the additional term
\begin{equation}
\text{d}\mathbf{h}^{PR} = \frac{\beta_T}{3\tau_p}\frac{2 k_B T}{V}\mathbf{I}\,\text{d}t\,,
\end{equation}
as claimed in \cref{limit_PR}. Note that these equations can be obtained with the same derivation of the anisotropic SCR equations outlined in Appendix \ref{appendix_derivation}, but employing as target distribution
\begin{equation}
\mathcal{P}'_{N\mathbf{S}T}\big(\{\mathbf{q}_i,\mathbf{p}_i\},\mathbf{h}\big) \propto \exp\left[-\frac{1}{k_B T}\Big(K + U + P_0 \det \mathbf{h}\Big)\right]\,,
\end{equation}
namely neglecting the factor $\left(\det \mathbf{h}\right)^{-2}$ that was instead included in \cref{NPT_anisotropic}. In fact, neglecting this factor is equivalent to omit the term $(b)$ computed in \cref{term_b_derivation}, which is exactly the additional term appearing in this derivation but changed of sign.
In summary, the anisotropic SCR barostat can be seen as the Parrinello-Rahman barostat plus a Langevin thermostat applied to the components of $\mathbf{h}$, with a $\mathbf{h}$-dependent friction tensor $\bm{\gamma} = \bm{\gamma}(\mathbf{h},W)$ defined as in \cref{gamma} and in the high-friction and zero-mass limit described in \cite{high_friction_limit}.
\subsection*{High-friction limit with a generic external stress}
If we consider the most general case $\mathbf{S}\neq P_0\mathbf{I}$, the tensor $\mathbf{F}(\mathbf{h})$ in \cref{PR_plusfriction} gains an additional term, namely it has to be substituted by
\begin{equation}
\tilde{\mathbf{F}}(\mathbf{h}) = \mathbf{F}(\mathbf{h}) - \mathbf{h}\,\bm{\Sigma}\,.
\end{equation}
Taking the high-friction limit, this term contributes only to the term $(a)$ in \cref{hf_limit}:
\begin{subequations}
\begin{align}
\text{d}h^{\text{strain}}_{\alpha i} &= -\sum_{\beta j}\gamma_{\alpha i\beta j}^{-1}\left(\mathbf{h}\bm{\Sigma}\right)_{\beta j}\,\text{d}t \\
&= -\sum_{\beta j}\frac{\beta_T}{3V\tau_p}\delta_{\alpha\beta}\sum_{\gamma}h_{\gamma i}\,h_{\gamma j}\sum_k h_{\beta k}\,\Sigma_{kj}\,\text{d}t \\
&= -\frac{\beta_T}{3V\tau_p}\sum_{\gamma j k},h_{\alpha k}\,\Sigma_{kj}\,h_{\gamma j}\,h_{\gamma i}\,\text{d}t \\
&= -\frac{\beta_T}{3V\tau_p}\left(\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T \mathbf{h} \right)_{\alpha i}\,\text{d}t\,.
\end{align}
\end{subequations}
This is exactly the additional term that appears in \cref{crescale_eqs_strain} in presence of a generic external stress.
\chapter{Effective energy drift}\label{appendix_effenergy}
Let's consider a sampling algorithm based on a differential equation that satisfies the detailed balance condition with respect to the distribution $\mathcal{P}(x)$:
\begin{equation}
\mathcal{P}(x)\Pi(x\rightarrow x') = \mathcal{P}(x')\Pi(x'\rightarrow x)\,.
\end{equation}
Here, $\Pi(x\rightarrow x')$ is the transition probability of moving to $x'$ starting from $x$.
If the variable $x$ is a point in phase space, this condition has to be substituted with the generalized detailed balance described in \cref{hybrid_MC}.
When the equation satisfying this condition is integrated approximately - using for instance a finite time step propagation of the variable $x$ - detailed balance is violated. The amount of this violation, namely how much the ratio
\begin{equation}
\frac{\mathcal{P}(x')\Pi(x'\rightarrow x)}{\mathcal{P}(x)\Pi(x\rightarrow x')}
\end{equation}
moves away from $1$, can be used to evaluate the quality of the integration, in order to understand on the fly if the time step or other parameters of the integration algorithm were chosen correctly. The quantity that can be introduced with this purpose is the \emph{effective energy} $\widetilde{H}$ \cite{langevin_bussi}, defined through its finite increments:
\begin{equation}\label{effective_energy}
\widetilde{H}(t+\Delta t) - \widetilde{H}(t) = -k_B T \log\frac{\mathcal{P}(x')\Pi(x'\rightarrow x)}{\mathcal{P}(x)\Pi(x\rightarrow x')}\,.
\end{equation}
Clearly the transition probabilities embedded in $\Pi$ depend on the specific sampling algorithm.
Summing consecutive increments $\Delta \tilde{H}$ one typically observes a stationary \emph{effective energy drift}, whose slope increases with the time step. Hence, the effective energy plays the role of a conserved quantity whose conservation law is violated for any finite time step, and that can be monitored to detect problems in the simulation. As an example, considering the microcanonical ensemble and the velocity Verlet integrator, the effective energy is just the total energy of the system.
The effective energy variations $\Delta \widetilde{H}$ can alternatively be used to implement accept-reject algorithms using the Metropolis-Hastings rule, where the acceptance $\alpha$ is computed as
\begin{equation}
\alpha = \text{min}\left[1,\text{exp}(-\Delta \widetilde{H})\right]\,;
\end{equation}
in this way, finite step errors are by construction corrected without the need of changing the time step. This scheme defines what are typically called \emph{Metropolized integrators} \cite{metropolized_integrator}.
\section{Derivation of isotropic energy drift for SCR}\label{appendix:iso_effeng}
We discuss here the isotropic contribution to the effective energy drift associated to the time-reversible integrator of \cref{sec:TR}. The derivation is the same reported in \cite{crescale_iso}, except for the additional strain energy $E_s = \frac{1}{2}\text{Tr}(\bm{\Sigma}\mathbf{G})$ included in the target distribution. Indeed, in the limit of small $\Delta t$ it is possible to show that \cref{eq:lambda_equivalent_1+3} samples the volume distribution
\begin{equation}
\mathcal{P}_{1+3}(V) \propto\exp\left[-\frac{1}{k_B T}\Big(K + U + P_0 V + E_s\Big)\right]\,,
\end{equation}
or equivalently, as a function of $\lambda$:
\begin{equation}\label{eq:distr_1+3}
\mathcal{P}_{1+3}(\lambda) \propto\lambda \exp\left[-\frac{1}{k_B T}\Big(K + U + P_0 \lambda^2 + E_s\Big)\right]\,.
\end{equation}
In \cref{eq:distr_1+3} the additional factor $\lambda$ comes from the Jacobian of the change of variable. By condensing the two half-steps in a single step of size $\Delta t$, according to \cref{eq:lambda_equivalent_1+3}, we can write the forward and backward moves as
\begin{subequations}
\begin{align}
&\lambda^{i+1} = \lambda^i + \frac{D_\lambda}{k_B T}f({\lambda}^i) \Delta t + \sqrt{2D_\lambda \Delta t}\,\mathcal{R}^i\,,\label{eq:lambda_forward}\\
&\lambda^{i} = \lambda^{i+1} + \frac{D_\lambda}{k_B T}f({\lambda}^{i+1}) \Delta t + \sqrt{2D_\lambda \Delta t}\,\mathcal{R}^{i+1}\,,\label{eq:lambda_backward}
\end{align}
\end{subequations}
where $D_\lambda = \frac{k_B T\beta_T}{4\tau_p}$ and $f(\lambda) = -2\lambda\left(P_0 - P_{\text{int}} - \frac{k_BT}{2\lambda^2} + \frac{\text{Tr}\left(\mathbf{h}\,\bm{\Sigma}\,\mathbf{h}^T\right)}{3\lambda^2}\right)$. Then, the isotropic contribution to the effective energy drift is given by:
\begin{equation}\label{eq:1+3_effective_energy}
\Delta\widetilde{H}_{i\rightarrow i+1} = -k_B T \log\frac{\mathcal{P}_{1+3}(\lambda^{i+1})\Pi(\lambda^{i+1}\rightarrow \lambda^i)}{\mathcal{P}_{1+3}(\lambda^i)\Pi(\lambda^i\rightarrow \lambda^{i+1})}\,.
\end{equation}
The part due to the $\mathcal{P}_{1+3}$ probabilities is computed as:
\begin{subequations}
\begin{align}
\Delta\widetilde{H}_{i\rightarrow i+1}^\mathcal{P} = -k_B T \log\frac{\mathcal{P}_{1+3}(\lambda^{i+1})}{\mathcal{P}_{1+3}(\lambda^i)}
= \Delta K + \Delta U + \Delta E_s + P_0\Delta\lambda^2 - k_B T\Delta\log\lambda \,.
\end{align}
\end{subequations}
Recalling that $R^i$ and $R^{i+1}$ are zero-mean and unit-variance Gaussian numbers, the forward and backward transition probabilities have the following expressions:
\begin{subequations}
\begin{align}
\Pi(\lambda^i\rightarrow \lambda^{i+1}) &= \frac{1}{\sqrt{2\pi}}e^{- \frac{(\mathcal{R}^i)^2}{2}}\frac{\text{d}\mathcal{R}^i}{\text{d}\lambda^{i+1}}\,, \label{eq:lambda_forward_pi}\\
\Pi(\lambda^{i+1}\rightarrow \lambda^i) &= \frac{1}{\sqrt{2\pi}}e^{- \frac{(\mathcal{R}^{i+1})^2}{2}}\frac{\text{d}\mathcal{R}^{i+1}}{\text{d}\lambda^{i}}\,,\label{eq:lambda_backward_pi}
\end{align}
\end{subequations}
where the derivatives come from the changes of variable $\mathcal{R}^i\mapsto\lambda^{i+1}$, for the forward move, and $\mathcal{R}^{i+1}~\mapsto~\lambda^{i}$, for the backward one.
By inverting \cref{eq:lambda_forward,eq:lambda_backward} with respect to $R^i$ and $R^{i+1}$ and substituting their expressions in \cref{eq:lambda_forward_pi,eq:lambda_backward_pi}, it is straightforward to show that the contribution to the effective energy drift given by the transition probabilities is:
\begin{subequations}
\begin{align}
\Delta\widetilde{H}_{i\rightarrow i+1}^\Pi = -k_BT\log\frac{\Pi(\lambda^{i+1}\rightarrow \lambda^i)}{\Pi(\lambda^{i}\rightarrow \lambda^{i+1})}
= \Delta\lambda\left(\frac{f(\lambda^i)+f(\lambda^{i+1})}{2}\right) + \frac{\beta_T\Delta t}{16\tau_p}\Delta f^2\,.
\end{align}
\end{subequations}
Then, by summing $\Delta\widetilde{H}_{i\rightarrow i+1}^{\mathcal{P}}$ and $\Delta\widetilde{H}_{i\rightarrow i+1}^\Pi$ one recovers the result written in \cref{eq:eff_eng_iso}.
\section{Derivation of anisotropic energy drift for SCR}\label{appendix:aniso_effeng}
In this section we report the derivation of the anisotropic contribution to the effective energy drift, within the time-reversible integration scheme of \cref{sec:TR}. For this purpose, let's rewrite the second step of \cref{alg:trotterized_integrator}, namely the rescaling
\begin{equation}
\mathbf{h}^{i+1} = \exp\big(\mathbf{A}_2^i\,\Delta t + b\,\Delta \mathbf{W}_2^i\big)\,\mathbf{h}^{i}\,,
\end{equation}
by introducing an auxiliary momentum variable $\bm{\alpha}$ in a time-reversible fashion, shown in \cref{alg:TR_alpha}.\\
\begin{algorithm}[H]
\SetAlgoLined
$\bm{\alpha}\gets\bm{\mathcal{R}}_2$, where $\bm{\mathcal{R}}_2 = \bm{\mathcal{R}} - \text{Tr}\left(\bm{\mathcal{R}}\right)/3$\;
$\bm{\alpha}\gets\bm{\alpha}+\mathbf{A}_2\sqrt{\Delta t}/b$\;
$\mathbf{h} \gets \exp\big(\bm{\alpha}b\sqrt{\Delta t} \big)\,\mathbf{h}$\;
recompute $\mathbf{A}_2$\;
$\bm{\alpha}\gets\bm{\alpha}+\mathbf{A}_2\sqrt{\Delta t}/b$\;
$\bm{\alpha}\gets\bm{\mathcal{R}}_2'$, where $\bm{\mathcal{R}}_2' = \bm{\mathcal{R}}' - \text{Tr}\left(\bm{\mathcal{R}}'\right)/3$\;
\caption{Change of box shape with the auxiliary variable $\bm{\alpha}$.}
\label{alg:TR_alpha}
\end{algorithm}
\vspace{2.5cm}
We recall that $b$ only depends on $\mathbf{h}$ through its determinant, that is untouched in the rescaling at step 3. Therefore, in this context $b$ can be treated as a constant and its time index is omitted; for the same reason, only $\mathbf{A}_2$ is recomputed at step 4.\\
\noindent In steps 1 and 6, $\bm{\mathcal{R}}$ and $\bm{\mathcal{R}}'$ are 3$\times$3 matrices of i.i.d. zero-mean and unit-variance Gaussian numbers. The effective energy drift is formally defined according to \cref{effective_energy}, which here takes the form
\begin{equation}\label{eq:aniso_drift}
\Delta \widetilde{H}_{i\rightarrow i+1} = -k_B T \log\frac{\mathcal{P}\big(-\bm{\alpha}^{i+1},\mathbf{h}^{i+1}\big)\,\Pi\big((-\bm{\alpha}^{i+1},\mathbf{h}^{i+1})\rightarrow (-\bm{\alpha}^i,\mathbf{h}^i)\big)}{\mathcal{P}\big(\bm{\alpha}^i,\mathbf{h}^i\big)\,\Pi\big((\bm{\alpha}^i,\mathbf{h}^i)\rightarrow(\bm{\alpha}^{i+1},\mathbf{h}^{i+1})\big)}\,.
\end{equation}
Numerator and denominator in the argument of the logarithm can be evaluated by considering respectively the \emph{forward} and \emph{backward} moves in the scheme of \cref{alg:TR_alpha}, where the momentum-like variables $\bm{\alpha}$ gain a minus sign in the backward trajectory, exactly as discussed in the context of generalized detailed balance in \cref{hybrid_MC}. The forward move ($\bm{\alpha}^i,\mathbf{h}^i$)$\,\mapsto\,$($\bm{\alpha}^{i+1},\mathbf{h}^{i+1}$) can be written as
\begin{subequations}\label[pluralequation]{forward_traj}
\begin{align}
\bm{\alpha}^{i_+} &= \bm{\mathcal{R}}_{2F}^i + \mathbf{A}_2^i\frac{\sqrt{\Delta t}}{b}\,,\label{1st_eq_forward}\\
\mathbf{h}^{i+1} &= \exp\big(b\sqrt{\Delta t}\,\bm{\alpha}^{i_+}\big)\,\mathbf{h}^i\,, \label{2nd_eq_forward}\\
\bm{\alpha}^{i+1} &= \bm{\mathcal{R}}_{2F}^{i+1}\,,\label{3rd_eq_forward}
\end{align}
\end{subequations}
where the superscript $i_+$ identifies an intermediate step between $i$ and $i+1$. Similarly, calling $(i+1)_-$ an intermediate step in the backward trajectory, the move ($-\bm{\alpha}^{i+1},\mathbf{h}^i$)$\,\mapsto\,$($-\bm{\alpha}^{i},\mathbf{h}^{i}$) reads:
\begin{subequations}\label[pluralequation]{backward_traj}
\begin{align}
-\bm{\alpha}^{(i+1)_-} &= \bm{\mathcal{R}}_{2B}^{i+1} + \mathbf{A}_2^{i+1}\frac{\sqrt{\Delta t}}{b}\,,\\
\mathbf{h}^{i+1} &= \exp\big(-b\sqrt{\Delta t}\,\bm{\alpha}^{(i+1)_-}\big)\,\mathbf{h}^i\,,\label{2nd_eq_backward}\\
-\bm{\alpha}^{i} &= \bm{\mathcal{R}}_{2B}^{i}\,.
\end{align}
\end{subequations}
Note that by inverting \cref{2nd_eq_forward,2nd_eq_backward} with respect to $\bm{\alpha}^{i_+}$ and $\bm{\alpha}^{(i+1)_-}$ we obtain respectively:
\begin{subequations}
\begin{align}
\bm{\alpha}^{i_+} &= \frac{1}{b\sqrt{\Delta t}}\log\big(\mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}\big)\,,\\
\bm{\alpha}^{(i+1)_-} &= -\frac{1}{b\sqrt{\Delta t}}\log\big(\mathbf{h}^{i}(\mathbf{h}^{i+1})^{-1}\big) = \frac{1}{b\sqrt{\Delta t}}\log\big(\mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}\big)\,,\label{matrix_log_prop}
\end{align}
\end{subequations}
where the \emph{matrix logarithm} appearing in the two equations is defined as the inverse operation of the matrix exponential. In the last passage of \cref{matrix_log_prop} we have used the properties
\begin{subequations}\label[pluralequation]{generic_properties}
\begin{align}
-\log\big(\mathbf{M}_1\big) &= \log\big(\mathbf{M}_1^{-1}\big)\,, \\
\big(\mathbf{M}_1\,\mathbf{M}_2\big)^{-1} &= \mathbf{M}_2^{-1}\,\mathbf{M}_1^{-1}\,,
\end{align}
\end{subequations}
holding for two generic matrices $\mathbf{M}_1,\,\mathbf{M}_2$. As a consequence, $\bm{\alpha}^{i_+}$ and $\bm{\alpha}^{(i+1)_-}$, are actually the same matrix, which will be called $\bm{\alpha}^{i+1/2}$ in the following. Let's now evaluate the forward transition probability $\Pi_F = \Pi\big((\bm{\alpha}^i,\mathbf{h}^i)\rightarrow(\bm{\alpha}^{i+1},\mathbf{h}^{i+1})\big)$ appearing in \cref{eq:aniso_drift}. To write $\Pi_F$ correctly, we recall that $\bm{\mathcal{R}}_{2F}^i$ and $\bm{\mathcal{R}}_{2F}^{i+1}$ are obtained
as
\begin{subequations}
\begin{align}
\bm{\mathcal{R}}_{2F}^i &= \bm{\mathcal{R}}_{F}^i - \text{Tr}\big(\bm{\mathcal{R}}_{F}^i\big)/3\,,\\
\bm{\mathcal{R}}_{2F}^{i+1} &= \bm{\mathcal{R}}_{F}^{i+1} - \text{Tr}\big(\bm{\mathcal{R}}_{F}^{i+1}\big)/3\,,
\end{align}
\end{subequations}
where $\bm{\mathcal{R}}_{F}^i$ and $\bm{\mathcal{R}}_{F}^{i+1}$ are 3$\times$3 matrices of i.i.d. zero-mean and unit-variance Gaussian numbers. As a consequence, it is easy to show that the diagonal elements of $\bm{\mathcal{R}}_{2F}^i$ and $\bm{\mathcal{R}}_{2F}^{i+1}$ are Gaussian numbers with zero-mean and variance equal to 2/3, satisfying the constraints $\text{Tr}\big(\bm{\mathcal{R}}_{2F}^i\big) = \text{Tr}\big(\bm{\mathcal{R}}_{2F}^{i+1}\big) = 0$. In other words, $\bm{\mathcal{R}}_{2F}^i$ and $\bm{\mathcal{R}}_{2F}^{i+1}$ are random matrices containing only 8 independent elements, and a transformation followed by a marginalization over the redundant degree of freedom is necessary to write their joint probability distribution. With this regard, we introduce a linear transformation $\mathcal{G}$ that maps a generic 3$\times$3 matrix $\mathbf{M}$ into a 9-dimensional vector $\widetilde{\mathbf{M}}$ such that
\begin{subequations}\label[pluralequation]{eq:m0m1m2}
\begin{align}
\widetilde{M}_0 &= \text{Tr}\big(\mathbf{M}\big)/2\,, \label{first_tilde_var}\\
\widetilde{M}_1 &= \big(M_{xx} + M_{yy} - M_{zz}\big) / \sqrt{2}\,,\\
\widetilde{M}_2 &= \big(M_{xx} - M_{yy} + M_{zz}\big) / \sqrt{2}\,,
\end{align}
\end{subequations}
and the remaining six components $\widetilde{M}_j$ ($j = 3,...,8$) correspond to the off-diagonal elements of $\widetilde{\mathbf{M}}$, in an order that is irrelevant for the following reasoning. If we apply this transformation to $\bm{\mathcal{R}}_{2F}^i$ and $\bm{\mathcal{R}}_{2F}^{i+1}$, we can observe that the choice of the prefactors $1/\sqrt{2}$ in \cref{eq:m0m1m2} are such that the components $\widetilde{{\mathcal{R}}}_{2F,j}^i$ and $\widetilde{{\mathcal{R}}}_{2F,j}^{i+1}$ from $j=1$ to $j=8$ are zero-mean and unit-variance Gaussian numbers, and that in these new variables the constraints reported above simply read $\widetilde{{\mathcal{R}}}_{2F,0}^i = \widetilde{{\mathcal{R}}}_{2F,0}^{i+1} = 0$. Then, since the Jacobian of the transformation $\mathcal{G}$ is $1$ in absolute value, we can write the the forward transition probability as
\begin{equation}\label{Pi_f}
\Pi_F = \frac{1}{(2\pi)^9}\exp\left[-\frac{1}{2}\sum_{j=1}^8 \left(\left(\widetilde{{\mathcal{R}}}_{2F,j}^i\right)^2 +\left(\widetilde{{\mathcal{R}}}_{2F,j}^{i+1}\right)^2 \right)\right]\,\Big|\text{det}\,\mathbf{J}_F \Big|\,,
\end{equation}
where $\mathbf{J}_F$ is the Jacobian of the transformation $(\bm{\mathcal{R}}_{2F}^i,\bm{\mathcal{R}}_{2F}^{i+1})\mapsto(\bm{\alpha}^{i+1},\mathbf{h}^{i+1})$. Since $\bm{\mathcal{R}}_{2F}^i$ only depends on $\mathbf{h}^{i+1}$ and $\bm{\mathcal{R}}_{2F}^{i+1}$ only depends on $ \bm{\alpha}^{i+1}$, this matrix is block-diagonal and its determinant factorizes as
\begin{equation}
\text{det}\,\mathbf{J}_F = \left(\text{det}\,\mathbf{J}_{\bm{\mathcal{R}}_{2F}^i\mapsto\mathbf{h}^{i+1}} \right) \left(\text{det}\,\mathbf{J}_{\bm{\mathcal{R}}_{2F}^{i+1}\mapsto\bm{\alpha}^{i+1}} \right)\,.
\end{equation}
From \cref{3rd_eq_forward} we observe that the second determinant is simply 1, while from \cref{1st_eq_forward} we can write the components of the first Jacobian tensor as
\begin{equation}
\left(\mathbf{J}_{\bm{\mathcal{R}}_{2F}^i\mapsto\mathbf{h}^{i+1}}\right)_{\gamma k\alpha\beta} = \frac{\partial\, \mathcal{R}_{2F,\alpha\beta}^i}{\partial\, h_{\gamma k}^{i+1}} = \frac{1}{b\sqrt{\Delta t}}\frac{\partial}{\partial h^{i+1}_{\gamma k}}\left(\log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1} \right)_{\alpha\beta}
\end{equation}
Performing the same calculations on the time-reversed trajectory in \cref{backward_traj}, the expression that one finds for the backward transition probability $\Pi_B = \Pi\big((-\bm{\alpha}^{i+1},\mathbf{h}^{i+1})\rightarrow (-\bm{\alpha}^i,\mathbf{h}^i)\big)$ reads
\begin{equation}\label{Pi_b}
\Pi_B = \frac{1}{(2\pi)^9}\exp\left[-\frac{1}{2}\sum_{j=1}^8 \left(\left(\widetilde{{\mathcal{R}}}_{2B,j}^i\right)^2 +\left(\widetilde{{\mathcal{R}}}_{2B,j}^{i+1}\right)^2 \right)\right]\,\Big|\text{det}\,\mathbf{J}_B \Big|\,,
\end{equation}
and in this case the non-trivial part of the Jacobian is:
\begin{equation}
\left(\mathbf{J}_{\bm{\mathcal{R}}_{2B}^{i+1}\mapsto\mathbf{h}^{i}}\right)_{\gamma k\alpha\beta} = \frac{\partial\, \mathcal{R}_{2B,\alpha\beta}^{i+1}}{\partial\, h_{\gamma k}^{i}} = \frac{1}{b\sqrt{\Delta t}}\frac{\partial}{\partial h^{i}_{\gamma k}}\left(\log \mathbf{h}^{i}(\mathbf{h}^{i+1})^{-1} \right)_{\alpha\beta}\,.
\end{equation}
Therefore the contribution of the two Jacobians in the ratio $\Pi_B/\Pi_F$ in \cref{eq:aniso_drift} can be manipulated as it follows:
\begin{subequations}
\begin{align}
\frac{\text{det}\,\mathbf{J}_B }{\text{det}\,\mathbf{J}_F } &= \frac{\text{det}\left[\frac{\partial}{\partial \,\mathbf{h}^{i}}\log \mathbf{h}^{i}(\mathbf{h}^{i+1})^{-1} \right]}{\text{det}\left[\frac{\partial}{\partial\, \mathbf{h}^{i+1}}\log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1} \right]} = \\
&= \text{det}\left[\left(\frac{\partial}{\partial \,\mathbf{h}^{i}}\log \mathbf{h}^{i}(\mathbf{h}^{i+1})^{-1}\right)\left(\frac{\partial}{\partial\, \mathbf{h}^{i+1}}\log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}\right)^{-1} \right] \label{eq:d17b}\\
&= \text{det}\left[\left(\frac{\partial}{\partial \,\mathbf{h}^{i}}\log \mathbf{h}^{i}(\mathbf{h}^{i+1})^{-1}\right)\left(\frac{\partial\mathbf{h}^{i+1}}{\partial\, \log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}}\right) \right]\,. \label{eq:d17c}
\end{align}
\end{subequations}
\Cref{eq:d17b} is obtained using the well-known property $(\text{det}\,\mathbf{M})^{-1} = \text{det}\big(\mathbf{M}^{-1}\big)$, while \cref{eq:d17c} relies on the fact that
\begin{equation}
\frac{\partial \mathbf{X}}{\partial\mathbf{Y}} = \left(\frac{\partial\mathbf{Y}}{\partial \mathbf{X}}\right)^{-1}
\end{equation}
if $\mathbf{Y} = \mathbf{Y}(\mathbf{X})$ is continuous and differentiable in $X_{ij}$, if it is invertible and its inverse $\mathbf{X} = \mathbf{X}(\mathbf{Y})$ has the same properties with respect to $Y_{ij}$. In fact if these hypothesis hold, as in the case of $\mathbf{Y}=\log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}$ and $\mathbf{X}=\mathbf{h}^{i+1}$, then
\begin{subequations}
\begin{align}
\sum_{kl}\left(\frac{\partial \mathbf{Y}}{\partial\mathbf{X}} \right)_{ijkl}\left(\frac{\partial \mathbf{X}}{\partial\mathbf{Y}} \right)_{klmn} = \sum_{kl} \frac{\partial {Y}_{kl}}{\partial{X}_{ij}}\frac{\partial {X}_{mn}}{\partial{Y}_{kl}} = \delta_{mi}\,\delta_{jn}\,.
\end{align}
\end{subequations}
Starting again from \cref{eq:d17c} we can write:
\begin{subequations}
\begin{align}
\frac{\text{det}\,\mathbf{J}_B }{\text{det}\,\mathbf{J}_F } &= \text{det}\left[\left(-\frac{\partial\,\log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}}{\partial \,\mathbf{h}^{i}}\right)\left(\frac{\partial\mathbf{h}^{i+1}}{\partial\, \log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}}\right) \right] \\
&= \text{det}\left[-\left(\frac{\partial\,\log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}}{\partial\, \log \mathbf{h}^{i+1}(\mathbf{h}^{i})^{-1}}\right)\left(\frac{\partial\mathbf{h}^{i+1}}{\partial \,\mathbf{h}^{i}}\right) \right] \\
&= \text{det}\left[-\left(\frac{\partial\mathbf{h}^{i+1}}{\partial \,\mathbf{h}^{i}}\right) \right]\,.
\end{align}
\end{subequations}
Here we have used both the properties in \cref{generic_properties} and the fact that
\begin{equation}
\frac{\partial\mathbf{A}}{\partial\mathbf{B}}\frac{\partial\mathbf{C}}{\partial\mathbf{A}} = \frac{\partial\mathbf{C}}{\partial\mathbf{B}}
\end{equation}
if these matrix derivatives are well defined. In fact, writing explicitly the product:
\begin{subequations}
\begin{align}
\left(\frac{\partial\mathbf{A}}{\partial\mathbf{B}}\frac{\partial\mathbf{C}}{\partial\mathbf{A}}\right)_{ijmn} &= \sum_{kl}\left(\frac{\partial\mathbf{A}}{\partial\mathbf{B}}\right)_{ijkl}\left(\frac{\partial\mathbf{C}}{\partial\mathbf{A}}\right)_{klmn} = \sum_{kl}\frac{\partial{A}_{kl}}{\partial{B}_{ij}}\frac{\partial{C}_{mn}}{\partial{A}_{kl}} = \frac{\partial{C}_{mn}}{\partial{B}_{ij}}\,.
\end{align}
\end{subequations}
Finally, the remaining derivative can be easily computed using \cref{2nd_eq_forward}:
\begin{equation}
\frac{\partial\mathbf{h}^{i+1}}{\partial\,\mathbf{h}^{i}} = \frac{\partial}{\partial\,\mathbf{h}^{i}}\exp\big(b\sqrt{\Delta t}\,\bm{\alpha}^{i+1/2}\big)\,\mathbf{h}^i = \exp\big(b\sqrt{\Delta t}\,\bm{\alpha}^{i+1/2}\big)\,.
\end{equation}
Since this exponential matrix has determinant equal to 1, as discussed in \cref{sec:TR}, the ratio of the Jacobians in $\Pi_B/\Pi_F$ is simply
\begin{equation}
\left| \frac{\text{det}\,\mathbf{J}_B }{\text{det}\,\mathbf{J}_F }\right| = 1\,.
\end{equation}
Then, using \cref{Pi_f,Pi_b} the contribution of the transitions probabilities to the anisotropic drift reads:
\begin{subequations}
\begin{align}
\Delta \widetilde{H}_{i\rightarrow i+1}^{\Pi} &= -k_B T\log\frac{\Pi_B}{\Pi_F} \\
&= \frac{k_B T}{2}\sum_{j=1}^8\left[\left(\widetilde{\mathcal{R}}_{2B,j}^{i+1}\right)^2+\left(\widetilde{\mathcal{R}}_{2B,j}^{i}\right)^2 - \left(\widetilde{\mathcal{R}}_{2F,j}^{i+1}\right)^2 -\left(\widetilde{\mathcal{R}}_{2F,j}^{i}\right)^2\right]
\end{align}
\end{subequations}
Substituting the expressions of these four random vectors in terms of the $\mathcal{G}$-transformed quantities appearing in \cref{forward_traj} and \cref{backward_traj} one finds:
\begin{equation}\label{deltaH_Pi_tilde}
\Delta \widetilde{H}_{i\rightarrow i+1}^{\Pi} = \frac{k_B T}{2}\sum_{j=1}^8 \left[\frac{\Delta t}{b^2}\Delta \widetilde{A}_{2j}^{\,2} -\Delta\widetilde{\alpha}_{j}^2 + 2\widetilde{\alpha}_j^{i+1/2}\frac{\sqrt{\Delta t}}{b}\left(\widetilde{A}_{2j}^{\,i} + \widetilde{A}_{2j}^{\,i+1} \right)\right]\,.
\end{equation}
Note that we can include for each term in the sum also the component $j=0$ as defined in \cref{first_tilde_var}, since all the matrices considered here are traceless. As a consequence, all the terms in \cref{deltaH_Pi_tilde} can be interpreted as squared moduli or scalar products of 9-dimensional vectors obtained from the transformation $\mathcal{G}$. Since this transformation is unitary and then preserves all scalar products, it is possible to map back \cref{deltaH_Pi_tilde} to the original matrix quantities by applying the inverse transformation $\mathcal{G}^{-1}$:
\begin{equation}
\Delta \widetilde{H}_{i\rightarrow i+1}^{\Pi} = \frac{k_B T}{2}\sum_{\alpha\beta} \left[\frac{\Delta t}{b^2}\Delta {A}_{2,\alpha\beta}^{\,2} -\Delta{\alpha}_{\alpha\beta}^2 + 2{\alpha}_{\alpha\beta}^{i+1/2}\frac{\sqrt{\Delta t}}{b}\left({A}_{2,\alpha\beta}^{\,i} + {A}_{2,\alpha\beta}^{\,i+1} \right)\right]\,.
\end{equation}
Let's now evaluate the remaining contribution to \cref{eq:aniso_drift}, starting from the following ratio:
\begin{equation}
\frac{\mathcal{P}\big(-\bm{\alpha}^{i+1},\mathbf{h}^{i+1}\big)}{\mathcal{P}\big(\bm{\alpha}^{i},\mathbf{h}^{i}\big)} = \frac{\mathcal{P}\big(-\bm{\alpha}^{i+1}\big)\,\mathcal{P}\big(\mathbf{h}^{i+1}\big)}{\mathcal{P}\big(\bm{\alpha}^{i}\big)\,\mathcal{P}\big(\mathbf{h}^{i}\big)}\,.
\end{equation}
The equality holds because the distributions of $\mathbf{h}$ and $\bm{\alpha}$ are independent. According to the $N\mathbf{S}T$ distribution in \cref{NPT_anisotropic}, the contribution from the ratio of the $\mathbf{h}$-distributions results in a sum of energy increments, only due to the change of shape:
\begin{equation}
\Delta \widetilde{H}_{i\rightarrow i+1}^{\mathcal{P}(\mathbf{h})} = -k_B T \log\frac{\mathcal{P}\big(\mathbf{h}^{i+1}\big)}{\mathcal{P}\big(\mathbf{h}^{i}\big)} = \Delta K + \Delta U + \Delta E_s\,.
\end{equation}
The only remaining term is
\begin{subequations}\label{last_effort}
\begin{align}
\Delta \widetilde{H}_{i\rightarrow i+1}^{\mathcal{P}(\bm{\alpha})} = -k_B T \log\frac{\mathcal{P}\big(-\bm{\alpha}^{i+1}\big)}{\mathcal{P}\big(\bm{\alpha}^{i}\big)}\,.
\end{align}
\end{subequations}
Recalling from \cref{forward_traj} and \cref{backward_traj} that $\bm{\alpha}^{i+1} = \bm{\mathcal{R}}_{2F}^{i+1}$ and $\bm{\alpha}^{i} = -\bm{\mathcal{R}}_{2B}^{i}$, it is possible to evaluate the distribution of $\alpha$ and the ratio in \cref{last_effort} using exactly the same procedure shown before to compute the transition probabilities. The final result,
\begin{equation}
\Delta \widetilde{H}_{i\rightarrow i+1}^{\mathcal{P}(\bm{\alpha})} = \frac{k_B T}{2}\sum_{\alpha\beta}\Delta\alpha_{\alpha\beta}^2\,,
\end{equation}
cancels the second term in \cref{deltaH_Pi_tilde}. Then, calling $\Delta\bm{\varepsilon}$ the argument of the exponential matrix responsible for the change of shape, namely $\Delta\bm{\varepsilon} = b\sqrt{\Delta t}\bm{\alpha}^{i+1/2}$, the sum $\Delta \widetilde{H}_{i\rightarrow i+1}^{\Pi} + \Delta \widetilde{H}_{i\rightarrow i+1}^{\mathcal{P}(\mathbf{h})} + \Delta \widetilde{H}_{i\rightarrow i+1}^{\mathcal{P}(\bm{\alpha})}$ results in the expression reported in \cref{eq:eff_eng_aniso}.
\chapter{PBCs and Bravais lattices}\label{PBCs}
Periodic boundary conditions (PBCs) are widely used in MD simulations to eliminate effects due to the boundaries, where the physical behaviour of the system could be substantially different than in the bulk. The implementation of these conditions is based on a geometric construction called \emph{Bravais lattice}, which is built using as \emph{primitive vectors} the box vectors $\mathbf{a},\mathbf{b},\mathbf{c}$. This periodic construction is mathematically defined as the infinite set of points $\{\mathbf{r}_{n_1 n_2 n_3}\}$ that one can generate by means of integer linear combinations of the primitive vectors:
\begin{equation}
\mathbf{r}_{n_1 n_2 n_3} = n_1 \mathbf{a} + n_2 \mathbf{b} + n_3 \mathbf{c}\,,
\end{equation}
where $n_1,\,n_2$ and $n_3$ are integer numbers. The points $\{\mathbf{r}_{n_1 n_2 n_3}\}$ define a periodic structure throughout the space where the fundamental unit, called \emph{unit cell}, has the geometry of the box defined by the three vectors $\mathbf{a},\mathbf{b},\mathbf{c}$. If we duplicate and translate every atom according to all the Bravais lattice vectors $\{\mathbf{r}_{n_1 n_2 n_3}\}$, all the space will be filled by periodic copies of the system of interest (see \cref{fig:pbcs}).
\begin{figure}[h!]
\centering
\includegraphics[width = 0.35\textwidth]{core/appendices/pbcs.png}
\caption{Graphical representation of PBCs in a two-dimensional system within a square box. Source: \cite{pbcs_img}.}
\label{fig:pbcs}
\end{figure}
Then, PBCs can be applied by making each atom interact with all the periodic copies of the remaining ones. Since it is not possible to deal with infinite interactions, a scheme called \emph{minimal image convention} is typically employed: the (short-range) interactions are considered only within a cut-off $r_{\text{cut}}$ such that, with a suitable choice of the box dimensions, each atom turns out to interact with only the nearest copy of each other atom. In particular, the necessary condition to apply this convention is
\begin{equation}
r_{\text{cut}} < \frac{1}{2}\text{min}\left(|\mathbf{a}|,|\mathbf{b}|,|\mathbf{c}| \right)\,.
\end{equation}
It is relevant to observe that the box vectors generating a given Bravais lattice structure are not unique (see \cref{fig:bravais_lattice_equiv}), and independent choices are connected by a transformation of the form \cref{change_cell_vectors}. Since the results of any MD simulation should be independent on this choice, this explains the check for the invariance of the anisotropic SCR equations under a redefinition of the box vectors, and the reason why this property should be satisfied by any equation describing anisotropic volume fluctuations in finite systems.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.45\textwidth]{core/appendices/bravais_lattice.png}
\caption{Equivalent choices of the primitive vectors for a two-dimensional Bravais lattice.}
\label{fig:bravais_lattice_equiv}
\end{figure}
We finally report in \cref{tab:boxes_def} some of the most employed box shapes in MD simulations, using the convention of upper triangular box matrices.
\begin{table}[h!]
\centering
\caption{Definitions of common box shapes.}
\label{tab:boxes_def}
\begin{tabular}{p{5cm} p{4cm} c}
\toprule[0.5pt]\toprule[0.5pt]
\textbf{Box type} & \textbf{Box matrix} $\mathbf{h}$ & \textbf{Volume} \\\midrule
Cubic & $\begin{pmatrix}
d & 0 & 0\\
0 & d & 0\\
0 & 0 & d
\end{pmatrix}$ & $d^3$ \\
Orthorhombic & $\begin{pmatrix}
d_x & 0 & 0\\
0 & d_y & 0\\
0 & 0 & d_z
\end{pmatrix}$ & $d_x\,d_y\,d_z$ \\
Rhombic dodecahedron & $\begin{pmatrix}
d & 0 & d/2\\
0 & d & d/2\\
0 & 0 & \sqrt{2}\,d/2
\end{pmatrix}$ & $\frac{\sqrt{2}}{2}d^3 \simeq 0.71\,d^3$ \\
Truncated octahedron & $\begin{pmatrix}
d & d/3 & -d/3\\
0 & 2\sqrt{2}\,d/3 & \sqrt{2}\,d/3\\
0 & 0 & \sqrt{2}\,d/3
\end{pmatrix}$ & $\frac{4\sqrt{3}}{9}d^3 \simeq 0.77\,d^3$ \\\bottomrule[0.5pt]\bottomrule[0.5pt]
\end{tabular}
\end{table}
\noindent Note that all these shapes are particular cases of \emph{triclinic} boxes, where $\mathbf{h}$ contains six independent non-zero elements, resulting in a parallelepiped with generic edge lengths $|\mathbf{a}|,|\mathbf{b}|,|\mathbf{c}|$ and generic angles $\alpha,\beta,\gamma$,
\chapter{First attempt for a time-reversible integrator}\label{appendix_epsvariables}
We discuss here an attempt to construct an integrator of \cref{crescale_aniso} with a time-reversible behaviour in the limit of small time steps (i.e. only considering the stochastic part of the equations). Let's introduce nine new variables as the components of a matrix $\bm{\varepsilon}$ defined as:
\begin{equation}
\mathbf{h} = e^{\bm{\varepsilon}}\,\mathbf{h}_0\,.
\end{equation}
Here $\mathbf{h}_0$ is a reference box matrix and $e^{\bm{\varepsilon}}$ is the \emph{matrix exponential} of $\bm{\varepsilon}$, defined by the power series
\begin{equation}
e^{\bm{\varepsilon}} = \sum_{k= 0}^\infty \frac{1}{k!}\,\bm{\varepsilon}^k\,.
\end{equation}
Note that $\bm{\varepsilon}$ can be seen as the generalization of the logarithmic volume $\varepsilon = \log(V/V_0)$ defined in the isotropic case. Since the expansion above contains infinite terms, $\exp(\bm{\varepsilon})$ cannot be practically computed in general, but in case of small matrix increments $\exp(\text{d}\bm{\varepsilon})$ can be approximated efficiently by truncating the expansion after a few terms.
With a relevant \emph{caveat} highlighted below, the equations for $\bm{\varepsilon}$ can be obained by means of It\^{o}'s chain rule. Then, using the well-known property $\text{det}\big(\text{exp}(\bm{\varepsilon})\big) = \text{exp}\big(\text{Tr}(\bm{\varepsilon})\big)$ the isotropic degree of freedom can be embedded in the variable
\begin{equation}\label{det_of_exp}
\lambda = \sqrt{V_0}\exp\left(\frac{1}{2}\text{Tr}(\bm{\varepsilon})\right) = \sqrt{V}\,,
\end{equation}
which is decoupled from the other degrees of freedom and can be propagated as in the isotropic case with \cref{lambda_eq_SCR}, where the noise prefactor is constant. The other eight variables $\Phi_k$ ($k=1,...,8$) can be chosen as independent linear combinations of the box matrix components and they are propagated according to eight SDEs with a $\lambda$-dependent noise prefactor $b(\lambda)$, which can be symmetrized by means of a geometric mean between the values of $\lambda$ at the current and the next steps:
\begin{equation}
b(\lambda_{t}) \longmapsto b\left(\sqrt{\lambda_{t}\,\lambda_{t+\Delta t}}\,\right)\,.
\end{equation}
This operation, which requires $\lambda$ to be propagated before the other eight variables at each step, should enhance the time-reversibility of the generated trajectory, and as a consequence the "good scaling" of the effective energy drift with the integration time step. However, this integrator actually appears to work worse than the Euler one (see \cref{euler_integrator}) for at least two reasons:
\begin{itemize}
\item The derivation of the equations for $\bm{\varepsilon}$ requires to know the derivatives
\begin{equation}
\frac{\partial \varepsilon_{\alpha\beta}}{\partial h_{\gamma i}} = \frac{\partial }{\partial h_{\gamma i}}\log(\mathbf{h}\,\mathbf{h}_0^{-1})_{\alpha\beta}\,,
\end{equation}
where the \emph{matrix logarithm} appearing above is the inverse of the matrix exponential previously defined. Unfortunately, these derivatives do not admit any closed-form solution and they can only be computed in the (wrong) hypothesis that the following commutator is zero:
\begin{equation}
\left[\frac{\partial }{\partial h_{\gamma i}}(\mathbf{h}\,\mathbf{h}_0^{-1}), (\mathbf{h}\,\mathbf{h}_0^{-1}) \right] = 0\,.
\end{equation}
As a consequence, the equations for $\bm{\varepsilon}$ can only be obtained within a certain degree of approximation, which cannot be clearly quantified.
\item After computing the increments of the variables $\{\lambda,\Phi_k\}$ and mapping them back to the increments of the $\bm{\varepsilon}$ variables, namely $\text{d}\bm{\varepsilon}$, the rescaling matrix $\bm{\mu}$ should be computed as
\begin{equation}
\bm{\mu} = \mathbf{h}^{-1}\left(\mathbf{h}+\text{d}\mathbf{h} \right) = e^{-\bm{\varepsilon}}\left(e^{\bm{\varepsilon}\,+\,\text{d}\bm{\varepsilon}} \right) \simeq e^{\text{d}\bm{\varepsilon}}\,.
\end{equation}
The last passage includes a relevant approximation, coming from the fact that the matrix $\bm{\varepsilon}$ and its increment do not commute in general: $\left[\bm{\varepsilon},\text{d}\bm{\varepsilon} \right]\neq 0$. However, this approximation is unavoidable to implement the algorithm, as only the matrix exponential $\text{exp}(\text{d}\bm{\varepsilon})$ can be computed with a sufficient accuracy. As a consequence, the rescaling matrix is $\bm{\mu}$ necessarily obtained with a further error.
\end{itemize}
Because of these issues, the tests of this time-reversible integrator have shown a bad behaviour of the effective energy drift, together with additional problems related to the ACFs of the box matrix components. A well-behaved time-reversible integrator for the anisotropic SCR equations is discussed in \cref{sec:TR}.
\chapter{Integration to simulation analysis}
\section{Error estimation}\label{appendix_block_bootstrap}
In the analysis reported in \cref{implementations}, errors are computed with \emph{block-bootstrap analysis}, which is a technique to estimate the uncertainty of statistics calculated over time series, namely by employing correlated samplings. To illustrate the procedure, let's consider a stationary time series $\mathbf{x}=\{x_1,...,x_n\}$, from which a statistics of interest $s(\mathbf{x})$ has been computed. If the samplings were uncorrelated, the error associated to $s(\mathbf{x})$ could be estimated with the \emph{bootstrap} approach \cite{efron}, which consists in the following steps:
\begin{itemize}
\item $B$ new time series $\mathbf{x}^{\prime\,b}$ of length $n$ are generated by resampling with replacement the original series;
\item the statistics of interest is recomputed over each bootstrap sample $\mathbf{x}^{\prime\, b}$;
\item the standard error of $s(\mathbf{x})$ is calculated as the standard deviation of the bootstrap estimates, namely as
\begin{equation}
\text{SE}\left[s(\mathbf{x}) \right] = \sqrt{\frac{1}{B}\sum_{b=1}^B\Big(s(\mathbf{x}^{\prime\,b })-\bar{s}\Big)^2}\,,
\end{equation}
\end{itemize}
where $\bar{s} = \sum_b s(\mathbf{x}^{\prime\,b })/B$. However, this method only works under the assumption of i.i.d. samplings, and brings to underestimate the actual error when this condition is not satisfied.
A possibility to circumvent this problem is to divide the time series in $N_B$ non-ovelapping blocks, each one containing $n_b$ samplings, and to apply the bootstrap resampling on the blocks, studying the standard error obtained as a function of $n_b$. Each bootstrap trajectory is composed in this way by a sequence of blocks coming from the original series, each one preserving its internal order. For large values of $n_b$ the standard error is expected to saturate, and the value of the plateau gives a meaningful estimate of $\text{SE}\left[s(\mathbf{x}) \right]$. \Cref{fig:block_bootstrap_example} shows an example of this systematic procedure applied to a volume time series, produced in one of the the simulations illustrated in \cref{LJ_crystal}.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.75\textwidth]{core/third_chapter/block_bootstrap_example_conservative.png}
\caption{Example of block bootstrap analysis, applied to the standard error of the average volume from a simulation performed with SimpleMD. In the bootstrap procedure, the number of resampled trajectories is $B=200$. The optimal value of $\text{SE}\left[\langle V\rangle\right])$ is determined conservatively using the largest estimate in the plateau region.}
\label{fig:block_bootstrap_example}
\end{figure}
\section{Autocorrelation time and error of the sample mean}\label{appendix_autocorrelation}
We show here that, if a statistics is estimated as the average over a finite-length time series, the uncertainty of the estimate depends both on the length of the series and on its autocorrelation. We consider the case of a continuous time series, but the same reasoning can be applied to discretized MD trajectories just replacing integrals with sums. If $x(t)$ is a stationary stochastic process, its autocorrelation time $\tau_x$ is defined as
\begin{equation}
\tau_x = \int_0^\infty \text{d}t\, C_x(t)\,,
\end{equation}
where $C_x(t)$ is the (normalized) ACF of the process:
\begin{equation}
C_x(t) = \frac{\langle x(t)x(0) \rangle - \langle x\rangle^2}{\langle x^2 \rangle - \langle x\rangle^2} = \frac{\langle x(t)x(0)\rangle -\langle x\rangle^2}{\sigma^2_x}\,.
\end{equation}
Note that $\langle x\rangle$ and $\langle x^2\rangle$ are time-independent, since the process is assumed to be stationary. Let's now consider the average value of $x$ estimated in a trajectory of finite length $T$, where the system is supposed to be already equilibrated at time $t=0$:
\begin{equation}
\overbar{x} = \frac{1}{T}\int_0^T \text{d}t\,x(t)\,.
\end{equation}
The variance of $\overbar{x}$ as an estimator of the real average value $\langle x \rangle$ can be written as:
\begin{equation}
\sigma_{\overbar{x}}^2 = \langle \overbar{x}^2\rangle - \langle \overbar{x}\rangle^2\,,
\end{equation}
where these averages can be thought over infinite simulations of length $T$ of the process. Then:
\begin{subequations}
\begin{align}
\sigma_{\overbar{x}}^2 &= \Big\langle \frac{1}{T^2}\int_0^T \text{d}t\,x(t)\int_0^T \text{d}t'\,x(t')\Big\rangle - \langle x\rangle^2 \\
&= \frac{1}{T^2}\int_0^T\int_0^T \text{d}t\,\text{d}t'\,\langle x(t) x(t')\rangle - \langle x\rangle^2 \label{intx2}\\
&= \frac{2}{T^2}\int_0^T\int_0^t \text{d}t\,\text{d}t'\,\langle x(t) x(t')\rangle - \langle x\rangle^2 \,.
\end{align}
\end{subequations}
The last passage is performed observing that the integrand function is symmetric with respect to the exchange $x\leftrightarrow x'$, and the integral in \cref{intx2} contains for each pair of integration points $(t_1,t_2)$ its symmetric $(t_2,t_1)$. With other manipulations we can wite:
\begin{subequations}
\begin{align}
\sigma_{\overbar{x}}^2 &= \frac{2}{T^2}\int_0^T\int_0^t \text{d}t\,\text{d}t'\,\Big[\langle x(t) x(t')\rangle - \langle x\rangle^2\Big] \\
&= \frac{2\sigma_x^2}{T^2}\int_0^T\int_0^t \text{d}t\,\text{d}t'\,C_x(t-t') \\
&\simeq \frac{2\sigma_x^2}{T^2}\int_0^T \text{d}t \,\tau_x\,,
\end{align}
\end{subequations}
where the last approximation is meaningful if $T\gg\tau_x$. Then the final result reads
\begin{equation}\label{standard_error_autocorr}
\sigma_{\overbar{x}}^2 \simeq \frac{2\tau_x}{T}\sigma_x^2\,,
\end{equation}
namely the uncertainty of an average $\overbar{x}$ computed over a correlated time series:
\begin{itemize}
\item decreases with the length of the series;
\item increases with its autocorrelation time.
\end{itemize}
\vfil
\section{Reference ACFs of volume and its variance}\label{appendix_relaxation_autocorrelation}
The aim of this section is to justify the dashed lines in \cref{fig:ACFs_LJ},
\begin{subequations}\label[pluralequation]{ref_acfs}
\begin{align}
C_{V}(t) &= e^{-t/\tau_p}\,, \\
C_{\sigma_V^2}(t) &= e^{-2t/\tau_p}\,,
\end{align}
\end{subequations}
which are the analytical ACFs of the volume and its variance in the limit case of a Gaussian-distributed volume following a Langevin dynamics:
\begin{equation}
\text{d}V = -\frac{1}{\tau_p}\left(V-\overbar{V}\right) \text{d}t + \sqrt{\frac{2\sigma_V^2}{\tau_p}}\,\text{d}W\,.
\end{equation}
Note that the SCR dynamics for the volume, given by \cref{crescale_iso_V}, is a first-order stochastic dynamics that resembles to the above Langevin equation if we neglect the additional dependencies on $V$ in the internal pressure and in the noise prefactor; moreover, due to the central limit theorem the volume distribution is expected to approach a Gaussian in the thermodynamic limit, unless the system is in a critical point where different phases coexist. As a consequence, \cref{ref_acfs} represent the reference behaviours for the ACFs of interest when $N$ is large.
In order to derive \cref{ref_acfs}, let's first apply It\^{o}'s lemma to rewrite the Langevin equation in terms of $\Delta V = V-\overbar{V}$:
\begin{equation}
\text{d}\,\Delta V = -\frac{\Delta V}{\tau_p} \text{d}t + \sqrt{\frac{2\sigma_V^2}{\tau_p}}\,\text{d}W\,.
\end{equation}
Using a different formalism, this equation can be equivalently written as
\begin{equation}
\frac{\text{d}\Delta V}{\text{d}t} = -\frac{\Delta V}{\tau_p} + \sqrt{\frac{2\sigma_V^2}{\tau_p}}\,\eta(t) \,,
\end{equation}
where $\eta(t)$ is a Gaussian white noise satisfying \cref{gaussian_white_noise}. The formal solution of this equation, decomposed as the sum of the general homogeneous solution and a particular solution of the inhomogenous problem, can be written as
\begin{equation}
\Delta V(t) = \Delta V(0)e^{-t/\tau_p} + \sqrt{\frac{2\sigma_V^2}{\tau_p}}\int_0^t \text{d}t'\, e^{-(t-t')/\tau_p}\,\eta(t')\,.
\end{equation}
Representing with $\langle \cdot\rangle$ the average over different realization of the stochastic process $\Delta V(t)$, we can evaluate the following correlation:
\begin{subequations}\label{deltaVdeltaV0}
\begin{align}
\langle \Delta V(t)\Delta V(0)\rangle = \langle \Delta V(0)^2\rangle e^{-t/\tau_p} +
\sqrt{\frac{2\sigma_V^2}{\tau_p}}\int_0^t \text{d}t'\, e^{-(t-t')/\tau_p}\,\langle \Delta V(0)\eta(t')\rangle\,.
\end{align}
\end{subequations}
We can now observe that:
\begin{itemize}
\item $\langle \Delta V(0)^2\rangle = \sigma_V^2$, since the process is stationary;
\item $\langle \Delta V(0)\eta(t')\rangle = \langle \Delta V(0)\rangle \langle\eta(t')\rangle = 0$, since $\Delta V(t)$ is a non-anticipating function of $\eta(t)$ and is then independent on future realizations of the noise.
\end{itemize}
As a consequence, \cref{deltaVdeltaV0} becomes
\begin{equation}
\langle \Delta V(t)\Delta V(0)\rangle = \sigma_V^2\,e^{-t/\tau_p}\,,
\end{equation}
and the ACF of the volume can be simply computed as:
\begin{subequations}
\begin{align}
C_V(t) &= \frac{\langle V(t) V(0)\rangle - \langle V\rangle^2 }{\langle \Delta V^2\rangle} =\frac{\langle \Delta V(t)\Delta V(0)\rangle}{ \sigma_V^2} = e^{-t/\tau_p}\,.
\end{align}
\end{subequations}
In order to compute the ACF of the volume variance, let's start by evaluating the following correlation:
\begin{subequations}
\begin{align}
\langle \Delta V(t)^2 &\Delta V(0)^2\rangle = \langle \Delta V(0)^4\rangle e^{-2t/\tau_p} \\
&+ 2\langle \Delta V(0)^3\rangle \sqrt{\frac{2\sigma_V^2}{\tau_p}}e^{-t/\tau_p}\int_0^t \text{d}t'\, e^{-(t-t')/\tau_p}\,\langle\eta(t')\rangle \nonumber\\
&+ \frac{2\sigma_V^2}{\tau_p}\langle \Delta V(0)^2\rangle \int_0^t \text{d}t'\int_0^t \text{d}t''\, e^{-(t-t')/\tau_p}e^{-(t-t'')/\tau_p}\,\langle\eta(t')\eta(t'')\rangle\,. \nonumber
\end{align}
\end{subequations}
Using the properties of $\eta(t)$ in \cref{gaussian_white_noise} we get:
\begin{equation}
\langle \Delta V(t)^2 \Delta V(0)^2\rangle = \langle \Delta V^4\rangle e^{-2t/\tau_p} + \sigma_V^4 \left( 1-e^{-2t/\tau_p}\right)\,.
\end{equation}
Hence, the ACF of the volume variance can be written as:
\begin{subequations}
\begin{align}
C_{\sigma_V^2}(t) &= \frac{\langle \Delta V(t)^2 \Delta V(0)^2\rangle - \langle \Delta V^2\rangle^2}{\langle \Delta V^4\rangle - \langle \Delta V^2\rangle^2} \\
&= \frac{\langle \Delta V^4\rangle e^{-2t/\tau_p} + \sigma_V^4 \left( 1-e^{-2t/\tau_p}\right)-\sigma_V^4}{\langle \Delta V^4\rangle - \sigma_V^4} \\
&= \frac{\left(\langle \Delta V^4\rangle - \sigma_V^4 \right)e^{-2t/\tau_p}}{\langle \Delta V^4\rangle - \sigma_V^4} = e^{-2t/\tau_p}\,.
\end{align}
\end{subequations}
This concludes the derivation of the ACFs in \cref{ref_acfs}.
\section{Pathological volume distributions in GROMACS}\label{appendix_pathological_distributions}
We show in this section some examples of problematic volume distributions obtained with the barostats available in GROMACS 2021.2. \Cref{fig:PR_aniso_prob,fig:PR_iso_prob,fig:MTTK_prob} show respectively three comparisons between "well behaving" and "pathological" volume distributions obtained from the simulations of the Argon crystal system described in \cref{LJ_crystal}.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/appendices/PR_aniso_ok.png}\\[1 ex]
\label{fig:PR_aniso_ok}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/appendices/PR_aniso_problem.png}\\[1 ex]
\label{fig:PR_aniso_notok}
\end{subfigure}
\caption{Volume distributions generated by the Parrinello-Rahman (PR) anisotropic barostat in GROMACS.}
\label{fig:PR_aniso_prob}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/appendices/PR_iso_ok.png}\\[1 ex]
\label{fig:PR_iso_ok}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/appendices/PR_iso_problem.png}\\[1 ex]
\label{fig:PR_iso_notok}
\end{subfigure}
\caption{Volume distributions generated by the Parrinello-Rahman (PR) isotropic barostat in GROMACS.}
\label{fig:PR_iso_prob}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/appendices/MTTK_ok.png}\\[1 ex]
\label{fig:MTTK_ok}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/appendices/MTTK_problem.png}\\[1 ex]
\label{fig:MTTK_notok}
\end{subfigure}
\caption{Volume distributions generated by the Martyna-Tuckerman-Tobias-Klein (MTTK) isotropic barostat in GROMACS.}
\label{fig:MTTK_prob}
\end{figure}
\section{Supplementary results}\label{appendix_supplementaryresults}
\begin{figure}[h!]
\centering
\includegraphics[width = 1.\textwidth]{core/third_chapter/box_comp_LJ.png}
\caption{Results from the simulations of the LJ crystal in SimpleMD. The average squared moduli of the cell vectors $\mathbf{a},\mathbf{b},\mathbf{c}$ and their average scalar products are shown respectively in the upper and lower plots, as functions of the relaxation time $\tau_p$. Deviations in the large-$\tau_p$ regime are related to the large statistical errors of the estimates, due to trajectories that are too short with respect to the autocorrelation times of the analyzed quantities. Note that these deviations are not so evident when observing the volume distributions (see \cref{fig:taups_LJ}), because the autocorrelation time of the single box matrix components is typically larger than $\tau_p$ (see for instance \cref{fig:ACFs_mod2_LJ,fig:ACFs_mod2_IceIh}).}
\label{fig:box_comp_LJ}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = 1.\textwidth]{core/third_chapter/box_vars_LJ.png}
\caption{Results from the simulations of the LJ crystal in SimpleMD. The variances of the distributions considered in \cref{fig:box_comp_LJ} are shown as functions of $\tau_p$.}
\label{fig:box_vars_LJ}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/third_chapter/ACFs_volume_LJ_TR.png}\\[1 ex]
\caption{\RaggedLeft Volume ACFs$\hspace{1.4cm}$}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[height = 4.4 cm]{core/third_chapter/ACF_var_LJ_TR.png}\\[1 ex]
\caption{\RaggedLeft ACFs of volume variance$\hspace{0.6cm}$}
\end{subfigure}
\caption{ACFs of the volume (left panel) and its variance (right panel) from time series from the simulations of a LJ crystal with the anisotropic SCR barostat, using the time-reversible integrator in SimpleMD. The dashed lines are the same exponential functions reported in \cref{fig:ACFs_LJ}.}
\label{fig:ACFs_LJ_TR}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = 1.0\textwidth]{core/third_chapter/ACFs_mod1_LJ.png}
\caption{Results from the simulation of the LJ crystal system in SimpleMD, using the anisotropic SCR barostat. The left panel shows the ACF of the squared modulus of the cell vector $\mathbf{a}$; the right panel shows the ACF of its variance. The dashed lines are exponential functions $e^{-t/(c\tau_{p,\text{exp}})}$, where the coefficient $c$ is set by hand ($c=1.75$ for the left plot, $c=0.9$ for the right one) to show that the autocorrelation time of interest appears linearly dependent on $\tau_p$, as in the case of the volume and its variance. However, extending this analysis to the other systems discussed in \cref{implementations}, the values of $c$ appear system-dependent and cannot be exactly predicted $\emph{a priori}$ (see for instance \cref{fig:ACFs_mod2_IceIh}). In the case of the LJ system, identical behaviours are observed for the ACFs of the remaining squared moduli and the ACFs of the scalar products, with the same parameters $c$ in both the classes of functions.}
\label{fig:ACFs_mod2_LJ}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = 1.0\textwidth]{core/third_chapter/ACFs_var_LJ_GMX.png}
\caption{Comparison between the ACFs of the volume variance from GROMACS simulations of the Argon crystal system, using the anisotropic SCR and PR barostats. Variance ACFs are computed as explained in \cref{LJ_crystal}, and dashed curves represent the exponentially decaying functions $\exp\left(-2t/\tau_{p,\text{exp}} \right)$. Comparing these ACFs with the volume ones in \cref{fig:ACFs_volume_LJ_GMX}, it is possible to conclude that a calculation of the volume variance converges faster than a calculation of the average volume using the SCR method. The ACFs obtained with the PR barostats show no symmetry with respect to zero, resulting in an integrated autocorrelation time that is larger than the one from SCR simulations.}
\label{fig:ACFs_var_LJ_GMX}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = 1.\textwidth]{core/third_chapter/box_comp_Ice.png}
\caption{Results from the simulations of the ice I$_h$ crystal. The average squared moduli of the cell vectors $\mathbf{a},\mathbf{b},\mathbf{c}$ and their average scalar products are shown respectively in the upper and lower plots, as functions of the relaxation time $\tau_p$.}
\label{fig:box_comp_ice}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = 1.\textwidth]{core/third_chapter/box_vars_Ice.png}
\caption{Results from the simulations of the ice I$_h$ crystal. The variances of the distributions considered in \cref{fig:box_comp_ice} are shown as functions of $\tau_p$.}
\label{fig:box_vars_ice}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = 0.9\textwidth]{core/third_chapter/ACFs_var_Ice.png}
\caption{ACFs of volume variance from the simulations of the Ice I$_h$ system. Dashed lines represent the expected decay $\exp\left(-2t/\tau_p\right)$.}
\label{fig:ACFs_vars_Ice}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = 1.0\textwidth]{core/third_chapter/ACFs_mod1_IceIh.png}
\caption{Results from the simulation of the Ice I$_h$ system in GROMACS, using the anisotropic SCR barostat. Left and right panels show respectively the ACF of the squared modulus of the cell vector $\mathbf{a}$ and the ACF of its variance. The dashed lines are the exponential functions $e^{-t/(c\tau_{p,\text{exp}})}$ with $c = 8$ and $c=4$ respectively, showing as a first approximation a linear scaling between the autocorrelation time of the quantity of interest and $\tau_p$. The ACFs of the other squared moduli and of the three scalar products between the cell vectors follow a similar behaviour, but with different parameters $c$.}
\label{fig:ACFs_mod2_IceIh}
\end{figure}
\begin{table}[h!]
\centering
\caption{Results from the simulations of the gypsum crystal in LAMMPS, with relaxation time $\tau_p = 1$ ps for both the SCR and the MTTK barostats. The first six raws are related to the distributions of the squared moduli and the scalar products of the three cell vectors $\mathbf{a},\mathbf{b},\mathbf{c}$, while the last six raws concern the distributions of the independent components of the internal pressure tensor $\mathbf{P}_{\text{int}}$. }
\begin{threeparttable}
\label{tab:gypsum_lammps_aniso_box}
\begin{tabular}{l c c c c}
\toprule[0.5pt]\toprule[0.5pt]
{} & \small\textbf{Anisotropic SCR} & \small\textbf{Anisotropic MTTK} \\\midrule
$\langle|\mathbf{a}|^2\rangle\,$ (nm$^2$) & $11.3993\pm0.0005$ & $11.39903\pm0.00010$ \\
$\sigma_{|\mathbf{a}|^2}\,$ & $0.0309\pm0.0002$ & $0.03080\pm0.00009$ \\ \midrule
$\langle|\mathbf{b}|^2\rangle\,$ (nm$^2$) & $9.3914\pm0.0007$ & $9.3915\pm0.0002$\\
$\sigma_{|\mathbf{b}|^2}\,$ & $0.0290\pm0.0003$ & $0.02925\pm0.00009$ \\ \midrule
$\langle|\mathbf{c}|^2\rangle\,$ (nm$^2$) & $16.5031\pm0.0015$ & $16.5041\pm0.0007$ \\
$\sigma_{|\mathbf{c}|^2}\,$ & $0.0691\pm0.0007$ & $0.0694\pm 0.0004$ \\ \midrule
$\langle\mathbf{a}\cdot\mathbf{b}\rangle\,$ (nm$^2$) & $0.00106\pm0.00010$ & $0.00001\pm0.00016$ \\
$\sigma_{\mathbf{a}\cdot\mathbf{b}}\,$ & $0.0366\pm0.0005$ & $0.03350\pm0.00019$ \\ \midrule
$\langle\mathbf{a}\cdot\mathbf{c}\rangle\,$ (nm$^2$) & $-6.5865\pm0.0007$ & $-6.58668\pm0.00014$ \\
$\sigma_{\mathbf{a}\cdot\mathbf{c}}\,$ & $0.0338\pm0.0003$ & $0.03409\pm0.00014$ \\ \midrule
$\langle\mathbf{b}\cdot\mathbf{c}\rangle\,$ (nm$^2$) & $-0.0007\pm0.0012$ & $-0.0000\pm0.0003$ \\
$\sigma_{\mathbf{b}\cdot\mathbf{c}}\,$ & $0.0480\pm0.0006$ & $0.0454\pm0.0005$ \\\midrule
$\langle P_{\text{int},xx} \rangle\,$ (bar) & $-14\pm10$ & $1\pm2$ \\
$\sigma_{P_{\text{int},xx}}\,$ & $1490\pm5$ & $1473\pm10$ \\ \midrule
$\langle P_{\text{int},yy}\rangle\,$ (bar) & $-8\pm10$ & $2\pm2$\\
$\sigma_{P_{\text{int},yy}}\,$ & $1545\pm4$ & $1534\pm12$ \\ \midrule
$\langle P_{\text{int},zz}\rangle\,$ (bar) & $-0.0 \pm9.6$ & $0.6\pm1.1$ \\
$\sigma_{P_{\text{int},zz}}\,$ & $1139\pm4$ & $1144\pm 4$ \\ \midrule
$\langle P_{\text{int},xy}\rangle\,$ (bar) & $5.112\pm6.57$ & $-0.3\pm 1.3$ \\
$\sigma_{P_{\text{int},xy}}\,$ & $887\pm3$ & $883\pm5$ \\ \midrule
$\langle P_{\text{int},xz}\rangle\,$ (bar) & $-2\pm8$ & $-0.3\pm 1.3$ \\
$\sigma_{P_{\text{int},xz}}\,$ & $850\pm3$ & $851\pm3$ \\ \midrule
$\langle P_{\text{int},yz}\rangle\,$ (bar) & $-7\pm7$ & $-0.2\pm1.3$ \\
$\sigma_{P_{\text{int},yz}}\,$ & $816\pm3$ & $813\pm6$
\\\bottomrule[0.5pt]\bottomrule[0.5pt]
\end{tabular}
\end{threeparttable}
\end{table}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 3.9 cm]{core/third_chapter/SE_avgs_Ice.png}\\[1 ex]
\caption{\RaggedLeft Errors of volume average$\hspace{0.3cm}$}
\end{subfigure}
\hfil
\begin{subfigure}[b]{0.44\textwidth}
\centering
\includegraphics[height = 3.9 cm]{core/third_chapter/SE_vars_Ice.png}\\[1 ex]
\caption{\RaggedLeft Errors of volume variance$\hspace{0.2cm}$ }
\end{subfigure}
\caption{Standard errors of volume average (left panel) and variance (right panel) computed with block bootstrap analysis, referred to the simulations of the Ice I$_h$ crystal. Dashed lines represent the expected behaviour of the errors as a function of $\tau_p$, computed with \cref{standard_error_autocorr}. For the SCR barostat, errors of the average volume show deviations from the expected behaviour when $\tau_p$ is small, resulting in an effective autocorrelation time that is larger than the barostat relaxation time. As commented for the volume ACFs in \cref{fig:ACFs_avgs_Ice}, this limit in the volume dynamics is given by the autocorrelation time in the rearrangement of atoms.}
\vspace{12cm}
\label{fig:SE_Ice}
\end{figure}
\end{appendices}
\pagestyle{plain}
\renewcommand*{\bibfont}{\small}
\printbibheading
\addcontentsline{toc}{chapter}{Bibliography}
\printbibliography[heading = none]
\end{document} |
\section{Introduction}
Many phenomena may be described through networks, including investment interactions between bidders and firms in venture capital (VC) markets \cite{liang2016} and professional relationships among firms \cite{bonaventura2020}.
Risk capital is an essential resource for the formation and growth of entrepreneurial venture and venture capital firms are often linked together in a network by their joint investments in portfolio companies \cite{bygrave1988}. Through connections in such a network, they exchange resources and investment opportunities with one another. Many studies show the impact of network dynamics on investments, raising efficiency \cite{wetzel1987informal} and providing precious information when there is a great level of information asymmetry \cite{fiet1995reliance}. Also, differentiating connection types and avoiding tight cliques appear to help the success of an investor by providing more diverse information and reducing confirmation bias \cite{bygrave1988}.
CB Insights \cite{cbins} provides records of all transactions in venture capital markets from 1948. Since data until 2000 are partial and discontinuous, we focus on the period 2000-2020, in order to minimize the impact of missing data on our analysis. Additionally, since different sectors may be characterized by different investment dynamics \cite{dushnitsky2006does}, we focus on the healthcare sector, which is of great importance and has shown to be less sensitive to market oscillations \cite{pisano2006}. This stability is also shared by returns of life science VC, where investments have a lower failure rate but are at the same time less likely to generate "black-swan" returns \cite{booth2011defense}, offering more consistency but a lower likelihood of achieving billion-dollars evaluations.
While the number of exits through an IPO or through a trade sale can be seen as a proxy for the success of an investor \cite{hege2003determinants}, there are instead different definitions of "success" for startups, but a common factor seems to be the growth rate of the company \cite{santi2017lit}. Our work aims to understand whether network features may affect "success" of investments in healthcare firms. In order to investigate this, we introduce progressively more nuanced definitions of "success", and analyze them with increasingly sophisticated statistical tools.
The paper is organized as follows. Section \ref{sec:char} introduces and characterizes a network of investors and firms, describing its structure and salient properties, including the communities emerging from its topology. Then, Section \ref{sec:success} focuses on the definition and analysis of "successful" firms. We first characterize "success" by looking at the funding trajectories of each firm, clustering these trajectories into two broad groups capturing a high and a low funding regime. The binary cluster membership labels provide a first, rough definition of "success". We run a logistic regression in order to explain "success" defined in this fashion with statistics computed on the network itself. We then move to more complex characterizations of "success": the total amount of money raised (a scalar) and the funding trajectory itself (a functional outcome). We run regressions also on these outcomes, to validate and refine our previous results. Finally, we discuss main findings and provide some concluding remarks in Section \ref{sec:disc}.
\section{Network characterization}
\label{sec:char}
The 83258 agents in the healthcare sector are divided into two broad categories: 32796 bidders, or investors, and 50462 firms. Companies open investment calls in order to collect funds; investors answer such calls and finance firms. Each deal, i.e. each transaction from an investor to a company, is recorded in the CB Insights' database. This market dynamics can be described by a \textit{bipartite network}, which indeed is built on the notion of dichotomous heterogeneity among its nodes. In our case, each node may be a firm or an investor, respectively. An undirected link exists between two nodes of different kinds when a bidder has invested into a firm. Of course, given the possibility for an investor to finance the same firm twice, the bipartite network is also a \textit{multi-graph}. By knowing the date in which investments are made, we can produce yearly snapshots of the bipartite network. A company (investor) is included in a snapshot of a certain year only when it receives (makes) an investment that year. By projecting the bipartite network onto investors and firms, we produce the two projected graphs which are used to compute all the node statistics described in Table \ref{tab:variables}. As the bipartite network is a multi-graph, defining projections on a subset of nodes requires an additional assumption. Specifically, we project the bipartite graph onto firms by linking them in a cumulative fashion: we iteratively add to each yearly projected snapshot a link between two companies in which a bidder has invested during that year. Concerning the projection of the bipartite network onto investors, we link two bidders whenever they invest in the same company in the same financing round.
\begin{table}[t]
\centering
\caption{Statistics computed on the projected graphs of investors and firms. Before running regressions in Section~\ref{sec:success}, left-skewed variables are normalized through log-transformation.}
\label{tab:variables}
\begin{tabularx}{\textwidth}{XX}
\toprule
\textbf{Variable} & \textbf{Network meaning}\\
\midrule
Degree centrality & Influence \\
Betweenness centrality \cite{hannan1977population} & Role within flow of information \\
Eigenvector centrality \cite{bonacich1987power} & Influence \\
VoteRank \cite{zhang2016identifying} & Best spreading ability \\
PageRank \cite{page1999pagerank}& Influence \\
Closeness centrality \cite{freeman1978centrality} & Spreading power (short average distance from all other nodes) \\
Subgraph centrality \cite{estrada2005subgraph} & Participation in subgraphs across the network \\
Average neighbor degree \cite{barrat2004architecture} & Affinity between neighbor nodes \\
Current flow betweenness centrality \cite{newman2005measure} & Role within flow of information\\
\bottomrule
\end{tabularx}
\end{table}
Roughly 75\% of the companies in the network projected onto firms are North American and European (around 55\% belong to the US market), while the remaining 25\% is mostly composed of Asian companies. Around 60\% of the companies operate within the sub-sectors of medical devices, medical facilities and biotechnology -- the pharmaceutical sub-sector alone accounts for 20\% of the network. As of August, 2021, roughly of 80\% the companies in the network are either active or acquired, with the remaining portion being inactive or having completed an IPO. We witness turnover of the active companies through the years, but this is expected: a company's status is evaluated as of 2021, and it is more likely to observe a dead company among those that received investments in 1999 than in 2018. Indeed, both death and IPO represent the final stage of the evolution of a company, so those that received funding in earlier years are more likely to have already reached their final stage. Finally, we do not observe marked changes in terms of graph sub-sectoral composition: the relative share of each sub-sector is rather stable through the years, with the exception of an increase in the shares of the internet software and mobile software sub-sectors (from 1\% in 1999 to 8\% in 2019 and from 0\% in 1999 to 5\% in 2019, respectively).
\subsection{Communities}
By employing the Louvain method \cite{blondel2008fast}, we identify meso-scale structures for each yearly snapshot of the network projected onto firms. For each year, we rank communities by their size, from the largest to singletons. We then compare the largest communities across years, by looking at their relative sub-sectors, status and geographical composition.
While the specific nodes in the biggest communities may vary throughout the years, we notice a relative stability in their features. The largest communities (which contain between 13\% and 20\% of the nodes) reflect the status composition of the general network, downplaying unsuccessful companies and giving higher relative weight to IPO ones, showing just a variation between acquired and active companies across years (i.e.~active companies are relatively over-represented in more recent largest communities than in older ones). Considering geographical information, the largest communities comprise mainly US companies, with an under-representation of other continents. This trait is quite consistent through the years, with the exception of
two years (2013-2014). With respect to sub-sectors, the largest communities mainly contain medical device and biotechnology companies, and they are quite consistent through the years in terms of sub-sectoral composition.
The second largest communities (containing between 10\% and 14\% of nodes in the network) have a less consistent sub-sectoral composition through the years, although it is worth highlighting that they comprise companies operating within software and technology. Geographically, we are still witnessing communities of mostly US-based companies, although 5 years out of 20 show a remarkable (roughly 80\%) presence of European companies. Finally, status composition is balanced between active and acquired until the later years, when active companies predominate within the second largest communities. IPOs are not present, while there are, in a small percentage (between 5\% and 20\%), dead startups.
Finally, the third largest communities (containing between 7\% and 12\% of the nodes) present a clear change within the period considered: in the first ten years, they mostly comprise failed or acquired European companies within the fields of biotechnology and drug development, while, in the second decade, they comprise active US companies within the fields of medical devices and medical facilities.
\section{Success analysis}
\label{sec:success}
Given the bipartite network and its projections, we now turn to the analysis of success and of its main drivers. Because of the elusiveness of the definition of "success", we proceed in stages -- considering progressively more refined outcomes and comparing our findings. Moreover, since many of the records available in the CB Insights' data set are incomplete, and our aim is to capture the temporal dynamics leading a firm to succeed, we further restrict attention to those companies for which full information is available on birth year, healthcare market sub-sector and investment history for the first 10 years from founding. Although this filtering may introduce some biases, it still leaves us with a sizeable set of 3663 firms belonging to 22 different sub-sectors.
Notably, we restrict our focus also in terms of potential predictors, due to the fact that our collection of network features exhibits strong multicollinearities. By building a feature dendrogram (Pearson correlation distance, complete linkage) and by evaluating the correlation matrix, we reduce the initial set to four representatives. In particular, we select two features related to the investors' projection (the maximum among the degree centralities of the investors in a company and the maximum among their current flow betweenness centralities, both computed in the company's birth year) and two features computed on the firms' projection (a company's eigenvector and closeness centralities, computed in the year in which the company received its first funding).
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{aligned_example_pharma.eps}
\caption{Money raised cumulatively as a function of time, shown for 319 firms in the pharmaceuticals and drugs sub-sector. Funding trajectories are constructed over a period of $10$ years since birth, and aligned using birth years as registration landmarks.}
\label{fig:aligned_pharma}
\end{figure}
Each firm has its own funding history: after its birth, it collects funds over the years, building a \textit{trajectory} of the amount of money it is able to attract. We treat these trajectories as a specific kind of structured data, by exploiting tools from a field of statistics called \textit{Functional Data Analysis} (FDA) \cite{ramsey2005functional}, which studies observations that come in the form of functions taking shape over a continuous domain. In particular, we focus on the \textit{cumulative} function of the money raised over time by each company. As an example, Figure \ref{fig:aligned_pharma} shows 319 such cumulative functions, for the firms belonging to the pharmaceuticals and drugs sub-sector. Trajectories are \textit{aligned}, so that their domain ("time") starts at each company's birth (regardless of the calendar year it corresponds to). By construction, these functions exhibit two characterizing properties: first, they are monotonically non-decreasing; second, they are step functions, with jumps indicating investment events.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\textwidth]{clustering_pharma.eps}
\caption{$k$-means clustering ($k=2$) of the funding trajectories of firms belonging to the pharmaceuticals and drugs sub-sector. The green and red dashed lines represent firms in the high ("successful") and low regimes, respectively. Bold curves represent cluster centroids. To aid their visualization, centroids are shown again in the right panel with individual trajectories in gray.}
\label{fig:clustered_pharma}
\end{figure}
Our first definition of success is based on separating these trajectories into two regimes characterized by high (successful) vs. low investment patterns: the first runs at high levels, indicating successful patterns, and the second at low levels. Because of heterogeneity among healthcare sub-sectors, we accomplish this by running a \textit{functional k-means clustering} algorithm \cite{jacques2014func, hartigan1979} with $k=2$, separately on firms belonging to each sub-sector.
As an example, companies belonging to the sub-sector of pharmaceuticals and drugs are clustered in Figure \ref{fig:clustered_pharma}. Throughout all sub-sectors, the algorithm clusters $89$ firms in the high-regime group and $3574$ in the low-regime one.
This binary definition of "success" turns out to be rather conservative; very few firms are labeled as belonging to the high investment regime. Consider the logistic regression
\begin{equation}
\label{eq:log}
\log\left(\frac{P(y_i=1)}{1-P(y_i=1)}\right)=\beta_0+\sum_{j=1}^{p} \beta_j x_{ij} \quad i=1,\dots n
\end{equation}
where $n$ is the number of observations, $y_i$, $i=1,\dots n$, are the binary responses indicating membership to the high ($y_i=1$) or low ($y_i=0$) regime clusters; $\beta_0$ is an intercept and $x_{ij}$, $i=1,\dots n$ and $j=1,\dots,p$ ($p=4$), are the previously selected scalar covariates.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{significance.eps}
\caption{Scatter plots of logistic regression coefficient estimates (horizontal) and significance (vertical; $-log(p$-$value)$). Each point represents one of $1000$ fits run on data balanced by subsampling the most abundant class. Orange solid line mark averages across the fits, and orange dashed lines $\pm1$ standard deviations about them. Green solid lines mark $0$ on horizontal axes. Blue line mark significance values associated to a p-value of $0.1$.}
\label{fig:est_coef}
\end{figure}
If we fit this regression on our unbalanced data, results are bound to be unsatisfactory and driven by the most abundant class. Running such a fit, one obtains an explained deviance of only $0.10$. To mitigate the effects of unbalanced data \cite{haibo2009}, we randomly subsample the most abundant class (the low-regime firms) as to enforce balance between the two classes, and then run the logistic regression in Equation \ref{eq:log}. We repeat this procedure 1000 times, recording estimated coefficients, associated p-values and explained deviances. The average of the latter across the 1000 replications is substantially higher than on the unbalanced fit, reaching $0.18$
(some fits produce deviance explained as high as $0.45$). Moreover, we can investigate significance and stability of the coefficient estimates through their distribution across the repetitions. Figure \ref{fig:est_coef} shows scatter-plots of these quantities, suggesting that the two variables related to the firms' centrality have a modest yet stable, positive impact on the probability of belonging to the high-regime cluster. This is not the case for the variables related to the investors' centrality.
This first evidence of a positive relationship between the success of a firm and its centrality, or importance (in a network sense) is promising. However, the binary definition of "success" we employed is very rough -- and the unbalance in the data forced us to run the analysis relying on reduced sample sizes ($89+89=178$ observations in each repeated run). Thus, we next consider a scalar proxy for "success", which may provide a different and potentially richer perspective. Specifically, we consider the cumulative end point of a firm's funding trajectory, i.e. the total value of the investment received through its temporal domain.
For this scalar response, we run a \textit{best subset selection} \cite{friedman2005elements} considering all the network features in our initial set -- not just the $4$ selected to mitigate multicollinearity prior to the logistic regression exercise. Notably, despite the substantial change in the definition of "success", results are in line with those from the logistic regression. Indeed, the first selected variable, when the predictor subset is forced to contain only one feature, is the eigenvector centrality of firms. When the predictor subset size is allowed to reach $4$, the features selected are the closeness and the VoteRank of the firm, and the maximum current flow betweenness centrality among its investors (computed on the firm's birth year). Thus, the only difference compared to our previous choice is the selection of the firms' VoteRank centrality instead of the maximum among the investors' degree centrality. We compare the two alternative selections of four features as predictors of the scalar "success" response fitting two linear models of the form:
\begin{equation}
y_i =\beta_0+\sum_{j=1}^{p} \beta_j x_{ij} +\epsilon_i \quad i=1,\dots n
\end{equation}
where $n$ is the number of observations, $y_i$, $i=1,\dots n$, are the scalar responses (aggregate amount of money raised); $\beta_0$ is an intercept; $x_{ij}$, $i=1,\dots n$ and $j=1,\dots,p$ ($p=4$), are the scalar covariates belonging to one or the other subset and $\epsilon_i$, $i=1,\dots n$, are i.i.d. Gaussian model errors. As shown in Table \ref{tab:linreg}, the maximum degree centrality among a firm's investors is not statistically significant.
Surprisingly, the maximum among investors' current flow betweenness centralities is significantly negative, but its magnitude is close to 0. In contrast, the firms' closeness and eigenvector centralities are positive, statistically significant and sizeable. This is in line with what we expected, since it is reasonable to think that knowledge may indirectly flow from other startups through common investors, increasing the expected aggregate money raised. Finally, the firms' VoteRank centrality appears to have a negative, statistically significant impact on the aggregate money raised. This should not be surprising, given that the higher the VoteRank centrality is, the less influential the node will be. The variance explained by the two models is similar and still relatively low ($R^2 \approx 0.13)$, which may be simply due to the fact that network characteristics are only one among the many factors involved in a firm's success \cite{dosilimmancabile1994}. Nevertheless, the results obtained here through the scalar "success" outcome are consistent with those obtained through the binary one and logistic regression.
\begin{table}[t]
\centering
\caption{Linear regressions of aggregate money raised on
two sets of predictors. All variables are scaled and some are log-transformed (as indicated parenthetically).}
\label{tab:linreg}
\begin{tabular}{@{\extracolsep{5pt}}lcc}
\toprule
& \multicolumn{2}{c}{\textit{Dependent variable:}} \\
\cline{2-3}
\\[-1.8ex] & \multicolumn{2}{c}{Aggregate money raised (log)} \\
\\[-1.8ex] & (1) & (2)\\
\midrule
newman\_max & $-$0.065$^{**}$ & $-$0.072$^{*}$ \\
& (0.030) & (0.041) \\
& & \\
voterank (log) & $-$0.140$^{***}$ & \\
& (0.033) & \\
& & \\
degcen\_max (log) & & 0.050 \\
& & (0.040) \\
& & \\
closeness & 0.126$^{***}$ & 0.130$^{***}$ \\
& (0.037) & (0.030) \\
& & \\
eigenvector (log) & 0.214$^{***}$ & 0.255$^{***}$ \\
& (0.034) & (0.028) \\
& & \\
Constant & 0.113$^{***}$ & 0.062$^{**}$ \\
& (0.030) & (0.025) \\
& & \\
\midrule \\[-1.8ex]
Observations & 1,118 & 1,364 \\
R$^{2}$ & 0.136 & 0.127 \\
Adjusted R$^{2}$ & 0.133 & 0.125 \\
Residual Std. Error & 0.992 (df = 1113) & 0.923 (df = 1359) \\
F Statistic & 43.951$^{***}$ (df = 4; 1113) & 49.458$^{***}$ (df = 4; 1359) \\
\bottomrule
\textit{Note:} & \multicolumn{2}{r}{$^{*}$p$<$0.1; $^{**}$p$<$0.05; $^{***}$p$<$0.01} \\
\end{tabular}
\end{table}
Our scalar outcome (aggregate money raised) has its own drawbacks. In particular, it implicitly assumes that the right time to evaluate success and investigate its dependence on network features is, cumulatively, at the end of the period considered (10 years). Note that this translates into a 10-year gap between the measurement of network features and financial success.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{fos_plot.eps}
\caption{Function-on-scalar regression, coefficient curve estimates. (a) intercept function (this can be interpreted as the sheer effect of time on the response); (b) maximum degree centrality among investors (company's birth year); (c) maximum across investors' current flow betweenness centrality (company's birth year); (d) company's eigenvector centrality; (e) company's closeness centrality. Dotted lines represent confidence bands. All the covariates are standardized.}
\label{fig:fos_coef}
\end{figure}
Although this issue could be approached relying on additional economic assumptions, we tackle it refining the target outcome and considering the full funding trajectories -- instead of just their end point. This requires the use of a more sophisticated regression framework from FDA; that is, function-on-scalar regression \cite{kokoszka2017introduction}. In particular, we regress the funding trajectories on the same two sets of covariates considered in the scalar case above. The equation used for function-on-scalar regression is:
\begin{equation}
\label{eq:fun_reg}
Y_i(t) = \beta_0(t) + \sum_{j=1}^{p} \beta_j(t) x_{ij} + \epsilon_i(t) \quad i=1,\dots n
\end{equation}
where $n$ is the number of observations; $Y_i(t)$, $i=1,\dots n$, are the aligned funding trajectories; $\beta_0(t)$ is a functional intercept; $x_{ij}$, $i=1,\dots n$ and $j=1,\dots,p$ ($p=4$), are the scalar covariates belonging to the one or the other set, and $\epsilon_i(t)$, $i=1,\dots n$, are i.i.d. Gaussian model errors.
The regression coefficient of a scalar covariate in this model, $\beta_j(t)$, is itself a curve describing the time-varying relationship between the covariate and the functional response along its domain. Together with the functional coefficients, we also estimate their standard errors, which we use to build confidence bands around the estimated functional coefficients \cite{refund2016}. Coefficient curve estimates for the covariate set including the maximum investors' degree centrality are shown in Figure~\ref{fig:fos_coef} (results are very similar with the other set of covariates). The impacts of an increase in the maximum among the degree centralities and in the maximum among the current flow betweenness centralities of the investors in a firm are not statistically significant. Conversely, eigenvector and closeness centralities of firms have positive and significant impacts. The impact of the eigenvector centrality seems to be increasing during the first five years, reaching a "plateau" in the second half of the domain. These findings reinforce those obtained with the binary and scalar outcomes previously considered, confirming a role for firms' centrality in shaping their success.
\section{Discussion}
\label{sec:disc}
This paper exploits techniques from the fields of network and functional data analysis. We build a network of investors and firms in the healthcare sector and characterize its largest communities. Next, we progressively shape the concept of a firm's "success" using various definitions, and associate it to different network features. Our findings show a persistent positive relationship between the importance of a firm (measured by its centrality in the network) and various (binary, scalar and functional) definitions of "success". In particular, we cluster funding trajectories into a high ("successful") and a low regime, and find significant associations between the cluster memberships and firms' centrality measures. Then, we switch from this binary outcome to a scalar and then a functional one, which allow us to confirm and enrich the previous findings. Among centralities computed on the two network projections, our results suggest a preeminent role for those computed in the companies' projection. In particular, both a firm high closeness centrality, indicating a small shortest distances to other firms, and its eigenvector centrality, which may account for a firm's reputation, seem to be related to the propensity to concentrate capital.
Our analysis can be expanded in several ways. First, we limit our study to the healthcare sector, while it may be interesting to investigate other fields, or more healthcare firms based on the availability of more complete records. It would also be interesting to account for external data (e.g. country, sub-sector, etc.) in two ways. One the one hand, these information would be useful as to compute more informative statistics on the network topology. On the other hand, they may be used in our regression, to control for these factors.
Moreover, meso-scale communities may be analyzed in terms of their longitudinal evolution, as to characterize "successful" clusters of firms from a topological point of view.
\section*{Acknowledgments}
F.C., C.E., G.F., A.M. and L.T. acknowledge support from the Sant'Anna School of Advanced Studies.
F.C. acknowledges support from Penn State University.
G.R. acknowledges support from the scheme "INFRAIA-01-2018-2019: Research and Innovation action", Grant Agreement n. 871042 "SoBigData++: European Integrated Infrastructure for Social Mining and Big Data Analytics".
|
\section{Introduction}
Discriminative models, especially those based on deep learning, have been proven effective in various medical image analysis tasks~\cite{shen2017deep}. However, such models primarily focus on discovering distinguishable patterns and features existing in medical images for down-stream analysis tasks, thus may neglect patterns that are characteristic of the images but not distinct enough for discriminative tasks.
Meanwhile, generative models provide a complementary way of learning all image patterns by modeling the entire data distribution. Towards a better comprehension of medical image attributes, we propose an attribute-guided generative adversarial network, termed AttributeGAN, to model the data distribution conditioned on different attributes and link the attribute values with image patterns and characteristics. Different from existing generative models proposed in the medical image domain for applications such as cross-modality translation~\cite{armanious2020medgan}, synthetic augmentation~\cite{xue2021selective} and image reconstruction~\cite{quan2018compressed}, we investigate the problem of synthesizing histopathology images conditioned on different image attributes to build a more controllable and interpretable medical image generative model.
Existing literature on controllable and interpretable image synthesis models~\cite{shen2020interpreting,shoshan2021gan} focus on noticeable attributes such as human body pose, hair color, age of human face, among others. However, attributes of medical images are more nuanced and harder to model and thus the problem of generating medical images based on controllable attributes is more challenging to solve. For conditional image synthesis, conditional GANs (cGANs)~\cite{odena2017conditional,miyato2018cgans} have utilized various types of discriminator networks to help the models capture the relationships between input conditions and image features. However, few of them work on multiple attribute inputs or are studied for medical image applications.
In this work, our goal is to develop an attribute-guided medical image synthesis model which can generate high-resolution and realistic images as well as make sure the generated images accurately reflect the attributes given to the model. We build upon a successful unsupervised generative model, leverage a carefully designed attribute-attention model, and employ a conditional contrastive learning strategy to efficiently model the conditional data distribution. Multiple attributes are one-hot encoded and concatenated with the noise vector and fed into different stages of the proposed model. Our proposed model generates photo-realistic histopathology images while being more controllable and interpretable than unconditional generative models. We conduct experiments on a histopathology dataset containing stained H\&E images of urothelial carcinoma and compare our proposed AttributeGAN with the state-of-the-art cGAN as well as different variants of our model. We summarize our contributions in this work as follows:
\begin{itemize}
\item[*] We propose a multi-attribute controllable generative model for high quality histopathology image synthesis. To the best of our knowledge, our work is the first to develop an attribute-aware GAN model with the capability to precisely control cellular features while preserving photo-realism for synthesized images.
\item[*] We incorporate efficient attention modules and conditional contrastive learning in both the generator and the discriminator to significantly improve quality as well as achieve better attribute-awareness of the generated images. Experiments on a histopathology dataset show better image quality using our proposed AttributeGAN than the state-of-the-art conditional GAN model.
\end{itemize}
\section{Methodology}
\begin{figure}
\includegraphics[width=\textwidth]{architecture2.pdf}
\caption{The architecture of our proposed controllable cellular attribute-aware generative model. Each color block of the attribute vector input represents a corresponding cellular attribute feature: cell crowding, cell polarity, mitosis, prominence of nucleoli and state of nuclear pleomorphism. Different colors in the feature space for contrastive learning refers to the label constructed by combination of 5 cellular attribute levels (\textit{e.g. (cell-crowding-severe, cell-polarity-completely-lacking, mitosis-frequent, nucleoli-prominent, pleomorphism-moderate)})} \label{architecture}
\end{figure}
To guarantee the quality of synthesized images, we build our model upon a recent unsupervised backbone generative model introduced by \cite{Liu2021TowardsFA}. For attribute-aware and controllable generation, we incorporate multi-attribute annotations of each image as the additional condition information to explicitly control the generation process. With attribute conditions inserted, the synthesized results are expected to
maintain sufficiently photo-realistic while accurately capturing the distinguishable image feature patterns within attributes. To fulfill the first goal, we adopt a skip-layer channel-wise excitation (SLE) module and include additional reconstruction loss in discriminator as in \cite{Liu2021TowardsFA}. SLE leverages learned feature patterns from a lower abstract level to further re-calibrate the channel-wise features map of higher scale. As demonstrated in the architecture of our proposed controllable cellular attribute-aware generative model in Fig.~\ref{architecture}, in addition to the basic structure of SLE, we further improve the backbone by incorporating a global attention pooling for context modeling~\cite{Cao2020GlobalCN} at earlier stages of upsampling before the transformation through the bottleneck blocks to capture channel-wise dependencies.
For the second goal of attribute learning, while existing conditional GANs (cGANs)~\cite{miyato2018cgans} concatenate noise vectors with the conditional vectors and leverage projection discriminator for condition correctness, such models may not be able to capture the nuanced changes in attribute levels of medical images.
In addition to input concatenation and projection discriminator, we integrate conditional contrastive losses~\cite{kang2020contragan} to both discriminator and generator to exploit the relation between images and the attributes they contain.
Integrating a self-supervised learning based module to exploit data-to-data and data-to-attribute relations within a mini-batch of proper size comes with two merits. First, with known attributes available for reference, the performance no longer heavily relies on the hard negative and positive samples mining. We consider the anchor image itself together with real images with the same attribute combination as positive samples, while real images with different attribute combinations in the same mini-batch as negative samples. Second, the performance of capturing the explicitly distinguishable feature representations in a fine-grained manner is substantially enhanced. Theoretically,
this is achieved by minimizing the feature-level distances between positive samples while maximizing the distances between positive and negative samples. During training, the knowledge of attribute-dependent feature distinction learned by the discriminator is then passed to the generator for synthesizing images that are more sensitive to inter/intra-attribute characteristics. The effectiveness is further proven empirically in the qualitative ablation study of model architecture with and without the contrastive learning objective as demonstrated in Fig.~\ref{ablation}.
To elaborate, first we denote $X = \{x_1, x_2, ..., x_n\}$ as extracted features from the intermediate stage of discriminator, and $Y = \{y_1, y_2, ..., y_n\}$ as the combination of multiple attributes. Intuitively, after mapping data to the hypersphere via feature and attribute projectors $f$ and $l$, our goal is to push the inter-attribute samples further and pull intra-attribute ones closer at the feature level. Thus, our conditional contrastive loss is formulated as:
\begin{equation}
\mathcal{L}(x_i, y_i; t) = - \log \left ( \frac{
\exp(\frac{f(x_i)^\top l(y_i)}{t}) +
\sum^n_{k=1} \mathbbm{1}_{y_k = y_i} \cdot \exp(\frac{f(x_i)^\top f(x_k)}{t})
}{
\exp(\frac{f(x_i)^\top l(y_i)}{t}) +
\sum^n_{k=1} \mathbbm{1}_{k \neq i} \cdot \exp(\frac{f(x_i)^\top f(x_k)}{t})
} \right ),
\end{equation}
where $\mathbbm{1}$ is the indicator function and the scalar value $t$ plays the role as the regularizer to balance the push and pull force among samples across different and within the same group of attributes.
Recent GAN models~\cite{zhang2019self} for image synthesis have adopted the self-attention module~\cite{vaswani2017attention} to capture long-range dependencies within the image. However, the dot-product based self-attention can quadratically increase computational complexity and constrain the number of images inside each batch. Meanwhile, the aforementioned contrastive learning efficiency heavily relies on a relatively large batch size as both data-to-data and data-to-attribute relation learning would be seriously compromised with a small batch size and insufficient number of positive/negative pairs. Hence, in order to free up more space to accommodate a larger volume of data in each batch and train with lower computational complexity, we apply a more efficient equivalence~\cite{shen2021efficient} of self-attention. As illustrated in the efficient attention module in Fig.~\ref{architecture}, feature vectors at intermediate stages in both generator and discriminator are projected onto three latent spaces through convolution operations termed as query, key and value as in the original self-attention~\cite{vaswani2017attention}, and denoted as $Q, K, V$, respectively. Here, $Q \in \mathbb{R}^{(H*W) \times d_q}, K \in \mathbb{R}^{(H*W) \times d_k}, V \in \mathbb{R}^{(H*W) \times d_v}$, $d_q = d_k$ and $H, W$ refer to the height and width of the image. Leveraging the associative property of matrix multiplication, rather than start with the multiplication of $QK^T$ as formulated in \cite{vaswani2017attention} to measure the pair-wise similarity exhaustively, instead we begin with the multiplication between $K^T$ and $V$. It is feasible because $(\frac{QK^T}{n})V = \frac{Q}{\sqrt{n}}(\frac{K^T}{\sqrt{n}}V)$.
Following this procedure, we obtain a matrix $g \in \mathbb{R}^{d_k \times d_v}$, representing the intermediate global context vector with dimension of $d_v$ in $d_k$ channels after aggregating from $H*W$ positions through weighted summation. At the next step, the context vector is acquired by having each pixel gathering positional features from all $d_k$ channels for $d_v$ dimensions, by multiplying $Q$ and the result of $(K^TV)$. With the efficient attention, the memory complexity is reduced to $\mathcal{O}(d_k*d_v)$ from the original $\mathcal{O}(n^2)$, escalating convergence speed and freeing up more space, making it possible for conditional contrastive learning to deliver its performance to the fullest.
More specifically, for conditional attributes, we encode the input condition into a one-hot vector with attribute level labels for all five cellular features. The attribute vector is later concatenated with the input noise vector after the initial stage of upsampling both vectors using transposed convolution operators. For synthesizing images with resolution $512 \times 512$, efficient attention modules are applied in two intermediate upsampling stages at $32 \times 32$ and $64 \times 64$ resolutions as shown in Fig.~\ref{architecture}. For each upsampling block without attention module, input images first go through an upsampling layer with scale factor set as 2, immediately followed by a gaussian blurring kernel for antialiasing. Next, to enlarge the feature learning space channel-wise, a basic block including a convolutional layer, a batch normalization layer and an activation layer is added as another major component in each individual upsampling block. Gated Linear Units (GLU) is utilized for every activation layer in the AttributeGAN architecture, as it has shown quality-improving potential over the commonly used ReLU or GELU activations~\cite{shazeer2020glu}. Additionally, three skip-layer connections are applied at the resolutions of $16\times16$, $32 \times 32$, and $64 \times 64$ to $128 \times 128$, $256\times 256$ and $512 \times 512$ in order to strengthen the gradient signals between layers.
For the discriminator, the conditional attributes are required together with either synthesized or real images to be further utilized in a projection based discrimination. Attribute vectors are fed into a feed-forward layer before being incorporated into the output of discriminator. As shown in Fig.~\ref{architecture}, at the resolution of $128 \times 128$ the feature vectors and the attribute level information are projected to an embedded space for contrastive learning, which is later included in the losses for the discriminator. To further refine the discriminator's capability of capturing a more comprehensive feature map to be differentiated from the fakes, two auxiliary reconstruction losses are added. We utilize two additional simple decoders trained within the discriminator for the $8 \times 8$ and $16\times16$ feature vectors, and calculate the mean squared error (MSE) for both in the reconstruction loss.
\begin{figure}[tbp]
\includegraphics[width=0.96\textwidth]{result.pdf}
\caption{The AttributeGAN generated histopathology patches based on different levels of cell crowding and the state of nuclear pleomorphism. The input noise vector is set to be identical for each sample, which explains the resemblance of general shape and texture shared within each column and rich diversity across columns. Zoom in for better view.}
\label{results}
\end{figure}
\section{Experiments and Results}
\noindent \textbf{Dataset}. We conduct comprehensive experiments on a histopathology dataset representing patients with bladder cancer collected by \cite{zhang2019pathologist}. The dataset contains $4,253$ histopathology image patches with $512\times 512$ resolution. Each patch is accompanied with a paragraph of pathology report descriptions provided by multiple experienced pathologists.
Each report follows a template format that describes 5 types of key morphological visual cellular features essential for classifying urothelial carcinoma, including cell crowding, cell polarity, mitosis, prominence of nucleoli and state of nuclear pleomorphism.
To achieve a more concise representation of attributes and their levels, we extract feature-describing keywords in the report as annotations (see Table 1-5 in Supplementary Materials).
Converting raw reports to categorical levels for each cellular attribute facilitates the manipulation of semantic editing in our experiments, as demonstrated in Fig.~\ref{results}. There are 4, 3, 3, 2, 4 levels assigned to describe different degrees of cell crowding, cell polarity, mitosis, nucleoli and pleomorphism, respectively. Following this procedure, each patch is paired with a combination of levels from all 5 cellular attributes. To accelerate the learning of attribute-relevant patterns, we discard the combinations with frequency less than the $20^{th}$ percentile since most of those merely appear in the dataset once or twice.
\noindent \textbf{Implementation Details}.
AttributeGAN is trained on two NVIDIA Quadro RTX 6000 GPUs each with 24GB RAM in parallel by applying the PyTorch DistributedDataParallel module together with SyncBatchNorm. The GPU space freed up from the attention module efficiency enables a larger batch size. In our experiments, the batch size is set as 64, and each device processes half of the inputs from the current batch. The learning rate is fixed to be $2e-4$ throughout the entire 50000 steps of training.
\begin{figure}[tbp]
\includegraphics[width=0.9\textwidth]{ablation.pdf}
\caption{Generated images from ablation study. The selected attribute for ablation study is cell crowding. We present the comparison among images synthesized using our baseline model BigGAN, our proposed AttributeGAN w/o the efficient attention module, our proposed AttributeGAN w/o the conditional contrastive loss, and real images for 4 levels of cell crowding: normal, mild, moderate and severe. Zoom in for better view.}
\label{ablation}
\end{figure}
We present example images generated by our AttributeGAN in Fig.~\ref{results}. To demonstrate the smooth progression through different attribute levels and showcase disentanglement among attributes, the input attribute is framed as a 5-dimensional vector where we only alter one attribute at a time inside each result batch. Other than the attribute whose level is being varied, the remaining dimensions are fixed to be a combination of the other four attributes that frequently appear in the dataset. With attribute conditions given in such manner, the generated images show clear progressions in cellular pattern in accordance with the changes in input attribute condition.
To examine the effectiveness of our proposed AttributeGAN and its components, we compare images generated by different models as well as the real images in Fig.~\ref{ablation}.
Various well-developed and extensively-used models are relevant to attribute-controlling, such as Fader Networks \cite{lample2017fader}, StyleGAN \cite{karras2019style} and StyleGAN v2 \cite{karras2020analyzing}. Although the aforementioned models present state-of-the-art results on photo-realism and attribute accuracy, they are not suitable to be directly compared with our approach for the conditional histopathology image synthesis task, because they are designed for slightly different goals such as semantic editing of assigned attributes (e.g. the Fader Networks), or unconditional image synthesis (e.g. StyleGAN, StyleGAN v2).
Hence we consider the state-of-the-art conditional GAN model, BigGAN~\cite{brock2018large}, as the most appropriate baseline model. Since BigGAN can only handle single-dimensional condition, we train 5 different BigGAN models for different attributes. Considering that BigGAN consumes larger memory and requires longer time to converge, we train all baseline BigGAN models with image resolution $256 \times 256$. During comparison, we resize all images to the same size for fair comparison. One can observe that images generated by our models show superb realism. Compared with the BigGAN model, different variants of AttributeGAN model keep the global shape and texture well inside each column. On the contrary, the global image pattern changes for BigGAN given different attribute level inputs. For variants of our AttributeGAN, our proposed model without the attention module generates less realistic images, and the model without the conditional contrastive learning reacts less responsively to the changes in attribute level. The full AttributeGAN model respects the changes in attribute level and retains the global patterns well.
\begin{table*}[tbp]
\begin{center}
\begin{tabular}{l|c|c|c|c|c|c}
\hline
\multirow{3}{*}{\textbf{Methods}} &
\multirow{3}{*}{\textbf{FID}$\downarrow$} &
\multicolumn{5}{c}{\textbf{Attribute Error}$\downarrow$} \\
\cline{3-7}
{}&{}& Cell & Cell & \multirow{2}{*}{Mitosis} & \multirow{2}{*}{Nucleoli} & \multirow{2}{*}{Pleomorphism} \\
{}&{}& Crowding & Polarity & {} & {} & {} \\%[1ex]
\hline
Real Images* & - & .011 & .034 & .037 & .018 & .014 \\%[3ex]
\hline
BigGAN \cite{brock2018large} & 158.39 & .112 & .080 & .104 & \textbf{.049} & .065 \\
\hline
AttributeGAN (Ours) & \multirow{2}{*}{142.015} & \multirow{2}{*}{.035} & \multirow{2}{*}{\textbf{.078}} & \multirow{2}{*}{.208} & \multirow{2}{*}{.056} & \multirow{2}{*}{\textbf{.023}} \\
\enspace \textit{w$\slash$o} EA & & & & & & \\%[3ex]
\hline
AttributeGAN (Ours) & \multirow{2}{*}{55.772} & \multirow{2}{*}{.094} & \multirow{2}{*}{.112} & \multirow{2}{*}{.111} & \multirow{2}{*}{.056} & \multirow{2}{*}{.070} \\
\enspace \textit{w$\slash$o} CCL & & & & & & \\%[3ex]
\hline
AttributeGAN (Ours) & \textbf{53.689} & \textbf{.021} & .098 & \textbf{.088} & .081 & .063 \\%[3ex]
\hline
\end{tabular}
\end{center}
\caption{Quantitative evaluation results of different methods.
Real images are from a holdout validation set during fine-tuning of the pre-trained classifier. Note that we report BigGAN results from five independently trained BigGAN models for five attributes as it can only work with single attribute inputs. EA refers to Efficient Attention module and CCL refers to the Conditional Contrastive Loss.}\label{result_table}
\end{table*}
In Table~\ref{result_table}, we show quantitative comparison results between different models. Following conventions in image synthesis works~\cite{zhang2019self,brock2018large}, we adopt Fréchet Inception Distance (FID)~\cite{heusel2017gans} score which has shown to correlate well with human perception of realism. FID measures the Fréchet distance between two multivariate Gaussians fit to features of generated and real images extracted by the pre-trained Inception V3~\cite{szegedy2016rethinking} model. Compared with BigGAN whose FID score is averaged from five BigGAN models, all AttributeGAN variants achieve better FID score indicating better realism. After including the attention module, the FID score improved significantly for the AttributeGAN model. To better evaluate the correctness of represented attributes, we further calculate an Attribute Error to measure the discrepancy between attribute levels predicted by an ImageNet pre-trained ResNet18~\cite{he2016deep} model fine-tuned on the histopathology dataset and the groundtruth attribute levels. Images generated by all models are first normalized to same resolution $224 \times 224$ for fair comparison. All attribute levels are normalized to the range $[0,1]$ and the MSE of the predicted attributes and the groundtruth attributes are computed as the attribute error value. During fine-tuning of the ResNet18 model, we keep a holdout validation set and the corresponding attribute error evaluated on the holdout real images are also reported in Table~\ref{result_table}. For BigGAN and our proposed AttributeGAN without attention, although they achieve small attribute errors for certain attributes, the quality of generated images are lower which makes them differ more from real images, thus the attribute prediction model trained on real images may not be able to correctly predict the attribute level for such images. Compared to AttributeGAN without contrastive learning, the full AttributeGAN generally gets lower attribute error, especially on cell crowding. Based on both the qualitative and quantitative comparisons, we prove the necessity of the attention module and the conditional contrastive loss, and show that one multi-attribute AttributeGAN model can generate images with better quality than multiple BigGAN models for conditional histopathology image synthesis.
\section{Discussion}
To assess the quality of the generated images and how well the images correspond to the input attribute levels, we presented five sets of images that were generated based on different cellular attribute levels to two expert pathologists. Both pathologists commented that the synthetic images are remarkably good in resembling routinely stained H\&E images of urothelial carcinoma. In the set of images generated according to different levels of cell crowding (see examples in Fig.~\ref{results}-Left), the crowding of nuclei occurs appropriately overall at each of the described levels, and the degree of crowding remains within the realm of reality, although for a few images, the increase in crowding seems to be by increasing the epithelial/stromal ratio, rather than increasing the density of cells within the same amount of epithelium. For the set of images generated according to different levels of pleomorphism (see examples in Fig.~\ref{results}-Right), an increase in nuclear pleomorphism was observed as the images progress through the pleomorphism prominence levels. For the other three sets of images generated based on different levels of cell polarity, mitosis, and prominence of nucleoli (see Figures 1-3 in Supplementary Materials), the pathologists commented that no obvious progression was observed through those sequences of images. We plan to further investigate these three attributes in our future work, study whether the attributes are correlated in some fashion in real images and learn how to improve the responsiveness of generated images to varying input conditions.
\section{Conclusion}
In this work, we present a multi-attribute guided generative model, AttributeGAN, for synthesizing highly realistic histopathology images. Images generated by the proposed model show smooth progression through different input attribute levels and contain photo-realistic patterns. With the quality of synthesized images, AttributeGAN can be potentially used for medical education or training and support various medical imaging applications.
\bibliographystyle{splncs04}
|
\section{Introduction}
The occurrence of solitons was first described in the literature as early as 1834 by John Scott Russell. Theoretical studies of topologically stable solitons began in earnest in the mid-1960s,
see e.g.\!\! the collection~\cite{soliani1984solitons}. The discovery of instantons~\cite{BPST} and 't Hooft-Polyakov monopoles~\cite{hooft1974magnetic,polyakov1974particle} after the advent of Yang-Mills theories provided an impetus to further rapid advances in this area.
In this paper we will investigate the close interrelationship of several notions of topological charge in systems with three spatial dimensions. In particular we will focus on a lattice spin system which supports three-dimensional Skyrmions in the original sense of Skyrme~\cite{skyrme,WITTEN1983433,adkins1983static} and connect these to Hopfions~\cite{faddeev1997stable} (reviewed in~\cite{radu2008stationary}) which are widely discussed as a generalization of the two-dimensional magnetic Skyrmions (also known as baby Skyrmions) which are more familiar in a condensed matter context (see e.g.~\cite{Nagaosa2013,han2017skyrmions,Back2020}).
While much work on magnetic Skyrmions involves chiral ferromagnets with a Dzyaloshinskii-Moriya interaction~\cite{DzMo}, more recently there has been a lot of interest in topological spin textures in inversion-symmetric magnets~\cite{okubo2012multiple,leonov2015multiply,lin2016ginzburg,Wang2020} where the magnetic Skyrmions are stabilized by competing magnetic interactions. The three-dimensional model we will consider here is a frustrated spin system of this type, and in one limit is very close to the effective theory studied by Sutcliffe~\cite{sutcliffe2017skyrmion} as a medium for Hopfions.
A Hopfion is a topological defect which may be considered to be a magnetic Skyrmion extended in the third dimension to form a loop of string. Such magnetic Skyrmion strings have been observed experimentally~\cite{milde2013unwinding,seki2021direct}, and moreover Hopfions themselves have recently been constructed in magnetic systems~\cite{kent2021creation}. {There has also been much discussion of Hopfions in two-component superconductors \cite{babaev2002hidden,gorsky2013revisiting}, which although in the simplest case are not energetically stable \cite{jaykka2011supercurrent} they may perhaps be stabilized by a current-current interaction \cite{rybakov2019stable}.} The topological stability of such Hopfions is provided for by the Hopf invariant~\cite{whitehead1947expression} which is closely analogous to the notion of helicity in hydrodynamic systems~\cite{kamchatnov1982topological,moffatt2014helicity}, and in this sense knotted `Hopfions' have even been constructed experimentally in fluids~\cite{kleckner2013creation}.
Such knotted topological solitons have captured the interest of many working with the model of Faddeev and Niemi~\cite{faddeev1997stable,faddeev1976some,faddeev1997toroidal,battye1998knots,sutcliffe2007knots}\footnote{{This model is often referred to as the Skyrme-Faddeev model, but in order to more clearly distinguish it from the Skyrme model proper we will refer to it as the \emph{Faddeev-Niemi} model.}}, and in such papers attention is often brought to the old idea of Lord Kelvin suggesting knotted vortices might be related to elementary particles~\cite{LordKelvin}. Of course there is a different but related model involving topological defects which really is believed to have some connection to baryons, and that is the \emph{Skyrme model}~\cite{skyrme,WITTEN1983433,adkins1983static} (see also the review~\cite{makhan1992skyrme}).
Skyrmions proper in the sense of the three-dimensional {(3D)} Skyrme model are not as familiar in a condensed matter context, but
{they have} been predicted to emerge near the Lifshitz point (commensurate to incommensurate transition) of 3D magnetic systems with non-collinear ground states~\cite{BatistaEtAl2018}.
For certain choices of parameters,
the {second-derivative terms of the} effective field theory of
{such frustrated lattice models are} equivalent to the $SU(2)$ principal chiral model which is a major component of the Skyrme model. The simplified lattice model considered here
{can be regarded as the low-energy effective model for a broad class of 3D frustrated magnets with non-collinear ground states,}
and it too reduces to the principal chiral model in one limit of parameters. While the higher order terms in the effective field theory expansion of the frustrated lattice models differ from the Skyrme term, they too
{were predicted to}
stabilize 3D Skyrmions {near the Lifshitz point}~\cite{BatistaEtAl2018}, as will be shown here explicitly.
The lattice model considered here involves multiple spins at each lattice site of a cubic lattice, much like how a pyrochlore lattice involves four spins at each tetrahedral cell of an outer face-centered-cubic lattice. Large scale rigid rotations of these spins lead to $SO(3)$ Goldstone modes much as in the principal chiral model. Typically the low-energy theory of realistic frustrated magnetic models involves both these rigid rotations as well as other modes that modify the relative angles between neighboring spins. However these extra modes can be gapped out by certain interactions, for instance the biquadratic term on the pyrochlore lattice model~\cite{BatistaEtAl2018}. So to provide a low-energy effective model for a large class of realistic 3D non-collinear magnets we may simply take the relative angles between spins on the same site to be fixed. A key idea used here is that when the relative angle between spins takes a special value then the effective description is much like the Skyrme model, i.e. a non-linear sigma model on a three-sphere $S^3$. As the vacuum configuration of spins becomes more and more colinear the target space deforms to a \emph{squashed sphere} \cite{azaria1995massive,squashedsigma}, and in the limit of perfect colinearity the model becomes equivalent to a non-linear sigma model with an $S^2$ target space much as is in the Faddeev-Niemi model or the frustrated magnetic models previously considered as a host to Hopfions \cite{sutcliffe2017skyrmion}.
A very similar continuum model deforming the Skyrme model to the Faddeev-Niemi model has been considered previously by Nasir and Niemi~\cite{nasir2002effective} and Ward and Silva Lobo~\cite{ward2004skyrmions,lobo2011generalized,silva2011lattices}. We will also reconsider this model here as a close analogy to the lattice model, and show that it obeys a linear Bogomol’nyi-Prasad-Sommerfeld (BPS)~\cite{BPS} bound on the energy. Note however that the connection between 3D Skyrmions and Hopfions in these models is entirely different from another notion of Hopfions in the Skyrme model pointed out by Meissner~\cite{meissner1985toroidal} and Cho~\cite{cho2001monopoles,cho2008new}. In our case we wish to stress that the natural projection map from $S^3$ to $S^2$ implies that 3D Skyrmions themselves may be considered as Hopfions and vice versa. This is an idea that occasionally appears in the literature and in particular underlies Ward's treatment of the continuum model~\cite{ward2004skyrmions}. The squashing of the sphere changes the energy functional and thus the quantitative features of the minimum energy soliton, but there is no dramatic qualitative difference between the topological defects of the Skyrme model and the Faddeev-Niemi model.
The analogue of this statement will also be shown explicitly in numerical simulations of the frustrated magnetic model which we introduce here. First it will be shown that the model has 3D Skyrmion solutions much like the Skyrme model. The unit Skyrmion will be seen to be quantitatively very close to the spherically symmetric hedgehog solution, and small clusters of Skyrmions may be approximated by the same rational map ansatz used for both the Skyrme model and BPS monopoles in $SU(2)$ Yang-Mills~\cite{houghton1998rational,battye2002skyrmions}. However as the charge increases the Skyrmions in the lattice model will exhibit new shapes departing from the rational map ansatz. And as the squashing parameter of the model increases the Skyrmion clusters, which are similar in some respects to models of nuclei, will be shown to deform to twisted{, linked} or knotted strings, much like a modern incarnation of Lord Kelvin's idea.
\subsection{Outline}
This paper is divided into two main sections. Sec.~\ref{sec 2} deals with the effective theory of the squashed sphere sigma model and it introduces the notation and necessary topological concepts in that context in Sec.~\ref{sec Squashed sphere} and \ref{sec Top charge} respectively. The continuum model of Nasir, Niemi~\cite{nasir2002effective} and Ward~\cite{ward2004skyrmions} interpolating between the Skyrme model and the Faddeev-Niemi model is reconsidered with some new numerical simulations and a new theoretical result on energy bounds in Sec.~\ref{sec Ward model}. Finally the notion of the equivalence between 3D Skyrmions and Hopfions is discussed a bit further in Sec.~\ref{sec Position curves and strings}, where an ansatz for Skyrmion strings which have baryon charge per length is also introduced and compared to previous results in the Faddeev-Niemi model.
Then in Sec.~\ref{sec 3} the main frustrated magnetic system is considered. Most of the discussion in Sec.~\ref{sec 2} will be applicable to the continuum description of this model as well. The lattice model is introduced in Sec.~\ref{sec Lattice model intro} and its continuum description is found in Sec.~\ref{sec Lattice model effective theory}. The details of the numerical simulation are introduced in Sec.~\ref{sec Lattice model numerical results} and results on the unit charge Skyrmion are compared to the hedgehog ansatz in the continuum description. In Sec.~\ref{sec Rational map} higher charge Skyrmion configurations are considered and compared to the rational map ansatz in the continuum description.
{In Sec.~\ref{sec charge-10} we show numerical results for the charge-10 Skyrmion, with emphasis on the interpolation of the topological charge isosurfaces and the position curves. }
Finally, in the concluding Sec.~\ref{sec Conclusion} the possible connection to experiment and further investigation of the theoretical model in terms of Skyrmion lattices is discussed.
\section{Squashing the Skyrme model}\label{sec 2
Since our aim is to introduce a model which is closely related to both the Skyrme model and the Faddeev-Niemi model, let us begin by reviewing these continuum models and illustrating the connection between them. The notation and discussion on topological charge in this context will be directly applicable to the lattice model which is our main focus in the next section.
\subsection{The squashed sphere non-linear sigma model} \label{sec Squashed sphere}
The terms of the Skyrme model which are quadratic in derivatives are identical to the $SU(2)$ principal chiral model (PCM), which is expressed in terms of a matrix field $U\in SU(2)$ and a parameter $f_\pi$ with dimensions of energy,
\begin{align}
\mathcal{L}_{PCM}=\frac{f_\pi^2}{4}\text{Tr}\left(\partial_\mu U^{-1}\partial_\mu U\right).\label{lagr PCM U}
\end{align}
This has a global symmetry under right multiplication $U\rightarrow UV_R$, and the three independent Noether currents $J^i_\mu$ corresponding to this symmetry are
\begin{align}
J^i_\mu = \frac{1}{2}\text{Tr}\left(J_\mu \sigma^i\right),\qquad J_\mu \equiv -i U^{-1}\partial_\mu U,\label{def J currents}
\end{align}
where $\sigma^i$ are the standard Pauli matrices with normalization $\text{Tr}\left(\sigma^i\sigma^j\right)=2\delta^{ij}$. Momentarily we will consider models where the global symmetry associated with $J$ is explicitly broken to a $U(1)$ subgroup (although the full global symmetry under \emph{left} multiplication will be maintained) but these quantities $J$ will still be very useful, and the Lagrangian may be expressed in terms of them,
\begin{align}
\mathcal{L}_{PCM}=\frac{f_\pi^2}{2}\sum_{i=1,2,3}\left(J^i_\mu\right)^2.\label{lagr PCM J}
\end{align}
On the other hand, we also wish to consider the Faddeev-Niemi model which involves a three-component real unit vector $\left(S^i\right)^2=1$, and the quadratic terms in the action are just that of the $O(3)$ non-linear sigma model. As usual, the action may instead be expressed in the form of a complex two-component unit vector $z^\alpha$, which is connected to real unit vector $S^i$ through the Pauli matrices $\sigma^i$,
\begin{align}
S^i\equiv -\bar{z}^\alpha \sigma^i_{\alpha\beta}z^\beta, \qquad \bar{z}^\alpha z^\alpha=1.\label{def S and z}
\end{align}
This change of fields leads to the $CP^1$ form of the non-linear sigma model,
\begin{align}
\mathcal{L}_{CP^1}=\frac{f_\pi^2}{2}\left(\partial \bar{z}\cdot\partial z +\left(\bar{z}\cdot\partial z\right)^2\right)=\frac{f_\pi^2}{8}\left(\partial {S}\right)^2, \label{lagr CP1 S and z}
\end{align}
where the indices will be suppressed where obvious, and a dot may be used to clarify contraction of internal indices.
The $CP^1$ model above may be related to the PCM by expressing the action in terms of the special unitary matrix $U$ which is uniquely determined by $z$,
\begin{align}
U= \left(\begin{array}{cc}
\bar{z}^1& z^0\\
-\bar{z}^0 & z^1
\end{array}\right), \label{def U and z}
\end{align}
and then further in terms of the $J$ currents defined above in \eqref{def J currents},
\begin{align}
\mathcal{L}_{CP^1}=\frac{f_\pi^2}{2}\sum_{a=1,2}\left(J^a_\mu\right)^2. \label{lagr CP1 J}
\end{align}
The only difference from the PCM case \eqref{lagr PCM J} is that the sum only runs over two components. To better distinguish the two cases even in the absence of an explicit summation symbol, a Latin index from the beginning of the alphabet will run over $1,2$, and a Latin index from the middle of the alphabet will run over $1,2,3$.
In this form there is an obvious interpolation between the PCM and $CP^1$ model which may be constructed in terms of a parameter $\beta$ ranging from $0$ to $1$, respectively\footnote{$\beta$ may also be continued to negative values, which is relevant in e.g. \cite{dombre1989nonlinear}.},
\begin{align}
\mathcal{L}_\beta = \frac{f_\pi^2}{2}\left[\left(J^i_\mu\right)^2-\beta\left(J^3_\mu\right)^2\right]. \label{lagr Squashed J}
\end{align}
This Lagrangian actually has a clear geometric interpretation as a non-linear sigma model with a target space which is a \emph{squashed sphere} homeomorphic to $S^3$ but with a less symmetric metric. Such a model (with a specific negative value of $\beta$) has been shown by Dombre and Read to arise in 2D as an effective theory of a Heisenberg antiferromagnet on a triangular lattice~\cite{dombre1989nonlinear}, and the renormalization of the continuum model has previously been considered in detail~\cite{azaria1995massive,squashedsigma}. The 3D frustrated magnetic model introduced below in Sec.~\ref{sec 3} was specifically chosen to produce this squashed sphere sigma model action in the continuum limit much as was done by Dombre and Read in the 2D case. As we will show there, terms involving higher order derivatives will naturally arise in the continuum approximation to the lattice model, and these terms will allow for the presence of stable topological defects.
\subsection{Gauge symmetry and topological charges}\label{sec Top charge}
In the $\beta=1$ limit, the model reduces to the $CP^1$ model \eqref{lagr CP1 S and z} which has a $U(1)$ gauge symmetry under transformations $z(x)\rightarrow e^{-i\phi(x)}z(x)$, where $\phi(x)$ is an arbitrary function of the spatial coordinate $x$. In terms of $U$ in \eqref{def U and z}, this gauge symmetry corresponds to right multiplication by the unitary matrix $V_R=\text{diag}\left(e^{i\phi(x)}, e^{-i\phi(x)}\right)$, from which it is easily shown that the $J$ currents transform as,
\begin{align}J^3\rightarrow J^3+\partial \phi,\qquad \left(\begin{array}{c}
J^1 \\
J^2
\end{array}\right)\rightarrow\left(\begin{array}{cc}
\cos 2\phi & -\sin 2\phi\\
\sin 2\phi & \cos 2\phi
\end{array}\right)\left(\begin{array}{c}
J^1 \\
J^2
\end{array}\right).\label{eq J gauge transf}
\end{align}
Clearly the structure in the $CP^1$ action, $\left(J^1\right)^2+\left(J^2\right)^2$, is gauge invariant since there is no explicit $J^3$ dependence and the quadratic form is invariant under rotations of $J^1, J^2$. There are two other obvious gauge invariant structures that may be constructed. The two-form $J^1_\mu J^2_\nu-J^2_\mu J^1_\nu$ is also invariant under rotations of $J^1, J^2$. And given that $J^3$ transforms like a vector potential
\begin{align}
A_\mu\equiv J^3_\mu,\label{def A}
\end{align}
the gauge invariant field strength tensor $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$ may also be constructed. In fact by expressing the $J$ currents in terms of the $z$ field it can be quickly shown that these two quantities are not independent,
\begin{align}
F_{\mu\nu}\equiv \partial_\mu J^3_\nu -\partial_\nu J^3_\mu= 2\left(J^1_\mu J^2_\nu -J^2_\mu J^1_\nu\right).\label{def F tensor}
\end{align}
Just as the gauge invariant $CP^1$ action \eqref{lagr CP1 S and z} may be written entirely in terms of the real unit vector field $S^i$, so may the gauge invariant $F_{\mu\nu}$ tensor.\footnote{This equality can be shown by exploiting the global symmetry to choose $S^i=(0,0,1)$ and expressing $F$ in terms of the complex field $z^\alpha=(z^0,0)$.}
\begin{align}
F_{\mu\nu}=\frac{1}{2}\epsilon_{ijk}S^i\partial_\mu S^j \partial_\nu S^k.\label{eq F tensor S}
\end{align}
This $F$ tensor is directly related to the notion of topological charge for two-dimensional magnetic (baby) Skyrmion field configurations.\footnote{In a high-energy context this same topological charge might be referred to as the instanton charge of the 2D $CP^1$ model.} As a two-form, $F$ may be integrated over an arbitrary two-dimensional surface $\Sigma$, and the result will be the magnetic Skyrmion charge counting the number of times the map $S^i:\Sigma\rightarrow S^2$ wraps around the $S^2$ target space of the $S^i$ field (up to a $2\pi$ difference in normalization).
Furthermore, $F$ may be used to define a $U(1)$ Chern-Simons three-form $A\wedge F$ which represents a distinct notion of topological charge density which is integrated over 3D volume rather than a 2D surface. This is just the Hopf charge $Q$~\cite{whitehead1947expression},
\begin{align}
Q= -\frac{1}{8\pi^2}\int d^3x \,\epsilon^{\lambda\mu\nu}A_\lambda F_{\mu\nu}. \label{def Q Hopf}
\end{align}
Roughly speaking, a field configuration with non-zero Hopf charge may be described as a 2D magnetic Skyrmion (or $CP^1$ instanton) extended as a string in the third spatial direction, and then tied back in a loop. For this loop to be topologically distinct from the $Q=0$ vacuum it must be twisted or knotted in a non-trivial way~\cite{moffatt1995helicity}.
So far we have been considering two forms of topological charge which, due to the gauge invariance of the quantities involved, are able to be expressed in terms of the unit vector $S^i$ field which maps physical space to the $S^2$ target space. However, we began with the squashed sphere sigma model \eqref{lagr Squashed J} and the currents $J$ which are expressible in terms of the $U\in SU(2)$ field. Since $SU(2)$ is homeomorphic to the three sphere $S^3$, there is another seemingly distinct form of topological charge which describes the windings of the $S^3$ base space\footnote{Our base space, i.e. ordinary physical space, is $R^3$ but due to the boundary condition at infinity it may be considered topologically equivalent to $S^3$. This boundary condition must also be applied to the gauge field $A$. } around the $S^3$ target space. This is referred to as the Skyrme charge or baryon charge,
\begin{align}
Q= -\frac{1}{2\pi^2}\int d^3x \,\epsilon^{\lambda\mu\nu}J^1_\lambda J^2_\mu J^3_\nu. \label{def Q Skyrme}
\end{align}
But as can be seen from a direct substitution of the definitions of $A$ and $F$ in \eqref{def A},\eqref{def F tensor}, this is actually identical to the Hopf charge! This is the main point that we wish to stress, \emph{a Hopfion may be considered to be a three-dimensional Skyrmion, and vice-versa}. The $z$ field description of a Skyrmion may be directly mapped to the $S$ field using \eqref{def S and z}, and the result will have Hopf charge equal to its original baryon charge. On the other hand, a Hopfion involves a map to a $S^2$ target space which may be lifted to a $S^3$ target space by identifying the $A$ field with the $J^3$ field and integrating. Although this construction is not unique since $A$ is only defined up to a gauge transformation, for any choice of $A$ the lifted map will have baryon charge equal to its original Hopf charge.
This notion of the equivalence between the baryon charge and Hopf charge is not a new idea, it is clearly discussed in~\cite{radu2008stationary} and~\cite{han2017skyrmions} for instance. The idea also underlies the model of Ward which will be discussed further in the next subsection. Note however that this is distinct from a completely different notion of Hopfions in the Skyrme model~\cite{meissner1985toroidal,cho2001monopoles,cho2008new}, where a field configuration $U$ is restricted to only take values in a subspace $S^2\subset SU(2)$. In that case since the $U$ field does not cover $SU(2)$ the baryon charge vanishes, but a different notion of Hopf charge\footnote{The $F$ tensor for this second notion of Hopf charge is defined in terms of the unit vector $n$ which is considered in Sec.~\ref{sec Rational map} rather than $S$.} may still be defined in terms of the $S^2$ subset.
Finally, let us briefly comment on a third way in which the charge $Q$ may be understood which is more familiar from Yang-Mills theory. The current $J_\mu$ in \eqref{def J currents} may also be understood as a non-Abelian gauge field associated to gauge symmetry under right multiplication by $SU(2)$ matrices. It is pure gauge and the non-Abelian field strength tensor vanishes. If we consider the Chern-Simons three-form associated to this \emph{non-Abelian} gauge symmetry (as opposed to the Abelian gauge symmetry involved in the definition of the Hopf charge), we have the charge
$$\mathcal{K}=-\frac{1}{16\pi^2}\int d^3 x \, \epsilon^{\lambda\mu\nu}\left(J^k_\lambda \partial_\mu J^k_\nu + \frac{1}{3}\epsilon_{ijk}J_\lambda^iJ_\mu^jJ_\nu^k\right).$$
Rewriting $\partial_{[\mu}J^k_{\nu]}$ as a product of two gauge fields in a manner similar to \eqref{def F tensor}\footnote{This relation may also be understood as arising from the vanishing of non-Abelian field strength tensor.} we see that this just reduces to the expression for the baryon charge \eqref{def Q Skyrme}, so $\mathcal{K}=Q$. This non-Abelian Chern-Simons charge $\mathcal{K}$ is interesting in 4D Yang-Mills theory because 4D instantons can be understood as interpolating between 3D vacua with different values of $\mathcal{K}$~\cite{jackiw1976vacuum,CALLAN1976334,shifman1994instantons}.
This perspective on the charge is well-illustrated by the construction of Atiyah and Manton, where Skyrmion configurations with non-zero $Q$ are generated from an initial trivial configuration by integrating over $SU(2)$ instantons~\cite{Atiyah1989,Atiyah1993}. This gives a reasonably good approximation to the minimal energy configuration, and further work by Sutcliffe explained the success of the Atiyah-Manton approximation. In~\cite{Sutcliffe2010}, a novel BPS model is derived from the pure Yang-Mills theory in one higher dimension, obtaining a Skyrme field coupled to an infinite tower of vector mesons. Interestingly, when all vector mesons are considered, the BPS property is fulfilled with a Skyrme field given exactly by the holonomy of the instanton. Nevertheless, the restriction to the lowest vector meson already improves the Skyrmion description of nuclei, with low binding energies and nuclear cluster structures arising~\cite{Naya2018}.
\subsection{The squashed Skyrme model and energy bounds}\label{sec Ward model}
So far we have discussed the terms in the squashed sphere sigma model which are quadratic in derivatives, but due to Derrick's theorem~\cite{derrickstheorem} higher order terms are necessary to stabilize the topological defects with non-zero $Q$ which were discussed above. The Lagrangian of the full Skyrme model~\cite{skyrme} is
\begin{align}
\mathcal{L}_{Skyrme}=\frac{f_\pi^2}{2}\left[\left(J^k_\lambda\right)^2 + \frac{1}{2M^2}\epsilon_{ijk}\epsilon_{klm}J_\mu^iJ_\nu^jJ_\mu^lJ_\nu^m\right],
\end{align}
where $M$ is some new dimensionful parameter often written as $ef_\pi$. Using \eqref{def Q Skyrme}, the energy may be written as,
\begin{gather*}
E_{Skyrme}=\int d^3x\,\mathcal{L}_{Skyrme}=E_{BPS}|Q|+ \frac{f_\pi^2}{2}\int d^3x\,\left(J^k_\lambda \pm \frac{1}{2M}\epsilon_{ijk}\epsilon^{\lambda\mu\nu}J_\mu^iJ_\nu^j\right)^2,\nonumber\\\qquad E_{BPS}\equiv 6\pi^2 \frac{f_\pi^2}{M}.
\end{gather*}
This form of the Skyrme energy functional clearly shows the BPS bound $E\geq |Q|E_{BPS}$. This expression for the energy functional may easily be generalized to the squashed sphere case,
\begin{align*}
E_{BPS}|Q|+ \frac{f_\pi^2}{2}\int d^3x\left[\left(J^a_\lambda \pm \frac{1}{2M}\epsilon_{ija}\epsilon^{\lambda\mu\nu}J_\mu^iJ_\nu^j\right)^2+(1-\beta)\left(J^3_\lambda \pm \frac{1}{2M(1-\beta)}\epsilon_{ij3}\epsilon^{\lambda\mu\nu}J_\mu^iJ_\nu^j\right)^2\right],
\end{align*}
where $a$ is only summed over $1,2$. Note that the $\epsilon_{ij3}J^iJ^j$ expression in the $\beta$ dependent term is proportional to the $F$ tensor \eqref{def F tensor} defined above. Expanding the squares leads to the Lagrangian,
\begin{align}
\mathcal{L}=\frac{f_\pi^2}{2}\left[\left(J^k_\lambda\right)^2 -\beta\left(J^3_\lambda\right)^2+ \frac{1}{2M^2}\epsilon_{ijk}\epsilon_{klm}J_\mu^iJ_\nu^jJ_\mu^lJ_\nu^m+\frac{\beta}{8M^2(1-\beta)}\left(F_{\mu\nu}\right)^2\right],\label{lagr Ward}
\end{align}
which also satisfies the BPS bound
\begin{align}
E\geq 12\pi^2\frac{f^2_\pi}{2M}|Q|\equiv E_{BPS}|Q|. \label{eq BPS bound}
\end{align}
The new term quartic in derivatives is exactly that of the Faddeev-Niemi model, so squashing the target space of the Skyrme model while maintaining the BPS bound naturally leads to an interpolation between the Skyrme and Faddeev-Niemi models. This generalized Skyrme system was considered earlier by Nasir and Niemi~\cite{nasir2002effective} and Ward and Silva Lobo~\cite{ward2004skyrmions,lobo2011generalized,silva2011lattices}.
It may seem that there is a difficulty in extending to the limit $\beta=1$ due to the prefactor $(1-\beta)^{-1}$ of the Faddeev term. If $f_\pi$ and $M$ are taken fixed as $\beta$ is varied this is indeed the case. This parametrization will be referred to as the \emph{fixed bound parametrization} since the energy satisfies the BPS inequality with an energy $E_{BPS}$ that is constant with $\beta$.
But if $M^2$ is allowed to vary with $\beta$, then there is no problem taking the $\beta=1$ limit. In particular, the \emph{Ward parametrization}~\cite{ward2004skyrmions},
\begin{align*}
\frac{f_\pi^2}{2}=\frac{1}{4\pi^2\left(3-\beta\right)},\qquad \frac{f_\pi^2}{2M^2}=\frac{1-\beta}{4\pi^2\left(3-2\beta\right)},
\end{align*}
is based on requiring that the identity map from a base space with spherical $S^3$ geometry to the $S^3$ target space has unit energy for all $\beta$, and it leads to a fairly constant dependence on $\beta$ of the energy of a $Q=1$ Skyrmion in flat space as well. No matter which parametrization for $f_\pi$ and $M$ is chosen, the results for any other parametrization may be recovered by adjusting the energy and length scales. Table \ref{table SqSkyrme} involves a simulation in the fixed bound parametrization, but the rescaled results agree with Ward up to an error of $\sim0.1\%$ from finite size effects.
\begin{table}
\begin{center}
\begin{tabular}{c|c|c|c|c}
$\beta$ & $E/E_{BPS}$ & $E_{h}/E_{BPS}$ & $E_{W}$ & $\frac{\langle (J_\mu^1)^2 \rangle}{\langle (J_\mu^3)^2 \rangle}$ \\
\hline \hline
0.0 &1.2323 &1.2331 & 1.2323& 1.0 \\
0.1 &1.2339 &1.2348& 1.2324& 0.9873 \\
0.2 &1.2392 &1.2403 &1.2324 & 0.9737 \\
0.3 &1.2497 &1.2513 & 1.2322& 0.9592 \\
0.4 &1.2679 &1.2702 & 1.2319& 0.9439 \\
0.5 &1.2981 &1.3015 & 1.2315& 0.9279 \\
0.6 &1.3486 &1.3535 & 1.2311& 0.9103 \\
0.7 &1.4370 &1.4442 &1.2309 & 0.8912 \\
0.8 &1.6111 &1.6224 & 1.2316& 0.8695 \\
0.9 &2.0530 &2.0650 & 1.2269& 0.8519\\
\end{tabular}
\caption{A simulation of a $Q=1$ soliton in the squashed Skyrme model. $E$ is the energy in the fixed bound parametrization. The simulation was carried out on a cubic lattice with $100^3$ sites (except for $\beta=0.9$ where the length was doubled to $200^3$) and lattice spacing $a=0.2$ in units where $M=1$. An arrested Newton flow method was used for the minimization as described in~\cite{battye2002skyrmions}, with the time evolution implemented by a fourth-order Runge-Kutta method with time step $\Delta t = 0.1$. $E_h$ is the optimal energy in the spherically symmetric hedgehog ansatz for this same parametrization. $E_W$ is the energy in the Ward parametrization which was found by rescaling $E$. To better indicate the departure from the hedgehog ansatz, the values of $\left(J^1\right)^2$ and $\left(J^3\right)^2$ are averaged over the domain of the simulation and compared.}
\label{table SqSkyrme}
\end{center}
\end{table}
Any parametrization which allows for a well-defined $\beta=1$ limit will involve the energy $E_{BPS}$ in the BPS bound \eqref{eq BPS bound} tending to zero. This makes sense since in the Faddeev-Niemi model the minimal energy solutions obey a weaker $E\geq K Q^{3/4}$ inequality for some value of $K$~\cite{vakulenko1979stability,ward1999hopf} and moreover the minimal energy Hopfions found numerically~\cite{battye1998knots,battye1999solitons,sutcliffe2007knots} appear to come close to saturating this bound. For $\beta$ close to but less than $1$, the energies of solitons with small values of $Q$ may be very close to the energies in the Faddeev-Niemi model, and this is not disallowed by \eqref{eq BPS bound} since the value of $E_{BPS}$ may be very small. But no matter how small $E_{BPS}$ may be, eventually for large enough $Q$, $E_{BPS}Q> K Q^{3/4}$. So for $\beta<1$ the energies of the large $Q$ solitons can not scale asymptotically as $Q^{3/4}$, and thus if the Faddeev-Niemi model indeed has this asymptotic behavior there must be a dramatic difference for large $Q$ solitons if $\beta$ is even slightly below $1$.
\subsection{Position curves and baryon strings}\label{sec Position curves and strings}
Intuitively a Hopfion is often described as a loop of string whereas a single Skyrmion in the Skyrme model is spherically symmetric and multiple Skyrmions form polyhedral clusters. While we have shown that the baryon charge and Hopf charge are identical, let us comment a bit more on how these two pictures are resolved.
The $Q=1$ Skyrmion in the Skyrme model satisfies the hedgehog ansatz,
\begin{align}
U(x^\mu) = \cos f(r)\,I+ i \sin f(r)\frac{x^i}{r}\sigma^i,\label{def Hedgehog}
\end{align}
for some radial profile function $f(r)$ which equals $\pi$ at $r=0$ and vanishes at infinity. Considering \eqref{def U and z} and \eqref{def S and z}, the third component of the unit vector $S$ field in the hedgehog ansatz is,
$$S^3=\cos^2 f -\frac{x^2+y^2-z^2}{r^2}\sin^2 f.$$
The boundary condition on the $S$ field at infinity $S^3=+1$ is also satisfied along the $z$ axis, and the furthest departure from the boundary condition $S^3=-1$ is satisfied in a loop in the $xy$-plane with radius $r_0$ such that $f(r_0)=\pi/2$. A curve such as this where $S^3=-1$ is referred to as the \emph{position curve}, and it may be thought of as the core of the Hopfion.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{BlackHopfions.pdf}
\caption{On the left is a cross-section in the $yz$-plane of the hedgehog Skyrmion in the Skyrme model ($\beta=0$), mapped to the unit vector $S$ field, as discussed in the text. The
{color scheme follows the typical conventions in Lorentz transmission electron microscopy, where the}
hue of a color denotes the azimuthal angle of $S$, and the brightness denotes the polar angle. The limiting case of black represents the boundary condition at infinity, and white denotes the center of the position curve describing the core of a Hopfion. On the right is the minimal energy $Q=1$ soliton in the squashed Skyrme model at $\beta=0.9$ in the fixed bound parametrization. Both plots may be compared to similar plots for true Hopfions with gauge invariance such as Fig. 1b of~\cite{sutcliffe2017skyrmion}, and Fig. 1b and Fig. 3 in~\cite{kent2021creation}.}
\label{fig Hedgehog cross section}
\end{figure}
A cross-section in the $yz$-plane of the Skyrmion at both $\beta=0$ and $\beta=0.9$ is plotted in Fig. \ref{fig Hedgehog cross section}. The two intersections of the position curve loop with the plane are clearly seen, and it may be seen from the colors representing the orientation of $S$ how the 2D magnetic Skyrmion charge in the $yz$-plane (quantified by $F_{23}$ \eqref{eq F tensor S}) is concentrated around the position curve. Note that due to the dependence on the $J^3$ field the energy density and baryon charge of the hedgehog Skyrmion are actually spherically symmetric and not concentrated near the position curve. But even the $Q=1$ Hopfion in the Faddeev-Niemi model at $\beta=1$ is approximately spherically symmetric in this sense as well, as was noted by Ward~\cite{ward2004skyrmions} and is seen by the extent to which the hedgehog ansatz fits the data in Table \ref{table SqSkyrme}.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{Charge7-1.pdf}
\caption{In (a) a position curve of a $Q=7$ Skyrmion at $\beta=0$ is shown. For small $\beta>0$ this settles to a three loop configuration as in (b). This solution may be tracked for increasing $\beta$, but it develops an instability at $\beta=0.77$ and settles to a distinct buckled loop configuration (c). This buckled loop branch may be continued towards decreasing $\beta$ (indicated by an asterisk) in (d) and (e), and eventually it becomes degenerate in energy with the solution (a). The energy per charge in the Ward parametrization is plotted in (f), with red indicating the three loop branch shown in (b) and blue indicating the buckled loop branch of (c,d,e).}\label{fig Charge 7}
\end{figure}
The position curve may also be considered for higher charge solitons in the Skyrme model, such as the $Q=7$ case in Fig. \ref{fig Charge 7}. At $\beta=0$ the position curve self-intersects much like the $\chi$ solutions found by Sutcliffe in the Faddeev-Niemi model~\cite{sutcliffe2007knots}. {At $\beta=0.2$ the position curve transforms to three loops and we were originally anticipating that this solution would smoothly deform to a trefoil knot as was seen at $\beta=1$ in the Faddeev-Niemi model~\cite{battye1998knots,battye1999solitons}. However it appears that there are actually many locally stable solution branches which may exchange roles as the global minimum as the deforming parameter is adjusted.
In particular, we found that, at $\beta = 0.77$, a three loop configuration becomes unstable and evolves to a buckled loop with lower energy. We explored this new configuration by decreasing $\beta$ all the way back to $\beta=0$, in which case the buckled loop transforms to a self-intersecting solution which is degenerate with our original solution in Fig. \ref{fig Charge 7}(a).
This wealth of local minima will be seen again in the results of the lattice model in \ref{sec charge-10}, and has also been seen in systems such as the Skyrme model with a non-zero pion mass \cite{bjarke2022smorgaasbord}.}
Note that at $\beta=0$ the global symmetry of system under right multiplication is increased from $U(1)$ to $SU(2)$. Any soliton configuration $U$ may be transformed to a new field $U \rightarrow V^{-1}U V$ with the same boundary conditions at infinity and the same energy. For a general $V\in SU(2)$ this transformation will not leave the position curve invariant. For the hedgehog configuration this ends up being equivalent to the degeneracy of the solution under spatial rotations, but for higher charge configurations the shape of the position curve itself may change. However for $\beta>0$ the symmetry is reduced to a {$U(1)\rtimes Z_2$} subgroup\footnote{ {The extra discrete $Z_2$ global symmetry arises from those internal $SU(2)_R$ transformations which flip the third axis in isospin space. Concretely the $Z_2$ subgroup may be chosen as $\{ I,\sigma^1\}\subset SU(2)_R$.}} which leaves the locus of the position curve unchanged, and only translates the field along the position curve.
Some insight may be gained by considering the structure of the $U$ field around the position curve in the hedgehog ansatz in \eqref{def Hedgehog} and Fig. \ref{fig Hedgehog cross section}, and abstracting this to a new ansatz of a cyllindrically symmetric straight string with a position curve aligned with the $z$-axis,
\begin{align}
U(\rho,z,\phi)= \cos g(\rho) e^{-i\phi\sigma^3}+i\sin g(\rho)\left(\cos\left(\frac{2\pi z}{L}\right)\sigma^1+\sin\left(\frac{2\pi z}{L}\right)\sigma^2\right).
\end{align}
Here $g(\rho)$ is some new profile function depending on $\rho=\sqrt{x^2+y^2}$, and $L$ is some parameter describing the rate of twisting along the string. The profile function vanishes at infinity and $g(0)=\pi/2$. The baryon charge \eqref{def Q Skyrme} integrated over a length $\Delta z$ is found to be,
\begin{align*}
Q= -\frac{1}{L}\int dz d\rho \,g^\prime(\rho)\sin \left(2g(\rho)\right) =\frac{\Delta z}{L},
\end{align*}
so every segment of length $L$ has baryon charge $1$.
Outside the core of the string, where $g\approx 0$, $U$ is restricted to a $U(1)$ subgroup, and the principal chiral model effectively reduces to a 3D XY model. Unless the $U(1)$ subgroup is gauged (as it is at $\beta=1$) the energy per length of an isolated straight string will be logarithmically divergent.
To better understand the energy per length due to the core of the string, the structures involved in the energy density \eqref{lagr Ward} may be expressed in terms of the ansatz,
\begin{gather}
(J^1)^2+(J^2)^2=g^{\prime2}+\left[\left(\frac{2\pi}{L}\right)^2+\frac{1}{\rho^2}\right]\cos^2g \sin^2 g \nonumber\\
(J^3)^2=\left(\frac{2\pi}{L}\right)^2 \sin^4g + \frac{1}{\rho^2}\cos^4g \nonumber\\
\frac{1}{8}\left(F_{\mu\nu}\right)^2=g^{\prime 2}\left[\left(\frac{2\pi}{L}\right)^2+\frac{1}{\rho^2}\right]\cos^2g \sin^2 g\nonumber\\
-\frac{1}{4}\text{Tr}[J_\mu,J_\nu]^2=\left[\sin^2g \left(\frac{2\pi}{L}\right)^2 +\cos^2g \frac{1}{\rho^2}\right]g^{\prime 2}+\left(\frac{2\pi}{L}\right)^2\frac{1}{\rho^2}\cos^2g \sin^2g.\label{eq Energy string ansatz}
\end{gather}
The $\rho^{-2}$ term in the $\left(J^3\right)^2$ structure is what causes the infrared divergence of an isolated string. Of course this is typical for strings or vortices and does not preclude either a network of oppositely oriented long strings or strings forming closed loops with radius of curvature much larger than the string thickness, in which cases this ansatz may still be useful. In the latter case, due to the reduction of the model to a global $U(1)$ theory outside the string core, and due to the expression \eqref{def F tensor} relating the curl of $J^3$ to the dual of $F$ tensor which has a constant $2\pi$ flux across the string core, the $J^3$ field may be calculated outside the string core using the Biot-Savart formula, much as is done calculating the fluid velocity outside knotted loops of vorticity~\cite{kleckner2013creation}.
The energy inside the core may be calculated by optimizing the above expressions for the energy \eqref{lagr Ward} and \eqref{eq Energy string ansatz} over profile functions $g$ and length per baryon charge $L$. An important special case is that of Faddeev-Niemi model itself at $\beta=1$, in which case the energy contribution from the $J^3$ field outside the core vanishes. To compare with previous results, we may use the parametrization in Battye and Sutcliffe~\cite{battye1999solitons} where $f^2_\pi/2=4$ and $M\sqrt{1-\beta}\rightarrow 1/2$. Then the optimum energy of the straight string ansatz is found to be $E \approx 396$ per length $L\approx 3.95$. The energy per charge already agrees reasonably well with the unstable toroidal solutions of Battye and Sutcliffe in Table 1 and Figure 9 of~\cite{battye1999solitons}, which may be expected to become closer to the straight Skyrmion string ansatz as the charge increases.
\section{A toy model of a frustrated magnet}\label{sec 3}
Now we will introduce a simple spin system which at lowest order in the continuum approximation reduces to the same squashed sphere non-linear sigma model discussed in the previous section. The higher order derivative terms which may stabilize topological defects will be different from the rotationally symmetric Skyrme and Faddeev-Niemi terms, but are in many respects qualitatively similar.
\subsection{A description of the lattice model} \label{sec Lattice model intro}
The system is defined on an ordinary cubic lattice with lattice spacing $a$, and each site $x$ has three real unit vector spins $S^i_r(x)$, where the $i$ index refers to the three components of the unit vector, and the $r$ index labels the distinct spins at the site. The dot product between any two spins at a given site is constrained to be equal to a parameter $\kappa$ which is fixed for the entire system, i.e. $S_r(x)\cdot S_s(x)=\kappa$ for $r\neq s$ and all $x$. So the three spins at each site act like a rigid body with an orientation which may be described by a matrix $R(x)\in SO(3)$.
The spins may be written in terms of this $R(x)$ and a basis $e_r$ which does not depend on $x$,
\begin{gather}
S^i_r(x)=R(x)^i_{\,j}e^{j}_r,\qquad
e^j_{r}\equiv
\left(\begin{array}{ccc}
\frac{\sqrt{3}}{2}\sin\theta& 0 &-\frac{\sqrt{3}}{2}\sin\theta\\
-\frac{1}{2}\sin\theta &\sin\theta&-\frac{1}{2}\sin\theta\\
\cos\theta & \cos\theta & \cos\theta
\end{array}\right),\label{def R e}
\end{gather}
where the three vectors $e_r$ are represented as a matrix with $r$ referring to different columns. The fixed parameter $\theta$ in this basis is directly related to the parameter $\kappa$,
\begin{align}
S_r(x)\cdot S_s(x)=\kappa \equiv \frac{3}{2}\cos^2\theta-\frac{1}{2}.\label{def Kappa theta}
\end{align}
The spins interact as a frustrated classical Heisenberg model, with a ferromagnetic coupling $K_1<0$ between nearest neighbors at a distance of one lattice spacing $a$, and an antiferromagnetic coupling $K_2>0$ between sites at a distance of $2a$, which are indicated by doubled angled brackets in a slight abuse of notation,\footnote{For simplicity in this toy model, sites at the nearer distances of $\sqrt{2}a$ and $\sqrt{3}a$ are not taken to interact.}
\begin{align}
H&=K_1\sum_{r,\,\langle x,y\rangle}S_r(x) \cdot S_r(y)+K_2\sum_{r,\,\langle\!\langle x,y\rangle\!\rangle}S_r(x) \cdot S_r(y).\label{ham S}
\end{align}
Note that a given spin $S_r(x)$ only interacts with spins $S_r(y)$ with the same `species' $r$.
This model {was originally inspired} by the treatment of spins interacting on a pyrochlore lattice in~\cite{BatistaEtAl2018}, where in that case $r$ takes four values corresponding to the four sites of the tetrahedral cells of the pyrochlore lattice. The dot product between spins $\kappa$ in that case is fixed so that the spins are in an `all-in-all-out' configuration which is preferred {in the presence of biquadratic spin interactions}. If we formally allow $\kappa$ to be a tunable parameter and restrict the interaction to third-nearest-neighbor sites so only spins with the same value of $r$ interact we obtain a very similar model to that considered here.
{The motivation for making these abstractions was to create a simple lattice model that still captures the main qualitative features of a broad class of realistic 3D noncollinear magnets which involve $SO(3)$ Goldstone modes. The `squashing' parameter $\kappa$ that measures the degree of collinearity} is expected to vary for different 3D lattices and different magnetic anisotropies. So the model studied in this paper is expected to provide a unified qualitative description for a series of realistic materials that are described by different values of $\kappa$.
Also note that at the limiting value $\kappa=1$ where the spins $S_r$ {are perfectly collinear}, the lattice model reduces to a 3D version of inversion-symmetric frustrated magnets
which have been previously considered in 2D as a host to magnetic Skyrmions~\cite{leonov2015multiply,lin2016ginzburg}. A 3D extension of these frustrated magnets has already been considered {in the collinear case} \cite{sutcliffe2017skyrmion}, and Hopfions were investigated and the analogy to the Faddeev-Niemi model was pointed out.
But in the opposite limit of $\kappa=0$ this model will instead be shown to be closely analogous to the Skyrme model, so this toy model bears the same relationship to the effective theory of frustrated magnets in~\cite{sutcliffe2017skyrmion} as the squashed Skyrme model~\cite{nasir2002effective,ward2004skyrmions} discussed in the previous section bears to the Faddeev-Niemi model.
\subsection{Effective theory in the continuum limit}\label{sec Lattice model effective theory}
To show that this analogy is valid, let us now turn to the effective continuum description of the model. Following a similar procedure to Dombre and Read's continuum description of the triangular antiferromagnet~\cite{dombre1989nonlinear}, the Hamiltonian can be described up to fourth order in derivatives in terms of continuous fields $S^i_r(x)$,
\begin{align}
H=-\frac{1}{2a}\left(K_1+4K_2\right)\sum_\mu \int d^3 x \left(\partial_\mu S_r\right)^2 +\frac{a}{24}\left(K_1+16 K_2\right)\sum_\mu \int d^3 x \left(\partial^2_\mu S_r\right)^2. \label{ham Continuum}
\end{align}
Now since rotational symmetry is broken by the fourth order terms, any sums over the spatial index $\mu$ will always be indicated explicitly, although sums over internal indices like $r$ or $i$ are still implied by the summation convention or context. The lack of rotational symmetry in the fourth order terms is the main difference between the effective description of this toy model and that considered by Lin and Hayami~\cite{lin2016ginzburg}. Here the interaction between the neighbors at distances $\sqrt{2}a$ and $\sqrt{3}a$ was set to zero whereas in~\cite{lin2016ginzburg,sutcliffe2017skyrmion} it was implicitly tuned to maintain rotational symmetry in the fourth order terms. Note that in the absence of any tuning such cubic anisotropies would generically be expected to be present, {and the presence of isotropy in these higher derivative terms is not essential for the stabilization of topological defects.}
For this Hamiltonian to have stable topological defects it is easily shown by an argument along the lines of Derrick's theorem~\cite{derrickstheorem} that the coefficients of both the second and fourth order terms must be positive,
$$-K_1 > 4 K_2 > -\frac{1}{4}K_1.$$
Moreover, for the Skyrmion size to be much larger than the lattice spacing and this continuum description to be valid we must be close to the Lifshitz transition $K_2= -\frac{1}{4}K_1$ where the sign of the quadratic term changes from positive to negative. Suppose that a Skyrmion field configuration has some length scale $L$ representing the radius, and the parameters are displaced from the Lifshitz transition by some small positive quantity $\epsilon$,
\begin{align}
K_2= -\left(\frac{1}{4}-\epsilon\right)K_1.\label{def epsilon}
\end{align}
Then it can be shown that radius of the Skyrmion is on the order
$L\sim \epsilon^{-1/2}a,$ where the exact coefficient depends on dimensionless integrals over the field configuration.
Now to proceed and better illustrate the connection to the squashed sphere sigma model in Sec.~\ref{sec Squashed sphere}, the spins $S_r$ may be written in terms of the rotation matrix field $R(x)$ using \eqref{def R e}. The quadratic terms become
$$\sum_\mu \int d^3 x \left(\partial_\mu S_r\right)^2 = \sum_\mu \int d^3 x \text{Tr}\left[\partial_\mu R^{-1}\partial_\mu R \,e_r\otimes e_r\right],$$
where
\begin{align}
e_r\otimes e_r = \text{diag}\left(1-\kappa,\,1-\kappa,\, 1+2\kappa\right).
\end{align}
For $\kappa=0$ this is clearly equivalent to the principal chiral model \eqref{lagr PCM U}, except that it is expressed in terms of $R\in SO(3)$ rather than $U\in SU(2)$. For $\kappa \neq 0$, the components of the diagonal matrix $e_r\otimes e_r$ will take different values and this will become a squashed sphere model. This can be seen by expressing the model in terms of the currents $J$ \eqref{def J currents}, which may also be expressed in terms of the $SO(3)$ matrix,
\begin{align}
\left(R^{-1}\partial_\mu R\right)_{ij} = 2\epsilon_{ijk}J^k_\mu.\label{eq J R}
\end{align}
Using this identity, the quadratic terms become
\begin{align}
\sum_\mu \int d^3 x \left(\partial_\mu S_r\right)^2 = 4(\kappa+2)\sum_\mu \int d^3 x \left[\left(J^i_\mu\right)^2-\frac{\kappa}{\frac{1}{3}\left(\kappa+2\right)}\left(J^3_\mu\right)^2\right].\label{lagr Quadr}
\end{align}
This is precisely the squashed sphere model in \eqref{lagr Squashed J}, with the parameter $\beta$ expressed in terms of $\kappa$. The overall dimensionfull parameter $f_\pi$ in the squashed sphere model depends on the prefactor of the quadratic terms given in the full Hamiltonian \eqref{ham Continuum}, and it is seen to be on the order $f_\pi \sim \left(\epsilon|K_1|a^{-1} \right)^{1/2}$.
Exactly the same chain of steps may now be followed to express the quartic terms of the Hamiltonian in terms of the $J$ fields and the parameter $\kappa$. After some calculation,
\begin{align}
\sum_\mu \left(\partial^2_\mu S_r\right)^2 = &8(1-\kappa)\sum_\mu \left[\left(\partial_\mu J_\mu^i\right)^2+4\left(J_\mu^i J_\mu^i\right)^2\right]+ 12\kappa \sum_\mu \left[\left(\mathcal{D}_\mu J_\mu^a\right)^2+4\left(J_\mu^a J_\mu^a\right)^2\right],\label{lagr Quartic}
\end{align}
where as discussed previously, $i$ runs over all components $1, 2, 3$, and $a$ is only taken over $1, 2$. The covariant derivative with respect to the gauge symmetry defined in \eqref{eq J gauge transf} is
$$\mathcal{D} J^a\equiv \partial J^a + 2\epsilon_{ab3}J^3\,J^b.$$
Note that the continuum model is completely gauge symmetric at $\kappa=1$, which must be the case considering that in the lattice model all three spins at each site are pointing in the same direction, so the rotation field $R(x)$ is only fixed up to rotations about the spin axis.
This continuum description of the model in equations \eqref{ham Continuum}\eqref{lagr Quadr}\eqref{lagr Quartic} will later be applied to calculate the energies of highly symmetric ansatzes for Skyrmion configurations, and the results will be compared with Skyrmions found in a numerical simulation of the lattice model described below.
\subsection{Numerical simulation and unit charge Skyrmions}\label{sec Lattice model numerical results}
In this section, we perform numerical simulations directly on the lattice model~\eqref{ham S}, which captures the higher order terms and spatial anisotropies that we neglected in the continuum model~\eqref{ham Continuum}. Strictly speaking, the energy barriers between different topological sectors are no longer infinity on a discrete lattice. Consequently, when the Skyrmion size is not significantly larger than the lattice spacing $a$, it could be unstable towards tunneling into the vacuum state. In such situations, it is beneficial to fully relax the assumed Skyrmion configuration and check the stability.
Typically, the local minima of the classical {spin} models can be found by the Landau-Lifshitz-Gilbert dynamics~\cite{LandauLifshitz1992,Gilbert2004} that works directly on the spin degrees of freedom. To enforce the constraint $S_r(x)\cdot S_s(x)=\kappa$, a penalty term can be included in the model, which slightly complicates the computation.
To avoid such complication, we work directly with the rotation matrix $R(x)$. The spin-spin interaction between site $x$ and $y$ can be written as
\begin{equation}
\sum_r S_r(x) \cdot S_r(y) = \text{Tr} \left[
R^{-1}(x)R(y) e_r\otimes e_r
\right]. \label{eq:spin-spin_rotation}
\end{equation}
There are a few representations that can be used for the rotation matrix. To avoid ``Gimbal lock'', we use the quaternion representation in this work:
\begin{equation}
R=\begin{pmatrix}q_{0}^{2}+q_{1}^{2}-q_{2}^{2}-q_{3}^{2} & 2q_{1}q_{2}-2q_{0}q_{3} & 2q_{1}q_{3}+2q_{0}q_{2}\\
2q_{1}q_{2}+2q_{0}q_{3} & q_{0}^{2}-q_{1}^{2}+q_{2}^{2}-q_{3}^{2} & 2q_{2}q_{3}-2q_{0}q_{1}\\
2q_{1}q_{3}-2q_{0}q_{2} & 2q_{2}q_{3}+2q_{0}q_{1} & q_{0}^{2}-q_{1}^{2}-q_{2}^{2}+q_{3}^{2}
\end{pmatrix},
\end{equation}
where the quaternion $\bm{q}=(q_0,q_1,q_2,q_3)^T$ is a four-vector satisfying
\begin{equation}
q_0^2+q_1^2+q_2^2+q_3^2=1.
\end{equation} Note that the quaternion $\bm{q}$ is related to the complex numbers $z^0$ and $z^1$ introduced in Eq.~\eqref{def U and z} by
\begin{equation}
z^0 = -q_2 - i q_1,\quad z^1=q_0 + i q_3.
\end{equation}
The Hamiltonian~\eqref{ham S} is now expressed in terms of the quaternions:
\begin{equation}
\begin{split}
H &= 3 K_1 \sum_{\langle x,y \rangle} \left\{ q_0^2(x,y) - \cos^2 \theta \left[ q_1^2(x,y) + q_2^2(x,y) \right] + \cos (2\theta) q_3^2(x,y) \right\} \\
&\quad + 3 K_2 \sum_{\langle\!\langle x,y \rangle\!\rangle} \left\{ q_0^2(x,y) - \cos^2 \theta \left[ q_1^2(x,y) + q_2^2(x,y) \right] + \cos (2\theta) q_3^2(x,y) \right\},
\end{split}\label{ham quaternion}
\end{equation}
where the quaternion product is defined as
\begin{equation}
\bm{q}(x,y)\equiv \overline{\bm{q}(x)} \bm{q}(y)
=
\begin{pmatrix}
\quad q_{0}(x)q_{0}(y)+q_{1}(x)q_{1}(y)+q_{2}(x)q_{2}(y)+q_{3}(x)q_{3}(y)\\
-q_{1}(x)q_{0}(y)+q_{0}(x)q_{1}(y)-q_{2}(x)q_{3}(y)+q_{3}(x)q_{2}(y)\\
-q_{2}(x)q_{0}(y)+q_{0}(x)q_{2}(y)-q_{3}(x)q_{1}(y)+q_{1}(x)q_{3}(y)\\
-q_{3}(x)q_{0}(y)+q_{0}(x)q_{3}(y)-q_{1}(x)q_{2}(y)+q_{2}(x)q_{1}(y)
\end{pmatrix},\label{quaternion prod}
\end{equation}
{and the quaternion conjugate is $\overline{\bm{q}}\equiv (q_0,-q_1,-q_2,-q_3)^T$}.
The Skyrmion solutions are local minima of the lattice model. Consequently, they can be obtained by local minimization algorithms from initial spin configurations not too far away from the minima. In this work, we use the low-storage BFGS method~\cite{nlopt, Nocedal1980, Liu1989} for the minimization. When the spin configurations get close enough to the minima, we switch to the overdamped Langevin dynamics to avoid being trapped in saddle points:
\begin{equation}
\frac{d \bm{q}(x)}{dt} = \bm{f}(x) - \left[ \bm{f}(x) \cdot \bm{q}(x)\right] \bm{q}(x), \label{langevin}
\end{equation}
where the force $\bm{f}(x)$ is defined as
\begin{equation}
\bm{f}(x)= -\frac{dH}{d \bm{q}(x)}.
\end{equation}
The overdamped Langevin dynamics~\eqref{langevin} is integrated by the explicit fourth order Adams-Bashforth method as predictor and the implicit fourth order Adams-Moulton method as corrector. The time step $dt$ is chosen as $dt=0.01/|K_1|$ for $\kappa=0$ and $0.625$, and $dt=0.005 /|K_1|$ for $\kappa \approx 0.955\, (\theta=\pi/18)$.
Before discussing the numerical solutions of Skyrmions, we discuss the energy of the vacuum here. For the ferromagnetic state $\bm{q}(x)=\bm{q}(y)$, the quaternion product~\eqref{quaternion prod} is $\bm{q}(x,y)=(1,0,0,0)^T$. Consequently, the energy is
\begin{equation}
E_{\text{FM}}=
\begin{cases}
9N\left(K_{1}+K_{2}\right), & \text{periodic boundary condition (PBC)}\\
9N\left(K_{1}+K_{2}\right)+9L^{2}(K_{1}+2K_{2}), & \text{fixed boundary condition (FBC)}
\end{cases}
\end{equation}
where $N=L^3$ is the total number of lattice sites. In this paper, we fix $L=128$. Here, the two types of boundary conditions {(BC)} differ in how a quaternion $\bm{q}(x)$ with $x$ inside the boundary is connected to another quaternion $\bm{q}(x+\delta x)$ with $x+\delta x$ outside the boundary: in PBC $x+\delta x$ is translated back to the lattice by displacement vector $(mL,nL,lL)$ where $\{m,n,l\}$ are integers; in FBC we simply set $\bm{q}(x+\delta x) = (1,0,0,0)^T$. {In the following text, we always define the energy $E$ as the total energy where $E_\text{FM}$ has been subtracted.}
Now we move to the discussion of the unit charge Skyrmion on the lattice. Figure~\ref{fig profile lattice charge1} shows the relaxed unit charge solutions of the lattice model at $\kappa = 0$, where local minima are obtained by the combination of minimization and Langevin dynamics. The Hopf charge $Q=1$ may be immediately read out from the linking of the two curves in Fig.~\ref{fig profile lattice charge1}(e). The red curve is the \emph{position curve} which was discussed in Sec.\,\ref{sec Position curves and strings}, and is defined as the curve where the spin $S\equiv R\hat{z}$ takes the value $(0,0,-1)$. The blue curve will be referred to as the \emph{linking curve} and is instead where $S$ takes the value $(0,-1/\sqrt{2},-1/\sqrt{2})$.
The linking of the position curve and linking curve gives a clear definition of the Hopf charge on a lattice, but there is an alternative method for defining the topological charge from a finite difference approximation to \eqref{def Q Skyrme}, where the $J$ fields are expressed in terms of the quaternion components. The latter method does not produce exact integer values for $Q$ {numerically}, but the departure from an integer value may be used as a rough estimate of the intrinsic `discretization error' that may be expected from using a continuum Hamiltonian \eqref{ham Continuum} in place of the exact lattice Hamiltonian \eqref{ham S}. {Indeed, the departure from the exact integer is found to be more significant when $K_2$ moves away from the Lifshitz point ($K_2/|K_1|=1/4$) causing a reduction of the Skyrmion size relative to the lattice constant.}
\begin{figure}[tbp]
\centering
\includegraphics[width=1\textwidth]{Fig_lattice_charge1-300dpi.png}
\caption{The relaxed unit charge Skyrmion solutions on a $128\times 128\times 128$ lattice with $K_1=-1$, $\kappa=0$, and PBC. (a) The profile functions for different choices of $K_2$, where [1,1,1] is chosen as the radial direction. We have also computed the curves for $K_2=0.248$ with FPC, and the profile functions appear to be the same as the PBC ones by eye. (b) The profile functions for $K_2=0.248$ along different high-symmetry directions. (c) The profile functions for $K_2=0.245$ along different high-symmetry directions, along with the solution of the continuum theory Eq.~\eqref{ham Continuum}. (d) The topological charge density isosurface with $K_2=0.248$. (e) The position curve (red) and the linking curve (blue).}
\label{fig profile lattice charge1}
\end{figure}
\begin{table}[tbp]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\diagbox[width=3.5cm]{$\qquad\kappa$}{$K_2$} & 0.245 (PBC) & 0.246 (PBC) & 0.247 (PBC) & 0.248 (PBC) & 0.248 (FBC) \\
\hline
0 & 91.713 & 82.670 & 72.184 & 59.530 & 59.734 \\
\hline
0.625 & \slashbox{}{} & \slashbox{}{} & \slashbox{}{} & {58.746} & {58.930} \\ \hline
0.955 & \slashbox{}{} & \slashbox{}{} & \slashbox{}{} & {57.739} & {57.896} \\ \hline
\end{tabular}
\caption{The relaxed energy {$E$} of the {unit charge} Skyrmion on a $128\times 128\times 128$ lattice with $K_1=-1$. The boundary conditions are indicated in the parentheses.}\label{table:E_Q1}
\end{table}
\begin{table}[tbp]
\centering
\begin{tabular}{|c|c|c|}
\hline
\diagbox[width=2.5cm]{$\qquad\kappa$}{BC} & PBC & FBC \\
\hline
0 & 1.0 & 1.0 \\
\hline
0.625 & \quad 0.885 \quad & \quad 0.886 \quad \\ \hline
0.955 & \quad 0.834 \quad & \quad 0.837 \quad \\ \hline
\end{tabular}
\caption{ { $\langle \left( J_y^1\right)^2 \rangle/ \langle \left( J_y^3\right)^2 \rangle$ of the unit charge Skyrmion on a $128\times 128\times 128$ lattice with $K_1=-1$ and $K_2=0.248$.}}\label{table:J_Q1}
\end{table}
These unit charge solutions at $\kappa=0$ are actually well described by the continuum hedgehog ansatz \eqref{def Hedgehog}, and their approximate rotational symmetry may be seen in the topological charge density isosurface plotted in Fig.~\ref{fig profile lattice charge1}(d). The profile function may be found directly by minimizing the energy functional in the continuum theory, as discussed further in the next section. As shown in Fig.~\ref{fig profile lattice charge1}(c), the continuum profile function at $K_2=0.245$ agrees very closely with the profile function found by extracting the rotation angles along various directions in the lattice simulation. As $K_2$ is increased towards the Lifshitz point the Skyrmion size increases, as may be seen from the profile functions in Fig.~\ref{fig profile lattice charge1}(a). As shown in Fig.~\ref{fig profile lattice charge1}(b), at $K_2=0.248$ the Skyrmion size is comparable to the box size, and cubic anisotropies from the boundary conditions lead to the profile function being slightly different when calculated along three different high-symmetry directions on the cubic lattice.
{The parameter $\kappa$ (or equivalently $\theta$) allows us to interpolate between the $\text{SO}(3)$ and the $S^2$ target spaces. The energy $E$ of the relaxed unit charge Skyrmion for $\theta = \pi/6$ ($\kappa=0.625$) and $\theta = \pi/18$ ($\kappa \approx 0.955$) are recorded in Table~\ref{table:E_Q1}, which deviate only slightly from the $\kappa=0$ case. Small distortion of the topological charge isosurface also appears for $\kappa \neq 0$. To better indicate the departure from the $\kappa = 0$ limit, we also compare the average of $\left( J_\mu^1\right)^2$ to $\left( J_\mu^3\right)^2$ (see Table~\ref{table:J_Q1} for $\mu=y$). The results turn out to be quite similar to the ones shown in Table~\ref{table SqSkyrme} for the squashed Skyrme model. }
\subsection{Higher charge Skyrmions and rational maps} \label{sec Rational map}
For Skyrmions with relatively low charge, we can create them by the method of ``merging''. For $Q=\{2,3,4\}$, we follow Ref.~\cite{MantonBook} by putting multiple $Q=1$ Skyrmions in the attractive channel and wait until the energy is fully minimized. For higher charge, to avoid missing the lowest energy solution, we use multiple ways of merging. In particular, for $Q=5$, we try two possible combinations: $Q=1+4$ and $Q=2+3$, which are found to relax to the same state; For $Q=6$, we try $Q=1+5$, $Q=2+4$, and $Q=3+3$, where two solutions are found; For $Q=7$, we try $Q=1+6$, $Q=2+5$ and $Q=3+4$, which all relax to the same solution. The energies of the solutions can be found in Table~\ref{table multi-charge}, and the charge density isosurfaces can be found in Fig.~\ref{fig high charge}.
\begin{table}[tbp]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
\diagbox[width=2.0cm]{BC}{Q} & 1 & 2 & 3 & 4 & 5 & \multicolumn{3}{c|}{6} & 7\\
\hline
PBC & 59.530 & 56.274 & 53.883 & 52.608 & 52.411 & 52.043 & 52.100 & 52.261 & 51.316 \\
\hline
FBC & 59.734 & 56.469 & 53.974 & 52.742 & 52.571 & 52.264 & 52.350 & NA & 51.544 \\
\hline
\end{tabular}
\caption{The relaxed energy {per charge $E/Q$} of the $Q=\{1,2,\ldots,7\}$ Skyrmions on a $128\times 128\times 128$ lattice with $K_1=-1$, $K_2=0.248$ and $\kappa=0$. Note for $Q=6$: we {have found} a few extra stable local minima: the first two columns are results from merging, the 3rd column of PBC is a minimum relaxed from rational map, and the rational map with FBC is relaxed to $E/Q=52.264$.}
\label{table multi-charge}
\end{table}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.8\textwidth]{Fig_charge_high_150dpi.png}
\caption{Charge density isosurfaces of the relaxed higher charge Skyrmion solutions on a $128\times 128\times 128$ lattice with $K_1=-1$, $K_2=0.248$, and $\kappa=0$. The shapes for PBC and FBC are found to be the same by eye. Note for $Q=6$: all three isosurfaces are found to be stable for PBC, while only the first two on the left are found to be stable for FBC. }
\label{fig high charge}
\end{figure}
The charge density isosurfaces displayed in Fig.~\ref{fig high charge} are found to be the same as in the Skyrme model for $Q=\{1,2,3\}$, and for $Q=4$ we start seeing small deviations~\cite{houghton1998rational}. For higher charge $Q=\{5,6\}$, the isosurfaces are found to have very different symmetries compared to the Skyrme model, but for $Q=7$ the isosurface is again only slightly distorted from the one in the Skyrme model. {One exceptional case} is a higher-energy $Q=6$ state we find with PBC, which {has the same charge density isosurface} as the lowest energy $Q=6$ solution in the Skyrme model. The reason for not seeing it with FBC is suspected to be that the energy barrier to other $Q=6$ states is very low, so it is easy to miss this solution {in the numerical relaxation}.
It is apparent from Fig.~\ref{fig high charge} that the lower charge solitons are qualitatively similar to the corresponding Skyrmions in the Skyrme model~\cite{BRAATEN1990147} as well BPS monopoles in $SU(2)$ Yang-Mills~\cite{hitchin1995symmetric}, both of which may be approximated by the rational map ansatz~\cite{houghton1998rational,battye1997symmetric,battye2002skyrmions},
\begin{align}
U(r,w) = \cos f(r)\,I+ i \sin f(r)n(w)\cdot\sigma.\label{def Rational map ansatz}
\end{align}
Here $n^i$ is a unit vector, which is a generalization of $n^i=x^i/r$ in the hedgehog ansatz \eqref{def Hedgehog}. The spatial coordinates $x^i=(x,y,z)$ are expressed in a spherical coordinate system $r,w$, where $w$ is a complex coordinate which is a function of the angles $\theta,\phi$,
\begin{align}
w\equiv\quad\tan \frac{\theta}{2}e^{i\phi}=\quad\frac{x+iy}{r+z}. \label{def W}
\end{align}
The dependence of $n$ on $w$ may be expressed in terms of an analytic function $R(w)$,
\begin{align}
n=\frac{1}{1+|R|^2}\left(2\,\text{Re} R, \,2\,\text{Im} R,\, 1-|R|^2\right).\label{eq n definition}
\end{align}
This function $R$ is the rational map from which the ansatz gets its name. It is a rational function $R=p/q$ where $p,q$ are polynomials with no common roots. The degree of $R$ is defined as the maximum degree of $p$ or $q$, and it turns out that the degree is simply equal to the baryon charge $Q$ of the ansatz $U$.
To determine how well the lower charge solitons found in the direct lattice simulation fit the rational map ansatz, the continuum description of the Hamiltonian \eqref{ham Continuum} was used to optimize the profile function $f$ given some rational map $R$. The quartic terms \eqref{lagr Quartic} in this case are a bit more complicated than the Skyrme model, where all integrals over angle either lead to an expression for the charge
\begin{align}
Q=\frac{r^2}{4\pi}\int d\Omega\,\frac{1}{2}\sum_\mu \left(\partial_\mu n\right)^2,\label{eq Q n integral}
\end{align}
or a single non-trivial integral $\mathcal{I}_0$,
\begin{align*}
\mathcal{I}_0\equiv \frac{r^4}{4\pi}\int d\Omega\left[\frac{1}{2}\sum_\mu \left(\partial_\mu n\right)^2\right]^2.
\end{align*}
$\mathcal{I}_0$ may be easily expressed in terms of $R$ and minimized independently of the profile function~\cite{houghton1998rational}. On the other hand, the present model leads to four distinct angular integrals \eqref{def I} which are all coupled to the profile function and rather complicated if expressed in terms of $R$. In practice we simply took $R(w)$ to have the same discrete symmetry as it does in the Skyrme model, and for $Q\leq 4$ that completely fixes $R(w)$ so no minimization is necessary~\cite{houghton1998rational}. For $Q>4$, the parameters of the rational map were minimized directly in the lattice simulation as will be discussed below.
The structures of the Hamiltonian \eqref{ham Continuum} at $\kappa=0$ expressed in terms of the ansatz and averaged over solid angle are
\begin{gather}
\frac{1}{4\pi}\sum_\mu\int d\Omega \left(J^i_\mu\right)^2 =\left(f'\right)^2+\frac{2Q\sin^2 f}{r^2},\label{lagr Quad f}\\
\frac{1}{4\pi}\sum_\mu\int d\Omega \left( \left(J^i_\mu\right)^2\right)^2=\frac{3}{5}\left(f^\prime\right)^4+\frac{2\sin^2 f}{r^2}\left(f^\prime\right)^2\mathcal{I}_1+\frac{\sin^4 f}{r^4}\mathcal{I}_2\label{lagr Quart J4 f}
\end{gather}
\begin{align}
\frac{1}{4\pi}\sum_{\mu}\int d\Omega \left(\partial_\mu J^i_\mu\right)^2&= \frac{3}{5}\left(f^{\prime\prime}\right)^2+\frac{4}{5r}f^\prime f^{\prime\prime}+\frac{8}{5r^2}\left(f^{\prime}\right)^2\nonumber\\
&\qquad +4\frac{\cos^2 f}{r^2}\left(f^{\prime}\right)^2\mathcal{I}_1+\frac{\sin^2 f}{r^4}\mathcal{I}_3-\frac{\sin^4 f}{r^4}\mathcal{I}_2\nonumber\\
&\qquad -2\frac{\cos f\, \sin f}{r^2}f^{\prime\prime}\mathcal{I}_1+2\frac{\cos f\, \sin f}{r^3}f^{\prime}\left(-2Q+\mathcal{I}_1+2\mathcal{I}_4\right),\label{lagr Quart dJdJ f}
\end{align}
with the integrals defined as,
\begin{align}
\mathcal{I}_1&\equiv \frac{r^2}{4\pi}\sum_\mu\int d\Omega \left(\left(x^\mu\right)^2 \left(\partial_\mu n^i\right)^2\right) \no
\mathcal{I}_2&\equiv \frac{r^4}{4\pi}\sum_\mu\int d\Omega\left( \left(\partial_\mu n^i\right)^2\right)^2 \nonumber\\
\mathcal{I}_3&\equiv \frac{r^4}{4\pi}\sum_\mu\int d\Omega \left(\partial^2_\mu n^i\right)^2\nonumber\\
\mathcal{I}_4&\equiv \frac{r^3}{4\pi}\sum_\mu\int d\Omega \left(x^\mu\,\partial_\mu n^i\partial^2_\mu n^i\right).\label{def I}
\end{align}
The values of the $\mathcal{I}$ integrals are given in Table \ref{table I integrals}, including the hedgehog special case $R(w)=w$, which was used in the previous section. It is seen that the energy of optimal rational map ansatz in the continuum comes fairly close to the energy of the more general low charge solitons in the lattice model.
\begin{table}
\begin{center}
\begin{tabular}{ c||c|c|c|c|c||c|c|c| }
$Q$ & $R(w)$ & $\mathcal{I}_1$ & $\mathcal{I}_2$ & $\mathcal{I}_3$ & $\mathcal{I}_4$& $E_{ansatz}/Q$ & $E/Q$ & $E/Q_{num}$ \\
\hline
$1$ & $w$ & $0.4$ & $1.6$ & $3.2$ & $-0.8$ & $59.81$ & $59.530$ & $61.308$\\
$2$ & $w^2$ & $0.74926$ & $9.64307$ & $16.09$ & $-1.62537$& $59.92$ & $56.274$ & $57.687$ \\
$3$ & $\frac{\pm i \sqrt{3} w^2-1}{w \left(w^2\mp i \sqrt{3}\right)}$ & $1.24335$ & $20.7566$ & $33.5133$ & $-2.37832$& $57.92$ & $53.883$ & $55.189$ \\
$4$ & $\frac{w^4+2 i \sqrt{3} w^2+1}{w^4-2 i \sqrt{3} w^2+1}$ & $1.97218$ & $30.2953$ & $48.4568$ & $-3.01391$& $55.25$ & $52.608$ & $53.791$ \\
\hline
\end{tabular}
\caption{A comparison of the energy $E_{ansatz}$ of the rational map ansatz to the energy $E$ of the solitons found in the lattice simulation {with PBC}. All values are taken at $K_1=-1, K_2= 0.248$, {$\kappa=0$}, and the profile function was minimized in a finite volume of $r\leq 120$. $E$ is divided by both the exact charge $Q$ and a numerical charge $Q_{num}$ found from a finite difference approximation to \eqref{def Q Skyrme}.}\label{table I integrals}
\end{center}
\end{table}
Unlike the situation
in the Skyrme model, the profile function $f(r)$ and the rational function $R$ can not be minimized independently. {In this paper, we minimize them} simultaneously using a simulated annealing algorithm. Compared to local minimization, the simulated annealing method is advantageous in overcoming local minima with the help of thermal fluctuations. Typically we parametrize $f(r)$ by 20 to 30 discrete points and interpolate between them via Steffen's method which guarantees monotonicity. Our unit Monte Carlo (MC) step consists of updating $f(r)$ at each discrete point once, and updating each parameter in $R$ thirty times.
The initial temperature is $T_0=0.5 |K_1|$ and we bring it down to $T=0.001|K_1|$ in 3000 MC steps, then we use another 1000 MC steps for further equilibration at $T=0.001|K_1|$. We note that while the energy of the lattice model has to be evaluated at each MC update, there are only a few parameters to be minimized. This is in contrast to the full relaxation of the lattice model, where all $L^3$ quaternions $\bm{q}(x)$ are to be optimized.
{To illustrate these procedures, now} we consider the charge-5 Skyrmion of the lattice model. The rational map ansatz with $D_{2d}$ symmetry is~\cite{MantonBook}
\begin{equation}
R(w)= \frac{w (a+ibw^2+w^4)}{1+ibw^2+aw^4},\label{rational B5}
\end{equation}
where parameters $a$ and $b$ have to be optimized together with the profile function.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.9\textwidth]{Fig_rational_B5_300dpi.png}
\caption{Results of the charge-5 rational map ansatz~\eqref{rational B5} on a $L=128^3$ cubic lattice with $K_1=-1$, $K_2=0.248$, $\kappa=0$, and PBC. (a)(b) The evolution of energy and \{$a$, $b$\} during simulated annealing. (c)(d) The profile function and the charge density isosurface at the end of the simulated annealing (MC step$=4000$). The squares in (c) are the discrete points of $f(r)$ and the dashed line is the interpolation. (e) The charge density isosurface after fully relaxing the rational map ansatz.}
\label{fig rational B5}
\end{figure}
Figure~\ref{fig rational B5} shows the simulated annealing results of the charge-5 rational map. The optimized values $a\approx -2.99$ and $b \approx 3.97$ are quite close to the numbers of the Skyrme model ($a=-3.07$, $b=3.94$)~\cite{MantonBook}. Indeed, the charge density isosurface is the same as the one in the Skyrme model [Fig.~\ref{fig rational B5}(d)]. After further full relaxation, the energy and charge density isosurface converge to the results from merging [see Fig.~\ref{fig high charge}]. This result demonstrates both the usefulness and limitations of the rational map ansatz. On one hand, it allows us to construct Skyrmions with higher charge without going through the merging process. On the other hand, the rational map ansatz is sometimes incompatible with the anisotropies, and can become unstable towards the lower energy solution after full relaxation.
We have also performed simulated annealing on the $Q=6$ and $Q=7$ rational maps. For $Q=6$, we use the rational map with $D_{4d}$ symmetry:
\begin{equation}
R(w) = \frac{w^4+ia}{w^2 (iaw^4+1)},
\end{equation}
where both the parameter $a$ and the profile function $f(r)$ are optimized. After further full relaxation, the result is found to be stable (the 3rd $Q=6$ plot in Fig.~\ref{fig high charge}) with PBC, while it tunnels to the lowest energy solution from merging (the 1st $Q=6$ plot in Fig.~\ref{fig high charge}) with FBC. The tunneling is suspected to be caused by the small energy barrier between different $Q=6$ states.
For $Q=7$, we use the rational map with $Y_h$ symmetry:
\begin{equation}
R(w) = \frac{w^7-7w^5-7w^2-1}{w^7+7w^5-7w^2+1}.
\end{equation}
In this case, only the profile function $f(r)$ has to be optimized. The result of simulated annealing with full relaxation is found to also deviate slightly from the one in Skyrme model (Fig.~\ref{fig high charge}).
\begin{figure}[tbp]
\centering
\includegraphics[width=0.97\textwidth]{Fig_Q10_FBC-150dpi.png}
\caption{The charge density isosurfaces and the position curves (red) and linking curves (blue), for fully relaxed charge $Q=10$ Skyrmion solutions on a $128\times 128\times 128$ cubic lattice with $K_1=-1$, $K_2=0.248$, $\kappa=\{ 0,\, 0.625,\,0.955 \}$, and FBC. The position curves in (a)(b) are all simple 1-rings except the bottom rows. The position curves in (c) include linked rings and knots, where the relative positions are illustrated by the crosses.}
\label{fig Q10}
\end{figure}
\subsection{Charge-10 Skyrmions and Position Curves} \label{sec charge-10}
{
In the $CP^1$ limit ($\beta = \kappa = 1$), the position curves were shown to have nontrivial structures including {linked rings} and knots~\cite{battye1998knots,sutcliffe2017skyrmion}. Here we show that such structures also appear naturally in our lattice model \eqref{ham S} when we interpolate to large $\kappa$.
}
{As we discussed in Sec.~\ref{sec Position curves and strings}, the position curves are not uniquely defined at the PCM limit ($\beta=\kappa=0$), due to the increased symmetry from $U(1)$ to $SU(2)$. In the quaternion representation, the energy \eqref{ham quaternion} is invariant under a global rotation \begin{equation}
\tilde{\bm{q}}(x)\equiv \bm{p} \bm{q}(x) \overline{\bm{p}}
\end{equation} for any given quaternion $\bm{p}$, {but this transformation does not leave the position curve invariant}. To get around this problem, in the following we always first find the optimal $\bm{p}$ that minimizes the energy for $\kappa \rightarrow 0+$ when plotting the position curves for $\kappa=0$.}
{Away from the $\kappa=0$ limit, both the energy and the third spin component $S^3(x)=q_0^2(x) + q_3^2(x)-q_1^2(x)-q_2^2(x)$ are invariant if $\bm{p}$ corresponds to rotation around the $\hat{z}$ axis. In other words, the position curve defined by ${S}=(0,0,-1)$ is unique under such global transformations, but position curves along other spin directions are not generally invariant. For this reason, we mainly consider the ${S}=(0,0,-1)$ position curves in this section; while the linking curves at ${S}=(0,-1/\sqrt{2},-1/\sqrt{2})$ are not unique, we also plot them together with the position curves to show their relative linking structures.}
{Figure~\ref{fig Q10} shows the fully relaxed $Q=10$ solutions of the lattice model for three different choices of $\kappa$. For each $\kappa$, since we used multiple ways of ``merging'' and rational maps, multiple local minima are obtained. The lowest energy $\kappa =0$ solutions [Fig.~\ref{fig Q10}(a), first two rows] shows clear deviation from the relaxed rational map with $D_{4d}$ symmetry [Fig.~\ref{fig Q10}(a), bottom row]. The position curves are all simple 1-rings except the relaxed rational map in which the position curve forms a net. We note that such net-like structure was also observed for low charge Skyrmions at $\beta=0$ in the Skyrme model~\cite{ward2004skyrmions}. }
{Both the topological charge isosurfaces and the position curves change dramatically as $\kappa$ is tuned away from zero. For $\kappa=0.625$, the position curves are found to be either a simple 1-ring, or several 1-rings which are disjointed [see Fig.~\ref{fig Q10}(b)]. For a large value $\kappa\approx 0.955$, the position curves start developing nontrivial topological features, including both linked rings and knots [see Fig.~\ref{fig Q10}(c)].
This is indeed as expected since such topological structures were known to exist in the $\kappa=1$ limit~\cite{sutcliffe2017skyrmion}.}
\section{Conclusion and discussion} \label{sec Conclusion}
{3D Skyrmions proper in the sense of the Skyrme model are shown to be stabilized in a frustrated spin model on the cubic lattice.
By tuning a parameter that describes the ``collinearity'' of the magnetic ground state, {the model interpolates between two limits with $S^3$ and $S^2$ target spaces}. In the $S^3$ limit, the Skyrmion solutions are found to be qualitatively the same as in the Skyrme model for small $Q$, and they start to deviate for $Q\gsim 4$. Near the $S^2$ limit, the position curves of the Skyrmions are found to develop nontrivial topological structures including {linked rings} and knots.
}
{Since the lattice model considered in this paper can be regarded as the low-energy effective model for a broad class of 3D non-collinear magnets where $SO(3)$ rotation is the only low-energy mode, it is expected that 3D Skyrmions should also appear in realistic models that have {SO(3) Goldstone modes, whether squashed or not.}
To this end, it is worth emphasizing}
a few necessary ingredients in the search of 3D Skyrmion excitations (defects) in magnetic systems~\cite{BatistaEtAl2018}. First, a non-collinear ground state is required to ensure a target space homeomorphic to $S^3$, which can be commonly realized in frustrated spin systems. Second, the Skyrmion size has to be much larger than the lattice spacing for the excitation to be topologically protected (energy barrier {large enough} between different topological sectors). For the toy model considered in this paper, the Skyrmion size becomes large when $K_2/|K_1|\lesssim 1/4$. More generally, this condition is satisfied for systems near a Lifshitz point (a commensurate to incommensurate transition).
{Similar to the 2D Skyrmion crystals that are commonly studied in condensed matter systems,}
3D Skyrmion crystals are expected to be realized {as the ground state (vacuum of the theory)} on the other side of the Lifshitz transtion.
{We note that the precise definition of ``Skyrmion crystal'' is slightly different in the condensed matter and high-energy literatures: the condensed matter community often refers ``Skyrmion crystal'' as the $T=0$ ground state or the finite-$T$ equilibrium state (Skyrmion crystal becomes the new vacuum), while in high-energy ``Skyrmion crystal'' is often referred as an excited state in the original vacuum.
In both cases, the Skyrmion crystal can be described as ``multi-$\bm{Q}$'' states (linear combination of multiple incommensurate spirals), whose energy can be quite close to other single- or multi-$\bm{Q}$ states.}
Such degeneracy {is often} lifted by spin anisotropy, magnetic field, thermal and {quantum} fluctuations.
{Besides magnetic systems, we note that 3D Skyrmion crystals were also predicted to be realized in cold atom systems described by multicomponent imbalanced superfluids~\cite{Samoilenka2020}.}
{Lorentz transmission electron microscopy is often used for direct visualization of 2D magnetic Skyrmions.}
More recently, magnetic
X-ray tomography was successfully applied to 3D systems for the visualization of Skyrmion strings and Hopfions~\cite{seki2021direct}. In principle, 3D Skyrmions could also be detected by the same X-ray tomography methods. As we noted in this paper, the close connection between 3D Skyrmions and Hopfions implies that visualization of
{the underlying position curves}
{can be}
{strong evidence of 3D Skyrmion formation, if the ground state is known to be non-collinear.}
{Small angle neutron scattering is also a useful tool to see the multi-$\bm{Q}$ structure of the underlying spin arrangements, which serves as indirect evidence of Skyrmion crystal formation.}
{Finally, let us note that the picture we have presented of the continuity between Skyrmions and string-like Hopfions may have some relevance in high-energy physics to the study of the Skyrme model and its various extensions and modifications. As we have discussed in Sec. \ref{sec Position curves and strings} a solution with non-zero baryon charge in the Skyrme model may equivalently be considered as a knotted or twisted loop of string with long range interactions associated with the $J^3$ field that winds around the string core. It may be difficult to make use of this picture in a concrete way since for minimum energy configurations the radius of curvature of the position curve is on the same order of magnitude as the string thickness. But at the very least in the limit of the Faddeev-Niemi model there are unstable configurations which are well described by a thin string ansatz along these lines \cite{battye1999solitons}.
The main qualitative difference between the Faddeev-Niemi model at $\beta=1$ and the squashed Skyrme model for $\beta<1$ in this point of view is that former involves local strings whereas the latter involves global strings with long range interactions. It may be interesting to explore whether this is connected to the difference which must be present in large $Q$ solutions given the linear energy bound \eqref{eq BPS bound} we have found here. While these considerations are certainly more speculative than the possibility of direct detection of 3D Skyrmions in condensed matter systems discussed above, it may also be fruitful to investigate these analogies between the Skyrme model and systems of stringy topological defects in further detail.}
\begin{acknowledgments}
We would like to acknowledge Cristian Batista for helpful discussions. D.S. was supported in part by the U of MN Doctoral Dissertation Fellowship. During the writing of this paper, Z.W. was supported by the U.S. Department of Energy through the University of Minnesota Center for Quantum Materials, under Award No. DE-SC-0016371. C.N. was supported by the Olle Engkvist foundation, Grant No 204-0185. M.S. is supported in part by DOE Grant No. DE- SC0011842.
\end{acknowledgments}
|
\section{Introduction}
Social media platforms are strongly criticised nowadays for enabling the spread of hate speech--the expression of hate or encouraging violence based on race, nationality, religion, and sexual orientation, among others. While social media providers take some measures for fighting hate speech, such as the sampling and screening of suspicious content by dedicated systems and workforce, these measures fail to identify and mitigate the manifestation of hate speech at scale. To this end, automatic methods for detecting hate speech are being developed, where many challenges are yet to be addressed.
\textcolor{black}{In accordance with the advances of neural computing methods for text processing, recently proposed methods of automatic hate detection employ high-performing large neural network architectures, mainly, the transformer-based BERT~\citep{devlin2018bert} and its variants, e.g.,~\citep{tranEMNLP20,sarkarEMNLP21,miok2021ban}. These models are pretrained as contextual encoders of token and sentence meaning using vast amounts of general unannotated texts. In classification, the underlying meaning of a given text is encoded by these models onto a low-dimension vector space, which is then mapped to specified target categories given labeled examples in another learning step named as {\it fine-tuning}.}
One key challenge of learning automatic models of hate detection is the paucity and lack of diversity of labeled hate speech examples. Since the vast majority of social media content is not offensive or hateful, a massive amount of data needs to be screened in order to collect a sufficient and representative sample of hate speech examples. Researchers therefore revert to collecting hate speech samples that contain specified terms, or focus on the content posted by accounts that are known to be hateful. Quality labeling demands reaching an inter-coder agreement regarding the correct label for every sample, making the data annotation process complicated, time consuming and expensive. For these reasons, hate speech datasets are in most cases small, topically-biased, and imbalanced, including relatively few hate speech examples~\citep{wiegand2019detection,wullach2020towards}. \textcolor{black}{Several recent works investigated methods for automatically expanding the existing datasets by means of data augmentation~\citep{fengACL21}, for example by means of back translation and paraphrasing~\citep{rizos2019augment,beddiarJO21}}. In this work, we follow the approach by Wullach et al.~\citeyearpar{wullach2020towards}, augmenting the available labeled datasets using synthetic hate and non-hate examples generated by GPT-2~\citep{radford2019language}, a pretrained generative language model which we fine-tune to synthesize text sequences that are similar yet diversify the available relevant examples. The datasets that we synthesized in this work are balanced and an order of magnitude larger than any previously explored hate speech datasets, accumulating in total to 10M generated text sequences.
Another main challenge of identifying hate speech automatically is that the textual content posted on social media platforms (e.g., Twitter) in general, and hateful expressions in particular, \iffalse Hate speech data obtained from \fi are often noisy and diverse language-wise, as users tend to quickly adopt new terms and use creative and personalized language styles. \iffalse as hate speech has many flavors. This property imposes another challenge in training a \fi State-of-the-art deep-learning hate speech detectors exhibit vocabulary limitations, using a memory expensive lookup table which maps each text \textcolor{black}{token} to a high-dimensional embedded vector~\citep{mikolov2013distributed}. These models are therefore challenged by the rich diversity and irregularity of social media language~\citep{achilles20}.
Finally, a challenge of practical importance is that state-of-the-art deep learning networks are extremely large, reaching up to hundreds of millions, or billions of parameters. \textcolor{black}{The large size of these models and their high latency prevent their deployment at large scale, or in conditions that are resource-limited, such as mobile edge computing~\citep{7879258}.} Ideally, hate detection would be performed in real time at the end device, allowing to alert the user prior to posting hateful content, yet high-performing compact hate detection models are required to achieve this goal~\citep{tranEMNLP20,mitraICSC21}. In order to alleviate the associated deployment costs, researchers have recently proposed to distill the large pretrained transformer-based models into smaller neural architectures, typically reducing the original model size by up to one magnitude of order, e.g.,~\citep{sanh2019distilbert,tinyBERT,mobileBert}.
In this work, we introduce new learning architectures of hate speech detection that are based on HyperNetworks, a special class of deep learning networks, which utilize weight sharing across layers~\citep{ha2016hypernetworks}. The proposed architectures process the text at character-level--as opposed to \textcolor{black}{token}-level. This results in a significant reduction in the total number of learnable parameters, introducing hate detection classifiers that are extremely efficient and compact.
\textcolor{black}{We assess the proposed models alongside several deep learning networks, including the popular large transformer-based BERT and RoBERTA~\citep{roberta19} architectures, which model text at wordpiece level. We further consider in our experiments the smaller transformer-based models of ALBERT~\citep{albert} and MobileBERT~\citep{mobileBert}, a recently introduced distilled variant of BERT, as well as the recently proposed model of CharBERT~\citep{charbert}, which introduces character embeddings alongside the wordpiece information into BERT. In addition, we experiment with CNN-GRU~\citep{zhang2018detecting}, a smaller neural network model that has previously shown strong results on the task of hate detection, modeling text at word level. Our empirical evaluation applies to five public datasets labeled for hate speech detection. We train and assess the models using examples drawn from from a single dataset, as well as report cross-dataset evaluation results, where the models are trained and tested using examples drawn from different datasets. The latter scenario is more challenging, and more realistic in assuming a possible shift in the underlying data distribution. In order to improve generalization, we train the models in both setups using increasing amounts of training examples, leveraging our large-scale resource of synthetic examples.}
\textcolor{black}{Our findings extend previous research~\citep{wullachEMNLP21}, showing that all of the above networks benefit from data augmentation, especially in data shift conditions, and mainly due to consistent improvements in recall. We find that while the HyperNetwork models are inferior to the larger models given limited amounts of labeled data, training the HyperNetworks using increasing amounts of generated labeled data, up to 1-2M text sequences, consistently improves and achieves competitive and in some cases even better performance than the popular deep learning methods, while requiring only a fraction of the number of parameters.}
Based on these results, we believe that the proposed networks, which include only tens of thousands of parameters, pave the way to \textcolor{black}{mobile edge computing~\citep{7879258}, allowing} automatic hate speech deployment in end-user devices with limited computation resources. \iffalse including smartphones and tablets.\fi
The main contributions of this paper are three-fold:
(1) the paper presents for the first time the utilization of HyperNetworks for hate speech detection. The proposed networks operate at character-level, and have an exceptionally low number of parameters.
(2) by increasing significantly the training data set size via text generation, the proposed solutions are demonstrated to achieve competitive, or better performance in some cases, than state-of-the-art deep learning models, which are orders of magnitude larger than the proposed solutions. (3) the paper is accompanied by a new hate speech corpus that is the largest ever created (10M sequences).\footnote{The created dataset will be shared for research purposes upon request to the authors.} This new corpus results from data generation using a state-of-the-art deep generative language model, fine-tuned to approximate the distributions of five hate speech public datasets. \textcolor{black}{Our results indicate that all hate detection models benefit greatly from training using this resource.}
\textcolor{black}{The rest of the paper is organized as follows: Section 2 reviews hate speech detection methods and datasets, Section 3 presents the proposed HyperNetworks-based approach, Section 4 provides a detailed performance evaluation, and Section 5 concludes the paper.}
\section{Background: Hate Speech Methods and Data Resources}
Research has proposed several approaches for hate speech detection over the past years, recently employing deep learning approaches. Deep learning classification models are known to perform and generalize well when a large, diverse, and high-quality data resource is available for training. In the lack of such a resource, previous related works constructed datasets for this purpose, which were manually labeled, and are therefore strictly limited in size.
In this Section, we describe several popular public labeled datasets that have been used by researchers for training and evaluating hate speech detectors. Following previous works~\citep{wullach2020towards,wullachEMNLP21}, we exploit a large-scale corpus of generated hate and non-hate sequences that extend these datasets for training the various hate classification methods included in our experiments.
This Section further includes a general overview of existing approaches of hate speech detection, including a detailed description of the hate speech classification methods that we apply in this work. Later, we review related literature concerning character-based convolution neural networks and parameter generation via HyperNetworks, as we utilize this methodology in proposing compact and efficient character-based hate detectors.
\subsection{Hate speech datasets}
\label{ssec:datasets}
The typical process that has been traditionally applied by researchers and practitioners for constructing labeled datasets in general, and of hate speech in particular, involves the identification of authentic hate speech sentences, as well as counter (non-hate) examples. Twitter is often targeted as a source of relevant data; this (and other) public social media platform allows users to share their thoughts and interact with each other freely, making it a fertile ground for expressing all kinds of agendas, some of which may be racist or hateful. The initial retrieval of hate speech examples from Twitter is based on keyword matching, specifying terms that are strongly associated with hate.\footnote{e.g., \url{https://www.hatebase.org/}} Once candidate tweets are collected, they are assessed and labeled by human annotators into pre-specified categories.
The manual annotation of the examples is intended to result in high-quality ground truth labeled datasets. Yet, manual annotation is costly~\citep{modhaESWA20}, and accordingly, the available datasets each include only a few thousands of labeled examples. Due to their size limit, and biases involved in the dataset collection process, e.g., keyword selection and labeling guidelines, these datasets are under-representative of the numerous forms and shapes in which hate speech may be manifested~\citep{wiegand2019detection}.\footnote{Another caveat of referring to authentic content within a dataset concerns the discontinued availability of the collected texts by the relevant provider over time. For example, tweets must be stored by their identifier number, where access to the tweet's content may be defined, impairing the dataset.}
Table~\ref{tab:datasets} details the statistics of five popular public datasets of hate speech, which we experiment with in this work. All of the datasets were manually curated and are of modest size, including a few thousands of labeled examples, out of which a minority contain hate speech.
Let us describe the individual datasets listed in Table~\ref{tab:datasets} in more detail. The dataset due to Davidson et al.~\citeyearpar{davidson2017automated} (\textbf{DV}) includes tweets labeled by CrowdFlower\footnote{https://www.welcome.ai/crowdflower} workers into three categories: {\it hate speech}, {\it offensive}, or {\it neither}. For the purposes of this work, we only consider the examples of the first and latter categories. Waseem and Hovy~\citeyearpar{waseem2016hateful} created another dataset (\textbf{WS}), considering tweets of accounts which frequently used slurs and terms related to religious, sexual, gender and ethnic minorities; those tweets were manually labeled into the categories of {\it racism}, {\it sexism} or {\it neither}. Again, as we focus on hate speech, and strive at compatible labels across datasets, we only consider the first and latter categories as examples of hate and non-hate, respectively. Another dataset was constructed by SemEval conference organizers~\citep{basile2019semeval} for the purpose of promoting hate detection (\textbf{SE}). They considered the historical posts of identified hateful Twitter users, narrowed down to tweets that included hateful terms, and had examples labeled by CrowdFlower workers. Many of the tweets labeled as hateful in this dataset target women and immigrants. Founta et al.~\citeyearpar{founta2018large} (\textbf{FN}) performed iterative sampling and exploration while having tweets annotated using crowdsourcing. Their resulting dataset is relatively large ($\sim$80K examples), and distinguishes between multiple flavors of offensive speech, namely {\it offensive}, {\it abusive}, {\it hateful}, {\it aggressive}, {\it cyber bullying}, {\it spam} and {\it none}. In order to maintain our focus on hate speech, we consider the labeled examples associated with the {\it hateful} and {\it none} categories. Finally, the dataset due to de-Gibert et al.~\citeyearpar{de-gibert18} (\textbf{WH}) was extracted from the extremist StormFront Internet forum,\footnote{https://en.wikipedia.org/wiki/Stormfront\_(website)}. This dataset aims to gauge hate in context, considering also cases where a sentence does not qualify as hate speech on its own, but is interpreted as hateful within a larger context comprised of several sentences.
We leveraged these datasets for (a) training and evaluating several types of hate speech detectors, (b) fine-tuning a pre-trained generative model and automatically producing a large number of additional similar hate speech sequences for training purposes. In our experiments, we split the available examples into fixed train (80\%) and test (20\%) sets, while maintaining similar class proportions. Only the train examples are used in the sequence generation process. Additional details about these datasets, as well as examples of the tweets generated per dataset, are available in Wullach {\it et al.}~\citeyearpar{wullach2020towards}.
\textcolor{black}{While our experiments concern hate speech detection in English, our approach is language-independent. Being semi-supervised, it requires relevant labeled examples in the target language, as well as a language-specific model decoder such as GPT, for expanding the manually labeled datasets by generating additional examples synthetically. Such resources in non-English languages are currently scarce but are being developed~\citep{multilingual,deVriesACL21}.}
\begin{table}[t]
\small
\centering
\caption{Publicly available hate speech datasets: size statistics}
\label{tab:datasets}
\setlength\tabcolsep{0pt}
\begin{tabular*}{0.7\columnwidth}{@{\extracolsep{\fill}} lrc}
Dataset & Size & Hate class \%\\\hline\hline
{~\cite{davidson2017automated} (\textbf{DV})} & {6K} & {24\%}\\
{~\cite{founta2018large} (\textbf{FN})} & {53K} & {11\%}\\
{~\cite{waseem2016hateful} (\textbf{WS})} & {13K} & {15\%}\\
{~StormFront~\citep{de-gibert18} (\textbf{WH})} & {9.6K} & {11\%}\\
{~SemEval2019~\citep{basile2019semeval} (\textbf{SE})} & {10K} & {40\%}\\
\hline
\end{tabular*}
\end{table}
\subsection{Hate Speech Detection}
Deep learning (DL) can effectively exploit large-scale data, learning latent representations using multi-layered neural network architectures. Various modern DL architectures of text classification consist of a word embedding layer \textcolor{black}{(a dimensionality reduction (\cite{ding2021dimensionality}) technique that is commonly utilized in DL-based natural language processing models)} that is intended to capture generalized semantic meaning of words, mapping each word in the input sentence into a vector of low-dimension~\citep{mikolov2013distributed}. The following layers learn relevant latent feature representations, where the processed information is fed into a classification layer that predicts the label of the input sentence.
For example, previously, Founta et al.~\citeyearpar{founta2019unified} employed the following DL architecture for detecting hate speech (and other types of offensive texts). They transformed the input words into GloVe word embeddings~\citep{pennington2014glove}. They then used a recurrent layer comprised of Gated Recurrent Unit (GRU) for generating contextual and sequential word representations, having each word processed given the representations of previous words in the input sentence. Following a dropout layer (intended to prevent over-fitting), the final dense layer outputs the probability that the sentence belongs to each of the targets using a softmax activation function. Other works on hate speech detection employed Convolution neural networks (CNN) for modeling contextual information, e.g.,~\citep{badjatiya2017deep}. CNN applies a filter over the input word representations, processing local features per fixed-size word subsequences. Hierarchical processing, which involves aggregation and down-sampling, consolidates the local features into global representations, which are fed into the final layer that predicts the probability distribution over the target classes.
In this work, we consider a popular DL classification architecture due to Zhang et al.~\citeyearpar{zhang2018detecting}, which comprises both convolution and recurrent processing layers. \textcolor{black}{This architecture has been previously shown to yield top-performance results on the task of hate speech classification, and is therefore evaluated as one of the architectures of choice in our experiments, similar to other related studies~\citep{hatexplain}.} In brief, this text classification network consists of a convolution layer applied to the input word embeddings, which is down-sampled using a subsequent max-pooling layer. The following layer is recurrent (GRU), producing hidden state representations per time step. Finally, a global max-pooling layer is applied, and a softmax layer produces a probability distribution over the target classes for the given input. Further implementation details are provided later in this paper.
\subsection{Hate Speech Detection with Pre-trained Language Models}
\textcolor{black}{More recently, larger and deeper language models (LMs), that have been pretrained on massive heterogeneous corpora, were shown to yield state-of-the-art contextual text representations, leading to further improvement in text classification. We focus our attention on the popular transformer-based language encoder of BERT~\citep{devlin2018bert} and its variants, as described below.}
\paragraph{BERT}
Text classification using pretrained language models like BERT~\citep{devlin2018bert} has been shown to give state-of-the-art performance on a variety of text processing tasks. While BERT generates task-agnostic contextual word embeddings, it can be optimized to target tasks via {\it fine-tuning}~\citep{devlin2018bert}. We follow the common practice of turning BERT into a text classifier by adding a final feed-forward network, which receives as input the embedding of the input text as processed by BERT (the `[CLS]' token embedding, which has been tuned to represent the meaning of the whole input text sequence), and outputs the target class probabilities via a softmax layer. Given labeled training data, the extended network parameters, including the weights of the terminal network, as well as BERT parameters, are fine-tuned jointly to optimize classification performance.
\paragraph{RoBERTa}
This model applies the same architecture as BERT, but has been trained on ten times more data, including news articles and Web content. Due to this augmentation of training data, and other modifications to the pretraining procedure and cost function, RoBERTa has been shown to outperform BERT on multiple benchmark datasets~\citep{roberta19}.
\paragraph{ALBERT}
The architecture of ALBERT~\citep{albert} was designed as a light variant of the transformer-based BERT model. It enhances BERT in several ways, including a factorization of the embedding parameters and cross-layer parameter sharing; both measures are intended to improve parameter efficiency as well as a form of regularization. ALBERT also replaces the next-sentence-prediction loss that is used in training BERT with sentence-order prediction loss, which focuses on modeling inter-sentence
coherence. As a result, ALBERT has been shown to outperform BERT on several multi-sentence encoding tasks~\citep{albert}. \\
It has been shown that while BERT has been pretrained using vast amounts of textual data, namely, all of Wikipedia and Google Books, the word encodings it produces can benefit from further adaptation to the target domain and task~\citep{gururACL2020}. Indeed, there exist multiple works that aim to adapt pretrained models like BERT to the task of hate detection using additional relevant data. Isaksen and Gamb{\"a}ck~\citeyearpar{isaksen2020} report results of fine-tuning the base and large variants of BERT on each of the \textbf{DV} and \textbf{FN} datasets. In addition to the pretrained version of these methods, they adapted the models by further training them using unlabeled examples from additional hate detection datasets. Overall, they found the performance of BERT-base to be comparable or better than BERT-large, but failed to improve the model results using the unlabeled data. Another work introduced HateBERT~\citep{hatebert}, a model of BERT (base) which was further trained on Reddit comments extracted from communities banned for being offensive, abusive, or hateful. A comparison of BERT and HateBERT on hate detection yielded mixed and non-conclusive results. \textcolor{black}{More recently, Sarkar et al.~\citeyearpar{sarkarEMNLP21} introduced {\it fBERT}, a variant of BERT (base) which they continued to train using large amounts of tweets that were labeled as offensive using semi-supervised classification. Unlike HateBERT, they achieved performance improvements on several experimental datasets.}
\textcolor{black}{In our experiments, we fine-tune and evaluate BERT, RoBERTa and ALBERT as hate detection classifiers on multiple datasets. Similar to other works~\citep{wiegand2019detection,mozafariPLOS2020,wullachEMNLP21}, we consider both within- and cross-dataset evaluation setups, testing the models generalization when trained and tested on examples drawn from different data distributions. Here and elsewhere~\citep{wullachEMNLP21}, we show that augmenting the available training datasets with additional generated examples results in substantial performance gains and improved generalization.}
\iffalse We propose an approach for detecting hate speech using character-based models. Previous works demonstrated the advantages of utilizing a character-based models, achieving competitive performance by utilizing character-based for detecting hate speech. Such models are composed of a significantly lower number of parameters, making them less complex and better fitted on memory-limited devices. Moreover, character-based models are less sensitive to misspelled and out-of-vocabulary terms compared to models that use a word (or sub-word) vocabulary, an valuable attribute when dealing with data drawn from Tweeter. We also experiment with "Weight Generating Networks"~\citep{ha2016hypernetworks, choi2019adaptive}, allowing the model to share weights across layers and generate model weights that are conditioned on the input.\fi
\textcolor{black}{Admittedly, while transformer-based pretrained language models yield state-of-the-art performance, they are characterised with large model sizes and high latency. Concretely, here we apply the base configurations of BERT and RoBERTa, which both include 110 million parameters. Ongoing efforts aim to design smaller versions of these models with faster inference times by means of distillation. The distilled models of DistilBERT~\citep{sanh2019distilbert} and TinyBERT~\citep{tinyBERT} have $\sim$40\% and 7.5x less parameters than BERT base, respectively. Here, we experiment with ALBERT, a light and competitive variant of BERT that has 17 million parameters, i.e., $\sim$9x less parameters than BERT base. We further experiment with MobileBERT~\citep{mobileBert}, another recently introduced model that distills the pretrained BERT into a deep and thin model that includes 24 million parameters, being 4.3× smaller compared with BERT base.}
\textcolor{black}{As detailed in Table~\ref{tab:params}, the character-based networks that are proposed and evaluated in this work are smaller by orders of magnitude compared with the transformer-based models, as well as compared with CNN-GRU. Yet, we show that these compact models, when trained using large amounts of relevant data, can achieve competitive levels of performance.}
\subsection{Character-based models}
\textcolor{black}{Character-level CNNs have been previously shown to be successful on several natural language processing (NLP) tasks, e.g.,~\citep{kim2015character, zhang2015character,mehdad2016characters,charSemEval}. Mehdad and Tetreault~\citeyearpar{mehdad2016characters} considered the task of abusive language detection, and showed that light-weight and simple character-based approaches might be superior to token-based modeling using adequate methods on this task; specifically, they experimented with recurrent neural network architectures.} Compared with the character-level deep learning networks proposed previously~\citep{zhang2015character}, our solution is more shallow and compact, as it incorporates HyperNetworks~\citep{ha2016hypernetworks}. The HyperNetwork architecture utilizes a relaxed-form of weight sharing across the network layers, enabling adaptive tuning of the network weights according to specific input text sequences. It therefore promotes generalization, while further reducing the parameter space.
The popular deep learning architectures that we consider in this work model language at word-level (CNN-GRU) or sub-word level (BERT and its variants). \textcolor{black}{Sub-word lexicons are typically generated by applying a learning algorithm over the training set, aiming to find the most likely vocabulary representing the training data. A prominent subword representation used by modern language models is WordPiece, which assembles a subword lexicon by applying an iterative token merging procedure~\citep{kudoACL18}.} \iffalse and SentencePiece, which treats the text as a stream of characters including the spaces between characters~\citep{kudo2018sentencepiece}}.\fi
An important advantage of character-level as opposed to word or sub-word language processing is its flexibility in handling unknown out-of-vocabulary terms, morphological inflections, and noisy word variants that are prevalent on social media. Conversely, it has been shown that the BERT model is highly sensitive to noise in the data, such as spelling mistakes and word variations~\citep{achilles20}. \textcolor{black}{Previously, several deep contextual language modeling architectures, such as ELMO~\citep{elmo}, incorporated character-level information. The recently proposed model of {\it CharBERT}~\citep{charbert} enhances the BERT and RoBERTa models by fusing the representations of characters and subwords, and applies a new pre-training task named NLM (Noisy LM) for unsupervised character representation learning. Overall, CharBERT adds 5M parameters to BERT or RoBERTa, modeling a character channel in addition to the token channel. We experiment with the CharBERT architecture in this work. Unlike CharBERT, the proposed HyperNetwork architectures model the text merely as character sequences, and are orders of magnitude smaller compared with the alternative approaches.}
While character-level text processing may place less emphasis on encoding high-level relationships between words~\citep{zhang2018detecting}, this approach is substantially more compact, and requires modest memory resources in comparison to word-level deep networks.
We describe and motivate the use of several variants of character-level HyperNetworks for hate detection (Section~\ref{sec:charhyper}).
We further show that such light-weight modeling yields high performance when provided with sufficient amounts of task-specific labeled data, that allows to learn relevant semantic and grammatical phenomena, as well as reduce generalization error~\citep{zhang2015character}. Our approach of generating large amounts of hate- and non-hate text sequences for training these models is described in Section~\ref{sec:gen}.
\section{The Proposed Approach: Character-level HyperNetworks}
\label{sec:approach}
\begin{figure*}[t]
\centering
\includegraphics[trim={0.0cm 13cm 0cm 1cm},clip,scale=0.5]{HyperNet.png}
\caption{The proposed HyperHate-Static architecture: The main network (lower part) computes the posterior probability that an input text sequence is hateful. The weights of the two 1D-convolutional layers are generated by the auxiliary network (upper part). The figure shows for illustration purposes a separate auxiliary network per convolutional layer, however, only a single auxiliary network is implemented and the weights for each convolutional layer are determined by the corresponding layer embedding vector $\mathbf{z}_j$ for $j=1,2$, respectively. The output of the auxiliary network is utilized as weights of each 1D-Convolutional layer, according to (\ref{eqref:static}).}
\label{fig:HyperHate}
\end{figure*}
\subsection{Character-level HyperNetworks for Hate Speech Detection}
\label{sec:charhyper}
HyperNetworks generally include two sub-networks: a main network--the hate speech detector in our task, and an auxiliary network that generates the weights of the main network.\footnote{In some publications, the auxiliary network is referred to as the \textit{hypernetwork}, yet, in order to avoid ambiguity we refer only to the complete architecture, composed of the main and auxiliary networks, as the \textit{hypernetwork}.} Formally, let $F(\mathbf{x};\mathbf{w}): X\times W \rightarrow Y$ denote the main network, where $\mathbf{x}$ is the input text to be classified, $\mathbf{w}$ are the parameters of the main network and $\mathbf{y}$ is the class label (capital letters represent the space of each variable). The auxiliary network is defined as $G(\mathbf{z};\theta): Z\times \Theta \rightarrow W$, where $\mathbf{z}$ is the auxiliary network input and $\theta$ are the parameters of the auxiliary network. Thus, the auxiliary network computes the weights of the main network, and provides a regularization mechanism over the main network's weights~\citep{ha2016hypernetworks}. The weight computation mechanism can be either independent of the input text during inference, resulting in a \textit{static} hypernetwork, or input-dependent (i.e. adaptive) during inference, resulting in a \textit{dynamic} hypernetwork.
Next, we first describe the details of the main network, and then present two hate speech detection solutions: (i) a static hypernetwork; and (ii) a dynamic hypernetwork.
\paragraph{The main network}
Our main network is a character-level CNN, which computes the posterior probability that the input text sequence is hateful. This architecture is inspired by the CNNs proposed in~\citep{zhang2015character}, however, our model is more shallow, comprising of two convolutional layers, compared to six in their work.\\
The text at the network input is represented using an alphabet of 69 characters, comprised of lower-case English letters, digits and other characters, as detailed in Table \ref{tab:alphabet}. We conducted a preliminary study to find the optimal letter-casing setting, as the repeated usage of capital letters may be more sensitive, but enlarges the vocabulary composing the training data and prohibits generalization. Our study revealed that using a lower-case setting is preferable in terms of classification performance over both upper- and lower-case character representation in the embedding layer. Therefore, as a pre-processing step we convert the data to a lower-cased representation. Characters that are not in the alphabet are mapped to a vector representation of a designated \textit{unknown} character. We denote the alphabet of characters as $V_{c}$. An embedding layer is utilized to map each one of the characters in $V_{c}$ to a lower-dimensional vector. \textcolor{black}{We set a limit of 120 characters to the input sequences. The length of the sequences varies between 80 to 150 characters, while approximately 75\% of the sequences are 120 characters or shorter. We found that a 120 character limit was sufficient for classification, as there was no clear advantage for using longer inputs. We applied padding and truncated sequences that were shorter or longer than this length limit, respectively. Finally, we evaluated several possible embedding dimensions, and chose a dimension of $d_{c} = 50$, which provided the best results. }
\begin{table}
\caption{The characters comprising the alphabet used in the proposed character-based HyperNetworks.}
\small
\setlength\tabcolsep{0pt}
\label{tab:alphabet}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}} ccccccccccccccc}
\hline \hline
a & b & c & d & e & f & g & h & i & j & k & l & m & n & o \\\hline
p & q & r & s & t & u & v & w & x & y & z & 0 & 1 & 2 & 3 \\\hline
4 & 5 & 7 & 8 & 9 & - & , & ; & . & ! & ? & : & ' & " & / \\\hline
\ {|} & {\_} & {@} & {\#} & {\$} & {\%} & {\textasciicircum} & {\&} & {*} & {\textasciitilde} & {`} & {+} & {-} & {=} & {\textless}\\\hline
\textgreater & ( & ) & [ & ] & \{ & \} & \textbackslash\\\hline\hline
\end{tabular*}
\end{table}
The character embedding layer is followed by two convolutional blocks. Each block is comprised of a 1-dimensional convolutional layer with 64 filters and kernel width of 7 with ReLU activation, and a 1-dimensional max pooling layer with a pool size of 4. Next, two consecutive fully-connected (FC) layers are applied with 128 and 32 perceptrons, respectively. A dropout layer is applied to the output of each FC layer, with a dropout probability of 0.5. Finally, an FC layer with a single perceptron and Sigmoid activation is applied to produce the posterior probability that the input text is hateful. We experimented with several hyper-parameters settings, including the number of convolutional layers, and selected the above set of parameter values, which provided the best performance. The complete main network, termed \textit{CharCNN}, is depicted in the lower part of Figure \ref{fig:HyperHate}.
\begin{figure*}[ht]
\centering
\includegraphics[trim={1.75cm 10cm 0cm 2cm},clip,scale=0.5]{DynamicHyperNN.png}
\caption{The proposed HyperHate-Dynamic architecture: The main network (lower part) computes the posterior probability that an input text sequence is hateful. The weights of the two 1D-convolutional layers are generated by the auxiliary network (upper part). The figure shows for illustration purposes a separate auxiliary network per convolutional layer, however, only a single auxiliary network is implemented and the weights for each convolutional layer are determined by the corresponding context vector computed by a bi-directional GRU (Bi-GRU) layer. The auxiliary network utilizes the context vector for generating input-adaptive weights for the corresponding convolutional layer, according to (\ref{eqref:static}) - (\ref{eq:context_vec}).}
\label{fig:DynamicCNN}
\end{figure*}
\paragraph{HyperHate-Static} Our first hate speech detection solution utilizes CharCNN as the main network $F(\mathbf{x};\mathbf{w})$, and a light-weight auxiliary network $G(\mathbf{z};\theta)$ for generating the weights of the convolutional layers in the main network. The input to the auxiliary network $G$ is an embedding vector $\mathbf{z}_{j} \in \mathbb{R}^{Z}$, which is learned during network training and encodes the information of the $j$-\textit{th} convolutional layer of the main network. In our experiments, a layer embedding vector of dimension $Z=10$ was found to provide good results. Inspired by the static architecture introduced in Ha et al.~\citeyearpar{ha2016hypernetworks}, our auxiliary network $G$ consists of two non-linear layers. The first, denoted as $f_1$, performs a matrix-vector multiplication between the input embedding vector $\mathbf{z}_{j}$ and the learned matrix $\mathbf{W_{in}} \in \mathbb{R}^{(Z C_{in})\times Z}$, followed by a ReLU activation, where $C_{in}=64$ is the input depth into both convolutional layers of the main network.\footnote{For the first convolutional layer, the embedding layer output (of dimension 50) is padded by zeros to obtain a dimension of 64.} The $Z C_{in}$ dimensional vector $\mathbf{o}_1$ computed by $f_1$ is reshaped to a matrix $\mathbf{O}_1$ of dimensions $C_{in}\times Z$ and layer $f_2$ multiplies it with a learned matrix $\mathbf{W_{out}} \in \mathbb{R}^{Z\times (kC_{out})}$, where $k=7$ is the length of each 1D convolutional kernel and $C_{out}=64$ is the number of kernels (and consequently feature maps) in both convolutional layers of the main network. The output of layer $f_2$ is a matrix $\mathbf{O}_2 \in \mathbb{R}^{C_{in}\times (kC_{out})}$ after applying element-wise ReLU activation, yielding the required $C_{in}\times k\times C_{out}$ weights for each convolutional layer of the main network.
The auxiliary network workflow is described as follows:
\begin{subequations}
\label{eqref:static}
\begin{align}
&\mathbf{o}_1=f_1(\mathbf{z}_j)=ReLU(\mathbf{W}_{in}\mathbf{z}_j) \\
&\mathbf{O}_1 = reshape(\mathbf{o}_1, [C_{in}\times Z])\\
&\mathbf{O}_2=f_2(\mathbf{O}_1)=ReLU(\mathbf{O}_1\mathbf{W}_{out}).
\end{align}
\end{subequations}
The complete HyperHate-Static architecture is depicted in Figure \ref{fig:HyperHate}, where the auxiliary network is illustrated in the upper part of the figure: note that the figure shows for illustration purposes a separate auxiliary network per convolutional layer, however, only a single instance of the auxiliary network is implemented and the weights for each convolutional layer are determined by the unique layer embedding vector $\mathbf{z}_j$ for $j=1,2$. The total number of coefficients of the HyperHate-Static network is 76K, as detailed in Table \ref{tab:params}.
\iffalse
In our experiments using HyperHate-Static, we use a layer embedding vector of dimension $Z=10$ and a weight generating hidden layer $f_1$ with 10 perceptrons. The rest of the model hyper-parameters are identical to CharCNN.
\fi
\iffalse The Static setting, shown in Figure~\ref{fig:HyperHate}, is similar to the static Hypernetwork described in~\citep{ha2016hypernetworks}. A network composed of two FC layers is utilized to generate parameters for CharCNN's convolutional layers. In order to produce parameters, a layer embedding vector, representing a convolutional layer within CharCNN, is fed to the generating network and projected to a \latin{$N_{dim}$} dimension, where \latin{$N_{dim}$} is the number of convolutional channels. Then, the next layer in the generating network projects the intermediate representation into the number of parameters required for the convolutional layer. We refer to this model as \textbf{HyperHate}.\fi
\paragraph{HyperHate-Dynamic}
\iffalse In addition, we propose a dynamic setting intended to produce parameters according to layer's input.
In this setting, each input sample {\it $x_i = [x^{1}_1, x^{1}_2,...,x^{1}_L]$} ({\it L} being the maximum sequence length) is fed into a parameter generating network, producing parameters that are better suited for that input. We refer to this setting as \textbf{HyperHate-Dynamic} (Figure~\ref{fig:DynamicCNN}). \fi
\begin{table}[t]
\caption{Size and input type of the evaluated hate speech detectors.}
\small
\centering
\label{tab:detectors}
\setlength\tabcolsep{0pt}
\begin{tabular*}{0.6\columnwidth}{@{\extracolsep{\fill}} lrl}
Model & Params & Embedding Granularity\\\hline
{CharBERT\textsubscript{$BERT$}} & {130M} & {Subword and Character}\\
{RoBERTa\textsubscript{$base$}} & {125M} & {WordPiece}\\
{BERT\textsubscript{$base-uncased$}} & {110M} & {WordPiece}\\
{MobileBERT} & {24M} & {WordPiece}\\
{ALBERT\textsubscript{$large-v2$}} & {18M} & {SentencePiece}\\
{CNN-GRU} & {2.5M} & {Word}\\
{HyperHate-Dynamic} & {129K} & {Character}\\
{HyperHate-Static} & {76K} & {Character}\\
\hline
\end{tabular*}
\end{table}
While the HyperHate-Static model has an exceptionally low number of coefficients, it lacks the capability of adapting the main network coefficients to each different input text sequence, which can facilitate better generalization performance. Therefore, our second solution is a dynamic HyperNetwork, in which the auxiliary network has a similar architecture as in the HyperHate-Static network, excluding the replacement of the static embedding vector $\mathbf{z}_j$ by a dynamic context vector $\mathbf{h}_j$, computed online per each input text sequence and convolutional layer $j=1,2$. The context vector is computed as follows: the input sequence of vectors to the \textit{j}-th convolutional layer of the main network {\it $X_j = [x^{j}_1, x^{j}_2,...,x^{j}_L]$} , where {\it L} is the maximal sequence length, is processed by a bidirectional GRU (Bi-GRU) layer, resulting in two final hidden states, one per direction:
\begin{subequations}
\label{eq:dynmaic}
\begin{align}
\overrightarrow{\mathbf{h}_{out}} &= \overrightarrow{GRU}(X_{j})\\
\overleftarrow{\mathbf{h}_{out}} &= \overleftarrow{GRU}(X_{j}),
\end{align}
\end{subequations}
and the context vector is given by:
\begin{equation}
\label{eq:context_vec}
{h}_j =
\begin{bmatrix}
\overleftarrow{\mathbf{h}_{out}}\\
\overrightarrow{\mathbf{h}_{out}}
\end{bmatrix}.
\end{equation}
The context vector $\mathbf{h}_j$ is utilized as the input to $f_1$, replacing $\mathbf{z}_j$ in Equation \ref{eqref:static}(a), for dynamically generating the main network weights. The output states of the Bi-GRU consist of 32 units per GRU direction, resulting in a context vector of dimension 64. We applied a recurrent dropout with a drop probability of 0.1, and initialized the GRU weights using a uniform glorot~\citep{glorot2010understanding} initialization.
The complete HyperHate-Dynamic architecture is depicted in Figure~\ref{fig:DynamicCNN}, where the auxiliary network is illustrated in the upper part of the figure: note that the figure shows for illustration purposes a separate auxiliary network per convolutional layer, however, only a single instance of the auxiliary network is implemented and the weights for each convolutional layer are determined by the dynamic context vector $\mathbf{h}_j$ for $j=1,2$. The total number of coefficients of the HyperHate-Dynamic network is 129K, as detailed in Table \ref{tab:params}.
\iffalse Note that the main advantage of utilizing HyperNetworks is that the number of parameters it requires for generating weights is independent of the number of convolutional layers composing the main network, whereas a standard convolutional layer relies on a unique set of parameters for each convolutional layer.\\
\begin{table}[ht]
\caption{The number of parameters required to produce weights for a single Conv. layer in our character-based models. The Conv. layers use a $k$ width filter, and operate on $C_{in}$ and $C_{out}$ channels. The HyperHate-Static model obtain weights for Conv. layer $j$ by feeding an embedding vector $z^{j} \in \mathbb{R}^{Z}$ to a HyperNetwork comprised of two fully-connected (FC) layers. HyperHate-Dynamic utilize a bidirectional-GRU to compress the input into a context-vector, which is then received by a HyperNetwork, creating input-conditioned weights.}
\small
\centering
\label{tab:params}
\setlength\tabcolsep{0pt}
\begin{tabular*}{0.8\columnwidth}{@{\extracolsep{\fill}} lcc}
\toprule
Model & Single Conv. Layer Params & Tot. Conv. Layers Params\\\hline\hline
{CharCNN} & {$k \times C_{in} \times C_{out}$} & {$2 \times 28,672$}\\
{HyperHate-Static} & {$Z \times (Z \times C_{in}) \times (k \times C_{out})$} & {10,880}\\
{HyperHate-Dynamic} & {$W_{GRU}+2 \times Z \times (Z \times C_{in}) \times (k \times C_{out}) $}\ & {91,584}\\
\hline
\hline
\end{tabular*}
\end{table}
\fi
\begin{table*}[ht]
\centering
\caption{The number of parameters in the proposed HyperHate-Static and HyperHate-Dynamic architectures}
\resizebox{0.7\textwidth}{!}{
\small
\centering
\begin{tabular}{p{1.3cm}|p{1.5cm}|p{3cm}|p{4cm}}
\hfil Network & \hfil Layer & \hfil HyperHate-Static & \hfil HyperHate-Dynamic\\
\hline\hline
{\vfill Auxiliary } & {\vfill \hfil All} & {\hspace*{0.1cm}$Z +Z\times (Z\times C_{in})$ \newline \hspace*{0.35cm}$+Z \times (k\times C_{out})$ \newline \hspace*{1.0cm}(10,954)} & {$W_{GRU}+2\times Z\times (Z\times C_{in})$ \newline \hspace*{1cm} $+Z\times (k\times C_{out})$ \newline \hspace*{1.5cm}(64,320)}\\\hline
{\vfill \hfil Main} & {Embedding} & {\hspace*{1.0cm}$V_{c}\times d_{c}$ \newline \hspace*{1.0cm}(3,500)} & {\hspace*{1.5cm}$V_{c}\times d_{c}$ \newline \hspace*{1.5cm}(3,500)}\\\hline
{\vfill \hfil Main} & {\vfill \hfil Conv. 1} & \hfil Generated by the\newline \hspace*{0.15cm} Auxiliary Network & \hfil Generated by the\newline \hspace*{0.6cm} Auxiliary Network \\\hline
{\vfill \hfil Main} & {\vfill \hfil Conv. 2} & \hfil Generated by the\newline \hspace*{0.15cm} Auxiliary Network & \hfil Generated by the\newline \hspace*{0.7cm}Auxiliary Network\\\hline
{\vfill \hfil Main} & {\vfill \hfil FC 1} & {\hspace*{0.8cm}128 units \newline \hspace*{0.8cm}(57,472)} & {\hspace*{1.4cm}128 units \newline \hspace*{1.5cm}(57,472)}\\\hline
{\vfill \hfil Main} & {\vfill \hfil FC 2} & {\hspace*{1.0cm}32 units \newline \hspace*{1.0cm}(4,128)} & {\hspace*{1.4cm}32 units \newline \hspace*{1.5cm}(4,128)}\\\hline
{\vfill \hfil Main} & {\vfill \hfil FC 3} & {\hspace*{1.0cm}1 unit \newline \hspace*{1.1cm}(33)} & {\hspace*{1.5cm}1 unit \newline \hspace*{1.6cm}(33)}\\\hline
& {\hfil Total} & {\hspace*{0.9cm}(76,087)} & {\hspace*{1.4cm}(129,453)}\\
\hline
\end{tabular}
}
\label{tab:params}
\end{table*}
\subsection{Assessing the Impact of Training Data Augmentation}
\label{sec:gen}
In this work, we assess the performance of the proposed solutions using varying amounts of training examples. The publicly available hate speech datasets are relatively small and imbalanced. As can be observed in Table~\ref{tab:datasets}, these gold-labeled (i.e., manually annotated) datasets include only thousands of examples, where a small minority of the text sequences are hate speech examples. Such a modest-sized selection can hardly represent the rich diversity of hate speech, leading to a large generalization gap ~\citep{wullach2020towards}. Aiming at state-of-the-art text classification performance, and a high level of generalization, we opt for generating a very large number of synthetic hate- and non-hate speech text sequences for training purposes. We follow the data generation workflow proposed in~\citep{wullach2020towards}, which is based on fine-tuning the GPT-2~\citep{radford2019language} deep generative language model as described below. Compared with this previous work, we upscale the example generation effort, reaching \iffalse utilized in our study a very large-scale dataset~\citep{BANKO2001}, with\fi a total of 10 million synthetic examples, balanced over the two classes of hate and non-hate. The generated sequences are comprised of five sub-datasets of 2 million examples each, corresponding to the data distributions of the five hate speech datasets in Table~\ref{tab:datasets}. Thus, the dataset is diverse and allows to improve learning generalization.\footnote{The hate and non-hate generated sequences are available to the research community upon request from the authors.}
Concretely, the first step of the generation workflow, illustrated in Figure~\ref{fig:GPT_2_WF}, is fine-tuning two GPT-2 generative models using the training set of each dataset in Table~\ref{tab:datasets}--one using hate speech class examples and another using the non-hate class examples, resulting with a total of 10 fine-tuned GPT-2 models. We fine-tuned a separate GPT-2 model for each dataset and class (hate/non-hate) in order to learn the unique data distribution characterized by each human-labeled dataset and class. The second step of the data generation included sampling of millions of text samples from the 10 fine-tuned models. In order to ensure that the hate speech generated sequences indeed belong to each hate speech class examples, we fine-tuned BERT (base and uncased version) to classify each generated hate speech sequences as hate or non-hate. Post-filtering, a total of 10M sequences comprised our generated corpus, corresponding to 2M per dataset: 1M per each of the hate and non-hate classes.
\begin{figure*}[t]
\centerline{
\includegraphics[trim={0cm 0.0cm 0cm 0cm},clip,scale=0.8]{GPT_2_WF_1M.png}}
\caption{The data generation workflow, adapted from~\citep{wullach2020towards}. Each dataset in Table~\ref{tab:datasets} is used to generate 2M synthetic sequences, and a total of 10M sequences using the five available datasets.}
\label{fig:GPT_2_WF}
\end{figure*}
Importantly, we note that each one of the datasets in Table~\ref{tab:datasets} was randomly split into disjoint training (80\%) and test (20\%) sets. \iffalse as follows: each dataset was randomly split to a training set (80\%) and a held-out testing set (20\%).\fi The data generation processes and all the training procedures described in Section~\ref{sec:results} were performed using the training sets, whereas hate detection performance was evaluated exclusively using the corresponding held-out testing sets.
\section{Performance Evaluation}
\label{sec:results}
This Section provides detailed performance evaluation of the proposed character-based HyperNetwork solutions compared to state-of-the-art deep learning architectures of hate detection, including the word-based CNN-GRU hate speech detector by Zhang et al.~\citeyearpar{zhang2018detecting}, and several pretrained Transformer-based models, namely BERT, RoBERTa and ALBERT. Concretely, we experiment with the large variant of the ALBERT model, which was shown to yield comparable results to BERT-base, but is much smaller, including only 18M parameters versus 110M. Both BERT and ALBERT were pretrained using the same data--the BookCorpus and the whole of English Wikipedia, while RoBERTa has been trained using additional data. In addition, we experiment with the newly introduced model of CharBERT, which incorporates character representations into BERT. We train our proposed character-based solutions and the CNN-GRU, and similarly fine-tune the pretrained models, using the available labeled training examples, where we test and compare the performance of the different methods on the held-out labeled examples. We further evaluate learning performance using increasing amounts of synthetic examples, augmenting the labeled data with up to 2M additional generated examples per dataset. While all the models may benefit from this procedure~\citep{wullach2020towards}, we expect the extended training to improve significantly the character-based models by exposing them to relevant language statistics.
Placing emphasis on learning generalization, our analysis includes two types of experiments, which we refer to as \textit{intra-domain} and \textit{cross-domain} hate speech detection. In the {\it intra-domain} experiments we follow the common practice of applying the DL detectors to each dataset independently: the performance of each DL detector is measured using the held-out gold-labeled test examples of every dataset $D_i$, having trained the detector with either the baseline training set, or with the augmented training set available for $D_i$, across the range of 0-2M generated sequences per dataset. In addition, we formed a \textit{combined} dataset comprised of the union of all datasets (having joined all training and test sets, respectively). The {\it cross-domain} experiments aim to evaluate the realistic condition of data distribution shift, where the hate detector is applied to text sequences sampled from a data distribution that is different from the one used for training~\citep{wiegand2019detection,wullach2020towards,wullachEMNLP21}. In these experiments, having the detector trained using the training set of dataset $\mathcal{D}_s$, it is evaluated on the held-out test set of dataset $\mathcal{D}_t\neq \mathcal{D}_s$. Again, we trained each DL detector using either the baseline or the augmented training set available for dataset $D_s$, across the range of 0-2M generated sequences per dataset.
We report classification performance with respect to the hate class in term of Recall, Precision and \textit{F}1 scores. Recall corresponds to the proportion of true hate speech examples that were automatically identified as hate speech, and Precision is the proportion of correct predictions within the examples identified as hate speech by the detector. \textit{F}1 is the harmonic mean of those measures, assigning them equal importance. Let us note that assuming that only a small proportion of the data is automatically identified as hate speech, false positive predictions may be tracked relatively easily by means of further human inspection, whereas it is impossible to track false negatives at scale. Therefore, increasing recall by means of improved generalization is of great importance in practice. \\
\subsection{Implementation Details}
We implemented all the DL models in TensorFlow~\citep{tensorflow2015-whitepaper} and utilized the NVIDIA K-80 GPU for training and testing. In training the proposed HyperHate models, we applied the Adam optimizer with a mini-batch size of 32 samples and an early stopping mechanism, minimizing the binary cross-entropy loss. Additional implementation choices of the proposed HyperNetwork models are discussed in Section~\ref{sec:approach}.\footnote{We make our code available at https://github.com/tomerwul/CharLevelHyperNetworks.}
The CNN-GRU detector was trained using the Adam optimizer~\citep{Goodfellow-et-al-2016}, minimizing the binary cross-entropy loss, with early stopping and mini-batch size of 32. The 1D-Convolutional layer includes 100 filters with kernel size of 4, the 1D-MaxPooling layer with pool size of 4 and the GRU output dimension of 100.
Our implementation of BERT and the other pretrained models makes use of the popular HuggingFace Transformers repository.\footnote{https://github.com/huggingface/transformers} \iffalse which provides a variety of pre-trained transformer-based models.\fi In all cases, we fine-tuned the pretrained model in its uncased setting, with an additional feed-forward layer on top of the final output of the [CLS] token, which yields a probability distribution for a binary classification task. Following a validation step, we fine-tuned the models for 3 epochs with a mini-batch size of 32, and used Adam optimizer with an initial learning rate of 2e-5, performing 1000 warm up steps prior to reaching the initial learning rate. We set a fixed sentence length of 30 tokens, padding input sentences that are shorter and truncating longer sentences. An attention mask was employed to avoid including the padded tokens in the self-attention calculations.
\iffalse Gururangan et al.~\citep{gururACL2020} suggests sampling examples that are similar to the classification training data from a task-related unlabeled corpus, and adapt the model using those examples. Here, we use generated labeled examples, which are artificial and noisy to a certain extent, for extending the gold-labeled, yet biased and small, existing datasets of hate speech. We evaluate the impact of train set augmentation in fine-tuning BERT with increasing amounts of our generated data. \fi
Our implementation specifications of ALBERT are similar to those utilized for fine-tuning BERT, with the exception of using a text tokenizer that is suited for the ALBERT model.\\
\begin{figure*}[p]
\centerline{
\includegraphics[width=24cm]{intra_curves_a_charbert.png}}
\caption{Intra-dataset results (part 1/2): hate-class F1, using increasing amounts of 0-2M generated sequences added as data augmentation.}
\label{fig:intra-Curves}
\end{figure*}
\begin{figure*}[p]
\centerline{
\includegraphics[trim={0cm 0cm 0cm 0cm},clip,scale=1.4]{intra_curves_b_charbert.png}}
\caption{Intra-dataset results (part 2/2): hate-class F1, using increasing amounts of 0-2M generated sequences added as data augmentation.}
\label{fig:intra-Curves_b}
\end{figure*}
\iffalse
\begin{figure*}[t]
\centerline{
\includegraphics[trim={1cm 0.5cm 1.25cm 0.15cm},clip,scale=1.2]{cross_curves_4on1.png}}
\caption{Combined cross-domain results. Plots (a)-(d) show the results of training the detectors on a combined dataset consists of four hate speech datasets (augmented with the corresponding generated sequences) and evaluated on a test set taken from a fifth dataset. }
\label{fig:4on1-Curves}
\end{figure*}
\fi
\begin{figure*}[p]
\centerline{
\includegraphics[trim={1cm 0.0cm 1.25cm 0cm},clip,scale=1.4]{cross_curves_a_charbert.png}}
\caption{Cross-dataset results (part 1/2): hate-class F1, using increasing amounts of 0-2M generated sequences added as data augmentation.
The detectors were trained on the left dataset, and tested on the held-out examples of the right dataset, of each dataset pair. The best {\it F}1 result per dataset pair are highlighted in boldface.}
\label{fig:cross-Curves}
\end{figure*}
\begin{figure*}[p]
\centerline{
\includegraphics[trim={1cm 0.0cm 1.25cm 0cm},clip,scale=1.4]{cross_curves_b_charbert.png}}
\caption{Cross-dataset results (part 2/2): hate-class F1, using increasing amounts of 0-2M generated sequences added as data augmentation.
The detectors were trained on the left dataset, and tested on the held-out examples of the right dataset, of each dataset pair. The best {\it F}1 result per dataset pair are highlighted in boldface.}
\label{fig:cross-Curves_b}
\end{figure*}
\subsection{Intra-domain Experiments: Results}
Figures~\ref{fig:intra-Curves} and~\ref{fig:intra-Curves_b} report the results of the intra-domain experiments in terms of $F1$ scores, computed with respect to the {\it hate} class. The two figures present the results for the various methods, trained and tested using each of the five source datasets and their union, denoted as the \textit{combined} dataset. The classifiers' performance using no (0) augmentation corresponds to training using merely the gold-labeled training portion (80\%) of the source datasets. The figures show the results of augmenting this gold-labeled data with increasing amounts of up to 2M class-balanced generated examples.
\textcolor{black}{Several trends are observed in the results. First, we observe that the larger token-based models, including BERT, RoBERTa, and ALBERT, as well as CNN-GRU, dominate the proposed character-based HyperNetwork models when the amount of labeled examples is small (i.e., with no data augmentation). Indeed, it has been previously shown that the models that are pretrained using large amounts of text are advantageous given limited labeled examples~\citep{wullachEMNLP21}. It is further observed that, for most of the models and datasets, a boost in the performance is achieved once as few as 5-10K generated examples are added to the gold-labeled data. As the original datasets contain only a few thousands of gold-labeled examples, adding a similar amount of synthetic data increases data diversity, roughly doubling the train set size, where this has a large impact on learning performance.}
\textcolor{black}{We further observe in Figures~\ref{fig:intra-Curves} and~\ref{fig:intra-Curves_b} that the respective $F1$ Performance levels of the token-based models typically continues to rise slowly or stabilizes as more generated labeled examples are provided. In comparison, the HyperNetwork models that are inferior initially leverage the additional generated training data, where the peak performance using these models is obtained in the range of 1-2 million generated examples. That is, large amounts of data is required to train these models, which were not pretrained, to identify hate- and non-hate related character sequences. Crucially, the figures show that the gap in performance between the large token-based and compact character-based models narrows down with increased levels of data augmentation.}
Detailed results including precision, recall and \textit{F}1 scores using full vs. no augmentation per method and dataset are given in Table~\ref{tab:intra_dataset_results} (Appendix~\ref{sec:appendix_results}). \textcolor{black}{The best results in terms of \textit{F}1 are highlighted in boldface per dataset in the table.}
As detailed in Table~\ref{tab:intra_dataset_results}, all of the methods benefit from data augmentation by 2M generated sequences. Mainly, we observe significant and consistent gains in recall with data augmentation for all methods. The recall improvements are more pronounces for the HyperNetwork models, peaking at +153.58$\%$ for the HyperHate-Dynamic model and StormFront (WH) dataset. Notably, the introduction of large amounts of synthetic text sequences may introduce a drift in the data distribution, or noise. As shown, in some cases the precision scores are negative, with a minimum of -31.54$\%$ for the ALBERT model and Founta (FN) dataset.
Nevertheless, precision declines to a lesser degree compared to the high gains in recall. And, in some cases, both precision and recall benefit from data augmentation (e.g., see the results for most methods on the WS dataset). Overall, hate detection improves with data augmentation, as reflected by the \textit{F}1 scores, which are overall positive. We note that while \textit{F}1 performances rise between $\sim$0-20\% for the larger models and the individual datasets, the respective improvements for the character-based HyperNetworks are significantly higher, ranging between $\sim$14-70\%.
\textcolor{black}{Overall, as shown in Table~\ref{tab:intra_dataset_results}, following data augmentation, the HyperNetwork architectures achieve preferable or competitive results compared with some of the large networks following data augmentation, e.g., achieving best results along with CNN-GRU on the Davidson (DV) test set; outperforming CNN-GRU, BERT, CharBERT and MobileBERT on StormFront (WH), CNN-GRU and ALBERT on the Founta (FN) dataset; and, the model of ALBERT on the Waseem (WS) dataset. Also, these models achieve comparable performance to the large pretrained token-based models, while slightly outperforming ALBERT, MobileBERT and CNN-GRU, on the combined dataset. This result has striking implications considering the difference in size of these types of models.}
\subsection{Cross-domain Experiments: Results}
The results of the cross-domain experiments are displayed in Figures~\ref{fig:cross-Curves} and~\ref{fig:cross-Curves_b} for eight representative source-target dataset pairs in terms of \textit{F}1 scores. In the cross-domain setup, the models are trained using the gold-labeled train set of a source dataset $\mathcal{D}_s$, being augmented with synthetic examples generated from the same train data distribution of $\mathcal{D}_s$, and are then evaluated over the held-out test set of a target dataset $\mathcal{D}_t$. The figures show the results of augmenting the gold-labeled training data with increasing amounts of 0-2M class-balanced generated examples.
The cross-domain setup is more challenging learning-wise and is more realistic than the intra-domain setup as it introduces a shift in data distribution. Accordingly, the overall results are dramatically lower compared with the intra-dataset experiments. The extent of drop in performance varies across the dataset pairs, depending on the gap in their characteristics.
Overall, we observe consistent significant performance gains with data augmentation for all methods. In particular, there is a steep rise of recall in most of the experiments. In terms of precision scores, there are substantial gains in some cases due to data augmentation (peaking at +332.25$\%$ using HyperNetwork-static, and +229.9$\%$ using ALBERT), where in a few cases precision decreased following augmentation (reaching a minimum of -48.21$\%$ in one of the experiments, namely the ALBERT model and WS-FN pair; yet, also in that case the overall {\it F}1 performance improved due to the larger boost in recall). The boost in recall and overall {\it F}1 performance is especially prominent for the HyperNetwork models, e.g., reaching up to +2925$\%$ improvement in recall and +1477$\%$ in {\it F}1 for the HyperHate-Static model and DV-WS pair.
Detailed results, including precision, recall and \textit{F}1 scores are given in Table~\ref{tab:cross_dataset_results} (Appendix~\ref{sec:appendix_results}), comparing the performance per dataset and method with and without data augmentation by 2M sequences. The best hate detection performance across methods is highlighted in boldface in the table for each dataset pair. \textcolor{black}{It is an overwhelming result that the compact HyperNetwork models reach better performance compared to the larger models in some cases following data augmentation. Specifically, the dynamic HyperNetwork variant yields the best performance score for the DV-SE and the SE-FN pairs (on par with ALBERT, for the latter pair). The Static HyperNetwork model provides the best performance for the FN-SE and WH-FN pairs and is second-best to RoBERTa on the WH-SE pair. In general, it is observed in the table that while all models improve from data augmentation, both of the HyperNetwork configurations gain more in performance following data augmentation in almost all of the experiments as compared to the larger token-based models, both in terms of recall and precision. }
\textcolor{black}{Overall, the static and dynamic HyperNetwork variants perform comparably. The best choice of model may depend on the characteristics of the data, and can be determined using validation experiments, following the common practice in machine learning.}
To conclude, the reported results demonstrate strong performance of the proposed character-level HyperNetwork solutions on hate detection across data distributions. Presumably, these compact models are less expressive compared to large pretrained language models such as BERT, RoBERTa and CharBERT. \textcolor{black}{Yet, they achieved superior or competitive performance in a substantial portion of our cross-dataset experiments. We conjecture that character-level processing may be less overfitting to the language statistics observed in training, being more adaptive to word variants and inflections which vary across datasets.}
\iffalse In terms of recall scores, we observe consistently significant gains with data augmentation for all methods, peaking at +2925.0$\%$ for the HyperHate-Static model and DV-WS pair. In terms of precision scores, the gains due to data augmentation are typically lower, peaking at +332.25$\%$ and in few cases even negative with a minimum of -48.21$\%$ for the ALBERT model and WS-FN pair.\fi
\section{Conclusions}
In this paper, we proposed light-weight network architectures that make use of efficient mechanisms of weight sharing, namely HyperNetworks, as the backbone of character-based models trained to detect hate speech. Concretely, we described and experimented with a static auxiliary network where the generated weights are fixed during inference, and proposed a novel dynamic auxiliary network that produces input-conditioned weights.
We evaluated hate detection performance using gold-labeled examples of five public datasets, both using intra- and cross-dataset settings, while performing data augmentation using varying amounts of generated synthetic data. \textcolor{black}{Alongside the suggested methods, we reported hate detection results using BERT and RoBERTa, popular pretrained transfomer-based models which yield high performance of text classification but are large and computationally expensive. We further evaluated CharBERT, a variant of BERT that incorporates character embeddings, and CNN-GRU, a well-performing token-based hate detection model.} The results indicate that the proposed methods, while producing inferior results at first when no augmentation is added, can effectively leverage additional synthetic data and achieve competitive results compared to these state-of-the-art models. Furthermore, we obtained superior results using the proposed architectures in \textcolor{black}{some cases in} the challenging cross-dataset setup, simulating the data shift that typically occurs in practice across domains and over time. Thus, the proposed character-level HyperNetwork solutions present a combination of high performance as well as compactness, being smaller by orders of magnitude compared with state-of-the-art architectures, making them suitable for hate detection in general, and applicable for low-memory end devices in particular.
\textcolor{black}{The proposed character-based architectures are language independent. We believe that these methods will prove particularly advantageous in morphology-rich languages, where character sequences can model the variation that is observed in social media text, as well as related word forms. This would require relevant resources of labeled data, and mechanisms of generating additional high-quality relevant texts in the target language~\citep{deVriesACL21,fengACL21}. Another interesting direction to explore is the use of data that is automatically labeled using semi-supervised classification~\citep{sarkarEMNLP21}, as alternative or in addition to synthetically generated data, in training the character-based hate detection models.}
\section*{Acknowledgement}
This research was supported partly by Facebook Content Policy Research on Social Media
Platforms Research Award.
\bibliographystyle{model5-names}
|
\section{Introduction}
\label{sec1}
Autonomous driving has received significant research interests in the past two decades due to its many potential societal and economical benefits. Compared to traditional vehicles, autonomous vehicles (AVs) not only promise fewer emissions \cite{Paden} but are also expected to improve safety and efficiency. Despite remarkable progress, high-level decision-making in AVs remains a big challenge due to the complex and dynamic traffic environment, especially in mixed traffic co-existing with other road users. Lane changing is one such challenging high-level decision-making in AVs, which has significant influences on traffic safety and efficiency \cite{desiraju2014minimizing, li2020cooperative}, and is the focus of this paper.
The considered lane-changing scenario is illustrated in Fig.~\ref{lane-changing scene}, where AVs and HDVs coexist on a one-way highway with two lanes. The AVs aim to safely travel through the traffic while making necessary lane changes to overtake slow-moving vehicles for improved efficiency. Furthermore, in the presence of multiple AVs, the AVs are expected to collaboratively learn a policy to adapt to HDVs and enable safe and efficient lane changes. As HDVs bring unknown/uncertain behaviors, planning and control in such mixed traffic to realize safe and efficient maneuvers is a challenging task \cite{chen2020autonomous}.
\begin{figure}[!ht]
\centering
\includegraphics[ width=0.7\textwidth]{imgs/Fig1.pdf}
\caption{Illustration of the considered lane-changing scenario (green: AVs, blue: HDVs, arrow curve: a possible trajectory of the ego vehicle AV1 to make the lane change).}
\label{lane-changing scene}
\end{figure}
Recently, reinforcement learning (RL) has emerged as a promising framework for autonomous driving due to its online adaptation capabilities and
the ability to solve complex problems \cite{wang2019continuous, xi2020efficient}.
Several recent studies have explored the use of RL in AV lane-changing \cite{wang2021harmonious, du2020cooperative, chen2020autonomous}, which consider a single AV setting where the ego vehicle learns a lane-changing behavior by taking all other vehicles as part of the driving environment for decision making. While completely scalable, this single-agent approach will lead to unsatisfactory performance in the complex environment like multi-AV lane-changing in mixed traffic that requires close collaboration and coordination among AVs \cite{Hoel}.
On the other hand, multi-agent reinforcement learning (MARL) has been greatly advanced and successfully applied to a variety of complex multi-agent systems such as games \cite{vinyals2019grandmaster}, traffic light control \cite{Chu} and fleet management \cite{lin2018efficient}. The applications of MARL to autonomous driving also exist \cite{shalev2016safe, wangmulti, ha2020leveraging, palanisamy2020multi}, with the objective of accomplishing autonomous driving tasks cooperatively and reacting timely to HDVs.
In particular, previous works \cite{chen2021graph, ha2020leveraging} also apply the MARL to highway lane change tasks and show promising and scalable performance, in which AVs learn cooperatively via sharing the same objective (i.e., reward/cost function) that considers safety and efficiency. However, those reward designs often ignore the passengers' comfort, which may lead to sudden acceleration and deceleration that can cause ride discomfort. In addition, they assume that the HDVs follow unchanged, universal human driver behaviors, which is clearly oversimplified and impractical in the real world as different human drivers tend to behave quite differently. Learning algorithms should thus work with different human driving behaviors, e.g., aggressive or mild.
To address the above issues, we develop a multi-agent reinforcement learning algorithm by employing a multi-agent advantage actor-critic network (MA2C) for multi-AV lane-changing decision making, featuring a novel local reward design that incorporates the safety, efficiency and passenger comfort as well as a parameter sharing scheme to foster inter-agent collaborations. The main contributions and the technical advancements of this paper are summarized as follows.
\begin{enumerate}
\item We formulate the multi-AV highway lane changing in mixed traffic is modeled as a decentralized cooperative MARL problem, where agents cooperatively learn a safe and efficient driving policy.
\item We develop a novel, efficient, and scalable MARL algorithm, multi-agent advantage actor-critic network, by introducing a parameter-sharing mechanism and effective reward function design.
\item We conduct comprehensive experiments on three different traffic densities and two levels of drivers' behavior modes, and the results show that the proposed approach consistently outperforms several state-of-the-art algorithms in terms of driving safety, efficiency and driver comfort.
\end{enumerate}
The rest of the paper is organized as follows. Section~\ref{section:RELATED WORK} reviews the state-of-the-art dynamics-based and RL/MARL algorithms for autonomous driving tasks. The preliminaries of RL and the proposed MARL algorithm are introduced in Section~\ref{section:PROPOSED SCHEMES}. Experiments, results, and discussions are presented in Section~\ref{section:NUMERICAL EXPERIMENTS AND DISCUSSION}. Finally, we summarize the paper and discuss future work in Section~\ref{section:Conclusion}.
\section{Related Work}
\label{section:RELATED WORK}
In this section, we survey the existing literature on decision-making tasks in autonomous driving, which can be mainly classified into two categories: non-data-driven and data-driven methods.
\subsection{Non-data-driven Methods}
Conventional rule-based or model-based approaches \cite{ho2009lane, nilsson2016if, nilsson2016lane} rely on hard-coded rules or dynamical models to construct predefined logic mechanisms to determine the behaviors of ego vehicles under different situations. For instance, in \cite{ho2009lane}, lane-changing guidance is provided by establishing virtual trajectory references for every vehicle, and a safe trajectory is then planned by considering the trajectories of other vehicles. In \cite{nilsson2016if}, a low-complexity lane-changing algorithm is developed by following heuristic rules such as keeping appropriate inter-vehicle traffic gaps and time instances to perform the maneuver. In addition, an optimization-based lane change approach is proposed in \cite{nilsson2016lane}, which formulates the trajectory planning problem as coupled longitudinal and lateral predictive control problems and is then solved via Quadratic Programs under specific system constraints. However, the rules and optimization criteria for real-world driving problems, especially in mixed-traffic scenarios with unknown and stochastic driver's behaviors, may become too complex to be explicitly formulated for all scenarios.
\subsection{Data-Driven Methods}
Recently, data-driven methods, such as reinforcement learning (RL), have received great attention and been widely explored for autonomous driving tasks. Particularly, a model-free RL approach based on deep deterministic policy gradient (DDPG) is proposed in \cite{wang2019continuous} to learn a continuous control policy efficient lane changing. In \cite{chen2020autonomous}, a safe RL framework is presented by integrating a lane-changing regret model into a safety supervisor based on an extended double deep Q-network (DDQN). In \cite{chen2019attention}, a hierarchical RL algorithm is developed to learn lane-changing behaviors in dense traffic by applying the designed temporal and spatial attention strategies, and promising performance are demonstrated in the TORCS simulator under various lane change scenarios. However, the aforementioned methods are designed for the single-agent (i.e., one ego vehicle) scenarios, treating all other vehicles as part of the environment, which makes them implausible for the considered multi-agent lane-changing setting where collaboration and coordination among AVs are required.
On the other hand, multi-agent reinforcement learning (MARL) has also been explored for autonomous driving tasks \cite{shalev2016safe, dong2020drl, wangmulti, chen2021deep}. In particular, in \cite{shalev2016safe}, a MARL algorithm with hard-coded safety constraints is proposed to solve the double-merge problem. Also, a hierarchical temporal abstraction method is applied to reduce the effective horizon and the variance of the gradient estimation error. In \cite{chen2021deep}, a MARL algorithm is proposed to solve the on-ramp merging problem with safety enhancement by a novel priority-based safety supervisor. In addition, the authors in \cite{dong2020drl} propose a novel MARL approach combining Graphic Convolution Neural Network (GCN) \cite{kipf2016semi} and Deep Q Network (DQN) \cite{mnih2013playing} to better fuse the acquired information from collaborative sensing, showing promising results on a 3-lane freeway containing 2 off-ramps highway environment. While these MARL algorithms only consider the efficiency and safety in their designed reward function, another important factor, the passenger comfort, is not considered in their reward function design. Furthermore, those approaches assume the HDVs follow a constant, universal driving behavior, which has limited implications for real-world applications as different human drivers may behave totally differently.
In this paper, we formulate the decision making of multiple AVs on highway lane changing as a MARL problem, where a multi-objective reward function is proposed to simultaneously promote safety, efficiency and passenger comfort. A parameter-sharing scheme is exploited to foster inter-agent collaborations. Experimental results on three different traffic densities with two levels of driver aggressiveness show the proposed MARL performs well on different lane change scenarios.
\section{Problem Formulation}
\label{section:PROPOSED SCHEMES}
In this section, we review the preliminaries of RL and formulate the considered highway lane-changing problem as a partially observable Markov decision process (POMDP). Then we present the proposed multi-agent actor-critic algorithm, featuring a parameter-sharing mechanism and efficient reward function design, to solve the formulated POMDP.
\subsection{Preliminary of RL}
In the standard RL setting, the agent aims to learn an optimal policy $\pi^*$ to maximize the accumulated future reward $R_t = \sum_{k=0}^{T} \gamma ^k r_{t+k}$ from the time step $t$ with discount factor $\gamma\in (0,1]$ by continuous interacting with the environment. Specially, at time step $t$, the agent receives a state $s_t \in \mathcal{R}^{n}$ from the environment and selects an action $a_t \in\mathcal{A}^{m}$ according to its policy $\pi:\,\mathcal{S}\rightarrow\Pr(\mathcal{A})$. As a result, the agent receives the next state $s_{t+1}$ and receives a scalar reward $r_t$. If the agent can only observe a part of the state $s_t$, the underlying dynamics becomes a POMDP \cite{spaan2012partially} and the goal is then to learn a policy that maps from the partial observation to an appropriate action to maximize the rewards.
The action-value function $Q^{\pi}(s, a) = E[R_t{\mid}{s=s_t}, a]$ is defined as the expected return obtained by selecting an action $a$ in state $s_t$ and following policy $\pi$ afterwards. The optimal Q-function is given by $Q^{*}(s,a) = \max_{\pi} Q^{\pi}(s,a)$ for state $s$ and action $a$. Similarly, the state-value function is defined as $V^{\pi}(s_t) = E_{\pi}{[R_t{\mid}{s=s_t}]}$ representing the expected return for following the policy $\pi$ from state $s_t$. In model-free RL methods, the policy is often represented by a neural network denoted as $\pi_{\theta}(a_t{\mid}s_t)$, where $\theta$ is the learnable parameters. In actor-critic (A2C) algorithms \cite{mnih2016asynchronous}, a critic network, parameterized by $\omega$, learns the state-value function $V_{\omega}^{\pi_\theta}(s_t)$ and an actor network $\pi_{\theta}(a_t{\mid}s_t)$ parameterized by $\theta$ is applied to update the policy distribution in the direction suggested by the critic network as follows:
\begin{equation}\label{eqn:advantagepolicygradient}
\theta \leftarrow \theta + E_{\pi_{\theta}} \left[ \Big (\nabla_{\theta} \log \pi_{\theta}(a_t{\mid}s_t) \Big)
A_t \right],
\end{equation}
where the advantage function $A_t= Q^{\pi_\theta}(s,a) - V_{\omega}^{\pi_\theta}(s_t)$ \cite{mnih2016asynchronous} is introduced to reduce the sample variance.
The parameters of the state-value function are then updated by minimizing the following loss function:
\begin{equation}\label{eqn:valueloss}
\min_{\omega} E_{\mathcal{B}}\Big (R_t + \gamma V_{\omega'} (s_{t+1}) - V_{\omega}(s_t)\Big )^2,
\end{equation}
where $\mathcal{B}$ is the experience replay buffer that stores previously encountered trajectories and $\omega'$ denotes the parameters of the target network \cite{mnih2013playing}.
\subsection{Lane Changing as MARL}
In this subsection, we develop a decentralized, MARL-based approach for highway lane-changing of multiple AVs. In particular, we model the mixed-traffic lane-changing environment as a multi-agent network: $\mathcal{G} = (\text{$\nu$}, \text{$\varepsilon$})$, where each agent (i.e., ego vehicle) $i \in \text{$\nu$}$ communicates with its neighbors $\mathcal{N}_i$ via the communication link $\varepsilon_{ij}\in\large\text{$\varepsilon$}$. The corresponding POMDP is characterized as $(\{\mathcal{A}_i, \mathcal{O}_i, \mathcal{R}_i\}_{i\subseteq \nu}, \mathcal{T})$, where $\mathcal{O}_i \in \mathcal{S}_i$ is the partial description of the environment state as stated in \cite{chu2020multi}. In a multi-agent POMDP, each agent $i$ follows a decentralized policy $\pi_i: \mathcal{O}_i \times \mathcal{S}_i \rightarrow [0, 1]$ to choose the action $a_t$ at time step $t$. The described POMDP is defined as:
\begin{enumerate}
\item \textit{State Space}:
The state space $\mathcal{O}_{i}$ of Agent $i$ is defined as a matrix $\mathcal{N}_{N_{i}}\times \mathcal{F}$, where $\mathcal{N}_{N_{i}}$ is the number of detected vehicles, and $\mathcal{F}$ is the number of features, which is used to represent the current state of vehicles. It includes the longitudinal position $x$, the lateral position $y$ of the observed vehicle relative to the ego vehicle, the longitudinal speed $v_x$, and the lateral speed $v_y$ of the observed vehicle relative to the ego vehicle.
\item \textit{Action Space}:
The action space $\mathcal{A}_i$ of agent $i$ is defined as a set of high-level control decisions, including speed up, slow down, cruising, turn left, and turn right. The action space combination for AVs is defined as $\mathcal{A}=\mathcal{A}_{1}\times \mathcal{A}_{2}\times \cdot\cdot\cdot\times \mathcal{A}_{N}$, where $N$ is the total number of vehicles in the scene.
\item \textit{Reward Function}:
Multiple metrics including safety, traffic efficiency and passenger's comfort are considered in the reward function design:
\begin{itemize}
\item[$\bullet$] safety evaluation $r_{s}$: The vehicle should operate without collisions.
\item[$\bullet$] headway evaluation $r_{d}$: The vehicle should maintain a safe distance from the preceding vehicles during driving to avoid collisions.
\item[$\bullet$] speed evaluation $r_{v}$: Under the premise of ensuring safety, the vehicle is expected to drive at a high and stable speed.
\item[$\bullet$] driving comfort $r_{c}$: Smooth acceleration and deceleration are expected to ensure safety and comfort. In addition, frequent lane changes should be avoided.
\end{itemize}
As such, a multi-objective reward $r_{i,t}$ at the time step $t$ is defined as:
\begin{equation}
\label{eqn:reward}
r_{i,t}=\omega_{s}r_{s}+\omega_{d}r_{d}+\omega_{v}r_{v}-\omega_{c}r_{c},
\end{equation}
where $\omega_{s}$, $\omega_{d}$, $\omega_{v}$ and $\omega_{c}$ are the weighting coefficients. We set the safety factor $\omega_{s}$ to a large value, because safety is the most important criterion during driving. The details of the four performance measurements are discussed next:
\begin{enumerate}[(1)]
\item If there is no collision, the collision evaluation $r_{s}$ is set to 0, otherwise, $r_{s}$ is set as -1.
\item The headway evaluation is defined as
\begin{equation}
\label{equation:rd}
r_{d}=\log\frac{d_{headway}}{v_{t}t_{d}},
\end{equation}
where $d_{headway}$ is the distance to the preceding vehicle, and $v_t$ and $t_d$ are the current vehicle speed and time headway threshold, respectively.
\item The speed evaluation $r_{v}$ is defined as
\begin{equation}
r_{v}=\min\left \{\frac{v_{t}-v_{min}}{v_{max}-v_{min}},1\right\},
\end{equation}
where $v_{t}$, $v_{min}$ and $v_{max}$ are the current, minimum, and maximum speeds of the ego vehicle, respectively. Within the specified speed range, higher speed is preferred to improve the driving efficiency.
\item The driving comfort $r_{c}$ is defined as
\begin{equation}
\label{eqn:comfort}
r_{c}=r_{a}+r_{lc},
\end{equation}
where $$r_{a}=\left\{
\begin{array}{rcl}
-1,&\vert a_{t}\vert \geq a_{th}\\
0,&\vert a_{t}\vert < a_{th}
\end{array}
\right.$$ is the penalty term of rapid acceleration and deceleration than a given threshold $a_{th}$. Here $a_{t}$ presents the acceleration at time $t$.
$$r_{lc}=\left\{
\begin{array}{rcl}
-1,&\rm\normalsize change&\rm\normalsize lane\\
0,&\rm\normalsize keep&\rm\normalsize lane
\end{array}
\right.$$ is defined as the lane change penalty. Excessive lane changes can cause discomfort and safety issues.
\end{enumerate}
\item \textit{Transition Probability}: The transition probability $T(s^{,}\mid s,a)$ characterizes the transition from one state to another. Since our MARL algorithm is a model-free design, we do not assume any prior knowledge about transition probability.
\end{enumerate}
\subsection{MA2C for AVs}
In this paper, we extend the actor-critic network \cite{mnih2016asynchronous} to the multi-agent setting as a multi-agent actor-critic network (i.e., MA2C). MA2C improves the stability and scalability of the learning process by allowing certain communication among agents \cite{chu2020multi}. To take the advantage of homogeneous agents in the considered MARL setting, we assume all the agents share the same network structure and parameters, while they are still able to make different maneuvers according to different input states.
The goal in cooperative MARL setting is to maximize the global reward of all the agents. To overcome the communication overhead and the credit assignment problem \cite{sutton2018reinforcement}, we adopt the local reward design \cite{chen2021deep} as follows:
\begin{equation}\label{eqn:local_reward}
r_{i, t} = \frac{1}{\mid{\nu_i}\mid} \sum_{j\in\nu_i} r_{j,t},
\end{equation}
where $\mid \nu_i \mid$ denotes the cardinality of a set containing the ego vehicle and its close neighbors. Compared to the global reward design previously used in \cite{kaushik2018parameter, dong2020drl}, the designed local reward design mitigates the impact of remote agents.
The backbone of the proposed MA2C network is shown in Fig.~\ref{Backbone of MA2C Network}, in which states separated by physical units are first processed by separate 64-neuron fully connected (FC) layers. Then all hidden units are combined and fed into the 128-neuron FC layer. Then the shared actor-critic network will update the policy and value networks with the extracted features. As mentioned in \cite{chen2021deep}, the adopted parameter sharing scheme \cite{lin2018efficient} between the actor and value networks can greatly improve the learning efficiency.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.92\textwidth]{imgs/Fig2.pdf}
\caption{The architecture of the proposed MA2C network with shared actor-critic network design, where $x$ and $y$ are the longitudinal and lateral position of the observed vehicle relative to the ego vehicle, and $v_x$ and $v_y$ are the longitudinal and lateral speed of the observed vehicle relative to the ego vehicle.}
\label{Backbone of MA2C Network}
\end{figure}
The pseudo-code of the proposed MA2C algorithm is shown in Algorithm \ref{algo:marl_algo}. The hyperparameters include: the (time)-discount factor $\gamma$, the learning rate $\eta$, the politeness coefficient $p$ and the epoch length $T$. Specifically, the agent receives the observation $O_{i,t}$ from the environment and updates the action by its policy (Line 3-6). After each episode is completed, the network parameters are updated accordingly (Line 9-11). If an episode is completed or a collision occurs, the ``DONE" signal is released and the environment will be reset to its initial state to start a new epoch (Line 13-14).
\begin{algorithm}[htb]
\caption{MARL for AVs.}
\label{algo:marl_algo}
\begin{algorithmic}[1]
\Require $\gamma, \eta, p, T$.
\Ensure $\theta$.
\State {\bf Initialize} $o_0, t \leftarrow 0$.
\Repeat
\For{$i \in V$}
\State Observe $o_{i,t}$;
\State Update $a_{i,t} \sim \pi_{\theta_{i,t}}$;
\EndFor
\State Update $t=t+1$;
\If{DONE}
\For{$i \in V$}
\State Update $\theta_i \leftarrow \theta_i + \eta \nabla_{\theta_i} {J(\theta_i)}$;
\EndFor
\EndIf
\If{$t=T$}
\State Initialize $o_{0},t \leftarrow 0$;
\EndIf
\Until{Stop condition is reached}
\end{algorithmic}
\end{algorithm}
\section{Experiments and Discussion}
\label{section:NUMERICAL EXPERIMENTS AND DISCUSSION}
In this section, we evaluate the performance of the proposed MARL algorithm in terms of training efficiency, safety and driving comfort in the considered highway lane changing scenario shown in Fig.~\ref{lane-changing scene}.
\subsection{HDV Models}
In this experiment, we assume that the longitudinal control of HDVs follows the Intelligent Driver Model (IDM) \cite{Treiber}, which is a deterministic continuous-time model describing the dynamics of the position and speed of each vehicle. It takes into account the expected speed, distance between the vehicles and the behavior of the acceleration/deceleration process caused by the different driving habits. In addition, Minimize Overall Braking Induced By Lane Change model (MOBIL) ~\cite{Kesting} is adopted for the lateral control. It takes vehicle acceleration as the input variable of the model and can work well with most car-following models. The acceleration expression is defined as follows:
\begin{equation}
\label{equation:safe}
\tilde{a}_{n} \ge -b_{safe},
\end{equation}
where $\tilde{a}_{n}$ is the acceleration of the new follower after the lane change, and $b_{safe}$ is the maximum braking imposed to the new follower. If the inequality in Eqn.~\ref{equation:safe} is satisfied, the ego vehicle is able to change lanes. The incentive condition is defined as:
\begin{equation}
\label{equation:incentive}
{\underbrace{\tilde{a}_{c}-a_{c}}_{\text{\rm ego~vehicle}}}+p\big({\underbrace{\tilde{a}_{n}-a_{n}}_{\text{\rm new~follower}}}+{\underbrace{\tilde{a}_{o}-a_{o}}_{\text{\rm old~follower}}}\big) \ge \Delta a_{th},
\end{equation}
where $a$ and $\tilde{a}$ are the acceleration of the ego vehicle before and after the lane change, respectively, $\Delta a_{th}$ is the threshold that determines whether to trigger the lane change or not, and $p$ is a politeness coefficient that controls how much effect we want to take into account for the followers, where $p=1$ represents the most considerate drivers whose decision on change lanes may give way to the following blocked vehicles whereas $p=0$ characterizes the most aggressive drivers where the HDV makes selfish lane-changing decisions by only considering their own speed gains and ignoring other vehicles. The performance evaluation of different $p$ values is discussed in Section~\ref{subsection:verification of Driving Comfort}.
\subsection{Experimental Settings}
The simulation environment is modified from the gym-based highway-env simulator~\cite{highway-env}. We set the highway road length to $520~m$, and the vehicles beyond the road are ignored. The vehicles are randomly spawned on the highway with different initial speeds $25-30 ~m/s$ ($56~mph-67~mph$).
The vehicle control sampling frequency is set as the default value of $5~Hz$. The motions of HDVs follow the IDM and MOBIL model, where the maximum deceleration for safety purposes is limited by $b_{safe}=-9~m/s^{2}$, politeness factor $p$ is 0, and the lane-changing threshold $\Delta a_{th}$ is set as $0.1~m/s^{2}$.
To evaluate the effectiveness of the proposed methods, three traffic density levels are employed, which correspond to low, middle, high levels of traffic congestion, respectively. The number of vehicles in different traffic modes is shown in Table~\ref{tab:Traffic density modes}.
\begin{table}[!ht]
\centering
\label{tab:Traffic density modes}
\caption{Traffic density modes.}
\begin{tabular}{cccc}
\hline
Traffic density modes & AVs & HDVs &Explanation \\ \hline
1 & 1-3 & 1-3 & low level \\
2 & 2-4 & 2-4 & middle level \\
3 & 4-6 & 4-6 & high level\\ \hline
\end{tabular}
\end{table}
We train the MARL algorithms for 1 million steps (10,000 epochs) by applying two different random seeds and the same random seed is shared among agents. We evaluate each model 3 times every 200 training episodes. The parameters $\gamma$ and learning rate $\eta$ are set as $0.99$ and $5 \times 10^{-4}$, respectively. The weighting coefficients in the reward function are set as $\omega_{s}=200$, $\omega_{d}=4$, $\omega_{v}=1$ and $\omega_{c}=1$, respectively. These experiments are conducted on a macOS server with a 2.7 GHz Intel Core i5 processor and 8GB of memory.
\subsection{Results \& Analysis}
\subsubsection{Local v.s. Global Reward Designs}
\label{subsection:Local vs Global Rewards}
Fig.~\ref{Global vs Local} shows the performance comparison between the proposed local reward and the global reward design \cite{kaushik2018parameter, dong2020drl} (with shared actor-critic parameters). In all three traffic modes, the local reward design consistently outperforms the global reward design in terms of larger evaluation rewards and smaller variance. In addition, the performance gaps are enlarged as the number of vehicles increases. This is due to the fact that the global reward design is more likely to cause credit assignment issues as mentioned in \cite{sutton2018reinforcement}.
\begin{figure*}[!ht]
\centering
\includegraphics[width=1\textwidth]{imgs/Fig4.pdf}
\caption{Performance comparisons between local and global reward designs. The shaded region denotes the standard deviation over 2 random seeds.}
\label{Global vs Local}
\end{figure*}
\subsubsection{Sharing v.s. Separate Actor-critic Network}
\label{subsection:Sharing vs No-sharing multi-agent A2C Network}
Fig.~\ref{Sharing vs No-sharing} shows the performance comparison between strategies with or without sharing the actor-critic network parameters during training. Obviously, sharing an actor-critic network has better performance than without sharing. Specifically, sharing actor-critic parameters in all three modes results in higher rewards and lower variance. The reason is that, in separate actor-critic networks, the critic network can only guide the actor network to the correct training direction until the critic network is well-trained which may take a long time to achieve. In contrast, the actor network can benefit from the shared state representation via the critic network in a shared actor-critic network \cite{chen2021deep, graesser2019foundations}.
\begin{figure*}[!ht]
\centering
\includegraphics[width=1\textwidth]{imgs/Fig5.pdf}
\caption{Performance comparisons between with and without actor-critic network sharing.}
\label{Sharing vs No-sharing}
\end{figure*}
\subsubsection{Verification of Driving Comfort}
\label{subsection:verification of Driving Comfort}
In this subsection, we evaluate the effectiveness of the proposed multi-objective reward function with the driving comfort in Eqn.~\ref{eqn:reward}. Fig.~\ref{acc} shows the acceleration and deceleration of the AV with or without the comfort measurement defined in Eqn.~\ref{eqn:comfort}. It is clear that the proposed reward design with the comfort measurement has a low variance (average deviation: $0.237 m/s^{2}$) and is more smooth than the reward design without comfort term (average deviation: $0.582m/s^{2}$), which shows the proposed reward design presents good driving comfort.
\begin{figure*}[!ht]
\centering
\includegraphics[width=0.78\textwidth]{imgs/Fig6.pdf}
\caption{Performance comparisons of acceleration between the reward design with or without comfort measurement.}
\label{acc}
\end{figure*}
\subsubsection{Adaptability of the Proposed Method}
\label{subsection:Adaptability of the proposed method}
In this subsection, we evaluate the proposed MA2C under different HDV behaviors, which is controlled by the politeness coefficient $p$ denoted in Eqn.~\ref{equation:incentive}, in which $p=0$ means the most aggressive behavior while $p=1$ represents the most polite behavior.
Fig.~\ref{polite} shows the training performance of two different HDV models (i.e., aggressive or politeness) under different traffic densities. It is clear that the proposed algorithm achieves scalable and stable performance whenever the HDVs take aggressive or courteous behaviors.
\begin{figure*}[!ht]
\centering
\includegraphics[width=1\textwidth]{imgs/Fig7.pdf}
\caption{Performance comparisons on different politeness coefficients $p$ under different traffic densities.}
\label{polite}
\end{figure*}
\subsubsection{Comparison with the state-of-the-art benchmarks}
\label{subsection:Comparison with Other Methods}
In order to demonstrate the performance of the proposed MARL approach, we compared it with several state-of-the-art MARL methods:
\begin{enumerate}
\item \textit{Multi-agent Deep Q-Network} (MADQN) \cite{Ji}: This is the multi-agent version of Deep Q-Network (DQN) \cite{mnih2013playing}, which is an off-policy RL method by applying a deep neural network to approximate the value function and an experience replay buffer to break the correlations between samples to stabilize the training.
\item \textit{Multi-agent actor-critic using Kronecker-Factored Trust Region} (MAACKTR): This is the multi-agent version of actor-critic using Kronecker-Factored Trust Region (ACKTR) \cite{Yuhuaiwu}, which is an on-policy RL algorithm by optimizing both the actor and the critic using Kronecker-factored approximate curvature (K-FAC) with trust region.
\item \textit{Multi-agent Proximal Policy Optimization (MAPPO)} \cite{Schulman}: This is a multi-agent version of Proximal Policy Optimization (PPO) \cite{schulman2017proximal}, which improves the trust region policy optimization (TRPO) \cite{schulman2015trust} by using a clipped surrogate objective and adaptive KL penalty coefficient.
\item \textit{The Proposed MA2C}: This is our proposed method with the designed multi-objective reward function design, parameter sharing and local reward design schemes.
\end{enumerate}
Table~\ref{tab:Mean episode reward in different traffic flow scenario} shows the average return for the MARL algorithms during the evaluation. Obviously, the proposed MA2C algorithm shows the best performance under the density1 scenario than other MARL algorithms. It also shows promising results on the density2 and density3 scenarios and outperforms MAACKTR and MAPPO algorithms. Note that even though MADQN shows a better average reward than the MA2C algorithm, it shows larger reward deviations which may cause unstable training and safety issues. Similarly, the evaluation curves during the training process are shown in Fig.~\ref{Comparison with Other Methods}. As expected, the proposed MA2C algorithm outperforms other benchmarks in terms of evaluation reward and reward standard deviations.
\begin{figure*}[!ht]
\centering
\includegraphics[width=1\linewidth]{imgs/Fig8.pdf}
\caption{Performance comparisons on accumulated rewards in MADQN, MA2C, MAACKTR, and MAPPO.}
\label{Comparison with Other Methods}
\end{figure*}
\begin{table}[!ht]
\centering
\caption{Mean episode reward in different traffic flow scenario.}
\label{tab:Mean episode reward in different traffic flow scenario}
\begin{tabular}{cccc}
\hline
Method & Density 1 & Density 2 &Density 3 \\ \hline
MADQN &47.451 &51.568 &48.509 \\
&($\pm 27.948$)& ($\pm 32.943$) & ($\pm 24.078$) \\
MA2C & 58.000 & 44.744 & 32.579 \\
& ($\pm 9.308$) & ($\pm 10.895$) & ( $\pm 8.160$) \\
MAACKTR &8.812 &3.759 &4.892 \\
&($\pm 6.217$) &($\pm 10.858$)&($\pm 10.986$)\\
MAPPO & 31.988 &19.300 &5.073 \\
&($\pm 6.567$) &($\pm 16.097$) &($\pm 19.762$)\\ \hline
\end{tabular}
\end{table}
\subsubsection{Policy Interpretation}
In this subsection, we attempt to interpret the learned AVs' behavior. Fig.~\ref{Lane Change} shows the snapshots during testing at time steps 20, 28, and 40. As shown in Fig.~\ref{inital_state}, ego vehicle \ding{197} attempts to make a lane change to achieve a higher speed. To make a safe lane change, ego vehicle \ding{197} and ego vehicle \ding{198} are expected to work cooperatively. Specially, the ego vehicle \ding{198} should slow down to make space for the ego vehicle \ding{197} to avoid collisions, which is also represented in Fig.~\ref{Speeds of the AVs}, where the ego vehicle \ding{198} starts to slow down at about 20-time steps. Then the ego vehicle \ding{197} begins to speed up to make the lane change as shown in Fig.~\ref{changing_lanes} and Fig.~\ref{Speeds of the AVs}. Meanwhile, the ego vehicle \ding{198} continues to slow down to ensure a safe headway distance with ego vehicle \ding{197} as shown in Fig.~\ref{Speeds of the AVs}. Fig.~\ref{lane_change_completed} shows the completed lane changes, at which time the ego vehicle \ding{198} starts to speed up. This demonstration shows the proposed MARL framework learns a reasonable and cooperative policy for ego vehicles.
\begin{figure}[!ht]
\centering
\subfigure[initial state]{
\begin{minipage}[t]{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/Fig9.pdf}
\label{inital_state}
\end{minipage}%
}%
\quad
\subfigure[changing lanes]{
\begin{minipage}[t]{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/Fig10.pdf}
\label{changing_lanes}
\end{minipage}%
}%
\quad
\subfigure[lane change completed]{
\begin{minipage}[t]{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/Fig11.pdf}
\label{lane_change_completed}
\end{minipage}
}%
\centering
\caption{Lane change in simulation environment (vehicles \ding{192}-\ding{194}: HDVs, vehicles \ding{197}-\ding{199}: AVs).}
\label{Lane Change}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{imgs/Fig12.pdf}
\caption{Speeds of the AVs \ding{197}, \ding{198} and HDVs \ding{192}.}
\label{Speeds of the AVs}
\end{figure}
\section{Conclusion}
\label{section:Conclusion}
In this paper, we formulated the highway lane-changing problem in mixed traffic as an on-policy MARL problem, and extended the A2C into the multi-agent setting, featuring a novel local reward design and parameter sharing schemes. Specifically, a multi-objective reward function is proposed to simultaneously promote the driving efficiency, comfort, and safety of autonomous driving. Comprehensive experimental results, conducted on three different traffic densities under different levels of HDV aggressiveness, show that our proposed MARL framework consistently outperforms several state-of-the-art benchmarks in terms of efficiency, safety and driver comfort.
\bmhead{Acknowledgments}
The authors are grateful for the efforts of our colleagues in the Sino-German Center of Intelligent Systems, Tongji University. We are grateful for the suggestions on our manuscript from Dr. Qi Deng.
\bmhead{Authors' Contribution}
Wei Zhou, Dong Chen, Jun Yan, and Prof. Zhaojian Li participated in the framework design and manuscript writing, and Wei Zhou implemented experiments inspired by the encouragement and guidance of Dong Chen. Prof. Huilin Yin helped revise the manuscript. Prof. Wancheng Ge is the master supervisor of Wei Zhou in Tongji University which provides the opportunity to complete the research work.
\bmhead{Funding}
Jun Yan is supported by the National Natural Science
Foundation of China under Grant No. 61701348 hosted by Pro. Huilin Yin. Jun Yan, Prof. Huilin Yin, and Prof. Wancheng Ge are grateful to the generous support of T{\"{U}}V S{\"{U}}D. Wei Zhou is supported by the DAAD scholarship for a dual-degree program between Tongji University and Technical University of Munich.
\bmhead{Availability of data and materials}
Not applicable.
\bmhead{Code availability}
Not applicable.
\bmhead{Competing interests}
There are no conflicts of interest for this paper.
|
\section{Introduction}
\label{sec:intro}
\begin{figure}[t]\centering
\includegraphics[width=0.98\linewidth]{fig/arch}
\caption{\textbf{Our MAE architecture}. During pre-training, a large random subset of image patches (\eg, 75\%) is masked out. The encoder is applied to the small subset of \emph{visible patches}. Mask tokens are introduced \textit{after} the encoder, and the full set of encoded patches and mask tokens is processed by a small decoder that reconstructs the original image in pixels. After pre-training, the decoder is discarded and the encoder is applied to uncorrupted images (full sets of patches) for recognition tasks.}
\label{fig:arch}
\end{figure}
Deep learning has witnessed an explosion of architectures of continuously growing capability and capacity \cite{Krizhevsky2012, He2016, Vaswani2017}. Aided by the rapid gains in hardware, models today can easily overfit one million images \cite{Deng2009} and begin to demand hundreds of millions of---often publicly inaccessible---\textit{labeled} images \cite{Dosovitskiy2021}.
This appetite for data has been successfully addressed in natural language processing (NLP) by self-supervised pre-training. The solutions, based on autoregressive language modeling in GPT \cite{Radford2018, Radford2019, Brown2020} and \emph{masked autoencoding} in BERT \cite{Devlin2019}, are conceptually simple: they remove a portion of the data and learn to predict the removed content. These methods now enable training of generalizable NLP models containing over one hundred billion parameters \cite{Brown2020}.
The idea of masked autoencoders, a form of more general denoising autoencoders \cite{Vincent2008}, is natural and applicable in computer vision as well. Indeed, closely related research in vision \cite{Vincent2010,Pathak2016} preceded BERT. However, despite significant interest in this idea following the success of BERT, progress of autoencoding methods in vision lags behind NLP. We ask: \textit{what makes masked autoencoding different between vision and language}? We attempt to answer this question from the following perspectives:
\begin{figure*}[t]\centering\vspace{-.5em}
\includegraphics[width=0.98\linewidth]{fig/samples}\vspace{-.5em}
\caption{Example results on ImageNet \emph{validation} images. For each triplet, we show the masked image (left), our MAE reconstruction$^\dagger$ (middle), and the ground-truth (right). The masking ratio is {80\%}, leaving only 39 out of 196 patches. More examples are in the appendix.\\ \textit{\footnotesize $^\dagger$As no loss is computed on visible patches, the model output on visible patches is qualitatively worse. One can simply overlay the output with the visible patches to improve visual quality. We intentionally opt not to do this, so we can more comprehensively demonstrate the method's behavior.}}
\label{fig:samples}\vspace{-.5em}
\end{figure*}
\begin{figure*}[t]\centering
\includegraphics[width=0.98\linewidth]{fig/samples_coco}\vspace{-.5em}
\caption{Example results on COCO validation images, using an MAE trained on ImageNet (the same model weights as in Figure~\ref{fig:samples}). Observe the reconstructions on the two right-most examples, which, although different from the ground truth, are semantically plausible.}
\label{fig:samples_coco}\vspace{-2mm}
\end{figure*}
\textbf{(i)} Until recently, architectures were different. In vision, convolutional networks \cite{LeCun1989} were dominant over the last decade \cite{Krizhevsky2012}. Convolutions typically operate on regular grids and it is not straightforward to integrate `indicators' such as mask tokens \cite{Devlin2019} or positional embeddings \cite{Vaswani2017} into convolutional networks. This architectural gap, however, has been addressed with the introduction of Vision Transformers (ViT) \cite{Dosovitskiy2021} and should no longer present an obstacle.
\textbf{(ii)} Information density is different between language and vision. Languages are human-generated signals that are highly semantic and information-dense. When training a model to predict only a few missing words per sentence, this task appears to induce sophisticated language understanding. Images, on the contrary, are natural signals with heavy spatial redundancy---\eg, a missing patch can be recovered from neighboring patches with little high-level understanding of parts, objects, and scenes. To overcome this difference and encourage learning useful features, we show that a simple strategy works well in computer vision: masking a \textit{very high} portion of random patches. This strategy largely reduces redundancy and creates a challenging self-supervisory task that requires holistic understanding beyond low-level image statistics. To get a qualitative sense of our reconstruction task, see Figures~\ref{fig:samples} -- \ref{fig:mask_generalization}.
\textbf{(iii)} The autoencoder's \textit{decoder}, which maps the latent representation back to the input, plays a different role between reconstructing text and images. In vision, the decoder reconstructs \emph{pixels}, hence its output is of a lower \mbox{semantic} level than common recognition tasks. This is in contrast to language, where the decoder predicts missing \emph{words} that contain rich semantic information. While in BERT the decoder can be trivial (an MLP) \cite{Devlin2019}, we found that for images, the decoder design plays a key role in determining the semantic level of the learned latent representations.
Driven by this analysis, we present a simple, effective, and scalable form of a masked autoencoder (MAE) for visual representation learning. Our MAE masks random patches from the input image and reconstructs the missing patches in the pixel space. It has an \textit{asymmetric} encoder-decoder design. Our encoder operates only on the visible subset of patches (without mask tokens), and our decoder is lightweight and reconstructs the input from the latent representation along with mask tokens (Figure~\ref{fig:arch}). Shifting the mask tokens to the small decoder in our asymmetric encoder-decoder results in a large reduction in computation. Under this design, a very high masking ratio (\eg, 75\%) can achieve a win-win scenario: it optimizes accuracy while allowing the encoder to process only a small portion (\eg, 25\%) of patches. This can reduce overall pre-training time by 3${\times}$ or more and likewise reduce memory consumption, enabling us to easily scale our MAE to large models.
Our MAE learns very high-capacity models that generalize well. With MAE pre-training, we can train data-hungry models like ViT-Large/-Huge \cite{Dosovitskiy2021} on ImageNet-1K with improved generalization performance. With a vanilla \mbox{ViT-Huge} model, we achieve 87.8\% accuracy when fine-tuned on ImageNet-1K. This outperforms all previous results that use only ImageNet-1K data. We also evaluate transfer learning on object detection, instance segmentation, and semantic segmentation. In these tasks, our pre-training achieves better results than its supervised pre-training counterparts, and more importantly, we observe significant gains by scaling up models. These observations are aligned with those witnessed in self-supervised pre-training in NLP \cite{Devlin2019, Radford2018, Radford2019, Brown2020} and we hope that they will enable our field to explore a similar trajectory.
\begin{figure}[t]\centering
\includegraphics[width=0.995\linewidth]{fig/mask_generalization}\vspace{-.7em}
\caption{Reconstructions of ImageNet \textit{validation} images using an MAE pre-trained with a masking ratio of 75\% but applied on inputs with higher masking ratios. The predictions differ plausibly from the original images, showing that the method can generalize. }
\label{fig:mask_generalization}
\end{figure}
\section{Related Work}\label{sec:related}
\paragraph{Masked language modeling} and its autoregressive counterparts, \eg, BERT \cite{Devlin2019} and GPT \cite{Radford2018, Radford2019, Brown2020}, are highly successful methods for pre-training in NLP. These methods hold out a portion of the input sequence and train models to predict the missing content. These methods have been shown to scale excellently \cite{Brown2020} and a large abundance of evidence indicates that these pre-trained representations generalize well to various downstream tasks.
\paragraph{Autoencoding} is a classical method for learning representations. It has an encoder that maps an input to a latent representation and a decoder that reconstructs the input. For example, PCA and k-means are autoencoders \cite{Hinton1994}. Denoising autoencoders (DAE) \cite{Vincent2008} are a class of autoencoders that corrupt an input signal and learn to reconstruct the original, uncorrupted signal. A series of methods can be thought of as a generalized DAE under different corruptions, \eg, masking pixels \cite{Vincent2010, Pathak2016, Chen2020c} or removing color channels \cite{Zhang2016}. Our MAE is a form of denoising autoencoding, but different from the classical DAE in numerous ways.
\paragraph{Masked image encoding} methods learn representations from images corrupted by masking. The pioneering work of \cite{Vincent2010} presents masking as a noise type in DAE. Context Encoder \cite{Pathak2016} inpaints large missing regions using convolutional networks. Motivated by the success in NLP, related recent methods \cite{Chen2020c, Dosovitskiy2021, Bao2021} are based on Transformers \cite{Vaswani2017}. iGPT \cite{Chen2020c} operates on sequences of pixels and predicts unknown pixels. The ViT paper \cite{Dosovitskiy2021} studies masked patch prediction for self-supervised learning. Most recently, BEiT \cite{Bao2021} proposes to predict discrete tokens \cite{Oord2017, Ramesh2021}.
\paragraph{Self-supervised learning} approaches have seen significant interest in computer vision, often focusing on different pretext tasks for pre-training \cite{Doersch2015, Wang2015a, Noroozi2016, Zhang2016, Pathak2017, Gidaris2018}. Recently, contrastive learning \cite{Becker1992, Hadsell2006} has been popular, \eg, \cite{Wu2018a, Oord2018, He2020, Chen2020}, which models image similarity and dissimilarity (or only similarity \cite{Grill2020, Chen2021}) between two or more views. Contrastive and related methods strongly depend on data augmentation \cite{Chen2020, Grill2020, Chen2021}. Autoencoding pursues a conceptually different direction, and it exhibits different behaviors as we will present.
\vspace{1mm}\section{Approach}\vspace{0.5mm}
\label{sec:approach}
Our masked autoencoder (MAE) is a simple autoencoding approach that reconstructs the original signal given its partial observation. Like all autoencoders, our approach has an encoder that maps the observed signal to a latent representation, and a decoder that reconstructs the original signal from the latent representation. Unlike classical autoencoders, we adopt an \emph{asymmetric} design that allows the encoder to operate only on the partial, observed signal (without mask tokens) and a lightweight decoder that reconstructs the full signal from the latent representation and mask tokens. Figure~\ref{fig:arch} illustrates the idea, introduced next.
\paragraph{Masking.} Following ViT \cite{Dosovitskiy2021}, we divide an image into regular non-overlapping patches. Then we sample a subset of patches and mask (\ie, remove) the remaining ones. Our sampling strategy is straightforward: we sample random patches without replacement, following a uniform distribution. We simply refer to this as ``random sampling".
Random sampling with a \textit{high} masking ratio (\ie, the ratio of removed patches) largely eliminates redundancy, thus creating a task that cannot be easily solved by extrapolation from visible neighboring patches (see Figures~\ref{fig:samples} -- \ref{fig:mask_generalization}). The uniform distribution prevents a potential center bias (\ie, more masked patches near the image center). Finally, the highly sparse input creates an opportunity for designing an efficient encoder, introduced next.
\paragraph{MAE encoder.} Our encoder is a ViT \cite{Dosovitskiy2021} but applied only on \emph{visible, unmasked patches}. Just as in a standard ViT, our encoder embeds patches by a linear projection with added positional embeddings, and then processes the resulting set via a series of Transformer blocks. However, our encoder only operates on a small subset (\eg, 25\%) of the full set. Masked patches are removed; no mask tokens are used. This allows us to train very large encoders with only a fraction of compute and memory. The full set is handled by a lightweight decoder, described next.
\paragraph{MAE decoder.} The input to the MAE decoder is the full set of tokens consisting of (i) encoded visible patches, and (ii) mask tokens. See Figure~\ref{fig:arch}. Each mask token \cite{Devlin2019} is a shared, learned vector that indicates the presence of a missing patch to be predicted. We add positional embeddings to all tokens in this full set; without this, mask tokens would have no information about their location in the image. The decoder has another series of Transformer blocks.
The MAE decoder is only used during pre-training to perform the image reconstruction task (only the encoder is used to produce image representations for recognition). Therefore, the decoder architecture can be flexibly designed in a manner that is \emph{independent} of the encoder design. We experiment with very small decoders, narrower and shallower than the encoder. For example, our default decoder has $<$10\% computation per token \vs the encoder. With this asymmetrical design, the full set of tokens are only processed by the lightweight decoder, which significantly reduces pre-training time.
\paragraph{Reconstruction target.} Our MAE reconstructs the input by predicting the \textit{pixel} values for each masked patch. Each element in the decoder's output is a vector of pixel values representing a patch. The last layer of the decoder is a linear projection whose number of output channels equals the number of pixel values in a patch. The decoder's output is reshaped to form a reconstructed image. Our loss function computes the mean squared error (MSE) between the reconstructed and original images in the pixel space. We compute the loss only on \mbox{masked} patches, similar to BERT \cite{Devlin2019}.\footnotemark
\footnotetext{Computing the loss only on masked patches differs from traditional denoising autoencoders \cite{Vincent2008} that compute the loss on all pixels. This choice is purely result-driven: computing the loss on all pixels leads to a slight decrease in accuracy (\eg, \app0.5\%).}
We also study a variant whose reconstruction target is the normalized pixel values of each masked patch. Specifically, we compute the mean and standard deviation of all pixels in a patch and use them to normalize this patch. Using normalized pixels as the reconstruction target improves representation quality in our experiments.
\paragraph{Simple implementation.} Our MAE pre-training can be implemented efficiently, and importantly, does not require any specialized sparse operations. First we generate a token for every input patch (by linear projection with an added positional embedding). Next we \emph{randomly shuffle} the list of tokens and \emph{remove} the last portion of the list, based on the masking ratio. This process produces a small subset of tokens for the encoder and is equivalent to sampling patches without replacement. After encoding, we append a list of mask tokens to the list of encoded patches, and \emph{unshuffle} this full list (inverting the random shuffle operation) to align all tokens with their targets. The decoder is applied to this full list (with positional embeddings added). As noted, no sparse operations are needed. This simple implementation introduces negligible overhead as the shuffling and unshuffling operations are fast.
\section{ImageNet Experiments}
\label{sec:exp}
We do self-supervised pre-training on the ImageNet-1K (IN1K) \cite{Deng2009} training set. Then we do supervised training to evaluate the representations with (i) end-to-end fine-tuning or (ii) linear probing. We report top-1 validation accuracy of a single 224$\times$224 crop. Details are in Appendix~\ref{app:impl_mae}.
\paragraph{Baseline: ViT-Large.} We use {ViT-Large} (ViT-L/16) \cite{Dosovitskiy2021} as the backbone in our ablation study. ViT-L is very big (an order of magnitude bigger than ResNet-50 \cite{He2016}) and tends to overfit. The following is a comparison between ViT-L trained from scratch \vs fine-tuned from our baseline MAE:
\begin{center}\vspace{-.2em}
\tablestyle{4pt}{1.05}
\begin{tabular}{x{68}x{60}x{60}}
scratch, original \cite{Dosovitskiy2021} & scratch, our impl. & baseline MAE \\
\shline
76.5 & 82.5 & 84.9
\end{tabular}\vspace{-.2em}
\end{center}
We note that it is nontrivial to train \textit{supervised} ViT-L from scratch and a good recipe with strong regularization is needed (82.5\%, see Appendix \ref{app:supervised_vit_large}). Even so, our MAE pre-training contributes a big improvement. Here fine-tuning is only for 50 epochs (\vs 200 from scratch), implying that the fine-tuning accuracy heavily depends on pre-training.
\begin{figure}[t]\centering
\vspace{-1em}
\includegraphics[width=.9\linewidth]{fig/ratio_ft}\\
\scriptsize masking ratio (\%) \\
\includegraphics[width=.9\linewidth]{fig/ratio_linear}\\
\scriptsize masking ratio (\%) \\
\vspace{-.7em}
\caption{\textbf{Masking ratio}. A high masking ratio (75\%) works well for both fine-tuning (top) and linear probing (bottom). The y-axes are ImageNet-1K validation accuracy (\%) in all plots in this paper.
}
\label{fig:mask_ratio}
\vspace{-1em}
\end{figure}
\begin{table*}[t]
\vspace{-.2em}
\centering
\subfloat[
\textbf{Decoder depth}. A deep decoder can improve linear probing accuracy.
\label{tab:decoder_depth}
]{
\centering
\begin{minipage}{0.29\linewidth}{\begin{center}
\tablestyle{4pt}{1.05}
\begin{tabular}{x{18}x{24}x{24}}
blocks & ft & lin \\
\shline
1 & 84.8 & 65.5 \\
2 & \textbf{84.9} & 70.0 \\
4 & \textbf{84.9} & 71.9 \\
8 & \baseline{\textbf{84.9}} & \baseline{\textbf{73.5}} \\
12 & 84.4 & 73.3 \\
\end{tabular}
\end{center}}\end{minipage}
}
\hspace{2em}
\subfloat[
\textbf{Decoder width}. The decoder can be narrower than the encoder (1024-d).
\label{tab:decoder_width}
]{
\begin{minipage}{0.29\linewidth}{\begin{center}
\tablestyle{4pt}{1.05}
\begin{tabular}{x{18}x{24}x{24}}
dim & ft & lin \\
\shline
128 & \textbf{84.9} & 69.1 \\
256 & 84.8 & 71.3 \\
512 & \baseline{\textbf{84.9}} & \baseline{\textbf{73.5}} \\
768 & 84.4 & 73.1 \\
1024 & 84.3 & 73.1 \\
\end{tabular}
\end{center}}\end{minipage}
}
\hspace{2em}
\subfloat[
\textbf{Mask token}.
An encoder without mask tokens is more accurate and faster (Table~\ref{tab:wallclock}).
\label{tab:mask_token}
]{
\begin{minipage}{0.29\linewidth}{\begin{center}
\tablestyle{1pt}{1.05}
\begin{tabular}{y{56}x{24}x{24}z{24}}
case & ft & lin & FLOPs \\
\shline
{encoder w/ \texttt{[M]}} & 84.2 & 59.6 & 3.3$\times$ \\
{encoder {w/o} \texttt{[M]}} & \baseline{\textbf{84.9}} & \baseline{\textbf{73.5}} & \baseline{\textbf{1$\times$}} \\
\multicolumn{4}{c}{~}\\
\multicolumn{4}{c}{~}\\
\multicolumn{4}{c}{~}\\
\end{tabular}
\end{center}}\end{minipage}
}
\\
\centering
\vspace{.3em}
\subfloat[
\textbf{Reconstruction target}. Pixels as reconstruction targets
\label{tab:mae_target} are effective.
]{
\begin{minipage}{0.29\linewidth}{\begin{center}
\tablestyle{6pt}{1.05}
\begin{tabular}{y{54}x{24}x{24}}
case & ft & lin \\
\shline
pixel (w/o norm) & \baseline{84.9} & \baseline{73.5} \\
pixel (w/ norm) & \textbf{85.4} & \textbf{73.9} \\
PCA & 84.6 & 72.3 \\
dVAE token & 85.3 & 71.6 \\
\end{tabular}
\end{center}}\end{minipage}
}
\hspace{2em}
\subfloat[
\textbf{Data augmentation}. Our MAE works with minimal or no augmentation.
\label{tab:aug}
]{
\centering
\begin{minipage}{0.29\linewidth}{\begin{center}
\tablestyle{4pt}{1.05}
\begin{tabular}{y{54}x{22}x{22}}
case & ft & lin \\
\shline
none & 84.0 & 65.7 \\
crop, fixed size & 84.7 & 73.1 \\
crop, rand size & \baseline{\textbf{84.9}} & \baseline{\textbf{73.5}} \\
crop + color jit & 84.3 & 71.9 \\
\end{tabular}
\end{center}}\end{minipage}
}
\hspace{2em}
\subfloat[
\textbf{Mask sampling}. Random sampling works the best. See Figure~\ref{fig:mask_sampling} for visualizations.
\label{tab:mask_types}
]{
\begin{minipage}{0.29\linewidth}{\begin{center}
\tablestyle{1pt}{1.05}
\begin{tabular}{y{28}x{24}x{24}x{24}}
case & ratio & ft & lin \\
\shline
random & 75 & \baseline{\textbf{84.9}} & \baseline{\textbf{73.5}} \\
block & 50 & 83.9 & 72.3 \\
block & 75 & 82.8 & 63.9 \\
grid & 75 & 84.0 & 66.0 \\
\end{tabular}
\end{center}}\end{minipage}
}
\vspace{-.1em}
\caption{\textbf{MAE ablation experiments} with ViT-L/16 on ImageNet-1K. We report fine-tuning (ft) and linear probing (lin) accuracy (\%). If not specified, the default is: the decoder has depth 8 and width 512, the reconstruction target is unnormalized pixels, the data augmentation is random resized cropping, the masking ratio is 75\%, and the pre-training length is 800 epochs. Default settings are marked in \colorbox{baselinecolor}{gray}.}
\label{tab:ablations} \vspace{-.5em}
\end{table*}
\subsection{Main Properties}
We ablate our MAE using the default settings in Table~\ref{tab:ablations} (see caption). Several intriguing properties are observed.
\paragraph{Masking ratio.} Figure~\ref{fig:mask_ratio} shows the influence of the masking ratio. The optimal ratios are surprisingly high. The ratio of 75\% is good for both linear probing and fine-tuning. This behavior is in contrast with BERT \cite{Devlin2019}, whose typical masking ratio is 15\%. Our masking ratios are also much higher than those in related works \cite{Chen2020c,Dosovitskiy2021,Bao2021} in computer vision (20\% to 50\%).
The model \textit{infers} missing patches to produce different, yet plausible, outputs (Figure~\ref{fig:mask_generalization}). It makes sense of the gestalt of objects and scenes, which cannot be simply completed by extending lines or textures. We hypothesize that this reasoning-like behavior is linked to the learning of useful representations.
Figure~\ref{fig:mask_ratio} also shows that linear probing and fine-tuning results follow \textit{different} trends. For linear probing, the accuracy increases steadily with the masking ratio until the sweet point: the accuracy gap is up to $\raise.17ex\hbox{$\scriptstyle\sim$}$20\% (54.6\% \vs 73.5\%). For fine-tuning, the results are less sensitive to the ratios, and a wide range of masking ratios (40--80\%) work well. All fine-tuning results in Figure~\ref{fig:mask_ratio} are better than training from scratch (82.5\%).
\paragraph{Decoder design.} Our MAE decoder can be flexibly designed, as studied in Table~\ref{tab:decoder_depth} and~\ref{tab:decoder_width}.
Table~\ref{tab:decoder_depth} varies the decoder depth (number of Transformer blocks). A sufficiently deep decoder is important for linear probing. This can be explained by the gap between a pixel reconstruction task and a recognition task: the last several layers in an autoencoder are more specialized for reconstruction, but are less relevant for recognition. A reasonably deep decoder can account for the reconstruction specialization, leaving the latent representations at a more abstract level. This design can yield up to 8\% improvement in linear probing (Table~\ref{tab:decoder_depth}, `lin'). However, if fine-tuning is used, the last layers of the encoder can be tuned to adapt to the recognition task. The decoder depth is less influential for improving fine-tuning (Table~\ref{tab:decoder_depth}, `ft').
Interestingly, our MAE with a \textit{single}-block decoder can perform strongly with fine-tuning (84.8\%). Note that a single Transformer block is the minimal requirement to propagate information from visible tokens to mask tokens. Such a small decoder can further speed up training.
In Table~\ref{tab:decoder_width} we study the decoder width (number of channels). We use 512-d by default, which performs well under fine-tuning and linear probing. A narrower decoder also works well with fine-tuning.
Overall, our default MAE decoder is lightweight. It has 8 blocks and a width of 512-d (\colorbox{baselinecolor}{gray} in Table~\ref{tab:ablations}). It only has 9\% FLOPs per token \vs ViT-L (24 blocks, 1024-d). As such, while the decoder processes all tokens, it is still a small fraction of the overall compute.
\begin{table}
\tablestyle{2pt}{1.1}
\begin{tabular}{y{56}x{36}x{32}x{28}x{28}}
encoder & dec. depth & ft acc & hours & speedup \\
\shline
\gc{ViT-L, w/ \texttt{[M]}} & \gc{8} & \gc{84.2} & \gc{42.4} & \gc{-} \\
ViT-L & 8 & 84.9 & 15.4 & 2.8$\times$ \\
ViT-L & 1 & 84.8 & 11.6 & \textbf{3.7}$\times$ \\
\hline
\gc{ViT-H, w/ \texttt{[M]}} & \gc{8} & \gc{-} & \gc{119.6$^\dagger$} & \gc{-} \\
ViT-H & 8 & 85.8 & 34.5 & 3.5$\times$ \\
ViT-H & 1 & 85.9 & 29.3 & \textbf{4.1}$\times$ \\
\end{tabular}
\vspace{-.7em}
\caption{\textbf{Wall-clock time} of our MAE training (800 epochs), benchmarked in 128 TPU-v3 cores with TensorFlow. The speedup is relative to the entry whose encoder has mask tokens (\gc{gray}). The decoder width is 512, and the mask ratio is 75\%. $^\dagger$: This entry is estimated by training ten epochs.}
\label{tab:wallclock}
\vspace{-.5em}
\end{table}
\paragraph{Mask token.} An important design of our MAE is to skip the mask token \texttt{[M]} in the encoder and apply it later in the lightweight decoder. Table~\ref{tab:mask_token} studies this design.
If the encoder \textit{uses} mask tokens, it performs \textit{worse}: its accuracy drops by 14\% in linear probing. In this case, there is a gap between pre-training and deploying: this encoder has a large portion of mask tokens in its input in pre-training, which does not exist in uncorrupted images. This gap may degrade accuracy in deployment. By removing the mask token from the encoder, we constrain the encoder to always see \textit{real} patches and thus improve accuracy.
Moreover, by skipping the mask token in the encoder, we greatly reduce training computation. In Table~\ref{tab:mask_token}, we reduce the overall training FLOPs by 3.3$\times$. This leads to a 2.8$\times$ wall-clock speedup in our implementation (see Table~\ref{tab:wallclock}). The wall-clock speedup is even bigger (3.5--4.1$\times$), for a smaller decoder (1-block), a larger encoder (\mbox{ViT-H}), or both. Note that the speedup can be $>$4$\times$ for a masking ratio of 75\%, partially because the self-attention complexity is quadratic. In addition, memory is greatly reduced, which can enable training even larger models or speeding up more by large-batch training. The time and memory efficiency makes our MAE favorable for training very large models.
\paragraph{Reconstruction target.} We compare different reconstruction targets in Table~\ref{tab:mae_target}. Our results thus far are based on pixels without (per-patch) normalization. Using pixels \textit{with} normalization improves accuracy. This per-patch normalization enhances the contrast locally. In another variant, we perform PCA in the patch space and use the largest PCA coefficients (96 here) as the target. Doing so degrades accuracy. Both experiments suggest that the high-frequency components are useful in our method.
We also compare an MAE variant that predicts \textit{tokens}, the target used in BEiT \cite{Bao2021}. Specifically for this variant, we use the DALLE pre-trained dVAE \cite{Ramesh2021} as the tokenizer, following \cite{Bao2021}. Here the MAE decoder predicts the token indices using cross-entropy loss. This tokenization improves fine-tuning accuracy by 0.4\% \vs unnormalized pixels, but has no advantage \vs normalized pixels. It also reduces linear probing accuracy. In \mbox{\S\ref{sec:transfer}} we further show that tokenization is not necessary in transfer learning.
Our \textit{pixel}-based MAE is much simpler than tokenization. The dVAE tokenizer requires one more pre-training stage, which may depend on extra data (250M images \cite{Ramesh2021}). The dVAE encoder is a large convolutional network (40\% FLOPs of ViT-L) and adds nontrivial overhead. Using pixels does not suffer from these problems.
\begin{figure}[t]
\centering
\includegraphics[width=0.99\linewidth]{fig/mask_sampling}
\vspace{-.3em}
\caption{\textbf{Mask sampling strategies} determine the pretext task difficulty, influencing reconstruction quality and representations (Table~\ref{tab:mask_types}). Here each output is from an MAE trained with the specified masking strategy. {Left}: random sampling (our default). {Middle}: block-wise sampling \cite{Bao2021} that removes large random blocks. {Right}: grid-wise sampling that keeps one of every four patches. Images are from the validation set.}
\label{fig:mask_sampling}\vspace{.5em}
\end{figure}
\paragraph{Data augmentation.} Table~\ref{tab:aug} studies the influence of data augmentation on our MAE pre-training.
Our MAE works well using \textit{cropping-only} augmentation, either fixed-size or random-size (both having random horizontal flipping). Adding color jittering degrades the results and so we do not use it in other experiments.
Surprisingly, our MAE behaves decently even if using \textit{no data augmentation} (only center-crop, no flipping). This property is dramatically different from contrastive learning and related methods \cite{Wu2018a,He2020,Chen2020,Grill2020}, which heavily rely on data augmentation. It was observed \cite{Grill2020} that using cropping-only augmentation reduces the accuracy by 13\% and 28\% respectively for BYOL \cite{Grill2020} and SimCLR \cite{Chen2020}. In addition, there is no evidence that contrastive learning can work without augmentation: the two views of an image are the same and can easily satisfy a trivial solution.
In MAE, the role of data augmentation is mainly performed by random masking (ablated next). The masks are different for each iteration and so they generate new training samples regardless of data augmentation. The pretext task is made difficult by masking and requires less augmentation to regularize training.
\paragraph{Mask sampling strategy.} In Table~\ref{tab:mask_types} we compare different mask sampling strategies, illustrated in Figure~\ref{fig:mask_sampling}.
The \textit{block-wise} masking strategy, proposed in \cite{Bao2021}, tends to remove large blocks (Figure~\ref{fig:mask_sampling} middle). Our MAE with block-wise masking works reasonably well at a ratio of 50\%, but degrades at a ratio of 75\%. This task is harder than that of random sampling, as a higher training loss is observed. The reconstruction is also blurrier.
We also study \textit{grid-wise} sampling, which regularly keeps one of every four patches (Figure~\ref{fig:mask_sampling} right). This is an easier task and has lower training loss. The reconstruction is sharper. However, the representation quality is lower.
Simple random sampling works the best for our MAE. It allows for a higher masking ratio, which provides a greater speedup benefit while also enjoying good accuracy.
\paragraph{Training schedule.} Our ablations thus far are based on 800-epoch pre-training. Figure~\ref{fig:schedule} shows the influence of the training schedule length. The accuracy improves steadily with longer training. Indeed, we have not observed saturation of linear probing accuracy even at 1600 epochs. This behavior is unlike contrastive learning methods, \eg, MoCo~v3 \cite{Chen2021a} saturates at 300 epochs for ViT-L. Note that the MAE encoder only sees 25\% of patches per epoch, while in contrastive learning the encoder sees 200\% (two-crop) or even more (multi-crop) patches per epoch.
\begin{figure}[t]\centering
\includegraphics[width=0.99\linewidth]{fig/schedule_ft.pdf}\\
\scriptsize epochs (log-scale) \\
\includegraphics[width=0.99\linewidth]{fig/schedule_linear.pdf}\\
\scriptsize epochs (log-scale) \\
\vspace{-.5em}
\caption{\textbf{Training schedules}. A longer training schedule gives a \mbox{noticeable} improvement. Here each point is a full training schedule. The model is ViT-L with the default setting in Table~\ref{tab:ablations}.}
\label{fig:schedule}
\vspace{-.5em}
\end{figure}
\begin{table}
\vspace{-.5em}
\tablestyle{3pt}{1.1}
\begin{tabular}{l l x{24}x{24}x{24}x{24}}
\multirow{1}{*}{method} &
\multirow{1}{*}{pre-train data}
& \multicolumn{1}{c}{ViT-B} & \multicolumn{1}{c}{ViT-L} & \multicolumn{1}{c}{ViT-H} & \multicolumn{1}{c}{ViT-H$_\text{448}$} \\
\shline
\gc{scratch, our impl.} & \gc{-} & \gc{82.3} & \gc{82.6} & \gc{83.1} & \gc{-} \\
DINO \cite{Caron2021} & \scriptsize IN1K & 82.8 & - & - & - \\
MoCo v3 \cite{Chen2021a} & \scriptsize IN1K & 83.2 & 84.1 & - & - \\
BEiT \cite{Bao2021} & \scriptsize IN1K+DALLE & 83.2 & 85.2 & - & - \\
\hline
MAE & \scriptsize IN1K & \underline{83.6} & \underline{85.9} & \underline{86.9} & \textbf{87.8} \\
\end{tabular}
\vspace{-.8em}
\caption{\textbf{Comparisons with previous results on ImageNet-1K}. The pre-training data is the ImageNet-1K training set (except the tokenizer in BEiT was pre-trained on 250M DALLE data \cite{Ramesh2021}). All self-supervised methods are evaluated by end-to-end fine-tuning. The ViT models are B/16, L/16, H/14 \cite{Dosovitskiy2021}. The best for each column is underlined. All results are on an image size of 224, except for ViT-H with an extra result on 448. Here our MAE reconstructs normalized pixels and is pre-trained for 1600 epochs.}
\label{tab:imagenet_e2e}
\end{table}
\subsection{Comparisons with Previous Results}
\paragraph{Comparisons with self-supervised methods.} In Table~\ref{tab:imagenet_e2e} we compare the fine-tuning results of self-supervised ViT models. For ViT-B, all methods perform closely. For \mbox{ViT-L}, the gaps among methods are bigger, suggesting that a challenge for bigger models is to reduce overfitting.
\begin{figure}[t]\centering
\vspace{-.7em}
\begin{overpic}[percent,width=.98\linewidth]{fig/model_size}
\put(86.8,9.8){\tiny \cite{Dosovitskiy2021}}
\put(90.8,6.8){\tiny \cite{Dosovitskiy2021}}
\end{overpic}
\vspace{-.5em}
\scriptsize params (M) \\
\vspace{-.5em}
\caption{\textbf{MAE pre-training \vs supervised pre-training}, evaluated by fine-tuning in ImageNet-1K (224 size). We compare with the original ViT results \cite{Dosovitskiy2021} trained in IN1K or JFT300M.}
\label{fig:model_size} \vspace{-.7em}
\end{figure}
Our MAE can scale up easily and has shown steady improvement from bigger models. We obtain 86.9\% accuracy using \mbox{ViT-H} (224 size). By fine-tuning with a 448 size, we achieve \textbf{87.8}\% accuracy, \textit{using only IN1K data}. The previous best accuracy, among all methods using only IN1K data, is 87.1\% (512 size) \cite{Yuan2021}, based on advanced networks. We improve over the state-of-the-art by a nontrivial margin in the highly competitive benchmark of IN1K (no external data). Our result is based on \textit{vanilla} ViT, and we expect advanced networks will perform better.
Comparing with BEiT \cite{Bao2021}, our MAE is \textit{more accurate} while being \textit{simpler} and \textit{faster}. Our method reconstructs pixels, in contrast to BEiT that predicts tokens: BEiT reported a 1.8\% degradation \cite{Bao2021} when reconstructing pixels with \mbox{ViT-B}.\footnotemark~We do not need dVAE pre-training. Moreover, our MAE is considerably faster (3.5$\times$ per epoch) than BEiT, for the reason as studied in Table~\ref{tab:mask_token}.
\footnotetext{We observed the degradation also in BEiT with ViT-L: it produces 85.2\% (tokens) and 83.5\% (pixels), reproduced from the official code.}
The MAE models in Table~\ref{tab:imagenet_e2e} are pre-trained for 1600 epochs for better accuracy (Figure~\ref{fig:schedule}). Even so, our total pre-training time is \textit{less} than the other methods when trained on the same hardware. For example, training \mbox{ViT-L} on 128 TPU-v3 cores, our MAE's training time is 31 hours for 1600 epochs and MoCo v3's is 36 hours for 300 epochs \cite{Chen2021a}.
\paragraph{Comparisons with supervised pre-training.} In the original ViT paper \cite{Dosovitskiy2021}, ViT-L degrades when trained in IN1K. Our implementation of supervised training (see \ref{app:supervised_vit_large}) works better, but accuracy saturates. See Figure~\ref{fig:model_size}.
Our MAE pre-training, using only IN1K, can generalize better: the gain over training from scratch is bigger for higher-capacity models. It follows a trend similar to the \mbox{JFT-300M} \textit{supervised} pre-training in \cite{Dosovitskiy2021}. This comparison shows that our MAE can help scale up model sizes.
\begin{figure}[t]\centering
\vspace{-.7em}
\includegraphics[width=.9\linewidth]{fig/partial_ft} \\
\vspace{-.5em}
{\scriptsize \# blocks fine-tuned} \\
\vspace{-.3em}
\caption{\textbf{Partial fine-tuning} results of ViT-L \wrt the number of fine-tuned Transformer blocks under the default settings from Table~\ref{tab:ablations}. Tuning 0 blocks is linear probing; 24 is full fine-tuning. Our MAE representations are less linearly separable, but are consistently better than MoCo v3 if one or more blocks are tuned.}
\label{fig:partial_ft} \vspace{-1em}
\end{figure}
\subsection{Partial Fine-tuning}
\label{sec:partial_ft}
Table~\ref{tab:ablations} shows that linear probing and fine-tuning results are largely \textit{uncorrelated}. Linear probing has been a popular protocol in the past few years; however, it misses the opportunity of pursuing \textit{strong but non-linear} features---which is indeed a strength of deep learning. As a middle ground, we study a \textit{partial fine-tuning} protocol: fine-tune the last several layers while freezing the others. This protocol was also used in early works, \eg, \cite{Yosinski2014,Zhang2016,Noroozi2016}.
Figure~\ref{fig:partial_ft} shows the results. Notably, fine-tuning only \textit{one} Transformer block boosts the accuracy significantly from 73.5\% to 81.0\%. Moreover, if we fine-tune only ``half" of the last block (\ie, its MLP sub-block), we can get 79.1\%, much better than linear probing. This variant is essentially fine-tuning an MLP head. Fine-tuning a few blocks (\eg, 4 or 6) can achieve accuracy close to full fine-tuning.
In Figure~\ref{fig:partial_ft} we also compare with MoCo v3 \cite{Chen2021a}, a contrastive method with ViT-L results available. MoCo v3 has higher linear probing accuracy; however, all of its partial fine-tuning results are worse than MAE. The gap is 2.6\% when tuning 4 blocks. While the MAE representations are less linearly separable, they are stronger \textit{non-linear} features and perform well when a non-linear head is tuned.
These observations suggest that linear separability is not the sole metric for evaluating representation quality. It has also been observed (\eg, \cite{Chen2021}) that linear probing is not well \mbox{correlated} with transfer learning performance, \eg, for object detection. To our knowledge, linear evaluation is not often used in NLP for benchmarking pre-training.
\section{Transfer Learning Experiments}\label{sec:transfer}
\begin{table}[t]
\vspace{-1.5em}
\tablestyle{5pt}{1.05}
\begin{tabular}{llcccc}
& & \multicolumn{2}{c}{AP$^\text{box}$} & \multicolumn{2}{c}{\gc{AP$^\text{mask}$}} \\
method & pre-train data & ViT-B & ViT-L & \gc{ViT-B} & \gc{ViT-L} \\
\shline
supervised & \scriptsize IN1K w/ labels & 47.9 & 49.3 & \gc{42.9} & \gc{43.9} \\
MoCo v3 & \scriptsize IN1K & 47.9 & 49.3 & \gc{42.7} & \gc{44.0} \\
BEiT & \scriptsize IN1K+{DALLE} & 49.8 & \textbf{53.3} & \gc{44.4} & \gc{47.1} \\
\hline
MAE & \scriptsize IN1K & \textbf{50.3} & \textbf{53.3} & \gc{\textbf{44.9}} & \gc{\textbf{47.2}} \\
\end{tabular}
\vspace{-.7em}
\caption{\textbf{COCO object detection and segmentation} using a ViT Mask R-CNN baseline. All entries are based on our implementation. Self-supervised entries use IN1K data \textit{without} labels. Mask AP follows a similar trend as box AP.}
\label{tab:coco} \vspace{-1em}
\end{table}
We evaluate transfer learning in downstream tasks using the pre-trained models in Table~\ref{tab:imagenet_e2e}.
\paragraph{Object detection and segmentation.} We fine-tune Mask R-CNN \cite{He2017} end-to-end on COCO \cite{Lin2014}. The ViT backbone is adapted for use with FPN~\cite{Lin2017} (see \ref{app:coco}). We apply this approach for all entries in Table~\ref{tab:coco}. We report box AP for object detection and mask AP for instance segmentation.
Compared to supervised pre-training, our MAE performs better under all configurations (Table~\ref{tab:coco}). With the smaller ViT-B, our MAE is 2.4 points higher than \textit{supervised} pre-training (50.3 \vs 47.9, AP$^\text{box}$). More significantly, with the larger ViT-L, our MAE pre-training outperforms supervised pre-training by 4.0 points (53.3 \vs 49.3).
The \textit{pixel}-based MAE is better than or on par with the \textit{token}-based BEiT, while MAE is much simpler and faster. Both MAE and BEiT are better than MoCo v3 and MoCo v3 is on par with supervised pre-training.
\paragraph{Semantic segmentation.} We experiment on ADE20K \cite{Zhou2019} using UperNet \cite{Xiao2018} (see \ref{app:ade20k}). Table~\ref{tab:ade20k} shows that our pre-training significantly improves results over \textit{supervised} pre-training, \eg, by 3.7 points for ViT-L. Our pixel-based MAE also outperforms the token-based BEiT. These observations are consistent with those in COCO.
\paragraph{Classification tasks.} \mbox{Table~\ref{tab:cls_transfer}} studies transfer learning on the iNaturalists \cite{VanHorn2018} and Places \cite{Zhou2014} tasks (see \ref{app:class}). On iNat, our method shows strong scaling behavior: accuracy improves considerably with bigger models. Our results surpass the previous best results \textit{by large margins}. On Places, our MAE outperforms the previous best results \cite{Goyal2021,Mahajan2018}, which were obtained via pre-training on billions of images.
\paragraph{Pixels \vs tokens.} Table~\ref{tab:pixel_vs_token} compares pixels \vs tokens as the MAE reconstruction target. While using dVAE tokens is better than using \textit{unnormalized} pixels, it is statistically similar to using \textit{normalized} pixels across all cases we tested. It again shows that tokenization is not necessary for our MAE.
\begin{table}[t]
\vspace{-.5em}
\tablestyle{8pt}{1.05}
\begin{tabular}{llcc}
method & pre-train data & ViT-B & ViT-L \\
\shline
supervised & \scriptsize IN1K w/ labels & 47.4 & 49.9 \\
MoCo v3 & \scriptsize IN1K & 47.3 & 49.1 \\
BEiT & \scriptsize IN1K+{DALLE} & 47.1 & 53.3 \\
\hline
MAE & \scriptsize IN1K & \textbf{48.1} & \textbf{53.6} \\
\end{tabular}
\vspace{-.7em}
\caption{\textbf{ADE20K semantic segmentation} (mIoU) using UperNet. BEiT results are reproduced using the official code. Other entries are based on our implementation. Self-supervised entries use IN1K data \textit{without} labels.}
\label{tab:ade20k} \vspace{-.5em}
\end{table}
\begin{table}[t]
\tablestyle{5pt}{1.05}
\begin{tabular}{l x{24}x{24}x{24}x{24}y{36}}
\multirow{1}{*}{dataset}
& \multicolumn{1}{c}{ViT-B} & \multicolumn{1}{c}{ViT-L} & \multicolumn{1}{c}{ViT-H} & \multicolumn{1}{c}{ViT-H$_\text{448}$} & \gc{prev best} \\
\shline
iNat 2017 & 70.5 & 75.7 & 79.3 & \textbf{83.4} & \gc{75.4} \cite{Touvron2019} \\
iNat 2018 & 75.4 & 80.1 & 83.0 & \textbf{86.8} & \gc{81.2} \cite{Touvron2021b} \\
iNat 2019 & 80.5 & 83.4 & 85.7 & \textbf{88.3} & \gc{84.1} \cite{Touvron2021b} \\
Places205 & 63.9 & 65.8 & 65.9 & \textbf{66.8} & \gc{66.0} \cite{Goyal2021}$^\dagger$ \\
Places365 & 57.9 & 59.4 & 59.8 & \textbf{60.3} & \gc{58.0} \cite{Mahajan2018}$^\ddagger$ \\
\end{tabular}
\vspace{-.8em}
\caption{\textbf{Transfer learning accuracy on classification datasets}, using MAE pre-trained on IN1K and then fine-tuned.
We provide system-level comparisons with the previous best results.
\\{\scriptsize $^\dagger$: pre-trained on 1 billion images. $^\ddagger$: pre-trained on 3.5 billion images.}
}
\label{tab:cls_transfer}
\vspace{-.5em}
\end{table}
\begin{table}[t]
\tablestyle{3.5pt}{1.05}
\begin{tabular}{l|rrr|rr|rr}
& \multicolumn{3}{c|}{\scriptsize IN1K}
& \multicolumn{2}{c|}{\scriptsize COCO}
& \multicolumn{2}{c}{\scriptsize ADE20K} \\
& \scriptsize ViT-B & \scriptsize ViT-L & \scriptsize ViT-H
& \scriptsize ViT-B & \scriptsize ViT-L
& \scriptsize ViT-B & \scriptsize ViT-L \\
\shline
pixel (w/o norm) & 83.3 & 85.1 & 86.2 & 49.5 & 52.8 & 48.0 & 51.8 \\
pixel (w/ norm) & 83.6 & 85.9 & 86.9 & 50.3 & 53.3 & 48.1 & 53.6 \\
\hline
dVAE token & 83.6 & 85.7 & 86.9 & 50.3 & 53.2 & 48.1 & 53.4 \\
$\triangle$ & 0.0 & -0.2 & 0.0 & 0.0 & -0.1 & 0.0 & -0.2
\end{tabular}
\vspace{-.7em}
\caption{\textbf{Pixels \vs tokens} as the MAE reconstruction target. $\triangle$ is the difference between using dVAE tokens and using normalized pixels. The difference is statistically insignificant.}
\label{tab:pixel_vs_token} \vspace{-0.5em}
\end{table}
\section{Discussion and Conclusion}
Simple algorithms that scale well are the core of deep learning. In NLP, simple self-supervised learning methods (\eg, \cite{Radford2018, Devlin2019, Radford2019, Brown2020}) enable benefits from exponentially scaling models. In computer vision, practical pre-training paradigms are dominantly supervised (\eg \cite{Krizhevsky2012,Simonyan2015,He2016,Dosovitskiy2021}) despite progress in self-supervised learning. In this study, we observe on ImageNet and in transfer learning that an autoencoder---a simple self-supervised method similar to techniques in NLP---provides scalable benefits. Self-supervised learning in vision may now be embarking on a similar trajectory as in NLP.
On the other hand, we note that images and languages are \textit{signals of a different nature} and this difference must be addressed carefully. Images are merely recorded light \mbox{\textit{without}} a semantic decomposition into the visual analogue of words. Instead of attempting to remove objects, we remove random patches that most likely do \textit{not} form a semantic segment. Likewise, our MAE reconstructs pixels, which are \emph{not} semantic entities. Nevertheless, we observe (\eg, Figure \ref{fig:mask_generalization}) that our MAE infers complex, holistic reconstructions, suggesting it has learned numerous visual concepts, \ie, semantics. We hypothesize that this behavior occurs by way of a rich hidden representation inside the MAE. We hope this perspective will inspire future work.
\paragraph{Broader impacts.} The proposed method predicts content based on learned statistics of the training dataset and as such will reflect biases in those data, including ones with negative societal impacts. The model may generate inexistent content. These issues warrant further research and consideration when building upon this work to generate images.
{
\fontsize{8.2pt}{9.84pt}\selectfont
\bibliographystyle{ieee_fullname} |
\section{Introduction}
The black hole interior is a mysterious region of spacetime where non-perturbative quantum gravity effects are sometimes important.
Despite the importance of such non-perturbative effects, work from the past few decades supports the idea that ``plain'' gravity (e.g. a sum over metrics and possibly a few light fields) knows a lot about fine-grained quantum information.
The paradigmatic examples of this include the geometrization of von Neumann entropy in holographic systems, including its suitable generalization to generate the Page curve of Hawking radiation of an evaporating black hole \cite{Penington:2019npb,Almheiri:2019psf}.
Such features were thought to require a UV complete theory of gravity, such as string theory, and obtaining them from gravity came as a pleasant surprise.
Gravity, however, doesn't know everything. It appears to know of the underlying random unitary dynamics, but it fails to pin down a particular realization of those dynamics. In particular, while gravity is able to reproduce arbitrary moments of the signal drawn from an ensemble of random dynamics, it fails to capture the large fluctuations that come with a given realization. This is a pretty big miss since the size of those fluctuations is of order the signal.
A related point is that whereas the gravity calculations seem reliable for some universal quantities, it seems to know only statistical properties about the non-perturbative effects that are highly theory-dependent. For example, the spectral form factor, which at large times probes the detailed energy spectrum of the black hole, is a highly erratic function that depends sensitively on the couplings\cite{Cotler:2016fpe,Saad:2018bqo}. Similarly, the black hole $S$-matrix that governs the formation and evaporation of a black hole is suspected to be an erratic and possibly pseudo-random matrix \cite{Polchinski:2015cea}. Presumably a precise computation of such quantities from the bulk point of view will involve strings, branes, half-wormholes \cite{Saad:2021rcu}, etc, and the answers would depend on the particular string vacua.
The dependence of these non-perturbative effects on the couplings of the theory suggests that the interior of the black hole is in some sense highly theory dependent. The goal of our work will be to make this more precise in the context of bulk reconstruction. We will consider models which admit an ensemble of boundary theories parameterized by a set of boundary couplings, and analyze the sensitivity of bulk reconstruction on the level of precision in specifying those couplings. This means we will look for instances where the reconstruction fails, and map out which bulk regions are most sensitive to this. Hence, those bulk regions require exquisite knowledge of the couplings to reconstruct.
{\it Note added}: as we were finishing this work, we became aware of work by Qi, Shangnan, and Yang \cite{Qi:2021oni}. We have arranged to coordinate our preprints. See also \cite{Renner:2021qbe}.
\subsection{Knowing the couplings -- an operational definition}
\def{\mathsf{sys}}{{\mathsf{sys}}}
\def{\mathsf{journal}}{{\mathsf{journal}}}
\def\mathsf{ptr}{\mathsf{ptr}}
\def\mathsf{env}{\mathsf{env}}
\def\, \cup \, {\, \cup \, }
\def\mathcal{N}=4{\mathcal{N}=4}
\def{\lambda}{{\boldsymbol{\lambda}}}
\def{\lambda}{{\mathbf{J}}}
The first point we make precise is the notion of ``knowing the couplings." Suppose we have a system whose Hamiltonian depends on a set of parameters ${\lambda} = \{\lambda_i \}$. For example, in SYK it is natural to choose ${\lambda} = \{ J_{ijkl} \}$ to be the set of random couplings. However, in general the couplings could be non-random, for example we could consider $\mathcal{N}=4$ and take ${\lambda} = \{ g^2_{YM} \} $.
We denote the state of the system prepared with those couplings as
\begin{align}
\ket{ \psi; {\lambda} }_{\mathsf{sys}}.
\end{align}
As a concrete example, we could take $ \ket{ \psi; {\lambda} }_{\mathsf{sys}} = \ket{\beta+i2T; {\lambda}}_{\mathsf{sys}}$ to be the thermofield double state associated to the Hamiltonian $H({\lambda})$ at some temperature $\beta$ and some Lorentzian time $T$.
Let's imagine that these couplings are drawn form some distribution $P({\lambda})$. We can keep track of this by using a standard method in quantum information theory of entangling the system with a reference system labelled by the couplings
\begin{align}
\ket{ \Psi}_{{\mathsf{sys}} \, \cup \, {\mathsf{journal}}} = \sum_{{\lambda}} \ \sqrt{P\left ( {\lambda} \right ) } \, | \psi; {\lambda} \rangle_{\mathsf{sys}} | {\lambda} \rangle_{{\mathsf{journal}}}. \label{global:pure}
\end{align}
We call the reference system ${\mathsf{journal}}$ since in the SYK context it records the $J$'s.
Tracing over the ${\mathsf{journal}}$, we get a density matrix
\eqn{\rho_{\mathsf{sys}} = \sum_{{\lambda}} P({\lambda}) \rho_\psi({\lambda}) = \ev{\rho_\psi}_{\lambda}, \quad \rho_\psi({\lambda}) = \ketbra{\psi; {\lambda}}{\psi; {\lambda}}_{\mathsf{sys}}. \label{eq:entangle}}
This state represents the situation where we have no information about the couplings. This ignorance represented in equation \nref{eq:entangle} captured by the non-zero von Neumann entropy between ${\mathsf{sys}}$ and ${\mathsf{journal}}$. If we think of ${\lambda}$ as parameters in a disorder average, we may use condensed matter jargon and refer to the von-Neumann entropy as the {\it annealed} entropy $S\left ( \ev{\rho_\psi}_\lambda\right ) $. In the above discussion we started with a pure state density matrix $\rho_\psi$ but after averaging over ${\lambda}$ we get a mixed state. As a slight generalization, we could consider any subsytem of ${\mathsf{sys}}$, and get a similar formula, where $\rho_\psi$ is replaced by some partial trace of $\ketbra{\psi; {\lambda}}{\psi; {\lambda}}_{\mathsf{sys}}$.
The main question we'd like to answer is the following: How much of the bulk can be reconstructed given the density matrix of the system after tracing out the reference? Or in other words, how much of the bulk is contained within the entanglement wedge of ${\mathsf{sys}}$? In a theory with a semiclassical holographic dual, the answer is given by the quantum extremal surface (QES) formula, which states that the boundary of the entanglement wedge is given by the QES responsible for the von Neumann entropy $S({\mathsf{sys}})$.
An alternative version of this question is: how much of the bulk is contained within the entanglement wedge of the reference? Since the $S({\mathsf{sys}}) = S({\mathsf{journal}})$ for a pure state, the entanglement wedge is simply the complement of the entanglement wedge of ${\mathsf{sys}}$.
Let us make some preliminary comments on what we mean by bulk reconstruction.
First notice that semi-classically, the entropy $S({\mathsf{sys}}) > 0$. The bulk matter is not pure, since it is entangled with ${\mathsf{journal}}$.
This semi-classical entropy can ``pollute'' the bulk and lead to problems with reconstructing any operator using traditional methods like HKLL\cite{Hamilton:2005ju}.
Any matter in the bulk will generically interact (at least weakly) with the fields that are sourced by the couplings we turn on at the boundary of AdS; this will cause problems with reconstruction methods such as HKLL. We will not be discussing such semiclassical problems in this paper. Instead, we will ask for bulk reconstruction in the modern sense of entanglement wedge reconstruction \cite{Jafferis:2015del, Dong:2016eik}. The failure of our ability to do reconstruction will be a non-perturbative effect, signaled by replica wormholes and the appearance of an island.
In the above discussion, we entangled ${\mathsf{sys}}$ to ${\mathsf{journal}}$ in such a way that the global state is pure.
However, one could also consider a setup where instead of entangling ${\mathsf{sys}}$ to a ${\mathsf{journal}}$, we instead classically correlate ${\mathsf{sys}}$ to a ``pointer'' system $\mathsf{ptr}$:
\eqn{\rho_{{\mathsf{sys}}{} \, \cup \, \mathsf{ptr}{}} = \sum_{{\lambda}} P({\lambda}) \rho_\psi({\lambda}) \otimes \ketbra{{\lambda}}{{\lambda}}_\mathsf{ptr}. \label{densitypo} }
This mixed state would be the appropriate description of a setup where Alice flips some coins and uses the outcome to decide which couplings to prepare the system in.
In fact, following the standard discussion of decoherence/measurement theory, this state can be purified by adding an auxiliary system $\mathsf{env}$. Often one adopts the interpretation that the system $\mathsf{ptr}$ is a pointer or measurement device, and $\mathsf{env}$ is the environment. The purification of this state is simply \eqref{global:pure}, where ${\mathsf{journal}} = \mathsf{ptr} \cup \mathsf{env}$ and $\ket{{\lambda}}_{\mathsf{journal}} = \ket{{\lambda}}_{\mathsf{ptr}{}} \ket{{\lambda}}_\mathsf{env}$. Tracing over $\mathsf{env}$ gives \eqref{densitypo}.
Alternatively, we can go the other direction: we start from the global pure state of \eqref{global:pure} and perform a complete measurement in the $\lambda$ basis:
\eqn{\ketbra{ \Psi}{\Psi}_{{\mathsf{sys}} \, \cup \, {\mathsf{journal}}} \to \sum_{\lambda} \Pi_{\lambda}^{{\mathsf{journal}}} \ketbra{ \Psi}{\Psi}_{{\mathsf{sys}} \, \cup \, {\mathsf{journal}}}\Pi_{\lambda}^{{\mathsf{journal}}}.}
Then relabelling ${\mathsf{journal}} \to \mathsf{ptr}$, we obtain \eqref{densitypo}. Thus including the $\mathsf{ptr}$ system makes precise the idea of having classical (e.g. decohered) knowledge of the couplings.
To ask about bulk reconstruction given perfect classical knowledge of the couplings is to ask for the entanglement wedge of ${\mathsf{sys}} \cup \mathsf{ptr}$.
Following the QES rules, we should compute the von Neumann entropy of ${\mathsf{sys}} \cup \mathsf{ptr}$:
\eqn{S({\mathsf{sys}} \, \cup \, \mathsf{ptr}) = \sum_{\lambda} P({\lambda}) S(\rho_\psi({\lambda}) )-\sum_{\lambda} P({\lambda}) \log P({\lambda}). \label{union}
The first term is the von Neumann entropy, averaged over the couplings.
If $\rho_{\lambda}$ is pure, this vanishes. But this formula also applies to an initially mixed state, or a subsystem, e.g., we could take ${\mathsf{sys}}$ to be the left side of the thermofield double. Then the first term would be the disorder-averaged thermal entropy, known as the \emph{quenched} entropy, $\ev{ S\left ( \rho_\psi\right )}_\lambda $.
The second term in $\eqref{union}$ is the Shannon entropy of the couplings $S(\mathsf{ptr})$. It is an entirely classical entropy and does not play a role in determining the QES. To see this more explicitly, note that \eqref{union} is a formula for the exact entropy of a boundary subsystem, but in a holographic system we could also consider the reduced density matrix $\rho_{A \cup \mathsf{ptr}}$ of the semiclassical quantum fields in some bulk region $A$. This density matrix will have the same form as \nref{densitypo}, so a similar formula holds for the semi-classical bulk matter entropy:
\eqn{ S_\text{matter}(A \cup \mathsf{ptr}) = \sum_{\lambda} P({\lambda}) S(\rho^A_{\lambda})-\sum_{\lambda} P({\lambda}) \log P({\lambda}), \quad \rho_{\lambda}^A = \Pi_{\lambda} \rho_A^\text{semi} \Pi_{\lambda}. \label{eq:matterent}}
The second term is independent of the choice of $A$, so it will never contribute to the derivative of the matter entropy (relevant for the QES).
The conclusion is that the entanglement wedge of ${\mathsf{sys}} \cup \mathsf{ptr}$ can be diagnosed by computing $S({\mathsf{sys}}|\mathsf{ptr}) = S({\mathsf{sys}} \cup \mathsf{ptr}) - S(\mathsf{ptr})$, which is simply the quenched entropy. In this context, we have a QES formula for the conditional entropy:
\def\mathsf{matter}{\mathsf{matter}}
\eqn{S({\mathsf{sys}}|\mathsf{ptr}) = \mathrm{min} \; \mathrm{ext}_I \left [ \mathrm{Area}(\partial I)/(4G) + S_\mathsf{matter}(I|\mathsf{ptr}) \right ] \label{eq:matterent2} .}
To summarize, if we are interested in the entanglement wedge of ${\mathsf{sys}}$ with no knowledge of the couplings whatsoever, we should compute the annealed entropy $S(\ev{\rho_\psi}_{\lambda})$. If we are interested in bulk reconstruction when we have perfect classical knowledge of the couplings ${\mathsf{sys}} \cup \mathsf{ptr}$, we should compute the quenched entropy $\ev { S(\rho_\psi) }_{\lambda}$.
Note that whether we choose to entangle ${\mathsf{sys}}$ with a reference or classically correlate ${\mathsf{sys}}$ to a pointer makes no difference for the density matrix $\rho_{\mathsf{sys}}$. A crucial difference is for the complement of ${\mathsf{sys}}$: the density matrix of $\rho_\mathsf{ptr}$ is diagonal in the $\lambda$ basis whereas $\rho_{\mathsf{journal}}$ has off-diagonal elements. Furthermore, the entanglement wedge of ${\mathsf{journal}}$ can contain an island (as we will show), whereas a quick argument \cite{Qi:2021sxb} rules out any possible island for the entanglement wedge of $\mathsf{ptr}$: for the pointer system, the matter conditional entropy is positive
$S(A |\mathsf{ptr}) = S(A \cup \mathsf{ptr}) - S(\mathsf{ptr}) \ge 0$ for any bulk region $A$, so adding any the bulk region $A$ can only increase the entropy. Therefore, any island QES must be non-minimal.
\begin{comment}
\AAl{Older version}: Obtaining definite knowledge of the values of a subset of couplings amounts to the acting with a projection operator on the reference. For instance, knowing the first coupling is represented by the operation
\begin{align}
|\Psi \rangle \langle \Psi| \rightarrow \Pi_{\lambda_1} \, |\Psi \rangle \langle \Psi| \Pi_{\lambda_1}
\end{align}
Some of the models we will study below are simpler to analyze when we average over the values of the known couplings. This amounts to summing over the possible results of the $\lambda_1$ measurement in the example above:
\begin{align}
|\Psi \rangle \langle \Psi| \rightarrow \sum_{\lambda_1} \Pi_{\lambda_1} \, |\Psi \rangle \langle \Psi| \Pi_{\lambda_1}
\end{align}
The same result can be obtained by introducing a pointer system $\mathsf{ptr}$ which is entangled with the couplings, and then is traced out. Note that following this procedure, the system has no entanglement with the $\lambda_1$ register of the reference or with the (traced out) pointer system individually, but only classical information.
\end{comment}
\subsection{Both known and unknown couplings}
\def\mathsf{unknown}{\mathsf{unknown}}
\def\mathsf{known}{\mathsf{known}}
\def\kappa{\kappa}
\def{\mu}{{\mu}}
More generally, we can imagine that all the couplings in ${\mathsf{journal}}$ can be divided into $\mathsf{known}$ and $\mathsf{unknown}$. We write ${\lambda} = \{ \kappa, {\mu}\}$ with $\kappa$ the known parameter(s) and ${\mu}$ the unknown, ${\mathsf{journal}} = \mathsf{known} \, \cup \, \mathsf{unknown}$. The global state is a density matrix on ${\mathsf{sys}} \cup {\mathsf{journal}}$:
\eqn{
\rho = \sum_{\kappa, {\mu}, {\mu}'} \sqrt{P({\mu}, \kappa) P({\mu}', \kappa)}\ketbra{\psi; \kappa, {\mu}}{\psi; \kappa, {\mu}'}_{\mathsf{sys}} \otimes \ketbra{\kappa}{\kappa}_\mathsf{known} \otimes \ketbra{{\mu}}{{\mu}'}_\mathsf{unknown}
\label{eq:joint}}
Here we are treating the known couplings as ``classical'' pointer states $\mathsf{ptr}$, whereas the unknown parameters are entangled.
The reduced density matrices
\eqn{
\rho_{ {{\mathsf{journal}}}} = \sum_{ {\mu}, {\mu}',\kappa } \ketbra{ {\mu}}{{\mu}'}_{\mathsf{unknown}} \ketbra{\kappa}{\kappa}_\mathsf{known} \sqrt{P\left ( {\mu}, \kappa \right ) P\left ( {\mu}', \kappa \right ) } \, \langle \psi ; {\mu}, \kappa | \psi ; {\mu}', \kappa \rangle_{\mathsf{sys}} \label{eq:rho_unk}
}
\eqn{
\rho_{ {\mathsf{known}}} = \sum_{\kappa } \ketbra{\kappa}{\kappa}_\mathsf{known} P(\kappa), \quad P(\kappa) = \sum_\mu P(\mu, \kappa) \label{eq:rho_k}
}
We see that the density matrix of $\mathsf{known}$ is completely classical as before.
Once we include ``known'' couplings, the global state is not pure. Therefore, $S({\mathsf{journal}}) \ne S({\mathsf{sys}})$. Nevertheless, a close analog is given by the conditional entropies:
\boxedeq{eq:cond}{S({\mathsf{sys}}|\mathsf{known}) = S(\mathsf{unknown}|\mathsf{known}).}
This follows from \eqref{eq:rho_unk}.
As we have already argued, the entanglement wedge of ${\mathsf{sys}} \, \cup \, \mathsf{known}$ may be diagnosed by computing the conditional entropy $S({\mathsf{sys}} \cup \mathsf{known})$, and similarly for the entanglement wedge of $\mathsf{known} \cup \mathsf{unknown}$. The equality of the above conditional entropies essentially shows that the entanglement wedges are in fact complementary even though the global state is not pure.
The fact that we are including the known couplings when computing the entanglement wedges reflects the fact that the known couplings are completely classical, so there is no obstruction to cloning. The experimentalist simply publishes in a journal the values of the couplings in which she prepared her system.
\pagebreak
\section{The entanglement wedge with uncertain couplings \label{matterentropy}}
In this section we analyze the candidate entanglement wedges of the reference keeping track of the unknown couplings. This will be done in a bulk model of JT gravity \cite{Almheiri:2014cka,Jensen:2016pah,Maldacena:2016upp,Engelsoy:2016xyb} coupled to general conformal matter:
\eqn{-I[g] =- S_0 \chi +\int_{\Sigma_{2}} \frac{\phi}{4 \pi}(R+2)+\frac{\phi_{b}}{4 \pi} \int_{\partial \Sigma_{2}} 2 K+\log Z_{\mathsf{CFT}}[g],}
where $\chi$ is the Euler characteristic.
We will specialize to a concrete BCFT when needed.
The couplings in this model will be the choice of CFT boundary conditions along the AdS boundary, which we label $\ket{J}$ and interpret as arising from a holographic boundary Hamiltonian $H_J$. We require the state $\ket{J}$ to be a conformally invariant Cardy state.
Here $J$ could be a discrete variable (e.g., if the CFT is a minimal model) or continuous. We will study the state given by entangling the thermofield double to the ${\mathsf{journal}}$ as in \eqref{global:pure}:
\begin{align}
| \Psi \rangle_{{\mathsf{sys}} \, \cup \, {\mathsf{journal}} } &= \sum_{J} \sqrt{P(J)} | \beta+2iT, J \rangle_{{\mathsf{sys}}} | J \rangle_{{\mathsf{journal}}} \label{eq:comp}
\end{align}
where the time-evolved thermofield double $\ket{\beta+ 2iT ,J}_{\mathsf{sys}} = e^{-iH_L T}\ket{\beta,J}_{\mathsf{sys}} $. We will frequently drop the $2iT$ and think of $\beta$ as a complex number. The reduced density matrix
\begin{align}
\rho_{\mathsf{journal}} = \sum_{J, J'} \sqrt{P(J) P(J')} \langle \beta, J' | \beta, J \rangle \ | J \rangle \langle J' |_{\mathsf{journal}}. \label{rho_unk}
\end{align}
To compute the entanglement wedge of the journal, we will use the QES formula \cite{Faulkner:2013ana, Engelhardt:2014gca}:
\eqn{S({\mathsf{journal}}) =\operatorname{min} \left\{ \operatorname{ext}_{I} \left[ \sum_{\partial I} {\phi(\partial I)}{}+ S_{m}(I \cup {\mathsf{journal}})\right] \right\} \label{eq:qes}.}
Here we are instructed to compute the generalized entropy of extremal islands, and then pick the smallest extremized entropy. Note that for marginal deformations of the boundary theory, we expect the dilaton to be independent of the boundary state $J$. Hence the non-trivial calculation is just the matter entropy in \eqref{eq:qes}. We will compute this in some special cases in this section.
Before getting into the weeds, let us clarify somewhat the interpretation of the calculation.
We are thinking of the above theory as an effective theory of the bulk. It is not UV complete due to divergences when wormholes get narrow \cite{Saad:2019lba}.
We do not know what the precise boundary dual of this bulk model is (or even if it exists).
The simplest possibility is that the boundary dual is a theory with a small number of couplings.
For example, one might be able to embed such a setup in a traditional higher dimensional example of AdS/CFT by considering near extremal black holes in AdS.
In this case, we assume that all of the couplings besides $J$ are fixed, and the state of the combined system is given by \eqref{eq:comp}.
However, it is also possible that the gravity description only arises after a disorder average over other random couplings, like in pure JT gravity \cite{Saad:2019lba} or in SYK \cite{Maldacena:2016hyu}.
In this case, the above \eqref{eq:comp} and \eqref{rho_unk} are not quite right; instead, we assume that the additional couplings are ``known'' while the $J$ couplings are unknown, and use \eqref{eq:joint}.
The density matrix of ${\mathsf{sys}} \cup \mathsf{unknown}$ would be
\eqn{
\rho_{{\mathsf{sys}} \, \cup \, \mathsf{unknown}} = \sum_{\kappa, J, J'}P(\kappa) \sqrt{P( J ) P(J')}\ketbra{\beta; \kappa,J }{\beta; \kappa, J'}_{\mathsf{sys}} \otimes \ketbra{J}{J'}_\mathsf{unknown}
\label{eq:jointt}}
Then following \eqref{eq:matterent} we would interpret the QES computation as giving the conditional entropy $S(\mathsf{unknown}| \mathsf{known})$:
\eqn{S(\mathsf{unknown}|\mathsf{known}) =\operatorname{min} \left\{ \operatorname{ext}_{I} \left[ \sum_{\partial I} {\phi(\partial I)}{}+\ev{ S_{\text {m}}(I \cup \mathsf{unknown} )}\right] \right\} \label{eqn:qes2}.}
Here we have used $S_{\text {matter }}(I \, \cup \, \mathsf{unknown} | \mathsf{known}) = \ev{S_{\text {matter }}(I \cup \mathsf{unknown})}$, where $\ev{ \cdots} $ is a disorder average with respect to the known couplings, see \eqref{eq:matterent} and \eqref{eq:matterent2}.
\begin{comment}
conditions, or boundary states, of the bulk matter CFT. We will label these by $J$, and interpret these as arising from different boundary Hamiltonians $H_J$.
In this section we consider JT gravity $+$ a matter BCFT.
We will assume that the matter CFT has various boundary states, which we label by $J$. We assume that the boundary dual of the theory has a different Hamiltonian $H_J$ for each $J$. We will entangle the state of the boundary theory $\ket{\beta+iT, J}$ with a journal, which records $J$.
To find the entanglement wedge of either the boundary system or the journal, we use the QES prescription. This requires an understanding of the matter entropy of an interval. In our setup, even when the interval includes all of AdS, the matter entropy does not vanish: the bulk semiclassical state is not pure because it is entangled with the journal.
Our first order of business is therefore to compute the matter entropy of (1) the entire AdS and (2) an interval in AdS.
Suggested outline for this section:
\begin{itemize}
\item Searching for QES
\begin{itemize}
\item dilaton profile (this can be moved to the same section of applying the QES formula. There's a question about how to split things for zero and non-zero temperature.)
\item matter entropy in general
\item applying QES prescription, case by case (details of cases can be left to the appendix)
\end{itemize}
\item Renyi entropies, examples with irrelevant couplings that would be missed in a naive application of the island formula
\end{itemize}
new suggested outline:
\begin{itemize}
\item Perfect reconstructability absent perfect knowledge is inconsistent with unitarity. Show the unitarity paradox with a general CFT. Check with a compact boson.
\item Island contribution. General setup, OPE approximation, and check with compact boson.
\end{itemize}
\AAl{My take:}
In this section we consider \AAl{a System composed of} JT gravity $+$ a matter BCFT.
We will assume that the matter CFT has various boundary states, which we label by $J$. \AAl{We take this as model of having a different boundary} Hamiltonian $H_J$ for each $J$. We will entangle the state of the boundary theory $\ket{\beta+iT, J}$ with a journal, which records $J$.
\AAl{We consider a family of states of the combined System plus journal describing various levels of knowledge of the boundary conditions, which we write as}
\begin{align}
\rho = \sum_{J_0, J} P_0(J_0) \sqrt{P(J) P(J')} | \beta + i T, J_0 + J \rangle \langle \beta - i T, J_0 + J|_{\mathsf{sys}} \otimes |J_0\rangle \langle J_0|_{\mathsf{known}} \otimes |J \rangle \langle J' |_{\mathsf{unknown}}
\end{align}
\AAl{This state describes perfect knowledge of the coupling $J_0$ since its Hilbert space factor is completely decohered. $P_0(J_0)$ is the probability of having found the value $J_0$. Our ignorance is restricted to the value of $J$, the degree of which is controlled by the probability distribution $P$. A quick measure of this is the Shannon entropy $-\int dJ P(J) \ln P(J)$. We will specify $P_0$ and $P$ in a case by case basis.}
Our first task is to show that any ignorance is detrimental to bulk reconstruction when using only the boundary system. Since full bulk reconstruction requires the entire bulk be contained within the entanglement wedge of the boundary, we probe this failure by showing that the trivial surface cannot be the dominant contribution the entropy of the \AAl{journal}. We will find a Hawking-like result of a continuously growing entropy leading to the usual bags-of-gold issues in the bulk.
We then show that there is an island contribution to this entropy, which we find via applying the QES prescription of extremizing the dilaton and the entropy of the bulk matter. We find...
We end this section on explcit computations of Renyi entropies, etc...
\HL{One question about this section: the bulk model we wrote down (JT + CFT) is not obviously dual to an ensemble. I am not sure if we want to assume it is either dual to a theory with fixed couplings or not.}
\AAl{I see your point. The current perspective (reflected above) is that each boundary condition is dual to a single Hamiltonian. However, that being said, we could also think of each boundary condition being dual to a different ensemble. Thus, averaging of the boundary condition is like averaging over one of the couplings in the total ensemble (the hidden couplings + the boundary condition). Does that make sense?}
\HL{I am tempted to suggest the following. In this section, we simply ignore the hidden couplings and just assume that there is a single coupling dual to the boundary condition of the free boson. We can then add some comments at the end of the section saying that our calculations also apply if there are hidden couplings. Then in the next section, we can be a bit more explicit about the hidden couplings, since they are more important in SYK.}
\end{comment}
\subsection{Inconsistency of the trivial surface \label{trivialsurface}}
Here we analyze the contribution from the trivial surface, namely where the entanglement wedge of the reference doesn't include any part of the gravitational system. The entire bulk is encoded on the boundary. We will find that this contribution to the entanglement between the boundary and the reference leads to an ever growing entropy, producing to a Hawking-like information paradox. This signals that at late times, the couplings should contain an island, to prevent the entropy from growing.
To compute the entropy of the trivial surface, all we need is the bulk matter entropy. The most straightforward way of getting this is by computing the entropy of the density matrix \eqref{rho_unk} using the replica trick while freezing the gravitational saddle to be the product of Euclidean cigars on the $n$ copies. The $n$-th Renyi entropy\footnote{In the standard quantum information literature, the Renyi's usually refer to $S_n = \frac{1}{1-n} \log R_n$; here we will refer to both $S_n$ and $R_n$ as Renyi entropy.} of the journal is therefore given by the path integral with the boundary conditions $Z_n = \tr \rho^n_\mathsf{unknown}$
\eqn{Z_n \quad &= \quad \eqfig{0.7\columnwidth}{renyiNall.pdf} \\
&= \quad \int \prod_{i = 1}^n P(J_i) \, d J_i \ \prod_{i = 1}^n \langle \beta, J_i | \beta, J_{i+1} \rangle,
}
where $J_{n + 1} = J_1$.
In the above drawing, we are supposed to sum over all the indices, weighted by the probability distribution $P(J_i)$.
The overlaps can be represented as the BCFT disk partition function in the presence of boundary changing operators.
\begin{align}
\langle \beta , J_i | \beta, J_{i+1} \rangle = \ev{O_{J_i,J_{i+1}}(0) O_{J_{i+1},J_i}({\tau}) }_\mathsf{disk}, \quad \tau = \beta/2.
\end{align}
where $O_{J_i,J_{i + 1}}(\tau)$ changes the boundary condition by $J_{i + 1} - J_i$ as $\tau$ is crossed along the path integral contour. They satisfy $\left( O_{J_i,J_{i + 1}}(\tau)\right)^\dagger = O_{J_{i+1},J_{i }}({-\tau})$, and behave as primary operators with a dimension that depends on the boundary conditions $\Delta[{J_{i + 1}, J_{i}}]$, and their two point function is
\begin{align}
\ev{O_{J_i,J_{i+1}}(0) O_{J_{i+1},J_i}(\tau) }_\mathsf{disk} = \left[ {\pi \epsilon \over \beta \sin (\pi \tau / \beta)} \right]^{2 \Delta[{J_{i + 1} , J_{i}}]}
\end{align}
Here the $\epsilon$ comes from the Weyl factor evaluated on the boundary.
Instead of consisting a product of $n$ disks, we may equivalently consider the quotient picture where we instead consider the $n$-fold tensor product of the CFT on a single disk, with only 2 boundary condition changing operators:
\eqn{Z_n \quad &= \quad \eqfig{0.3\columnwidth}{renyiquotient-all.pdf}
}
\defO_{\{J\}}{O_{\{J\}}}
\noindent In this picture, the black dot represents a composite boundary condition changing operator $O_{\{J\}} = O_{J_1, J_2} \otimes O_{J_2, J_3} \otimes \cdots \otimes O_{J_n, J_1}$ which shifts the boundary condition $\{J_1, J_2 \cdots J_n\} \to \{J_2, J_3 \cdots J_1\} $.
This quotient picture is a bit overkill for the no-island computation, but we are introducing it now since it will be crucial when the QES is non-trivial.
Since we are interested in the time evolution of the entropy, we need to consider the Renyi computation in the time evolved state
$ | \beta + 2 i T, J \rangle = e^{-i H_{ J}^L T} | \beta, J \rangle$.
We can achieve this by setting $\tau = \beta/2 + i T$. The overlaps become
\begin{align}
\langle \beta + 2 i T , J_i | \beta + 2i T , J_{i+1} \rangle =\left[ {\pi \epsilon \over \beta \cosh (\pi T / \beta)} \right]^{2 \Delta[{J_{i + 1} , J_{i}}]}
\end{align}
Notice that all off-diagonal matrix elements are decaying to zero at large $T$; only the diagonal terms where $\Delta = 0$, e.g., $J_i = J_{i+1}$ do not decay. This implies that at late times the matter is close to maximally mixed.
Putting this back into the Renyi entropy, we get
\begin{align}
Z_n = \int \prod_{i = 1}^n d J_i \, p(J_i) \ \left[ {\pi \epsilon \over \beta \cosh ( \pi T / \beta)} \right]^{\sum_{i=1}^n 2 \Delta[{J_{i + 1} , J_{i}}]} \label{eq:renyi}
\end{align}
So the only dynamical input we need from the particular CFT to evaluate these Renyi entropies is the boundary dimensions $\Delta[{J_{i + 1}, J_{i}}]$.
In Appendix \ref{app:2bd} we consider a generic BCFT with 2 boundary states. Here we will consider a non-compact boson, with action
\eqn{S = {1 \over 2 \pi \alpha'} \int d^2 z \partial X \bar{\partial} X \label{eq:freebosonaction}}
The boundary conditions we consider are simply Dirichlet conditions on the free field, labeled by $X$. The boundary condition changing operator\footnote{A quick way to see that changing Dirichlet conditions behaves as a local boundary primary is to use T-duality, which relates this setup to the insertion of a vertex operator with momentum $k$ on the boundary (with standard Neumann condtions.) We thank Juan Maldacena for pointing this out.} changes the value of the field from $X_1$ to $X_2$. Its dimension is equal to the energy on the theory on the strip, which is given by
\eqn{\Delta_b = {1 \over 4 \pi \alpha'} \int_0^\pi (\partial_\sigma X)^2 d \sigma = {1 \over \alpha'} \left ({X_1 - X_2 \over 2\pi}\right )^2 \label{eq:bcc} }
This is familiar from string theory. The mass of a bosonic open string of level $N=1$ which stretches between two D-branes is $M^2 = (X_1 - X_2)^2/(2\pi \alpha')^2$.
We will take a Gaussian measure over the boundary conditions $p(X) = \frac{m}{\sqrt{2\pi}} e^{-m^2 X^2/2}$.
In terms of the boundary dual, we are considering an ensemble of boundary Hamiltonians parameterized by a marginal coupling $J$ that changes the Dirichlet condition of the bulk field $X$; the ensemble for this boundary coupling is Gaussian distributed.
Then Equation \eqref{eq:renyi} becomes
\eqn{ Z_n &= \left ( \frac{1}{\sqrt{2\pi} } \right )^n \int \prod_{i=1}^n {dx_i} \exp \left [ - \frac{1}{2} \left ( x_i^2 +a (x_i - x_{i+1})^2 \right ) \right ] \\
&=\left ( \det M_n \right )^{-1/2}, \quad a = {1 \over 2\pi^2 \alpha' m^2} \log \left [ \frac{\beta}{\epsilon \pi} \cosh \frac{\pi T}{\beta} \right ], \quad x_i = m X_i, \label{eq:noncompact}}
Here $M_n$ is an $n \times n$ matrix; to take $n=5$ for example:
\eqn{M_5 = \left(
\begin{array}{ccccc}
2 a+1 & -a & 0 & 0 & -a \\
-a & 2 a+1 & -a & 0 & 0 \\
0 & -a & 2 a+1 & -a & 0 \\
0 & 0 & -a & 2 a+1 & -a \\
-a & 0 & 0 & -a & 2 a+1 \\
\end{array}
\right)}
Here the determinant is a polynomial in $a$:
\eqn{\det M_n = \sum_{k=0}^{n-1} {2n-k \choose k } \frac{n}{2n-k} a^k \label{eq:detpoly} }
This sum can be analytically continued to complex $n$:
\eqn{\det M_n = \left ( d_+^{2}\right )^n + \left ( d_-^2\right )^n-2 a^n, \quad d_\pm = {1\pm \sqrt{4 a+1} \over 2 } \label{eq:carlson} }
One can show that for integer $n$ the Taylor series of this expression in $a$ reproduces the polynomial \eqref{eq:detpoly}. Furthermore, this function satisfies Carlson's theorem, since it grows exponentially in $n$ on the real axis, but only oscillates along the imaginary axis.
Then differentiating and taking the $n\to 1$ limit of the normalized Renyi entropy we get
\boxedeq{eq:replica}{S_m &=-\left.\partial_{n} Z_{n} \right|_{n=1} = \frac{\log (a)}{2}+\sqrt{4 a+1} \coth ^{-1}\left(\sqrt{4 a+1}\right)}
At late times, $a \gg 1$ so\footnote{Note that we could derive the large time expansion by simply taking $\det M \approx n^2 a^{n-1}$.}
\eqn{S_m \approx \frac{1}{2} \log a +2 \sim \frac{1}{2} \log \left ( \frac{T }{\beta \alpha' m^2} \right ) .}
In general, the entropy monotonically increases. This growth results from states with different boundary conditions becoming more orthogonal under time evolution. If we had $c$ independent free bosons, each with uncorrelated boundary conditions, the entropy would simply be $c$ times the above answer.
This entropy eventually competes with the thermal entropy of the wormhole, which at low temperatures is $S = 2 S_0 + O(1/\beta)$, and produces a unitarity problem when
\eqn{T/\beta \sim 4 \pi \alpha' m^2 e^{4 S_0/c -4}.}
This signals that the trivial surface eventually becomes subdominant in the full non-perturbative calculation of the entropy of the unknown couplings entropy, or equivalently of the system.\footnote{Note that whether we choose to regulate the problem by compactifying the boson $X \sim X+ 2\pi R$ or by changing the measure so that $X^2 \sim 1/m$, the net result is quite similar. See Appendix \ref{windings}. } It suggests that some part of the bulk might become inaccessible to the boundary system, falling out of it's entanglement wedge as a result of the uncertainty in the couplings.
It is also interesting to note the dependence of this time scale on the uncertainty in the couplings. For that, we use the Shannon entropy of the dimensionless variable $(\alpha')^{1/2} X$, which for a Gaussian distribution i
\eqn{S_\mathrm{Sh}( X) = \frac{c}{2} \left ( 1+ \log\frac{2\pi}{\alpha' m^2}\right ) }
Recall that $m$ is the inverse width of the distribution, and the increase of $S_\mathrm{Sh}(X)$ as $m$ decreases is consistent with there being more uncertainty. The time scale is
\eqn{T/\beta \sim e^{{2 \over c} [2S_0 - S_\mathrm{Sh}(X)]}.}
Thus, the transition time is sooner for higher uncertainty. This can be used to make the transition time less than exponential in the entropy, and seemingly as small as desired by considering a wide enough distribution for the couplings.
\subsection{Matter entropy for semi-infinite interval \label{sec:peninsula}}
\def\bar{y}{\bar{y}}
\def {\bar{w}} {\bar{w}}
Before addressing the matter entropy of an island, let us consider the matter entropy of a semi-infinite interval (a ``peninsula''). This may be viewed as a warmup to the island calculation, which we will see reduces to the peninsula calculation at late times. It is also the relevant computation for the entanglement wedge of the left side of the boundary system, tracing out the right side and the ${\mathsf{journal}}$. If the journal is empty, the QES lives at the bifurcate horizon; with a journal, we expect the QES to shift to the left.
To compute the matter entropy, we will follow the same strategy as above. First note that the left density matrix
\begin{align}
\rho_L = \sum_J P(J) \frac{e^{-\beta H_J}}{Z_J(\beta)}
\end{align}
For a free boson, the disk partition function $Z_J(\beta)$ is independent of $J$, so we can ignore the denominator (it will just contribute to an overall normalization of the density matrix). More generally, for a BCFT on a disk one can we write $Z_J(\beta) = g(J) Z(\beta)$, where $\log g(J)$ is the ``boundary entropy'' which is the boundary analog of the central charge \cite{Cardy:2004hm}.\footnote{ $Z(\beta)$ is generated from the conformal anomaly, which only depends on the central charge.} Therefore we can absorb $g(J)$ into the definition of $P(J)$.
Then the computation of the unnormalized density matrix $\tr \rho_L^n$ is given by
\eqn{Z_n \quad &= \quad \raisebox{-0.3\totalheight}{\includegraphics[width = 0.75\columnwidth]{renyiNdisk.pdf}}
}
\noindent This is a path integral on a cone with boundary total length $\beta n $. Each segment of length $\beta$ has some (generically different) boundary conditions $J_i$. So there are $n$ boundary condition changing operators on the disk. We can quotient the picture to obtain:
\eqn{Z_n \quad &= \quad \eqfig{0.25\columnwidth}{renyiquotient.pdf}
}
\noindent In the quotient picture, the CFT is again the $n$-fold tensor product of the seed theory, now with only a single boundary condition changing operator $O_{\{J\}}$ and a twist operator $\sigma$. We will compute the matter entropy as a function of the position of the twist operator. The virtue of the quotient picture is that whereas before we had an $n$-pt function on the disk, now we only have a bulk-to-boundary 2-pt function, which is fixed by conformal symmetry:
\eqn{\ev{\sigma(z,\bar{z}) O_b(y) }_\mathsf{disk}
\propto \frac{ 1 }{(|1-|z|^2)^{2h-h_b} |1 -z \bar{y}|^{2h_b} } }
We can obtain this by mapping the disk to the upper half plane, and then using the doubling trick to relate the correlator to a chiral three-pt function on the plane.
This is the 2-pt function on a flat disk. We are interested in the \pcr{} disk
$ds^2 = 4 dz d\bar{z} /(1-|z|^2)^2 $ with a circular boundary of circumference $\beta/\epsilon$ at $(1-|z|^2)/2 = 2\pi \epsilon/ \beta$, so a Weyl transformation leaves us with
\eqn{\ev{\sigma(z,\bar{z}) O_{\{J\}}(\theta') }_{\mathsf{AdS}} = c_n(\{J\}) \left ( \frac{2\pi \epsilon}{\beta } \frac{ (1-|z|^2) }{ |1 -z \bar{y}|^{2} }\right )^{h_b}. \label{eq:twist1} }
Note that this expression is invariant under rotational (boost) symmetry $z \to z e^{i\theta}, y \to y e^{i \theta}$.
In our problem, the boundary condition changing operator is on the left side at $\bar{y} = -1$ and $z$ is real.
The dependence on $X$ and comes in via both $h_B$ and the BOE coefficient $c_n(\{ J \})$.
In principle one can evaluate this coefficient, which is related to an $n$-pt function of boundary vertex operators on a disk:
\eqn{ c_n(\{J\}) = \prod_{i < j}^n \left| w_i-w_j\right|^{-2\left ( \frac{X_i - X_j}{2\pi} \right )^2/\alpha' } }
where $w_k = e^{2\pi i (k-1)/n}$. Together $h_b \propto -2\left ( \frac{X_i - X_j}{2\pi} \right )^2/\alpha'$, \eqref{eq:twist1} reduces to ``just'' a Gaussian integral. Nevertheless, the $c_n(\{J\})$ factors lead to a sufficiently complicated determinant that analytically continuing the answer in $n$ is not easy. However, defining
\def\tilde{a}{\tilde{a}}
\eqn{\tilde{a}&= {1 \over 4\pi^2 \alpha' m^2} \log \left ( \frac{\beta}{2\pi \epsilon } \frac{ 1+z }{1-z } \right ) ={d_\mathsf{twist} \over 2\pi^2 \alpha' m^2} ,}
where $d_\mathsf{twist} $ is the distance from the boundary to the twist operator.2
we expect that when $\tilde{a} \gg 1$, the integral will be dominated by the second factor in \eqref{eq:twist1}, which is much more sharply peaked in $X^2$. The entropy with then be of the form \eqref{eq:noncompact} where $a \to \tilde{a}$.
The entropy in the limit of large $\Tilde{a}$ is
\begin{align}
S_m &\approx {1 \over 2} \ln \Tilde{a} + ...
\end{align}
At fixed $\alpha' m^2$, this answer is expected\footnote{The only possible complication is that the $n\to 1$ limit does not commute with $\tilde{a} \to \infty$} to be valid when $z+1 \gg \epsilon^{1/2}$, e.g., as long as the twist operator is far from the boundary cutoff.
To find the QES, we need the profile of the dilaton, which in these coordinates is
\eqn{\phi = \frac{2\pi \phi_r}{\beta} \frac{ 1 + |z|^2}{1-|z|^2}}
In general, there will be a QES to the left of the horizon as long as $\phi_r/\epsilon \gg 1$, e.g., as long as $\phi_r$ is fixed in the $\epsilon \to 0$ limit. This is shown in figure \ref{extqes}. Balancing out the derivatives of the $\phi$ and $S_m$ places the QES at
\begin{align}
z_\mathsf{QES} = -{\beta \over 2 \pi \phi_r \ln \beta/(2\pi \epsilon)}
\end{align}
This is just outside the left horizon. To leading order in $1/\Tilde{a}$ we find no dependence of the location of the QES on the degree of uncertainty of the couplings $m^2$.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.6]{1sided-bulkentropy.pdf}
\caption{ Generalized entropy $S_m + \phi - \phi_0$ as a function of the real coordinate $z$. Here $S_m$ is the matter entropy of an interval $[-1+\epsilon^{1/2},z]$. For large values of $\phi_h$, there is a QES near the horizon. As $\phi_r$ gets smaller, the QES shifts closer to the left side.
\label{extqes}}
\end{center}
\end{figure}
\def {\bar{w}} { {\bar{w}} }
\subsection{Entropy of the island \label{sec:island}}
\def\Delta_{\partial}{\Delta_{\partial}}
Let us finally consider the matter entropy of an island $S_m(I \, \cup \, {\mathsf{journal}})$.
Again consider the $n$-th Renyi entropy of the semiclassical theory.
Like in the case of the trivial surface, there are 2 boundary condition changing operators; now, there are also 2 twist operators in the bulk:
\eqn{Z_n \quad &= \quad \eqfig{0.25\columnwidth}{renyiquotient-island.pdf}
}
One can ask what happens when we bring one of the twist operators very close to a boundary operator. This is a bulk-boundary OPE (sometimes referred to as a BOE) limit. More explicitly, we can expand both twist operators in terms of boundary primaries and their descendants. Then we are left with a boundary 4-pt function. In the limit that two of the points are close, only the boundary identity will contribute:
\eqn{\langle O_{\{J\}}(y_1) \sigma(z_1, \bar{z}_1) \sigma(z_2, \bar{z}_2) O_{\{J\}}^\dagger(y_2)\rangle \quad &= \quad \eqfig{0.35\columnwidth}{boe.pdf}
}
Let us consider the entropy of a finite interval with both endpoints near the horizon at very late times. This is a bulk-boundary OPE limit: the distance to the horizon is fixed whereas the length of the wormhole is growing linearly with time. So we expect the above four point function to factorize:
\eqn{\langle O_{\{J\}}(y_1) \sigma(z_1, \bar{z}_1) \rangle \langle \sigma(z_2, \bar{z}_2) O_{\{J\}}^\dagger(y_2)\rangle \quad }
The Renyi entropy, however, doesn't factorize, but is instead given by correlated sum of the above product:
\eqn{\int \prod_{i = 1}^n d J_i \, p(J_i) \ \langle O_{\{J\}}(y_1) \sigma(z_1, \bar{z}_1) \rangle \langle \sigma(z_2, \bar{z}_2) O_{\{J\}}^\dagger(y_2)\rangle \quad }
In the frame where we evolve in time symmetrically on the left and right, the problem has a symmetry that interchanges the left and right sides. This allows us to consider instead the quantity
\eqn{\int \prod_{i = 1}^n d J_i \, p(J_i) \ \langle \sigma(z_2, \bar{z}_2) O_{\{J\}}^\dagger(y_2)\rangle^2 \quad }
This is identical to the single interval case, but with $\Tilde{a} \rightarrow 2 \Tilde{a}$. This minor modification does not change the location of the QES. See Figure \ref{fig:EC}.
This case demonstrates the strong sensitivity of the black hole interior to the values of the couplings. The Page-like transition gives a limit to the allowed uncertainty in the couplings, as measured by the entropy between the journal and the boundary, after which the entanglement wedge snaps and the interior falls outside the entanglement wedge of the boundary.
\pagebreak
\section{Renyi Entropies in gravity and SYK \label{sec:renyi}}
In the previous section, we discussed a model where the reconstructability absent precise knowledge of the couplings could be directly probed by finding the QES of the unknown couplings. Not all models enjoy this level solvability. However, it is sometimes easier to compute the Renyi entropy in these more general models. Thankfully, there are two signatures of the presence of a QES from the Renyi entropy. The first is a unitarity paradox where the semi-classical saddle produces a Renyi entropy too small to be consistent with the dimensionality of the Hilbert space of the system. The second signature (and what fixes the first) is the presence of Replica wormholes. In solvable models, the QES prescription falls out of the $n \rightarrow 1$ limit of the $n$-th Renyi entropy computation.
In this section, we study the $n = 2$ Renyi entropy in both JT gravity and SYK. This will be done using various methods/regimes, including the low temperature Schwarzian regime, at large $q$, and also numerically via exact diagonalization of the SYK Hamiltonian. We will again find signatures of the failure of reconstruction due to insufficient knowledge of the couplings. We will also consider the case where the unknown couplings correspond to an irrelevant deformation of the system.
\subsection{Disk contribution}
As an intermediate step for the Renyi computation, we will compute elements of the journal density matrix. In any holographic theory, this is determined by a computation like
\eqn{\bra{J} \rho \ket{J'} \quad &\propto \quad \eqfig{0.25\columnwidth}{renyi2disk.pdf} \label{eq:disk}
}
where we fill in some gravity solution (gray) with the appropriate boundary conditions that correspond to evolution by $H + \chi(u) O_\Delta$. $\chi(u)$ takes the value $J$ for $0<u < \tau$ (blue) and the value $J'$ (red) for the other part of the circle $\tau < u < \beta$, which determines the boundary conditions of the bulk scalar field $\chi$. By taking $\tau = \beta/2+iT$ one can study the matrix elements as a function of real time $T$.
To make further progress, we will now assume that field dual to $O_\Delta$ is a free field in AdS$_2$. We can then integrate out the field to get an effective matter action
\eqn{ -I_m = D\int du_1 \, du_2 \left [ \frac{t'(u_1) t'(u_2)}{2 \sin^2 \left ( \frac{t(u_1) - t(u_2)}{2}\right ) } \right ]^{\Delta} \chi_r(u) \chi_r(u') \label{eq:matter} }
where $D=\frac{\left(\Delta-\frac{1}{2}\right) \Gamma(\Delta)}{\sqrt{\pi} \Gamma\left(\Delta-\frac{1}{2}\right)}$.
In computing the overlap, we should divide by the norms of the states $\sqrt{\braket{J}{J} \braket{J'}{J'}}$; but these are not time dependent and do not play an important conceptual role.
Following the discussion of the free boson in the previous section, we will now specialize to the case of a marginal deformation $\Delta = 1$, where the computation simplifies significantly and the full gravitational backreaction can be computed.
For one thing, we may shift the $J \to J + a$, $J' \to J' + a$ without changing the value of the overlap, so without loss of generality we may set $J'=0$. In Appendix \ref{app:matter} we evaluate this integral for the special case of a marginal deformation $\Delta = 1$, giving an answer that is remarkably bi-local in $u_1$ and $u_2$:
\eqn{e^{-I_m} &= \left [ \frac{\epsilon^2 t'(u_1) t'(u_2)}{ \sin^{2} \left ( \frac{t(u) -t(0)}{2}\right ) }\right ]^{\delta}, \quad \delta = \frac{(J-J')^2}{2\pi} \label{eq:matt2} }
where $\epsilon$ is a UV regulator. This precisely agrees with the dimension of the boundary condition changing operator we found in \nref{eq:bcc} if we set $\alpha' = 1/(2\pi)$ so that the scalar is canonically normalized. Note that the final form is consistent with a picture where we have inserted a bulk ``domain wall'' of mass $\sim \delta$ that separates the $J$ and $J'$ vacua. The domain wall is where the gradient of the bulk field is appreciable; it is not a thin wall.
This matter action is valid for an off-shell $t(u)$. In other words, we can include the gravitational backreaction by integrating over $t(u)$ with the Schwarzian action appropriate for the disk. This is the same action we would have gotten from inserting two local operators of NCFT$_1$ dimension $\delta$ at $u=0$ and $u$. For a free scalar field in the bulk, these expressions combined with the results for the Schwarzian $n$-pt functions \cite{Mertens:2017mtv, Lam:2018pvp, Saad:2019pqd, Yang:2018gdb} give us the exact disk contribution, summing over all quantum fluctuations of the boundary mode.
The fact that we have an exact quantum expression for the Renyi entropy gives us confidence that there is really an information paradox if we just focus on the disk. Although in the classical approximation, the disk contribution decays exponentially, without the quantum expressions, we would not be confident that the disk decays to a value smaller than what is required by unitarity $\sim e^{-S_0}$. Indeed, the quantum modifications show that the exponential decay is replaced by a power law decay $\sim T^{-3}$, with the exponent independent of $\delta$.
Here we would also like to comment that the above result also applies to SYK in the Schwarzian limit. We will use the following conventions for SYK:
\eqn{H_\mathsf{SYK}=i^{q / 2} \!\!\! \!\!\! \sum_{1 \leq i_{1} \ldots \leq i_{a} \leq N} \!\!\!\!\!\!\! J_{i_{1} \ldots i_{q}} \psi_{i_{1}} \ldots \psi_{i_{q}}, \quad\left\langle J_{i_{1} \ldots i_{q}}^{2}\right\rangle=\frac{N J^{2}}{q {N \choose q}}=\frac{N \mathcal{J}^{2}}{2 q^2 {N \choose q} \label{eq:sykdef} },
\quad\left\{\psi_{i}, \psi_{j}\right\}=2 \delta_{i j}}
To mimic the above expressions, we can imagine turning on a deformation by another SYK Hamiltonian,
\eqn{H=H_\mathsf{SYK}(J_1,q_1) + \chi(u) H_\mathsf{SYK}(J_2,q_2) \label{syksum} }
Here $J_1, J_2$ are distributed like in \eqref{eq:sykdef}. We can introduce a $G,\Sigma$ action by following the usual steps, only integrating out both $J_1$ and $J_2$ to obtain an action $I_0 + I_m$, where
\eqn{-I_0/N &= \log \operatorname{Pf}\left(\partial_{t}-{\Sigma}\right)-\frac{1}{2} \int d \tau_{1} d \tau_{2}\left[{\Sigma}\left(\tau_{1}, \tau_{2}\right) {G}\left(\tau_{1}, \tau_{2}\right)-\frac{J^{2}}{q} {G}\left(\tau_{1}, \tau_{2}\right)^{q} \right]\\
&\approx \frac{ \alpha_{S}}{\mathcal{J}} \int d u \left\{\tan \frac{\pi t(u)}{\beta}, u\right\}\\
-I_m/N &= \frac{J^2 }{2q'} \int du_1 \, du_2 \, \left [ G(u_1, u_2) \right ] ^{q'} \chi(u_1) \chi(u_2)\\
&\approx \int du_1 \, du_2 \, \left [ \frac{t'(u_1) t'(u_2)}{2 \sin^2 \left ( \frac{t(u_1) - t(u_2)}{2}\right ) } \right ]^{q'/q} \chi_r(u_1) \chi_r(u_2), \quad \chi_r = \chi(u) b^{q'/2} \label{eq:sykaction}
}
where $J^2 b^q \pi = \left ( \frac{1}{2} - {1 \over q} \right ) \tan (\pi/q)$.
Here $I_0$ comes from $J_1$ and $I_m$ comes from the $J_2$ couplings. The relative normalization of the two terms is controlled by the magnitude of $\chi$.
Now in the low temperature, large $N$ limit $I_0[G,\Sigma] \to \mathsf{Sch}[t(u)]$ is replaced by the Schwarzian action, where we integrate over $t(u)$ instead of $G,\Sigma$. Similarly, the action $I_m$ can be re-written in terms of the Schwarzian mode.
Note that in writing these expressions, we also assume that $\chi$ is small so that we can simply integrate over the near-zero mode. This is a similar approximation to what is discussed in \cite{Maldacena:2018lmt}.
We expect that the disk partition function at early times to be self-averaging both in $J_1$ and $J_2$. Therefore, this computation has multiple interpretations. The first interpretation is that we draw some particular choice of $J_2$ and view the term $H_\mathsf{SYK}(J_2)$ as a single operator that is deforming the original theory. The journal records not the values of $J_2$ but merely a single coupling $\chi_0$ which is the overall normalization of the deformation. Then the above action would govern the matrix element of the journal density matrix, e.g., an overlap between $\mathsf{TFD}$'s with different values of $\chi_0$. Since the disk answer (for early times) is self-averaging, we can average over $J_1, J_2$ in this computation and derive the above effective action in terms of $G,\Sigma$.
A second physically different setup is when $\chi_0$ is fixed, and the journal instead records the values of all ${N \choose q'}$ couplings $J_2$.
Then we think of $J_1$ as ``known'' or fixed. We will see in the next section that \eqref{eq:sykaction} will still be relevant for computing Renyi entropies, or the overlaps averaged over $J_2$.
\subsection{Renyi-2 wormhole}
\def\bar{\tau}{\bar{\tau}}
\def\mathcal{C}{\mathcal{C}}
The Renyi-2 entropy of the journal is given by
\eqn{\tr \rho^2_{\mathsf{journal}} = \int p(J)p(J') |\bra{J} \rho \ket{J'}|^2.}
As an intermediate step, we should compute the squares of density matrix elements.
If we plug in our answer for the density matrix elements \eqref{eq:matter}, there will be a unitarity paradox at late times. In particular, $\tr \rho^2_{\mathsf{journal}}$ will decay to zero, which is impossible since it is bounded by $2^{-N}$.
The resolution to this paradox is that the squares of the density matrix elements have an additional contribution given by wormholes. In a theory with other ``known'' couplings, this is an acceptable resolution, but such wormholes would violate factorization in a theory with fixed couplings.
To obtain the norm of the overlaps, we start with a Euclidean computation real $\tau, \bar{\tau}$ and then analytically continuing $\tau = \beta/2 + iT$, $\bar{\tau} = \beta/2-iT$:
\eqn{|\braket{\beta + 2 i T, J}{\beta + 2i T, J'} |^2 &= \tr \left ( e^{-\bar{\tau} H_J}e^{- \tau H_{J'}} \right ) \tr \left ( e^{- \tau H_J}e^{-\bar{\tau} H_{J'}} \right ) \\
&\supset \quad \eqfig{0.3\columnwidth}{renyi2wormhole.pdf} \label{eq:wormhole}}
In addition to the disk topology which we have already discussed, we have drawn a wormhole contribution. We can think of this wormhole as being supported by the bulk ``domain walls'' that separate the $J$ and $J'$ region. Such a wormhole is closely related to the one described by Douglas Stanford in Appendix B of \cite{StanfordMore}.
Let us outline the steps to obtaining the wormhole, leaving a more thorough discussion to Appendix \ref{app:matter}. We will start with the double trumpet geometry, with a cutout parameterized by two boundary times $T_L(u_L)$ and $T_R(u_R)$. First, we turn on a source $\chi_L$ on the left side from times in $(0,\tau)$ and a source $\chi_R$ which is on from times in $(\tau,\beta)$. This leads to a matter action that has $LL$ correlators, $RR$ correlators, as well as $LR$ cross terms.
\def{\mathsf{wormhole}}{{\mathsf{wormhole}}}
\eqn{ -I_m^{\mathsf{wormhole}} &= D\int du_1 \, du_2 \Bigg\{ \left [ \frac{T_L'(u_1) T_R'(u_2)}{\cosh^2 \left ( \frac{T_L - T_R}{2}\right ) } \right ]^{\Delta} \chi_L(u_1) \chi_R(u_2) +\\
& \left [ \frac{T_L'(u_1) T_L'(u_2)}{\sinh^2 \left ( \frac{T_L(u_1)- T_L(u_2)}{2}\right ) } \right ]^{\Delta} \chi_L(u_1) \chi_L(u_1)+ \left [ \frac{T_R'(u_1) T_R'(u_2)}{\sinh^2 \left ( \frac{T_R(u_1) - T_R(u_2)}{2}\right ) } \right ]^{\Delta} \chi_R(u_1) \chi_R(u_2) \Bigg\} }
In principle, one should evaluate this integral for off-shell $T_L,T_R$, and then find the classical solution.
For $\Delta = 1$ the integral can be performed but the answer is a bit complicated. However, large Lorentzian times $T$, a sensible ansatz is that the distance between the quench sites on opposite sides of the wormhole and $u_2$ and $u_3$. In the above picture \eqref{eq:wormhole}, we are saying that the domain walls that cross the wormhole on the ``front'' and ``back'' sides have fixed length at large $T$. With this ansatz, one can show that the above matter integral becomes the insertion of two domain wall operators:
\eqn{e^{-I_m} = \mathcal{C}^{LR}(u_1, u_4) \mathcal{C}^{LR}(u_2, u_3), \quad \mathcal{C} = \left [ \frac{T_L'(u_1) T_R'(u_2)}{\cosh^2 \left ( \frac{T_L - T_R}{2}\right ) } \right ]^{\delta}.}
Since the overall effect of the matter is local, the on-shell solution must again be semi-circular arcs, with junctions at the quench sites. Therefore the wormhole can be obtained by starting with two copies of the disk geometry in \eqref{eq:disk}, one for the ``front side'' of the wormhole and the other for the ``back side.''
Then one cuts both of these disks and pastes them together along two geodesics so that there is no bulk discontinuity.
The key equation we will need is the effective energy of the solution, or equivalently the circumference $\beta_E$ of the pieces of the disk that we use to make the solution. This is the same energy on all pieces of the solution away from the quench sites.
It is determined by imposing $\mathrm{SL}(2,\R)$ charge conservation at the quench sites, see \ref{app:matter}. The result is that for any $T$,
\eqn{\tan \left(\frac{\pi \beta}{2 \beta_{E}}\right)=\frac{\delta \beta_{E}}{2 \pi}.\label{eq:tanb} }
For small values of $\delta \beta$, we get $\beta_E^2 = \pi^2 \beta/\delta$.
Several aspects of the wormhole geometry are discussed in \ref{app:matter}; here we will just check that the wormhole resolves the unitarity problem. To do so in a semi-classical approximation, we need to evaluate the on-shell value of
\eqn{\exp \left ( {-I^{\mathsf{wormhole}}} \right ) \mathcal{C}^{LR}(u_1, u_4) \mathcal{C}^{LR}(u_2, u_3).}
Let us now evaluate $I^{\mathsf{wormhole}}$ in the limit of small $\delta \beta$ and show that this resolves the potential unitarity paradox at large $T$. There are two Schwarzian boundaries in the computation, both of which have the same action. The left boundary consists of two arcs:
\eqn{
-I_\mathsf{Sch} &= 2 \int du_L \{ \tan \frac{\pi t(u)}{\beta}, t\} = 2 \left [ \tau + \bar{\tau} \right ] \left ( \frac{\pi }{\beta_E}\right )^2 \\
&= 2 \pi^2 \beta / \beta_E^2.
}
The correlators are essentially on opposite sides of an effective thermofield double $\beta_E$, so they take thermal values:
\eqn{Z_\text{wormhole} \sim \exp \left ( {- 4 \pi^2 \beta / \beta_E^2} \right ) \beta_E^{-4\delta}. }
We see that the classical action approaches a value independent of $T$ at late times, which resolves the potential unitarity paradox for the Renyi-2 entropy.
More generally, from the results in Appendix \ref{app:matter}, we can see that the solution at general $iT$ will give an action independent of $T$ as both the distance across the wormhole and the action, including corner contributions, will be time-independent.
\subsection{Numerics for finite $N$ SYK }
\defn_\mathrm{trials}{n_\mathrm{trials}}
In this section, we resort to the numerical analysis of the $n = 2$ Renyi entropy of the journal by exact diagonalization of the SYK Hamiltonian at finite $N$. We will present evidence of the wormholes. The ``known" couplings in all of our setups will be in a standard SYK Hamiltonian with ${N \choose q}$ couplings. The ``unknown'' couplings will be additional unknown couplings.
We argued around \eqref{eq:cond} that the relevant entropy that diagnoses the entanglement wedge given the ``known'' couplings is the relative entropy $S(\mathsf{unknown} | \mathsf{known})$. The ``conditional'' Renyi entropy whose $n\rightarrow 1$ limit is equal to this conditional entropy is given by
\eqn{\tilde{R}_n &= \sum_{\kappa} P(\kappa) \sum_{{\mu}_1, \cdots, {\mu}_n} \prod_{i=1}^n P({\mu}_i) \braket{ \psi ; {\mu}_i, \kappa }{ \psi ; {\mu}_{i+1}, \kappa}_{\mathsf{sys}}.}
\eqn{ S(\mathsf{unknown} | \mathsf{known}) = -\partial_n \tilde{R}_n\big|_{n=1}. }
An efficient numerical way to estimate the above quantity is to Monte Carlo sample the quantity $\prod_{i=1}^n \braket{ \psi ; {\mu}_i, \kappa }{ \psi ; {\mu}_{i+1}, \kappa}_{\mathsf{sys}}$ from the distribution $P(\mu_1) \cdots P(\mu_n) P(\kappa)$. The conditional Renyi $\tilde{R}_n$ has the appealing feature that it is linear in $P(\kappa)$, as opposed to the standard Renyi entropy of the full journal, which comes with a $P^n(\kappa)$ factor.
We will now employ this method in studying $\tilde{R}_2$ for various ``unknown'' deformations of the SYK model. The first case we will consider is the $q=4$ SYK, where the couplings are known with only limited precision. To model this, we decompose the usual SYK Hamiltonian into two terms:
\eqn{ H = \left ( K_{i jkl } + J_{ijkl}\right ) \psi_i \psi_j \psi_k \psi_{l}, \quad }
Here $K$ are the known couplings, and $J$ are the unknown couplings (the ones stored in the journal). We can think of this setup as modeling a situation where the usual SYK couplings are known but with some Gaussian errors. This is similar to \eqref{syksum}, except that there we were viewing the single coupling as an overall normalization of the second term; here, there are ${N \choose 4}$ couplings in the journal. The conditional $n = 2$ Renyi entropy is given by
\eqn{\tilde{R}_2 &= \int dJ_0 \, p(J_0) \int dJ dJ' \, p(J) p(J') \, |\ev{\beta/2+iT; J_0, J|\beta/2+iT; J_0,J'}|^2.}
We would like to test whether wormholes contribute to the above calculation. To do so, we consider the following ``disk'' approximation to $\tilde{R}_2$:
\eqn{\tilde{R}_2^\mathsf{disk} &= \Bigg|\int dJ_0 dJ dJ' \,p(J_0) p(J) p(J') \, \ev{\beta/2+iT; J_0, J|\beta/2+iT; J_0,J'}\Bigg|^2.}
In this approximation, we have averaged first, before squaring. In the large $N$ analysis, only disconnected solutions can contribute to the above quantity. If $\tilde{R}_2 $ is significantly larger than $\tilde{R}_2^\mathsf{disk} $, we interpret this as evidence that a wormhole is dominating over the disconnected solution.
To Monte Carlo sample these integrals, we draw two independent SYK Hamiltonians $H_1$ and $H_2$ from Gaussian distributions. Then $H_\pm = (\cos \theta) H_1 \pm (\sin \theta) H_2$ are also normalized Hamiltonians such that $\ev{\tr H_\pm^2 } = J^2, \quad \ev{\tr H_+ H_-} = \lambda^2 J^2$, where $\lambda^2 = \cos (2\theta)$.
The results are displayed in Figure \ref{fig:2}. The disk approximation is a good one at small $JT$ but at large values, we see that it significantly underestimates $\tilde{R}_2$. In fact, at very late times, the underestimate is so bad that even without computing $\tilde{R}_2$ exactly, we can rule it out on the grounds that it would violate unitarity.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.7]{figs_couplings/diagonalization-randomJ.pdf}
\caption{ The conditional Renyi $\tilde{R}_2$ as a function of Lorentzian time evolution $J T$, obtained by numerically diagonalizing the SYK Hamiltonian for $N=20$, $\beta J = 16$ and $n_\mathrm{trials} = 3000$. $1 \sigma$ error bars are displayed. The solid curve is the full answer while the dashed curve is the ``disk'' approximation $\tilde{R}_2^\mathsf{disk}$. We see that the disk approximation is good for early times, but decays too rapidly at late times. Even without computing the full answer (in orange), the unitarity bound $\tr \rho^2 \ge 2^{-N}$ would rule out the disk answer at large times. \label{fig:2}}
\end{center}
\end{figure}
We also consider deforming a standard $q = 4$ SYK Hamiltonian with a {\it single} unknown coupling $\chi$:
\eqn{ H = \left ( K_{i jkl } \right ) \psi_i \psi_j \psi_k \psi_{l} + i^{q_2/2} \left ( \psi_1 \cdots \psi_{q_2}\right ) \chi . }
Here $K_{ijkl}$ are drawn from a Gaussian as usual with a variance set by \eqref{eq:sykdef}; we also choose $\chi$ to be normally distributed, with a variance $\ev{\chi^2} = \epsilon^2 J^2$.
This may be thought of as a variant of the problem in \eqref{syksum}. The results are shown in figure \ref{fig:q4q6}.
We see clear evidence of a wormhole, even when $q_2 > 4$. In fact, at the moderate values of $N$ that we computed, there does not seem to be a large qualitative difference between $q_2=4$ and $q_2>4$. This is perhaps surprising since in the conformal regime, $q_2 > 4$ would correspond to an {\it irrelevant} coupling whereas $q_2 = 4$ would be marginal.
To further test this interpretation, we compute the $G_{\mathsf{LL}}$ correlators numerically and see that they qualitatively agree with the predictions given in the subsequent section on the large $q$ wormhole, see Figure \ref{fig:l0c} {for the case of many marginal couplings} and \ref{fig:q6c} {for the case of a single irrelevant coupling}. We also computed $G_{\mathsf{LR}}$ numerically and saw that it is nearly zero at early times $Jt \sim 1$ but becomes appreciable at times that correspond to the wormhole/disk transition. More specifically, in Figures \ref{fig:l0c} and \ref{fig:q6c} we compare
\eqn{ |G_{\mathsf{LL}}^\mathsf{disk}| &= \frac{1}{N'}\sum_i \frac{\overline{ \left| \bra{ \beta+2 iT, J} \psi_i(\beta/2) \psi_i(t) \ket{\beta+2i T,J'} \right|} }{\overline{|\braket{\beta+2iT,J}{\beta+2iT,J'}|} },\\
|G_{\mathsf{LL}}^\mathsf{WH}| &= \frac{ \overline{ \bra{ \beta+2 iT, J} \psi_i(\beta/2) \psi_i(t) \ket{\beta+2i T,J'}\braket{\beta+2iT,J'}{\beta+2iT,J} }}{ \overline{|\braket{ \beta+2 iT, J'}{\beta+2i T,J}|^2} }, }
where in the above expressions the overbar indicates averaging with respect to $J$ and $J'$. {For the single coupling case, the fermion index $i$ is summed over all $N' = N - q_2$ fermions except those appearing in the deformation, whereas in the case of many marginal couplings, the index $i$ is summed over all $N' = N$ fermions.}
Of course our limited numerics cannot conclusively test the conformal regime ($N \to \infty$, $\beta \mathcal{J} \gg 1$) where the term ``irrelevant'' has a sharp meaning. Nevertheless, we conjecture that in the conformal/JT regime, there exist wormhole solutions for irrelevant couplings; we hope to report progress in this direction in a future publication.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.6]{q4q6.pdf}
\caption{$N = 20$, $n_\mathrm{trials} = 1000$, $\beta J = 16, \epsilon = 0.1$. We show $q_2 = 4,6$ deformations and $1\sigma$ error bars. Note that $q_2=6$ is an irrelevant deformation, but the curves seem qualitatively similar to the marginal case. \label{fig:q4q6}}
\end{center}
\end{figure}
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.6]{l0c.pdf}
\caption{The 2-pt correlator $|G_{LL}|$ as a function of Lorentzian time for the $\lambda^2=0$ disk and wormhole for the case of many marginal couplings. We take $N = 20$, $n_\mathrm{trials} = 600, \beta \mathcal{J} = 1, JT=10^3$. Notice that as $t \to T$ we reach the second quench site. For the disk solution, a prediction from the large q analysis is that the correlator becomes large on the disk, whereas it remains small on the wormhole. This seems to be in rough agreement with the modest $N,q= 4$ numerics. In computing the correlator, we divide by the average norm $\braket{\beta+2iT,J}{\beta+2iT,J'}$ for the disk and $|\braket{\beta+2iT,J}{\beta+2iT,J'}|^2$ for the wormhole. \label{fig:l0c}}
\end{center}
\end{figure}
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.6]{f5v2.pdf}
\caption{
The 2-pt correlator $|G_{LL}|$ as a function of Lorentzian time, where a single irrelevant $q=6$ parameter is varied, as in Figure \ref{fig:q4q6}. We take $N = 20$, $n_\mathrm{trials} = 2000, \beta \mathcal{J} = 1, JT=10^5$. The correlators are qualitatively similar to the case with many marginal parameters in Figure \ref{fig:l0c}.
\label{fig:q6c}}
\end{center}
\end{figure}
\subsection{Large $q$ SYK \label{sec:largeq} }
\def\hat{\mu}{\mu}
\defg{g}
\def {\hat{g} }{ {\hat{g} }}
\def\Omega_{\tau}{\Omega_{\tau}}
\def\beta_{E}{\beta_{E}}
In this section, we will consider the disk contribution to the Renyi-2 entropy in the $N\to \infty$ large $q$ SYK model. We will show that the disk contribution decays to zero at large Lorentzian times, in conflict with unitarity. This suggests that there should be a wormhole that dominates at late times. We show that the wormhole in the low temperature limit is closely related to the Schwarzian wormhole described in the gravity section.
\begin{figure}[H]
\begin{center}
\includegraphics[width=0.5\columnwidth]{drawing2.pdf}
\caption{ Contour relevant for matrix element of journal. This can be obtained by considering a red segment of length $\tau$ and a blue segment of length $\beta - \tau$, and then setting $\tau = \beta/2 + i T$. \label{fig:drawing2} }
\end{center}
\end{figure}
We will consider the disk solution for the thermal circle of length $\beta$, with quench sites at times $0$ and $\tau$. The couplings are constant in the segments $(0,\tau)$ and $(\tau, \beta)$ and are partially correlated between these two segments.
Let us assume the correlation between the couplings on the two sides is $\lambda^2 = e^{- \mu}$. When $\lambda= 1$ there is no change in the coupling at the quench sites; when $\lambda = 0$ the two sides are uncorrelated. The large $q$ equations of motion are
\eqn{
\partial_1 \partial_2 g_\mathsf{toc} + 2 \mathcal{J}^2 e^{g_\mathsf{toc}} = 0, \quad &t_1, t_2 \text{ on same side}\\
\partial_1 \partial_2 g_\mathsf{otoc} + 2 e^{-\hat{\mu}} \mathcal{J}^2 e^{g_\mathsf{otoc}} =0, \quad &t_1, t_2 \text{ on opposite sides} \label{eq:verbose} }
We will sometimes use the subscripts $(\mathsf{toc}, \mathsf{otoc})$ to emphasize when an equation applies when both times are on the same side or on opposite sides of the quench.
It is convenient to define
\eqn{
\Omega_\tau &=\Theta(\tau - t_1) \Theta(t_2 - \tau) + \Theta(\tau-t_2) \Theta(t_1 - \tau)\\
\Omega_\mathsf{toc} &= 0, \quad \Omega_\mathsf{otoc} = 1.
}
Then, we can summarize \eqref{eq:verbose} as
\eqn{
\partial_1 \partial_2 g + 2 \mathcal{J}^2 \exp \left ( {g- \hat{\mu}\Omega_{\tau} } \right ) = 0.}
In addition to the usual UV boundary condition $g(\tau, \tau) = 0$, we impose continuity at the quench site:
\eqn{g(t,-\epsilon) &= g(t,+\epsilon) \\
g(t,\tau-\epsilon) &= g(t,\tau+\epsilon) . \label{eq:cont}}
These boundary conditions follows from the fact that for very short times $\epsilon$, we can neglect the interactions from $\mathcal{J}$.
A useful trick is the following. Consider the field redefinition
\eqn{g = {\hat{g} } + \mu \Omega. \label{hatg} }
Then the equations of motion and the boundary conditions are
\eqn{
\partial_1 \partial_2 {\hat{g} } &+ 2 \mathcal{J}^2 e^ {\hat{g} } =0,\\
{\hat{g} }_\mathsf{toc} &\to {\hat{g} }_\mathsf{otoc} - \mu .
}
By using the variable $ {\hat{g} }$ the equations of motion is uniform over the entire circle, at the price of a discontinuous boundary condition.
This discontinuous boundary condition is considered in \cite{Qi:2018bje, Streicher:2019wek, Eberlein:2017wah}. For general $\tau$, see Streicher \cite{Streicher:2019wek}. In our conventions,
\begin{align}
e^{g_\mathsf{toc}} &= \left( {\alpha_1 \over \mathcal{J} \sin \left[ \alpha_1 t_{12} + \gamma_1 \right]} \right)^2 , \quad 0 < t_1 < t_2 < \tau \label{eqn:gtop} \\
e^{g_\mathsf{toc}} &= \left( {\alpha_2 \over \mathcal{J} \sin \left[ \alpha_2 t_{12} + \gamma_2 \right]} \right)^2, \quad t_3 < t_1 < \bar{\tau} \label{eqn:gbot} \\
e^{g_\mathsf{otoc}} &= \left ( { \alpha_1 \alpha_2 \mathcal{J}^{-2} \over \lambda^2 \sin \left[ \alpha_1 (t_1 -\tau) \right]\sin \left[ \alpha_2 (t_2 -\bar{\tau}) \right] - \sin \left[ \alpha_1 (t_1 -\tau) + \gamma_1 \right] \sin \left[ \alpha_2 (t_2 -\bar{\tau}) - \gamma_2 \right]} \right )^2. \label{eqn:gotoc}
\end{align}
We also have the UV boundary conditions:
\begin{align}
&\alpha_i = {\cal J} \sin \gamma_i \\
&\sin \left( \frac{ \alpha_1 \tau \pm \alpha_2 \bar{\tau}}{2} + \gamma_1 \pm \gamma_2 \right) = \lambda^2 \sin \left( \frac{ \alpha_1 \tau \pm \alpha_2 \bar{\tau} }{2} \right)\label{streich}
\end{align}
Here we are imagining solving the above equations for real $\tau, \bar{\tau}$ and then analytically continuing and setting $\tau = \beta/2+iT, \bar{\tau} = \beta/2-iT$.
The physical branch of the above equations is the one that is continuously connected to the $T=0, \lambda = 1$ solution.
In figure \ref{fig:gamma} we solve for $\gamma$ numerically at $T=0$ for varying $\beta \mathcal{J}$. We also solve for $\gamma$ as a function of $T$ for fixed $\beta \mathcal{J}$ in figure \ref{fig:gamma2}.
It is interesting to study these constraints for small $\gamma \approx \alpha/\mathcal{J}$:
\eqn{ -\frac{\gamma_1 \pm \gamma_2 }{1-\lambda^2} \approx \tan \left ( \frac{\alpha_1 \tau \pm \alpha_2 \bar{\tau}}{2} \right ). \label{eq:tana}}
This can be compared with a classical Schwarzian analysis for the insertion of a heavy operator of dimension $\delta$ on the disk, which gives (see \cite{Goel:2018ubv} equation 2.9):
\eqn{-\frac{k_1 \pm k_2}{\Delta} = \tan \left ( \frac{k_1 \tau \pm k_2 \bar{\tau} }{4 C} \right ). \label{eq:tanSch} }
This means that we may identify $k_i/2C = \alpha_i $ and $(1-\lambda^2) 2 C \mathcal{J} = \delta$. Here $C$ is the coefficient of the Schwarzian action, which for large $q$ SYK is $C \sim N /\mathcal{J} q^2$. In the Schwarzian theory, $k_i = 2\pi C/(\beta_E)_i$, where $\beta_E$ is the length of the circle that is pieced together to form the disk solution.
The left hand side of the above equation is of order $(\beta \Delta)^{-1} \sim (\beta \mathcal{J} (1-\lambda^2))^{-1}$. So for large $\beta \Delta$, we have $\tan \left [ \frac{1}{2} \left ( \alpha_1 \tau \pm \alpha_2 \bar{\tau}\right ) \right ] = 0$. The correct branch is to take $\alpha_1 \tau +\alpha_2 \bar{\tau} = 2 \pi$ and $\alpha_1 \tau - \alpha_2 \bar{\tau} = 0$, which yields
\eqn{\alpha_1 = \frac{ \pi}{\tau}, \quad \alpha_2 = \frac{\pi}{\bar{\tau}} . \label{eq:pinch}}
This has a pleasing geometric interpretation for real $\tau, \bar{\tau}$. As we have noted, $\alpha_i = \pi/(\beta_{E})_i$ where $\beta_{E}$ is the effective inverse temperature (e.g. correlators on the same side of the quench agree with thermal correlators at a temperature $\beta_E^{-1}$.) The solution tells us to set $(\beta_{E})_1 = \tau$ and $(\beta_{E}) = \bar{\tau}$. This means the geometry has the form of two circles which are completely ``pinched.'' Indeed, if we plug in this solution, we can compute the correlator \eqref{eqn:gtop} at $t_1 = 0$ and $t_2 = \tau$. We see that $e^g= 1$, its maximal value. From the Schwarzian point of view, the domain wall is so massive that it pinches the disk together. {This is depicted in Figure \ref{fig:infbeta}.}
\begin{figure}[H]
\begin{center}
\includegraphics[width=0.7 \columnwidth]{infinitebetapinch.pdf}
\caption{ Left: the ordinary extremal or zero temperature solution. Since the boundary particle intercepts the asymptotic boundary, the total proper time is infinite. Right: if we take $\beta = \infty$ while holding the mass of the brane constant, the solution pinches. This means that the correlator between the quench sites is getting large. In the large $q$ theory, we showed that the correlator in fact becomes maximal $e^{g} = 1$. \label{fig:infbeta} }
\end{center}
\end{figure}
This should be contrasted with the case when $\Delta \beta \ll 1$ in which case $\alpha_i = 2 \pi/\beta$. So for $\tau =\bar{\tau}$ large but general $\lambda$, we expect that $\pi /(\beta \mathcal{J}) \le \gamma \le 2\pi /(\beta \mathcal{J})$. These bounds are shown in figure \ref{fig:gamma}.
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{gamma-lambda.pdf}
\caption{ Here we show $\gamma$ as a function of $\lambda$ for $\beta\mathcal{J} = 200$ and $\beta \mathcal{J} = 800$ in blue and orange respectively. We also show in dashed lines the upper and lower bounds on $\gamma$ obtained analytically (in a $1/\beta \mathcal{J}$ expansion.) Note that for larger $\beta \mathcal{J}$, $\gamma$ quickly obtains its maximal value $ 2\pi/\beta$ away from $\lambda = 1$. \label{fig:gamma} }
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{gamma-pinch.pdf}
\caption{ $\gamma_2 = \gamma_1^*$ as a function of $\mathcal{J} T$ for $\beta\mathcal{J} = 100$. On the left, $\lambda = 0.9$, whereas on the right $\lambda = 0$. Solid lines are obtained by numerically solving equation \eqref{streich}. The dashed lines are the ``pinching'' approximation derived in \eqref{eq:pinch}. At large $\beta \mathcal{J}$, pinching is a reasonable approximation even when $1-\lambda^2$ is not small (left). \label{fig:gamma2} }
\end{center}
\end{figure}
We can use our results for $\gamma_1 = \gamma_2^*$ to compute the 2-pt correlator $e^{g(\beta, \tau)}$ as a function of $\tau$, see figure \ref{fig:correlator}. This will be needed for the evaluation of the on-shell action for the Renyi-2 entropy in the next subsection.
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{figs_couplings/disk-correlator-largeq.pdf}
\caption{ Here we show the disk contribution to the 2-pt correlator $e^{g(t_1,t_2)}$ evaluated at the quench sites ($t_1 = 0$, $t_2 = \beta/2+iT$) for $\lambda = 0.3,0.4,0.5$, and $\beta\mathcal{J}=5$. Note that at small values of $\lambda$ the correlator stays large, which indicates that the distance between the quench sites is staying relatively small, even at large Lorentzian times. This is qualitatively similar to the Schwarzian behavior where the geometry gets pinched by the brane. \label{fig:correlator} }
\end{center}
\end{figure}
\begin{comment}
For $\lambda = 0$,
\begin{align}
\left ( \alpha_1 \pm \alpha_2 \right ) \frac{\beta}{2} + 2 \left ( -\alpha_1 \pm \alpha_2 \right ) i T + \gamma_1 \pm \gamma_2 = n^\pm \pi
\end{align}
For $T= 0$, we have $\alpha_1=\alpha_2$. Comparing to the previous case, $n^+ = 1$ and $n^- = 0$. We may solve this equation for small $\gamma$,
\eqn{ \gamma_1 = \frac{\pi}{2 + (\beta -2iT)\mathcal{J} }, \quad \gamma_2 = \frac{ \pi}{(\beta+2iT)\mathcal{J}}. }
This approximation is valid if $\beta \mathcal{J} \gg 1$.
In our conventions, their solution give
\eqn{
\exp \left ( {g_\mathsf{toc}}\right ) &= \left ( \frac{\alpha }{\mathcal{J} \sin \left ( \alpha (t_1 -t_2) + \gamma \right ) }\right )^2 \\
\exp \left ( {g_\mathsf{otoc}}\right ) &=
\left[ e^{-g_\mathsf{toc}/2} -\frac{ \mathcal{J}^2 \left(1 - \lambda^{2} \right) }{\alpha^2 } \sin \left(\alpha t_1 \right) \sin \left(\alpha t_2\right)\right]^{-2 }
}
Here the physical correlator is $G = {1 \over N}\ev{\psi_i(t) \psi_i(0)} = G_0 e^{g/q}$.
\eqn{ \alpha = \mathcal{J} \sin \gamma, \quad \sin (\alpha \beta/2 + 2 \gamma) = \lambda^2 \sin (\alpha \beta/2)\label{uvbd}
}
Note that when $\lambda=1$, we just get $g_\mathsf{toc} = g_\mathsf{otoc}$ as expected.
The equations \eqref{uvbd} have many solutions. Working in units where $\mathcal{J} = 1$, we are solving $\sin( \sin \gamma \beta/2 + 2 \gamma)) = \lambda^2 \sin(\beta/2 \sin \gamma )$. The physical solution corresponds to the smallest positive $\gamma$ solution.
As a simple case, consider $\lambda^2 = 0$, so that we have completely decorrelated both sides of the circle. We get $(\beta/2) \sin \gamma + 2 \gamma = \pi$. This can be solved analytically when $\beta$ is large, $\alpha = \gamma = 2\pi/\beta$. Note the factor of 2.
If we instead solve for $\lambda = 1$ at large $\beta$, we get $\alpha = \gamma= \pi/\beta$. Note that $\alpha$ is a measure of the local effective temperature, e.g., $\alpha = \pi/\beta_{E}$. In the Schwarzian picture, the solution is to cut and paste two thermal disks, each with a temperature $\beta_{E}$, so that the total boundary length is $\beta$. So when $\lambda = 0$ and $\beta_{E} = \beta/2$, the thermal circles are joined at a single point. This means the geometry is totally pinched, which is reasonable given that at $\lambda = 0$ the ``domain wall'' is infinitely massive.
Actually, the solution we just discussed for $\lambda = 0$, $\beta \to \infty$ also holds when we fix $\lambda < 1$ and take $\beta \to \infty$. This follows from \eqref{uvbd}, since $\sin(\alpha \beta /2) = 0$. From the Schwarzian point of view, we are fixing the mass of the domain wall, but the Schwarzian action scales like $\sim N/(\beta \mathcal{J})$. So at large $\beta \mathcal{J}$ the Schwarzian action always loses in the tug of war and the disk ``pinches.'' We plot $\gamma$ as a function of $\lambda$ for large $\beta$ in Figure \ref{fig:gamma}.
We can also evaluate the correlator at $t = \beta/2$ for $\lambda = 0$, but general $\beta$. We get $e^g = (\sin(\gamma)/\sin(\alpha \beta/2 + \gamma) )^2 = (\sin \gamma/\sin( \pi - \gamma))^2 = 1$. So we see that when the two pieces of the disk are completely decorrelated, the two quench sites have a maximal correlation.
From the above we can also evaluate the correlator at $t=\beta/2$ for large $\beta$ and general $\lambda<1$. By the same logic, we also have a maximal correlator.
$$
\begin{aligned}
&\frac{k_{1} \tau_{1}}{C}+2 \arctan \frac{k_{1}+k_{2}}{\ell}+2 \arctan \frac{k_{1}-k_{2}}{\ell}=2 \pi \\
&\frac{k_{2} \tau_{2}}{C}+2 \arctan \frac{k_{1}+k_{2}}{\ell}-2 \arctan \frac{k_{1}-k_{2}}{\ell}=2 \pi
\end{aligned}
$$
\end{comment}
\subsubsection{Evaluation of the action}
Now we would like to evaluate the on-shell action of the disk. We start with the large-$q$ Liouville action \cite{Cotler:2016fpe}:
\eqn{ S= {N \over 8q^2} \int dt_1 dt_2 \; \partial_1 g \partial_2 g - 4 \mathcal{J}^2 e^{-\mu \Omega} e^g }
Then in terms of $ {\hat{g} }$ \eqref{hatg}:
\eqn{ S = {N \over 8q^2} \int dt_1 dt_2 \; \partial_1 \hat{g} \partial_2 \hat{g} + \mu^2 \partial_1 \Omega \partial_2 \Omega - \mu \partial_1 \hat{g} \partial_2 \Omega - \mu \partial_1 \Omega \partial_2 \hat{g} - 4 \mathcal{J}^2 e^{\hat{g}} }
Let us consider $\partial S/\partial \mu$. Since we are evaluating the action on-shell, only the explicit $\mu$ dependence contributes:
\eqn{ -\partial S/\partial \mu &= {N \over 8q^2} \int \partial_1 \hat{g} \partial_2 \Omega + \partial_2 \hat{g} \partial_1 \Omega \\
&= -{N \over 8q^2} \int dt_1 \partial_1 {\hat{g} } (t_1, \tau) \mathrm{sgn}(t_1-\tau) + \int dt_2 \partial_2 {\hat{g} }(\tau,t_2) \mathrm{sgn}(t_2-\tau)\\
&={N \over 2q^2} \left [ {\hat{g} }(\tau,\tau) - {\hat{g} }(\beta, \tau) \right ]
}
So to compute the action, we may simply integrate the correlator as a function of $\lambda$
\def\tilde{\lambda}{\tilde{\lambda}}
\eqn{I(\tau, \bar{\tau}) = -{N \over q^2} \int_{\tilde{\lambda} = \lambda}^1 2 \tilde{\lambda}^3 d\tilde{\lambda} \log \left [ \frac{\alpha_1(\tilde{\lambda}) }{\sin (\alpha_1(\tilde{\lambda}) \tau + \gamma_1(\tilde{\lambda}) )} \right ] }
In this expression, we have subtracted the action at $\lambda = 1$, which is required when we normalize the thermofield double.
This expression can be computed numerically. We display the result in Figure \ref{fig:action}.
We see that the answer grows at late times $T$. This suggests that the disk contribution alone will lead to a violation of unitarity at late times. (It does not prove that there is a violation since in principle quantum corrections could in principle stop the growth, but based on both the quantum Schwarzian computations and the finite $N$ numerics, this seems unlikely.)
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{action-calc2.pdf}
\caption{ The large $N$ on-shell action of the disk as a function of $\mathcal{J} T$ for $\beta \mathcal{J} = 5$ for various values of $\lambda$. The growth of the action implies that the typical overlap $\braket{\mathsf{TFD},J}{\mathsf{TFD},J'} \sim e^{-I}$ is shrinking as a function of time. Smaller values of $\lambda$ lead to more decorrelated states, which agrees with the larger values of the action. The growth at late times suggests that there is a unitarity problem if we only include the disk solution. \label{fig:action} }
\end{center}
\end{figure}
\pagebreak
\subsection{Wormhole in large q SYK}
\def{\mathsf{LR}}{{\mathsf{LR}}}
\def{\mathsf{LL}}{{\mathsf{LL}}}
\def{\mathsf{RR}}{{\mathsf{RR}}}
\def{\mathsf{RL}}{{\mathsf{RL}}}
\defg_{\lr} {g_{{\mathsf{LR}}} }
\defg_{\lll} {g_{{\mathsf{LL}}} }
\def\beta_\mathrm{aux}{\beta_\mathrm{aux}}
\begin{figure}[H]
\begin{center}
\includegraphics[width=1\columnwidth]{doublecontour2.pdf}
\caption{ Contour relevant for the Renyi-2 wormhole, with our conventions labeled. The black arrows start at the points where $t_i = 0$ and they point in the direction of positive $t_i$. The correlator between points $t_1$ and $t_2$ behaves as if they are on two sides of a thermofield double. \label{figc2} }
\end{center}
\end{figure}
Here we present a preliminary exploration of the wormhole in large $q$ SYK, leaving a more thorough analysis for the future. When we introduce a second side, the large $N$ variables include a $g_{\mathsf{LR}}$. Remarkably in the standard large $q$ approximation, the $g_{\mathsf{LL}}$ and $g_{\mathsf{LR}}$ variables are uncoupled in the equations of motion:
\eqn{
\partial_1 \partial_2 g_{\lll} =- 2 \mathcal{J}^2 \exp \left ( {g_{\lll} - \hat{\mu}\Omega_{\mathsf{LL}} } \right ) , \quad \partial_1 \partial_2 g_{\lr} = 2 \mathcal{J}^2 \exp \left ( {g_{\lr} - \hat{\mu}\Omega_{\mathsf{LR}} } \right ). \label{eq:eom}}
Furthermore, the equations of motion just differ by a sign, which can be accounted for by changing the direction of time.
The basic idea is that using the same tricks as above, we will have a $g_{LR}$ that obeys a Liouville like equation. So we expect a solution that is very similar to the Schwarzian wormhole, except that we ``build'' the solution from the large $q$ disk solution. More explicitly, consider an ansatz:
\eqn{\exp\left ( {g_\mathsf{toc}^{\mathsf{LR}}(u_1,u_2)} \right ) &= \left [ {\alpha \over \mathcal{J} \sin \left[ \alpha \left ( u_1 -\tilde{u}_2\right ) + \gamma \right]} \right ]^2, \quad \tilde{u}_2 = u_2 - u_b \\
\exp \left ( {g_\mathsf{otoc}^{\mathsf{LR}}(u_3,u_2) }\right ) &=
\frac{\alpha^2}{\mathcal{J}^2 }\left[ { \sin \left ( \alpha \left ( u_3-\tilde{u}_2 \right ) + \gamma \right ) } -\frac{ \mathcal{J} \left(1 - \lambda^{2} \right) }{\alpha } \sin \left(\alpha u_3 \right) \sin \left(\alpha \tilde{u}_2\right)\right]^{-2 } \label{eq:glr}\\
\exp \left ( {g_\mathsf{otoc}^{\mathsf{LL}}(u_1,u_3) }\right ) &=
\frac{\alpha^2}{\mathcal{J}^2 }\left[ { \sin \left ( \alpha \left ( u_1-u_3\right ) + \gamma \right ) } -\frac{ \mathcal{J} \left(1 - \lambda^{2} \right) }{\alpha } \sin \left(\alpha u_1 \right) \sin \left(\alpha u_3\right)\right]^{-2 }\\
\exp \left ( {g_\mathsf{toc}^{\mathsf{LL}}(u_1,u_1') }\right ) &= \left [ {\alpha \over \mathcal{J} \sin \left[ \alpha \left ( u_1 -u_1' \right ) + \gamma \right]} \right ]^2.
}
We are adopting a convention where $u_i = 0$ corresponds to a quench site, and the second quench site on the left (right) side is at $u_1 = \tau$ ($u_2 = \tau$).
Furthermore, $u_1, u_3$ runs clockwise on the LHS, whereas $u_2$ runs counterclockwise on the RHS. So for example, the solution $g_\mathsf{toc}^{\mathsf{LR}}$ is valid when $u_1 u_2 <0$ before $u_1$ or $u_2$ cross another quench site. This accounts for the sign difference in the equations of motion of \eqref{eq:eom}.
To analytically continue, we set $u_i = it_i$.
This gives a sign convention $t_i$ illustrated in Figure \ref{figc2}.
One can confirm that the equations of motion are satisfied if $\alpha = \sin \gamma$. The remaining task is to determine $\alpha$ and the time shift $u_b$.
As a warm-up, it is useful to reconsider the Schwarzian wormhole from a slightly different perspective. There, too, the equations of motion are locally the same as for the disk solution. The main idea is that we can build the wormhole by considering the disk solution with some auxiliary temperature $\beta_\mathrm{aux}$.
The disk solution in the Schwarzian limit is obtained by taking two disks of circumference $\beta_E$ and joining them together at the quench sites.
The wormhole is constructed by cutting along the diameter of each partial disk and then gluing to another copy of the solution, see figure \ref{fig:stanfordworm} and Appendix \ref{app:r2wormhole}. Pasting together the two solutions gives a new solution that is a topological cylinder.
Now before cutting, the total length of the doubled disk solution is $2\beta_\mathrm{aux}$. When we cut along the diameter of the partial disks and join the two copies, we remove a portion of the boundary that has length $4 \times (\beta_{E}/2)$. Then requiring that the total length of the boundary is $2\beta$ (a factor of $\beta$ for each side) gives
\eqn{ 2\beta_\mathrm{aux} - 2\beta_{E} = 2\beta\label{eq:trivial} }
Now for the disk solution, we have from \eqref{eq:tanSch}
\eqn{ -\tan \left ( \frac{\beta_\mathrm{aux} \pi}{2 \beta_{E}}\right ) = \frac{4\pi C}{\delta \beta_{E} } .}
Eliminating $\beta_\mathrm{aux}$ reproduces \eqref{eq:tanb}, as desired:
\eqn{\tan \left(\frac{\pi \beta}{2 \beta_{E}}\right)=\frac{\delta \beta_{E}}{4 \pi C} \approx \frac{(1-\lambda^2) \beta_E \mathcal{J}}{2\pi} .\label{eq:tanb2} }
The last equality is based on the approximate relation derived in the previous subsection, around \eqref{eq:tana}.
Now we can follow the same strategy to find $\beta_E$ as a function of $\beta$ in the large $q$ solution. The key point is that although we do not have a clear geometric picture like in the Schwarzian case, nevertheless the $\mathsf{toc}$ correlator is exactly thermal at an effective temperature $1/\beta_{E}$, so the procedure is entirely analogous. In particular, just as we cut $\beta_{E}/2$ of the boundary particle on the partial disk, we remove $\beta_{E}/2$ of the $\mathsf{toc}$ correlator, and paste it to a second copy. If we enforce equation \eqref{eq:trivial}, this pasting procedure gives a smooth correlator.
Indeed, the special thing about removing exactly $\beta_{E}/2$ of the solution is that the $\mathsf{toc}$ 2-pt function is minimized at this time. So when we continue the 2-pt function $g_{{\mathsf{LR}}}$ onto the ``back side'' of the solution, there will be no discontinuity in any of its derivatives.
Now while \eqref{eq:trivial} holds for both the Schwarzian and the large $q$ theory, the relation between $\beta_{E}$ and $\beta$ will be modified. For the $\mathsf{toc}$ correlator with any $\alpha$, we can define an effective inverse temperature $\beta_E$:
\eqn{\frac{\alpha}{\mathcal{J}} = \cos \left ( \alpha \beta_E \over 2 \right ).}
In general, $\beta_{E}$ differs from $\beta_\mathrm{aux}$ due to the added energy from the insertion of the operator.
Together with the relation between $\beta_\mathrm{aux}$ and $\alpha$ derived in the previous subsection
\eqn{\sin \left ( \alpha \beta_\mathrm{aux}/2 + 2\gamma\right ) = \lambda^2 \sin \left ( \alpha \beta_\mathrm{aux}/2\right ), \alpha/\mathcal{J} = \sin \alpha}
and \eqref{eq:trivial}, we can eliminate $\beta_\mathrm{aux}$ and obtain $\beta_E$ as a function of $\beta$ for the wormhole solution. We show the results for in \ref{fig:betabeta}.
This solution for $\alpha(\beta_{E})$ can be substituted into \eqref{eq:glr} along with
\eqn
u_b = \beta_\mathrm{aux}/2 = (\beta+\beta_{E})/2.}
The last requirement just says that in the auxiliary disk solution, the second quench site appears after a boundary time $\beta_\mathrm{aux}/2$.
We can check that with the above requirements, all the $\mathsf{toc}$ correlators obey the correct boundary conditions, e.g. $g_{\lr} (0,0) = g_{\lr} (\tau,\bar{\tau}), g_{\lr} (0,\bar{\tau}) = g_{\lr} (\tau,0)$.
We expect that the wormhole solution discussed above to be an approximate solution to the large $N$ equations of motion when the Lorentzian time $T$ is large.
To borrow the Schwarzian language, the reason why we need this condition is that we have essentially ignored ``windings.'' In the Schwarzian description, there are multiple geodesics that connect two points; in general, a correlator will receive contributions from all of these geodesics unless the wormhole throat $b$ is very long (which happens at large $T$, see Appendix \ref{app:matter}).
To state the issue in the large $q$ formalism, note that we have described the solution in patches \eqref{eq:glr}, but actually the patches overlap. While the $\mathsf{toc}$ solutions are valid whenever the two times are on the same side of a quench, the $\mathsf{otoc}$ solutions are supposed to be valid ``near'' the quench site at $u_1 = u_3 = 0$. When the $u_i$ are near the opposite quench site, we should use a similar ansatz:
\eqn{\exp \left ( {g_\mathsf{otoc}^{\mathsf{LL}}(u_1,u_3) }\right ) &=
\frac{\alpha^2}{\mathcal{J}^2 }\Bigg[ { \sin \left ( \alpha \left ( u_1-u_3\right ) + \gamma \right ) }
-\frac{ \mathcal{J} \left(1 - \lambda^{2} \right) }{\alpha } \sin \left(\alpha (u_1-\tau) \right) \sin \left(\alpha (u_3-\tau)\right)\Bigg]^{-2 }\\
\exp \left ( {g_\mathsf{otoc}^{\mathsf{LR}}(u_3,u_2) }\right ) &=
\frac{\alpha^2}{\mathcal{J}^2 }\Bigg[ { \sin \left ( \alpha \left ( u_3-\tilde{u}_2 \right ) + \gamma \right ) }
-\frac{ \mathcal{J} \left(1 - \lambda^{2} \right) }{\alpha } \sin \left(\alpha (u_3-\tau) \right) \sin \left(\alpha (\tilde{u}_2-\tau) \right)\Bigg]^{-2 } \label{eq:glr2}\\
}
We believe that the full solution to the large $N$ equations of motion is close to a sum of the two ansatzs the large $\alpha T$ limit, with a sign appropriate for the periodicity/anti-periodicity of the correlators. Summing the two ansatz only makes sense if the correlators are small in the overlap region. For fixed $\beta \mathcal{J}, \lambda<1$ this is satisfied if $\alpha T \gg 1$, since the correlators decay exponentially in the overlapping region $e^g \sim \alpha^2 e^{-2\alpha T}$. A similar issue arises in finding the double cone in SYK \cite{Saad:2018bqo} and also for the finite temperature wormhole in the coupled SYK model see also \cite{Maldacena:2018lmt}. To check this more carefully, one would need to use a different large $q$ approximation, discussed in \cite{Maldacena:2018lmt} that is valid when the correlators $G_{\mathsf{LL}}$ are small (and therefore $|g_{\mathsf{LL}}/q| \gg 1$.) It would also be interesting to solve the finite $q$ $G,\Sigma$ equations numerically.
A regime where the solution simplifies is when $\beta \mathcal{J} \gg 1$ while keeping $\lambda<1$ fixed. The wormhole correlators become
\eqn
\exp\left ( g_\mathsf{otoc}^{\mathsf{LL}}(u_1,u_3) \right ) &= {1 \over \left ( -(1-\lambda^2) \mathcal{J}^2 u_1 u_3 + \mathcal{J} u_{13} + 1\right )^2} \\
\exp \left ( {g_\mathsf{toc}^{\mathsf{LR}}(u_1,u_2) }\right ) &= {1 \over \left ( \mathcal{J} (u_1 - u_2) + 1 \right )^2} \\
\exp \left ( {g_\mathsf{otoc}^{\mathsf{LR}}(u_3,u_2) }\right ) &= {1 \over \left ( -(1-\lambda^2) \mathcal{J}^2 u_3 u_2 + \mathcal{J} u_{32} + 1\right )^2} \label{eq:syklowtemp}
}
Notice that the correlator across the wormhole is maximal at $u_1 = u_2 = 0$! This means that the wormhole is basically what is depicted in the right side of Figure \ref{fig:infbeta}. Since both circles intersect the asymptotic boundary, we can interpret the figure as either having a single boundary, or two boundaries that are infinitely long, where the two boundaries are separated at asymptotic infinity.
This wormhole solution can be contrasted with the disk solution in the same limit.
At large $(1-\lambda)^2 \beta \gg 1$, the wormhole relation is $\beta_{E} \approx \beta$. On the other hand, the disk in this limit gives $\beta_{E} \approx \beta/2$. This means that for the disk, the 1-sided correlator $G_{\mathsf{LL}}(0,\beta/2+iT) \approx 1$ is large whereas $G_{\mathsf{LR}} =0$. But for the wormhole $G_{\mathsf{LL}}(0,\beta/2+iT)$ is small, but the two-sided correlator $G_{\mathsf{LR}}(0,0) \approx 1$. Such qualitative behavior of the correlators seems to be roughly in agreement with the $q=4, N=20$ numerics, see Figure \ref{fig:l0c}.
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{be-beta-wormh.pdf}
\caption{ The effective inverse temperature $\beta_E$ as a function of $\beta$ in units of $\mathcal{J}$. We show the results for $\lambda = 0.9$ and $\lambda=0.5$. The dashed lines are the Schwarzian predictions given by \eqref{eq:tanb2}. At very large $\beta\mathcal{J}$, the growth is approximately linear $\beta_{E} \approx \beta$. \label{fig:betabeta} }
\end{center}
\end{figure}
Note that our setup has various exact discrete global symmetries, which implies that there are multiple wormhole solutions.
These symmetries are the same as the ones for the ``ramp'' in SYK \cite{Saad:2018bqo}, so we will be brief
First note that $(-1)^F_L$ and $(-1)^F_R$ generate a $\mathbb{Z}_2 \times \mathbb{Z}_2 $ global symmetry,
that leaves $G_{\mathsf{LL}}$ and $G_{\mathsf{RR}}$ invariant but takes $G_{\mathsf{LR}} \to -G_{\mathsf{LR}}$ when only one $(-1)^F$ is applied.
Thus any wormhole solution $G_{LR} \ne 0$ spontaneously breaks one of the two $\mathbb{Z}_2$ symmetries; further, the solutions come in pairs. For the large $q$ solutions, we should write $G_{\mathsf{LR}} = \pm e^{g_{\mathsf{LR}}/q}$
When $q$ is a multiple of 4, there is the additional symmetry $G_{\mathsf{LR}}\left(t, t^{\prime}\right) \rightarrow i G_{\mathsf{LR}}\left(t,-t^{\prime}\right), \quad G_{\mathsf{RL}}\left(t, t^{\prime}\right) \rightarrow i G_{\mathsf{RL}}\left(-t, t^{\prime}\right), \quad G_{\mathsf{RR}}\left(t, t^{\prime}\right) \rightarrow-G_{\mathsf{RR}}\left(-t,-t^{\prime}\right)$ that should be treated similarly.
Another setup in which one can study the wormhole of the journal analytically is Brownian SYK. There, the journal contains a whole function's worth of couplings $\{ J_{ijkl}(t) \}$. See Appendix \ref{app:brown}.
\pagebreak
\section{Reconstruction with erroneous knowledge of the couplings}
The goal of this section is to illustrate how bulk reconstruction using the boundary system is affected by erroneous or incomplete knowledge of the couplings. The discussion will be mostly qualitative. We will keep things simple and use the Petz Lite protocol of \cite{westcoast}.
In the simplest setting of bulk reconstruction, one considers a ``code subspace'' of the boundary Hilbert space that share the same bulk background geometry but differ in the state of the matter. In our case, we consider a code subspace composed of perturbations of the time evolved thermofield double state $| \beta + i T; J \rangle$, where $J$ labels the couplings of the Hamiltonian $H_J$ used to prepare the state. We use $|i ; J \rangle$ to denote the different matter states spanning the code subspace. If we like, we can view $i$ as some species index for a particle in AdS. Note that we are not yet averaging over couplings.
We will take implementing successful bulk reconstruction to mean that there's an operator that can transition between any two states of the code subspace; for $i,j,k,l$ in some orthonormal basis, there exists an operator ${\cal O}_{ji}$ such that
\begin{align}
\langle k; J | {\cal O}_{ji} |l; J \rangle \approx \delta_{kj}\delta_{li}. \label{reccond}
\end{align}
We use the approximate signs $\approx$ to indicate that we are ignoring non-perturbatively small effects from non-factorization wormholes, and also to account for the approximation of using Petz lite. This condition is easily satisfied by constructing the ``global'' $J$ operator,
\begin{align}
{\cal O}^J_{ji} \equiv |j; J \rangle\langle i; J |,
\end{align}
constructed out of the basis states of the code states. This gives the square of the norm of the states:
\begin{align}
\langle k; J | {\cal O}^J_{ji}|l; J \rangle = \eqfig{0.12\columnwidth}{petzSame.pdf} \times \delta_{jk}\delta_{li}
\end{align}
In this equation, the orange and green colors indicate the particle species (or equivalently the index $i$ or $j$).
The setting we want to consider is where we don't know the value of the coupling $J$ in the Hamiltonian that's used in preparing the code subspace, and we are tasked with finding an operator that transitions between two specified code words $i$ and $j$. This is risky because using a wrong value of the coupling, say $J'$ instead of $J$, leads to an eventual breakdown of the reconstruction:
\begin{align}
\langle j; J | {\cal O}^{J'}_{ji} |i; J \rangle = \langle j; J |j; J' \rangle \langle i; J' |i; J \rangle = \eqfig{0.14\columnwidth}{petzDifferent.pdf}
\end{align}
For $J \neq J'$, the right hand side decays in time due to the backreaction of the shocks (depicted in black) created by the sudden change in the boundary conditions. We showed that in the case of JT + free boson BCFT model of section \ref{sec:renyi}, that this decays exponentially
\begin{align}
\langle j; J | {\cal O}^{J'}_{ji} |i; J \rangle \approx \left[\pi \epsilon \over \beta \cosh(\pi T/\beta)\right]^{(J-J')^2 \over 2 \alpha' \pi^2 }.
\end{align}
(We are ignoring the gravitational backreaction of the orange and green particles and quantum Schwarzian effects, hence the $\approx$ sign.) The reconstruction therefore fails at late times, since the states prepared with different couplings become more orthogonal under time evolution. This is the same effect that resulted in a unitarity problem, which hinted at the presence of an island.
If we are given the knowledge of the distribution of the original couplings, then we can do better. A natural procedure is to average the couplings in the operator over this distribution. We can estimate the failure on average by also averaging over the couplings of the system:
\begin{align}
\int dJ dJ' P(J) P(J') \langle j; J | {\cal O}^{J'}_{ji} |i; J \rangle &\approx {1\over 1 + {2 \over \pi^2 \alpha' m^2} \ln\left[ {\beta \over \pi \epsilon} \cosh(\pi T/\beta)\right]} \approx { \beta \pi \alpha' m^2 \over {2 T } } \label{avgPetz}
\end{align}
At first sight it appears that the reconstruction fails, even on average, albeit more slowly than if we don't average. However, we note that this decay only depends on the statistical properties of the couplings, and also weakly on $i, j$ (not shown here). This means we can improve our reconstruction by simply scaling the operator by a time dependent factor $C_{ij, m}(T)$. We will take this to mean that the reconstruction is successful, at least on average.
However, there's another, more severe way in which this operator can fail: the sum over couplings enhances the contribution of a wormhole that connects the bra/ket of the operator/state its ket/bra. This in particular means that ``flipped'' matrix elements which should vanish actually get a contribution:
\begin{align}
\int dJ dJ' P(J) P(J') \langle i; J | {\cal O}^{J'}_{ji} |j; J \rangle \approx \eqfig{0.12\columnwidth}{petzAvgInsideOD.pdf} = \ \ \eqfig{0.25\columnwidth}{petzAvgInsideWormholeOD.pdf}
\end{align}
Averaging over the same value of the boundary coupling is indicated by like colors. At early times, this is a small contribution that only mildly violates our condition \eqref{reccond}, but at large times, when the disk is highly suppressed, it is comparable to the ``good'' matrix elements $\langle j; J | {\cal O}^{J'}_{ji} |i; J \rangle$ that come from the disk (which are decaying). So, if we rescale the operator to set the ``good'' elements $\approx 1$, we will also rescale these ``flipped'' elements $\langle j; J | {\cal O}^{J'}_{ji} |i; J \rangle \sim 1$. Thus we see that this reconstruction attempt fails badly at late times.
Furthermore, even when the disk contribution to the ``good'' matrix elements is small, there is not a substantial contribution from the wormhole since the bulk states in the throat are orthogonal, and winding contributions decay exponentially in $b \sim T$.
One can see graphically that the overlap is zero by considering
\begin{align}
\int dJ dJ' P(J) P(J') &\langle j; J | {\cal O}^{J'}_{ji} |i; J \rangle \approx \eqfig{0.12\columnwidth}{petzAvgInside.pdf} &= \eqfig{0.25\columnwidth}{petzAvgInsideWormhole.pdf} = 0.
\end{align}
The red crosses above indicate the orthogonality of the bulk states. A similar use for wormholes was first pointed out in \cite{westcoast}. However, failure is not guaranteed because the particles can pair up outside the branch cut:
\begin{align}
\int dJ dJ' P(J) P(J') \langle j; J | {\cal O}^{J'}_{ji} |i; J \rangle \approx \eqfig{0.14\columnwidth}{petzAvgOutsideNoJ.pdf}
\end{align}
For such operators outside the horizon, the wormhole and the disk contribution give similar answers, so there are no significant contributions to the ``flipped'' elements and rescaling the operator with time should succeed.
The right interpretation of this is that the entanglement wedge of the boundary develops a blind spot in the bulk; the reconstruction fails because the particle falls into the island. This island is the entanglement wedge of reference purifying the system, namely what we've been calling the journal in the previous sections. With this interpretation, the averaged operator used in \eqref{avgPetz} is nothing but the Petz lite operator, obtained by tracing out the journal,
\begin{align}
O_{ji}^\mathsf{Petz \, Lite} &= \Tr_\mathrm{journal} \left[ \int \, dJ \sqrt{P(J) P(J')} |j; J \rangle \langle i; J' |_\mathsf{Sys} \otimes | J \rangle \langle J' |_\mathsf{journal} \right] \\
&= \int \, dJ P(J) \, |j; J \rangle \langle i; J |_\mathsf{Sys}
\end{align}
We can adapt the above discussion to the case of having partial knowledge by defining a reconstruction controlled by the result of measuring the couplings. This is just a controlled Petz Lite,
\begin{align}
{\cal O}_{ji} = \sum_{\kappa} \underbrace{\sum_{{\mu}} P({\mu}, \kappa)\ketbra{j; \kappa, {\mu}}{i; \kappa, {\mu}}_{\mathsf{sys}}}_{ O_{ji}^\mathsf{Petz \, Lite}(\kappa)} \otimes \ketbra{\kappa}{\kappa}_\mathsf{known}
\end{align}
We note that this can be generalized to the case of full Petz with the replacement
\begin{align}
O_{ji}^\mathsf{Petz \, Lite}(\kappa) \rightarrow \left[\Pi^\mathsf{code}_\mathsf{Sys}(\kappa)\right]^{-{1\over 2}} O_{ji}^\mathsf{Petz \, Lite}(\kappa) \left[\Pi_\mathsf{Sys}^\mathsf{code}(\kappa)\right]^{-{1\over 2} }
\end{align}
where
\begin{align}
\Pi^\mathsf{code}_\mathsf{Sys}(\kappa) \equiv \Tr_{{\mu}, \kappa'} \left[ \Pi_{\kappa'} \Pi^\mathsf{code} \Pi_{\kappa'} \right]
\end{align}
is the trace over the couplings of the projector onto the code subspace when projected onto a given value of the known couplings.
It is interesting to speculate how the failure of reconstruction due to islands manifests itself in a theory with fixed couplings. Given that it arises from the emergence of an island and wormholes (as a result of averaging) it suggests that something like half wormholes \cite{Saad:2021rcu} are important for reconstructing the black hole interior. This suggests that we need to go beyond semi-classical gravity to address the firewall question \cite{Almheiri:2012rt,Almheiri:2013hfa}.
Besides the Petz map, one can also try to use simpler methods of bulk reconstruction. One option is to use the Maldacena-Qi Hamiltonian \cite{Maldacena:2018lmt} which directly couples the left and right sides\footnote{We thank Juan Maldacena for suggesting this direction.}. In \cite{Maldacena:2018lmt}, they showed that the wormhole does not require perfect correlation between couplings. This suggests that at early times, as long as the couplings are not too uncertain, evolution with the coupled Hamiltonian will still lead to an eternal traversable wormhole, which implies that the entanglement wedge of the two sides includes all of AdS. At late times, the wormhole is very long, so the correlations between the two sides is not strong enough to prevent the wormhole from growing. It would be interesting to understand this more quantitatively.
\begin{comment}
\newpage
\AAl{OLd below}
The upshot of this section was to demonstrate the sensitivity of bulk reconstruction on the values of the couplings of the theory. The connection between reconstruction and entanglement wedges, islands, and so forth came about when analyzing how well the reconstruction works on average. For a given realization we expect that
The discussion of this section is meant to reinforce the results of the previous sections, and to give some more intuitive understanding of the reconstruction properties in the presence of uncertainty in the values of the couplings. What we are learning is that the exterior of the black hole is within reach even with coarse knowledge of the couplings, while the interior requires knowing the fine details.
These ideas can be recast in terms of quantum channels as follows. Starting with an arbitrary state of the boundary system, we define an encoding map ${\cal E}$ that encodes a state of an abstract code subspace, $| i\rangle \in {\cal H}_\mathsf{code}$ into the boundary system and that depends on the state of a journal,
\begin{align}
{\cal E}_i(\, \cdot \,) = \sum_{J,J'} U_i^J(\, \cdot \,) U_i^{J' \dagger} \otimes |J \rangle \langle J'|_\mathsf{journal}
\end{align}
where the unitaries satisfy an orthogonality $U_i^{J \dagger}U^J_j = \delta_{i j}$ relation. Hence this prepares orthogonal states $\tilde{\rho}_{i}$ on the union of the boundary system and the journal. This orthogonality allows for a decoding assuming we have access to the journal.
The problem of reconstruction in the presence of known and unknown couplings is a bit more constrained. Uncertainty in the couplings can be introduced through defining a measurement channel ${\cal M}$ that decoheres the couplings that we have access to and traces over the unknown couplings
\begin{align}
{\cal M}(\, \cdot \,) = \Tr_{{\mu}} \circ \sum_{\kappa} \Pi_\kappa (\, \cdot \, ) \Pi_\kappa
\end{align}
This can be thought of as the error channel from the perspective of quantum error correction. The reconstruction condition \eqref{reccond} would follow if ${\cal M}$ maintains the orthogonality of encoded code states,
\begin{align}
\Tr[ {\cal M}(\Tilde{\rho}_i) {\cal M}(\tilde{\rho}_j)] \ll 1
\end{align}
If this condition holds then one can recosntruct a recovery channel ${\cal R}$ that satisfies
\begin{align}
{\cal R} \circ {\cal M} \circ {\cal E}_i (\, \cdot \,) = {\cal E}_i (\, \cdot \,)
\end{align}
We consider the starting point of one state, say the TFD, entangled with a journal that keeps track of the couplings used in preparing the TFD:
\begin{align}
\rho = \sum_{J,J'}
\end{align}
To put these in terms of quantum channels, we start
These ideas can be rephrased in terms of quantum channels. The recovery channel which decodes the code subspace can be implemented as the controlled unitary
\begin{align}
{\cal R}(\, \cdot \,) = \sum_{J} {U}^J\otimes\Pi_J (\, \cdot \, ) { U}^{J\dagger} \otimes \Pi_J
\end{align}
Uncertainty in the couplings can be introduced by passing the state through a combination of a trace over the unknwown couplings, and a decoherence channel over the known,
\begin{align}
{\cal M}(\, \cdot \,) = \Tr_{{\mu}} \circ \sum_{\kappa} \Pi_\kappa (\, \cdot \, ) \Pi_\kappa
\end{align}
The condition for bulk reconstruction is
\eqn{\begin{raggedleft}
{\cal{R}} \circ {\cal M}(\, \cdot \,) =
\end{raggedleft}}
\AAl{I'll finish this in the morning. I'm getting sleepy... Feel free to change whatever.}
The discussion here is meant to reinforce the results of the previous sections, and to give some more intuitive understanding of the reconstruction properties in the presence of uncertainty in the values of the couplings. What we are learning is that the exterior of the black hole is within reach even with coarse knowledge of the couplings, while the interior requires knowing the fine details.
\HL{mention somewhere using traversable wormholes to do bulk reconstruction:}
Besides the Petz map, one can also try to use simpler methods of bulk reconstruction\footnote{We thank Juan Maldacena for suggesting this direction.}. One option is to use the Maldacena-Qi Hamiltonian \cite{Maldacena:2018lmt} which directly couples the left and right sides. In \cite{Maldacena:2018lmt}, they showed that the wormhole does not require perfect correlation between couplings. This suggests that at early times, as long as the couplings are not too uncertain, evolution with the coupled Hamiltonian will still lead to an eternal traversable wormhole, which implies that the entanglement wedge of the two sides includes all of AdS. At late times, the wormhole is very long, so the correlations between the two sides is not strong enough to prevent the wormhole from growing. It would be interesting to understand this more quantitatively.
\end{comment}
\pagebreak
\section{Discussion}
\begin{figure}[H]
\begin{center}
\includegraphics[width=1\columnwidth]{finite-temp.pdf}
\caption{ Island of the journal versus the island in the East coast setup. At late times, the separation between the twists from the island are nearly null separated from the boundary twists. In this OPE limit, one can approximate the entropy as the sum of two semi-infinite intervals discussed in section \ref{sec:peninsula}. \label{fig:EC} }
\end{center}
\end{figure}
\subsection{Bathing in the unknown \label{bathanalogy}}
In the introduction of this paper we motivated the entanglement entropy between the system and the journal as a measure of the uncertainty in the couplings, and we studied its growth under controlled time evolution and its effects on the entanglement wedge of the system.
Here we briefly motivate a different (but physically identical) picture of thinking of the journal as a bath with which the black hole system interacts. This makes the problem more analogous to the East Coast model \cite{eastcoast}, where the black hole interacts with a bath. Just as an example, we can consider the SYK system. The Hamiltonian \footnote{To reproduce a probability distribution over the couplings that is independent of $\beta$, we will generally need to include an explicit $\beta$ dependence in the Hamiltonian.} on the combined SYK and journal systems is
\eqn{H = \sum_{i < j < k < l } \hat{J}_{ijkl} \psi_i \psi_j \psi_k \psi_l + \hat{J}_{ijkl}^2 \frac{N^3}{6 J^2 }}
Here we are using the undergraduate quantum mechanics ``hat'' notation to emphasize that $\hat{J}$ is a quantum mechanical operator. The interaction is quite simple in a sense because there is no conjugate momentum to $\hat{J}$ in the Hamiltonian; however, it is sufficient to transfer quantum information to the journal.
Then the formation of the island for the journal is qualitatively similar to the formation of the island for the radiation, see Figure \ref{fig:EC}.
\subsection{Evaporating BH $+$ journal}
It would be interesting to study how the state of the Hawking radiation of an evaporating black hole is affected by the uncertainty in the couplings. A concrete case would be to consider the ``East coast'' model where we have JT gravity $+$ CFT, joined to a flat space region where the CFT continues. In addition to $\mathsf{CFT}_1$ in the bulk, we could also have a second $\mathsf{CFT}_2$ that lives only in the black hole region. In other words, we have a boundary condition $J$ for $\mathsf{CFT}_2$ that prevents any transmission into the flat space region. Furthermore, we assume that in the semi-classical picture there is no interaction between $\mathsf{CFT}_1$ and $\mathsf{CFT}_2$.
We get information about the state of the radiation by looking for the entanglement wedge of the radiation. In this setup there are two systems, the radiation and the journal, vying for ownership of the island. We can guess at the winner by comparing the rate of entropy growth of the two systems. The first to $~2 S_{BH}$ wins. We found in section \ref{trivialsurface} that the entropy of the journal grows only logarithmically, while it was shown in \cite{Almheiri:2019yqk} that the entropy of the radiation grows linearly time. Hence we believe that the radiation wins the race, at least at first.
An open question is whether ownership of the island is ever transferred to the journal. If it does get transferred, then one interpretation is that large uncertainty in the couplings reduces the system to the semi-classical state; this would support the recent ideas connecting premature ensemble averaging to Hawking's calculation \cite{Marolf:2020xie, Bousso:2019ykv, Bousso:2020kmy}. In the case that the island doesn't get transferred, then a possible interpretation is that the radiation alone purifies the journal. This translates physically to the statement that the couplings can be determined by the measurements on the Hawking radiation.
\subsection{Spectral form factor}
One kind of coupling that always exists is the overall normalization of the Hamiltonian $\lambda$. This case is interesting because we can make contact with the spectral form factor.
If the overall coupling is uncertain, time evolution with $H_{\lambda} = \lambda H$ and inverse time evolution with $H_{\lambda'} = \lambda' H$ will not cancel by an amount $T (\lambda-\lambda')$.
The Renyi-2 entropy of the journal is therefore non-trivial:
\eqn{\tr \rho^2 &= \int p(\lambda) p(\lambda') \left|\tr \left ( e^{i (\lambda-\lambda') H T - \frac{1}{2} (\lambda+\lambda') \beta H} \right )\right|^2 \\
&= \int p(\lambda) p(\lambda') \left |Z\left ( \frac{1}{2} (\lambda+\lambda') \beta + i (\lambda-\lambda') T \right )\right |^2 }
Here we see that the Renyi-2 is essentially the time-averaged spectral form factor. The specific kind of time-averaging depends on $p(\lambda)$.
The fact that the spectral form factor cannot decay to zero is just a consequence of the unitarity bound on this Renyi entropy. Hence the ``journal'' perspective conceptually unifies Maldacena's information paradox \cite{Maldacena:2001kr} with the information paradox of evaporating black holes.
Given that the plateau of the spectral form factor is not explained by a single wormhole but doubly non-perturbative effects \cite{Saad:2018bqo, Saad:2019lba}, one could wonder if such effects play a role in the entropy of the journal.
\subsection{Chaos and the Loschmidt Echo}
The lack of our ability to do bulk reconstruction even when there is a small uncertainty in couplings is closely related to quantum chaos.
In holography, one usually diagnoses chaos using the $\mathsf{otoc}$, e.g, we create a perturbation at some time in the past $W(-T)$, and watch it grow.
The perturbation is localized at some time $-T$.
An alternative diagnostic of chaos is the
\href{http://www.scholarpedia.org/article/Loschmidt_echo}{Loschmidt echo}.
Both diagonistics involve (backwards) time evolution $\exp( i H_1 T)$, but in the Loschmidt echo, one changes the Hamiltonian going forward $\exp( i H_2 T)$ by a small amount $H_2 = H_1 + \epsilon W$.
So whereas the $\mathsf{otoc}$ is a localized pertubation in time, the Loschmidt echo is completely de-localized.
The simplest diagnostic of chaos is the decay of the inner product
\eqn{ \bra{\psi'} \exp( -i H_2 T) \exp(i H_1 T) \ket{\psi} }
Clearly if $H_2 = H_1$ this product does not decay, but in general it should decay and then exhibit some erratic oscillations of order $e^{-S}$.
What we showed was that the disk contribution leads to a decay, but wormholes are needed to explain the erratic oscillations. We showed this for $\ket{\psi} = \ket{\beta, J_1}$, $\ket{\psi'}= \ket{\beta, J_2}$ in SYK and in JT gravity, but we believe that the lessons should be fairly general.
One can also study the size of the Loschmidt operator $\exp( -i H_2 T)\exp(i H_1 T) $. At infinite temperature, this is given by sandwiching the size operator $\sim \psi_L \psi_R$ with the state $\exp( -i H_2 T)\exp(i H_1 T) \ket{0}$. To define a finite temperature version of size, we would need to specify whether the thermal ensemble should correspond to $H_1$ or $H_2$, but for small perturbations this is a minor detail. For large $q$ SYK, it follows from the disk solution at $\lambda \approx 1$ in \ref{sec:largeq} that the size grows exponentially with Loschmidt Lyapunov exponent
\eqn{\lambda_L = 2 \alpha,}
e.g, the same Lyapunov exponent as the thermofield double. At low temperatures, the maximal chaos exponent follows from the bulk picture and the relation between size and the symmetry generators \cite{Lin:2019qwu}. Indeed, in the bulk, a change in the couplings at time $-T$ inserts a matter shockwave; the gravitational backreaction is responsible for the exponential growth in size. But we have calculated the exponent at finite temperature in the large $q$ theory, where SYK is not simply described by the Schwarzian mode.
An interesting question is whether one can argue for a bound on the Loschmidt Lypaunov exponent $\lambda_{L} \le 2\pi/\beta$ along the lines of \cite{Maldacena:2015waa}.
\subsection{Singularity?}
Our work shows that the black hole interior is quite sensitive to the precise values of the couplings of the theory. Even ignorance about the irrelevant couplings appears to be enough to prevent the interior from being reconstructed at late times.
This is surprising if we believe that irrelevant operators have small effects in the bulk. On the other hand, irrelevant operators can have a large effect near the black hole singularity. In JT gravity, one can show that the profile of a free scalar field with mass determined by $\Delta$ will diverge near the inner horizon except for integer values of $\Delta$ \cite{Lin:2019qwu}. It is tempting to speculate that the island associated to the journal that forms at late times is trying to censor us from accessing the region near the singularity. Adopting the spirit of Penrose's cosmic censorship conjecture \cite{penrose1969gravitational}, if we do not have access to the UV couplings of the boundary theory, it seems reasonable that we cannot reconstruct the region near the black hole singularity that is sensitive to those couplings. This obviously deserves more study.
\subsection{Janus's Journal
It would be interesting to consider the journal entangled to a ``conventional'' holographic theory which has no disorder average.
We could consider a situation the journal records the value of $\lambda$ in $\mathcal{N}=4$ SYM. An off diagonal element of the journal density matrix would involve half of the thermal cylinder with one value of the gauge coupling $g$ and the other half with a different value $g'$.
For the vacuum state, this problem is rather well-studied: one considers the Euclidean path integral on a sphere with gauge coupling $g$ on half of the sphere and $g'$ on the other half \cite{Bak:2003jk,Clark:2004sb,Clark:2005te,Gaiotto:2008sd}.
The problem at finite temperature would involve a path integral on a torus, with top and bottom halves of the torus differing \cite{Bak:2007jm}.
It would be interesting to find wormholes in such a setup (or in other holographic setups with well-defined string duals), as this would pose a sharp factorization problem in $\mathcal{N}=4$.
More generally, one can consider a holographic CFT$_1$ joined to a CFT$_2$ at some interface, see \cite{Bachas:2021fqo, Simidzija:2020ukv, May:2021xhz} for some recent discussions. Our paper suggests that the square of such quantites could have wormholes.
\subsection{Conclusion}
We are usually taught that gravity is a universal force, e.g., that spacetime is sourced by energy-momentum. What we seem to be learning here is that the spacetime in the interior of the black hole is in some sense highly non-universal.
This seems related to the following equation:
\eqn{P(\mathsf{firewall}) =\eqfig{0.6\columnwidth}{firewall.pdf} }
\begin{center} {\it What gravity doesn't know, neither do we...}
\end{center}
\section*{Acknowledgments}
We thank Alexey Milekhin, Adam Levine, Venkatesh Chandrasekar, Alex Streicher, Yiming Chen, Juan Maldacena, Edgar Shaghoulian, Douglas Stanford, Edward Witten, and Ying Zhao for discussions.
We thank Douglas Stanford for permission to reproduce his figure as our Figure \ref{fig:stanfordworm}, and Juan Maldacena for repeatedly threatening us to finish the draft.
|
\section{Introduction}
Within the next decade, several $O(\$1{\rm b})$ precision neutrino oscillation experiments
will be coming online.
These experiments target as of yet unknown parameter values associated with
neutrino flavor oscillation, such as the $CP$-violating phase
and the neutrino mass hierarchy.
In addition, neutrino oscillation experiments can search for proton decay events
and detect neutrinos from supernova sources.
The Deep Underground Neutrino Experiment (DUNE)~\cite{DUNE:2020mra}
plans to measure neutrinos over the first oscillation maximum,
which corresponds to an energy range of $1-10~{\rm GeV}$ for a baseline of 1300~km.
Many interaction topologies come into play in this energy range,
with primary neutrino interaction classes including quasielastic scattering,
resonance production, and deep inelastic scattering all in roughly equal proportions.
Quantification of the neutrino energy in events requires good control of cross sections
for all of the relevant interaction topologies and sophisticated nuclear modeling
to reconstruct the final statistical distributions and constrain oscillation parameters.
Of the interaction classes seen in neutrino oscillation experiments,
quasielastic scattering is a primary signal measurement process due to its simplicity.
Quasielastic scattering dominates the total cross section at low neutrino energies,
making it an important contribution for neutrino oscillation experiments such as
HyperK, which will have neutrino energies strongly peaked in the quasielastic regime,
and DUNE, which probes a large range of neutrino energies.
In this interaction class, a neutrino interacts with a
freely propagating nucleon within a nucleus and becomes an outgoing charged lepton.
For these reasons, quasielastic scattering cross section amplitudes have
stringent precision requirements and are a natural target for improving
cross section systematics.
Since the neutrinos interact via a weak current,
both vector and axial vector matrix elements are needed to compute the cross sections.
Unlike the vector matrix elements, the axial current contribution cannot be estimated
from electron-proton scattering experiments.
Constraints on the axial matrix elements must come from either
low statistics experimental measurements on elementary targets,
model-dependent estimations from pion electroproduction,
or large nuclear target neutrino scattering data with nuisance nuclear effects.
The uncertainty of the neutrino cross section amplitudes originating from
nucleon form factors is large enough to be a possible cause of theory-experiment discrepancies.
Reanalysis of neutrino scattering data on elementary targets using the model-independent
$z$ expansion reveals that the ubiquitous dipole model parameterization of the form
factor underestimates the uncertainty by nearly an order of magnitude~\cite{Meyer:2016oeg}.
In the absence of a modern neutrino-deuterium scattering experiment,
the most reasonable approach for quantifying and reducing uncertainties on the
axial form factor is instead to compute nucleon matrix elements using lattice QCD and to feed
the results into nuclear models.
Calculations are constructed to access the free nucleon form factor,
completely circumventing the need for nuclear modeling.
\section{Simulation Details}
The utilized lattice action is the same as used by CalLat to compute $g_A$ with sub-percent precision~\cite{Chang:2018uxx,Berkowitz:2018gqe,Walker-Loud:2019cif}: a mixed lattice action with
M\"obius Domain Wall Fermions (MDWF) in the valence sector, solved
in a sea of Highly-Improved Staggered Quarks (HISQ)~\cite{Follana:2006rc}
and one-loop Symanzik improved gauge action
background field configurations.
The links used in the valence fermion Dirac operator were smeared with the gradient-flow scheme to a flow-time of $t_{gf}/a^2=1$~\cite{Berkowitz:2017opd}. We are computing the axial form factor on approximately 30 ensembles, with seven pion masses ranging from $130\lesssim M_\pi \lesssim 400$~MeV, four lattice spacings of $0.06\lesssim a\lesssim0.15$~fm and multiple volumes~\cite{Miller:2020evg}.
In these proceedings, we present preliminary analysis on one ensemble, denoted a12m130, with $a\approx0.12$~fm and $M_\pi\approx130$~MeV, which was generated by the MILC collaboration~\cite{MILC:2012znn}.
Measurements are performed on 1000~configurations with 32 sources per configuration.
The two-point correlation functions are constructed from propagators
with point source quark fields smeared at both the source and sink
so that the correlation function is positive definite.
Additional quark propagators are sequentially solved with fixed source-sink time separations in the range $t/a\in\{3,...,12\}$, with sink-source spin polarizations up-up and down-down, for both forward time-propagating positive parity projectors and backwards time-propagating negative parity projectors.
The three-point correlation functions are constructed from the proper averaging of these four different correlation functions.
To reduce the cost of these sequential propagators, we use propagators from 8 sources to form a single coherent-sequential-sink~\cite{LHPC:2010jcs} as described in Ref.~\cite{He:2021yvm}.
The sink is projected to zero spatial momentum with the same quark smearing as at the source.
We then inject the current with all spatial momentum for all 16 quark bi-linear currents.
In this proceeding, we focus on the axial current aligned in the $z$ direction and the temporal vector current at 0 momentum (for normalization).
The axial form factor data are plotted as a ratio of a three-point over two-point
correlation function.
The two-point functions are denoted $C^{\text{2pt}}(t,\mathbf{p})$
with source-sink time $t$ and momentum $\mathbf{p}$, and
three-point functions are denoted $C^{\text{3pt}}_{{\cal A}_z}(t,\tau,\mathbf{q})$
with source-sink time $t$, current-insertion time $\tau$,
insertion momentum transfer $\mathbf{q}$, and source momentum $-\mathbf{q}$.
Using these definitions, a correlator ratio that isolates the axial form factor is
\begin{align}\label{eq:ratio}
{\cal R}_{{\cal A}_z}(t,\tau,\mathbf{q}) =&
\frac{
C^{\text{3pt}}_{{\cal A}_z}(t,\tau,\mathbf{q}) }{
\sqrt{ C^{\text{2pt}}(t-\tau,\mathbf{0}) C^{\text{2pt}}(\tau,\mathbf{q}) }
}
\sqrt{
\frac{ C^{\text{2pt}}( \tau,\mathbf{0}) }{ C^{\text{2pt}}(t,\mathbf{0}) }
\frac{ C^{\text{2pt}}(t-\tau,\mathbf{q}) }{ C^{\text{2pt}}(t,\mathbf{q}) }
}
\nonumber\\
\xrightarrow[t-\tau,\tau\to\infty]{}&
\frac{1}{\sqrt{2E_{\mathbf{q}} (E_{\mathbf{q}}+M)}}
\left[
-\frac{q_z^2}{2M} \mathring{\tilde{g}}_P(Q^2)
+(E_{\mathbf{q}}+M) \mathring{g}_A(Q^2)
\right]
\,,
\end{align}
where $\mathring{\tilde{g}}_{P}(Q^2)$ and $\mathring{{g}}_{A}(Q^2)$ are the respective unrenormalized induced pseudo scalar and axial form factors and $Q^2 = 2M^2(\sqrt{1+\frac{\mathbf{q}^2}{M^2}} -1)$.
In particular, for $q_z = 0$, the ground state of this ratio correlator is proportional to the unrenormalized axial form factor up to a computable kinematic factor.
In this analysis, a Bayesian framework is employed to fit the correlators.
The two-point and three-point correlation functions are fit with sums of exponentials with shared parameters,
\begin{align}
&C^{\text{2pt}}(t,\mathbf{p}) = \sum^N_n |z^{\mathbf{p}}_n|^2 e^{-E^{\mathbf{p}}_n t}\, ,&
&C^{\text{3pt}}_{{\cal A}_z}(t,\tau,\mathbf{q}) =
\sum^N_{m,n} z^{\mathbf{0}}_n z^{\mathbf{q}}_m A^{\mathbf{q}}_{nm}
e^{-E^{\mathbf{0}}_n (t-\tau)} e^{-E^{\mathbf{q}}_m \tau}\, ,&
\end{align}
with momentum $\mathbf{p}$ and current insertion momentum $\mathbf{q}$.
A tower of $N$ exponential contributions is included for each state,
with the choice of $N=3$ for all correlators.
The simultaneous fit includes the 0-momentum temporal vector current
in addition to the axial current in order to get the vector charge,
which provides an extra constraint on the spectrum of states.
With our lattice action, $Z_A/Z_V -1 \leq 5\times10^{-5}$ for all ensembles~\cite{Chang:2018uxx} and so
the ratio $\mathring{g}_A(Q^2)/\mathring{g}_V$ is normalized absolutely.
The posterior values for the ground state-to-ground state transitions
at various momentum transfers obtained from the correlation function fit
are the desired nucleon axial form factor data.
The form factor parameterization of choice is the $z$ expansion~\cite{Meyer:2016oeg}
\begin{align}
g_A(Q^2) = \sum_{k=0} a_k \big[ z(Q^2) \big]^k.
\label{eq:gazexp}
\end{align}
with a conformal mapping of the form,
\begin{align}
z(Q^2) = \frac{\sqrt{t_c +Q^2} -\sqrt{t_c -t_0}}{\sqrt{t_c +Q^2} +\sqrt{t_c -t_0}}.
\label{eq:zdef}
\end{align}
The parameter $t_c$ is a kinematic cutoff ($t_c=9M_\pi^2$ for the axial current)
and $t_0$ is a free parameter that may be set to improve convergence.
The form in Eq.~(\ref{eq:zdef}) ensures that $|z|<1$ for quasielastic scattering
so that the expansion parameter $z$ that appears in Eq.~(\ref{eq:gazexp})
is guaranteed to be small.
The sum in Eq.~(\ref{eq:gazexp}) is in practice truncated at finite order $k_{\rm max}$,
so the large-$Q^2$ behavior is controlled by including extra parameters
for $k=k_{\rm max}+n+1$ and enforcing sum rules of the form
\begin{align}
\biggr( \frac{\partial}{\partial z} \biggr)^{n}
\; \sum_{k=0}^{k_{\rm max}+4} a_k z^k \Big|_{z=1} = 0
\label{eq:sumrules}
\end{align}
with $n\in\{0,1,2,3\}$.
The coefficients $a_k$ in the form factor fit are given priors
\begin{align}
{\rm prior}\biggr[ \frac{a_k}{|a_0|} \biggr] = 0 \pm {\rm min} \biggr[ 5, \frac{25}{k} \biggr]
\end{align}
as done in Ref.~\cite{Meyer:2016oeg}.
\section{Results}
The ranges of the fit time have been chosen such that the minimum time,
$t_{\rm min}$, is consistent across both the two-point and three-point functions:
the two-point correlators are fit to the time range $t \in [t_{\rm min}, t_{\rm max,2}]$
and the three-point correlators with source-sink separation $t$ to the range
$\tau \in [t_{\rm min}, t-t_{\rm min}]$.
This choice ensures that the minimum time separation between any two operator insertions
is at least $t_{\rm min}/a$ timeslices apart.
All available three-point data that satisfy these temporal restrictions are included in the fit.
All discrete 3-momenta that satisfy
$|q_{x,y}| \leq 4\sqrt{2}\cdot (2\pi/L)$ and $q_z=0$
are simultaneously fit to extract posteriors,
which corresponds to four-momentum transfers up to about $1.06~{\rm GeV}$~\cite{Miller:2020evg}.
\newcommand{0.3\textwidth}{0.3\textwidth}
\begin{figure}[tbh!]
\begin{tikzpicture}
\node at (0,0) {
\setlength\tabcolsep{-1em}
\begin{tabular}{cccc}
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px0py0pz0-standalone.pdf} &
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px1py0pz0-standalone.pdf} &
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px1py1pz0-standalone.pdf} &
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px2py0pz0-standalone.pdf} \\[-2em]
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px2py1pz0-standalone.pdf} &
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px2py2pz0-standalone.pdf} &
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px3py0pz0-standalone.pdf} &
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px3py1pz0-standalone.pdf} \\[-2em]
&
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px3py2pz0-standalone.pdf} &
\includegraphics[width=0.3\textwidth]{standalone/pdm-a12m130-px3py3pz0-standalone.pdf} \\
\end{tabular}
};
\node[opacity=0.1,gray] at (0,0) {\scalebox{5.4}[20]{PRELIMINARY}};
\end{tikzpicture}
\caption{
Plot of the ratio in Eq.~(\ref{eq:ratio}) formed from the raw correlator data.
The horizontal axis is the source-insertion separation time minus half
the source-sink separation time to center the data.
The gray band is the posterior value for $\mathring{g}_A(Q^2)$
after a 3-exponential correlator fit, including all $m\rightarrow n$ transitions with $0\leq m,n<3$.
The data used in the fit are shown as filled-in circles and data outside the fit range
are shown as unfilled circles.
Different source-sink separations are plotted as different colors,
ranging from the shortest source-sink separation $t/a=3$ (purple)
to the longest source-sink separation with $t/a=12$ (red).
The posterior curves obtained from an exponential fit is plotted as the colored band,
with the color corresponding to the same source-sink separation as the data.
The different panels correspond to different 3-momentum transfers squared,
in units of the minimum lattice momentum $2\pi/L$ squared, for the lowest 10 momenta used in the analysis.
Since the data are preliminary, the units on the vertical axis are omitted.
\label{fig:pdm}
}
\end{figure}
The data and posterior fits are combined with the ratio written in Eq.~(\ref{eq:ratio}),
which is plotted in Fig.~\ref{fig:pdm} for a sub-set of the correlators.
This figure includes the lowest 10 momenta with $q_z=0$,
and all of the available three-point data are plotted with their appropriate ratios.
The fits to the correlator data show broad agreement across all ratio values plotted,
down to time separations as small as $t,\tau,(t-\tau)=2a$.
There are several qualities that suggest large excited-state contaminations at low-momentum transfers.
There is a strong curvature with $\tau/a$ at low momentum transfer,
even for the largest source-sink time separation (red), with a curvature that changes sign.
The smallest source-sink time separations are relatively smaller than the largest
source-sink time separations at low momentum, but the opposite is true at large momenta.
The gray band corresponding to the posterior matrix element connecting ground states
of ingoing and outgoing momenta is as much as $2\sigma$ from the ratio central value.
These qualities are suggestive of large excited state contaminations
in the low-momentum transfer data, particularly the axial charge.
As the momentum is increased, the agreement between the ground state posterior gray band
and the ratio value of the largest source-sink separation come into agreement.
The ratio data for the largest source-sink separation times fall on top of each other,
suggesting that the most harmful excited state contaminations have decayed away
and only the ground state contribution remains.
These two observations give confidence that the excited states are reasonably
controlled, and the posteriors give the axial matrix elements of interest.
The fit to the axial matrix elements is shown in Fig.~\ref{fig:gaq2}.
The green (lighter top) curve is a 5-parameter fit to the scatter points
with a $z$ expansion parameterization including 4 sum rule constraints,
which gives a good description of the data.
The gray (darker lower) curve is a 2-parameter dipole fit.
The precision of the data is particularly good,
with roughly constant uncertainties for
all of considered momentum transfers.
\renewcommand{0.3\textwidth}{0.75\textwidth}
\begin{figure}[tbh!]
\centering
\begin{tikzpicture}
\node at (0,0) {
\includegraphics[width=0.3\textwidth]{standalone/a12m130-A3-zexp-standalone.pdf}
};
\node[red] at (-2.0,-0.8) {\scalebox{1.0}{PRELIMINARY}};
\end{tikzpicture}
\caption{
Plot of the renormalized axial form factor as a function of $Q^2$.
The scatter points are the posteriors from the form factor obtained after fitting
to the correlation functions with their uncertainties.
The upper (green) band is the $z$ expansion fit to the scatter points.
The lower (gray) band is the dipole fit.
The dashed line is a dipole parameterization with $M_A=1.026$~GeV~\cite{Bernard:2001rs}.
Since the data are preliminary, the vertical axis scale is omitted
and the lower axis bound is set to a nonzero value.
\label{fig:gaq2}
}
\end{figure}
The form factor fits in this analysis achieve sub-percent precision on the form factor.
In particular, the axial charge is constrained with a relative precision of $0.6\%$, slightly better than CalLat's axial charge analysis~\cite{Chang:2018uxx,Berkowitz:2018gqe,Walker-Loud:2019cif}.
The form factor has a similar absolute precision out to larger momentum transfers.
Taking the axial radius squared as a definitive metric for the axial form factor,
with the definition
\begin{align}
r_A^2 = -\frac{6}{g_A(0)}
\frac{d g_A(Q^2)}{d Q^2} \Big|_{Q^2=0},
\end{align}
a relative precision of $\delta r_A^2 / r_A^2 \approx 0.13$ is achieved,
more than a factor of 3 more precise than the radius obtained from
neutrino scattering on deuterium~\cite{Meyer:2016oeg}.
In addition, the axial radius obtained is smaller than observed in experiments.
This points to a slower falloff with $Q^2$ than expected,
a trend that is consistent with other LQCD extractions.
This finding indicates that more weight should be given to larger
momentum transfer neutrino interactions,
which could change the relative frequency of quasielastic scattering
relative to other interaction topologies.
\section{Outlook}
Lattice QCD has the potential to make an impact on the neutrino oscillation
program by providing significantly more precise nucleon form factors
for weak interaction amplitudes.
Existing analyses of form factor data have the potential to reach percent-level
constraints on the form factors with a full error budget,
an order of magnitude more precise than the axial form factor constraints
obtained from neutrino scattering on elementary targets.
With this level of precision,
and barring existing tensions in the vector form factor parameterizations~\cite{Borah:2020gte},
nucleon-level form factor uncertainties can be made subdominant compared
to nuclear modeling uncertainties,
enabling robust inputs with realistic uncertainties for nuclear model calculations.
Excited states in the axial matrix elements have always been a difficulty
for nucleon matrix elements involving an axial current.
Incomplete characterization of excited states are believed to have historically
led to extractions of the axial charge that are low compared to experiments.
A significant contamination comes from transitions of the ground state nucleon
to excited states, possibly including multiparticle states involving
a nucleon and a pion.
Inclusion of many source-sink time separations down to short times
has been shown to be beneficial for control of excited states,
leading to more precise estimates of ground state
matrix element information~\cite{He:2021yvm}.
This analysis has included at least 8 source-sink separation times over 10 momenta
and has demonstrated reasonable control over a 3-state fit for all momentum combinations.
The axial form factor data on this single ensemble look very promising,
with an expected percent-level precision.
Scaling of uncertainties with the momentum transfer is better than expected
when considering 4-momentum transfers squared up to $0.7~{\rm GeV}^2$,
which use only a small fraction of the available data.
Data for 3-momentum-squared up to 5 times larger than those considered are available,
which would permit constraints on the form factor of $Q^2$ up to a few ${\rm GeV}^2$.
These data further feed back to better constraints on the ground state spectrum
and matrix elements, providing strong constraints on the axial charge
rest-frame spectrum.
\section{Acknowledgements}
This work was supported in part by
the NVIDIA Corporation (MAC),
the Alexander von Humboldt Foundation through a Feodor Lynen Research Fellowship (CK),
the RIKEN Special Postdoctoral Researcher Program (ER),
the Nuclear Physics Double Beta Decay Topical Collaboration (HMC, AN, AWL),
the U.S. Department of Energy, Office of Science, Office of Nuclear Physics under Award
Numbers DE-AC02-05CH11231 (CCC, CK, BH, AWL),
DEAC52-07NA27344 (DH, PV),
DE-FG02-93ER-40762 (EB),
DE-SC00046548 (ASM);
the DOE Early Career Award Program (AWL),
and the U.K. Science and Technology Facilities Council grants ST/S005781/1 and ST/T000945/1 (CB).
Computing time for this work was provided through the
Innovative and Novel Computational Impact on Theory and Experiment (INCITE) program
and the LLNL Multi programmatic and Institutional Computing program
for Grand Challenge allocations on the LLNL supercomputers.
This research utilized the NVIDIA GPU accelerated Summit supercomputer
at Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory,
which is supported by the Office of Science of the
U.S. Department of Energy under Contract No. DEAC05-00OR22725
as well as the Lassen supercomputer at Lawrence Livermore National Laboratory.
The computations were performed with \texttt{LaLiBe}~\cite{lalibe}, linked against \texttt{Chroma}~\cite{Edwards:2004sx} with \texttt{QUDA} solvers~\cite{Clark:2009wm,Babich:2011np} and HDF5~\cite{hdf5} for I/O~\cite{Kurth:2015mqa}.
They were efficiently managed with \texttt{METAQ}~\cite{Berkowitz:2017vcp,Berkowitz:2017xna} and EspressoDB~\cite{Chang:2019khk}.
The numerical analysis utilized \texttt{gvar}~\cite{gvar} and \texttt{lsqfit}~\cite{lsqfit}.
\bibliographystyle{JHEP}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.